lustre/obdclass/lu_object.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/obdclass/lu_object.c
  37  *
  38  * Lustre Object.
  39  * These are the only exported functions, they provide some generic
  40  * infrastructure for managing object devices
  41  *
  42  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  43  */
  44
  45 #define DEBUG_SUBSYSTEM S_CLASS
  46 #ifndef EXPORT_SYMTAB
  47 # define EXPORT_SYMTAB
  48 #endif
  49
  50 #include <libcfs/libcfs.h>
  51
  52 #ifdef __KERNEL__
  53 # include <linux/module.h>
  54 #endif
  55
  56 /* hash_long() */
  57 #include <libcfs/libcfs_hash.h>
  58 #include <obd_class.h>
  59 #include <obd_support.h>
  60 #include <lustre_disk.h>
  61 #include <lustre_fid.h>
  62 #include <lu_object.h>
  63 #include <libcfs/list.h>
  64 /* lu_time_global_{init,fini}() */
  65 #include <lu_time.h>
  66
  67 static void lu_object_free(const struct lu_env *env, struct lu_object *o);
  68
  69 /**
  70  * Decrease reference counter on object. If last reference is freed, return
  71  * object to the cache, unless lu_object_is_dying(o) holds. In the latter
  72  * case, free object immediately.
  73  */
  74 void lu_object_put(const struct lu_env *env, struct lu_object *o)
  75 {
  76         struct lu_object_header *top;
  77         struct lu_site          *site;
  78         struct lu_object        *orig;
  79         int                      kill_it;
  80
  81         top = o->lo_header;
  82         site = o->lo_dev->ld_site;
  83         orig = o;
  84         kill_it = 0;
  85         write_lock(&site->ls_guard);
  86         if (atomic_dec_and_test(&top->loh_ref)) {
  87                 /*
  88                  * When last reference is released, iterate over object
  89                  * layers, and notify them that object is no longer busy.
  90                  */
  91                 list_for_each_entry_reverse(o, &top->loh_layers, lo_linkage) {
  92                         if (o->lo_ops->loo_object_release != NULL)
  93                                 o->lo_ops->loo_object_release(env, o);
  94                 }
  95                 -- site->ls_busy;
  96                 if (lu_object_is_dying(top)) {
  97                         /*
  98                          * If object is dying (will not be cached), removed it
  99                          * from hash table and LRU.
 100                          *
 101                          * This is done with hash table and LRU lists
 102                          * locked. As the only way to acquire first reference
 103                          * to previously unreferenced object is through
 104                          * hash-table lookup (lu_object_find()), or LRU
 105                          * scanning (lu_site_purge()), that are done under
 106                          * hash-table and LRU lock, no race with concurrent
 107                          * object lookup is possible and we can safely destroy
 108                          * object below.
 109                          */
 110                         hlist_del_init(&top->loh_hash);
 111                         list_del_init(&top->loh_lru);
 112                         -- site->ls_total;
 113                         kill_it = 1;
 114                 }
 115         }
 116         write_unlock(&site->ls_guard);
 117         if (kill_it)
 118                 /*
 119                  * Object was already removed from hash and lru above, can
 120                  * kill it.
 121                  */
 122                 lu_object_free(env, orig);
 123 }
 124 EXPORT_SYMBOL(lu_object_put);
 125
 126 /**
 127  * Allocate new object.
 128  *
 129  * This follows object creation protocol, described in the comment within
 130  * struct lu_device_operations definition.
 131  */
 132 static struct lu_object *lu_object_alloc(const struct lu_env *env,
 133                                          struct lu_device *dev,
 134                                          const struct lu_fid *f,
 135                                          const struct lu_object_conf *conf)
 136 {
 137         struct lu_object *scan;
 138         struct lu_object *top;
 139         struct list_head *layers;
 140         int clean;
 141         int result;
 142         ENTRY;
 143
 144         /*
 145          * Create top-level object slice. This will also create
 146          * lu_object_header.
 147          */
 148         top = dev->ld_ops->ldo_object_alloc(env, NULL, dev);
 149         if (top == NULL)
 150                 RETURN(ERR_PTR(-ENOMEM));
 151         /*
 152          * This is the only place where object fid is assigned. It's constant
 153          * after this point.
 154          */
 155         LASSERT(fid_is_igif(f) || fid_ver(f) == 0);
 156         top->lo_header->loh_fid  = *f;
 157         layers = &top->lo_header->loh_layers;
 158         do {
 159                 /*
 160                  * Call ->loo_object_init() repeatedly, until no more new
 161                  * object slices are created.
 162                  */
 163                 clean = 1;
 164                 list_for_each_entry(scan, layers, lo_linkage) {
 165                         if (scan->lo_flags & LU_OBJECT_ALLOCATED)
 166                                 continue;
 167                         clean = 0;
 168                         scan->lo_header = top->lo_header;
 169                         result = scan->lo_ops->loo_object_init(env, scan, conf);
 170                         if (result != 0) {
 171                                 lu_object_free(env, top);
 172                                 RETURN(ERR_PTR(result));
 173                         }
 174                         scan->lo_flags |= LU_OBJECT_ALLOCATED;
 175                 }
 176         } while (!clean);
 177
 178         list_for_each_entry_reverse(scan, layers, lo_linkage) {
 179                 if (scan->lo_ops->loo_object_start != NULL) {
 180                         result = scan->lo_ops->loo_object_start(env, scan);
 181                         if (result != 0) {
 182                                 lu_object_free(env, top);
 183                                 RETURN(ERR_PTR(result));
 184                         }
 185                 }
 186         }
 187
 188         dev->ld_site->ls_stats.s_created ++;
 189         RETURN(top);
 190 }
 191
 192 /**
 193  * Free an object.
 194  */
 195 static void lu_object_free(const struct lu_env *env, struct lu_object *o)
 196 {
 197         struct list_head splice;
 198         struct lu_object *scan;
 199         struct lu_site          *site;
 200         struct list_head        *layers;
 201
 202         site   = o->lo_dev->ld_site;
 203         layers = &o->lo_header->loh_layers;
 204         /*
 205          * First call ->loo_object_delete() method to release all resources.
 206          */
 207         list_for_each_entry_reverse(scan, layers, lo_linkage) {
 208                 if (scan->lo_ops->loo_object_delete != NULL)
 209                         scan->lo_ops->loo_object_delete(env, scan);
 210         }
 211
 212         /*
 213          * Then, splice object layers into stand-alone list, and call
 214          * ->loo_object_free() on all layers to free memory. Splice is
 215          * necessary, because lu_object_header is freed together with the
 216          * top-level slice.
 217          */
 218         CFS_INIT_LIST_HEAD(&splice);
 219         list_splice_init(layers, &splice);
 220         while (!list_empty(&splice)) {
 221                 /*
 222                  * Free layers in bottom-to-top order, so that object header
 223                  * lives as long as possible and ->loo_object_free() methods
 224                  * can look at its contents.
 225                  */
 226                 o = container_of0(splice.prev, struct lu_object, lo_linkage);
 227                 list_del_init(&o->lo_linkage);
 228                 LASSERT(o->lo_ops->loo_object_free != NULL);
 229                 o->lo_ops->loo_object_free(env, o);
 230         }
 231         cfs_waitq_broadcast(&site->ls_marche_funebre);
 232 }
 233
 234 /**
 235  * Free \a nr objects from the cold end of the site LRU list.
 236  */
 237 int lu_site_purge(const struct lu_env *env, struct lu_site *s, int nr)
 238 {
 239         struct list_head         dispose;
 240         struct lu_object_header *h;
 241         struct lu_object_header *temp;
 242
 243         CFS_INIT_LIST_HEAD(&dispose);
 244         /*
 245          * Under LRU list lock, scan LRU list and move unreferenced objects to
 246          * the dispose list, removing them from LRU and hash table.
 247          */
 248         write_lock(&s->ls_guard);
 249         list_for_each_entry_safe(h, temp, &s->ls_lru, loh_lru) {
 250                 /*
 251                  * Objects are sorted in lru order, and "busy" objects (ones
 252                  * with h->loh_ref > 0) naturally tend to live near hot end
 253                  * that we scan last. Unfortunately, sites usually have small
 254                  * (less then ten) number of busy yet rarely accessed objects
 255                  * (some global objects, accessed directly through pointers,
 256                  * bypassing hash table). Currently algorithm scans them over
 257                  * and over again. Probably we should move busy objects out of
 258                  * LRU, or we can live with that.
 259                  */
 260                 if (nr-- == 0)
 261                         break;
 262                 if (atomic_read(&h->loh_ref) > 0)
 263                         continue;
 264                 hlist_del_init(&h->loh_hash);
 265                 list_move(&h->loh_lru, &dispose);
 266                 s->ls_total --;
 267         }
 268         write_unlock(&s->ls_guard);
 269         /*
 270          * Free everything on the dispose list. This is safe against races due
 271          * to the reasons described in lu_object_put().
 272          */
 273         while (!list_empty(&dispose)) {
 274                 h = container_of0(dispose.next,
 275                                  struct lu_object_header, loh_lru);
 276                 list_del_init(&h->loh_lru);
 277                 lu_object_free(env, lu_object_top(h));
 278                 s->ls_stats.s_lru_purged ++;
 279         }
 280         return nr;
 281 }
 282 EXPORT_SYMBOL(lu_site_purge);
 283
 284 /*
 285  * Object printing.
 286  *
 287  * Code below has to jump through certain loops to output object description
 288  * into libcfs_debug_msg-based log. The problem is that lu_object_print()
 289  * composes object description from strings that are parts of _lines_ of
 290  * output (i.e., strings that are not terminated by newline). This doesn't fit
 291  * very well into libcfs_debug_msg() interface that assumes that each message
 292  * supplied to it is a self-contained output line.
 293  *
 294  * To work around this, strings are collected in a temporary buffer
 295  * (implemented as a value of lu_cdebug_key key), until terminating newline
 296  * character is detected.
 297  *
 298  */
 299
 300 enum {
 301         /**
 302          * Maximal line size.
 303          *
 304          * XXX overflow is not handled correctly.
 305          */
 306         LU_CDEBUG_LINE = 256
 307 };
 308
 309 struct lu_cdebug_data {
 310         /**
 311          * Temporary buffer.
 312          */
 313         char lck_area[LU_CDEBUG_LINE];
 314 };
 315
 316 /* context key constructor/destructor: lu_global_key_init, lu_global_key_fini */
 317 LU_KEY_INIT_FINI(lu_global, struct lu_cdebug_data);
 318
 319 /**
 320  * Key, holding temporary buffer. This key is registered very early by
 321  * lu_global_init().
 322  */
 323 struct lu_context_key lu_global_key = {
 324         .lct_tags = LCT_MD_THREAD|LCT_DT_THREAD|LCT_CL_THREAD,
 325         .lct_init = lu_global_key_init,
 326         .lct_fini = lu_global_key_fini
 327 };
 328
 329 /**
 330  * Printer function emitting messages through libcfs_debug_msg().
 331  */
 332 int lu_cdebug_printer(const struct lu_env *env,
 333                       void *cookie, const char *format, ...)
 334 {
 335         struct lu_cdebug_print_info *info = cookie;
 336         struct lu_cdebug_data       *key;
 337         int used;
 338         int complete;
 339         va_list args;
 340
 341         va_start(args, format);
 342
 343         key = lu_context_key_get(&env->le_ctx, &lu_global_key);
 344         LASSERT(key != NULL);
 345
 346         used = strlen(key->lck_area);
 347         complete = format[strlen(format) - 1] == '\n';
 348         /*
 349          * Append new chunk to the buffer.
 350          */
 351         vsnprintf(key->lck_area + used,
 352                   ARRAY_SIZE(key->lck_area) - used, format, args);
 353         if (complete) {
 354                 libcfs_debug_msg(NULL, info->lpi_subsys, info->lpi_mask,
 355                                  (char *)info->lpi_file, info->lpi_fn,
 356                                  info->lpi_line, "%s", key->lck_area);
 357                 key->lck_area[0] = 0;
 358         }
 359         va_end(args);
 360         return 0;
 361 }
 362 EXPORT_SYMBOL(lu_cdebug_printer);
 363
 364 /*
 365  * Print object header.
 366  */
 367 static void lu_object_header_print(const struct lu_env *env,
 368                                    void *cookie, lu_printer_t printer,
 369                                    const struct lu_object_header *hdr)
 370 {
 371         (*printer)(env, cookie, "header@%p[%#lx, %d, "DFID"%s%s%s]",
 372                    hdr, hdr->loh_flags, atomic_read(&hdr->loh_ref),
 373                    PFID(&hdr->loh_fid),
 374                    hlist_unhashed(&hdr->loh_hash) ? "" : " hash",
 375                    list_empty(&hdr->loh_lru) ? "" : " lru",
 376                    hdr->loh_attr & LOHA_EXISTS ? " exist":"");
 377 }
 378
 379 /*
 380  * Print human readable representation of the @o to the @printer.
 381  */
 382 void lu_object_print(const struct lu_env *env, void *cookie,
 383                      lu_printer_t printer, const struct lu_object *o)
 384 {
 385         static const char ruler[] = "........................................";
 386         struct lu_object_header *top;
 387         int depth;
 388
 389         top = o->lo_header;
 390         lu_object_header_print(env, cookie, printer, top);
 391         (*printer)(env, cookie, "\n");
 392         list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 393                 depth = o->lo_depth + 4;
 394                 LASSERT(o->lo_ops->loo_object_print != NULL);
 395                 /*
 396                  * print `.' @depth times.
 397                  */
 398                 (*printer)(env, cookie, "%*.*s", depth, depth, ruler);
 399                 o->lo_ops->loo_object_print(env, cookie, printer, o);
 400                 (*printer)(env, cookie, "\n");
 401         }
 402 }
 403 EXPORT_SYMBOL(lu_object_print);
 404
 405 /*
 406  * Check object consistency.
 407  */
 408 int lu_object_invariant(const struct lu_object *o)
 409 {
 410         struct lu_object_header *top;
 411
 412         top = o->lo_header;
 413         list_for_each_entry(o, &top->loh_layers, lo_linkage) {
 414                 if (o->lo_ops->loo_object_invariant != NULL &&
 415                     !o->lo_ops->loo_object_invariant(o))
 416                         return 0;
 417         }
 418         return 1;
 419 }
 420 EXPORT_SYMBOL(lu_object_invariant);
 421
 422 static struct lu_object *htable_lookup(struct lu_site *s,
 423                                        const struct hlist_head *bucket,
 424                                        const struct lu_fid *f)
 425 {
 426         struct lu_object_header *h;
 427         struct hlist_node *scan;
 428
 429         hlist_for_each_entry(h, scan, bucket, loh_hash) {
 430                 s->ls_stats.s_cache_check ++;
 431                 if (likely(lu_fid_eq(&h->loh_fid, f) &&
 432                            !lu_object_is_dying(h))) {
 433                         /* bump reference count... */
 434                         if (atomic_add_return(1, &h->loh_ref) == 1)
 435                                 ++ s->ls_busy;
 436                         /* and move to the head of the LRU */
 437                         /*
 438                          * XXX temporary disable this to measure effects of
 439                          * read-write locking.
 440                          */
 441                         /* list_move_tail(&h->loh_lru, &s->ls_lru); */
 442                         s->ls_stats.s_cache_hit ++;
 443                         return lu_object_top(h);
 444                 }
 445         }
 446         s->ls_stats.s_cache_miss ++;
 447         return NULL;
 448 }
 449
 450 static __u32 fid_hash(const struct lu_fid *f, int bits)
 451 {
 452         /* all objects with same id and different versions will belong to same
 453          * collisions list. */
 454         return hash_long(fid_flatten(f), bits);
 455 }
 456
 457 /*
 458  * Search cache for an object with the fid @f. If such object is found, return
 459  * it. Otherwise, create new object, insert it into cache and return it. In
 460  * any case, additional reference is acquired on the returned object.
 461  */
 462 struct lu_object *lu_object_find(const struct lu_env *env,
 463                                  struct lu_site *s, const struct lu_fid *f)
 464 {
 465         struct lu_object     *o;
 466         struct lu_object     *shadow;
 467         struct hlist_head *bucket;
 468
 469         /*
 470          * This uses standard index maintenance protocol:
 471          *
 472          *     - search index under lock, and return object if found;
 473          *     - otherwise, unlock index, allocate new object;
 474          *     - lock index and search again;
 475          *     - if nothing is found (usual case), insert newly created
 476          *       object into index;
 477          *     - otherwise (race: other thread inserted object), free
 478          *       object just allocated.
 479          *     - unlock index;
 480          *     - return object.
 481          */
 482
 483         bucket = s->ls_hash + fid_hash(f, s->ls_hash_bits);
 484
 485         read_lock(&s->ls_guard);
 486         o = htable_lookup(s, bucket, f);
 487         read_unlock(&s->ls_guard);
 488
 489         if (o != NULL)
 490                 return o;
 491
 492         /*
 493          * Allocate new object. This may result in rather complicated
 494          * operations, including fld queries, inode loading, etc.
 495          */
 496         o = lu_object_alloc(env, s, f);
 497         if (unlikely(IS_ERR(o)))
 498                 return o;
 499
 500         LASSERT(lu_fid_eq(lu_object_fid(o), f));
 501
 502         write_lock(&s->ls_guard);
 503         shadow = htable_lookup(s, bucket, f);
 504         if (likely(shadow == NULL)) {
 505                 hlist_add_head(&o->lo_header->loh_hash, bucket);
 506                 list_add_tail(&o->lo_header->loh_lru, &s->ls_lru);
 507                 ++ s->ls_busy;
 508                 ++ s->ls_total;
 509                 shadow = o;
 510                 o = NULL;
 511         } else
 512                 s->ls_stats.s_cache_race ++;
 513         write_unlock(&s->ls_guard);
 514         if (o != NULL)
 515                 lu_object_free(env, o);
 516         return shadow;
 517 }
 518 EXPORT_SYMBOL(lu_object_find);
 519
 520 /*
 521  * Global list of all sites on this node
 522  */
 523 static CFS_LIST_HEAD(lu_sites);
 524 static DECLARE_MUTEX(lu_sites_guard);
 525
 526 /*
 527  * Global environment used by site shrinker.
 528  */
 529 static struct lu_env lu_shrink_env;
 530
 531 /*
 532  * Print all objects in @s.
 533  */
 534 void lu_site_print(const struct lu_env *env, struct lu_site *s, void *cookie,
 535                    lu_printer_t printer)
 536 {
 537         int i;
 538
 539         for (i = 0; i < s->ls_hash_size; ++i) {
 540                 struct lu_object_header *h;
 541                 struct hlist_node       *scan;
 542
 543                 read_lock(&s->ls_guard);
 544                 hlist_for_each_entry(h, scan, &s->ls_hash[i], loh_hash) {
 545
 546                         if (!list_empty(&h->loh_layers)) {
 547                                 const struct lu_object *obj;
 548
 549                                 obj = lu_object_top(h);
 550                                 lu_object_print(env, cookie, printer, obj);
 551                         } else
 552                                 lu_object_header_print(env, cookie, printer, h);
 553                 }
 554                 read_unlock(&s->ls_guard);
 555         }
 556 }
 557 EXPORT_SYMBOL(lu_site_print);
 558
 559 enum {
 560         LU_CACHE_PERCENT   = 20,
 561 };
 562
 563 /*
 564  * Return desired hash table order.
 565  */
 566 static int lu_htable_order(void)
 567 {
 568         unsigned long cache_size;
 569         int bits;
 570
 571         /*
 572          * Calculate hash table size, assuming that we want reasonable
 573          * performance when 20% of total memory is occupied by cache of
 574          * lu_objects.
 575          *
 576          * Size of lu_object is (arbitrary) taken as 1K (together with inode).
 577          */
 578         cache_size = num_physpages;
 579
 580 #if BITS_PER_LONG == 32
 581         /* limit hashtable size for lowmem systems to low RAM */
 582         if (cache_size > 1 << (30 - CFS_PAGE_SHIFT))
 583                 cache_size = 1 << (30 - CFS_PAGE_SHIFT) * 3 / 4;
 584 #endif
 585
 586         cache_size = cache_size / 100 * LU_CACHE_PERCENT *
 587                 (CFS_PAGE_SIZE / 1024);
 588
 589         for (bits = 1; (1 << bits) < cache_size; ++bits) {
 590                 ;
 591         }
 592         return bits;
 593 }
 594
 595 /*
 596  * Initialize site @s, with @d as the top level device.
 597  */
 598 int lu_site_init(struct lu_site *s, struct lu_device *top)
 599 {
 600         int bits;
 601         int size;
 602         int i;
 603         ENTRY;
 604
 605         memset(s, 0, sizeof *s);
 606         rwlock_init(&s->ls_guard);
 607         CFS_INIT_LIST_HEAD(&s->ls_lru);
 608         CFS_INIT_LIST_HEAD(&s->ls_linkage);
 609         s->ls_top_dev = top;
 610         top->ld_site = s;
 611         lu_device_get(top);
 612
 613         for (bits = lu_htable_order(), size = 1 << bits;
 614              (s->ls_hash =
 615               cfs_alloc_large(size * sizeof s->ls_hash[0])) == NULL;
 616              --bits, size >>= 1) {
 617                 /*
 618                  * Scale hash table down, until allocation succeeds.
 619                  */
 620                 ;
 621         }
 622
 623         s->ls_hash_size = size;
 624         s->ls_hash_bits = bits;
 625         s->ls_hash_mask = size - 1;
 626
 627         for (i = 0; i < size; i++)
 628                 INIT_HLIST_HEAD(&s->ls_hash[i]);
 629
 630         RETURN(0);
 631 }
 632 EXPORT_SYMBOL(lu_site_init);
 633
 634 /*
 635  * Finalize @s and release its resources.
 636  */
 637 void lu_site_fini(struct lu_site *s)
 638 {
 639         LASSERT(list_empty(&s->ls_lru));
 640         LASSERT(s->ls_total == 0);
 641
 642         down(&lu_sites_guard);
 643         list_del_init(&s->ls_linkage);
 644         up(&lu_sites_guard);
 645
 646         if (s->ls_hash != NULL) {
 647                 int i;
 648                 for (i = 0; i < s->ls_hash_size; i++)
 649                         LASSERT(hlist_empty(&s->ls_hash[i]));
 650                 cfs_free_large(s->ls_hash);
 651                 s->ls_hash = NULL;
 652         }
 653         if (s->ls_top_dev != NULL) {
 654                 s->ls_top_dev->ld_site = NULL;
 655                 lu_device_put(s->ls_top_dev);
 656                 s->ls_top_dev = NULL;
 657         }
 658 }
 659 EXPORT_SYMBOL(lu_site_fini);
 660
 661 /*
 662  * Called when initialization of stack for this site is completed.
 663  */
 664 int lu_site_init_finish(struct lu_site *s)
 665 {
 666         int result;
 667         down(&lu_sites_guard);
 668         result = lu_context_refill(&lu_shrink_env.le_ctx);
 669         if (result == 0)
 670                 list_add(&s->ls_linkage, &lu_sites);
 671         up(&lu_sites_guard);
 672         return result;
 673 }
 674 EXPORT_SYMBOL(lu_site_init_finish);
 675
 676 /*
 677  * Acquire additional reference on device @d
 678  */
 679 void lu_device_get(struct lu_device *d)
 680 {
 681         atomic_inc(&d->ld_ref);
 682 }
 683 EXPORT_SYMBOL(lu_device_get);
 684
 685 /*
 686  * Release reference on device @d.
 687  */
 688 void lu_device_put(struct lu_device *d)
 689 {
 690         atomic_dec(&d->ld_ref);
 691 }
 692 EXPORT_SYMBOL(lu_device_put);
 693
 694 /*
 695  * Initialize device @d of type @t.
 696  */
 697 int lu_device_init(struct lu_device *d, struct lu_device_type *t)
 698 {
 699         memset(d, 0, sizeof *d);
 700         atomic_set(&d->ld_ref, 0);
 701         d->ld_type = t;
 702         return 0;
 703 }
 704 EXPORT_SYMBOL(lu_device_init);
 705
 706 /*
 707  * Finalize device @d.
 708  */
 709 void lu_device_fini(struct lu_device *d)
 710 {
 711         if (d->ld_obd != NULL)
 712                 /* finish lprocfs */
 713                 lprocfs_obd_cleanup(d->ld_obd);
 714
 715         LASSERTF(atomic_read(&d->ld_ref) == 0,
 716                  "Refcount is %u\n", atomic_read(&d->ld_ref));
 717 }
 718 EXPORT_SYMBOL(lu_device_fini);
 719
 720 /*
 721  * Initialize object @o that is part of compound object @h and was created by
 722  * device @d.
 723  */
 724 int lu_object_init(struct lu_object *o,
 725                    struct lu_object_header *h, struct lu_device *d)
 726 {
 727         memset(o, 0, sizeof *o);
 728         o->lo_header = h;
 729         o->lo_dev    = d;
 730         lu_device_get(d);
 731         CFS_INIT_LIST_HEAD(&o->lo_linkage);
 732         return 0;
 733 }
 734 EXPORT_SYMBOL(lu_object_init);
 735
 736 /*
 737  * Finalize object and release its resources.
 738  */
 739 void lu_object_fini(struct lu_object *o)
 740 {
 741         LASSERT(list_empty(&o->lo_linkage));
 742
 743         if (o->lo_dev != NULL) {
 744                 lu_device_put(o->lo_dev);
 745                 o->lo_dev = NULL;
 746         }
 747 }
 748 EXPORT_SYMBOL(lu_object_fini);
 749
 750 /*
 751  * Add object @o as first layer of compound object @h
 752  *
 753  * This is typically called by the ->ldo_object_alloc() method of top-level
 754  * device.
 755  */
 756 void lu_object_add_top(struct lu_object_header *h, struct lu_object *o)
 757 {
 758         list_move(&o->lo_linkage, &h->loh_layers);
 759 }
 760 EXPORT_SYMBOL(lu_object_add_top);
 761
 762 /**
 763  * Add object \a o as a layer of compound object, going after \a before.
 764  *
 765  * This is typically called by the ->ldo_object_alloc() method of \a
 766  * before->lo_dev.
 767  */
 768 void lu_object_add(struct lu_object *before, struct lu_object *o)
 769 {
 770         list_move(&o->lo_linkage, &before->lo_linkage);
 771 }
 772 EXPORT_SYMBOL(lu_object_add);
 773
 774 /**
 775  * Initialize compound object.
 776  */
 777 int lu_object_header_init(struct lu_object_header *h)
 778 {
 779         memset(h, 0, sizeof *h);
 780         atomic_set(&h->loh_ref, 1);
 781         INIT_HLIST_NODE(&h->loh_hash);
 782         CFS_INIT_LIST_HEAD(&h->loh_lru);
 783         CFS_INIT_LIST_HEAD(&h->loh_layers);
 784         return 0;
 785 }
 786 EXPORT_SYMBOL(lu_object_header_init);
 787
 788 /*
 789  * Finalize compound object.
 790  */
 791 void lu_object_header_fini(struct lu_object_header *h)
 792 {
 793         LASSERT(list_empty(&h->loh_layers));
 794         LASSERT(list_empty(&h->loh_lru));
 795         LASSERT(hlist_unhashed(&h->loh_hash));
 796 }
 797 EXPORT_SYMBOL(lu_object_header_fini);
 798
 799 /*
 800  * Given a compound object, find its slice, corresponding to the device type
 801  * @dtype.
 802  */
 803 struct lu_object *lu_object_locate(struct lu_object_header *h,
 804                                    struct lu_device_type *dtype)
 805 {
 806         struct lu_object *o;
 807
 808         list_for_each_entry(o, &h->loh_layers, lo_linkage) {
 809                 if (o->lo_dev->ld_type == dtype)
 810                         return o;
 811         }
 812         return NULL;
 813 }
 814 EXPORT_SYMBOL(lu_object_locate);
 815
 816
 817
 818 /*
 819  * Finalize and free devices in the device stack.
 820  *
 821  * Finalize device stack by purging object cache, and calling
 822  * lu_device_type_operations::ldto_device_fini() and
 823  * lu_device_type_operations::ldto_device_free() on all devices in the stack.
 824  */
 825 void lu_stack_fini(const struct lu_env *env, struct lu_device *top)
 826 {
 827         struct lu_site   *site = top->ld_site;
 828         struct lu_device *scan;
 829         struct lu_device *next;
 830
 831         lu_site_purge(env, site, ~0);
 832         for (scan = top; scan != NULL; scan = next) {
 833                 next = scan->ld_type->ldt_ops->ldto_device_fini(env, scan);
 834                 lu_device_put(scan);
 835         }
 836
 837         /* purge again. */
 838         lu_site_purge(env, site, ~0);
 839
 840         if (!list_empty(&site->ls_lru) || site->ls_total != 0) {
 841                 /*
 842                  * Uh-oh, objects still exist.
 843                  */
 844                 static DECLARE_LU_CDEBUG_PRINT_INFO(cookie, D_ERROR);
 845
 846                 lu_site_print(env, site, &cookie, lu_cdebug_printer);
 847         }
 848
 849         for (scan = top; scan != NULL; scan = next) {
 850                 const struct lu_device_type *ldt = scan->ld_type;
 851                 struct obd_type             *type;
 852
 853                 next = ldt->ldt_ops->ldto_device_free(env, scan);
 854                 type = ldt->ldt_obd_type;
 855                 type->typ_refcnt--;
 856                 class_put_type(type);
 857         }
 858 }
 859 EXPORT_SYMBOL(lu_stack_fini);
 860
 861 enum {
 862         /*
 863          * Maximal number of tld slots.
 864          */
 865         LU_CONTEXT_KEY_NR = 16
 866 };
 867
 868 static struct lu_context_key *lu_keys[LU_CONTEXT_KEY_NR] = { NULL, };
 869
 870 static spinlock_t lu_keys_guard = SPIN_LOCK_UNLOCKED;
 871
 872 /*
 873  * Register new key.
 874  */
 875 int lu_context_key_register(struct lu_context_key *key)
 876 {
 877         int result;
 878         int i;
 879
 880         LASSERT(key->lct_init != NULL);
 881         LASSERT(key->lct_fini != NULL);
 882         LASSERT(key->lct_tags != 0);
 883         LASSERT(key->lct_owner != NULL);
 884
 885         result = -ENFILE;
 886         spin_lock(&lu_keys_guard);
 887         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
 888                 if (lu_keys[i] == NULL) {
 889                         key->lct_index = i;
 890                         atomic_set(&key->lct_used, 1);
 891                         lu_keys[i] = key;
 892                         result = 0;
 893                         break;
 894                 }
 895         }
 896         spin_unlock(&lu_keys_guard);
 897         return result;
 898 }
 899 EXPORT_SYMBOL(lu_context_key_register);
 900
 901 static void key_fini(struct lu_context *ctx, int index)
 902 {
 903         if (ctx->lc_value[index] != NULL) {
 904                 struct lu_context_key *key;
 905
 906                 key = lu_keys[index];
 907                 LASSERT(key != NULL);
 908                 LASSERT(key->lct_fini != NULL);
 909                 LASSERT(atomic_read(&key->lct_used) > 1);
 910
 911                 key->lct_fini(ctx, key, ctx->lc_value[index]);
 912                 atomic_dec(&key->lct_used);
 913                 LASSERT(key->lct_owner != NULL);
 914                 if (!(ctx->lc_tags & LCT_NOREF)) {
 915                         LASSERT(module_refcount(key->lct_owner) > 0);
 916                         module_put(key->lct_owner);
 917                 }
 918                 ctx->lc_value[index] = NULL;
 919         }
 920 }
 921
 922 /*
 923  * Deregister key.
 924  */
 925 void lu_context_key_degister(struct lu_context_key *key)
 926 {
 927         LASSERT(atomic_read(&key->lct_used) >= 1);
 928         LASSERT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
 929
 930         key_fini(&lu_shrink_env.le_ctx, key->lct_index);
 931
 932         if (atomic_read(&key->lct_used) > 1)
 933                 CERROR("key has instances.\n");
 934         spin_lock(&lu_keys_guard);
 935         lu_keys[key->lct_index] = NULL;
 936         spin_unlock(&lu_keys_guard);
 937 }
 938 EXPORT_SYMBOL(lu_context_key_degister);
 939
 940 /*
 941  * Return value associated with key @key in context @ctx.
 942  */
 943 void *lu_context_key_get(const struct lu_context *ctx,
 944                          struct lu_context_key *key)
 945 {
 946         LASSERT(ctx->lc_state == LCS_ENTERED);
 947         LASSERT(0 <= key->lct_index && key->lct_index < ARRAY_SIZE(lu_keys));
 948         return ctx->lc_value[key->lct_index];
 949 }
 950 EXPORT_SYMBOL(lu_context_key_get);
 951
 952 static void keys_fini(struct lu_context *ctx)
 953 {
 954         int i;
 955
 956         if (ctx->lc_value != NULL) {
 957                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i)
 958                         key_fini(ctx, i);
 959                 OBD_FREE(ctx->lc_value,
 960                          ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
 961                 ctx->lc_value = NULL;
 962         }
 963 }
 964
 965 static int keys_fill(const struct lu_context *ctx)
 966 {
 967         int i;
 968
 969         for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
 970                 struct lu_context_key *key;
 971
 972                 key = lu_keys[i];
 973                 if (ctx->lc_value[i] == NULL &&
 974                     key != NULL && key->lct_tags & ctx->lc_tags) {
 975                         void *value;
 976
 977                         LASSERT(key->lct_init != NULL);
 978                         LASSERT(key->lct_index == i);
 979
 980                         value = key->lct_init(ctx, key);
 981                         if (unlikely(IS_ERR(value)))
 982                                 return PTR_ERR(value);
 983                         LASSERT(key->lct_owner != NULL);
 984                         if (!(ctx->lc_tags & LCT_NOREF))
 985                                 try_module_get(key->lct_owner);
 986                         atomic_inc(&key->lct_used);
 987                         ctx->lc_value[i] = value;
 988                 }
 989         }
 990         return 0;
 991 }
 992
 993 static int keys_init(struct lu_context *ctx)
 994 {
 995         int result;
 996
 997         OBD_ALLOC(ctx->lc_value, ARRAY_SIZE(lu_keys) * sizeof ctx->lc_value[0]);
 998         if (likely(ctx->lc_value != NULL))
 999                 result = keys_fill(ctx);
1000         else
1001                 result = -ENOMEM;
1002
1003         if (result != 0)
1004                 keys_fini(ctx);
1005         return result;
1006 }
1007
1008 /*
1009  * Initialize context data-structure. Create values for all keys.
1010  */
1011 int lu_context_init(struct lu_context *ctx, __u32 tags)
1012 {
1013         memset(ctx, 0, sizeof *ctx);
1014         ctx->lc_state = LCS_INITIALIZED;
1015         ctx->lc_tags = tags;
1016         return keys_init(ctx);
1017 }
1018 EXPORT_SYMBOL(lu_context_init);
1019
1020 /*
1021  * Finalize context data-structure. Destroy key values.
1022  */
1023 void lu_context_fini(struct lu_context *ctx)
1024 {
1025         LASSERT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1026         ctx->lc_state = LCS_FINALIZED;
1027         keys_fini(ctx);
1028 }
1029 EXPORT_SYMBOL(lu_context_fini);
1030
1031 /*
1032  * Called before entering context.
1033  */
1034 void lu_context_enter(struct lu_context *ctx)
1035 {
1036         LASSERT(ctx->lc_state == LCS_INITIALIZED || ctx->lc_state == LCS_LEFT);
1037         ctx->lc_state = LCS_ENTERED;
1038 }
1039 EXPORT_SYMBOL(lu_context_enter);
1040
1041 /*
1042  * Called after exiting from @ctx
1043  */
1044 void lu_context_exit(struct lu_context *ctx)
1045 {
1046         int i;
1047
1048         LASSERT(ctx->lc_state == LCS_ENTERED);
1049         ctx->lc_state = LCS_LEFT;
1050         if (ctx->lc_value != NULL) {
1051                 for (i = 0; i < ARRAY_SIZE(lu_keys); ++i) {
1052                         if (ctx->lc_value[i] != NULL) {
1053                                 struct lu_context_key *key;
1054
1055                                 key = lu_keys[i];
1056                                 LASSERT(key != NULL);
1057                                 if (key->lct_exit != NULL)
1058                                         key->lct_exit(ctx,
1059                                                       key, ctx->lc_value[i]);
1060                         }
1061                 }
1062         }
1063 }
1064 EXPORT_SYMBOL(lu_context_exit);
1065
1066 /*
1067  * Allocate for context all missing keys that were registered after context
1068  * creation.
1069  */
1070 int lu_context_refill(const struct lu_context *ctx)
1071 {
1072         LASSERT(ctx->lc_value != NULL);
1073         return keys_fill(ctx);
1074 }
1075 EXPORT_SYMBOL(lu_context_refill);
1076
1077 static int lu_env_setup(struct lu_env *env, struct lu_context *ses,
1078                         __u32 tags, int noref)
1079 {
1080         int result;
1081
1082         LASSERT(ergo(!noref, !(tags & LCT_NOREF)));
1083
1084         env->le_ses = ses;
1085         result = lu_context_init(&env->le_ctx, tags);
1086         if (likely(result == 0))
1087                 lu_context_enter(&env->le_ctx);
1088         return result;
1089 }
1090
1091 static int lu_env_init_noref(struct lu_env *env, struct lu_context *ses,
1092                              __u32 tags)
1093 {
1094         return lu_env_setup(env, ses, tags, 1);
1095 }
1096
1097 int lu_env_init(struct lu_env *env, struct lu_context *ses, __u32 tags)
1098 {
1099         return lu_env_setup(env, ses, tags, 0);
1100 }
1101 EXPORT_SYMBOL(lu_env_init);
1102
1103 void lu_env_fini(struct lu_env *env)
1104 {
1105         lu_context_exit(&env->le_ctx);
1106         lu_context_fini(&env->le_ctx);
1107         env->le_ses = NULL;
1108 }
1109 EXPORT_SYMBOL(lu_env_fini);
1110
1111 static int lu_cache_shrink(int nr, unsigned int gfp_mask)
1112 {
1113         struct lu_site *s;
1114         struct lu_site *tmp;
1115         int cached = 0;
1116         int remain = nr;
1117         CFS_LIST_HEAD(splice);
1118
1119         if (nr != 0 && !(gfp_mask & __GFP_FS))
1120                 return -1;
1121
1122         down(&lu_sites_guard);
1123         list_for_each_entry_safe(s, tmp, &lu_sites, ls_linkage) {
1124                 if (nr != 0) {
1125                         remain = lu_site_purge(&lu_shrink_env, s, remain);
1126                         /*
1127                          * Move just shrunk site to the tail of site list to
1128                          * assure shrinking fairness.
1129                          */
1130                         list_move_tail(&s->ls_linkage, &splice);
1131                 }
1132                 read_lock(&s->ls_guard);
1133                 cached += s->ls_total - s->ls_busy;
1134                 read_unlock(&s->ls_guard);
1135                 if (remain <= 0)
1136                         break;
1137         }
1138         list_splice(&splice, lu_sites.prev);
1139         up(&lu_sites_guard);
1140         return cached;
1141 }
1142
1143 static struct shrinker *lu_site_shrinker = NULL;
1144
1145 /*
1146  * Initialization of global lu_* data.
1147  */
1148 int lu_global_init(void)
1149 {
1150         int result;
1151
1152         LU_CONTEXT_KEY_INIT(&lu_global_key);
1153         result = lu_context_key_register(&lu_global_key);
1154         if (result == 0) {
1155                 /*
1156                  * At this level, we don't know what tags are needed, so
1157                  * allocate them conservatively. This should not be too bad,
1158                  * because this environment is global.
1159                  */
1160                 down(&lu_sites_guard);
1161                 result = lu_env_init_noref(&lu_shrink_env, NULL, LCT_SHRINKER);
1162                 up(&lu_sites_guard);
1163                 if (result == 0) {
1164                         /*
1165                          * seeks estimation: 3 seeks to read a record from oi,
1166                          * one to read inode, one for ea. Unfortunately
1167                          * setting this high value results in lu_object/inode
1168                          * cache consuming all the memory.
1169                          */
1170                         lu_site_shrinker = set_shrinker(DEFAULT_SEEKS,
1171                                                         lu_cache_shrink);
1172                         if (result == 0)
1173                                 result = lu_time_global_init();
1174                 }
1175         }
1176         return result;
1177 }
1178
1179 /*
1180  * Dual to lu_global_init().
1181  */
1182 void lu_global_fini(void)
1183 {
1184         lu_time_global_fini();
1185         if (lu_site_shrinker != NULL) {
1186                 remove_shrinker(lu_site_shrinker);
1187                 lu_site_shrinker = NULL;
1188         }
1189
1190         lu_context_key_degister(&lu_global_key);
1191
1192         /*
1193          * Tear shrinker environment down _after_ de-registering
1194          * lu_global_key, because the latter has a value in the former.
1195          */
1196         down(&lu_sites_guard);
1197         lu_env_fini(&lu_shrink_env);
1198         up(&lu_sites_guard);
1199 }
1200
1201 struct lu_buf LU_BUF_NULL = {
1202         .lb_buf = NULL,
1203         .lb_len = 0
1204 };
1205 EXPORT_SYMBOL(LU_BUF_NULL);
1206
1207 /*
1208  * XXX: Functions below logically belong to fid module, but they are used by
1209  * dt_store_open(). Put them here until better place is found.
1210  */
1211
1212 void fid_pack(struct lu_fid_pack *pack, const struct lu_fid *fid,
1213               struct lu_fid *befider)
1214 {
1215         int recsize;
1216         __u64 seq;
1217         __u32 oid;
1218
1219         seq = fid_seq(fid);
1220         oid = fid_oid(fid);
1221
1222         /*
1223          * Two cases: compact 6 bytes representation for a common case, and
1224          * full 17 byte representation for "unusual" fid.
1225          */
1226
1227         /*
1228          * Check that usual case is really usual.
1229          */
1230         CLASSERT(LUSTRE_SEQ_MAX_WIDTH < 0xffffull);
1231
1232         if (fid_is_igif(fid) ||
1233             seq > 0xffffffull || oid > 0xffff || fid_ver(fid) != 0) {
1234                 fid_cpu_to_be(befider, fid);
1235                 recsize = sizeof *befider;
1236         } else {
1237                 unsigned char *small_befider;
1238
1239                 small_befider = (char *)befider;
1240
1241                 small_befider[0] = seq >> 16;
1242                 small_befider[1] = seq >> 8;
1243                 small_befider[2] = seq;
1244
1245                 small_befider[3] = oid >> 8;
1246                 small_befider[4] = oid;
1247
1248                 recsize = 5;
1249         }
1250         memcpy(pack->fp_area, befider, recsize);
1251         pack->fp_len = recsize + 1;
1252 }
1253 EXPORT_SYMBOL(fid_pack);
1254
1255 int fid_unpack(const struct lu_fid_pack *pack, struct lu_fid *fid)
1256 {
1257         int result;
1258
1259         result = 0;
1260         switch (pack->fp_len) {
1261         case sizeof *fid + 1:
1262                 memcpy(fid, pack->fp_area, sizeof *fid);
1263                 fid_be_to_cpu(fid, fid);
1264                 break;
1265         case 6: {
1266                 const unsigned char *area;
1267
1268                 area = pack->fp_area;
1269                 fid->f_seq = (area[0] << 16) | (area[1] << 8) | area[2];
1270                 fid->f_oid = (area[3] << 8) | area[4];
1271                 fid->f_ver = 0;
1272                 break;
1273         }
1274         default:
1275                 CERROR("Unexpected packed fid size: %d\n", pack->fp_len);
1276                 result = -EIO;
1277         }
1278         return result;
1279 }
1280 EXPORT_SYMBOL(fid_unpack);
1281
1282 const char *lu_time_names[LU_TIME_NR] = {
1283         [LU_TIME_FIND_LOOKUP] = "find_lookup",
1284         [LU_TIME_FIND_ALLOC]  = "find_alloc",
1285         [LU_TIME_FIND_INSERT] = "find_insert"
1286 };
1287 EXPORT_SYMBOL(lu_time_names);