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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/obdclass/lu_object.c
  33  *
  34  * Lustre Object.
  35  * These are the only exported functions, they provide some generic
  36  * infrastructure for managing object devices
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_CLASS
  42
  43 #include <linux/module.h>
  44 #include <linux/list.h>
  45 #include <libcfs/libcfs.h>
  46 #include <libcfs/libcfs_hash.h> /* hash_long() */
  47 #include <libcfs/linux/linux-mem.h>
  48 #include <obd_class.h>
  49 #include <obd_support.h>
  50 #include <lustre_disk.h>
  51 #include <lustre_fid.h>
  52 #include <lu_object.h>
  53 #include <lu_ref.h>
  54
  55 enum {
  56         LU_CACHE_PERCENT_MAX     = 50,
  57         LU_CACHE_PERCENT_DEFAULT = 20
  58 };
  59
  60 #define LU_CACHE_NR_MAX_ADJUST          512
  61 #define LU_CACHE_NR_UNLIMITED           -1
  62 #define LU_CACHE_NR_DEFAULT             LU_CACHE_NR_UNLIMITED
  63 #define LU_CACHE_NR_LDISKFS_LIMIT       LU_CACHE_NR_UNLIMITED
  64 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
  65 #define LU_CACHE_NR_ZFS_LIMIT           10240
  66
  67 #define LU_SITE_BITS_MIN    12
  68 #define LU_SITE_BITS_MAX    24
  69 #define LU_SITE_BITS_MAX_CL 19
  70 /**
  71  * total 256 buckets, we don't want too many buckets because:
  72  * - consume too much memory
  73  * - avoid unbalanced LRU list
  74  */
  75 #define LU_SITE_BKT_BITS    8
  76
  77
  78 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
  79 module_param(lu_cache_percent, int, 0644);
  80 MODULE_PARM_DESC(lu_cache_percent, "Percentage of memory to be used as lu_object cache");
  81
  82 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
  83 module_param(lu_cache_nr, long, 0644);
  84 MODULE_PARM_DESC(lu_cache_nr, "Maximum number of objects in lu_object cache");
  85
  86 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
  87 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx);
  88
  89 /**
  90  * Decrease reference counter on object. If last reference is freed, return
  91  * object to the cache, unless lu_object_is_dying(o) holds. In the latter
  92  * case, free object immediately.
  93  */
  94 void lu_object_put(const struct lu_env *env, struct lu_object *o)
  95 {
  96         struct lu_site_bkt_data *bkt;
  97         struct lu_object_header *top;
  98         struct lu_site *site;
  99         struct lu_object *orig;
 100         struct cfs_hash_bd bd;
 101         const struct lu_fid *fid;
 102
 103         top  = o->lo_header;
 104         site = o->lo_dev->ld_site;
 105         orig = o;
 106
 107         /*
 108          * till we have full fids-on-OST implemented anonymous objects
 109          * are possible in OSP. such an object isn't listed in the site
 110          * so we should not remove it from the site.
 111          */
 112         fid = lu_object_fid(o);
 113         if (fid_is_zero(fid)) {
 114                 LASSERT(top->loh_hash.next == NULL
 115                         && top->loh_hash.pprev == NULL);
 116                 LASSERT(list_empty(&top->loh_lru));
 117                 if (!atomic_dec_and_test(&top->loh_ref))
 118                         return;
 119                 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 120                         if (o->lo_ops->loo_object_release != NULL)
 121                                 o->lo_ops->loo_object_release(env, o);
 122                 }
 123                 lu_object_free(env, orig);
 124                 return;
 125         }
 126
 127         cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
 128         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
 129
 130         if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
 131                 if (lu_object_is_dying(top)) {
 132                         /*
 133                          * somebody may be waiting for this, currently only
 134                          * used for cl_object, see cl_object_put_last().
 135                          */
 136                         wake_up_all(&bkt->lsb_marche_funebre);
 137                 }
 138                 return;
 139         }
 140
 141         /*
 142          * When last reference is released, iterate over object
 143          * layers, and notify them that object is no longer busy.
 144          */
 145         list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 146                 if (o->lo_ops->loo_object_release != NULL)
 147                         o->lo_ops->loo_object_release(env, o);
 148         }
 149
 150         if (!lu_object_is_dying(top) &&
 151             (lu_object_exists(orig) || lu_object_is_cl(orig))) {
 152                 LASSERT(list_empty(&top->loh_lru));
 153                 list_add_tail(&top->loh_lru, &bkt->lsb_lru);
 154                 bkt->lsb_lru_len++;
 155                 percpu_counter_inc(&site->ls_lru_len_counter);
 156                 CDEBUG(D_INODE, "Add %p to site lru. hash: %p, bkt: %p, "
 157                        "lru_len: %ld\n",
 158                        o, site->ls_obj_hash, bkt, bkt->lsb_lru_len);
 159                 cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 160                 return;
 161         }
 162
 163         /*
 164          * If object is dying (will not be cached) then remove it
 165          * from hash table and LRU.
 166          *
 167          * This is done with hash table and LRU lists locked. As the only
 168          * way to acquire first reference to previously unreferenced
 169          * object is through hash-table lookup (lu_object_find()),
 170          * or LRU scanning (lu_site_purge()), that are done under hash-table
 171          * and LRU lock, no race with concurrent object lookup is possible
 172          * and we can safely destroy object below.
 173          */
 174         if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags))
 175                 cfs_hash_bd_del_locked(site->ls_obj_hash, &bd, &top->loh_hash);
 176         cfs_hash_bd_unlock(site->ls_obj_hash, &bd, 1);
 177         /*
 178          * Object was already removed from hash and lru above, can
 179          * kill it.
 180          */
 181         lu_object_free(env, orig);
 182 }
 183 EXPORT_SYMBOL(lu_object_put);
 184
 185 /**
 186  * Put object and don't keep in cache. This is temporary solution for
 187  * multi-site objects when its layering is not constant.
 188  */
 189 void lu_object_put_nocache(const struct lu_env *env, struct lu_object *o)
 190 {
 191         set_bit(LU_OBJECT_HEARD_BANSHEE, &o->lo_header->loh_flags);
 192         return lu_object_put(env, o);
 193 }
 194 EXPORT_SYMBOL(lu_object_put_nocache);
 195
 196 /**
 197  * Kill the object and take it out of LRU cache.
 198  * Currently used by client code for layout change.
 199  */
 200 void lu_object_unhash(const struct lu_env *env, struct lu_object *o)
 201 {
 202         struct lu_object_header *top;
 203
 204         top = o->lo_header;
 205         set_bit(LU_OBJECT_HEARD_BANSHEE, &top->loh_flags);
 206         if (!test_and_set_bit(LU_OBJECT_UNHASHED, &top->loh_flags)) {
 207                 struct lu_site *site = o->lo_dev->ld_site;
 208                 struct cfs_hash *obj_hash = site->ls_obj_hash;
 209                 struct cfs_hash_bd bd;
 210
 211                 cfs_hash_bd_get_and_lock(obj_hash, &top->loh_fid, &bd, 1);
 212                 if (!list_empty(&top->loh_lru)) {
 213                         struct lu_site_bkt_data *bkt;
 214
 215                         list_del_init(&top->loh_lru);
 216                         bkt = cfs_hash_bd_extra_get(obj_hash, &bd);
 217                         bkt->lsb_lru_len--;
 218                         percpu_counter_dec(&site->ls_lru_len_counter);
 219                 }
 220                 cfs_hash_bd_del_locked(obj_hash, &bd, &top->loh_hash);
 221                 cfs_hash_bd_unlock(obj_hash, &bd, 1);
 222         }
 223 }
 224 EXPORT_SYMBOL(lu_object_unhash);
 225
 226 /**
 227  * Allocate new object.
 228  *
 229  * This follows object creation protocol, described in the comment within
 230  * struct lu_device_operations definition.
 231  */
 232 static struct lu_object *lu_object_alloc(const struct lu_env *env,
 233                                          struct lu_device *dev,
 234                                          const struct lu_fid *f,
 235                                          const struct lu_object_conf *conf)
 236 {
 237         struct lu_object *scan;
 238         struct lu_object *top;
 239         struct list_head *layers;
 240         unsigned int init_mask = 0;
 241         unsigned int init_flag;
 242         int clean;
 243         int result;
 244         ENTRY;
 245
 246         /*
 247          * Create top-level object slice. This will also create
 248          * lu_object_header.
 249          */
 250         top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
 251         if (top == NULL)
 252                 RETURN(ERR_PTR(-ENOMEM));
 253         if (IS_ERR(top))
 254                 RETURN(top);
 255         /*
 256          * This is the only place where object fid is assigned. It's constant
 257          * after this point.
 258          */
 259         top->lo_header->loh_fid = *f;
 260         layers = &top->lo_header->loh_layers;
 261
 262         do {
 263                 /*
 264                  * Call ->loo_object_init() repeatedly, until no more new
 265                  * object slices are created.
 266                  */
 267                 clean = 1;
 268                 init_flag = 1;
 269                 list_for_each_entry(scan, layers, lo_linkage) {
 270                         if (init_mask & init_flag)
 271                                 goto next;
 272                         clean = 0;
 273                         scan->lo_header = top->lo_header;
 274                         result = scan->lo_ops->loo_object_init(env, scan, conf);
 275                         if (result != 0) {
 276                                 lu_object_free(env, top);
 277                                 RETURN(ERR_PTR(result));
 278                         }
 279                         init_mask |= init_flag;
 280 next:
 281                         init_flag <<= 1;
 282                 }
 283         } while (!clean);
 284
 285         list_for_each_entry_reverse(scan, layers, lo_linkage) {
 286                 if (scan->lo_ops->loo_object_start != NULL) {
 287                         result = scan->lo_ops->loo_object_start(env, scan);
 288                         if (result != 0) {
 289                                 lu_object_free(env, top);
 290                                 RETURN(ERR_PTR(result));
 291                         }
 292                 }
 293         }
 294
 295         lprocfs_counter_incr(dev->ld_site->ls_stats, LU_SS_CREATED);
 296         RETURN(top);
 297 }
 298
 299 /**
 300  * Free an object.
 301  */
 302 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
 303 {
 304         struct lu_site_bkt_data *bkt;
 305         struct lu_site          *site;
 306         struct lu_object        *scan;
 307         struct list_head        *layers;
 308         struct list_head         splice;
 309
 310         site   = o->lo_dev->ld_site;
 311         layers = &o->lo_header->loh_layers;
 312         bkt    = lu_site_bkt_from_fid(site, &o->lo_header->loh_fid);
 313         /*
 314          * First call ->loo_object_delete() method to release all resources.
 315          */
 316         list_for_each_entry_reverse(scan, layers, lo_linkage) {
 317                 if (scan->lo_ops->loo_object_delete != NULL)
 318                         scan->lo_ops->loo_object_delete(env, scan);
 319         }
 320
 321         /*
 322          * Then, splice object layers into stand-alone list, and call
 323          * ->loo_object_free() on all layers to free memory. Splice is
 324          * necessary, because lu_object_header is freed together with the
 325          * top-level slice.
 326          */
 327         INIT_LIST_HEAD(&splice);
 328         list_splice_init(layers, &splice);
 329         while (!list_empty(&splice)) {
 330                 /*
 331                  * Free layers in bottom-to-top order, so that object header
 332                  * lives as long as possible and ->loo_object_free() methods
 333                  * can look at its contents.
 334                  */
 335                 o = container_of0(splice.prev, struct lu_object, lo_linkage);
 336                 list_del_init(&o->lo_linkage);
 337                 LASSERT(o->lo_ops->loo_object_free != NULL);
 338                 o->lo_ops->loo_object_free(env, o);
 339         }
 340
 341         if (waitqueue_active(&bkt->lsb_marche_funebre))
 342                 wake_up_all(&bkt->lsb_marche_funebre);
 343 }
 344
 345 /**
 346  * Free \a nr objects from the cold end of the site LRU list.
 347  * if canblock is 0, then don't block awaiting for another
 348  * instance of lu_site_purge() to complete
 349  */
 350 int lu_site_purge_objects(const struct lu_env *env, struct lu_site *s,
 351                           int nr, int canblock)
 352 {
 353         struct lu_object_header *h;
 354         struct lu_object_header *temp;
 355         struct lu_site_bkt_data *bkt;
 356         struct cfs_hash_bd            bd;
 357         struct cfs_hash_bd            bd2;
 358         struct list_head         dispose;
 359         int                      did_sth;
 360         unsigned int             start = 0;
 361         int                      count;
 362         int                      bnr;
 363         unsigned int             i;
 364
 365         if (OBD_FAIL_CHECK(OBD_FAIL_OBD_NO_LRU))
 366                 RETURN(0);
 367
 368         INIT_LIST_HEAD(&dispose);
 369         /*
 370          * Under LRU list lock, scan LRU list and move unreferenced objects to
 371          * the dispose list, removing them from LRU and hash table.
 372          */
 373         if (nr != ~0)
 374                 start = s->ls_purge_start;
 375         bnr = (nr == ~0) ? -1 : nr / (int)CFS_HASH_NBKT(s->ls_obj_hash) + 1;
 376  again:
 377         /*
 378          * It doesn't make any sense to make purge threads parallel, that can
 379          * only bring troubles to us. See LU-5331.
 380          */
 381         if (canblock != 0)
 382                 mutex_lock(&s->ls_purge_mutex);
 383         else if (mutex_trylock(&s->ls_purge_mutex) == 0)
 384                 goto out;
 385
 386         did_sth = 0;
 387         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
 388                 if (i < start)
 389                         continue;
 390                 count = bnr;
 391                 cfs_hash_bd_lock(s->ls_obj_hash, &bd, 1);
 392                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
 393
 394                 list_for_each_entry_safe(h, temp, &bkt->lsb_lru, loh_lru) {
 395                         LASSERT(atomic_read(&h->loh_ref) == 0);
 396
 397                         cfs_hash_bd_get(s->ls_obj_hash, &h->loh_fid, &bd2);
 398                         LASSERT(bd.bd_bucket == bd2.bd_bucket);
 399
 400                         cfs_hash_bd_del_locked(s->ls_obj_hash,
 401                                                &bd2, &h->loh_hash);
 402                         list_move(&h->loh_lru, &dispose);
 403                         bkt->lsb_lru_len--;
 404                         percpu_counter_dec(&s->ls_lru_len_counter);
 405                         if (did_sth == 0)
 406                                 did_sth = 1;
 407
 408                         if (nr != ~0 && --nr == 0)
 409                                 break;
 410
 411                         if (count > 0 && --count == 0)
 412                                 break;
 413
 414                 }
 415                 cfs_hash_bd_unlock(s->ls_obj_hash, &bd, 1);
 416                 cond_resched();
 417                 /*
 418                  * Free everything on the dispose list. This is safe against
 419                  * races due to the reasons described in lu_object_put().
 420                  */
 421                 while (!list_empty(&dispose)) {
 422                         h = container_of0(dispose.next,
 423                                           struct lu_object_header, loh_lru);
 424                         list_del_init(&h->loh_lru);
 425                         lu_object_free(env, lu_object_top(h));
 426                         lprocfs_counter_incr(s->ls_stats, LU_SS_LRU_PURGED);
 427                 }
 428
 429                 if (nr == 0)
 430                         break;
 431         }
 432         mutex_unlock(&s->ls_purge_mutex);
 433
 434         if (nr != 0 && did_sth && start != 0) {
 435                 start = 0; /* restart from the first bucket */
 436                 goto again;
 437         }
 438         /* race on s->ls_purge_start, but nobody cares */
 439         s->ls_purge_start = i % CFS_HASH_NBKT(s->ls_obj_hash);
 440
 441 out:
 442         return nr;
 443 }
 444 EXPORT_SYMBOL(lu_site_purge_objects);
 445
 446 /*
 447  * Object printing.
 448  *
 449  * Code below has to jump through certain loops to output object description
 450  * into libcfs_debug_msg-based log. The problem is that lu_object_print()
 451  * composes object description from strings that are parts of _lines_ of
 452  * output (i.e., strings that are not terminated by newline). This doesn't fit
 453  * very well into libcfs_debug_msg() interface that assumes that each message
 454  * supplied to it is a self-contained output line.
 455  *
 456  * To work around this, strings are collected in a temporary buffer
 457  * (implemented as a value of lu_cdebug_key key), until terminating newline
 458  * character is detected.
 459  *
 460  */
 461
 462 enum {
 463         /**
 464          * Maximal line size.
 465          *
 466          * XXX overflow is not handled correctly.
 467          */
 468         LU_CDEBUG_LINE = 512
 469 };
 470
 471 struct lu_cdebug_data {
 472         /**
 473          * Temporary buffer.
 474          */
 475         char lck_area[LU_CDEBUG_LINE];
 476 };
 477
 478 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
 479 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
 480
 481 /**
 482  * Key, holding temporary buffer. This key is registered very early by
 483  * lu_global_init().
 484  */
 485 static struct lu_context_key lu_global_key = {
 486         .lct_tags = LCT_MD_THREAD | LCT_DT_THREAD |
 487                     LCT_MG_THREAD | LCT_CL_THREAD | LCT_LOCAL,
 488         .lct_init = lu_global_key_init,
 489         .lct_fini = lu_global_key_fini
 490 };
 491
 492 /**
 493  * Printer function emitting messages through libcfs_debug_msg().
 494  */
 495 int lu_cdebug_printer(const struct lu_env *env,
 496                       void *cookie, const char *format, ...)
 497 {
 498         struct libcfs_debug_msg_data *msgdata = cookie;
 499         struct lu_cdebug_data        *key;
 500         int used;
 501         int complete;
 502         va_list args;
 503
 504         va_start(args, format);
 505
 506         key = lu_context_key_get(&env->le_ctx, &lu_global_key);
 507         LASSERT(key != NULL);
 508
 509         used = strlen(key->lck_area);
 510         complete = format[strlen(format) - 1] == '\n';
 511         /*
 512          * Append new chunk to the buffer.
 513          */
 514         vsnprintf(key->lck_area + used,
 515                   ARRAY_SIZE(key->lck_area) - used, format, args);
 516         if (complete) {
 517                 if (cfs_cdebug_show(msgdata->msg_mask, msgdata->msg_subsys))
 518                         libcfs_debug_msg(msgdata, "%s\n", key->lck_area);
 519                 key->lck_area[0] = 0;
 520         }
 521         va_end(args);
 522         return 0;
 523 }
 524 EXPORT_SYMBOL(lu_cdebug_printer);
 525
 526 /**
 527  * Print object header.
 528  */
 529 void lu_object_header_print(const struct lu_env *env, void *cookie,
 530                             lu_printer_t printer,
 531                             const struct lu_object_header *hdr)
 532 {
 533         (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
 534                    hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
 535                    PFID(&hdr->loh_fid),
 536                    hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
 537                    list_empty((struct list_head *)&hdr->loh_lru) ? \
 538                    "" : " lru",
 539                    hdr->loh_attr & LOHA_EXISTS ? " exist" : "");
 540 }
 541 EXPORT_SYMBOL(lu_object_header_print);
 542
 543 /**
 544  * Print human readable representation of the \a o to the \a printer.
 545  */
 546 void lu_object_print(const struct lu_env *env, void *cookie,
 547                      lu_printer_t printer, const struct lu_object *o)
 548 {
 549         static const char ruler[] = "........................................";
 550         struct lu_object_header *top;
 551         int depth = 4;
 552
 553         top = o->lo_header;
 554         lu_object_header_print(env, cookie, printer, top);
 555         (*printer)(env, cookie, "{\n");
 556
 557         list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 558                 /*
 559                  * print `.' \a depth times followed by type name and address
 560                  */
 561                 (*printer)(env, cookie, "%*.*s%s@%p", depth, depth, ruler,
 562                            o->lo_dev->ld_type->ldt_name, o);
 563
 564                 if (o->lo_ops->loo_object_print != NULL)
 565                         (*o->lo_ops->loo_object_print)(env, cookie, printer, o);
 566
 567                 (*printer)(env, cookie, "\n");
 568         }
 569
 570         (*printer)(env, cookie, "} header@%p\n", top);
 571 }
 572 EXPORT_SYMBOL(lu_object_print);
 573
 574 /**
 575  * Check object consistency.
 576  */
 577 int lu_object_invariant(const struct lu_object *o)
 578 {
 579         struct lu_object_header *top;
 580
 581         top = o->lo_header;
 582         list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 583                 if (o->lo_ops->loo_object_invariant != NULL &&
 584                     !o->lo_ops->loo_object_invariant(o))
 585                         return 0;
 586         }
 587         return 1;
 588 }
 589
 590 static struct lu_object *htable_lookup(struct lu_site *s,
 591                                        struct cfs_hash_bd *bd,
 592                                        const struct lu_fid *f,
 593                                        __u64 *version)
 594 {
 595         struct lu_site_bkt_data *bkt;
 596         struct lu_object_header *h;
 597         struct hlist_node *hnode;
 598         __u64 ver = cfs_hash_bd_version_get(bd);
 599
 600         if (*version == ver)
 601                 return ERR_PTR(-ENOENT);
 602
 603         *version = ver;
 604         bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
 605         /* cfs_hash_bd_peek_locked is a somehow "internal" function
 606          * of cfs_hash, it doesn't add refcount on object. */
 607         hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
 608         if (!hnode) {
 609                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
 610                 return ERR_PTR(-ENOENT);
 611         }
 612
 613         h = container_of0(hnode, struct lu_object_header, loh_hash);
 614         cfs_hash_get(s->ls_obj_hash, hnode);
 615         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
 616         if (!list_empty(&h->loh_lru)) {
 617                 list_del_init(&h->loh_lru);
 618                 bkt->lsb_lru_len--;
 619                 percpu_counter_dec(&s->ls_lru_len_counter);
 620         }
 621         return lu_object_top(h);
 622 }
 623
 624 /**
 625  * Search cache for an object with the fid \a f. If such object is found,
 626  * return it. Otherwise, create new object, insert it into cache and return
 627  * it. In any case, additional reference is acquired on the returned object.
 628  */
 629 struct lu_object *lu_object_find(const struct lu_env *env,
 630                                  struct lu_device *dev, const struct lu_fid *f,
 631                                  const struct lu_object_conf *conf)
 632 {
 633         return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
 634 }
 635 EXPORT_SYMBOL(lu_object_find);
 636
 637 /*
 638  * Limit the lu_object cache to a maximum of lu_cache_nr objects.  Because
 639  * the calculation for the number of objects to reclaim is not covered by
 640  * a lock the maximum number of objects is capped by LU_CACHE_MAX_ADJUST.
 641  * This ensures that many concurrent threads will not accidentally purge
 642  * the entire cache.
 643  */
 644 static void lu_object_limit(const struct lu_env *env,
 645                             struct lu_device *dev)
 646 {
 647         __u64 size, nr;
 648
 649         if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
 650                 return;
 651
 652         size = cfs_hash_size_get(dev->ld_site->ls_obj_hash);
 653         nr = (__u64)lu_cache_nr;
 654         if (size <= nr)
 655                 return;
 656
 657         lu_site_purge_objects(env, dev->ld_site,
 658                               MIN(size - nr, LU_CACHE_NR_MAX_ADJUST), 0);
 659 }
 660
 661 static struct lu_object *lu_object_new(const struct lu_env *env,
 662                                        struct lu_device *dev,
 663                                        const struct lu_fid *f,
 664                                        const struct lu_object_conf *conf)
 665 {
 666         struct lu_object *o;
 667         struct cfs_hash *hs;
 668         struct cfs_hash_bd bd;
 669
 670         o = lu_object_alloc(env, dev, f, conf);
 671         if (unlikely(IS_ERR(o)))
 672                 return o;
 673
 674         hs = dev->ld_site->ls_obj_hash;
 675         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 676         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 677         cfs_hash_bd_unlock(hs, &bd, 1);
 678
 679         lu_object_limit(env, dev);
 680
 681         return o;
 682 }
 683
 684 /**
 685  * Core logic of lu_object_find*() functions.
 686  *
 687  * Much like lu_object_find(), but top level device of object is specifically
 688  * \a dev rather than top level device of the site. This interface allows
 689  * objects of different "stacking" to be created within the same site.
 690  */
 691 struct lu_object *lu_object_find_at(const struct lu_env *env,
 692                                     struct lu_device *dev,
 693                                     const struct lu_fid *f,
 694                                     const struct lu_object_conf *conf)
 695 {
 696         struct lu_object *o;
 697         struct lu_object *shadow;
 698         struct lu_site *s;
 699         struct cfs_hash *hs;
 700         struct cfs_hash_bd bd;
 701         __u64 version = 0;
 702
 703         /*
 704          * This uses standard index maintenance protocol:
 705          *
 706          *     - search index under lock, and return object if found;
 707          *     - otherwise, unlock index, allocate new object;
 708          *     - lock index and search again;
 709          *     - if nothing is found (usual case), insert newly created
 710          *       object into index;
 711          *     - otherwise (race: other thread inserted object), free
 712          *       object just allocated.
 713          *     - unlock index;
 714          *     - return object.
 715          *
 716          * For "LOC_F_NEW" case, we are sure the object is new established.
 717          * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
 718          * just alloc and insert directly.
 719          *
 720          * If dying object is found during index search, add @waiter to the
 721          * site wait-queue and return ERR_PTR(-EAGAIN).
 722          */
 723         if (conf && conf->loc_flags & LOC_F_NEW)
 724                 return lu_object_new(env, dev, f, conf);
 725
 726         s  = dev->ld_site;
 727         hs = s->ls_obj_hash;
 728         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 729         o = htable_lookup(s, &bd, f, &version);
 730         cfs_hash_bd_unlock(hs, &bd, 1);
 731         if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT)
 732                 return o;
 733
 734         /*
 735          * Allocate new object. This may result in rather complicated
 736          * operations, including fld queries, inode loading, etc.
 737          */
 738         o = lu_object_alloc(env, dev, f, conf);
 739         if (unlikely(IS_ERR(o)))
 740                 return o;
 741
 742         LASSERT(lu_fid_eq(lu_object_fid(o), f));
 743
 744         cfs_hash_bd_lock(hs, &bd, 1);
 745
 746         shadow = htable_lookup(s, &bd, f, &version);
 747         if (likely(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT)) {
 748                 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 749                 cfs_hash_bd_unlock(hs, &bd, 1);
 750
 751                 lu_object_limit(env, dev);
 752
 753                 return o;
 754         }
 755
 756         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
 757         cfs_hash_bd_unlock(hs, &bd, 1);
 758         lu_object_free(env, o);
 759         return shadow;
 760 }
 761 EXPORT_SYMBOL(lu_object_find_at);
 762
 763 /**
 764  * Find object with given fid, and return its slice belonging to given device.
 765  */
 766 struct lu_object *lu_object_find_slice(const struct lu_env *env,
 767                                        struct lu_device *dev,
 768                                        const struct lu_fid *f,
 769                                        const struct lu_object_conf *conf)
 770 {
 771         struct lu_object *top;
 772         struct lu_object *obj;
 773
 774         top = lu_object_find(env, dev, f, conf);
 775         if (IS_ERR(top))
 776                 return top;
 777
 778         obj = lu_object_locate(top->lo_header, dev->ld_type);
 779         if (unlikely(obj == NULL)) {
 780                 lu_object_put(env, top);
 781                 obj = ERR_PTR(-ENOENT);
 782         }
 783
 784         return obj;
 785 }
 786 EXPORT_SYMBOL(lu_object_find_slice);
 787
 788 int lu_device_type_init(struct lu_device_type *ldt)
 789 {
 790         int result = 0;
 791
 792         atomic_set(&ldt->ldt_device_nr, 0);
 793         if (ldt->ldt_ops->ldto_init)
 794                 result = ldt->ldt_ops->ldto_init(ldt);
 795
 796         return result;
 797 }
 798 EXPORT_SYMBOL(lu_device_type_init);
 799
 800 void lu_device_type_fini(struct lu_device_type *ldt)
 801 {
 802         if (ldt->ldt_ops->ldto_fini)
 803                 ldt->ldt_ops->ldto_fini(ldt);
 804 }
 805 EXPORT_SYMBOL(lu_device_type_fini);
 806
 807 /**
 808  * Global list of all sites on this node
 809  */
 810 static LIST_HEAD(lu_sites);
 811 static DECLARE_RWSEM(lu_sites_guard);
 812
 813 /**
 814  * Global environment used by site shrinker.
 815  */
 816 static struct lu_env lu_shrink_env;
 817
 818 struct lu_site_print_arg {
 819         struct lu_env   *lsp_env;
 820         void            *lsp_cookie;
 821         lu_printer_t     lsp_printer;
 822 };
 823
 824 static int
 825 lu_site_obj_print(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 826                   struct hlist_node *hnode, void *data)
 827 {
 828         struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
 829         struct lu_object_header  *h;
 830
 831         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
 832         if (!list_empty(&h->loh_layers)) {
 833                 const struct lu_object *o;
 834
 835                 o = lu_object_top(h);
 836                 lu_object_print(arg->lsp_env, arg->lsp_cookie,
 837                                 arg->lsp_printer, o);
 838         } else {
 839                 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
 840                                        arg->lsp_printer, h);
 841         }
 842         return 0;
 843 }
 844
 845 /**
 846  * Print all objects in \a s.
 847  */
 848 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
 849                    lu_printer_t printer)
 850 {
 851         struct lu_site_print_arg arg = {
 852                 .lsp_env     = (struct lu_env *)env,
 853                 .lsp_cookie  = cookie,
 854                 .lsp_printer = printer,
 855         };
 856
 857         cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
 858 }
 859 EXPORT_SYMBOL(lu_site_print);
 860
 861 /**
 862  * Return desired hash table order.
 863  */
 864 static unsigned long lu_htable_order(struct lu_device *top)
 865 {
 866         unsigned long cache_size;
 867         unsigned long bits;
 868         unsigned long bits_max = LU_SITE_BITS_MAX;
 869
 870         /*
 871          * For ZFS based OSDs the cache should be disabled by default.  This
 872          * allows the ZFS ARC maximum flexibility in determining what buffers
 873          * to cache.  If Lustre has objects or buffer which it wants to ensure
 874          * always stay cached it must maintain a hold on them.
 875          */
 876         if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
 877                 lu_cache_percent = 1;
 878                 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
 879                 return LU_SITE_BITS_MIN;
 880         }
 881
 882         if (strcmp(top->ld_type->ldt_name, LUSTRE_VVP_NAME) == 0)
 883                 bits_max = LU_SITE_BITS_MAX_CL;
 884
 885         /*
 886          * Calculate hash table size, assuming that we want reasonable
 887          * performance when 20% of total memory is occupied by cache of
 888          * lu_objects.
 889          *
 890          * Size of lu_object is (arbitrary) taken as 1K (together with inode).
 891          */
 892         cache_size = totalram_pages;
 893
 894 #if BITS_PER_LONG == 32
 895         /* limit hashtable size for lowmem systems to low RAM */
 896         if (cache_size > 1 << (30 - PAGE_SHIFT))
 897                 cache_size = 1 << (30 - PAGE_SHIFT) * 3 / 4;
 898 #endif
 899
 900         /* clear off unreasonable cache setting. */
 901         if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
 902                 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
 903                       " the range of (0, %u]. Will use default value: %u.\n",
 904                       lu_cache_percent, LU_CACHE_PERCENT_MAX,
 905                       LU_CACHE_PERCENT_DEFAULT);
 906
 907                 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 908         }
 909         cache_size = cache_size / 100 * lu_cache_percent *
 910                 (PAGE_SIZE / 1024);
 911
 912         for (bits = 1; (1 << bits) < cache_size; ++bits) {
 913                 ;
 914         }
 915
 916         return clamp_t(typeof(bits), bits, LU_SITE_BITS_MIN, bits_max);
 917 }
 918
 919 static unsigned lu_obj_hop_hash(struct cfs_hash *hs,
 920                                 const void *key, unsigned mask)
 921 {
 922         struct lu_fid  *fid = (struct lu_fid *)key;
 923         __u32           hash;
 924
 925         hash = fid_flatten32(fid);
 926         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 927         hash = hash_long(hash, hs->hs_bkt_bits);
 928
 929         /* give me another random factor */
 930         hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
 931
 932         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 933         hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
 934
 935         return hash & mask;
 936 }
 937
 938 static void *lu_obj_hop_object(struct hlist_node *hnode)
 939 {
 940         return hlist_entry(hnode, struct lu_object_header, loh_hash);
 941 }
 942
 943 static void *lu_obj_hop_key(struct hlist_node *hnode)
 944 {
 945         struct lu_object_header *h;
 946
 947         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
 948         return &h->loh_fid;
 949 }
 950
 951 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
 952 {
 953         struct lu_object_header *h;
 954
 955         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
 956         return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
 957 }
 958
 959 static void lu_obj_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
 960 {
 961         struct lu_object_header *h;
 962
 963         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
 964         atomic_inc(&h->loh_ref);
 965 }
 966
 967 static void lu_obj_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
 968 {
 969         LBUG(); /* we should never called it */
 970 }
 971
 972 static struct cfs_hash_ops lu_site_hash_ops = {
 973         .hs_hash        = lu_obj_hop_hash,
 974         .hs_key         = lu_obj_hop_key,
 975         .hs_keycmp      = lu_obj_hop_keycmp,
 976         .hs_object      = lu_obj_hop_object,
 977         .hs_get         = lu_obj_hop_get,
 978         .hs_put_locked  = lu_obj_hop_put_locked,
 979 };
 980
 981 void lu_dev_add_linkage(struct lu_site *s, struct lu_device *d)
 982 {
 983         spin_lock(&s->ls_ld_lock);
 984         if (list_empty(&d->ld_linkage))
 985                 list_add(&d->ld_linkage, &s->ls_ld_linkage);
 986         spin_unlock(&s->ls_ld_lock);
 987 }
 988 EXPORT_SYMBOL(lu_dev_add_linkage);
 989
 990 void lu_dev_del_linkage(struct lu_site *s, struct lu_device *d)
 991 {
 992         spin_lock(&s->ls_ld_lock);
 993         list_del_init(&d->ld_linkage);
 994         spin_unlock(&s->ls_ld_lock);
 995 }
 996 EXPORT_SYMBOL(lu_dev_del_linkage);
 997
 998 /**
 999   * Initialize site \a s, with \a d as the top level device.
1000   */
1001 int lu_site_init(struct lu_site *s, struct lu_device *top)
1002 {
1003         struct lu_site_bkt_data *bkt;
1004         struct cfs_hash_bd bd;
1005         char name[16];
1006         unsigned long bits;
1007         unsigned int i;
1008         int rc;
1009         ENTRY;
1010
1011         memset(s, 0, sizeof *s);
1012         mutex_init(&s->ls_purge_mutex);
1013
1014 #ifdef HAVE_PERCPU_COUNTER_INIT_GFP_FLAG
1015         rc = percpu_counter_init(&s->ls_lru_len_counter, 0, GFP_NOFS);
1016 #else
1017         rc = percpu_counter_init(&s->ls_lru_len_counter, 0);
1018 #endif
1019         if (rc)
1020                 return -ENOMEM;
1021
1022         snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
1023         for (bits = lu_htable_order(top);
1024              bits >= LU_SITE_BITS_MIN; bits--) {
1025                 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1026                                                  bits - LU_SITE_BKT_BITS,
1027                                                  sizeof(*bkt), 0, 0,
1028                                                  &lu_site_hash_ops,
1029                                                  CFS_HASH_SPIN_BKTLOCK |
1030                                                  CFS_HASH_NO_ITEMREF |
1031                                                  CFS_HASH_DEPTH |
1032                                                  CFS_HASH_ASSERT_EMPTY |
1033                                                  CFS_HASH_COUNTER);
1034                 if (s->ls_obj_hash != NULL)
1035                         break;
1036         }
1037
1038         if (s->ls_obj_hash == NULL) {
1039                 CERROR("failed to create lu_site hash with bits: %lu\n", bits);
1040                 return -ENOMEM;
1041         }
1042
1043         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1044                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1045                 INIT_LIST_HEAD(&bkt->lsb_lru);
1046                 init_waitqueue_head(&bkt->lsb_marche_funebre);
1047         }
1048
1049         s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1050         if (s->ls_stats == NULL) {
1051                 cfs_hash_putref(s->ls_obj_hash);
1052                 s->ls_obj_hash = NULL;
1053                 return -ENOMEM;
1054         }
1055
1056         lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1057                              0, "created", "created");
1058         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1059                              0, "cache_hit", "cache_hit");
1060         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1061                              0, "cache_miss", "cache_miss");
1062         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1063                              0, "cache_race", "cache_race");
1064         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1065                              0, "cache_death_race", "cache_death_race");
1066         lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1067                              0, "lru_purged", "lru_purged");
1068
1069         INIT_LIST_HEAD(&s->ls_linkage);
1070         s->ls_top_dev = top;
1071         top->ld_site = s;
1072         lu_device_get(top);
1073         lu_ref_add(&top->ld_reference, "site-top", s);
1074
1075         INIT_LIST_HEAD(&s->ls_ld_linkage);
1076         spin_lock_init(&s->ls_ld_lock);
1077
1078         lu_dev_add_linkage(s, top);
1079
1080         RETURN(0);
1081 }
1082 EXPORT_SYMBOL(lu_site_init);
1083
1084 /**
1085  * Finalize \a s and release its resources.
1086  */
1087 void lu_site_fini(struct lu_site *s)
1088 {
1089         down_write(&lu_sites_guard);
1090         list_del_init(&s->ls_linkage);
1091         up_write(&lu_sites_guard);
1092
1093         percpu_counter_destroy(&s->ls_lru_len_counter);
1094
1095         if (s->ls_obj_hash != NULL) {
1096                 cfs_hash_putref(s->ls_obj_hash);
1097                 s->ls_obj_hash = NULL;
1098         }
1099
1100         if (s->ls_top_dev != NULL) {
1101                 s->ls_top_dev->ld_site = NULL;
1102                 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1103                 lu_device_put(s->ls_top_dev);
1104                 s->ls_top_dev = NULL;
1105         }
1106
1107         if (s->ls_stats != NULL)
1108                 lprocfs_free_stats(&s->ls_stats);
1109 }
1110 EXPORT_SYMBOL(lu_site_fini);
1111
1112 /**
1113  * Called when initialization of stack for this site is completed.
1114  */
1115 int lu_site_init_finish(struct lu_site *s)
1116 {
1117         int result;
1118         down_write(&lu_sites_guard);
1119         result = lu_context_refill(&lu_shrink_env.le_ctx);
1120         if (result == 0)
1121                 list_add(&s->ls_linkage, &lu_sites);
1122         up_write(&lu_sites_guard);
1123         return result;
1124 }
1125 EXPORT_SYMBOL(lu_site_init_finish);
1126
1127 /**
1128  * Acquire additional reference on device \a d
1129  */
1130 void lu_device_get(struct lu_device *d)
1131 {
1132         atomic_inc(&d->ld_ref);
1133 }
1134 EXPORT_SYMBOL(lu_device_get);
1135
1136 /**
1137  * Release reference on device \a d.
1138  */
1139 void lu_device_put(struct lu_device *d)
1140 {
1141         LASSERT(atomic_read(&d->ld_ref) > 0);
1142         atomic_dec(&d->ld_ref);
1143 }
1144 EXPORT_SYMBOL(lu_device_put);
1145
1146 /**
1147  * Initialize device \a d of type \a t.
1148  */
1149 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1150 {
1151         if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1152             t->ldt_ops->ldto_start != NULL)
1153                 t->ldt_ops->ldto_start(t);
1154
1155         memset(d, 0, sizeof *d);
1156         d->ld_type = t;
1157         lu_ref_init(&d->ld_reference);
1158         INIT_LIST_HEAD(&d->ld_linkage);
1159
1160         return 0;
1161 }
1162 EXPORT_SYMBOL(lu_device_init);
1163
1164 /**
1165  * Finalize device \a d.
1166  */
1167 void lu_device_fini(struct lu_device *d)
1168 {
1169         struct lu_device_type *t = d->ld_type;
1170
1171         if (d->ld_obd != NULL) {
1172                 d->ld_obd->obd_lu_dev = NULL;
1173                 d->ld_obd = NULL;
1174         }
1175
1176         lu_ref_fini(&d->ld_reference);
1177         LASSERTF(atomic_read(&d->ld_ref) == 0,
1178                  "Refcount is %u\n", atomic_read(&d->ld_ref));
1179         LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1180
1181         if (atomic_dec_and_test(&t->ldt_device_nr) &&
1182             t->ldt_ops->ldto_stop != NULL)
1183                 t->ldt_ops->ldto_stop(t);
1184 }
1185 EXPORT_SYMBOL(lu_device_fini);
1186
1187 /**
1188  * Initialize object \a o that is part of compound object \a h and was created
1189  * by device \a d.
1190  */
1191 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1192                    struct lu_device *d)
1193 {
1194         memset(o, 0, sizeof(*o));
1195         o->lo_header = h;
1196         o->lo_dev = d;
1197         lu_device_get(d);
1198         lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1199         INIT_LIST_HEAD(&o->lo_linkage);
1200
1201         return 0;
1202 }
1203 EXPORT_SYMBOL(lu_object_init);
1204
1205 /**
1206  * Finalize object and release its resources.
1207  */
1208 void lu_object_fini(struct lu_object *o)
1209 {
1210         struct lu_device *dev = o->lo_dev;
1211
1212         LASSERT(list_empty(&o->lo_linkage));
1213
1214         if (dev != NULL) {
1215                 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1216                               "lu_object", o);
1217                 lu_device_put(dev);
1218                 o->lo_dev = NULL;
1219         }
1220 }
1221 EXPORT_SYMBOL(lu_object_fini);
1222
1223 /**
1224  * Add object \a o as first layer of compound object \a h
1225  *
1226  * This is typically called by the ->ldo_object_alloc() method of top-level
1227  * device.
1228  */
1229 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1230 {
1231         list_move(&o->lo_linkage, &h->loh_layers);
1232 }
1233 EXPORT_SYMBOL(lu_object_add_top);
1234
1235 /**
1236  * Add object \a o as a layer of compound object, going after \a before.
1237  *
1238  * This is typically called by the ->ldo_object_alloc() method of \a
1239  * before->lo_dev.
1240  */
1241 void lu_object_add(struct lu_object *before, struct lu_object *o)
1242 {
1243         list_move(&o->lo_linkage, &before->lo_linkage);
1244 }
1245 EXPORT_SYMBOL(lu_object_add);
1246
1247 /**
1248  * Initialize compound object.
1249  */
1250 int lu_object_header_init(struct lu_object_header *h)
1251 {
1252         memset(h, 0, sizeof *h);
1253         atomic_set(&h->loh_ref, 1);
1254         INIT_HLIST_NODE(&h->loh_hash);
1255         INIT_LIST_HEAD(&h->loh_lru);
1256         INIT_LIST_HEAD(&h->loh_layers);
1257         lu_ref_init(&h->loh_reference);
1258         return 0;
1259 }
1260 EXPORT_SYMBOL(lu_object_header_init);
1261
1262 /**
1263  * Finalize compound object.
1264  */
1265 void lu_object_header_fini(struct lu_object_header *h)
1266 {
1267         LASSERT(list_empty(&h->loh_layers));
1268         LASSERT(list_empty(&h->loh_lru));
1269         LASSERT(hlist_unhashed(&h->loh_hash));
1270         lu_ref_fini(&h->loh_reference);
1271 }
1272 EXPORT_SYMBOL(lu_object_header_fini);
1273
1274 /**
1275  * Given a compound object, find its slice, corresponding to the device type
1276  * \a dtype.
1277  */
1278 struct lu_object *lu_object_locate(struct lu_object_header *h,
1279                                    const struct lu_device_type *dtype)
1280 {
1281         struct lu_object *o;
1282
1283         list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1284                 if (o->lo_dev->ld_type == dtype)
1285                         return o;
1286         }
1287         return NULL;
1288 }
1289 EXPORT_SYMBOL(lu_object_locate);
1290
1291 /**
1292  * Finalize and free devices in the device stack.
1293  *
1294  * Finalize device stack by purging object cache, and calling
1295  * lu_device_type_operations::ldto_device_fini() and
1296  * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1297  */
1298 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1299 {
1300         struct lu_site   *site = top->ld_site;
1301         struct lu_device *scan;
1302         struct lu_device *next;
1303
1304         lu_site_purge(env, site, ~0);
1305         for (scan = top; scan != NULL; scan = next) {
1306                 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1307                 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1308                 lu_device_put(scan);
1309         }
1310
1311         /* purge again. */
1312         lu_site_purge(env, site, ~0);
1313
1314         for (scan = top; scan != NULL; scan = next) {
1315                 const struct lu_device_type *ldt = scan->ld_type;
1316                 struct obd_type             *type;
1317
1318                 next = ldt->ldt_ops->ldto_device_free(env, scan);
1319                 type = ldt->ldt_obd_type;
1320                 if (type != NULL) {
1321                         type->typ_refcnt--;
1322                         class_put_type(type);
1323                 }
1324         }
1325 }
1326
1327 enum {
1328         /**
1329          * Maximal number of tld slots.
1330          */
1331         LU_CONTEXT_KEY_NR = 40
1332 };
1333
1334 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1335
1336 DEFINE_RWLOCK(lu_keys_guard);
1337 static atomic_t lu_key_initing_cnt = ATOMIC_INIT(0);
1338
1339 /**
1340  * Global counter incremented whenever key is registered, unregistered,
1341  * revived or quiesced. This is used to void unnecessary calls to
1342  * lu_context_refill(). No locking is provided, as initialization and shutdown
1343  * are supposed to be externally serialized.
1344  */
1345 static unsigned key_set_version = 0;
1346
1347 /**
1348  * Register new key.
1349  */
1350 int lu_context_key_register(struct lu_context_key *key)
1351 {
1352         int result;
1353         unsigned int i;
1354
1355         LASSERT(key->lct_init != NULL);
1356         LASSERT(key->lct_fini != NULL);
1357         LASSERT(key->lct_tags != 0);
1358         LASSERT(key->lct_owner != NULL);
1359
1360         result = -ENFILE;
1361         write_lock(&lu_keys_guard);
1362         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1363                 if (lu_keys[i] == NULL) {
1364                         key->lct_index = i;
1365                         atomic_set(&key->lct_used, 1);
1366                         lu_keys[i] = key;
1367                         lu_ref_init(&key->lct_reference);
1368                         result = 0;
1369                         ++key_set_version;
1370                         break;
1371                 }
1372         }
1373         write_unlock(&lu_keys_guard);
1374         return result;
1375 }
1376 EXPORT_SYMBOL(lu_context_key_register);
1377
1378 static void key_fini(struct lu_context *ctx, int index)
1379 {
1380         if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1381                 struct lu_context_key *key;
1382
1383                 key = lu_keys[index];
1384                 LASSERT(key != NULL);
1385                 LASSERT(key->lct_fini != NULL);
1386                 LASSERT(atomic_read(&key->lct_used) > 1);
1387
1388                 key->lct_fini(ctx, key, ctx->lc_value[index]);
1389                 lu_ref_del(&key->lct_reference, "ctx", ctx);
1390                 atomic_dec(&key->lct_used);
1391
1392                 LASSERT(key->lct_owner != NULL);
1393                 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1394                         LINVRNT(module_refcount(key->lct_owner) > 0);
1395                         module_put(key->lct_owner);
1396                 }
1397                 ctx->lc_value[index] = NULL;
1398         }
1399 }
1400
1401 /**
1402  * Deregister key.
1403  */
1404 void lu_context_key_degister(struct lu_context_key *key)
1405 {
1406         LASSERT(atomic_read(&key->lct_used) >= 1);
1407         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1408
1409         lu_context_key_quiesce(key);
1410
1411         write_lock(&lu_keys_guard);
1412         ++key_set_version;
1413         key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1414
1415         /**
1416          * Wait until all transient contexts referencing this key have
1417          * run lu_context_key::lct_fini() method.
1418          */
1419         while (atomic_read(&key->lct_used) > 1) {
1420                 write_unlock(&lu_keys_guard);
1421                 CDEBUG(D_INFO, "lu_context_key_degister: \"%s\" %p, %d\n",
1422                        key->lct_owner ? key->lct_owner->name : "", key,
1423                        atomic_read(&key->lct_used));
1424                 schedule();
1425                 write_lock(&lu_keys_guard);
1426         }
1427         if (lu_keys[key->lct_index]) {
1428                 lu_keys[key->lct_index] = NULL;
1429                 lu_ref_fini(&key->lct_reference);
1430         }
1431         write_unlock(&lu_keys_guard);
1432
1433         LASSERTF(atomic_read(&key->lct_used) == 1,
1434                  "key has instances: %d\n",
1435                  atomic_read(&key->lct_used));
1436 }
1437 EXPORT_SYMBOL(lu_context_key_degister);
1438
1439 /**
1440  * Register a number of keys. This has to be called after all keys have been
1441  * initialized by a call to LU_CONTEXT_KEY_INIT().
1442  */
1443 int lu_context_key_register_many(struct lu_context_key *k, ...)
1444 {
1445         struct lu_context_key *key = k;
1446         va_list args;
1447         int result;
1448
1449         va_start(args, k);
1450         do {
1451                 result = lu_context_key_register(key);
1452                 if (result)
1453                         break;
1454                 key = va_arg(args, struct lu_context_key *);
1455         } while (key != NULL);
1456         va_end(args);
1457
1458         if (result != 0) {
1459                 va_start(args, k);
1460                 while (k != key) {
1461                         lu_context_key_degister(k);
1462                         k = va_arg(args, struct lu_context_key *);
1463                 }
1464                 va_end(args);
1465         }
1466
1467         return result;
1468 }
1469 EXPORT_SYMBOL(lu_context_key_register_many);
1470
1471 /**
1472  * De-register a number of keys. This is a dual to
1473  * lu_context_key_register_many().
1474  */
1475 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1476 {
1477         va_list args;
1478
1479         va_start(args, k);
1480         do {
1481                 lu_context_key_degister(k);
1482                 k = va_arg(args, struct lu_context_key*);
1483         } while (k != NULL);
1484         va_end(args);
1485 }
1486 EXPORT_SYMBOL(lu_context_key_degister_many);
1487
1488 /**
1489  * Revive a number of keys.
1490  */
1491 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1492 {
1493         va_list args;
1494
1495         va_start(args, k);
1496         do {
1497                 lu_context_key_revive(k);
1498                 k = va_arg(args, struct lu_context_key*);
1499         } while (k != NULL);
1500         va_end(args);
1501 }
1502 EXPORT_SYMBOL(lu_context_key_revive_many);
1503
1504 /**
1505  * Quiescent a number of keys.
1506  */
1507 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1508 {
1509         va_list args;
1510
1511         va_start(args, k);
1512         do {
1513                 lu_context_key_quiesce(k);
1514                 k = va_arg(args, struct lu_context_key*);
1515         } while (k != NULL);
1516         va_end(args);
1517 }
1518 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1519
1520 /**
1521  * Return value associated with key \a key in context \a ctx.
1522  */
1523 void *lu_context_key_get(const struct lu_context *ctx,
1524                          const struct lu_context_key *key)
1525 {
1526         LINVRNT(ctx->lc_state == LCS_ENTERED);
1527         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1528         LASSERT(lu_keys[key->lct_index] == key);
1529         return ctx->lc_value[key->lct_index];
1530 }
1531 EXPORT_SYMBOL(lu_context_key_get);
1532
1533 /**
1534  * List of remembered contexts. XXX document me.
1535  */
1536 static LIST_HEAD(lu_context_remembered);
1537
1538 /**
1539  * Destroy \a key in all remembered contexts. This is used to destroy key
1540  * values in "shared" contexts (like service threads), when a module owning
1541  * the key is about to be unloaded.
1542  */
1543 void lu_context_key_quiesce(struct lu_context_key *key)
1544 {
1545         struct lu_context *ctx;
1546
1547         if (!(key->lct_tags & LCT_QUIESCENT)) {
1548                 /*
1549                  * XXX memory barrier has to go here.
1550                  */
1551                 write_lock(&lu_keys_guard);
1552                 key->lct_tags |= LCT_QUIESCENT;
1553
1554                 /**
1555                  * Wait until all lu_context_key::lct_init() methods
1556                  * have completed.
1557                  */
1558                 while (atomic_read(&lu_key_initing_cnt) > 0) {
1559                         write_unlock(&lu_keys_guard);
1560                         CDEBUG(D_INFO, "lu_context_key_quiesce: \"%s\""
1561                                " %p, %d (%d)\n",
1562                                key->lct_owner ? key->lct_owner->name : "",
1563                                key, atomic_read(&key->lct_used),
1564                                atomic_read(&lu_key_initing_cnt));
1565                         schedule();
1566                         write_lock(&lu_keys_guard);
1567                 }
1568
1569                 list_for_each_entry(ctx, &lu_context_remembered,
1570                                     lc_remember)
1571                         key_fini(ctx, key->lct_index);
1572
1573                 ++key_set_version;
1574                 write_unlock(&lu_keys_guard);
1575         }
1576 }
1577
1578 void lu_context_key_revive(struct lu_context_key *key)
1579 {
1580         write_lock(&lu_keys_guard);
1581         key->lct_tags &= ~LCT_QUIESCENT;
1582         ++key_set_version;
1583         write_unlock(&lu_keys_guard);
1584 }
1585
1586 static void keys_fini(struct lu_context *ctx)
1587 {
1588         unsigned int i;
1589
1590         if (ctx->lc_value == NULL)
1591                 return;
1592
1593         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1594                 key_fini(ctx, i);
1595
1596         OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1597         ctx->lc_value = NULL;
1598 }
1599
1600 static int keys_fill(struct lu_context *ctx)
1601 {
1602         unsigned int i;
1603         unsigned pre_version;
1604
1605         /*
1606          * A serialisation with lu_context_key_quiesce() is needed, but some
1607          * "key->lct_init()" are calling kernel memory allocation routine and
1608          * can't be called while holding a spin_lock.
1609          * "lu_keys_guard" is held while incrementing "lu_key_initing_cnt"
1610          * to ensure the start of the serialisation.
1611          * An atomic_t variable is still used, in order not to reacquire the
1612          * lock when decrementing the counter.
1613          */
1614         read_lock(&lu_keys_guard);
1615         atomic_inc(&lu_key_initing_cnt);
1616         pre_version = key_set_version;
1617         read_unlock(&lu_keys_guard);
1618
1619 refill:
1620         LINVRNT(ctx->lc_value != NULL);
1621         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1622                 struct lu_context_key *key;
1623
1624                 key = lu_keys[i];
1625                 if (ctx->lc_value[i] == NULL && key != NULL &&
1626                     (key->lct_tags & ctx->lc_tags) &&
1627                     /*
1628                      * Don't create values for a LCT_QUIESCENT key, as this
1629                      * will pin module owning a key.
1630                      */
1631                     !(key->lct_tags & LCT_QUIESCENT)) {
1632                         void *value;
1633
1634                         LINVRNT(key->lct_init != NULL);
1635                         LINVRNT(key->lct_index == i);
1636
1637                         LASSERT(key->lct_owner != NULL);
1638                         if (!(ctx->lc_tags & LCT_NOREF) &&
1639                             try_module_get(key->lct_owner) == 0) {
1640                                 /* module is unloading, skip this key */
1641                                 continue;
1642                         }
1643
1644                         value = key->lct_init(ctx, key);
1645                         if (unlikely(IS_ERR(value))) {
1646                                 atomic_dec(&lu_key_initing_cnt);
1647                                 return PTR_ERR(value);
1648                         }
1649
1650                         lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1651                         atomic_inc(&key->lct_used);
1652                         /*
1653                          * This is the only place in the code, where an
1654                          * element of ctx->lc_value[] array is set to non-NULL
1655                          * value.
1656                          */
1657                         ctx->lc_value[i] = value;
1658                         if (key->lct_exit != NULL)
1659                                 ctx->lc_tags |= LCT_HAS_EXIT;
1660                 }
1661         }
1662
1663         read_lock(&lu_keys_guard);
1664         if (pre_version != key_set_version) {
1665                 pre_version = key_set_version;
1666                 read_unlock(&lu_keys_guard);
1667                 goto refill;
1668         }
1669
1670         ctx->lc_version = key_set_version;
1671
1672         atomic_dec(&lu_key_initing_cnt);
1673         read_unlock(&lu_keys_guard);
1674         return 0;
1675 }
1676
1677 static int keys_init(struct lu_context *ctx)
1678 {
1679         OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1680         if (likely(ctx->lc_value != NULL))
1681                 return keys_fill(ctx);
1682
1683         return -ENOMEM;
1684 }
1685
1686 /**
1687  * Initialize context data-structure. Create values for all keys.
1688  */
1689 int lu_context_init(struct lu_context *ctx, __u32 tags)
1690 {
1691         int     rc;
1692
1693         memset(ctx, 0, sizeof *ctx);
1694         ctx->lc_state = LCS_INITIALIZED;
1695         ctx->lc_tags = tags;
1696         if (tags & LCT_REMEMBER) {
1697                 write_lock(&lu_keys_guard);
1698                 list_add(&ctx->lc_remember, &lu_context_remembered);
1699                 write_unlock(&lu_keys_guard);
1700         } else {
1701                 INIT_LIST_HEAD(&ctx->lc_remember);
1702         }
1703
1704         rc = keys_init(ctx);
1705         if (rc != 0)
1706                 lu_context_fini(ctx);
1707
1708         return rc;
1709 }
1710 EXPORT_SYMBOL(lu_context_init);
1711
1712 /**
1713  * Finalize context data-structure. Destroy key values.
1714  */
1715 void lu_context_fini(struct lu_context *ctx)
1716 {
1717         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1718         ctx->lc_state = LCS_FINALIZED;
1719
1720         if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1721                 LASSERT(list_empty(&ctx->lc_remember));
1722                 keys_fini(ctx);
1723
1724         } else { /* could race with key degister */
1725                 write_lock(&lu_keys_guard);
1726                 keys_fini(ctx);
1727                 list_del_init(&ctx->lc_remember);
1728                 write_unlock(&lu_keys_guard);
1729         }
1730 }
1731 EXPORT_SYMBOL(lu_context_fini);
1732
1733 /**
1734  * Called before entering context.
1735  */
1736 void lu_context_enter(struct lu_context *ctx)
1737 {
1738         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1739         ctx->lc_state = LCS_ENTERED;
1740 }
1741 EXPORT_SYMBOL(lu_context_enter);
1742
1743 /**
1744  * Called after exiting from \a ctx
1745  */
1746 void lu_context_exit(struct lu_context *ctx)
1747 {
1748         unsigned int i;
1749
1750         LINVRNT(ctx->lc_state == LCS_ENTERED);
1751         ctx->lc_state = LCS_LEFT;
1752         if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1753                 /* could race with key quiescency */
1754                 if (ctx->lc_tags & LCT_REMEMBER)
1755                         read_lock(&lu_keys_guard);
1756
1757                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1758                         if (ctx->lc_value[i] != NULL) {
1759                                 struct lu_context_key *key;
1760
1761                                 key = lu_keys[i];
1762                                 LASSERT(key != NULL);
1763                                 if (key->lct_exit != NULL)
1764                                         key->lct_exit(ctx,
1765                                                       key, ctx->lc_value[i]);
1766                         }
1767                 }
1768
1769                 if (ctx->lc_tags & LCT_REMEMBER)
1770                         read_unlock(&lu_keys_guard);
1771         }
1772 }
1773 EXPORT_SYMBOL(lu_context_exit);
1774
1775 /**
1776  * Allocate for context all missing keys that were registered after context
1777  * creation. key_set_version is only changed in rare cases when modules
1778  * are loaded and removed.
1779  */
1780 int lu_context_refill(struct lu_context *ctx)
1781 {
1782         read_lock(&lu_keys_guard);
1783         if (likely(ctx->lc_version == key_set_version)) {
1784                 read_unlock(&lu_keys_guard);
1785                 return 0;
1786         }
1787
1788         read_unlock(&lu_keys_guard);
1789         return keys_fill(ctx);
1790 }
1791
1792 /**
1793  * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1794  * obd being added. Currently, this is only used on client side, specifically
1795  * for echo device client, for other stack (like ptlrpc threads), context are
1796  * predefined when the lu_device type are registered, during the module probe
1797  * phase.
1798  */
1799 __u32 lu_context_tags_default = 0;
1800 __u32 lu_session_tags_default = 0;
1801
1802 void lu_context_tags_update(__u32 tags)
1803 {
1804         write_lock(&lu_keys_guard);
1805         lu_context_tags_default |= tags;
1806         key_set_version++;
1807         write_unlock(&lu_keys_guard);
1808 }
1809 EXPORT_SYMBOL(lu_context_tags_update);
1810
1811 void lu_context_tags_clear(__u32 tags)
1812 {
1813         write_lock(&lu_keys_guard);
1814         lu_context_tags_default &= ~tags;
1815         key_set_version++;
1816         write_unlock(&lu_keys_guard);
1817 }
1818 EXPORT_SYMBOL(lu_context_tags_clear);
1819
1820 void lu_session_tags_update(__u32 tags)
1821 {
1822         write_lock(&lu_keys_guard);
1823         lu_session_tags_default |= tags;
1824         key_set_version++;
1825         write_unlock(&lu_keys_guard);
1826 }
1827 EXPORT_SYMBOL(lu_session_tags_update);
1828
1829 void lu_session_tags_clear(__u32 tags)
1830 {
1831         write_lock(&lu_keys_guard);
1832         lu_session_tags_default &= ~tags;
1833         key_set_version++;
1834         write_unlock(&lu_keys_guard);
1835 }
1836 EXPORT_SYMBOL(lu_session_tags_clear);
1837
1838 int lu_env_init(struct lu_env *env, __u32 tags)
1839 {
1840         int result;
1841
1842         env->le_ses = NULL;
1843         result = lu_context_init(&env->le_ctx, tags);
1844         if (likely(result == 0))
1845                 lu_context_enter(&env->le_ctx);
1846         return result;
1847 }
1848 EXPORT_SYMBOL(lu_env_init);
1849
1850 void lu_env_fini(struct lu_env *env)
1851 {
1852         lu_context_exit(&env->le_ctx);
1853         lu_context_fini(&env->le_ctx);
1854         env->le_ses = NULL;
1855 }
1856 EXPORT_SYMBOL(lu_env_fini);
1857
1858 int lu_env_refill(struct lu_env *env)
1859 {
1860         int result;
1861
1862         result = lu_context_refill(&env->le_ctx);
1863         if (result == 0 && env->le_ses != NULL)
1864                 result = lu_context_refill(env->le_ses);
1865         return result;
1866 }
1867 EXPORT_SYMBOL(lu_env_refill);
1868
1869 /**
1870  * Currently, this API will only be used by echo client.
1871  * Because echo client and normal lustre client will share
1872  * same cl_env cache. So echo client needs to refresh
1873  * the env context after it get one from the cache, especially
1874  * when normal client and echo client co-exist in the same client.
1875  */
1876 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1877                           __u32 stags)
1878 {
1879         int    result;
1880
1881         if ((env->le_ctx.lc_tags & ctags) != ctags) {
1882                 env->le_ctx.lc_version = 0;
1883                 env->le_ctx.lc_tags |= ctags;
1884         }
1885
1886         if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1887                 env->le_ses->lc_version = 0;
1888                 env->le_ses->lc_tags |= stags;
1889         }
1890
1891         result = lu_env_refill(env);
1892
1893         return result;
1894 }
1895 EXPORT_SYMBOL(lu_env_refill_by_tags);
1896
1897 static struct shrinker *lu_site_shrinker;
1898
1899 typedef struct lu_site_stats{
1900         unsigned        lss_populated;
1901         unsigned        lss_max_search;
1902         unsigned        lss_total;
1903         unsigned        lss_busy;
1904 } lu_site_stats_t;
1905
1906 static void lu_site_stats_get(struct cfs_hash *hs,
1907                               lu_site_stats_t *stats, int populated)
1908 {
1909         struct cfs_hash_bd bd;
1910         unsigned int  i;
1911
1912         cfs_hash_for_each_bucket(hs, &bd, i) {
1913                 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1914                 struct hlist_head       *hhead;
1915
1916                 cfs_hash_bd_lock(hs, &bd, 1);
1917                 stats->lss_busy  +=
1918                         cfs_hash_bd_count_get(&bd) - bkt->lsb_lru_len;
1919                 stats->lss_total += cfs_hash_bd_count_get(&bd);
1920                 stats->lss_max_search = max((int)stats->lss_max_search,
1921                                             cfs_hash_bd_depmax_get(&bd));
1922                 if (!populated) {
1923                         cfs_hash_bd_unlock(hs, &bd, 1);
1924                         continue;
1925                 }
1926
1927                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1928                         if (!hlist_empty(hhead))
1929                                 stats->lss_populated++;
1930                 }
1931                 cfs_hash_bd_unlock(hs, &bd, 1);
1932         }
1933 }
1934
1935
1936 /*
1937  * lu_cache_shrink_count() returns an approximate number of cached objects
1938  * that can be freed by shrink_slab(). A counter, which tracks the
1939  * number of items in the site's lru, is maintained in a percpu_counter
1940  * for each site. The percpu values are incremented and decremented as
1941  * objects are added or removed from the lru. The percpu values are summed
1942  * and saved whenever a percpu value exceeds a threshold. Thus the saved,
1943  * summed value at any given time may not accurately reflect the current
1944  * lru length. But this value is sufficiently accurate for the needs of
1945  * a shrinker.
1946  *
1947  * Using a per cpu counter is a compromise solution to concurrent access:
1948  * lu_object_put() can update the counter without locking the site and
1949  * lu_cache_shrink_count can sum the counters without locking each
1950  * ls_obj_hash bucket.
1951  */
1952 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
1953                                            struct shrink_control *sc)
1954 {
1955         struct lu_site *s;
1956         struct lu_site *tmp;
1957         unsigned long cached = 0;
1958
1959         if (!(sc->gfp_mask & __GFP_FS))
1960                 return 0;
1961
1962         down_read(&lu_sites_guard);
1963         list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage)
1964                 cached += percpu_counter_read_positive(&s->ls_lru_len_counter);
1965         up_read(&lu_sites_guard);
1966
1967         cached = (cached / 100) * sysctl_vfs_cache_pressure;
1968         CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
1969                cached, sysctl_vfs_cache_pressure);
1970
1971         return cached;
1972 }
1973
1974 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
1975                                           struct shrink_control *sc)
1976 {
1977         struct lu_site *s;
1978         struct lu_site *tmp;
1979         unsigned long remain = sc->nr_to_scan;
1980         LIST_HEAD(splice);
1981
1982         if (!(sc->gfp_mask & __GFP_FS))
1983                 /* We must not take the lu_sites_guard lock when
1984                  * __GFP_FS is *not* set because of the deadlock
1985                  * possibility detailed above. Additionally,
1986                  * since we cannot determine the number of
1987                  * objects in the cache without taking this
1988                  * lock, we're in a particularly tough spot. As
1989                  * a result, we'll just lie and say our cache is
1990                  * empty. This _should_ be ok, as we can't
1991                  * reclaim objects when __GFP_FS is *not* set
1992                  * anyways.
1993                  */
1994                 return SHRINK_STOP;
1995
1996         down_write(&lu_sites_guard);
1997         list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1998                 remain = lu_site_purge(&lu_shrink_env, s, remain);
1999                 /*
2000                  * Move just shrunk site to the tail of site list to
2001                  * assure shrinking fairness.
2002                  */
2003                 list_move_tail(&s->ls_linkage, &splice);
2004         }
2005         list_splice(&splice, lu_sites.prev);
2006         up_write(&lu_sites_guard);
2007
2008         return sc->nr_to_scan - remain;
2009 }
2010
2011 #ifndef HAVE_SHRINKER_COUNT
2012 /*
2013  * There exists a potential lock inversion deadlock scenario when using
2014  * Lustre on top of ZFS. This occurs between one of ZFS's
2015  * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2016  * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2017  * while thread B will take the ht_lock and sleep on the lu_sites_guard
2018  * lock. Obviously neither thread will wake and drop their respective hold
2019  * on their lock.
2020  *
2021  * To prevent this from happening we must ensure the lu_sites_guard lock is
2022  * not taken while down this code path. ZFS reliably does not set the
2023  * __GFP_FS bit in its code paths, so this can be used to determine if it
2024  * is safe to take the lu_sites_guard lock.
2025  *
2026  * Ideally we should accurately return the remaining number of cached
2027  * objects without taking the lu_sites_guard lock, but this is not
2028  * possible in the current implementation.
2029  */
2030 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2031 {
2032         int cached = 0;
2033         struct shrink_control scv = {
2034                  .nr_to_scan = shrink_param(sc, nr_to_scan),
2035                  .gfp_mask   = shrink_param(sc, gfp_mask)
2036         };
2037 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2038         struct shrinker* shrinker = NULL;
2039 #endif
2040
2041
2042         CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2043
2044         if (scv.nr_to_scan != 0)
2045                 lu_cache_shrink_scan(shrinker, &scv);
2046
2047         cached = lu_cache_shrink_count(shrinker, &scv);
2048         return cached;
2049 }
2050
2051 #endif /* HAVE_SHRINKER_COUNT */
2052
2053
2054 /*
2055  * Debugging stuff.
2056  */
2057
2058 /**
2059  * Environment to be used in debugger, contains all tags.
2060  */
2061 static struct lu_env lu_debugging_env;
2062
2063 /**
2064  * Debugging printer function using printk().
2065  */
2066 int lu_printk_printer(const struct lu_env *env,
2067                       void *unused, const char *format, ...)
2068 {
2069         va_list args;
2070
2071         va_start(args, format);
2072         vprintk(format, args);
2073         va_end(args);
2074         return 0;
2075 }
2076
2077 int lu_debugging_setup(void)
2078 {
2079         return lu_env_init(&lu_debugging_env, ~0);
2080 }
2081
2082 void lu_context_keys_dump(void)
2083 {
2084         unsigned int i;
2085
2086         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2087                 struct lu_context_key *key;
2088
2089                 key = lu_keys[i];
2090                 if (key != NULL) {
2091                         CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2092                                i, key, key->lct_tags,
2093                                key->lct_init, key->lct_fini, key->lct_exit,
2094                                key->lct_index, atomic_read(&key->lct_used),
2095                                key->lct_owner ? key->lct_owner->name : "",
2096                                key->lct_owner);
2097                         lu_ref_print(&key->lct_reference);
2098                 }
2099         }
2100 }
2101
2102 /**
2103  * Initialization of global lu_* data.
2104  */
2105 int lu_global_init(void)
2106 {
2107         int result;
2108         DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2109                          lu_cache_shrink_count, lu_cache_shrink_scan);
2110
2111         CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2112
2113         result = lu_ref_global_init();
2114         if (result != 0)
2115                 return result;
2116
2117         LU_CONTEXT_KEY_INIT(&lu_global_key);
2118         result = lu_context_key_register(&lu_global_key);
2119         if (result != 0)
2120                 return result;
2121
2122         /*
2123          * At this level, we don't know what tags are needed, so allocate them
2124          * conservatively. This should not be too bad, because this
2125          * environment is global.
2126          */
2127         down_write(&lu_sites_guard);
2128         result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2129         up_write(&lu_sites_guard);
2130         if (result != 0)
2131                 return result;
2132
2133         /*
2134          * seeks estimation: 3 seeks to read a record from oi, one to read
2135          * inode, one for ea. Unfortunately setting this high value results in
2136          * lu_object/inode cache consuming all the memory.
2137          */
2138         lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2139         if (lu_site_shrinker == NULL)
2140                 return -ENOMEM;
2141
2142         return result;
2143 }
2144
2145 /**
2146  * Dual to lu_global_init().
2147  */
2148 void lu_global_fini(void)
2149 {
2150         if (lu_site_shrinker != NULL) {
2151                 remove_shrinker(lu_site_shrinker);
2152                 lu_site_shrinker = NULL;
2153         }
2154
2155         lu_context_key_degister(&lu_global_key);
2156
2157         /*
2158          * Tear shrinker environment down _after_ de-registering
2159          * lu_global_key, because the latter has a value in the former.
2160          */
2161         down_write(&lu_sites_guard);
2162         lu_env_fini(&lu_shrink_env);
2163         up_write(&lu_sites_guard);
2164
2165         lu_ref_global_fini();
2166 }
2167
2168 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2169 {
2170 #ifdef CONFIG_PROC_FS
2171         struct lprocfs_counter ret;
2172
2173         lprocfs_stats_collect(stats, idx, &ret);
2174         return (__u32)ret.lc_count;
2175 #else
2176         return 0;
2177 #endif
2178 }
2179
2180 /**
2181  * Output site statistical counters into a buffer. Suitable for
2182  * lprocfs_rd_*()-style functions.
2183  */
2184 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2185 {
2186         lu_site_stats_t stats;
2187
2188         memset(&stats, 0, sizeof(stats));
2189         lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2190
2191         seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2192                    stats.lss_busy,
2193                    stats.lss_total,
2194                    stats.lss_populated,
2195                    CFS_HASH_NHLIST(s->ls_obj_hash),
2196                    stats.lss_max_search,
2197                    ls_stats_read(s->ls_stats, LU_SS_CREATED),
2198                    ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2199                    ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2200                    ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2201                    ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2202                    ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2203         return 0;
2204 }
2205 EXPORT_SYMBOL(lu_site_stats_seq_print);
2206
2207 /**
2208  * Helper function to initialize a number of kmem slab caches at once.
2209  */
2210 int lu_kmem_init(struct lu_kmem_descr *caches)
2211 {
2212         int result;
2213         struct lu_kmem_descr *iter = caches;
2214
2215         for (result = 0; iter->ckd_cache != NULL; ++iter) {
2216                 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2217                                                      iter->ckd_size,
2218                                                      0, 0, NULL);
2219                 if (*iter->ckd_cache == NULL) {
2220                         result = -ENOMEM;
2221                         /* free all previously allocated caches */
2222                         lu_kmem_fini(caches);
2223                         break;
2224                 }
2225         }
2226         return result;
2227 }
2228 EXPORT_SYMBOL(lu_kmem_init);
2229
2230 /**
2231  * Helper function to finalize a number of kmem slab cached at once. Dual to
2232  * lu_kmem_init().
2233  */
2234 void lu_kmem_fini(struct lu_kmem_descr *caches)
2235 {
2236         for (; caches->ckd_cache != NULL; ++caches) {
2237                 if (*caches->ckd_cache != NULL) {
2238                         kmem_cache_destroy(*caches->ckd_cache);
2239                         *caches->ckd_cache = NULL;
2240                 }
2241         }
2242 }
2243 EXPORT_SYMBOL(lu_kmem_fini);
2244
2245 /**
2246  * Temporary solution to be able to assign fid in ->do_create()
2247  * till we have fully-functional OST fids
2248  */
2249 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2250                           const struct lu_fid *fid)
2251 {
2252         struct lu_site          *s = o->lo_dev->ld_site;
2253         struct lu_fid           *old = &o->lo_header->loh_fid;
2254         struct cfs_hash         *hs;
2255         struct cfs_hash_bd       bd;
2256
2257         LASSERT(fid_is_zero(old));
2258
2259         /* supposed to be unique */
2260         hs = s->ls_obj_hash;
2261         cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2262 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
2263         {
2264                 __u64 version = 0;
2265                 struct lu_object *shadow;
2266
2267                 shadow = htable_lookup(s, &bd, fid, &version);
2268                 /* supposed to be unique */
2269                 LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2270         }
2271 #endif
2272         *old = *fid;
2273         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2274         cfs_hash_bd_unlock(hs, &bd, 1);
2275 }
2276 EXPORT_SYMBOL(lu_object_assign_fid);
2277
2278 /**
2279  * allocates object with 0 (non-assiged) fid
2280  * XXX: temporary solution to be able to assign fid in ->do_create()
2281  *      till we have fully-functional OST fids
2282  */
2283 struct lu_object *lu_object_anon(const struct lu_env *env,
2284                                  struct lu_device *dev,
2285                                  const struct lu_object_conf *conf)
2286 {
2287         struct lu_fid     fid;
2288         struct lu_object *o;
2289
2290         fid_zero(&fid);
2291         o = lu_object_alloc(env, dev, &fid, conf);
2292
2293         return o;
2294 }
2295 EXPORT_SYMBOL(lu_object_anon);
2296
2297 struct lu_buf LU_BUF_NULL = {
2298         .lb_buf = NULL,
2299         .lb_len = 0
2300 };
2301 EXPORT_SYMBOL(LU_BUF_NULL);
2302
2303 void lu_buf_free(struct lu_buf *buf)
2304 {
2305         LASSERT(buf);
2306         if (buf->lb_buf) {
2307                 LASSERT(buf->lb_len > 0);
2308                 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2309                 buf->lb_buf = NULL;
2310                 buf->lb_len = 0;
2311         }
2312 }
2313 EXPORT_SYMBOL(lu_buf_free);
2314
2315 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2316 {
2317         LASSERT(buf);
2318         LASSERT(buf->lb_buf == NULL);
2319         LASSERT(buf->lb_len == 0);
2320         OBD_ALLOC_LARGE(buf->lb_buf, size);
2321         if (likely(buf->lb_buf))
2322                 buf->lb_len = size;
2323 }
2324 EXPORT_SYMBOL(lu_buf_alloc);
2325
2326 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2327 {
2328         lu_buf_free(buf);
2329         lu_buf_alloc(buf, size);
2330 }
2331 EXPORT_SYMBOL(lu_buf_realloc);
2332
2333 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2334 {
2335         if (buf->lb_buf == NULL && buf->lb_len == 0)
2336                 lu_buf_alloc(buf, len);
2337
2338         if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2339                 lu_buf_realloc(buf, len);
2340
2341         return buf;
2342 }
2343 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2344
2345 /**
2346  * Increase the size of the \a buf.
2347  * preserves old data in buffer
2348  * old buffer remains unchanged on error
2349  * \retval 0 or -ENOMEM
2350  */
2351 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2352 {
2353         char *ptr;
2354
2355         if (len <= buf->lb_len)
2356                 return 0;
2357
2358         OBD_ALLOC_LARGE(ptr, len);
2359         if (ptr == NULL)
2360                 return -ENOMEM;
2361
2362         /* Free the old buf */
2363         if (buf->lb_buf != NULL) {
2364                 memcpy(ptr, buf->lb_buf, buf->lb_len);
2365                 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2366         }
2367
2368         buf->lb_buf = ptr;
2369         buf->lb_len = len;
2370         return 0;
2371 }
2372 EXPORT_SYMBOL(lu_buf_check_and_grow);