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