lustre/obdclass/local_storage.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,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License version 2 for more details.  A copy is
  14  * included in the COPYING file that accompanied this code.
  15
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2012, 2013, Intel Corporation.
  24  */
  25 /*
  26  * lustre/obdclass/local_storage.c
  27  *
  28  * Local storage for file/objects with fid generation. Works on top of OSD.
  29  *
  30  * Author: Mikhail Pershin <mike.pershin@intel.com>
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_CLASS
  34
  35 #include "local_storage.h"
  36
  37 /* all initialized local storages on this node are linked on this */
  38 static struct list_head ls_list_head = LIST_HEAD_INIT(ls_list_head);
  39 static DEFINE_MUTEX(ls_list_mutex);
  40
  41 static int ls_object_init(const struct lu_env *env, struct lu_object *o,
  42                           const struct lu_object_conf *unused)
  43 {
  44         struct ls_device        *ls;
  45         struct lu_object        *below;
  46         struct lu_device        *under;
  47
  48         ENTRY;
  49
  50         ls = container_of0(o->lo_dev, struct ls_device, ls_top_dev.dd_lu_dev);
  51         under = &ls->ls_osd->dd_lu_dev;
  52         below = under->ld_ops->ldo_object_alloc(env, o->lo_header, under);
  53         if (below == NULL)
  54                 RETURN(-ENOMEM);
  55
  56         lu_object_add(o, below);
  57
  58         RETURN(0);
  59 }
  60
  61 static void ls_object_free(const struct lu_env *env, struct lu_object *o)
  62 {
  63         struct ls_object        *obj = lu2ls_obj(o);
  64         struct lu_object_header *h = o->lo_header;
  65
  66         dt_object_fini(&obj->ls_obj);
  67         lu_object_header_fini(h);
  68         OBD_FREE_PTR(obj);
  69 }
  70
  71 struct lu_object_operations ls_lu_obj_ops = {
  72         .loo_object_init  = ls_object_init,
  73         .loo_object_free  = ls_object_free,
  74 };
  75
  76 struct lu_object *ls_object_alloc(const struct lu_env *env,
  77                                   const struct lu_object_header *_h,
  78                                   struct lu_device *d)
  79 {
  80         struct lu_object_header *h;
  81         struct ls_object        *o;
  82         struct lu_object        *l;
  83
  84         LASSERT(_h == NULL);
  85
  86         OBD_ALLOC_PTR(o);
  87         if (o != NULL) {
  88                 l = &o->ls_obj.do_lu;
  89                 h = &o->ls_header;
  90
  91                 lu_object_header_init(h);
  92                 dt_object_init(&o->ls_obj, h, d);
  93                 lu_object_add_top(h, l);
  94
  95                 l->lo_ops = &ls_lu_obj_ops;
  96
  97                 return l;
  98         } else {
  99                 return NULL;
 100         }
 101 }
 102
 103 static struct lu_device_operations ls_lu_dev_ops = {
 104         .ldo_object_alloc =     ls_object_alloc
 105 };
 106
 107 static struct ls_device *__ls_find_dev(struct dt_device *dev)
 108 {
 109         struct ls_device *ls, *ret = NULL;
 110
 111         list_for_each_entry(ls, &ls_list_head, ls_linkage) {
 112                 if (ls->ls_osd == dev) {
 113                         atomic_inc(&ls->ls_refcount);
 114                         ret = ls;
 115                         break;
 116                 }
 117         }
 118         return ret;
 119 }
 120
 121 struct ls_device *ls_find_dev(struct dt_device *dev)
 122 {
 123         struct ls_device *ls;
 124
 125         mutex_lock(&ls_list_mutex);
 126         ls = __ls_find_dev(dev);
 127         mutex_unlock(&ls_list_mutex);
 128
 129         return ls;
 130 }
 131
 132 static struct lu_device_type_operations ls_device_type_ops = {
 133         .ldto_start = NULL,
 134         .ldto_stop  = NULL,
 135 };
 136
 137 static struct lu_device_type ls_lu_type = {
 138         .ldt_name = "local_storage",
 139         .ldt_ops  = &ls_device_type_ops,
 140 };
 141
 142 struct ls_device *ls_device_get(struct dt_device *dev)
 143 {
 144         struct ls_device *ls;
 145
 146         ENTRY;
 147
 148         mutex_lock(&ls_list_mutex);
 149         ls = __ls_find_dev(dev);
 150         if (ls)
 151                 GOTO(out_ls, ls);
 152
 153         /* not found, then create */
 154         OBD_ALLOC_PTR(ls);
 155         if (ls == NULL)
 156                 GOTO(out_ls, ls = ERR_PTR(-ENOMEM));
 157
 158         atomic_set(&ls->ls_refcount, 1);
 159         INIT_LIST_HEAD(&ls->ls_los_list);
 160         mutex_init(&ls->ls_los_mutex);
 161
 162         ls->ls_osd = dev;
 163
 164         LASSERT(dev->dd_lu_dev.ld_site);
 165         lu_device_init(&ls->ls_top_dev.dd_lu_dev, &ls_lu_type);
 166         ls->ls_top_dev.dd_lu_dev.ld_ops = &ls_lu_dev_ops;
 167         ls->ls_top_dev.dd_lu_dev.ld_site = dev->dd_lu_dev.ld_site;
 168
 169         /* finally add ls to the list */
 170         list_add(&ls->ls_linkage, &ls_list_head);
 171 out_ls:
 172         mutex_unlock(&ls_list_mutex);
 173         RETURN(ls);
 174 }
 175
 176 void ls_device_put(const struct lu_env *env, struct ls_device *ls)
 177 {
 178         LASSERT(env);
 179         if (!atomic_dec_and_test(&ls->ls_refcount))
 180                 return;
 181
 182         mutex_lock(&ls_list_mutex);
 183         if (atomic_read(&ls->ls_refcount) == 0) {
 184                 LASSERT(list_empty(&ls->ls_los_list));
 185                 list_del(&ls->ls_linkage);
 186                 lu_site_purge(env, ls->ls_top_dev.dd_lu_dev.ld_site, ~0);
 187                 lu_device_fini(&ls->ls_top_dev.dd_lu_dev);
 188                 OBD_FREE_PTR(ls);
 189         }
 190         mutex_unlock(&ls_list_mutex);
 191 }
 192
 193 /**
 194  * local file fid generation
 195  */
 196 int local_object_fid_generate(const struct lu_env *env,
 197                               struct local_oid_storage *los,
 198                               struct lu_fid *fid)
 199 {
 200         LASSERT(los->los_dev);
 201         LASSERT(los->los_obj);
 202
 203         /* take next OID */
 204
 205         /* to make it unique after reboot we store
 206          * the latest generated fid atomically with
 207          * object creation see local_object_create() */
 208
 209         mutex_lock(&los->los_id_lock);
 210         fid->f_seq = los->los_seq;
 211         fid->f_oid = ++los->los_last_oid;
 212         fid->f_ver = 0;
 213         mutex_unlock(&los->los_id_lock);
 214
 215         return 0;
 216 }
 217
 218 int local_object_declare_create(const struct lu_env *env,
 219                                 struct local_oid_storage *los,
 220                                 struct dt_object *o, struct lu_attr *attr,
 221                                 struct dt_object_format *dof,
 222                                 struct thandle *th)
 223 {
 224         struct dt_thread_info   *dti = dt_info(env);
 225         int                      rc;
 226
 227         ENTRY;
 228
 229         /* update fid generation file */
 230         if (los != NULL) {
 231                 LASSERT(dt_object_exists(los->los_obj));
 232                 dti->dti_lb.lb_buf = NULL;
 233                 dti->dti_lb.lb_len = sizeof(struct los_ondisk);
 234                 rc = dt_declare_record_write(env, los->los_obj,
 235                                              &dti->dti_lb, 0, th);
 236                 if (rc)
 237                         RETURN(rc);
 238         }
 239
 240         rc = dt_declare_create(env, o, attr, NULL, dof, th);
 241         if (rc)
 242                 RETURN(rc);
 243
 244         dti->dti_lb.lb_buf = NULL;
 245         dti->dti_lb.lb_len = sizeof(dti->dti_lma);
 246         rc = dt_declare_xattr_set(env, o, &dti->dti_lb, XATTR_NAME_LMA, 0, th);
 247
 248         RETURN(rc);
 249 }
 250
 251 int local_object_create(const struct lu_env *env,
 252                         struct local_oid_storage *los,
 253                         struct dt_object *o, struct lu_attr *attr,
 254                         struct dt_object_format *dof, struct thandle *th)
 255 {
 256         struct dt_thread_info   *dti = dt_info(env);
 257         obd_id                   lastid;
 258         int                      rc;
 259
 260         ENTRY;
 261
 262         rc = dt_create(env, o, attr, NULL, dof, th);
 263         if (rc)
 264                 RETURN(rc);
 265
 266         if (los == NULL)
 267                 RETURN(rc);
 268
 269         LASSERT(los->los_obj);
 270         LASSERT(dt_object_exists(los->los_obj));
 271
 272         /* many threads can be updated this, serialize
 273          * them here to avoid the race where one thread
 274          * takes the value first, but writes it last */
 275         mutex_lock(&los->los_id_lock);
 276
 277         /* update local oid number on disk so that
 278          * we know the last one used after reboot */
 279         lastid = cpu_to_le64(los->los_last_oid);
 280
 281         dti->dti_off = 0;
 282         dti->dti_lb.lb_buf = &lastid;
 283         dti->dti_lb.lb_len = sizeof(lastid);
 284         rc = dt_record_write(env, los->los_obj, &dti->dti_lb, &dti->dti_off,
 285                              th);
 286         mutex_unlock(&los->los_id_lock);
 287
 288         RETURN(rc);
 289 }
 290
 291 /*
 292  * Create local named object (file, directory or index) in parent directory.
 293  */
 294 struct dt_object *__local_file_create(const struct lu_env *env,
 295                                       const struct lu_fid *fid,
 296                                       struct local_oid_storage *los,
 297                                       struct ls_device *ls,
 298                                       struct dt_object *parent,
 299                                       const char *name, struct lu_attr *attr,
 300                                       struct dt_object_format *dof)
 301 {
 302         struct dt_thread_info   *dti    = dt_info(env);
 303         struct lu_object_conf   *conf   = &dti->dti_conf;
 304         struct dt_insert_rec    *rec    = &dti->dti_dt_rec;
 305         struct dt_object        *dto;
 306         struct thandle          *th;
 307         int                      rc;
 308
 309         /* We know that the target object does not exist, to be created,
 310          * then give some hints - LOC_F_NEW to help low layer to handle
 311          * that efficiently and properly. */
 312         memset(conf, 0, sizeof(*conf));
 313         conf->loc_flags = LOC_F_NEW;
 314         dto = ls_locate(env, ls, fid, conf);
 315         if (unlikely(IS_ERR(dto)))
 316                 RETURN(dto);
 317
 318         LASSERT(dto != NULL);
 319         if (dt_object_exists(dto))
 320                 GOTO(out, rc = -EEXIST);
 321
 322         th = dt_trans_create(env, ls->ls_osd);
 323         if (IS_ERR(th))
 324                 GOTO(out, rc = PTR_ERR(th));
 325
 326         rc = local_object_declare_create(env, los, dto, attr, dof, th);
 327         if (rc)
 328                 GOTO(trans_stop, rc);
 329
 330         if (dti->dti_dof.dof_type == DFT_DIR) {
 331                 rc = dt_declare_ref_add(env, dto, th);
 332                 if (rc < 0)
 333                         GOTO(trans_stop, rc);
 334
 335                 rc = dt_declare_ref_add(env, parent, th);
 336                 if (rc < 0)
 337                         GOTO(trans_stop, rc);
 338         }
 339
 340         rec->rec_fid = fid;
 341         rec->rec_type = dto->do_lu.lo_header->loh_attr;
 342         rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
 343                                (const struct dt_key *)name, th);
 344         if (rc)
 345                 GOTO(trans_stop, rc);
 346
 347         rc = dt_trans_start_local(env, ls->ls_osd, th);
 348         if (rc)
 349                 GOTO(trans_stop, rc);
 350
 351         dt_write_lock(env, dto, 0);
 352         if (dt_object_exists(dto))
 353                 GOTO(unlock, rc = 0);
 354
 355         CDEBUG(D_OTHER, "create new object "DFID"\n",
 356                PFID(lu_object_fid(&dto->do_lu)));
 357         rc = local_object_create(env, los, dto, attr, dof, th);
 358         if (rc)
 359                 GOTO(unlock, rc);
 360         LASSERT(dt_object_exists(dto));
 361
 362         if (dti->dti_dof.dof_type == DFT_DIR) {
 363                 if (!dt_try_as_dir(env, dto))
 364                         GOTO(destroy, rc = -ENOTDIR);
 365
 366                 rec->rec_type = S_IFDIR;
 367                 rec->rec_fid = fid;
 368                 /* Add "." and ".." for newly created dir */
 369                 rc = dt_insert(env, dto, (const struct dt_rec *)rec,
 370                                (const struct dt_key *)".", th, BYPASS_CAPA, 1);
 371                 if (rc != 0)
 372                         GOTO(destroy, rc);
 373
 374                 dt_ref_add(env, dto, th);
 375                 rec->rec_fid = lu_object_fid(&parent->do_lu);
 376                 rc = dt_insert(env, dto, (const struct dt_rec *)rec,
 377                                (const struct dt_key *)"..", th, BYPASS_CAPA, 1);
 378                 if (rc != 0)
 379                         GOTO(destroy, rc);
 380         }
 381
 382         rec->rec_fid = fid;
 383         rec->rec_type = dto->do_lu.lo_header->loh_attr;
 384         dt_write_lock(env, parent, 0);
 385         rc = dt_insert(env, parent, (const struct dt_rec *)rec,
 386                        (const struct dt_key *)name, th, BYPASS_CAPA, 1);
 387         if (dti->dti_dof.dof_type == DFT_DIR)
 388                 dt_ref_add(env, parent, th);
 389         dt_write_unlock(env, parent);
 390         if (rc)
 391                 GOTO(destroy, rc);
 392 destroy:
 393         if (rc)
 394                 dt_destroy(env, dto, th);
 395 unlock:
 396         dt_write_unlock(env, dto);
 397 trans_stop:
 398         dt_trans_stop(env, ls->ls_osd, th);
 399 out:
 400         if (rc) {
 401                 lu_object_put_nocache(env, &dto->do_lu);
 402                 dto = ERR_PTR(rc);
 403         }
 404         RETURN(dto);
 405 }
 406
 407 /*
 408  * Look up and create (if it does not exist) a local named file or directory in
 409  * parent directory.
 410  */
 411 struct dt_object *local_file_find_or_create(const struct lu_env *env,
 412                                             struct local_oid_storage *los,
 413                                             struct dt_object *parent,
 414                                             const char *name, __u32 mode)
 415 {
 416         struct dt_thread_info   *dti = dt_info(env);
 417         struct dt_object        *dto;
 418         int                      rc;
 419
 420         LASSERT(parent);
 421
 422         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 423         if (rc == 0)
 424                 /* name is found, get the object */
 425                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 426                                 &dti->dti_fid, NULL);
 427         else if (rc != -ENOENT)
 428                 dto = ERR_PTR(rc);
 429         else {
 430                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 431                 if (rc < 0) {
 432                         dto = ERR_PTR(rc);
 433                 } else {
 434                         /* create the object */
 435                         dti->dti_attr.la_valid  = LA_MODE;
 436                         dti->dti_attr.la_mode   = mode;
 437                         dti->dti_dof.dof_type   = dt_mode_to_dft(mode & S_IFMT);
 438                         dto = __local_file_create(env, &dti->dti_fid, los,
 439                                                   dt2ls_dev(los->los_dev),
 440                                                   parent, name, &dti->dti_attr,
 441                                                   &dti->dti_dof);
 442                 }
 443         }
 444         return dto;
 445 }
 446 EXPORT_SYMBOL(local_file_find_or_create);
 447
 448 struct dt_object *local_file_find_or_create_with_fid(const struct lu_env *env,
 449                                                      struct dt_device *dt,
 450                                                      const struct lu_fid *fid,
 451                                                      struct dt_object *parent,
 452                                                      const char *name,
 453                                                      __u32 mode)
 454 {
 455         struct dt_thread_info   *dti = dt_info(env);
 456         struct dt_object        *dto;
 457         int                      rc;
 458
 459         LASSERT(parent);
 460
 461         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 462         if (rc == 0) {
 463                 dto = dt_locate(env, dt, &dti->dti_fid);
 464         } else if (rc != -ENOENT) {
 465                 dto = ERR_PTR(rc);
 466         } else {
 467                 struct ls_device *ls;
 468
 469                 ls = ls_device_get(dt);
 470                 if (IS_ERR(ls)) {
 471                         dto = ERR_PTR(PTR_ERR(ls));
 472                 } else {
 473                         /* create the object */
 474                         dti->dti_attr.la_valid  = LA_MODE;
 475                         dti->dti_attr.la_mode   = mode;
 476                         dti->dti_dof.dof_type   = dt_mode_to_dft(mode & S_IFMT);
 477                         dto = __local_file_create(env, fid, NULL, ls, parent,
 478                                                   name, &dti->dti_attr,
 479                                                   &dti->dti_dof);
 480                         /* ls_device_put() will finalize the ls device, we
 481                          * have to open the object in other device stack */
 482                         if (!IS_ERR(dto)) {
 483                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 484                                 lu_object_put_nocache(env, &dto->do_lu);
 485                                 dto = dt_locate(env, dt, &dti->dti_fid);
 486                         }
 487                         ls_device_put(env, ls);
 488                 }
 489         }
 490         return dto;
 491 }
 492 EXPORT_SYMBOL(local_file_find_or_create_with_fid);
 493
 494 /*
 495  * Look up and create (if it does not exist) a local named index file in parent
 496  * directory.
 497  */
 498 struct dt_object *local_index_find_or_create(const struct lu_env *env,
 499                                              struct local_oid_storage *los,
 500                                              struct dt_object *parent,
 501                                              const char *name, __u32 mode,
 502                                              const struct dt_index_features *ft)
 503 {
 504         struct dt_thread_info   *dti = dt_info(env);
 505         struct dt_object        *dto;
 506         int                      rc;
 507
 508         LASSERT(parent);
 509
 510         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 511         if (rc == 0) {
 512                 /* name is found, get the object */
 513                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 514                                 &dti->dti_fid, NULL);
 515         } else if (rc != -ENOENT) {
 516                 dto = ERR_PTR(rc);
 517         } else {
 518                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 519                 if (rc < 0) {
 520                         dto = ERR_PTR(rc);
 521                 } else {
 522                         /* create the object */
 523                         dti->dti_attr.la_valid          = LA_MODE;
 524                         dti->dti_attr.la_mode           = mode;
 525                         dti->dti_dof.dof_type           = DFT_INDEX;
 526                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 527                         dto = __local_file_create(env, &dti->dti_fid, los,
 528                                                   dt2ls_dev(los->los_dev),
 529                                                   parent, name, &dti->dti_attr,
 530                                                   &dti->dti_dof);
 531                 }
 532         }
 533         return dto;
 534
 535 }
 536 EXPORT_SYMBOL(local_index_find_or_create);
 537
 538 struct dt_object *
 539 local_index_find_or_create_with_fid(const struct lu_env *env,
 540                                     struct dt_device *dt,
 541                                     const struct lu_fid *fid,
 542                                     struct dt_object *parent,
 543                                     const char *name, __u32 mode,
 544                                     const struct dt_index_features *ft)
 545 {
 546         struct dt_thread_info   *dti = dt_info(env);
 547         struct dt_object        *dto;
 548         int                      rc;
 549
 550         LASSERT(parent);
 551
 552         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 553         if (rc == 0) {
 554                 /* name is found, get the object */
 555                 if (!lu_fid_eq(fid, &dti->dti_fid))
 556                         dto = ERR_PTR(-EINVAL);
 557                 else
 558                         dto = dt_locate(env, dt, fid);
 559         } else if (rc != -ENOENT) {
 560                 dto = ERR_PTR(rc);
 561         } else {
 562                 struct ls_device *ls;
 563
 564                 ls = ls_device_get(dt);
 565                 if (IS_ERR(ls)) {
 566                         dto = ERR_PTR(PTR_ERR(ls));
 567                 } else {
 568                         /* create the object */
 569                         dti->dti_attr.la_valid          = LA_MODE;
 570                         dti->dti_attr.la_mode           = mode;
 571                         dti->dti_dof.dof_type           = DFT_INDEX;
 572                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 573                         dto = __local_file_create(env, fid, NULL, ls, parent,
 574                                                   name, &dti->dti_attr,
 575                                                   &dti->dti_dof);
 576                         /* ls_device_put() will finalize the ls device, we
 577                          * have to open the object in other device stack */
 578                         if (!IS_ERR(dto)) {
 579                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 580                                 lu_object_put_nocache(env, &dto->do_lu);
 581                                 dto = dt_locate(env, dt, &dti->dti_fid);
 582                         }
 583                         ls_device_put(env, ls);
 584                 }
 585         }
 586         return dto;
 587 }
 588 EXPORT_SYMBOL(local_index_find_or_create_with_fid);
 589
 590 static int local_object_declare_unlink(const struct lu_env *env,
 591                                        struct dt_device *dt,
 592                                        struct dt_object *p,
 593                                        struct dt_object *c, const char *name,
 594                                        struct thandle *th)
 595 {
 596         int rc;
 597
 598         rc = dt_declare_delete(env, p, (const struct dt_key *)name, th);
 599         if (rc < 0)
 600                 return rc;
 601
 602         rc = dt_declare_ref_del(env, c, th);
 603         if (rc < 0)
 604                 return rc;
 605
 606         return dt_declare_destroy(env, c, th);
 607 }
 608
 609 int local_object_unlink(const struct lu_env *env, struct dt_device *dt,
 610                         struct dt_object *parent, const char *name)
 611 {
 612         struct dt_thread_info   *dti = dt_info(env);
 613         struct dt_object        *dto;
 614         struct thandle          *th;
 615         int                      rc;
 616
 617         ENTRY;
 618
 619         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 620         if (rc == -ENOENT)
 621                 RETURN(0);
 622         else if (rc < 0)
 623                 RETURN(rc);
 624
 625         dto = dt_locate(env, dt, &dti->dti_fid);
 626         if (unlikely(IS_ERR(dto)))
 627                 RETURN(PTR_ERR(dto));
 628
 629         th = dt_trans_create(env, dt);
 630         if (IS_ERR(th))
 631                 GOTO(out, rc = PTR_ERR(th));
 632
 633         rc = local_object_declare_unlink(env, dt, parent, dto, name, th);
 634         if (rc < 0)
 635                 GOTO(stop, rc);
 636
 637         rc = dt_trans_start_local(env, dt, th);
 638         if (rc < 0)
 639                 GOTO(stop, rc);
 640
 641         dt_write_lock(env, dto, 0);
 642         rc = dt_delete(env, parent, (struct dt_key *)name, th, BYPASS_CAPA);
 643         if (rc < 0)
 644                 GOTO(unlock, rc);
 645
 646         rc = dt_ref_del(env, dto, th);
 647         if (rc < 0) {
 648                 struct dt_insert_rec *rec = &dti->dti_dt_rec;
 649
 650                 rec->rec_fid = &dti->dti_fid;
 651                 rec->rec_type = dto->do_lu.lo_header->loh_attr;
 652                 rc = dt_insert(env, parent, (const struct dt_rec *)rec,
 653                                (const struct dt_key *)name, th, BYPASS_CAPA, 1);
 654                 GOTO(unlock, rc);
 655         }
 656
 657         rc = dt_destroy(env, dto, th);
 658 unlock:
 659         dt_write_unlock(env, dto);
 660 stop:
 661         dt_trans_stop(env, dt, th);
 662 out:
 663         lu_object_put_nocache(env, &dto->do_lu);
 664         return rc;
 665 }
 666 EXPORT_SYMBOL(local_object_unlink);
 667
 668 struct local_oid_storage *dt_los_find(struct ls_device *ls, __u64 seq)
 669 {
 670         struct local_oid_storage *los, *ret = NULL;
 671
 672         list_for_each_entry(los, &ls->ls_los_list, los_list) {
 673                 if (los->los_seq == seq) {
 674                         atomic_inc(&los->los_refcount);
 675                         ret = los;
 676                         break;
 677                 }
 678         }
 679         return ret;
 680 }
 681
 682 void dt_los_put(struct local_oid_storage *los)
 683 {
 684         if (atomic_dec_and_test(&los->los_refcount))
 685                 /* should never happen, only local_oid_storage_fini should
 686                  * drop refcount to zero */
 687                 LBUG();
 688         return;
 689 }
 690
 691 /* after Lustre 2.3 release there may be old file to store last generated FID
 692  * If such file exists then we have to read its content
 693  */
 694 int lastid_compat_check(const struct lu_env *env, struct dt_device *dev,
 695                         __u64 lastid_seq, __u32 *first_oid, struct ls_device *ls)
 696 {
 697         struct dt_thread_info   *dti = dt_info(env);
 698         struct dt_object        *root = NULL;
 699         struct los_ondisk        losd;
 700         struct dt_object        *o = NULL;
 701         int                      rc = 0;
 702
 703         rc = dt_root_get(env, dev, &dti->dti_fid);
 704         if (rc)
 705                 return rc;
 706
 707         root = ls_locate(env, ls, &dti->dti_fid, NULL);
 708         if (IS_ERR(root))
 709                 return PTR_ERR(root);
 710
 711         /* find old last_id file */
 712         snprintf(dti->dti_buf, sizeof(dti->dti_buf), "seq-"LPX64"-lastid",
 713                  lastid_seq);
 714         rc = dt_lookup_dir(env, root, dti->dti_buf, &dti->dti_fid);
 715         lu_object_put_nocache(env, &root->do_lu);
 716         if (rc == -ENOENT) {
 717                 /* old llog lastid accessed by FID only */
 718                 if (lastid_seq != FID_SEQ_LLOG)
 719                         return 0;
 720                 dti->dti_fid.f_seq = FID_SEQ_LLOG;
 721                 dti->dti_fid.f_oid = 1;
 722                 dti->dti_fid.f_ver = 0;
 723                 o = ls_locate(env, ls, &dti->dti_fid, NULL);
 724                 if (IS_ERR(o))
 725                         return PTR_ERR(o);
 726
 727                 if (!dt_object_exists(o)) {
 728                         lu_object_put_nocache(env, &o->do_lu);
 729                         return 0;
 730                 }
 731                 CDEBUG(D_INFO, "Found old llog lastid file\n");
 732         } else if (rc < 0) {
 733                 return rc;
 734         } else {
 735                 CDEBUG(D_INFO, "Found old lastid file for sequence "LPX64"\n",
 736                        lastid_seq);
 737                 o = ls_locate(env, ls, &dti->dti_fid, NULL);
 738                 if (IS_ERR(o))
 739                         return PTR_ERR(o);
 740         }
 741         /* let's read seq-NNNNNN-lastid file value */
 742         LASSERT(dt_object_exists(o));
 743         dti->dti_off = 0;
 744         dti->dti_lb.lb_buf = &losd;
 745         dti->dti_lb.lb_len = sizeof(losd);
 746         dt_read_lock(env, o, 0);
 747         rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 748         dt_read_unlock(env, o);
 749         if (rc == 0 && le32_to_cpu(losd.lso_magic) != LOS_MAGIC) {
 750                 CERROR("%s: wrong content of seq-"LPX64"-lastid file, magic %x\n",
 751                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq,
 752                        le32_to_cpu(losd.lso_magic));
 753                 rc = -EINVAL;
 754         } else if (rc < 0) {
 755                 CERROR("%s: failed to read seq-"LPX64"-lastid: rc = %d\n",
 756                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq, rc);
 757         }
 758         lu_object_put_nocache(env, &o->do_lu);
 759         if (rc == 0)
 760                 *first_oid = le32_to_cpu(losd.lso_next_oid);
 761         return rc;
 762 }
 763
 764 /**
 765  * Initialize local OID storage for required sequence.
 766  * That may be needed for services that uses local files and requires
 767  * dynamic OID allocation for them.
 768  *
 769  * Per each sequence we have an object with 'first_fid' identificator
 770  * containing the counter for OIDs of locally created files with that
 771  * sequence.
 772  *
 773  * It is used now by llog subsystem and MGS for NID tables
 774  *
 775  * Function gets first_fid to create counter object.
 776  * All dynamic fids will be generated with the same sequence and incremented
 777  * OIDs
 778  *
 779  * Returned local_oid_storage is in-memory representaion of OID storage
 780  */
 781 int local_oid_storage_init(const struct lu_env *env, struct dt_device *dev,
 782                            const struct lu_fid *first_fid,
 783                            struct local_oid_storage **los)
 784 {
 785         struct dt_thread_info   *dti = dt_info(env);
 786         struct ls_device        *ls;
 787         obd_id                   lastid;
 788         struct dt_object        *o = NULL;
 789         struct thandle          *th;
 790         __u32                    first_oid = fid_oid(first_fid);
 791         int                      rc = 0;
 792
 793         ENTRY;
 794
 795         ls = ls_device_get(dev);
 796         if (IS_ERR(ls))
 797                 RETURN(PTR_ERR(ls));
 798
 799         mutex_lock(&ls->ls_los_mutex);
 800         *los = dt_los_find(ls, fid_seq(first_fid));
 801         if (*los != NULL)
 802                 GOTO(out, rc = 0);
 803
 804         /* not found, then create */
 805         OBD_ALLOC_PTR(*los);
 806         if (*los == NULL)
 807                 GOTO(out, rc = -ENOMEM);
 808
 809         atomic_set(&(*los)->los_refcount, 1);
 810         mutex_init(&(*los)->los_id_lock);
 811         (*los)->los_dev = &ls->ls_top_dev;
 812         atomic_inc(&ls->ls_refcount);
 813         list_add(&(*los)->los_list, &ls->ls_los_list);
 814
 815         /* Use {seq, 0, 0} to create the LAST_ID file for every
 816          * sequence.  OIDs start at LUSTRE_FID_INIT_OID.
 817          */
 818         dti->dti_fid.f_seq = fid_seq(first_fid);
 819         dti->dti_fid.f_oid = LUSTRE_FID_LASTID_OID;
 820         dti->dti_fid.f_ver = 0;
 821         o = ls_locate(env, ls, &dti->dti_fid, NULL);
 822         if (IS_ERR(o))
 823                 GOTO(out_los, rc = PTR_ERR(o));
 824
 825         if (!dt_object_exists(o)) {
 826                 rc = lastid_compat_check(env, dev, fid_seq(first_fid),
 827                                          &first_oid, ls);
 828                 if (rc < 0)
 829                         GOTO(out_los, rc);
 830
 831                 th = dt_trans_create(env, dev);
 832                 if (IS_ERR(th))
 833                         GOTO(out_los, rc = PTR_ERR(th));
 834
 835                 dti->dti_attr.la_valid = LA_MODE | LA_TYPE;
 836                 dti->dti_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
 837                 dti->dti_dof.dof_type = dt_mode_to_dft(S_IFREG);
 838
 839                 rc = dt_declare_create(env, o, &dti->dti_attr, NULL,
 840                                        &dti->dti_dof, th);
 841                 if (rc)
 842                         GOTO(out_trans, rc);
 843
 844                 lastid = cpu_to_le64(first_oid);
 845
 846                 dti->dti_off = 0;
 847                 dti->dti_lb.lb_buf = &lastid;
 848                 dti->dti_lb.lb_len = sizeof(lastid);
 849                 rc = dt_declare_record_write(env, o, &dti->dti_lb, dti->dti_off,
 850                                              th);
 851                 if (rc)
 852                         GOTO(out_trans, rc);
 853
 854                 rc = dt_trans_start_local(env, dev, th);
 855                 if (rc)
 856                         GOTO(out_trans, rc);
 857
 858                 dt_write_lock(env, o, 0);
 859                 if (dt_object_exists(o))
 860                         GOTO(out_lock, rc = 0);
 861
 862                 rc = dt_create(env, o, &dti->dti_attr, NULL, &dti->dti_dof,
 863                                th);
 864                 if (rc)
 865                         GOTO(out_lock, rc);
 866
 867                 rc = dt_record_write(env, o, &dti->dti_lb, &dti->dti_off, th);
 868                 if (rc)
 869                         GOTO(out_lock, rc);
 870 out_lock:
 871                 dt_write_unlock(env, o);
 872 out_trans:
 873                 dt_trans_stop(env, dev, th);
 874         } else {
 875                 dti->dti_off = 0;
 876                 dti->dti_lb.lb_buf = &lastid;
 877                 dti->dti_lb.lb_len = sizeof(lastid);
 878                 dt_read_lock(env, o, 0);
 879                 rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 880                 dt_read_unlock(env, o);
 881                 if (rc == 0 && le64_to_cpu(lastid) > OBIF_MAX_OID) {
 882                         CERROR("%s: bad oid "LPU64" is read from LAST_ID\n",
 883                                o->do_lu.lo_dev->ld_obd->obd_name,
 884                                le64_to_cpu(lastid));
 885                         rc = -EINVAL;
 886                 }
 887         }
 888 out_los:
 889         if (rc != 0) {
 890                 list_del(&(*los)->los_list);
 891                 atomic_dec(&ls->ls_refcount);
 892                 OBD_FREE_PTR(*los);
 893                 *los = NULL;
 894                 if (o != NULL && !IS_ERR(o))
 895                         lu_object_put_nocache(env, &o->do_lu);
 896         } else {
 897                 (*los)->los_seq = fid_seq(first_fid);
 898                 (*los)->los_last_oid = le64_to_cpu(lastid);
 899                 (*los)->los_obj = o;
 900                 /* Read value should not be less than initial one
 901                  * but possible after upgrade from older fs.
 902                  * In this case just switch to the first_oid in memory and
 903                  * it will be updated on disk with first object generated */
 904                 if ((*los)->los_last_oid < first_oid)
 905                         (*los)->los_last_oid = first_oid;
 906         }
 907 out:
 908         mutex_unlock(&ls->ls_los_mutex);
 909         ls_device_put(env, ls);
 910         return rc;
 911 }
 912 EXPORT_SYMBOL(local_oid_storage_init);
 913
 914 void local_oid_storage_fini(const struct lu_env *env,
 915                             struct local_oid_storage *los)
 916 {
 917         struct ls_device *ls;
 918
 919         LASSERT(env);
 920         LASSERT(los->los_dev);
 921         ls = dt2ls_dev(los->los_dev);
 922
 923         /* Take the mutex before decreasing the reference to avoid race
 924          * conditions as described in LU-4721. */
 925         mutex_lock(&ls->ls_los_mutex);
 926         if (!atomic_dec_and_test(&los->los_refcount)) {
 927                 mutex_unlock(&ls->ls_los_mutex);
 928                 return;
 929         }
 930
 931         if (los->los_obj)
 932                 lu_object_put_nocache(env, &los->los_obj->do_lu);
 933         list_del(&los->los_list);
 934         OBD_FREE_PTR(los);
 935         mutex_unlock(&ls->ls_los_mutex);
 936         ls_device_put(env, ls);
 937 }
 938 EXPORT_SYMBOL(local_oid_storage_fini);