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, 2016, 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 <obd_class.h>
  48 #include <obd_support.h>
  49 #include <lustre_disk.h>
  50 #include <lustre_fid.h>
  51 #include <lu_object.h>
  52 #include <lu_ref.h>
  53
  54 enum {
  55         LU_CACHE_PERCENT_MAX     = 50,
  56         LU_CACHE_PERCENT_DEFAULT = 20
  57 };
  58
  59 #define LU_CACHE_NR_MAX_ADJUST          512
  60 #define LU_CACHE_NR_UNLIMITED           -1
  61 #define LU_CACHE_NR_DEFAULT             LU_CACHE_NR_UNLIMITED
  62 #define LU_CACHE_NR_LDISKFS_LIMIT       LU_CACHE_NR_UNLIMITED
  63 /** This is set to roughly (20 * OSS_NTHRS_MAX) to prevent thrashing */
  64 #define LU_CACHE_NR_ZFS_LIMIT           10240
  65
  66 #define LU_SITE_BITS_MIN    12
  67 #define LU_SITE_BITS_MAX    24
  68 #define LU_SITE_BITS_MAX_CL 19
  69 /**
  70  * total 256 buckets, we don't want too many buckets because:
  71  * - consume too much memory
  72  * - avoid unbalanced LRU list
  73  */
  74 #define LU_SITE_BKT_BITS    8
  75
  76
  77 static unsigned int lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
  78 module_param(lu_cache_percent, int, 0644);
  79 MODULE_PARM_DESC(lu_cache_percent, "Percentage of memory to be used as lu_object cache");
  80
  81 static long lu_cache_nr = LU_CACHE_NR_DEFAULT;
  82 module_param(lu_cache_nr, long, 0644);
  83 MODULE_PARM_DESC(lu_cache_nr, "Maximum number of objects in lu_object cache");
  84
  85 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
  86 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx);
  87
  88 /**
  89  * Decrease reference counter on object. If last reference is freed, return
  90  * object to the cache, unless lu_object_is_dying(o) holds. In the latter
  91  * case, free object immediately.
  92  */
  93 void lu_object_put(const struct lu_env *env, struct lu_object *o)
  94 {
  95         struct lu_site_bkt_data *bkt;
  96         struct lu_object_header *top;
  97         struct lu_site          *site;
  98         struct lu_object        *orig;
  99         struct cfs_hash_bd            bd;
 100         const struct lu_fid     *fid;
 101
 102         top  = o->lo_header;
 103         site = o->lo_dev->ld_site;
 104         orig = o;
 105
 106         /*
 107          * till we have full fids-on-OST implemented anonymous objects
 108          * are possible in OSP. such an object isn't listed in the site
 109          * so we should not remove it from the site.
 110          */
 111         fid = lu_object_fid(o);
 112         if (fid_is_zero(fid)) {
 113                 LASSERT(top->loh_hash.next == NULL
 114                         && top->loh_hash.pprev == NULL);
 115                 LASSERT(list_empty(&top->loh_lru));
 116                 if (!atomic_dec_and_test(&top->loh_ref))
 117                         return;
 118                 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
 119                         if (o->lo_ops->loo_object_release != NULL)
 120                                 o->lo_ops->loo_object_release(env, o);
 121                 }
 122                 lu_object_free(env, orig);
 123                 return;
 124         }
 125
 126         cfs_hash_bd_get(site->ls_obj_hash, &top->loh_fid, &bd);
 127         bkt = cfs_hash_bd_extra_get(site->ls_obj_hash, &bd);
 128
 129         if (!cfs_hash_bd_dec_and_lock(site->ls_obj_hash, &bd, &top->loh_ref)) {
 130                 if (lu_object_is_dying(top)) {
 131
 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                                        wait_queue_t *waiter,
 594                                        __u64 *version)
 595 {
 596         struct lu_site_bkt_data *bkt;
 597         struct lu_object_header *h;
 598         struct hlist_node       *hnode;
 599         __u64  ver = cfs_hash_bd_version_get(bd);
 600
 601         if (*version == ver)
 602                 return ERR_PTR(-ENOENT);
 603
 604         *version = ver;
 605         bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
 606         /* cfs_hash_bd_peek_locked is a somehow "internal" function
 607          * of cfs_hash, it doesn't add refcount on object. */
 608         hnode = cfs_hash_bd_peek_locked(s->ls_obj_hash, bd, (void *)f);
 609         if (hnode == NULL) {
 610                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_MISS);
 611                 return ERR_PTR(-ENOENT);
 612         }
 613
 614         h = container_of0(hnode, struct lu_object_header, loh_hash);
 615         if (likely(!lu_object_is_dying(h))) {
 616                 cfs_hash_get(s->ls_obj_hash, hnode);
 617                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_HIT);
 618                 if (!list_empty(&h->loh_lru)) {
 619                         list_del_init(&h->loh_lru);
 620                         bkt->lsb_lru_len--;
 621                         percpu_counter_dec(&s->ls_lru_len_counter);
 622                 }
 623                 return lu_object_top(h);
 624         }
 625
 626         /*
 627          * Lookup found an object being destroyed this object cannot be
 628          * returned (to assure that references to dying objects are eventually
 629          * drained), and moreover, lookup has to wait until object is freed.
 630          */
 631
 632         if (likely(waiter != NULL)) {
 633                 init_waitqueue_entry(waiter, current);
 634                 add_wait_queue(&bkt->lsb_marche_funebre, waiter);
 635                 set_current_state(TASK_UNINTERRUPTIBLE);
 636                 lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
 637         }
 638
 639         return ERR_PTR(-EAGAIN);
 640 }
 641
 642 /**
 643  * Search cache for an object with the fid \a f. If such object is found,
 644  * return it. Otherwise, create new object, insert it into cache and return
 645  * it. In any case, additional reference is acquired on the returned object.
 646  */
 647 struct lu_object *lu_object_find(const struct lu_env *env,
 648                                  struct lu_device *dev, const struct lu_fid *f,
 649                                  const struct lu_object_conf *conf)
 650 {
 651         return lu_object_find_at(env, dev->ld_site->ls_top_dev, f, conf);
 652 }
 653 EXPORT_SYMBOL(lu_object_find);
 654
 655 /*
 656  * Limit the lu_object cache to a maximum of lu_cache_nr objects.  Because
 657  * the calculation for the number of objects to reclaim is not covered by
 658  * a lock the maximum number of objects is capped by LU_CACHE_MAX_ADJUST.
 659  * This ensures that many concurrent threads will not accidentally purge
 660  * the entire cache.
 661  */
 662 static void lu_object_limit(const struct lu_env *env,
 663                             struct lu_device *dev)
 664 {
 665         __u64 size, nr;
 666
 667         if (lu_cache_nr == LU_CACHE_NR_UNLIMITED)
 668                 return;
 669
 670         size = cfs_hash_size_get(dev->ld_site->ls_obj_hash);
 671         nr = (__u64)lu_cache_nr;
 672         if (size <= nr)
 673                 return;
 674
 675         lu_site_purge_objects(env, dev->ld_site,
 676                               MIN(size - nr, LU_CACHE_NR_MAX_ADJUST), 0);
 677 }
 678
 679 static struct lu_object *lu_object_new(const struct lu_env *env,
 680                                        struct lu_device *dev,
 681                                        const struct lu_fid *f,
 682                                        const struct lu_object_conf *conf)
 683 {
 684         struct lu_object        *o;
 685         struct cfs_hash              *hs;
 686         struct cfs_hash_bd            bd;
 687
 688         o = lu_object_alloc(env, dev, f, conf);
 689         if (unlikely(IS_ERR(o)))
 690                 return o;
 691
 692         hs = dev->ld_site->ls_obj_hash;
 693         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 694         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 695         cfs_hash_bd_unlock(hs, &bd, 1);
 696
 697         lu_object_limit(env, dev);
 698
 699         return o;
 700 }
 701
 702 /**
 703  * Core logic of lu_object_find*() functions.
 704  */
 705 static struct lu_object *lu_object_find_try(const struct lu_env *env,
 706                                             struct lu_device *dev,
 707                                             const struct lu_fid *f,
 708                                             const struct lu_object_conf *conf,
 709                                             wait_queue_t *waiter)
 710 {
 711         struct lu_object      *o;
 712         struct lu_object      *shadow;
 713         struct lu_site        *s;
 714         struct cfs_hash            *hs;
 715         struct cfs_hash_bd          bd;
 716         __u64                  version = 0;
 717
 718         /*
 719          * This uses standard index maintenance protocol:
 720          *
 721          *     - search index under lock, and return object if found;
 722          *     - otherwise, unlock index, allocate new object;
 723          *     - lock index and search again;
 724          *     - if nothing is found (usual case), insert newly created
 725          *       object into index;
 726          *     - otherwise (race: other thread inserted object), free
 727          *       object just allocated.
 728          *     - unlock index;
 729          *     - return object.
 730          *
 731          * For "LOC_F_NEW" case, we are sure the object is new established.
 732          * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
 733          * just alloc and insert directly.
 734          *
 735          * If dying object is found during index search, add @waiter to the
 736          * site wait-queue and return ERR_PTR(-EAGAIN).
 737          */
 738         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
 739                 return lu_object_new(env, dev, f, conf);
 740
 741         s  = dev->ld_site;
 742         hs = s->ls_obj_hash;
 743         cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
 744         o = htable_lookup(s, &bd, f, waiter, &version);
 745         cfs_hash_bd_unlock(hs, &bd, 1);
 746         if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT)
 747                 return o;
 748
 749         /*
 750          * Allocate new object. This may result in rather complicated
 751          * operations, including fld queries, inode loading, etc.
 752          */
 753         o = lu_object_alloc(env, dev, f, conf);
 754         if (unlikely(IS_ERR(o)))
 755                 return o;
 756
 757         LASSERT(lu_fid_eq(lu_object_fid(o), f));
 758
 759         cfs_hash_bd_lock(hs, &bd, 1);
 760
 761         shadow = htable_lookup(s, &bd, f, waiter, &version);
 762         if (likely(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT)) {
 763                 cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
 764                 cfs_hash_bd_unlock(hs, &bd, 1);
 765
 766                 lu_object_limit(env, dev);
 767
 768                 return o;
 769         }
 770
 771         lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_RACE);
 772         cfs_hash_bd_unlock(hs, &bd, 1);
 773         lu_object_free(env, o);
 774         return shadow;
 775 }
 776
 777 /**
 778  * Much like lu_object_find(), but top level device of object is specifically
 779  * \a dev rather than top level device of the site. This interface allows
 780  * objects of different "stacking" to be created within the same site.
 781  */
 782 struct lu_object *lu_object_find_at(const struct lu_env *env,
 783                                     struct lu_device *dev,
 784                                     const struct lu_fid *f,
 785                                     const struct lu_object_conf *conf)
 786 {
 787         struct lu_site_bkt_data *bkt;
 788         struct lu_object        *obj;
 789         wait_queue_t           wait;
 790
 791         if (conf != NULL && conf->loc_flags & LOC_F_NOWAIT)
 792                 return lu_object_find_try(env, dev, f, conf, NULL);
 793
 794         while (1) {
 795                 obj = lu_object_find_try(env, dev, f, conf, &wait);
 796                 if (obj != ERR_PTR(-EAGAIN))
 797                         return obj;
 798                 /*
 799                  * lu_object_find_try() already added waiter into the
 800                  * wait queue.
 801                  */
 802                 schedule();
 803                 bkt = lu_site_bkt_from_fid(dev->ld_site, (void *)f);
 804                 remove_wait_queue(&bkt->lsb_marche_funebre, &wait);
 805         }
 806 }
 807 EXPORT_SYMBOL(lu_object_find_at);
 808
 809 /**
 810  * Find object with given fid, and return its slice belonging to given device.
 811  */
 812 struct lu_object *lu_object_find_slice(const struct lu_env *env,
 813                                        struct lu_device *dev,
 814                                        const struct lu_fid *f,
 815                                        const struct lu_object_conf *conf)
 816 {
 817         struct lu_object *top;
 818         struct lu_object *obj;
 819
 820         top = lu_object_find(env, dev, f, conf);
 821         if (IS_ERR(top))
 822                 return top;
 823
 824         obj = lu_object_locate(top->lo_header, dev->ld_type);
 825         if (unlikely(obj == NULL)) {
 826                 lu_object_put(env, top);
 827                 obj = ERR_PTR(-ENOENT);
 828         }
 829
 830         return obj;
 831 }
 832 EXPORT_SYMBOL(lu_object_find_slice);
 833
 834 int lu_device_type_init(struct lu_device_type *ldt)
 835 {
 836         int result = 0;
 837
 838         atomic_set(&ldt->ldt_device_nr, 0);
 839         if (ldt->ldt_ops->ldto_init)
 840                 result = ldt->ldt_ops->ldto_init(ldt);
 841
 842         return result;
 843 }
 844 EXPORT_SYMBOL(lu_device_type_init);
 845
 846 void lu_device_type_fini(struct lu_device_type *ldt)
 847 {
 848         if (ldt->ldt_ops->ldto_fini)
 849                 ldt->ldt_ops->ldto_fini(ldt);
 850 }
 851 EXPORT_SYMBOL(lu_device_type_fini);
 852
 853 /**
 854  * Global list of all sites on this node
 855  */
 856 static LIST_HEAD(lu_sites);
 857 static DECLARE_RWSEM(lu_sites_guard);
 858
 859 /**
 860  * Global environment used by site shrinker.
 861  */
 862 static struct lu_env lu_shrink_env;
 863
 864 struct lu_site_print_arg {
 865         struct lu_env   *lsp_env;
 866         void            *lsp_cookie;
 867         lu_printer_t     lsp_printer;
 868 };
 869
 870 static int
 871 lu_site_obj_print(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 872                   struct hlist_node *hnode, void *data)
 873 {
 874         struct lu_site_print_arg *arg = (struct lu_site_print_arg *)data;
 875         struct lu_object_header  *h;
 876
 877         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
 878         if (!list_empty(&h->loh_layers)) {
 879                 const struct lu_object *o;
 880
 881                 o = lu_object_top(h);
 882                 lu_object_print(arg->lsp_env, arg->lsp_cookie,
 883                                 arg->lsp_printer, o);
 884         } else {
 885                 lu_object_header_print(arg->lsp_env, arg->lsp_cookie,
 886                                        arg->lsp_printer, h);
 887         }
 888         return 0;
 889 }
 890
 891 /**
 892  * Print all objects in \a s.
 893  */
 894 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
 895                    lu_printer_t printer)
 896 {
 897         struct lu_site_print_arg arg = {
 898                 .lsp_env     = (struct lu_env *)env,
 899                 .lsp_cookie  = cookie,
 900                 .lsp_printer = printer,
 901         };
 902
 903         cfs_hash_for_each(s->ls_obj_hash, lu_site_obj_print, &arg);
 904 }
 905 EXPORT_SYMBOL(lu_site_print);
 906
 907 /**
 908  * Return desired hash table order.
 909  */
 910 static unsigned long lu_htable_order(struct lu_device *top)
 911 {
 912         unsigned long cache_size;
 913         unsigned long bits;
 914         unsigned long bits_max = LU_SITE_BITS_MAX;
 915
 916         /*
 917          * For ZFS based OSDs the cache should be disabled by default.  This
 918          * allows the ZFS ARC maximum flexibility in determining what buffers
 919          * to cache.  If Lustre has objects or buffer which it wants to ensure
 920          * always stay cached it must maintain a hold on them.
 921          */
 922         if (strcmp(top->ld_type->ldt_name, LUSTRE_OSD_ZFS_NAME) == 0) {
 923                 lu_cache_percent = 1;
 924                 lu_cache_nr = LU_CACHE_NR_ZFS_LIMIT;
 925                 return LU_SITE_BITS_MIN;
 926         }
 927
 928         if (strcmp(top->ld_type->ldt_name, LUSTRE_VVP_NAME) == 0)
 929                 bits_max = LU_SITE_BITS_MAX_CL;
 930
 931         /*
 932          * Calculate hash table size, assuming that we want reasonable
 933          * performance when 20% of total memory is occupied by cache of
 934          * lu_objects.
 935          *
 936          * Size of lu_object is (arbitrary) taken as 1K (together with inode).
 937          */
 938         cache_size = totalram_pages;
 939
 940 #if BITS_PER_LONG == 32
 941         /* limit hashtable size for lowmem systems to low RAM */
 942         if (cache_size > 1 << (30 - PAGE_SHIFT))
 943                 cache_size = 1 << (30 - PAGE_SHIFT) * 3 / 4;
 944 #endif
 945
 946         /* clear off unreasonable cache setting. */
 947         if (lu_cache_percent == 0 || lu_cache_percent > LU_CACHE_PERCENT_MAX) {
 948                 CWARN("obdclass: invalid lu_cache_percent: %u, it must be in"
 949                       " the range of (0, %u]. Will use default value: %u.\n",
 950                       lu_cache_percent, LU_CACHE_PERCENT_MAX,
 951                       LU_CACHE_PERCENT_DEFAULT);
 952
 953                 lu_cache_percent = LU_CACHE_PERCENT_DEFAULT;
 954         }
 955         cache_size = cache_size / 100 * lu_cache_percent *
 956                 (PAGE_SIZE / 1024);
 957
 958         for (bits = 1; (1 << bits) < cache_size; ++bits) {
 959                 ;
 960         }
 961
 962         return clamp_t(typeof(bits), bits, LU_SITE_BITS_MIN, bits_max);
 963 }
 964
 965 static unsigned lu_obj_hop_hash(struct cfs_hash *hs,
 966                                 const void *key, unsigned mask)
 967 {
 968         struct lu_fid  *fid = (struct lu_fid *)key;
 969         __u32           hash;
 970
 971         hash = fid_flatten32(fid);
 972         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
 973         hash = hash_long(hash, hs->hs_bkt_bits);
 974
 975         /* give me another random factor */
 976         hash -= hash_long((unsigned long)hs, fid_oid(fid) % 11 + 3);
 977
 978         hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
 979         hash |= (fid_seq(fid) + fid_oid(fid)) & (CFS_HASH_NBKT(hs) - 1);
 980
 981         return hash & mask;
 982 }
 983
 984 static void *lu_obj_hop_object(struct hlist_node *hnode)
 985 {
 986         return hlist_entry(hnode, struct lu_object_header, loh_hash);
 987 }
 988
 989 static void *lu_obj_hop_key(struct hlist_node *hnode)
 990 {
 991         struct lu_object_header *h;
 992
 993         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
 994         return &h->loh_fid;
 995 }
 996
 997 static int lu_obj_hop_keycmp(const void *key, struct hlist_node *hnode)
 998 {
 999         struct lu_object_header *h;
1000
1001         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1002         return lu_fid_eq(&h->loh_fid, (struct lu_fid *)key);
1003 }
1004
1005 static void lu_obj_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
1006 {
1007         struct lu_object_header *h;
1008
1009         h = hlist_entry(hnode, struct lu_object_header, loh_hash);
1010         atomic_inc(&h->loh_ref);
1011 }
1012
1013 static void lu_obj_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
1014 {
1015         LBUG(); /* we should never called it */
1016 }
1017
1018 static struct cfs_hash_ops 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         struct cfs_hash_bd bd;
1051         char name[16];
1052         unsigned long bits;
1053         unsigned int i;
1054         int rc;
1055         ENTRY;
1056
1057         memset(s, 0, sizeof *s);
1058         mutex_init(&s->ls_purge_mutex);
1059
1060 #ifdef HAVE_PERCPU_COUNTER_INIT_GFP_FLAG
1061         rc = percpu_counter_init(&s->ls_lru_len_counter, 0, GFP_NOFS);
1062 #else
1063         rc = percpu_counter_init(&s->ls_lru_len_counter, 0);
1064 #endif
1065         if (rc)
1066                 return -ENOMEM;
1067
1068         snprintf(name, sizeof(name), "lu_site_%s", top->ld_type->ldt_name);
1069         for (bits = lu_htable_order(top);
1070              bits >= LU_SITE_BITS_MIN; bits--) {
1071                 s->ls_obj_hash = cfs_hash_create(name, bits, bits,
1072                                                  bits - LU_SITE_BKT_BITS,
1073                                                  sizeof(*bkt), 0, 0,
1074                                                  &lu_site_hash_ops,
1075                                                  CFS_HASH_SPIN_BKTLOCK |
1076                                                  CFS_HASH_NO_ITEMREF |
1077                                                  CFS_HASH_DEPTH |
1078                                                  CFS_HASH_ASSERT_EMPTY |
1079                                                  CFS_HASH_COUNTER);
1080                 if (s->ls_obj_hash != NULL)
1081                         break;
1082         }
1083
1084         if (s->ls_obj_hash == NULL) {
1085                 CERROR("failed to create lu_site hash with bits: %lu\n", bits);
1086                 return -ENOMEM;
1087         }
1088
1089         cfs_hash_for_each_bucket(s->ls_obj_hash, &bd, i) {
1090                 bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, &bd);
1091                 INIT_LIST_HEAD(&bkt->lsb_lru);
1092                 init_waitqueue_head(&bkt->lsb_marche_funebre);
1093         }
1094
1095         s->ls_stats = lprocfs_alloc_stats(LU_SS_LAST_STAT, 0);
1096         if (s->ls_stats == NULL) {
1097                 cfs_hash_putref(s->ls_obj_hash);
1098                 s->ls_obj_hash = NULL;
1099                 return -ENOMEM;
1100         }
1101
1102         lprocfs_counter_init(s->ls_stats, LU_SS_CREATED,
1103                              0, "created", "created");
1104         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_HIT,
1105                              0, "cache_hit", "cache_hit");
1106         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_MISS,
1107                              0, "cache_miss", "cache_miss");
1108         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_RACE,
1109                              0, "cache_race", "cache_race");
1110         lprocfs_counter_init(s->ls_stats, LU_SS_CACHE_DEATH_RACE,
1111                              0, "cache_death_race", "cache_death_race");
1112         lprocfs_counter_init(s->ls_stats, LU_SS_LRU_PURGED,
1113                              0, "lru_purged", "lru_purged");
1114
1115         INIT_LIST_HEAD(&s->ls_linkage);
1116         s->ls_top_dev = top;
1117         top->ld_site = s;
1118         lu_device_get(top);
1119         lu_ref_add(&top->ld_reference, "site-top", s);
1120
1121         INIT_LIST_HEAD(&s->ls_ld_linkage);
1122         spin_lock_init(&s->ls_ld_lock);
1123
1124         lu_dev_add_linkage(s, top);
1125
1126         RETURN(0);
1127 }
1128 EXPORT_SYMBOL(lu_site_init);
1129
1130 /**
1131  * Finalize \a s and release its resources.
1132  */
1133 void lu_site_fini(struct lu_site *s)
1134 {
1135         down_write(&lu_sites_guard);
1136         list_del_init(&s->ls_linkage);
1137         up_write(&lu_sites_guard);
1138
1139         percpu_counter_destroy(&s->ls_lru_len_counter);
1140
1141         if (s->ls_obj_hash != NULL) {
1142                 cfs_hash_putref(s->ls_obj_hash);
1143                 s->ls_obj_hash = NULL;
1144         }
1145
1146         if (s->ls_top_dev != NULL) {
1147                 s->ls_top_dev->ld_site = NULL;
1148                 lu_ref_del(&s->ls_top_dev->ld_reference, "site-top", s);
1149                 lu_device_put(s->ls_top_dev);
1150                 s->ls_top_dev = NULL;
1151         }
1152
1153         if (s->ls_stats != NULL)
1154                 lprocfs_free_stats(&s->ls_stats);
1155 }
1156 EXPORT_SYMBOL(lu_site_fini);
1157
1158 /**
1159  * Called when initialization of stack for this site is completed.
1160  */
1161 int lu_site_init_finish(struct lu_site *s)
1162 {
1163         int result;
1164         down_write(&lu_sites_guard);
1165         result = lu_context_refill(&lu_shrink_env.le_ctx);
1166         if (result == 0)
1167                 list_add(&s->ls_linkage, &lu_sites);
1168         up_write(&lu_sites_guard);
1169         return result;
1170 }
1171 EXPORT_SYMBOL(lu_site_init_finish);
1172
1173 /**
1174  * Acquire additional reference on device \a d
1175  */
1176 void lu_device_get(struct lu_device *d)
1177 {
1178         atomic_inc(&d->ld_ref);
1179 }
1180 EXPORT_SYMBOL(lu_device_get);
1181
1182 /**
1183  * Release reference on device \a d.
1184  */
1185 void lu_device_put(struct lu_device *d)
1186 {
1187         LASSERT(atomic_read(&d->ld_ref) > 0);
1188         atomic_dec(&d->ld_ref);
1189 }
1190 EXPORT_SYMBOL(lu_device_put);
1191
1192 /**
1193  * Initialize device \a d of type \a t.
1194  */
1195 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
1196 {
1197         if (atomic_inc_return(&t->ldt_device_nr) == 1 &&
1198             t->ldt_ops->ldto_start != NULL)
1199                 t->ldt_ops->ldto_start(t);
1200
1201         memset(d, 0, sizeof *d);
1202         d->ld_type = t;
1203         lu_ref_init(&d->ld_reference);
1204         INIT_LIST_HEAD(&d->ld_linkage);
1205
1206         return 0;
1207 }
1208 EXPORT_SYMBOL(lu_device_init);
1209
1210 /**
1211  * Finalize device \a d.
1212  */
1213 void lu_device_fini(struct lu_device *d)
1214 {
1215         struct lu_device_type *t = d->ld_type;
1216
1217         if (d->ld_obd != NULL) {
1218                 d->ld_obd->obd_lu_dev = NULL;
1219                 d->ld_obd = NULL;
1220         }
1221
1222         lu_ref_fini(&d->ld_reference);
1223         LASSERTF(atomic_read(&d->ld_ref) == 0,
1224                  "Refcount is %u\n", atomic_read(&d->ld_ref));
1225         LASSERT(atomic_read(&t->ldt_device_nr) > 0);
1226
1227         if (atomic_dec_and_test(&t->ldt_device_nr) &&
1228             t->ldt_ops->ldto_stop != NULL)
1229                 t->ldt_ops->ldto_stop(t);
1230 }
1231 EXPORT_SYMBOL(lu_device_fini);
1232
1233 /**
1234  * Initialize object \a o that is part of compound object \a h and was created
1235  * by device \a d.
1236  */
1237 int lu_object_init(struct lu_object *o, struct lu_object_header *h,
1238                    struct lu_device *d)
1239 {
1240         memset(o, 0, sizeof(*o));
1241         o->lo_header = h;
1242         o->lo_dev = d;
1243         lu_device_get(d);
1244         lu_ref_add_at(&d->ld_reference, &o->lo_dev_ref, "lu_object", o);
1245         INIT_LIST_HEAD(&o->lo_linkage);
1246
1247         return 0;
1248 }
1249 EXPORT_SYMBOL(lu_object_init);
1250
1251 /**
1252  * Finalize object and release its resources.
1253  */
1254 void lu_object_fini(struct lu_object *o)
1255 {
1256         struct lu_device *dev = o->lo_dev;
1257
1258         LASSERT(list_empty(&o->lo_linkage));
1259
1260         if (dev != NULL) {
1261                 lu_ref_del_at(&dev->ld_reference, &o->lo_dev_ref,
1262                               "lu_object", o);
1263                 lu_device_put(dev);
1264                 o->lo_dev = NULL;
1265         }
1266 }
1267 EXPORT_SYMBOL(lu_object_fini);
1268
1269 /**
1270  * Add object \a o as first layer of compound object \a h
1271  *
1272  * This is typically called by the ->ldo_object_alloc() method of top-level
1273  * device.
1274  */
1275 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
1276 {
1277         list_move(&o->lo_linkage, &h->loh_layers);
1278 }
1279 EXPORT_SYMBOL(lu_object_add_top);
1280
1281 /**
1282  * Add object \a o as a layer of compound object, going after \a before.
1283  *
1284  * This is typically called by the ->ldo_object_alloc() method of \a
1285  * before->lo_dev.
1286  */
1287 void lu_object_add(struct lu_object *before, struct lu_object *o)
1288 {
1289         list_move(&o->lo_linkage, &before->lo_linkage);
1290 }
1291 EXPORT_SYMBOL(lu_object_add);
1292
1293 /**
1294  * Initialize compound object.
1295  */
1296 int lu_object_header_init(struct lu_object_header *h)
1297 {
1298         memset(h, 0, sizeof *h);
1299         atomic_set(&h->loh_ref, 1);
1300         INIT_HLIST_NODE(&h->loh_hash);
1301         INIT_LIST_HEAD(&h->loh_lru);
1302         INIT_LIST_HEAD(&h->loh_layers);
1303         lu_ref_init(&h->loh_reference);
1304         return 0;
1305 }
1306 EXPORT_SYMBOL(lu_object_header_init);
1307
1308 /**
1309  * Finalize compound object.
1310  */
1311 void lu_object_header_fini(struct lu_object_header *h)
1312 {
1313         LASSERT(list_empty(&h->loh_layers));
1314         LASSERT(list_empty(&h->loh_lru));
1315         LASSERT(hlist_unhashed(&h->loh_hash));
1316         lu_ref_fini(&h->loh_reference);
1317 }
1318 EXPORT_SYMBOL(lu_object_header_fini);
1319
1320 /**
1321  * Given a compound object, find its slice, corresponding to the device type
1322  * \a dtype.
1323  */
1324 struct lu_object *lu_object_locate(struct lu_object_header *h,
1325                                    const struct lu_device_type *dtype)
1326 {
1327         struct lu_object *o;
1328
1329         list_for_each_entry(o, &h->loh_layers, lo_linkage) {
1330                 if (o->lo_dev->ld_type == dtype)
1331                         return o;
1332         }
1333         return NULL;
1334 }
1335 EXPORT_SYMBOL(lu_object_locate);
1336
1337 /**
1338  * Finalize and free devices in the device stack.
1339  *
1340  * Finalize device stack by purging object cache, and calling
1341  * lu_device_type_operations::ldto_device_fini() and
1342  * lu_device_type_operations::ldto_device_free() on all devices in the stack.
1343  */
1344 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
1345 {
1346         struct lu_site   *site = top->ld_site;
1347         struct lu_device *scan;
1348         struct lu_device *next;
1349
1350         lu_site_purge(env, site, ~0);
1351         for (scan = top; scan != NULL; scan = next) {
1352                 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
1353                 lu_ref_del(&scan->ld_reference, "lu-stack", &lu_site_init);
1354                 lu_device_put(scan);
1355         }
1356
1357         /* purge again. */
1358         lu_site_purge(env, site, ~0);
1359
1360         for (scan = top; scan != NULL; scan = next) {
1361                 const struct lu_device_type *ldt = scan->ld_type;
1362                 struct obd_type             *type;
1363
1364                 next = ldt->ldt_ops->ldto_device_free(env, scan);
1365                 type = ldt->ldt_obd_type;
1366                 if (type != NULL) {
1367                         type->typ_refcnt--;
1368                         class_put_type(type);
1369                 }
1370         }
1371 }
1372
1373 enum {
1374         /**
1375          * Maximal number of tld slots.
1376          */
1377         LU_CONTEXT_KEY_NR = 40
1378 };
1379
1380 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
1381
1382 DEFINE_RWLOCK(lu_keys_guard);
1383 static atomic_t lu_key_initing_cnt = ATOMIC_INIT(0);
1384
1385 /**
1386  * Global counter incremented whenever key is registered, unregistered,
1387  * revived or quiesced. This is used to void unnecessary calls to
1388  * lu_context_refill(). No locking is provided, as initialization and shutdown
1389  * are supposed to be externally serialized.
1390  */
1391 static unsigned key_set_version = 0;
1392
1393 /**
1394  * Register new key.
1395  */
1396 int lu_context_key_register(struct lu_context_key *key)
1397 {
1398         int result;
1399         unsigned int i;
1400
1401         LASSERT(key->lct_init != NULL);
1402         LASSERT(key->lct_fini != NULL);
1403         LASSERT(key->lct_tags != 0);
1404         LASSERT(key->lct_owner != NULL);
1405
1406         result = -ENFILE;
1407         write_lock(&lu_keys_guard);
1408         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1409                 if (lu_keys[i] == NULL) {
1410                         key->lct_index = i;
1411                         atomic_set(&key->lct_used, 1);
1412                         lu_keys[i] = key;
1413                         lu_ref_init(&key->lct_reference);
1414                         result = 0;
1415                         ++key_set_version;
1416                         break;
1417                 }
1418         }
1419         write_unlock(&lu_keys_guard);
1420         return result;
1421 }
1422 EXPORT_SYMBOL(lu_context_key_register);
1423
1424 static void key_fini(struct lu_context *ctx, int index)
1425 {
1426         if (ctx->lc_value != NULL && ctx->lc_value[index] != NULL) {
1427                 struct lu_context_key *key;
1428
1429                 key = lu_keys[index];
1430                 LASSERT(key != NULL);
1431                 LASSERT(key->lct_fini != NULL);
1432                 LASSERT(atomic_read(&key->lct_used) > 1);
1433
1434                 key->lct_fini(ctx, key, ctx->lc_value[index]);
1435                 lu_ref_del(&key->lct_reference, "ctx", ctx);
1436                 atomic_dec(&key->lct_used);
1437
1438                 LASSERT(key->lct_owner != NULL);
1439                 if ((ctx->lc_tags & LCT_NOREF) == 0) {
1440                         LINVRNT(module_refcount(key->lct_owner) > 0);
1441                         module_put(key->lct_owner);
1442                 }
1443                 ctx->lc_value[index] = NULL;
1444         }
1445 }
1446
1447 /**
1448  * Deregister key.
1449  */
1450 void lu_context_key_degister(struct lu_context_key *key)
1451 {
1452         LASSERT(atomic_read(&key->lct_used) >= 1);
1453         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1454
1455         lu_context_key_quiesce(key);
1456
1457         ++key_set_version;
1458         write_lock(&lu_keys_guard);
1459         key_fini(&lu_shrink_env.le_ctx, key->lct_index);
1460
1461         /**
1462          * Wait until all transient contexts referencing this key have
1463          * run lu_context_key::lct_fini() method.
1464          */
1465         while (atomic_read(&key->lct_used) > 1) {
1466                 write_unlock(&lu_keys_guard);
1467                 CDEBUG(D_INFO, "lu_context_key_degister: \"%s\" %p, %d\n",
1468                        key->lct_owner ? key->lct_owner->name : "", key,
1469                        atomic_read(&key->lct_used));
1470                 schedule();
1471                 write_lock(&lu_keys_guard);
1472         }
1473         if (lu_keys[key->lct_index]) {
1474                 lu_keys[key->lct_index] = NULL;
1475                 lu_ref_fini(&key->lct_reference);
1476         }
1477         write_unlock(&lu_keys_guard);
1478
1479         LASSERTF(atomic_read(&key->lct_used) == 1,
1480                  "key has instances: %d\n",
1481                  atomic_read(&key->lct_used));
1482 }
1483 EXPORT_SYMBOL(lu_context_key_degister);
1484
1485 /**
1486  * Register a number of keys. This has to be called after all keys have been
1487  * initialized by a call to LU_CONTEXT_KEY_INIT().
1488  */
1489 int lu_context_key_register_many(struct lu_context_key *k, ...)
1490 {
1491         struct lu_context_key *key = k;
1492         va_list args;
1493         int result;
1494
1495         va_start(args, k);
1496         do {
1497                 result = lu_context_key_register(key);
1498                 if (result)
1499                         break;
1500                 key = va_arg(args, struct lu_context_key *);
1501         } while (key != NULL);
1502         va_end(args);
1503
1504         if (result != 0) {
1505                 va_start(args, k);
1506                 while (k != key) {
1507                         lu_context_key_degister(k);
1508                         k = va_arg(args, struct lu_context_key *);
1509                 }
1510                 va_end(args);
1511         }
1512
1513         return result;
1514 }
1515 EXPORT_SYMBOL(lu_context_key_register_many);
1516
1517 /**
1518  * De-register a number of keys. This is a dual to
1519  * lu_context_key_register_many().
1520  */
1521 void lu_context_key_degister_many(struct lu_context_key *k, ...)
1522 {
1523         va_list args;
1524
1525         va_start(args, k);
1526         do {
1527                 lu_context_key_degister(k);
1528                 k = va_arg(args, struct lu_context_key*);
1529         } while (k != NULL);
1530         va_end(args);
1531 }
1532 EXPORT_SYMBOL(lu_context_key_degister_many);
1533
1534 /**
1535  * Revive a number of keys.
1536  */
1537 void lu_context_key_revive_many(struct lu_context_key *k, ...)
1538 {
1539         va_list args;
1540
1541         va_start(args, k);
1542         do {
1543                 lu_context_key_revive(k);
1544                 k = va_arg(args, struct lu_context_key*);
1545         } while (k != NULL);
1546         va_end(args);
1547 }
1548 EXPORT_SYMBOL(lu_context_key_revive_many);
1549
1550 /**
1551  * Quiescent a number of keys.
1552  */
1553 void lu_context_key_quiesce_many(struct lu_context_key *k, ...)
1554 {
1555         va_list args;
1556
1557         va_start(args, k);
1558         do {
1559                 lu_context_key_quiesce(k);
1560                 k = va_arg(args, struct lu_context_key*);
1561         } while (k != NULL);
1562         va_end(args);
1563 }
1564 EXPORT_SYMBOL(lu_context_key_quiesce_many);
1565
1566 /**
1567  * Return value associated with key \a key in context \a ctx.
1568  */
1569 void *lu_context_key_get(const struct lu_context *ctx,
1570                          const struct lu_context_key *key)
1571 {
1572         LINVRNT(ctx->lc_state == LCS_ENTERED);
1573         LINVRNT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
1574         LASSERT(lu_keys[key->lct_index] == key);
1575         return ctx->lc_value[key->lct_index];
1576 }
1577 EXPORT_SYMBOL(lu_context_key_get);
1578
1579 /**
1580  * List of remembered contexts. XXX document me.
1581  */
1582 static LIST_HEAD(lu_context_remembered);
1583
1584 /**
1585  * Destroy \a key in all remembered contexts. This is used to destroy key
1586  * values in "shared" contexts (like service threads), when a module owning
1587  * the key is about to be unloaded.
1588  */
1589 void lu_context_key_quiesce(struct lu_context_key *key)
1590 {
1591         struct lu_context *ctx;
1592
1593         if (!(key->lct_tags & LCT_QUIESCENT)) {
1594                 /*
1595                  * XXX memory barrier has to go here.
1596                  */
1597                 write_lock(&lu_keys_guard);
1598                 key->lct_tags |= LCT_QUIESCENT;
1599
1600                 /**
1601                  * Wait until all lu_context_key::lct_init() methods
1602                  * have completed.
1603                  */
1604                 while (atomic_read(&lu_key_initing_cnt) > 0) {
1605                         write_unlock(&lu_keys_guard);
1606                         CDEBUG(D_INFO, "lu_context_key_quiesce: \"%s\""
1607                                " %p, %d (%d)\n",
1608                                key->lct_owner ? key->lct_owner->name : "",
1609                                key, atomic_read(&key->lct_used),
1610                                atomic_read(&lu_key_initing_cnt));
1611                         schedule();
1612                         write_lock(&lu_keys_guard);
1613                 }
1614
1615                 list_for_each_entry(ctx, &lu_context_remembered,
1616                                     lc_remember)
1617                         key_fini(ctx, key->lct_index);
1618                 write_unlock(&lu_keys_guard);
1619                 ++key_set_version;
1620         }
1621 }
1622
1623 void lu_context_key_revive(struct lu_context_key *key)
1624 {
1625         key->lct_tags &= ~LCT_QUIESCENT;
1626         ++key_set_version;
1627 }
1628
1629 static void keys_fini(struct lu_context *ctx)
1630 {
1631         unsigned int i;
1632
1633         if (ctx->lc_value == NULL)
1634                 return;
1635
1636         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
1637                 key_fini(ctx, i);
1638
1639         OBD_FREE(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1640         ctx->lc_value = NULL;
1641 }
1642
1643 static int keys_fill(struct lu_context *ctx)
1644 {
1645         unsigned int i;
1646
1647         /*
1648          * A serialisation with lu_context_key_quiesce() is needed, but some
1649          * "key->lct_init()" are calling kernel memory allocation routine and
1650          * can't be called while holding a spin_lock.
1651          * "lu_keys_guard" is held while incrementing "lu_key_initing_cnt"
1652          * to ensure the start of the serialisation.
1653          * An atomic_t variable is still used, in order not to reacquire the
1654          * lock when decrementing the counter.
1655          */
1656         read_lock(&lu_keys_guard);
1657         atomic_inc(&lu_key_initing_cnt);
1658         read_unlock(&lu_keys_guard);
1659
1660         LINVRNT(ctx->lc_value != NULL);
1661         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1662                 struct lu_context_key *key;
1663
1664                 key = lu_keys[i];
1665                 if (ctx->lc_value[i] == NULL && key != NULL &&
1666                     (key->lct_tags & ctx->lc_tags) &&
1667                     /*
1668                      * Don't create values for a LCT_QUIESCENT key, as this
1669                      * will pin module owning a key.
1670                      */
1671                     !(key->lct_tags & LCT_QUIESCENT)) {
1672                         void *value;
1673
1674                         LINVRNT(key->lct_init != NULL);
1675                         LINVRNT(key->lct_index == i);
1676
1677                         LASSERT(key->lct_owner != NULL);
1678                         if (!(ctx->lc_tags & LCT_NOREF) &&
1679                             try_module_get(key->lct_owner) == 0) {
1680                                 /* module is unloading, skip this key */
1681                                 continue;
1682                         }
1683
1684                         value = key->lct_init(ctx, key);
1685                         if (unlikely(IS_ERR(value))) {
1686                                 atomic_dec(&lu_key_initing_cnt);
1687                                 return PTR_ERR(value);
1688                         }
1689
1690                         lu_ref_add_atomic(&key->lct_reference, "ctx", ctx);
1691                         atomic_inc(&key->lct_used);
1692                         /*
1693                          * This is the only place in the code, where an
1694                          * element of ctx->lc_value[] array is set to non-NULL
1695                          * value.
1696                          */
1697                         ctx->lc_value[i] = value;
1698                         if (key->lct_exit != NULL)
1699                                 ctx->lc_tags |= LCT_HAS_EXIT;
1700                 }
1701                 ctx->lc_version = key_set_version;
1702         }
1703         atomic_dec(&lu_key_initing_cnt);
1704         return 0;
1705 }
1706
1707 static int keys_init(struct lu_context *ctx)
1708 {
1709         OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
1710         if (likely(ctx->lc_value != NULL))
1711                 return keys_fill(ctx);
1712
1713         return -ENOMEM;
1714 }
1715
1716 /**
1717  * Initialize context data-structure. Create values for all keys.
1718  */
1719 int lu_context_init(struct lu_context *ctx, __u32 tags)
1720 {
1721         int     rc;
1722
1723         memset(ctx, 0, sizeof *ctx);
1724         ctx->lc_state = LCS_INITIALIZED;
1725         ctx->lc_tags = tags;
1726         if (tags & LCT_REMEMBER) {
1727                 write_lock(&lu_keys_guard);
1728                 list_add(&ctx->lc_remember, &lu_context_remembered);
1729                 write_unlock(&lu_keys_guard);
1730         } else {
1731                 INIT_LIST_HEAD(&ctx->lc_remember);
1732         }
1733
1734         rc = keys_init(ctx);
1735         if (rc != 0)
1736                 lu_context_fini(ctx);
1737
1738         return rc;
1739 }
1740 EXPORT_SYMBOL(lu_context_init);
1741
1742 /**
1743  * Finalize context data-structure. Destroy key values.
1744  */
1745 void lu_context_fini(struct lu_context *ctx)
1746 {
1747         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1748         ctx->lc_state = LCS_FINALIZED;
1749
1750         if ((ctx->lc_tags & LCT_REMEMBER) == 0) {
1751                 LASSERT(list_empty(&ctx->lc_remember));
1752                 keys_fini(ctx);
1753
1754         } else { /* could race with key degister */
1755                 write_lock(&lu_keys_guard);
1756                 keys_fini(ctx);
1757                 list_del_init(&ctx->lc_remember);
1758                 write_unlock(&lu_keys_guard);
1759         }
1760 }
1761 EXPORT_SYMBOL(lu_context_fini);
1762
1763 /**
1764  * Called before entering context.
1765  */
1766 void lu_context_enter(struct lu_context *ctx)
1767 {
1768         LINVRNT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1769         ctx->lc_state = LCS_ENTERED;
1770 }
1771 EXPORT_SYMBOL(lu_context_enter);
1772
1773 /**
1774  * Called after exiting from \a ctx
1775  */
1776 void lu_context_exit(struct lu_context *ctx)
1777 {
1778         unsigned int i;
1779
1780         LINVRNT(ctx->lc_state == LCS_ENTERED);
1781         ctx->lc_state = LCS_LEFT;
1782         if (ctx->lc_tags & LCT_HAS_EXIT && ctx->lc_value != NULL) {
1783                 /* could race with key quiescency */
1784                 if (ctx->lc_tags & LCT_REMEMBER)
1785                         read_lock(&lu_keys_guard);
1786
1787                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1788                         if (ctx->lc_value[i] != NULL) {
1789                                 struct lu_context_key *key;
1790
1791                                 key = lu_keys[i];
1792                                 LASSERT(key != NULL);
1793                                 if (key->lct_exit != NULL)
1794                                         key->lct_exit(ctx,
1795                                                       key, ctx->lc_value[i]);
1796                         }
1797                 }
1798
1799                 if (ctx->lc_tags & LCT_REMEMBER)
1800                         read_unlock(&lu_keys_guard);
1801         }
1802 }
1803 EXPORT_SYMBOL(lu_context_exit);
1804
1805 /**
1806  * Allocate for context all missing keys that were registered after context
1807  * creation. key_set_version is only changed in rare cases when modules
1808  * are loaded and removed.
1809  */
1810 int lu_context_refill(struct lu_context *ctx)
1811 {
1812         return likely(ctx->lc_version == key_set_version) ? 0 : keys_fill(ctx);
1813 }
1814
1815 /**
1816  * lu_ctx_tags/lu_ses_tags will be updated if there are new types of
1817  * obd being added. Currently, this is only used on client side, specifically
1818  * for echo device client, for other stack (like ptlrpc threads), context are
1819  * predefined when the lu_device type are registered, during the module probe
1820  * phase.
1821  */
1822 __u32 lu_context_tags_default = 0;
1823 __u32 lu_session_tags_default = 0;
1824
1825 void lu_context_tags_update(__u32 tags)
1826 {
1827         write_lock(&lu_keys_guard);
1828         lu_context_tags_default |= tags;
1829         key_set_version++;
1830         write_unlock(&lu_keys_guard);
1831 }
1832 EXPORT_SYMBOL(lu_context_tags_update);
1833
1834 void lu_context_tags_clear(__u32 tags)
1835 {
1836         write_lock(&lu_keys_guard);
1837         lu_context_tags_default &= ~tags;
1838         key_set_version++;
1839         write_unlock(&lu_keys_guard);
1840 }
1841 EXPORT_SYMBOL(lu_context_tags_clear);
1842
1843 void lu_session_tags_update(__u32 tags)
1844 {
1845         write_lock(&lu_keys_guard);
1846         lu_session_tags_default |= tags;
1847         key_set_version++;
1848         write_unlock(&lu_keys_guard);
1849 }
1850 EXPORT_SYMBOL(lu_session_tags_update);
1851
1852 void lu_session_tags_clear(__u32 tags)
1853 {
1854         write_lock(&lu_keys_guard);
1855         lu_session_tags_default &= ~tags;
1856         key_set_version++;
1857         write_unlock(&lu_keys_guard);
1858 }
1859 EXPORT_SYMBOL(lu_session_tags_clear);
1860
1861 int lu_env_init(struct lu_env *env, __u32 tags)
1862 {
1863         int result;
1864
1865         env->le_ses = NULL;
1866         result = lu_context_init(&env->le_ctx, tags);
1867         if (likely(result == 0))
1868                 lu_context_enter(&env->le_ctx);
1869         return result;
1870 }
1871 EXPORT_SYMBOL(lu_env_init);
1872
1873 void lu_env_fini(struct lu_env *env)
1874 {
1875         lu_context_exit(&env->le_ctx);
1876         lu_context_fini(&env->le_ctx);
1877         env->le_ses = NULL;
1878 }
1879 EXPORT_SYMBOL(lu_env_fini);
1880
1881 int lu_env_refill(struct lu_env *env)
1882 {
1883         int result;
1884
1885         result = lu_context_refill(&env->le_ctx);
1886         if (result == 0 && env->le_ses != NULL)
1887                 result = lu_context_refill(env->le_ses);
1888         return result;
1889 }
1890 EXPORT_SYMBOL(lu_env_refill);
1891
1892 /**
1893  * Currently, this API will only be used by echo client.
1894  * Because echo client and normal lustre client will share
1895  * same cl_env cache. So echo client needs to refresh
1896  * the env context after it get one from the cache, especially
1897  * when normal client and echo client co-exist in the same client.
1898  */
1899 int lu_env_refill_by_tags(struct lu_env *env, __u32 ctags,
1900                           __u32 stags)
1901 {
1902         int    result;
1903
1904         if ((env->le_ctx.lc_tags & ctags) != ctags) {
1905                 env->le_ctx.lc_version = 0;
1906                 env->le_ctx.lc_tags |= ctags;
1907         }
1908
1909         if (env->le_ses && (env->le_ses->lc_tags & stags) != stags) {
1910                 env->le_ses->lc_version = 0;
1911                 env->le_ses->lc_tags |= stags;
1912         }
1913
1914         result = lu_env_refill(env);
1915
1916         return result;
1917 }
1918 EXPORT_SYMBOL(lu_env_refill_by_tags);
1919
1920 static struct shrinker *lu_site_shrinker;
1921
1922 typedef struct lu_site_stats{
1923         unsigned        lss_populated;
1924         unsigned        lss_max_search;
1925         unsigned        lss_total;
1926         unsigned        lss_busy;
1927 } lu_site_stats_t;
1928
1929 static void lu_site_stats_get(struct cfs_hash *hs,
1930                               lu_site_stats_t *stats, int populated)
1931 {
1932         struct cfs_hash_bd bd;
1933         unsigned int  i;
1934
1935         cfs_hash_for_each_bucket(hs, &bd, i) {
1936                 struct lu_site_bkt_data *bkt = cfs_hash_bd_extra_get(hs, &bd);
1937                 struct hlist_head       *hhead;
1938
1939                 cfs_hash_bd_lock(hs, &bd, 1);
1940                 stats->lss_busy  +=
1941                         cfs_hash_bd_count_get(&bd) - bkt->lsb_lru_len;
1942                 stats->lss_total += cfs_hash_bd_count_get(&bd);
1943                 stats->lss_max_search = max((int)stats->lss_max_search,
1944                                             cfs_hash_bd_depmax_get(&bd));
1945                 if (!populated) {
1946                         cfs_hash_bd_unlock(hs, &bd, 1);
1947                         continue;
1948                 }
1949
1950                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1951                         if (!hlist_empty(hhead))
1952                                 stats->lss_populated++;
1953                 }
1954                 cfs_hash_bd_unlock(hs, &bd, 1);
1955         }
1956 }
1957
1958
1959 /*
1960  * lu_cache_shrink_count() returns an approximate number of cached objects
1961  * that can be freed by shrink_slab(). A counter, which tracks the
1962  * number of items in the site's lru, is maintained in a percpu_counter
1963  * for each site. The percpu values are incremented and decremented as
1964  * objects are added or removed from the lru. The percpu values are summed
1965  * and saved whenever a percpu value exceeds a threshold. Thus the saved,
1966  * summed value at any given time may not accurately reflect the current
1967  * lru length. But this value is sufficiently accurate for the needs of
1968  * a shrinker.
1969  *
1970  * Using a per cpu counter is a compromise solution to concurrent access:
1971  * lu_object_put() can update the counter without locking the site and
1972  * lu_cache_shrink_count can sum the counters without locking each
1973  * ls_obj_hash bucket.
1974  */
1975 static unsigned long lu_cache_shrink_count(struct shrinker *sk,
1976                                            struct shrink_control *sc)
1977 {
1978         struct lu_site *s;
1979         struct lu_site *tmp;
1980         unsigned long cached = 0;
1981
1982         if (!(sc->gfp_mask & __GFP_FS))
1983                 return 0;
1984
1985         down_read(&lu_sites_guard);
1986         list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage)
1987                 cached += percpu_counter_read_positive(&s->ls_lru_len_counter);
1988         up_read(&lu_sites_guard);
1989
1990         cached = (cached / 100) * sysctl_vfs_cache_pressure;
1991         CDEBUG(D_INODE, "%ld objects cached, cache pressure %d\n",
1992                cached, sysctl_vfs_cache_pressure);
1993
1994         return cached;
1995 }
1996
1997 static unsigned long lu_cache_shrink_scan(struct shrinker *sk,
1998                                           struct shrink_control *sc)
1999 {
2000         struct lu_site *s;
2001         struct lu_site *tmp;
2002         unsigned long remain = sc->nr_to_scan;
2003         LIST_HEAD(splice);
2004
2005         if (!(sc->gfp_mask & __GFP_FS))
2006                 /* We must not take the lu_sites_guard lock when
2007                  * __GFP_FS is *not* set because of the deadlock
2008                  * possibility detailed above. Additionally,
2009                  * since we cannot determine the number of
2010                  * objects in the cache without taking this
2011                  * lock, we're in a particularly tough spot. As
2012                  * a result, we'll just lie and say our cache is
2013                  * empty. This _should_ be ok, as we can't
2014                  * reclaim objects when __GFP_FS is *not* set
2015                  * anyways.
2016                  */
2017                 return SHRINK_STOP;
2018
2019         down_write(&lu_sites_guard);
2020         list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
2021                 remain = lu_site_purge(&lu_shrink_env, s, remain);
2022                 /*
2023                  * Move just shrunk site to the tail of site list to
2024                  * assure shrinking fairness.
2025                  */
2026                 list_move_tail(&s->ls_linkage, &splice);
2027         }
2028         list_splice(&splice, lu_sites.prev);
2029         up_write(&lu_sites_guard);
2030
2031         return sc->nr_to_scan - remain;
2032 }
2033
2034 #ifndef HAVE_SHRINKER_COUNT
2035 /*
2036  * There exists a potential lock inversion deadlock scenario when using
2037  * Lustre on top of ZFS. This occurs between one of ZFS's
2038  * buf_hash_table.ht_lock's, and Lustre's lu_sites_guard lock. Essentially,
2039  * thread A will take the lu_sites_guard lock and sleep on the ht_lock,
2040  * while thread B will take the ht_lock and sleep on the lu_sites_guard
2041  * lock. Obviously neither thread will wake and drop their respective hold
2042  * on their lock.
2043  *
2044  * To prevent this from happening we must ensure the lu_sites_guard lock is
2045  * not taken while down this code path. ZFS reliably does not set the
2046  * __GFP_FS bit in its code paths, so this can be used to determine if it
2047  * is safe to take the lu_sites_guard lock.
2048  *
2049  * Ideally we should accurately return the remaining number of cached
2050  * objects without taking the lu_sites_guard lock, but this is not
2051  * possible in the current implementation.
2052  */
2053 static int lu_cache_shrink(SHRINKER_ARGS(sc, nr_to_scan, gfp_mask))
2054 {
2055         int cached = 0;
2056         struct shrink_control scv = {
2057                  .nr_to_scan = shrink_param(sc, nr_to_scan),
2058                  .gfp_mask   = shrink_param(sc, gfp_mask)
2059         };
2060 #if !defined(HAVE_SHRINKER_WANT_SHRINK_PTR) && !defined(HAVE_SHRINK_CONTROL)
2061         struct shrinker* shrinker = NULL;
2062 #endif
2063
2064
2065         CDEBUG(D_INODE, "Shrink %lu objects\n", scv.nr_to_scan);
2066
2067         if (scv.nr_to_scan != 0)
2068                 lu_cache_shrink_scan(shrinker, &scv);
2069
2070         cached = lu_cache_shrink_count(shrinker, &scv);
2071         return cached;
2072 }
2073
2074 #endif /* HAVE_SHRINKER_COUNT */
2075
2076
2077 /*
2078  * Debugging stuff.
2079  */
2080
2081 /**
2082  * Environment to be used in debugger, contains all tags.
2083  */
2084 static struct lu_env lu_debugging_env;
2085
2086 /**
2087  * Debugging printer function using printk().
2088  */
2089 int lu_printk_printer(const struct lu_env *env,
2090                       void *unused, const char *format, ...)
2091 {
2092         va_list args;
2093
2094         va_start(args, format);
2095         vprintk(format, args);
2096         va_end(args);
2097         return 0;
2098 }
2099
2100 int lu_debugging_setup(void)
2101 {
2102         return lu_env_init(&lu_debugging_env, ~0);
2103 }
2104
2105 void lu_context_keys_dump(void)
2106 {
2107         unsigned int i;
2108
2109         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
2110                 struct lu_context_key *key;
2111
2112                 key = lu_keys[i];
2113                 if (key != NULL) {
2114                         CERROR("[%d]: %p %x (%p,%p,%p) %d %d \"%s\"@%p\n",
2115                                i, key, key->lct_tags,
2116                                key->lct_init, key->lct_fini, key->lct_exit,
2117                                key->lct_index, atomic_read(&key->lct_used),
2118                                key->lct_owner ? key->lct_owner->name : "",
2119                                key->lct_owner);
2120                         lu_ref_print(&key->lct_reference);
2121                 }
2122         }
2123 }
2124
2125 /**
2126  * Initialization of global lu_* data.
2127  */
2128 int lu_global_init(void)
2129 {
2130         int result;
2131         DEF_SHRINKER_VAR(shvar, lu_cache_shrink,
2132                          lu_cache_shrink_count, lu_cache_shrink_scan);
2133
2134         CDEBUG(D_INFO, "Lustre LU module (%p).\n", &lu_keys);
2135
2136         result = lu_ref_global_init();
2137         if (result != 0)
2138                 return result;
2139
2140         LU_CONTEXT_KEY_INIT(&lu_global_key);
2141         result = lu_context_key_register(&lu_global_key);
2142         if (result != 0)
2143                 return result;
2144
2145         /*
2146          * At this level, we don't know what tags are needed, so allocate them
2147          * conservatively. This should not be too bad, because this
2148          * environment is global.
2149          */
2150         down_write(&lu_sites_guard);
2151         result = lu_env_init(&lu_shrink_env, LCT_SHRINKER);
2152         up_write(&lu_sites_guard);
2153         if (result != 0)
2154                 return result;
2155
2156         /*
2157          * seeks estimation: 3 seeks to read a record from oi, one to read
2158          * inode, one for ea. Unfortunately setting this high value results in
2159          * lu_object/inode cache consuming all the memory.
2160          */
2161         lu_site_shrinker = set_shrinker(DEFAULT_SEEKS, &shvar);
2162         if (lu_site_shrinker == NULL)
2163                 return -ENOMEM;
2164
2165         return result;
2166 }
2167
2168 /**
2169  * Dual to lu_global_init().
2170  */
2171 void lu_global_fini(void)
2172 {
2173         if (lu_site_shrinker != NULL) {
2174                 remove_shrinker(lu_site_shrinker);
2175                 lu_site_shrinker = NULL;
2176         }
2177
2178         lu_context_key_degister(&lu_global_key);
2179
2180         /*
2181          * Tear shrinker environment down _after_ de-registering
2182          * lu_global_key, because the latter has a value in the former.
2183          */
2184         down_write(&lu_sites_guard);
2185         lu_env_fini(&lu_shrink_env);
2186         up_write(&lu_sites_guard);
2187
2188         lu_ref_global_fini();
2189 }
2190
2191 static __u32 ls_stats_read(struct lprocfs_stats *stats, int idx)
2192 {
2193 #ifdef CONFIG_PROC_FS
2194         struct lprocfs_counter ret;
2195
2196         lprocfs_stats_collect(stats, idx, &ret);
2197         return (__u32)ret.lc_count;
2198 #else
2199         return 0;
2200 #endif
2201 }
2202
2203 /**
2204  * Output site statistical counters into a buffer. Suitable for
2205  * lprocfs_rd_*()-style functions.
2206  */
2207 int lu_site_stats_seq_print(const struct lu_site *s, struct seq_file *m)
2208 {
2209         lu_site_stats_t stats;
2210
2211         memset(&stats, 0, sizeof(stats));
2212         lu_site_stats_get(s->ls_obj_hash, &stats, 1);
2213
2214         seq_printf(m, "%d/%d %d/%d %d %d %d %d %d %d %d\n",
2215                    stats.lss_busy,
2216                    stats.lss_total,
2217                    stats.lss_populated,
2218                    CFS_HASH_NHLIST(s->ls_obj_hash),
2219                    stats.lss_max_search,
2220                    ls_stats_read(s->ls_stats, LU_SS_CREATED),
2221                    ls_stats_read(s->ls_stats, LU_SS_CACHE_HIT),
2222                    ls_stats_read(s->ls_stats, LU_SS_CACHE_MISS),
2223                    ls_stats_read(s->ls_stats, LU_SS_CACHE_RACE),
2224                    ls_stats_read(s->ls_stats, LU_SS_CACHE_DEATH_RACE),
2225                    ls_stats_read(s->ls_stats, LU_SS_LRU_PURGED));
2226         return 0;
2227 }
2228 EXPORT_SYMBOL(lu_site_stats_seq_print);
2229
2230 /**
2231  * Helper function to initialize a number of kmem slab caches at once.
2232  */
2233 int lu_kmem_init(struct lu_kmem_descr *caches)
2234 {
2235         int result;
2236         struct lu_kmem_descr *iter = caches;
2237
2238         for (result = 0; iter->ckd_cache != NULL; ++iter) {
2239                 *iter->ckd_cache = kmem_cache_create(iter->ckd_name,
2240                                                      iter->ckd_size,
2241                                                      0, 0, NULL);
2242                 if (*iter->ckd_cache == NULL) {
2243                         result = -ENOMEM;
2244                         /* free all previously allocated caches */
2245                         lu_kmem_fini(caches);
2246                         break;
2247                 }
2248         }
2249         return result;
2250 }
2251 EXPORT_SYMBOL(lu_kmem_init);
2252
2253 /**
2254  * Helper function to finalize a number of kmem slab cached at once. Dual to
2255  * lu_kmem_init().
2256  */
2257 void lu_kmem_fini(struct lu_kmem_descr *caches)
2258 {
2259         for (; caches->ckd_cache != NULL; ++caches) {
2260                 if (*caches->ckd_cache != NULL) {
2261                         kmem_cache_destroy(*caches->ckd_cache);
2262                         *caches->ckd_cache = NULL;
2263                 }
2264         }
2265 }
2266 EXPORT_SYMBOL(lu_kmem_fini);
2267
2268 /**
2269  * Temporary solution to be able to assign fid in ->do_create()
2270  * till we have fully-functional OST fids
2271  */
2272 void lu_object_assign_fid(const struct lu_env *env, struct lu_object *o,
2273                           const struct lu_fid *fid)
2274 {
2275         struct lu_site          *s = o->lo_dev->ld_site;
2276         struct lu_fid           *old = &o->lo_header->loh_fid;
2277         struct cfs_hash         *hs;
2278         struct cfs_hash_bd       bd;
2279
2280         LASSERT(fid_is_zero(old));
2281
2282         /* supposed to be unique */
2283         hs = s->ls_obj_hash;
2284         cfs_hash_bd_get_and_lock(hs, (void *)fid, &bd, 1);
2285 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
2286         {
2287                 __u64                    version = 0;
2288                 wait_queue_t             waiter;
2289                 struct lu_object        *shadow;
2290                 shadow = htable_lookup(s, &bd, fid, &waiter, &version);
2291                 /* supposed to be unique */
2292                 LASSERT(IS_ERR(shadow) && PTR_ERR(shadow) == -ENOENT);
2293         }
2294 #endif
2295         *old = *fid;
2296         cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
2297         cfs_hash_bd_unlock(hs, &bd, 1);
2298 }
2299 EXPORT_SYMBOL(lu_object_assign_fid);
2300
2301 /**
2302  * allocates object with 0 (non-assiged) fid
2303  * XXX: temporary solution to be able to assign fid in ->do_create()
2304  *      till we have fully-functional OST fids
2305  */
2306 struct lu_object *lu_object_anon(const struct lu_env *env,
2307                                  struct lu_device *dev,
2308                                  const struct lu_object_conf *conf)
2309 {
2310         struct lu_fid     fid;
2311         struct lu_object *o;
2312
2313         fid_zero(&fid);
2314         o = lu_object_alloc(env, dev, &fid, conf);
2315
2316         return o;
2317 }
2318 EXPORT_SYMBOL(lu_object_anon);
2319
2320 struct lu_buf LU_BUF_NULL = {
2321         .lb_buf = NULL,
2322         .lb_len = 0
2323 };
2324 EXPORT_SYMBOL(LU_BUF_NULL);
2325
2326 void lu_buf_free(struct lu_buf *buf)
2327 {
2328         LASSERT(buf);
2329         if (buf->lb_buf) {
2330                 LASSERT(buf->lb_len > 0);
2331                 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2332                 buf->lb_buf = NULL;
2333                 buf->lb_len = 0;
2334         }
2335 }
2336 EXPORT_SYMBOL(lu_buf_free);
2337
2338 void lu_buf_alloc(struct lu_buf *buf, size_t size)
2339 {
2340         LASSERT(buf);
2341         LASSERT(buf->lb_buf == NULL);
2342         LASSERT(buf->lb_len == 0);
2343         OBD_ALLOC_LARGE(buf->lb_buf, size);
2344         if (likely(buf->lb_buf))
2345                 buf->lb_len = size;
2346 }
2347 EXPORT_SYMBOL(lu_buf_alloc);
2348
2349 void lu_buf_realloc(struct lu_buf *buf, size_t size)
2350 {
2351         lu_buf_free(buf);
2352         lu_buf_alloc(buf, size);
2353 }
2354 EXPORT_SYMBOL(lu_buf_realloc);
2355
2356 struct lu_buf *lu_buf_check_and_alloc(struct lu_buf *buf, size_t len)
2357 {
2358         if (buf->lb_buf == NULL && buf->lb_len == 0)
2359                 lu_buf_alloc(buf, len);
2360
2361         if ((len > buf->lb_len) && (buf->lb_buf != NULL))
2362                 lu_buf_realloc(buf, len);
2363
2364         return buf;
2365 }
2366 EXPORT_SYMBOL(lu_buf_check_and_alloc);
2367
2368 /**
2369  * Increase the size of the \a buf.
2370  * preserves old data in buffer
2371  * old buffer remains unchanged on error
2372  * \retval 0 or -ENOMEM
2373  */
2374 int lu_buf_check_and_grow(struct lu_buf *buf, size_t len)
2375 {
2376         char *ptr;
2377
2378         if (len <= buf->lb_len)
2379                 return 0;
2380
2381         OBD_ALLOC_LARGE(ptr, len);
2382         if (ptr == NULL)
2383                 return -ENOMEM;
2384
2385         /* Free the old buf */
2386         if (buf->lb_buf != NULL) {
2387                 memcpy(ptr, buf->lb_buf, buf->lb_len);
2388                 OBD_FREE_LARGE(buf->lb_buf, buf->lb_len);
2389         }
2390
2391         buf->lb_buf = ptr;
2392         buf->lb_len = len;
2393         return 0;
2394 }