lustre/obdclass/lu_object.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdclass/lu_object.c
  37  *
  38  * Lustre Object.
  39  * These are the only exported functions, they provide some generic
  40  * infrastructure for managing object devices
  41  *
  42  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  43  */
  44
  45 #define DEBUG_SUBSYSTEM S_CLASS
  46
  47 #include <libcfs/libcfs.h>
  48 #include <linux/module.h>
  49 #include <libcfs/libcfs_hash.h> /* hash_long() */
  50 #include <obd_class.h>
  51 #include <obd_support.h>
  52 #include <lustre_disk.h>
  53 #include <lustre_fid.h>
  54 #include <lu_object.h>
  55 #include <lu_ref.h>
  56 #include <libcfs/list.h>
  57
  58 enum {
  59         LU_CACHE_PERCENT_MAX     = 50,
  60         LU_CACHE_PERCENT_DEFAULT = 20
  61 };
  62
  63 #define LU_CACHE_NR_MAX_ADJUST          128
  64 #define LU_CACHE_NR_UNLIMITED           -1
  65 #define LU_CACHE_NR_DEFAULT             LU_CACHE_NR_UNLIMITED
  66 #define LU_CACHE_NR_LDISKFS_LIMIT       LU_CACHE_NR_UNLIMITED
  67 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
  68 #define LU_CACHE_NR_ZFS_LIMIT           10240
  69
  70 #define LU_SITE_BITS_MIN    12
  71 #define LU_SITE_BITS_MAX    24
  72 /**
  73  * total 256 buckets, we don't want too many buckets because:
  74  * - consume too much memory
  75  * - avoid unbalanced LRU list
  76  */
  77 #define LU_SITE_BKT_BITS    8
  78
  79
  80 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
  81 CFS_MODULE_PARM(lu_cache_percent, "i", int, 0644,
  82                 "Percentage of memory to be used as lu_object cache");
  83
  84 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
  85 CFS_MODULE_PARM(lu_cache_nr, "l", long, 0644,
  86                 "Maximum number of objects in lu_object cache");
  87
  88 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
  89
  90 /**
  91  * Decrease reference counter on object. If last reference is freed, return
  92  * object to the cache, unless lu_object_is_dying(o) holds. In the latter
  93  * case, free object immediately.
  94  */
  95 void lu_object_put(const struct lu_env *env, struct lu_object *o)
  96 {
  97         struct lu_site_bkt_data *bkt;
  98         struct lu_object_header *top;
  99         struct lu_site          *site;
 100         struct lu_object        *orig;
 101         cfs_hash_bd_t            bd;
 102         const struct lu_fid     *fid;
 103
 104         top  = o->lo_header;
 105         site = o->lo_dev->ld_site;
 106         orig = o;
 107
 108         /*
 109          * till we have full fids-on-OST implemented anonymous objects
 110          * are possible in OSP. such an object isn't listed in the site
 111          * so we should not remove it from the site.
 112          */
 113         fid = lu_object_fid(o);
 114         if (fid_is_zero(fid)) {
 115                 LASSERT(top->loh_hash.next == NULL
 116                         && top->loh_hash.pprev == NULL);
 117                 LASSERT(list_empty(&top->loh_lru));
 118                 if (!atomic_dec_and_test(&top->loh_ref))
 119                         return;
 120                 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 121                         if (o->lo_ops->loo_object_release != NULL)
 122                                 o->lo_ops->loo_object_release(env, o);
 123                 }
 124                 lu_object_free(env, orig);
 125                 return;
 126         }
 127
 128         cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
 129         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
 130
 131         if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
 132                 if (lu_object_is_dying(top)) {
 133
 134                         /*
 135                          * somebody may be waiting for this, currently only
 136                          * used for cl_object, see cl_object_put_last().
 137                          */
 138                         wake_up_all(&bkt->lsb_marche_funebre);
 139                 }
 140                 return;
 141         }
 142
 143         LASSERT(bkt->lsb_busy > 0);
 144         bkt->lsb_busy--;
 145         /*
 146          * When last reference is released, iterate over object
 147          * layers, and notify them that object is no longer busy.
 148          */
 149         list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 150                 if (o->lo_ops->loo_object_release != NULL)
 151                         o->lo_ops->loo_object_release(env, o);
 152         }
 153
 154         if (!lu_object_is_dying(top)) {
 155                 LASSERT(list_empty(&top->loh_lru));
 156                 list_add_tail(&top->loh_lru, &bkt->lsb_lru);
 157                 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 158                 return;
 159         }
 160
 161         /*
 162          * If object is dying (will not be cached), removed it
 163          * from hash table and LRU.
 164          *
 165          * This is done with hash table and LRU lists locked. As the only
 166          * way to acquire first reference to previously unreferenced
 167          * object is through hash-table lookup (lu_object_find()),
 168          * or LRU scanning (lu_site_purge()), that are done under hash-table
 169          * and LRU lock, no race with concurrent object lookup is possible
 170          * and we can safely destroy object below.
 171          */
 172         if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
 173                 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
 174         cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 175         /*
 176          * Object was already removed from hash and lru above, can
 177          * kill it.
 178          */
 179         lu_object_free(env, orig);
 180 }
 181 EXPORT_SYMBOL(lu_object_put);
 182
 183 /**
 184  * Put object and don't keep in cache. This is temporary solution for
 185  * multi-site objects when its layering is not constant.
 186  */
 187 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
 188 {
 189         set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
 190         return lu_object_put(env, o);
 191 }
 192 EXPORT_SYMBOL(lu_object_put_nocache);
 193
 194 /**
 195  * Kill the object and take it out of LRU cache.
 196  * Currently used by client code for layout change.
 197  */
 198 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
 199 {
 200         struct lu_object_header *top;
 201
 202         top = o->lo_header;
 203         set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
 204         if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
 205                 cfs_hash_t *obj_hash = o->lo_dev->ld_site->ls_obj_hash;
 206                 cfs_hash_bd_t bd;
 207
 208                 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
 209                 list_del_init(&top->loh_lru);
 210                 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
 211                 cfs_hash_bd_unlock(obj_hash, &bd, 1);
 212         }
 213 }
 214 EXPORT_SYMBOL(lu_object_unhash);
 215
 216 /**
 217  * Allocate new object.
 218  *
 219  * This follows object creation protocol, described in the comment within
 220  * struct lu_device_operations definition.
 221  */
 222 static struct lu_object *lu_object_alloc(const struct lu_env *env,
 223                                          struct lu_device *dev,
 224                                          const struct lu_fid *f,
 225                                          const struct lu_object_conf *conf)
 226 {
 227         struct lu_object *scan;
 228         struct lu_object *top;
 229         struct list_head *layers;
 230         unsigned int init_mask = 0;
 231         unsigned int init_flag;
 232         int clean;
 233         int result;
 234         ENTRY;
 235
 236         /*
 237          * Create top-level object slice. This will also create
 238          * lu_object_header.
 239          */
 240         top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
 241         if (top == NULL)
 242                 RETURN(ERR_PTR(-ENOMEM));
 243         if (IS_ERR(top))
 244                 RETURN(top);
 245         /*
 246          * This is the only place where object fid is assigned. It's constant
 247          * after this point.
 248          */
 249         top->lo_header->loh_fid = *f;
 250         layers = &top->lo_header->loh_layers;
 251
 252         do {
 253                 /*
 254                  * Call ->loo_object_init() repeatedly, until no more new
 255                  * object slices are created.
 256                  */
 257                 clean = 1;
 258                 init_flag = 1;
 259                 list_for_each_entry(scan, layers, lo_linkage) {
 260                         if (init_mask & init_flag)
 261                                 goto next;
 262                         clean = 0;
 263                         scan->lo_header = top->lo_header;
 264                         result = scan->lo_ops->loo_object_init(env, scan, conf);
 265                         if (result != 0) {
 266                                 lu_object_free(env, top);
 267                                 RETURN(ERR_PTR(result));
 268                         }
 269                         init_mask |= init_flag;
 270 next:
 271                         init_flag <<= 1;
 272                 }
 273         } while (!clean);
 274
 275         list_for_each_entry_reverse(scan, layers, lo_linkage) {
 276                 if (scan->lo_ops->loo_object_start != NULL) {
 277                         result = scan->lo_ops->loo_object_start(env, scan);
 278                         if (result != 0) {
 279                                 lu_object_free(env, top);
 280                                 RETURN(ERR_PTR(result));
 281                         }
 282                 }
 283         }
 284
 285         lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
 286         RETURN(top);
 287 }
 288
 289 /**
 290  * Free an object.
 291  */
 292 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
 293 {
 294         struct lu_site_bkt_data *bkt;
 295         struct lu_site          *site;
 296         struct lu_object        *scan;
 297         struct list_head        *layers;
 298         struct list_head         splice;
 299
 300         site   = o->lo_dev->ld_site;
 301         layers = &o->lo_header->loh_layers;
 302         bkt    = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
 303         /*
 304          * First call ->loo_object_delete() method to release all resources.
 305          */
 306         list_for_each_entry_reverse(scan, layers, lo_linkage) {
 307                 if (scan->lo_ops->loo_object_delete != NULL)
 308                         scan->lo_ops->loo_object_delete(env, scan);
 309         }
 310
 311         /*
 312          * Then, splice object layers into stand-alone list, and call
 313          * ->loo_object_free() on all layers to free memory. Splice is
 314          * necessary, because lu_object_header is freed together with the
 315          * top-level slice.
 316          */
 317         INIT_LIST_HEAD(&splice);
 318         list_splice_init(layers, &splice);
 319         while (!list_empty(&splice)) {
 320                 /*
 321                  * Free layers in bottom-to-top order, so that object header
 322                  * lives as long as possible and ->loo_object_free() methods
 323                  * can look at its contents.
 324                  */
 325                 o = container_of0(splice.prev, struct lu_object, lo_linkage);
 326                 list_del_init(&o->lo_linkage);
 327                 LASSERT(o->lo_ops->loo_object_free != NULL);
 328                 o->lo_ops->loo_object_free(env, o);
 329         }
 330
 331         if (waitqueue_active(&bkt->lsb_marche_funebre))
 332                 wake_up_all(&bkt->lsb_marche_funebre);
 333 }
 334
 335 /**
 336  * Free \a nr objects from the cold end of the site LRU list.
 337  */
 338 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
 339 {
 340         struct lu_object_header *h;
 341         struct lu_object_header *temp;
 342         struct lu_site_bkt_data *bkt;
 343         cfs_hash_bd_t            bd;
 344         cfs_hash_bd_t            bd2;
 345         struct list_head         dispose;
 346         int                      did_sth;
 347         unsigned int             start;
 348         int                      count;
 349         int                      bnr;
 350         unsigned int             i;
 351
 352         if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
 353                 RETURN(0);
 354
 355         INIT_LIST_HEAD(&dispose);
 356         /*
 357          * Under LRU list lock, scan LRU list and move unreferenced objects to
 358          * the dispose list, removing them from LRU and hash table.
 359          */
 360         start = s->ls_purge_start;
 361         bnr = (nr == ~0) ? -1 : nr / (int)CFS_HASH_NBKT(s->ls_obj_hash) + 1;
 362  again:
 363         /*
 364          * It doesn't make any sense to make purge threads parallel, that can
 365          * only bring troubles to us. See LU-5331.
 366          */
 367         mutex_lock(&s->ls_purge_mutex);
 368         did_sth = 0;
 369         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 370                 if (i < start)
 371                         continue;
 372                 count = bnr;
 373                 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
 374                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 375
 376                 list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
 377                         LASSERT(atomic_read(&h->loh_ref) == 0);
 378
 379                         cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
 380                         LASSERT(bd.bd_bucket == bd2.bd_bucket);
 381
 382                         cfs_hash_bd_del_locked(s->ls_obj_hash,
 383                                                &bd2, &h->loh_hash);
 384                         list_move(&h->loh_lru, &dispose);
 385                         if (did_sth == 0)
 386                                 did_sth = 1;
 387
 388                         if (nr != ~0 && --nr == 0)
 389                                 break;
 390
 391                         if (count > 0 && --count == 0)
 392                                 break;
 393
 394                 }
 395                 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
 396                 cond_resched();
 397                 /*
 398                  * Free everything on the dispose list. This is safe against
 399                  * races due to the reasons described in lu_object_put().
 400                  */
 401                 while (!list_empty(&dispose)) {
 402                         h = container_of0(dispose.next,
 403                                           struct lu_object_header, loh_lru);
 404                         list_del_init(&h->loh_lru);
 405                         lu_object_free(env, lu_object_top(h));
 406                         lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
 407                 }
 408
 409                 if (nr == 0)
 410                         break;
 411         }
 412         mutex_unlock(&s->ls_purge_mutex);
 413
 414         if (nr != 0 && did_sth && start != 0) {
 415                 start = 0; /* restart from the first bucket */
 416                 goto again;
 417         }
 418         /* race on s->ls_purge_start, but nobody cares */
 419         s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
 420
 421         return nr;
 422 }
 423 EXPORT_SYMBOL(lu_site_purge);
 424
 425 /*
 426  * Object printing.
 427  *
 428  * Code below has to jump through certain loops to output object description
 429  * into libcfs_debug_msg-based log. The problem is that lu_object_print()
 430  * composes object description from strings that are parts of _lines_ of
 431  * output (i.e., strings that are not terminated by newline). This doesn't fit
 432  * very well into libcfs_debug_msg() interface that assumes that each message
 433  * supplied to it is a self-contained output line.
 434  *
 435  * To work around this, strings are collected in a temporary buffer
 436  * (implemented as a value of lu_cdebug_key key), until terminating newline
 437  * character is detected.
 438  *
 439  */
 440
 441 enum {
 442         /**
 443          * Maximal line size.
 444          *
 445          * XXX overflow is not handled correctly.
 446          */
 447         LU_CDEBUG_LINE = 512
 448 };
 449
 450 struct lu_cdebug_data {
 451         /**
 452          * Temporary buffer.
 453          */
 454         char lck_area[LU_CDEBUG_LINE];
 455 };
 456
 457 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
 458 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
 459
 460 /**
 461  * Key, holding temporary buffer. This key is registered very early by
 462  * lu_global_init().
 463  */
 464 struct lu_context_key lu_global_key = {
 465         .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
 466                     LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
 467         .lct_init = lu_global_key_init,
 468         .lct_fini = lu_global_key_fini
 469 };
 470
 471 /**
 472  * Printer function emitting messages through libcfs_debug_msg().
 473  */
 474 int lu_cdebug_printer(const struct lu_env *env,
 475                       void *cookie, const char *format, ...)
 476 {
 477         struct libcfs_debug_msg_data *msgdata = cookie;
 478         struct lu_cdebug_data        *key;
 479         int used;
 480         int complete;
 481         va_list args;
 482
 483         va_start(args, format);
 484
 485         key = lu_context_key_get(&env->le_ctx, &lu_global_key);
 486         LASSERT(key != NULL);
 487
 488         used = strlen(key->lck_area);
 489         complete = format[strlen(format) - 1] == '\n';
 490         /*
 491          * Append new chunk to the buffer.
 492          */
 493         vsnprintf(key->lck_area + used,
 494                   ARRAY_SIZE(key->lck_area) - used, format, args);
 495         if (complete) {
 496                 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
 497                         libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
 498                 key->lck_area[0] = 0;
 499         }
 500         va_end(args);
 501         return 0;
 502 }
 503 EXPORT_SYMBOL(lu_cdebug_printer);
 504
 505 /**
 506  * Print object header.
 507  */
 508 void lu_object_header_print(const struct lu_env *env, void *cookie,
 509                             lu_printer_t printer,
 510                             const struct lu_object_header *hdr)
 511 {
 512         (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
 513                    hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
 514                    PFID(&hdr->loh_fid),
 515                    hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
 516                    list_empty((struct list_head *)&hdr->loh_lru) ? \
 517                    "" : " lru",
 518                    hdr->loh_attr & LOHA_EXISTS ? " exist" : "");
 519 }
 520 EXPORT_SYMBOL(lu_object_header_print);
 521
 522 /**
 523  * Print human readable representation of the \a o to the \a printer.
 524  */
 525 void lu_object_print(const struct lu_env *env, void *cookie,
 526                      lu_printer_t printer, const struct lu_object *o)
 527 {
 528         static const char ruler[] = "........................................";
 529         struct lu_object_header *top;
 530         int depth = 4;
 531
 532         top = o->lo_header;
 533         lu_object_header_print(env, cookie, printer, top);
 534         (*printer)(env, cookie, "{\n");
 535
 536         list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 537                 /*
 538                  * print `.' \a depth times followed by type name and address
 539                  */
 540                 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
 541                            o->lo_dev->ld_type->ldt_name, o);
 542
 543                 if (o->lo_ops->loo_object_print != NULL)
 544                         (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
 545
 546                 (*printer)(env, cookie, "\n");
 547         }
 548
 549         (*printer)(env, cookie, "} header@%p\n", top);
 550 }
 551 EXPORT_SYMBOL(lu_object_print);
 552
 553 /**
 554  * Check object consistency.
 555  */
 556 int lu_object_invariant(const struct lu_object *o)
 557 {
 558         struct lu_object_header *top;
 559
 560         top = o->lo_header;
 561         list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 562                 if (o->lo_ops->loo_object_invariant != NULL &&
 563                     !o->lo_ops->loo_object_invariant(o))
 564                         return 0;
 565         }
 566         return 1;
 567 }
 568 EXPORT_SYMBOL(lu_object_invariant);
 569
 570 static struct lu_object *htable_lookup(struct lu_site *s,
 571                                        cfs_hash_bd_t *bd,
 572                                        const struct lu_fid *f,
 573                                        wait_queue_t *waiter,
 574                                        __u64 *version)
 575 {
 576         struct lu_site_bkt_data *bkt;
 577         struct lu_object_header *h;
 578         struct hlist_node       *hnode;
 579         __u64  ver = cfs_hash_bd_version_get(bd);
 580
 581         if (*version == ver)
 582                 return ERR_PTR(-ENOENT);
 583
 584         *version = ver;
 585         bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
 586         /* cfs_hash_bd_peek_locked is a somehow "internal" function
 587          * of cfs_hash, it doesn't add refcount on object. */
 588         hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
 589         if (hnode == NULL) {
 590                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
 591                 return ERR_PTR(-ENOENT);
 592         }
 593
 594         h = container_of0(hnode, struct lu_object_header, loh_hash);
 595         if (likely(!lu_object_is_dying(h))) {
 596                 cfs_hash_get(s->ls_obj_hash, hnode);
 597                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
 598                 list_del_init(&h->loh_lru);
 599                 return lu_object_top(h);
 600         }
 601
 602         /*
 603          * Lookup found an object being destroyed this object cannot be
 604          * returned (to assure that references to dying objects are eventually
 605          * drained), and moreover, lookup has to wait until object is freed.
 606          */
 607
 608         if (likely(waiter != NULL)) {
 609                 init_waitqueue_entry_current(waiter);
 610                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 611                 set_current_state(TASK_UNINTERRUPTIBLE);
 612                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
 613         }
 614
 615         return ERR_PTR(-EAGAIN);
 616 }
 617
 618 static struct lu_object *htable_lookup_nowait(struct lu_site *s,
 619                                               cfs_hash_bd_t *bd,
 620                                               const struct lu_fid *f)
 621 {
 622         struct hlist_node       *hnode;
 623         struct lu_object_header *h;
 624
 625         /* cfs_hash_bd_peek_locked is a somehow "internal" function
 626          * of cfs_hash, it doesn't add refcount on object. */
 627         hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
 628         if (hnode == NULL) {
 629                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
 630                 return ERR_PTR(-ENOENT);
 631         }
 632
 633         h = container_of0(hnode, struct lu_object_header, loh_hash);
 634         if (unlikely(lu_object_is_dying(h)))
 635                 return ERR_PTR(-ENOENT);
 636
 637         cfs_hash_get(s->ls_obj_hash, hnode);
 638         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
 639         list_del_init(&h->loh_lru);
 640         return lu_object_top(h);
 641 }
 642
 643 /**
 644  * Search cache for an object with the fid \a f. If such object is found,
 645  * return it. Otherwise, create new object, insert it into cache and return
 646  * it. In any case, additional reference is acquired on the returned object.
 647  */
 648 struct lu_object *lu_object_find(const struct lu_env *env,
 649                                  struct lu_device *dev, const struct lu_fid *f,
 650                                  const struct lu_object_conf *conf)
 651 {
 652         return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
 653 }
 654 EXPORT_SYMBOL(lu_object_find);
 655
 656 /*
 657  * Limit the lu_object cache to a maximum of lu_cache_nr objects.  Because
 658  * the calculation for the number of objects to reclaim is not covered by
 659  * a lock the maximum number of objects is capped by LU_CACHE_MAX_ADJUST.
 660  * This ensures that many concurrent threads will not accidentally purge
 661  * the entire cache.
 662  */
 663 static void lu_object_limit(const struct lu_env *env,
 664                             struct lu_device *dev)
 665 {
 666         __u64 size, nr;
 667
 668         if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
 669                 return;
 670
 671         size = cfs_hash_size_get(dev->ld_site->ls_obj_hash);
 672         nr = (__u64)lu_cache_nr;
 673         if (size > nr)
 674                 lu_site_purge(env, dev->ld_site,
 675                               MIN(size - nr, LU_CACHE_NR_MAX_ADJUST));
 676
 677         return;
 678 }
 679
 680 static struct lu_object *lu_object_new(const struct lu_env *env,
 681                                        struct lu_device *dev,
 682                                        const struct lu_fid *f,
 683                                        const struct lu_object_conf *conf)
 684 {
 685         struct lu_object        *o;
 686         cfs_hash_t              *hs;
 687         cfs_hash_bd_t            bd;
 688         struct lu_site_bkt_data *bkt;
 689
 690         o = lu_object_alloc(env, dev, f, conf);
 691         if (unlikely(IS_ERR(o)))
 692                 return o;
 693
 694         hs = dev->ld_site->ls_obj_hash;
 695         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 696         bkt = cfs_hash_bd_extra_get(hs, &bd);
 697         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 698         bkt->lsb_busy++;
 699         cfs_hash_bd_unlock(hs, &bd, 1);
 700
 701         lu_object_limit(env, dev);
 702
 703         return o;
 704 }
 705
 706 /**
 707  * Core logic of lu_object_find*() functions.
 708  */
 709 static struct lu_object *lu_object_find_try(const struct lu_env *env,
 710                                             struct lu_device *dev,
 711                                             const struct lu_fid *f,
 712                                             const struct lu_object_conf *conf,
 713                                             wait_queue_t *waiter)
 714 {
 715         struct lu_object      *o;
 716         struct lu_object      *shadow;
 717         struct lu_site        *s;
 718         cfs_hash_t            *hs;
 719         cfs_hash_bd_t          bd;
 720         __u64                  version = 0;
 721
 722         /*
 723          * This uses standard index maintenance protocol:
 724          *
 725          *     - search index under lock, and return object if found;
 726          *     - otherwise, unlock index, allocate new object;
 727          *     - lock index and search again;
 728          *     - if nothing is found (usual case), insert newly created
 729          *       object into index;
 730          *     - otherwise (race: other thread inserted object), free
 731          *       object just allocated.
 732          *     - unlock index;
 733          *     - return object.
 734          *
 735          * For "LOC_F_NEW" case, we are sure the object is new established.
 736          * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
 737          * just alloc and insert directly.
 738          *
 739          * If dying object is found during index search, add @waiter to the
 740          * site wait-queue and return ERR_PTR(-EAGAIN).
 741          */
 742         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
 743                 return lu_object_new(env, dev, f, conf);
 744
 745         s  = dev->ld_site;
 746         hs = s->ls_obj_hash;
 747         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 748         o = htable_lookup(s, &bd, f, waiter, &version);
 749         cfs_hash_bd_unlock(hs, &bd, 1);
 750         if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT)
 751                 return o;
 752
 753         /*
 754          * Allocate new object. This may result in rather complicated
 755          * operations, including fld queries, inode loading, etc.
 756          */
 757         o = lu_object_alloc(env, dev, f, conf);
 758         if (unlikely(IS_ERR(o)))
 759                 return o;
 760
 761         LASSERT(lu_fid_eq(lu_object_fid(o), f));
 762
 763         cfs_hash_bd_lock(hs, &bd, 1);
 764
 765         shadow = htable_lookup(s, &bd, f, waiter, &version);
 766         if (likely(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT)) {
 767                 struct lu_site_bkt_data *bkt;
 768
 769                 bkt = cfs_hash_bd_extra_get(hs, &bd);
 770                 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 771                 bkt->lsb_busy++;
 772                 cfs_hash_bd_unlock(hs, &bd, 1);
 773
 774                 lu_object_limit(env, dev);
 775
 776                 return o;
 777         }
 778
 779         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
 780         cfs_hash_bd_unlock(hs, &bd, 1);
 781         lu_object_free(env, o);
 782         return shadow;
 783 }
 784
 785 /**
 786  * Much like lu_object_find(), but top level device of object is specifically
 787  * \a dev rather than top level device of the site. This interface allows
 788  * objects of different "stacking" to be created within the same site.
 789  */
 790 struct lu_object *lu_object_find_at(const struct lu_env *env,
 791                                     struct lu_device *dev,
 792                                     const struct lu_fid *f,
 793                                     const struct lu_object_conf *conf)
 794 {
 795         struct lu_site_bkt_data *bkt;
 796         struct lu_object        *obj;
 797         wait_queue_t           wait;
 798
 799         if (conf != NULL && conf->loc_flags & LOC_F_NOWAIT)
 800                 return lu_object_find_try(env, dev, f, conf, NULL);
 801
 802         while (1) {
 803                 obj = lu_object_find_try(env, dev, f, conf, &wait);
 804                 if (obj != ERR_PTR(-EAGAIN))
 805                         return obj;
 806                 /*
 807                  * lu_object_find_try() already added waiter into the
 808                  * wait queue.
 809                  */
 810                 waitq_wait(&wait, TASK_UNINTERRUPTIBLE);
 811                 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
 812                 remove_wait_queue(&bkt->lsb_marche_funebre, &wait);
 813         }
 814 }
 815 EXPORT_SYMBOL(lu_object_find_at);
 816
 817 /**
 818  * Try to find the object in cache without waiting for the dead object
 819  * to be released nor allocating object if no cached one was found.
 820  *
 821  * The found object will be set as LU_OBJECT_HEARD_BANSHEE for purging.
 822  */
 823 void lu_object_purge(const struct lu_env *env, struct lu_device *dev,
 824                      const struct lu_fid *f)
 825 {
 826         struct lu_site          *s  = dev->ld_site;
 827         cfs_hash_t              *hs = s->ls_obj_hash;
 828         cfs_hash_bd_t            bd;
 829         struct lu_object        *o;
 830
 831         cfs_hash_bd_get_and_lock(hs, f, &bd, 1);
 832         o = htable_lookup_nowait(s, &bd, f);
 833         cfs_hash_bd_unlock(hs, &bd, 1);
 834         if (!IS_ERR(o)) {
 835                 set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
 836                 lu_object_put(env, o);
 837         }
 838 }
 839 EXPORT_SYMBOL(lu_object_purge);
 840
 841 /**
 842  * Find object with given fid, and return its slice belonging to given device.
 843  */
 844 struct lu_object *lu_object_find_slice(const struct lu_env *env,
 845                                        struct lu_device *dev,
 846                                        const struct lu_fid *f,
 847                                        const struct lu_object_conf *conf)
 848 {
 849         struct lu_object *top;
 850         struct lu_object *obj;
 851
 852         top = lu_object_find(env, dev, f, conf);
 853         if (IS_ERR(top))
 854                 return top;
 855
 856         obj = lu_object_locate(top->lo_header, dev->ld_type);
 857         if (unlikely(obj == NULL)) {
 858                 lu_object_put(env, top);
 859                 obj = ERR_PTR(-ENOENT);
 860         }
 861
 862         return obj;
 863 }
 864 EXPORT_SYMBOL(lu_object_find_slice);
 865
 866 /**
 867  * Global list of all device types.
 868  */
 869 static struct list_head lu_device_types;
 870
 871 int lu_device_type_init(struct lu_device_type *ldt)
 872 {
 873         int result = 0;
 874
 875         atomic_set(&ldt->ldt_device_nr, 0);
 876         INIT_LIST_HEAD(&ldt->ldt_linkage);
 877         if (ldt->ldt_ops->ldto_init)
 878                 result = ldt->ldt_ops->ldto_init(ldt);
 879
 880         if (result == 0) {
 881                 spin_lock(&obd_types_lock);
 882                 list_add(&ldt->ldt_linkage, &lu_device_types);
 883                 spin_unlock(&obd_types_lock);
 884         }
 885
 886         return result;
 887 }
 888 EXPORT_SYMBOL(lu_device_type_init);
 889
 890 void lu_device_type_fini(struct lu_device_type *ldt)
 891 {
 892         spin_lock(&obd_types_lock);
 893         list_del_init(&ldt->ldt_linkage);
 894         spin_unlock(&obd_types_lock);
 895         if (ldt->ldt_ops->ldto_fini)
 896                 ldt->ldt_ops->ldto_fini(ldt);
 897 }
 898 EXPORT_SYMBOL(lu_device_type_fini);
 899
 900 /**
 901  * Global list of all sites on this node
 902  */
 903 static struct list_head lu_sites;
 904 static DEFINE_MUTEX(lu_sites_guard);
 905
 906 /**
 907  * Global environment used by site shrinker.
 908  */
 909 static struct lu_env lu_shrink_env;
 910
 911 struct lu_site_print_arg {
 912         struct lu_env   *lsp_env;
 913         void            *lsp_cookie;
 914         lu_printer_t     lsp_printer;
 915 };
 916
 917 static int
 918 lu_site_obj_print(cfs_hash_t *hs, cfs_hash_bd_t *bd,
 919                   struct hlist_node *hnode, void *data)
 920 {
 921         struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
 922         struct lu_object_header  *h;
 923
 924         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
 925         if (!list_empty(&h->loh_layers)) {
 926                 const struct lu_object *o;
 927
 928                 o = lu_object_top(h);
 929                 lu_object_print(arg->lsp_env, arg->lsp_cookie,
 930                                 arg->lsp_printer, o);
 931         } else {
 932                 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
 933                                        arg->lsp_printer, h);
 934         }
 935         return 0;
 936 }
 937
 938 /**
 939  * Print all objects in \a s.
 940  */
 941 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
 942                    lu_printer_t printer)
 943 {
 944         struct lu_site_print_arg arg = {
 945                 .lsp_env     = (struct lu_env *)env,
 946                 .lsp_cookie  = cookie,
 947                 .lsp_printer = printer,
 948         };
 949
 950         cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
 951 }
 952 EXPORT_SYMBOL(lu_site_print);
 953
 954 /**
 955  * Return desired hash table order.
 956  */
 957 static unsigned long lu_htable_order(struct lu_device *top)
 958 {
 959         unsigned long cache_size;
 960         unsigned long bits;
 961
 962         /*
 963          * For ZFS based OSDs the cache should be disabled by default.  This
 964          * allows the ZFS ARC maximum flexibility in determining what buffers
 965          * to cache.  If Lustre has objects or buffer which it wants to ensure
 966          * always stay cached it must maintain a hold on them.
 967          */
 968         if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
 969                 lu_cache_percent = 1;
 970                 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
 971                 return LU_SITE_BITS_MIN;
 972         }
 973
 974         /*
 975          * Calculate hash table size, assuming that we want reasonable
 976          * performance when 20% of total memory is occupied by cache of
 977          * lu_objects.
 978          *
 979          * Size of lu_object is (arbitrary) taken as 1K (together with inode).
 980          */
 981         cache_size = totalram_pages;
 982
 983 #if BITS_PER_LONG == 32
 984         /* limit hashtable size for lowmem systems to low RAM */
 985         if (cache_size > 1 << (30 - PAGE_CACHE_SHIFT))
 986                 cache_size = 1 << (30 - PAGE_CACHE_SHIFT) * 3 / 4;
 987 #endif
 988
 989         /* clear off unreasonable cache setting. */
 990         if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
 991                 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
 992                       " the range of (0, %u]. Will use default value: %u.\n",
 993                       lu_cache_percent, LU_CACHE_PERCENT_MAX,
 994                       LU_CACHE_PERCENT_DEFAULT);
 995
 996                 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 997         }
 998         cache_size = cache_size / 100 * lu_cache_percent *
 999                 (PAGE_CACHE_SIZE / 1024);
1000
1001         for (bits = 1; (1 << bits) < cache_size; ++bits) {
1002                 ;
1003         }
1004         return bits;
1005 }
1006
1007 static unsigned lu_obj_hop_hash(cfs_hash_t *hs,
1008                                 const void *key, unsigned mask)
1009 {
1010         struct lu_fid  *fid = (struct lu_fid *)key;
1011         __u32           hash;
1012
1013         hash = fid_flatten32(fid);
1014         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
1015         hash = hash_long(hash, hs->hs_bkt_bits);
1016
1017         /* give me another random factor */
1018         hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
1019
1020         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
1021         hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
1022
1023         return hash & mask;
1024 }
1025
1026 static void *lu_obj_hop_object(struct hlist_node *hnode)
1027 {
1028         return hlist_entry(hnode, struct lu_object_header, loh_hash);
1029 }
1030
1031 static void *lu_obj_hop_key(struct hlist_node *hnode)
1032 {
1033         struct lu_object_header *h;
1034
1035         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1036         return &h->loh_fid;
1037 }
1038
1039 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
1040 {
1041         struct lu_object_header *h;
1042
1043         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1044         return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
1045 }
1046
1047 static void lu_obj_hop_get(cfs_hash_t *hs, struct hlist_node *hnode)
1048 {
1049         struct lu_object_header *h;
1050
1051         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1052         if (atomic_add_return(1, &h->loh_ref) == 1) {
1053                 struct lu_site_bkt_data *bkt;
1054                 cfs_hash_bd_t            bd;
1055
1056                 cfs_hash_bd_get(hs, &h->loh_fid, &bd);
1057                 bkt = cfs_hash_bd_extra_get(hs, &bd);
1058                 bkt->lsb_busy++;
1059         }
1060 }
1061
1062 static void lu_obj_hop_put_locked(cfs_hash_t *hs, struct hlist_node *hnode)
1063 {
1064         LBUG(); /* we should never called it */
1065 }
1066
1067 cfs_hash_ops_t lu_site_hash_ops = {
1068         .hs_hash        = lu_obj_hop_hash,
1069         .hs_key         = lu_obj_hop_key,
1070         .hs_keycmp      = lu_obj_hop_keycmp,
1071         .hs_object      = lu_obj_hop_object,
1072         .hs_get         = lu_obj_hop_get,
1073         .hs_put_locked  = lu_obj_hop_put_locked,
1074 };
1075
1076 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
1077 {
1078         spin_lock(&s->ls_ld_lock);
1079         if (list_empty(&d->ld_linkage))
1080                 list_add(&d->ld_linkage, &s->ls_ld_linkage);
1081         spin_unlock(&s->ls_ld_lock);
1082 }
1083 EXPORT_SYMBOL(lu_dev_add_linkage);
1084
1085 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
1086 {
1087         spin_lock(&s->ls_ld_lock);
1088         list_del_init(&d->ld_linkage);
1089         spin_unlock(&s->ls_ld_lock);
1090 }
1091 EXPORT_SYMBOL(lu_dev_del_linkage);
1092
1093 /**
1094   * Initialize site \a s, with \a d as the top level device.
1095   */
1096 int lu_site_init(struct lu_site *s, struct lu_device *top)
1097 {
1098         struct lu_site_bkt_data *bkt;
1099         cfs_hash_bd_t bd;
1100         char name[16];
1101         unsigned long bits;
1102         unsigned int i;
1103         ENTRY;
1104
1105         memset(s, 0, sizeof *s);
1106         mutex_init(&s->ls_purge_mutex);
1107         bits = lu_htable_order(top);
1108         snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
1109         for (bits = clamp_t(typeof(bits), bits,
1110                             LU_SITE_BITS_MIN, LU_SITE_BITS_MAX);
1111              bits >= LU_SITE_BITS_MIN; bits--) {
1112                 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1113                                                  bits - LU_SITE_BKT_BITS,
1114                                                  sizeof(*bkt), 0, 0,
1115                                                  &lu_site_hash_ops,
1116                                                  CFS_HASH_SPIN_BKTLOCK |
1117                                                  CFS_HASH_NO_ITEMREF |
1118                                                  CFS_HASH_DEPTH |
1119                                                  CFS_HASH_ASSERT_EMPTY |
1120                                                  CFS_HASH_COUNTER);
1121                 if (s->ls_obj_hash != NULL)
1122                         break;
1123         }
1124
1125         if (s->ls_obj_hash == NULL) {
1126                 CERROR("failed to create lu_site hash with bits: %lu\n", bits);
1127                 return -ENOMEM;
1128         }
1129
1130         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1131                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1132                 INIT_LIST_HEAD(&bkt->lsb_lru);
1133                 init_waitqueue_head(&bkt->lsb_marche_funebre);
1134         }
1135
1136         s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1137         if (s->ls_stats == NULL) {
1138                 cfs_hash_putref(s->ls_obj_hash);
1139                 s->ls_obj_hash = NULL;
1140                 return -ENOMEM;
1141         }
1142
1143         lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1144                              0, "created", "created");
1145         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1146                              0, "cache_hit", "cache_hit");
1147         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1148                              0, "cache_miss", "cache_miss");
1149         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1150                              0, "cache_race", "cache_race");
1151         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1152                              0, "cache_death_race", "cache_death_race");
1153         lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1154                              0, "lru_purged", "lru_purged");
1155
1156         INIT_LIST_HEAD(&s->ls_linkage);
1157         s->ls_top_dev = top;
1158         top->ld_site = s;
1159         lu_device_get(top);
1160         lu_ref_add(&top->ld_reference, "site-top", s);
1161
1162         INIT_LIST_HEAD(&s->ls_ld_linkage);
1163         spin_lock_init(&s->ls_ld_lock);
1164
1165         lu_dev_add_linkage(s, top);
1166
1167         RETURN(0);
1168 }
1169 EXPORT_SYMBOL(lu_site_init);
1170
1171 /**
1172  * Finalize \a s and release its resources.
1173  */
1174 void lu_site_fini(struct lu_site *s)
1175 {
1176         mutex_lock(&lu_sites_guard);
1177         list_del_init(&s->ls_linkage);
1178         mutex_unlock(&lu_sites_guard);
1179
1180         if (s->ls_obj_hash != NULL) {
1181                 cfs_hash_putref(s->ls_obj_hash);
1182                 s->ls_obj_hash = NULL;
1183         }
1184
1185         if (s->ls_top_dev != NULL) {
1186                 s->ls_top_dev->ld_site = NULL;
1187                 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1188                 lu_device_put(s->ls_top_dev);
1189                 s->ls_top_dev = NULL;
1190         }
1191
1192         if (s->ls_stats != NULL)
1193                 lprocfs_free_stats(&s->ls_stats);
1194 }
1195 EXPORT_SYMBOL(lu_site_fini);
1196
1197 /**
1198  * Called when initialization of stack for this site is completed.
1199  */
1200 int lu_site_init_finish(struct lu_site *s)
1201 {
1202         int result;
1203         mutex_lock(&lu_sites_guard);
1204         result = lu_context_refill(&lu_shrink_env.le_ctx);
1205         if (result == 0)
1206                 list_add(&s->ls_linkage, &lu_sites);
1207         mutex_unlock(&lu_sites_guard);
1208         return result;
1209 }
1210 EXPORT_SYMBOL(lu_site_init_finish);
1211
1212 /**
1213  * Acquire additional reference on device \a d
1214  */
1215 void lu_device_get(struct lu_device *d)
1216 {
1217         atomic_inc(&d->ld_ref);
1218 }
1219 EXPORT_SYMBOL(lu_device_get);
1220
1221 /**
1222  * Release reference on device \a d.
1223  */
1224 void lu_device_put(struct lu_device *d)
1225 {
1226         LASSERT(atomic_read(&d->ld_ref) > 0);
1227         atomic_dec(&d->ld_ref);
1228 }
1229 EXPORT_SYMBOL(lu_device_put);
1230
1231 /**
1232  * Initialize device \a d of type \a t.
1233  */
1234 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1235 {
1236         if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1237             t->ldt_ops->ldto_start != NULL)
1238                 t->ldt_ops->ldto_start(t);
1239
1240         memset(d, 0, sizeof *d);
1241         d->ld_type = t;
1242         lu_ref_init(&d->ld_reference);
1243         INIT_LIST_HEAD(&d->ld_linkage);
1244
1245         return 0;
1246 }
1247 EXPORT_SYMBOL(lu_device_init);
1248
1249 /**
1250  * Finalize device \a d.
1251  */
1252 void lu_device_fini(struct lu_device *d)
1253 {
1254         struct lu_device_type *t = d->ld_type;
1255
1256         if (d->ld_obd != NULL) {
1257                 d->ld_obd->obd_lu_dev = NULL;
1258                 d->ld_obd = NULL;
1259         }
1260
1261         lu_ref_fini(&d->ld_reference);
1262         LASSERTF(atomic_read(&d->ld_ref) == 0,
1263                  "Refcount is %u\n", atomic_read(&d->ld_ref));
1264         LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1265
1266         if (atomic_dec_and_test(&t->ldt_device_nr) &&
1267             t->ldt_ops->ldto_stop != NULL)
1268                 t->ldt_ops->ldto_stop(t);
1269 }
1270 EXPORT_SYMBOL(lu_device_fini);
1271
1272 /**
1273  * Initialize object \a o that is part of compound object \a h and was created
1274  * by device \a d.
1275  */
1276 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1277                    struct lu_device *d)
1278 {
1279         memset(o, 0, sizeof(*o));
1280         o->lo_header = h;
1281         o->lo_dev = d;
1282         lu_device_get(d);
1283         lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1284         INIT_LIST_HEAD(&o->lo_linkage);
1285
1286         return 0;
1287 }
1288 EXPORT_SYMBOL(lu_object_init);
1289
1290 /**
1291  * Finalize object and release its resources.
1292  */
1293 void lu_object_fini(struct lu_object *o)
1294 {
1295         struct lu_device *dev = o->lo_dev;
1296
1297         LASSERT(list_empty(&o->lo_linkage));
1298
1299         if (dev != NULL) {
1300                 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1301                               "lu_object", o);
1302                 lu_device_put(dev);
1303                 o->lo_dev = NULL;
1304         }
1305 }
1306 EXPORT_SYMBOL(lu_object_fini);
1307
1308 /**
1309  * Add object \a o as first layer of compound object \a h
1310  *
1311  * This is typically called by the ->ldo_object_alloc() method of top-level
1312  * device.
1313  */
1314 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1315 {
1316         list_move(&o->lo_linkage, &h->loh_layers);
1317 }
1318 EXPORT_SYMBOL(lu_object_add_top);
1319
1320 /**
1321  * Add object \a o as a layer of compound object, going after \a before.
1322  *
1323  * This is typically called by the ->ldo_object_alloc() method of \a
1324  * before->lo_dev.
1325  */
1326 void lu_object_add(struct lu_object *before, struct lu_object *o)
1327 {
1328         list_move(&o->lo_linkage, &before->lo_linkage);
1329 }
1330 EXPORT_SYMBOL(lu_object_add);
1331
1332 /**
1333  * Initialize compound object.
1334  */
1335 int lu_object_header_init(struct lu_object_header *h)
1336 {
1337         memset(h, 0, sizeof *h);
1338         atomic_set(&h->loh_ref, 1);
1339         INIT_HLIST_NODE(&h->loh_hash);
1340         INIT_LIST_HEAD(&h->loh_lru);
1341         INIT_LIST_HEAD(&h->loh_layers);
1342         lu_ref_init(&h->loh_reference);
1343         return 0;
1344 }
1345 EXPORT_SYMBOL(lu_object_header_init);
1346
1347 /**
1348  * Finalize compound object.
1349  */
1350 void lu_object_header_fini(struct lu_object_header *h)
1351 {
1352         LASSERT(list_empty(&h->loh_layers));
1353         LASSERT(list_empty(&h->loh_lru));
1354         LASSERT(hlist_unhashed(&h->loh_hash));
1355         lu_ref_fini(&h->loh_reference);
1356 }
1357 EXPORT_SYMBOL(lu_object_header_fini);
1358
1359 /**
1360  * Given a compound object, find its slice, corresponding to the device type
1361  * \a dtype.
1362  */
1363 struct lu_object *lu_object_locate(struct lu_object_header *h,
1364                                    const struct lu_device_type *dtype)
1365 {
1366         struct lu_object *o;
1367
1368         list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1369                 if (o->lo_dev->ld_type == dtype)
1370                         return o;
1371         }
1372         return NULL;
1373 }
1374 EXPORT_SYMBOL(lu_object_locate);
1375
1376 /**
1377  * Finalize and free devices in the device stack.
1378  *
1379  * Finalize device stack by purging object cache, and calling
1380  * lu_device_type_operations::ldto_device_fini() and
1381  * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1382  */
1383 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1384 {
1385         struct lu_site   *site = top->ld_site;
1386         struct lu_device *scan;
1387         struct lu_device *next;
1388
1389         lu_site_purge(env, site, ~0);
1390         for (scan = top; scan != NULL; scan = next) {
1391                 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1392                 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1393                 lu_device_put(scan);
1394         }
1395
1396         /* purge again. */
1397         lu_site_purge(env, site, ~0);
1398
1399         for (scan = top; scan != NULL; scan = next) {
1400                 const struct lu_device_type *ldt = scan->ld_type;
1401                 struct obd_type             *type;
1402
1403                 next = ldt->ldt_ops->ldto_device_free(env, scan);
1404                 type = ldt->ldt_obd_type;
1405                 if (type != NULL) {
1406                         type->typ_refcnt--;
1407                         class_put_type(type);
1408                 }
1409         }
1410 }
1411 EXPORT_SYMBOL(lu_stack_fini);
1412
1413 enum {
1414         /**
1415          * Maximal number of tld slots.
1416          */
1417         LU_CONTEXT_KEY_NR = 40
1418 };
1419
1420 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1421
1422 static DEFINE_SPINLOCK(lu_keys_guard);
1423
1424 /**
1425  * Global counter incremented whenever key is registered, unregistered,
1426  * revived or quiesced. This is used to void unnecessary calls to
1427  * lu_context_refill(). No locking is provided, as initialization and shutdown
1428  * are supposed to be externally serialized.
1429  */
1430 static unsigned key_set_version = 0;
1431
1432 /**
1433  * Register new key.
1434  */
1435 int lu_context_key_register(struct lu_context_key *key)
1436 {
1437         int result;
1438         unsigned int i;
1439
1440         LASSERT(key->lct_init != NULL);
1441         LASSERT(key->lct_fini != NULL);
1442         LASSERT(key->lct_tags != 0);
1443         LASSERT(key->lct_owner != NULL);
1444
1445         result = -ENFILE;
1446         spin_lock(&lu_keys_guard);
1447         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1448                 if (lu_keys[i] == NULL) {
1449                         key->lct_index = i;
1450                         atomic_set(&key->lct_used, 1);
1451                         lu_keys[i] = key;
1452                         lu_ref_init(&key->lct_reference);
1453                         result = 0;
1454                         ++key_set_version;
1455                         break;
1456                 }
1457         }
1458         spin_unlock(&lu_keys_guard);
1459         return result;
1460 }
1461 EXPORT_SYMBOL(lu_context_key_register);
1462
1463 static void key_fini(struct lu_context *ctx, int index)
1464 {
1465         if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1466                 struct lu_context_key *key;
1467
1468                 key = lu_keys[index];
1469                 LASSERT(key != NULL);
1470                 LASSERT(key->lct_fini != NULL);
1471                 LASSERT(atomic_read(&key->lct_used) > 1);
1472
1473                 key->lct_fini(ctx, key, ctx->lc_value[index]);
1474                 lu_ref_del(&key->lct_reference, "ctx", ctx);
1475                 atomic_dec(&key->lct_used);
1476
1477                 LASSERT(key->lct_owner != NULL);
1478                 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1479                         LINVRNT(module_refcount(key->lct_owner) > 0);
1480                         module_put(key->lct_owner);
1481                 }
1482                 ctx->lc_value[index] = NULL;
1483         }
1484 }
1485
1486 /**
1487  * Deregister key.
1488  */
1489 void lu_context_key_degister(struct lu_context_key *key)
1490 {
1491         LASSERT(atomic_read(&key->lct_used) >= 1);
1492         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1493
1494         lu_context_key_quiesce(key);
1495
1496         ++key_set_version;
1497         spin_lock(&lu_keys_guard);
1498         key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1499         if (lu_keys[key->lct_index]) {
1500                 lu_keys[key->lct_index] = NULL;
1501                 lu_ref_fini(&key->lct_reference);
1502         }
1503         spin_unlock(&lu_keys_guard);
1504
1505         LASSERTF(atomic_read(&key->lct_used) == 1,
1506                  "key has instances: %d\n",
1507                  atomic_read(&key->lct_used));
1508 }
1509 EXPORT_SYMBOL(lu_context_key_degister);
1510
1511 /**
1512  * Register a number of keys. This has to be called after all keys have been
1513  * initialized by a call to LU_CONTEXT_KEY_INIT().
1514  */
1515 int lu_context_key_register_many(struct lu_context_key *k, ...)
1516 {
1517         struct lu_context_key *key = k;
1518         va_list args;
1519         int result;
1520
1521         va_start(args, k);
1522         do {
1523                 result = lu_context_key_register(key);
1524                 if (result)
1525                         break;
1526                 key = va_arg(args, struct lu_context_key *);
1527         } while (key != NULL);
1528         va_end(args);
1529
1530         if (result != 0) {
1531                 va_start(args, k);
1532                 while (k != key) {
1533                         lu_context_key_degister(k);
1534                         k = va_arg(args, struct lu_context_key *);
1535                 }
1536                 va_end(args);
1537         }
1538
1539         return result;
1540 }
1541 EXPORT_SYMBOL(lu_context_key_register_many);
1542
1543 /**
1544  * De-register a number of keys. This is a dual to
1545  * lu_context_key_register_many().
1546  */
1547 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1548 {
1549         va_list args;
1550
1551         va_start(args, k);
1552         do {
1553                 lu_context_key_degister(k);
1554                 k = va_arg(args, struct lu_context_key*);
1555         } while (k != NULL);
1556         va_end(args);
1557 }
1558 EXPORT_SYMBOL(lu_context_key_degister_many);
1559
1560 /**
1561  * Revive a number of keys.
1562  */
1563 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1564 {
1565         va_list args;
1566
1567         va_start(args, k);
1568         do {
1569                 lu_context_key_revive(k);
1570                 k = va_arg(args, struct lu_context_key*);
1571         } while (k != NULL);
1572         va_end(args);
1573 }
1574 EXPORT_SYMBOL(lu_context_key_revive_many);
1575
1576 /**
1577  * Quiescent a number of keys.
1578  */
1579 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1580 {
1581         va_list args;
1582
1583         va_start(args, k);
1584         do {
1585                 lu_context_key_quiesce(k);
1586                 k = va_arg(args, struct lu_context_key*);
1587         } while (k != NULL);
1588         va_end(args);
1589 }
1590 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1591
1592 /**
1593  * Return value associated with key \a key in context \a ctx.
1594  */
1595 void *lu_context_key_get(const struct lu_context *ctx,
1596                          const struct lu_context_key *key)
1597 {
1598         LINVRNT(ctx->lc_state == LCS_ENTERED);
1599         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1600         LASSERT(lu_keys[key->lct_index] == key);
1601         return ctx->lc_value[key->lct_index];
1602 }
1603 EXPORT_SYMBOL(lu_context_key_get);
1604
1605 /**
1606  * List of remembered contexts. XXX document me.
1607  */
1608 static struct list_head lu_context_remembered;
1609
1610 /**
1611  * Destroy \a key in all remembered contexts. This is used to destroy key
1612  * values in "shared" contexts (like service threads), when a module owning
1613  * the key is about to be unloaded.
1614  */
1615 void lu_context_key_quiesce(struct lu_context_key *key)
1616 {
1617         struct lu_context *ctx;
1618         extern unsigned cl_env_cache_purge(unsigned nr);
1619
1620         if (!(key->lct_tags & LCT_QUIESCENT)) {
1621                 /*
1622                  * XXX layering violation.
1623                  */
1624                 cl_env_cache_purge(~0);
1625                 key->lct_tags |= LCT_QUIESCENT;
1626                 /*
1627                  * XXX memory barrier has to go here.
1628                  */
1629                 spin_lock(&lu_keys_guard);
1630                 list_for_each_entry(ctx, &lu_context_remembered,
1631                                     lc_remember)
1632                         key_fini(ctx, key->lct_index);
1633                 spin_unlock(&lu_keys_guard);
1634                 ++key_set_version;
1635         }
1636 }
1637 EXPORT_SYMBOL(lu_context_key_quiesce);
1638
1639 void lu_context_key_revive(struct lu_context_key *key)
1640 {
1641         key->lct_tags &= ~LCT_QUIESCENT;
1642         ++key_set_version;
1643 }
1644 EXPORT_SYMBOL(lu_context_key_revive);
1645
1646 static void keys_fini(struct lu_context *ctx)
1647 {
1648         unsigned int i;
1649
1650         if (ctx->lc_value == NULL)
1651                 return;
1652
1653         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1654                 key_fini(ctx, i);
1655
1656         OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1657         ctx->lc_value = NULL;
1658 }
1659
1660 static int keys_fill(struct lu_context *ctx)
1661 {
1662         unsigned int i;
1663
1664         LINVRNT(ctx->lc_value != NULL);
1665         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1666                 struct lu_context_key *key;
1667
1668                 key = lu_keys[i];
1669                 if (ctx->lc_value[i] == NULL && key != NULL &&
1670                     (key->lct_tags & ctx->lc_tags) &&
1671                     /*
1672                      * Don't create values for a LCT_QUIESCENT key, as this
1673                      * will pin module owning a key.
1674                      */
1675                     !(key->lct_tags & LCT_QUIESCENT)) {
1676                         void *value;
1677
1678                         LINVRNT(key->lct_init != NULL);
1679                         LINVRNT(key->lct_index == i);
1680
1681                         value = key->lct_init(ctx, key);
1682                         if (unlikely(IS_ERR(value)))
1683                                 return PTR_ERR(value);
1684
1685                         LASSERT(key->lct_owner != NULL);
1686                         if (!(ctx->lc_tags & LCT_NOREF))
1687                                 try_module_get(key->lct_owner);
1688                         lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1689                         atomic_inc(&key->lct_used);
1690                         /*
1691                          * This is the only place in the code, where an
1692                          * element of ctx->lc_value[] array is set to non-NULL
1693                          * value.
1694                          */
1695                         ctx->lc_value[i] = value;
1696                         if (key->lct_exit != NULL)
1697                                 ctx->lc_tags |= LCT_HAS_EXIT;
1698                 }
1699                 ctx->lc_version = key_set_version;
1700         }
1701         return 0;
1702 }
1703
1704 static int keys_init(struct lu_context *ctx)
1705 {
1706         OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1707         if (likely(ctx->lc_value != NULL))
1708                 return keys_fill(ctx);
1709
1710         return -ENOMEM;
1711 }
1712
1713 /**
1714  * Initialize context data-structure. Create values for all keys.
1715  */
1716 int lu_context_init(struct lu_context *ctx, __u32 tags)
1717 {
1718         int     rc;
1719
1720         memset(ctx, 0, sizeof *ctx);
1721         ctx->lc_state = LCS_INITIALIZED;
1722         ctx->lc_tags = tags;
1723         if (tags & LCT_REMEMBER) {
1724                 spin_lock(&lu_keys_guard);
1725                 list_add(&ctx->lc_remember, &lu_context_remembered);
1726                 spin_unlock(&lu_keys_guard);
1727         } else {
1728                 INIT_LIST_HEAD(&ctx->lc_remember);
1729         }
1730
1731         rc = keys_init(ctx);
1732         if (rc != 0)
1733                 lu_context_fini(ctx);
1734
1735         return rc;
1736 }
1737 EXPORT_SYMBOL(lu_context_init);
1738
1739 /**
1740  * Finalize context data-structure. Destroy key values.
1741  */
1742 void lu_context_fini(struct lu_context *ctx)
1743 {
1744         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1745         ctx->lc_state = LCS_FINALIZED;
1746
1747         if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1748                 LASSERT(list_empty(&ctx->lc_remember));
1749                 keys_fini(ctx);
1750
1751         } else { /* could race with key degister */
1752                 spin_lock(&lu_keys_guard);
1753                 keys_fini(ctx);
1754                 list_del_init(&ctx->lc_remember);
1755                 spin_unlock(&lu_keys_guard);
1756         }
1757 }
1758 EXPORT_SYMBOL(lu_context_fini);
1759
1760 /**
1761  * Called before entering context.
1762  */
1763 void lu_context_enter(struct lu_context *ctx)
1764 {
1765         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1766         ctx->lc_state = LCS_ENTERED;
1767 }
1768 EXPORT_SYMBOL(lu_context_enter);
1769
1770 /**
1771  * Called after exiting from \a ctx
1772  */
1773 void lu_context_exit(struct lu_context *ctx)
1774 {
1775         unsigned int i;
1776
1777         LINVRNT(ctx->lc_state == LCS_ENTERED);
1778         ctx->lc_state = LCS_LEFT;
1779         if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1780                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1781                         if (ctx->lc_value[i] != NULL) {
1782                                 struct lu_context_key *key;
1783
1784                                 key = lu_keys[i];
1785                                 LASSERT(key != NULL);
1786                                 if (key->lct_exit != NULL)
1787                                         key->lct_exit(ctx,
1788                                                       key, ctx->lc_value[i]);
1789                         }
1790                 }
1791         }
1792 }
1793 EXPORT_SYMBOL(lu_context_exit);
1794
1795 /**
1796  * Allocate for context all missing keys that were registered after context
1797  * creation. key_set_version is only changed in rare cases when modules
1798  * are loaded and removed.
1799  */
1800 int lu_context_refill(struct lu_context *ctx)
1801 {
1802         return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1803 }
1804 EXPORT_SYMBOL(lu_context_refill);
1805
1806 /**
1807  * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1808  * obd being added. Currently, this is only used on client side, specifically
1809  * for echo device client, for other stack (like ptlrpc threads), context are
1810  * predefined when the lu_device type are registered, during the module probe
1811  * phase.
1812  */
1813 __u32 lu_context_tags_default = 0;
1814 __u32 lu_session_tags_default = 0;
1815
1816 void lu_context_tags_update(__u32 tags)
1817 {
1818         spin_lock(&lu_keys_guard);
1819         lu_context_tags_default |= tags;
1820         key_set_version++;
1821         spin_unlock(&lu_keys_guard);
1822 }
1823 EXPORT_SYMBOL(lu_context_tags_update);
1824
1825 void lu_context_tags_clear(__u32 tags)
1826 {
1827         spin_lock(&lu_keys_guard);
1828         lu_context_tags_default &= ~tags;
1829         key_set_version++;
1830         spin_unlock(&lu_keys_guard);
1831 }
1832 EXPORT_SYMBOL(lu_context_tags_clear);
1833
1834 void lu_session_tags_update(__u32 tags)
1835 {
1836         spin_lock(&lu_keys_guard);
1837         lu_session_tags_default |= tags;
1838         key_set_version++;
1839         spin_unlock(&lu_keys_guard);
1840 }
1841 EXPORT_SYMBOL(lu_session_tags_update);
1842
1843 void lu_session_tags_clear(__u32 tags)
1844 {
1845         spin_lock(&lu_keys_guard);
1846         lu_session_tags_default &= ~tags;
1847         key_set_version++;
1848         spin_unlock(&lu_keys_guard);
1849 }
1850 EXPORT_SYMBOL(lu_session_tags_clear);
1851
1852 int lu_env_init(struct lu_env *env, __u32 tags)
1853 {
1854         int result;
1855
1856         env->le_ses = NULL;
1857         result = lu_context_init(&env->le_ctx, tags);
1858         if (likely(result == 0))
1859                 lu_context_enter(&env->le_ctx);
1860         return result;
1861 }
1862 EXPORT_SYMBOL(lu_env_init);
1863
1864 void lu_env_fini(struct lu_env *env)
1865 {
1866         lu_context_exit(&env->le_ctx);
1867         lu_context_fini(&env->le_ctx);
1868         env->le_ses = NULL;
1869 }
1870 EXPORT_SYMBOL(lu_env_fini);
1871
1872 int lu_env_refill(struct lu_env *env)
1873 {
1874         int result;
1875
1876         result = lu_context_refill(&env->le_ctx);
1877         if (result == 0 && env->le_ses != NULL)
1878                 result = lu_context_refill(env->le_ses);
1879         return result;
1880 }
1881 EXPORT_SYMBOL(lu_env_refill);
1882
1883 /**
1884  * Currently, this API will only be used by echo client.
1885  * Because echo client and normal lustre client will share
1886  * same cl_env cache. So echo client needs to refresh
1887  * the env context after it get one from the cache, especially
1888  * when normal client and echo client co-exist in the same client.
1889  */
1890 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1891                           __u32 stags)
1892 {
1893         int    result;
1894
1895         if ((env->le_ctx.lc_tags & ctags) != ctags) {
1896                 env->le_ctx.lc_version = 0;
1897                 env->le_ctx.lc_tags |= ctags;
1898         }
1899
1900         if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1901                 env->le_ses->lc_version = 0;
1902                 env->le_ses->lc_tags |= stags;
1903         }
1904
1905         result = lu_env_refill(env);
1906
1907         return result;
1908 }
1909 EXPORT_SYMBOL(lu_env_refill_by_tags);
1910
1911 static struct shrinker *lu_site_shrinker;
1912
1913 typedef struct lu_site_stats{
1914         unsigned        lss_populated;
1915         unsigned        lss_max_search;
1916         unsigned        lss_total;
1917         unsigned        lss_busy;
1918 } lu_site_stats_t;
1919
1920 static void lu_site_stats_get(cfs_hash_t *hs,
1921                               lu_site_stats_t *stats, int populated)
1922 {
1923         cfs_hash_bd_t bd;
1924         unsigned int  i;
1925
1926         cfs_hash_for_each_bucket(hs, &bd, i) {
1927                 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1928                 struct hlist_head       *hhead;
1929
1930                 cfs_hash_bd_lock(hs, &bd, 1);
1931                 stats->lss_busy  += bkt->lsb_busy;
1932                 stats->lss_total += cfs_hash_bd_count_get(&bd);
1933                 stats->lss_max_search = max((int)stats->lss_max_search,
1934                                             cfs_hash_bd_depmax_get(&bd));
1935                 if (!populated) {
1936                         cfs_hash_bd_unlock(hs, &bd, 1);
1937                         continue;
1938                 }
1939
1940                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1941                         if (!hlist_empty(hhead))
1942                                 stats->lss_populated++;
1943                 }
1944                 cfs_hash_bd_unlock(hs, &bd, 1);
1945         }
1946 }
1947
1948
1949 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
1950                                            struct shrink_control *sc)
1951 {
1952         lu_site_stats_t stats;
1953         struct lu_site *s;
1954         struct lu_site *tmp;
1955         unsigned long cached = 0;
1956
1957         if (!(sc->gfp_mask & __GFP_FS))
1958                 return 0;
1959
1960         mutex_lock(&lu_sites_guard);
1961         list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1962                 memset(&stats, 0, sizeof(stats));
1963                 lu_site_stats_get(s->ls_obj_hash, &stats, 0);
1964                 cached += stats.lss_total - stats.lss_busy;
1965         }
1966         mutex_unlock(&lu_sites_guard);
1967
1968         cached = (cached / 100) * sysctl_vfs_cache_pressure;
1969         CDEBUG(D_INODE, "%ld objects cached\n", cached);
1970         return cached;
1971 }
1972
1973 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
1974                                           struct shrink_control *sc)
1975 {
1976         struct lu_site *s;
1977         struct lu_site *tmp;
1978         unsigned long remain = sc->nr_to_scan;
1979         LIST_HEAD(splice);
1980
1981         if (!(sc->gfp_mask & __GFP_FS))
1982                 /* We must not take the lu_sites_guard lock when
1983                  * __GFP_FS is *not* set because of the deadlock
1984                  * possibility detailed above. Additionally,
1985                  * since we cannot determine the number of
1986                  * objects in the cache without taking this
1987                  * lock, we're in a particularly tough spot. As
1988                  * a result, we'll just lie and say our cache is
1989                  * empty. This _should_ be ok, as we can't
1990                  * reclaim objects when __GFP_FS is *not* set
1991                  * anyways.
1992                  */
1993                 return SHRINK_STOP;
1994
1995         mutex_lock(&lu_sites_guard);
1996         list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1997                 remain = lu_site_purge(&lu_shrink_env, s, remain);
1998                 /*
1999                  * Move just shrunk site to the tail of site list to
2000                  * assure shrinking fairness.
2001                  */
2002                 list_move_tail(&s->ls_linkage, &splice);
2003         }
2004         list_splice(&splice, lu_sites.prev);
2005         mutex_unlock(&lu_sites_guard);
2006
2007         return sc->nr_to_scan - remain;
2008 }
2009
2010 #ifndef HAVE_SHRINKER_COUNT
2011 /*
2012  * There exists a potential lock inversion deadlock scenario when using
2013  * Lustre on top of ZFS. This occurs between one of ZFS's
2014  * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2015  * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2016  * while thread B will take the ht_lock and sleep on the lu_sites_guard
2017  * lock. Obviously neither thread will wake and drop their respective hold
2018  * on their lock.
2019  *
2020  * To prevent this from happening we must ensure the lu_sites_guard lock is
2021  * not taken while down this code path. ZFS reliably does not set the
2022  * __GFP_FS bit in its code paths, so this can be used to determine if it
2023  * is safe to take the lu_sites_guard lock.
2024  *
2025  * Ideally we should accurately return the remaining number of cached
2026  * objects without taking the  lu_sites_guard lock, but this is not
2027  * possible in the current implementation.
2028  */
2029 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2030 {
2031         int cached = 0;
2032         struct shrink_control scv = {
2033                  .nr_to_scan = shrink_param(sc, nr_to_scan),
2034                  .gfp_mask   = shrink_param(sc, gfp_mask)
2035         };
2036 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2037         struct shrinker* shrinker = NULL;
2038 #endif
2039
2040
2041         CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2042
2043         lu_cache_shrink_scan(shrinker, &scv);
2044
2045         cached = lu_cache_shrink_count(shrinker, &scv);
2046         if (scv.nr_to_scan == 0)
2047                 CDEBUG(D_INODE, "%d objects cached\n", cached);
2048         return cached;
2049 }
2050
2051 #endif /* HAVE_SHRINKER_COUNT */
2052
2053
2054 /*
2055  * Debugging stuff.
2056  */
2057
2058 /**
2059  * Environment to be used in debugger, contains all tags.
2060  */
2061 struct lu_env lu_debugging_env;
2062
2063 /**
2064  * Debugging printer function using printk().
2065  */
2066 int lu_printk_printer(const struct lu_env *env,
2067                       void *unused, const char *format, ...)
2068 {
2069         va_list args;
2070
2071         va_start(args, format);
2072         vprintk(format, args);
2073         va_end(args);
2074         return 0;
2075 }
2076
2077 int lu_debugging_setup(void)
2078 {
2079         return lu_env_init(&lu_debugging_env, ~0);
2080 }
2081
2082 void lu_context_keys_dump(void)
2083 {
2084         unsigned int i;
2085
2086         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2087                 struct lu_context_key *key;
2088
2089                 key = lu_keys[i];
2090                 if (key != NULL) {
2091                         CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2092                                i, key, key->lct_tags,
2093                                key->lct_init, key->lct_fini, key->lct_exit,
2094                                key->lct_index, atomic_read(&key->lct_used),
2095                                key->lct_owner ? key->lct_owner->name : "",
2096                                key->lct_owner);
2097                         lu_ref_print(&key->lct_reference);
2098                 }
2099         }
2100 }
2101 EXPORT_SYMBOL(lu_context_keys_dump);
2102
2103 /**
2104  * Initialization of global lu_* data.
2105  */
2106 int lu_global_init(void)
2107 {
2108         int result;
2109         DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2110                          lu_cache_shrink_count, lu_cache_shrink_scan);
2111
2112         CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2113
2114         INIT_LIST_HEAD(&lu_device_types);
2115         INIT_LIST_HEAD(&lu_context_remembered);
2116         INIT_LIST_HEAD(&lu_sites);
2117
2118         result = lu_ref_global_init();
2119         if (result != 0)
2120                 return result;
2121
2122         LU_CONTEXT_KEY_INIT(&lu_global_key);
2123         result = lu_context_key_register(&lu_global_key);
2124         if (result != 0)
2125                 return result;
2126
2127         /*
2128          * At this level, we don't know what tags are needed, so allocate them
2129          * conservatively. This should not be too bad, because this
2130          * environment is global.
2131          */
2132         mutex_lock(&lu_sites_guard);
2133         result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2134         mutex_unlock(&lu_sites_guard);
2135         if (result != 0)
2136                 return result;
2137
2138         /*
2139          * seeks estimation: 3 seeks to read a record from oi, one to read
2140          * inode, one for ea. Unfortunately setting this high value results in
2141          * lu_object/inode cache consuming all the memory.
2142          */
2143         lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2144         if (lu_site_shrinker == NULL)
2145                 return -ENOMEM;
2146
2147         return result;
2148 }
2149
2150 /**
2151  * Dual to lu_global_init().
2152  */
2153 void lu_global_fini(void)
2154 {
2155         if (lu_site_shrinker != NULL) {
2156                 remove_shrinker(lu_site_shrinker);
2157                 lu_site_shrinker = NULL;
2158         }
2159
2160         lu_context_key_degister(&lu_global_key);
2161
2162         /*
2163          * Tear shrinker environment down _after_ de-registering
2164          * lu_global_key, because the latter has a value in the former.
2165          */
2166         mutex_lock(&lu_sites_guard);
2167         lu_env_fini(&lu_shrink_env);
2168         mutex_unlock(&lu_sites_guard);
2169
2170         lu_ref_global_fini();
2171 }
2172
2173 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2174 {
2175 #ifdef LPROCFS
2176         struct lprocfs_counter ret;
2177
2178         lprocfs_stats_collect(stats, idx, &ret);
2179         return (__u32)ret.lc_count;
2180 #else
2181         return 0;
2182 #endif
2183 }
2184
2185 /**
2186  * Output site statistical counters into a buffer. Suitable for
2187  * lprocfs_rd_*()-style functions.
2188  */
2189 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2190 {
2191         lu_site_stats_t stats;
2192
2193         memset(&stats, 0, sizeof(stats));
2194         lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2195
2196         return seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2197                           stats.lss_busy,
2198                           stats.lss_total,
2199                           stats.lss_populated,
2200                           CFS_HASH_NHLIST(s->ls_obj_hash),
2201                           stats.lss_max_search,
2202                           ls_stats_read(s->ls_stats, LU_SS_CREATED),
2203                           ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2204                           ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2205                           ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2206                           ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2207                           ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2208 }
2209 EXPORT_SYMBOL(lu_site_stats_seq_print);
2210
2211 int lu_site_stats_print(const struct lu_site *s, char *page, int count)
2212 {
2213         lu_site_stats_t stats;
2214
2215         memset(&stats, 0, sizeof(stats));
2216         lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2217
2218         return snprintf(page, count, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2219                         stats.lss_busy,
2220                         stats.lss_total,
2221                         stats.lss_populated,
2222                         CFS_HASH_NHLIST(s->ls_obj_hash),
2223                         stats.lss_max_search,
2224                         ls_stats_read(s->ls_stats, LU_SS_CREATED),
2225                         ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2226                         ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2227                         ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2228                         ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2229                         ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2230 }
2231 EXPORT_SYMBOL(lu_site_stats_print);
2232
2233 /**
2234  * Helper function to initialize a number of kmem slab caches at once.
2235  */
2236 int lu_kmem_init(struct lu_kmem_descr *caches)
2237 {
2238         int result;
2239         struct lu_kmem_descr *iter = caches;
2240
2241         for (result = 0; iter->ckd_cache != NULL; ++iter) {
2242                 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2243                                                      iter->ckd_size,
2244                                                      0, 0, NULL);
2245                 if (*iter->ckd_cache == NULL) {
2246                         result = -ENOMEM;
2247                         /* free all previously allocated caches */
2248                         lu_kmem_fini(caches);
2249                         break;
2250                 }
2251         }
2252         return result;
2253 }
2254 EXPORT_SYMBOL(lu_kmem_init);
2255
2256 /**
2257  * Helper function to finalize a number of kmem slab cached at once. Dual to
2258  * lu_kmem_init().
2259  */
2260 void lu_kmem_fini(struct lu_kmem_descr *caches)
2261 {
2262         for (; caches->ckd_cache != NULL; ++caches) {
2263                 if (*caches->ckd_cache != NULL) {
2264                         kmem_cache_destroy(*caches->ckd_cache);
2265                         *caches->ckd_cache = NULL;
2266                 }
2267         }
2268 }
2269 EXPORT_SYMBOL(lu_kmem_fini);
2270
2271 /**
2272  * Temporary solution to be able to assign fid in ->do_create()
2273  * till we have fully-functional OST fids
2274  */
2275 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2276                           const struct lu_fid *fid)
2277 {
2278         struct lu_site          *s = o->lo_dev->ld_site;
2279         struct lu_fid           *old = &o->lo_header->loh_fid;
2280         struct lu_site_bkt_data *bkt;
2281         struct lu_object        *shadow;
2282         wait_queue_t             waiter;
2283         cfs_hash_t              *hs;
2284         cfs_hash_bd_t            bd;
2285         __u64                    version = 0;
2286
2287         LASSERT(fid_is_zero(old));
2288
2289         hs = s->ls_obj_hash;
2290         cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2291         shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2292         /* supposed to be unique */
2293         LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2294         *old = *fid;
2295         bkt = cfs_hash_bd_extra_get(hs, &bd);
2296         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2297         bkt->lsb_busy++;
2298         cfs_hash_bd_unlock(hs, &bd, 1);
2299 }
2300 EXPORT_SYMBOL(lu_object_assign_fid);
2301
2302 /**
2303  * allocates object with 0 (non-assiged) fid
2304  * XXX: temporary solution to be able to assign fid in ->do_create()
2305  *      till we have fully-functional OST fids
2306  */
2307 struct lu_object *lu_object_anon(const struct lu_env *env,
2308                                  struct lu_device *dev,
2309                                  const struct lu_object_conf *conf)
2310 {
2311         struct lu_fid     fid;
2312         struct lu_object *o;
2313
2314         fid_zero(&fid);
2315         o = lu_object_alloc(env, dev, &fid, conf);
2316
2317         return o;
2318 }
2319 EXPORT_SYMBOL(lu_object_anon);
2320
2321 struct lu_buf LU_BUF_NULL = {
2322         .lb_buf = NULL,
2323         .lb_len = 0
2324 };
2325 EXPORT_SYMBOL(LU_BUF_NULL);
2326
2327 void lu_buf_free(struct lu_buf *buf)
2328 {
2329         LASSERT(buf);
2330         if (buf->lb_buf) {
2331                 LASSERT(buf->lb_len > 0);
2332                 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2333                 buf->lb_buf = NULL;
2334                 buf->lb_len = 0;
2335         }
2336 }
2337 EXPORT_SYMBOL(lu_buf_free);
2338
2339 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2340 {
2341         LASSERT(buf);
2342         LASSERT(buf->lb_buf == NULL);
2343         LASSERT(buf->lb_len == 0);
2344         OBD_ALLOC_LARGE(buf->lb_buf, size);
2345         if (likely(buf->lb_buf))
2346                 buf->lb_len = size;
2347 }
2348 EXPORT_SYMBOL(lu_buf_alloc);
2349
2350 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2351 {
2352         lu_buf_free(buf);
2353         lu_buf_alloc(buf, size);
2354 }
2355 EXPORT_SYMBOL(lu_buf_realloc);
2356
2357 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2358 {
2359         if (buf->lb_buf == NULL && buf->lb_len == 0)
2360                 lu_buf_alloc(buf, len);
2361
2362         if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2363                 lu_buf_realloc(buf, len);
2364
2365         return buf;
2366 }
2367 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2368
2369 /**
2370  * Increase the size of the \a buf.
2371  * preserves old data in buffer
2372  * old buffer remains unchanged on error
2373  * \retval 0 or -ENOMEM
2374  */
2375 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2376 {
2377         char *ptr;
2378
2379         if (len <= buf->lb_len)
2380                 return 0;
2381
2382         OBD_ALLOC_LARGE(ptr, len);
2383         if (ptr == NULL)
2384                 return -ENOMEM;
2385
2386         /* Free the old buf */
2387         if (buf->lb_buf != NULL) {
2388                 memcpy(ptr, buf->lb_buf, buf->lb_len);
2389                 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2390         }
2391
2392         buf->lb_buf = ptr;
2393         buf->lb_len = len;
2394         return 0;
2395 }
2396 EXPORT_SYMBOL(lu_buf_check_and_grow);
2397