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