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, 2014, 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 static 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 static 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         u64                      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 static 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,
 300                                              struct lu_attr *attr,
 301                                              struct dt_object_format *dof)
 302 {
 303         struct dt_thread_info   *dti    = dt_info(env);
 304         struct lu_object_conf   *conf   = &dti->dti_conf;
 305         struct dt_insert_rec    *rec    = &dti->dti_dt_rec;
 306         struct dt_object        *dto;
 307         struct thandle          *th;
 308         int                      rc;
 309
 310         /* We know that the target object does not exist, to be created,
 311          * then give some hints - LOC_F_NEW to help low layer to handle
 312          * that efficiently and properly. */
 313         memset(conf, 0, sizeof(*conf));
 314         conf->loc_flags = LOC_F_NEW;
 315         dto = ls_locate(env, ls, fid, conf);
 316         if (unlikely(IS_ERR(dto)))
 317                 RETURN(dto);
 318
 319         LASSERT(dto != NULL);
 320         if (dt_object_exists(dto))
 321                 GOTO(out, rc = -EEXIST);
 322
 323         th = dt_trans_create(env, ls->ls_osd);
 324         if (IS_ERR(th))
 325                 GOTO(out, rc = PTR_ERR(th));
 326
 327         rc = local_object_declare_create(env, los, dto, attr, dof, th);
 328         if (rc)
 329                 GOTO(trans_stop, rc);
 330
 331         if (dti->dti_dof.dof_type == DFT_DIR) {
 332                 rc = dt_declare_ref_add(env, dto, th);
 333                 if (rc < 0)
 334                         GOTO(trans_stop, rc);
 335
 336                 rc = dt_declare_ref_add(env, parent, th);
 337                 if (rc < 0)
 338                         GOTO(trans_stop, rc);
 339         }
 340
 341         rec->rec_fid = fid;
 342         rec->rec_type = dto->do_lu.lo_header->loh_attr;
 343         rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
 344                                (const struct dt_key *)name, th);
 345         if (rc)
 346                 GOTO(trans_stop, rc);
 347
 348         rc = dt_trans_start_local(env, ls->ls_osd, th);
 349         if (rc)
 350                 GOTO(trans_stop, rc);
 351
 352         dt_write_lock(env, dto, 0);
 353         if (dt_object_exists(dto))
 354                 GOTO(unlock, rc = 0);
 355
 356         CDEBUG(D_OTHER, "create new object "DFID"\n",
 357                PFID(lu_object_fid(&dto->do_lu)));
 358         rc = local_object_create(env, los, dto, attr, dof, th);
 359         if (rc)
 360                 GOTO(unlock, rc);
 361         LASSERT(dt_object_exists(dto));
 362
 363         if (dti->dti_dof.dof_type == DFT_DIR) {
 364                 if (!dt_try_as_dir(env, dto))
 365                         GOTO(destroy, rc = -ENOTDIR);
 366
 367                 rec->rec_type = S_IFDIR;
 368                 rec->rec_fid = fid;
 369                 /* Add "." and ".." for newly created dir */
 370                 rc = dt_insert(env, dto, (const struct dt_rec *)rec,
 371                                (const struct dt_key *)".", th, 1);
 372                 if (rc != 0)
 373                         GOTO(destroy, rc);
 374
 375                 dt_ref_add(env, dto, th);
 376                 rec->rec_fid = lu_object_fid(&parent->do_lu);
 377                 rc = dt_insert(env, dto, (const struct dt_rec *)rec,
 378                                (const struct dt_key *)"..", th, 1);
 379                 if (rc != 0)
 380                         GOTO(destroy, rc);
 381         }
 382
 383         rec->rec_fid = fid;
 384         rec->rec_type = dto->do_lu.lo_header->loh_attr;
 385         dt_write_lock(env, parent, 0);
 386         rc = dt_insert(env, parent, (const struct dt_rec *)rec,
 387                        (const struct dt_key *)name, th, 1);
 388         if (dti->dti_dof.dof_type == DFT_DIR)
 389                 dt_ref_add(env, parent, th);
 390         dt_write_unlock(env, parent);
 391         if (rc)
 392                 GOTO(destroy, rc);
 393 destroy:
 394         if (rc)
 395                 dt_destroy(env, dto, th);
 396 unlock:
 397         dt_write_unlock(env, dto);
 398 trans_stop:
 399         dt_trans_stop(env, ls->ls_osd, th);
 400 out:
 401         if (rc) {
 402                 lu_object_put_nocache(env, &dto->do_lu);
 403                 dto = ERR_PTR(rc);
 404         }
 405         RETURN(dto);
 406 }
 407
 408 /*
 409  * Look up and create (if it does not exist) a local named file or directory in
 410  * parent directory.
 411  */
 412 struct dt_object *local_file_find_or_create(const struct lu_env *env,
 413                                             struct local_oid_storage *los,
 414                                             struct dt_object *parent,
 415                                             const char *name, __u32 mode)
 416 {
 417         struct dt_thread_info   *dti = dt_info(env);
 418         struct dt_object        *dto;
 419         int                      rc;
 420
 421         LASSERT(parent);
 422
 423         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 424         if (rc == 0)
 425                 /* name is found, get the object */
 426                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 427                                 &dti->dti_fid, NULL);
 428         else if (rc != -ENOENT)
 429                 dto = ERR_PTR(rc);
 430         else {
 431                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 432                 if (rc < 0) {
 433                         dto = ERR_PTR(rc);
 434                 } else {
 435                         /* create the object */
 436                         dti->dti_attr.la_valid  = LA_MODE;
 437                         dti->dti_attr.la_mode   = mode;
 438                         dti->dti_dof.dof_type   = dt_mode_to_dft(mode & S_IFMT);
 439                         dto = __local_file_create(env, &dti->dti_fid, los,
 440                                                   dt2ls_dev(los->los_dev),
 441                                                   parent, name, &dti->dti_attr,
 442                                                   &dti->dti_dof);
 443                 }
 444         }
 445         return dto;
 446 }
 447 EXPORT_SYMBOL(local_file_find_or_create);
 448
 449 struct dt_object *local_file_find_or_create_with_fid(const struct lu_env *env,
 450                                                      struct dt_device *dt,
 451                                                      const struct lu_fid *fid,
 452                                                      struct dt_object *parent,
 453                                                      const char *name,
 454                                                      __u32 mode)
 455 {
 456         struct dt_thread_info   *dti = dt_info(env);
 457         struct dt_object        *dto;
 458         int                      rc;
 459
 460         LASSERT(parent);
 461
 462         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 463         if (rc == 0) {
 464                 dto = dt_locate(env, dt, &dti->dti_fid);
 465         } else if (rc != -ENOENT) {
 466                 dto = ERR_PTR(rc);
 467         } else {
 468                 struct ls_device *ls;
 469
 470                 ls = ls_device_get(dt);
 471                 if (IS_ERR(ls)) {
 472                         dto = ERR_PTR(PTR_ERR(ls));
 473                 } else {
 474                         /* create the object */
 475                         dti->dti_attr.la_valid  = LA_MODE;
 476                         dti->dti_attr.la_mode   = mode;
 477                         dti->dti_dof.dof_type   = dt_mode_to_dft(mode & S_IFMT);
 478                         dto = __local_file_create(env, fid, NULL, ls, parent,
 479                                                   name, &dti->dti_attr,
 480                                                   &dti->dti_dof);
 481                         /* ls_device_put() will finalize the ls device, we
 482                          * have to open the object in other device stack */
 483                         if (!IS_ERR(dto)) {
 484                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 485                                 lu_object_put_nocache(env, &dto->do_lu);
 486                                 dto = dt_locate(env, dt, &dti->dti_fid);
 487                         }
 488                         ls_device_put(env, ls);
 489                 }
 490         }
 491         return dto;
 492 }
 493 EXPORT_SYMBOL(local_file_find_or_create_with_fid);
 494
 495 /*
 496  * Look up and create (if it does not exist) a local named index file in parent
 497  * directory.
 498  */
 499 struct dt_object *local_index_find_or_create(const struct lu_env *env,
 500                                              struct local_oid_storage *los,
 501                                              struct dt_object *parent,
 502                                              const char *name, __u32 mode,
 503                                              const struct dt_index_features *ft)
 504 {
 505         struct dt_thread_info   *dti = dt_info(env);
 506         struct dt_object        *dto;
 507         int                      rc;
 508
 509         LASSERT(parent);
 510
 511         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 512         if (rc == 0) {
 513                 /* name is found, get the object */
 514                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 515                                 &dti->dti_fid, NULL);
 516         } else if (rc != -ENOENT) {
 517                 dto = ERR_PTR(rc);
 518         } else {
 519                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 520                 if (rc < 0) {
 521                         dto = ERR_PTR(rc);
 522                 } else {
 523                         /* create the object */
 524                         dti->dti_attr.la_valid          = LA_MODE;
 525                         dti->dti_attr.la_mode           = mode;
 526                         dti->dti_dof.dof_type           = DFT_INDEX;
 527                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 528                         dto = __local_file_create(env, &dti->dti_fid, los,
 529                                                   dt2ls_dev(los->los_dev),
 530                                                   parent, name, &dti->dti_attr,
 531                                                   &dti->dti_dof);
 532                 }
 533         }
 534         return dto;
 535
 536 }
 537 EXPORT_SYMBOL(local_index_find_or_create);
 538
 539 struct dt_object *
 540 local_index_find_or_create_with_fid(const struct lu_env *env,
 541                                     struct dt_device *dt,
 542                                     const struct lu_fid *fid,
 543                                     struct dt_object *parent,
 544                                     const char *name, __u32 mode,
 545                                     const struct dt_index_features *ft)
 546 {
 547         struct dt_thread_info   *dti = dt_info(env);
 548         struct dt_object        *dto;
 549         int                      rc;
 550
 551         LASSERT(parent);
 552
 553         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 554         if (rc == 0) {
 555                 /* name is found, get the object */
 556                 if (!lu_fid_eq(fid, &dti->dti_fid))
 557                         dto = ERR_PTR(-EINVAL);
 558                 else
 559                         dto = dt_locate(env, dt, fid);
 560         } else if (rc != -ENOENT) {
 561                 dto = ERR_PTR(rc);
 562         } else {
 563                 struct ls_device *ls;
 564
 565                 ls = ls_device_get(dt);
 566                 if (IS_ERR(ls)) {
 567                         dto = ERR_PTR(PTR_ERR(ls));
 568                 } else {
 569                         /* create the object */
 570                         dti->dti_attr.la_valid          = LA_MODE;
 571                         dti->dti_attr.la_mode           = mode;
 572                         dti->dti_dof.dof_type           = DFT_INDEX;
 573                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 574                         dto = __local_file_create(env, fid, NULL, ls, parent,
 575                                                   name, &dti->dti_attr,
 576                                                   &dti->dti_dof);
 577                         /* ls_device_put() will finalize the ls device, we
 578                          * have to open the object in other device stack */
 579                         if (!IS_ERR(dto)) {
 580                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 581                                 lu_object_put_nocache(env, &dto->do_lu);
 582                                 dto = dt_locate(env, dt, &dti->dti_fid);
 583                         }
 584                         ls_device_put(env, ls);
 585                 }
 586         }
 587         return dto;
 588 }
 589 EXPORT_SYMBOL(local_index_find_or_create_with_fid);
 590
 591 static int local_object_declare_unlink(const struct lu_env *env,
 592                                        struct dt_device *dt,
 593                                        struct dt_object *p,
 594                                        struct dt_object *c, const char *name,
 595                                        struct thandle *th)
 596 {
 597         int rc;
 598
 599         rc = dt_declare_delete(env, p, (const struct dt_key *)name, th);
 600         if (rc < 0)
 601                 return rc;
 602
 603         rc = dt_declare_ref_del(env, c, th);
 604         if (rc < 0)
 605                 return rc;
 606
 607         return dt_declare_destroy(env, c, th);
 608 }
 609
 610 int local_object_unlink(const struct lu_env *env, struct dt_device *dt,
 611                         struct dt_object *parent, const char *name)
 612 {
 613         struct dt_thread_info   *dti = dt_info(env);
 614         struct dt_object        *dto;
 615         struct thandle          *th;
 616         int                      rc;
 617
 618         ENTRY;
 619
 620         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 621         if (rc == -ENOENT)
 622                 RETURN(0);
 623         else if (rc < 0)
 624                 RETURN(rc);
 625
 626         dto = dt_locate(env, dt, &dti->dti_fid);
 627         if (unlikely(IS_ERR(dto)))
 628                 RETURN(PTR_ERR(dto));
 629
 630         th = dt_trans_create(env, dt);
 631         if (IS_ERR(th))
 632                 GOTO(out, rc = PTR_ERR(th));
 633
 634         rc = local_object_declare_unlink(env, dt, parent, dto, name, th);
 635         if (rc < 0)
 636                 GOTO(stop, rc);
 637
 638         rc = dt_trans_start_local(env, dt, th);
 639         if (rc < 0)
 640                 GOTO(stop, rc);
 641
 642         dt_write_lock(env, dto, 0);
 643         rc = dt_delete(env, parent, (struct dt_key *)name, th);
 644         if (rc < 0)
 645                 GOTO(unlock, rc);
 646
 647         rc = dt_ref_del(env, dto, th);
 648         if (rc < 0) {
 649                 struct dt_insert_rec *rec = &dti->dti_dt_rec;
 650
 651                 rec->rec_fid = &dti->dti_fid;
 652                 rec->rec_type = dto->do_lu.lo_header->loh_attr;
 653                 rc = dt_insert(env, parent, (const struct dt_rec *)rec,
 654                                (const struct dt_key *)name, th, 1);
 655                 GOTO(unlock, rc);
 656         }
 657
 658         rc = dt_destroy(env, dto, th);
 659 unlock:
 660         dt_write_unlock(env, dto);
 661 stop:
 662         dt_trans_stop(env, dt, th);
 663 out:
 664         lu_object_put_nocache(env, &dto->do_lu);
 665         return rc;
 666 }
 667 EXPORT_SYMBOL(local_object_unlink);
 668
 669 struct local_oid_storage *dt_los_find(struct ls_device *ls, __u64 seq)
 670 {
 671         struct local_oid_storage *los, *ret = NULL;
 672
 673         list_for_each_entry(los, &ls->ls_los_list, los_list) {
 674                 if (los->los_seq == seq) {
 675                         atomic_inc(&los->los_refcount);
 676                         ret = los;
 677                         break;
 678                 }
 679         }
 680         return ret;
 681 }
 682
 683 void dt_los_put(struct local_oid_storage *los)
 684 {
 685         if (atomic_dec_and_test(&los->los_refcount))
 686                 /* should never happen, only local_oid_storage_fini should
 687                  * drop refcount to zero */
 688                 LBUG();
 689         return;
 690 }
 691
 692 /* after Lustre 2.3 release there may be old file to store last generated FID
 693  * If such file exists then we have to read its content
 694  */
 695 static int lastid_compat_check(const struct lu_env *env, struct dt_device *dev,
 696                                __u64 lastid_seq, __u32 *first_oid,
 697                                struct ls_device *ls)
 698 {
 699         struct dt_thread_info   *dti = dt_info(env);
 700         struct dt_object        *root = NULL;
 701         struct los_ondisk        losd;
 702         struct dt_object        *o = NULL;
 703         int                      rc = 0;
 704
 705         rc = dt_root_get(env, dev, &dti->dti_fid);
 706         if (rc)
 707                 return rc;
 708
 709         root = ls_locate(env, ls, &dti->dti_fid, NULL);
 710         if (IS_ERR(root))
 711                 return PTR_ERR(root);
 712
 713         /* find old last_id file */
 714         snprintf(dti->dti_buf, sizeof(dti->dti_buf), "seq-"LPX64"-lastid",
 715                  lastid_seq);
 716         rc = dt_lookup_dir(env, root, dti->dti_buf, &dti->dti_fid);
 717         lu_object_put_nocache(env, &root->do_lu);
 718         if (rc == -ENOENT) {
 719                 /* old llog lastid accessed by FID only */
 720                 if (lastid_seq != FID_SEQ_LLOG)
 721                         return 0;
 722                 dti->dti_fid.f_seq = FID_SEQ_LLOG;
 723                 dti->dti_fid.f_oid = 1;
 724                 dti->dti_fid.f_ver = 0;
 725                 o = ls_locate(env, ls, &dti->dti_fid, NULL);
 726                 if (IS_ERR(o))
 727                         return PTR_ERR(o);
 728
 729                 if (!dt_object_exists(o)) {
 730                         lu_object_put_nocache(env, &o->do_lu);
 731                         return 0;
 732                 }
 733                 CDEBUG(D_INFO, "Found old llog lastid file\n");
 734         } else if (rc < 0) {
 735                 return rc;
 736         } else {
 737                 CDEBUG(D_INFO, "Found old lastid file for sequence "LPX64"\n",
 738                        lastid_seq);
 739                 o = ls_locate(env, ls, &dti->dti_fid, NULL);
 740                 if (IS_ERR(o))
 741                         return PTR_ERR(o);
 742         }
 743         /* let's read seq-NNNNNN-lastid file value */
 744         LASSERT(dt_object_exists(o));
 745         dti->dti_off = 0;
 746         dti->dti_lb.lb_buf = &losd;
 747         dti->dti_lb.lb_len = sizeof(losd);
 748         dt_read_lock(env, o, 0);
 749         rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 750         dt_read_unlock(env, o);
 751         if (rc == 0 && le32_to_cpu(losd.lso_magic) != LOS_MAGIC) {
 752                 CERROR("%s: wrong content of seq-"LPX64"-lastid file, magic %x\n",
 753                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq,
 754                        le32_to_cpu(losd.lso_magic));
 755                 rc = -EINVAL;
 756         } else if (rc < 0) {
 757                 CERROR("%s: failed to read seq-"LPX64"-lastid: rc = %d\n",
 758                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq, rc);
 759         }
 760         lu_object_put_nocache(env, &o->do_lu);
 761         if (rc == 0)
 762                 *first_oid = le32_to_cpu(losd.lso_next_oid);
 763         return rc;
 764 }
 765
 766 /**
 767  * Initialize local OID storage for required sequence.
 768  * That may be needed for services that uses local files and requires
 769  * dynamic OID allocation for them.
 770  *
 771  * Per each sequence we have an object with 'first_fid' identificator
 772  * containing the counter for OIDs of locally created files with that
 773  * sequence.
 774  *
 775  * It is used now by llog subsystem and MGS for NID tables
 776  *
 777  * Function gets first_fid to create counter object.
 778  * All dynamic fids will be generated with the same sequence and incremented
 779  * OIDs
 780  *
 781  * Returned local_oid_storage is in-memory representaion of OID storage
 782  */
 783 int local_oid_storage_init(const struct lu_env *env, struct dt_device *dev,
 784                            const struct lu_fid *first_fid,
 785                            struct local_oid_storage **los)
 786 {
 787         struct dt_thread_info   *dti = dt_info(env);
 788         struct ls_device        *ls;
 789         u64                      lastid;
 790         struct dt_object        *o = NULL;
 791         struct thandle          *th;
 792         __u32                    first_oid = fid_oid(first_fid);
 793         int                      rc = 0;
 794
 795         ENTRY;
 796
 797         ls = ls_device_get(dev);
 798         if (IS_ERR(ls))
 799                 RETURN(PTR_ERR(ls));
 800
 801         mutex_lock(&ls->ls_los_mutex);
 802         *los = dt_los_find(ls, fid_seq(first_fid));
 803         if (*los != NULL)
 804                 GOTO(out, rc = 0);
 805
 806         /* not found, then create */
 807         OBD_ALLOC_PTR(*los);
 808         if (*los == NULL)
 809                 GOTO(out, rc = -ENOMEM);
 810
 811         atomic_set(&(*los)->los_refcount, 1);
 812         mutex_init(&(*los)->los_id_lock);
 813         (*los)->los_dev = &ls->ls_top_dev;
 814         atomic_inc(&ls->ls_refcount);
 815         list_add(&(*los)->los_list, &ls->ls_los_list);
 816
 817         /* Use {seq, 0, 0} to create the LAST_ID file for every
 818          * sequence.  OIDs start at LUSTRE_FID_INIT_OID.
 819          */
 820         dti->dti_fid.f_seq = fid_seq(first_fid);
 821         dti->dti_fid.f_oid = LUSTRE_FID_LASTID_OID;
 822         dti->dti_fid.f_ver = 0;
 823         o = ls_locate(env, ls, &dti->dti_fid, NULL);
 824         if (IS_ERR(o))
 825                 GOTO(out_los, rc = PTR_ERR(o));
 826
 827         if (!dt_object_exists(o)) {
 828                 rc = lastid_compat_check(env, dev, fid_seq(first_fid),
 829                                          &first_oid, ls);
 830                 if (rc < 0)
 831                         GOTO(out_los, rc);
 832
 833                 th = dt_trans_create(env, dev);
 834                 if (IS_ERR(th))
 835                         GOTO(out_los, rc = PTR_ERR(th));
 836
 837                 dti->dti_attr.la_valid = LA_MODE | LA_TYPE;
 838                 dti->dti_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
 839                 dti->dti_dof.dof_type = dt_mode_to_dft(S_IFREG);
 840
 841                 rc = dt_declare_create(env, o, &dti->dti_attr, NULL,
 842                                        &dti->dti_dof, th);
 843                 if (rc)
 844                         GOTO(out_trans, rc);
 845
 846                 lastid = cpu_to_le64(first_oid);
 847
 848                 dti->dti_off = 0;
 849                 dti->dti_lb.lb_buf = &lastid;
 850                 dti->dti_lb.lb_len = sizeof(lastid);
 851                 rc = dt_declare_record_write(env, o, &dti->dti_lb, dti->dti_off,
 852                                              th);
 853                 if (rc)
 854                         GOTO(out_trans, rc);
 855
 856                 rc = dt_trans_start_local(env, dev, th);
 857                 if (rc)
 858                         GOTO(out_trans, rc);
 859
 860                 dt_write_lock(env, o, 0);
 861                 if (dt_object_exists(o))
 862                         GOTO(out_lock, rc = 0);
 863
 864                 rc = dt_create(env, o, &dti->dti_attr, NULL, &dti->dti_dof,
 865                                th);
 866                 if (rc)
 867                         GOTO(out_lock, rc);
 868
 869                 rc = dt_record_write(env, o, &dti->dti_lb, &dti->dti_off, th);
 870                 if (rc)
 871                         GOTO(out_lock, rc);
 872 out_lock:
 873                 dt_write_unlock(env, o);
 874 out_trans:
 875                 dt_trans_stop(env, dev, th);
 876         } else {
 877                 dti->dti_off = 0;
 878                 dti->dti_lb.lb_buf = &lastid;
 879                 dti->dti_lb.lb_len = sizeof(lastid);
 880                 dt_read_lock(env, o, 0);
 881                 rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 882                 dt_read_unlock(env, o);
 883                 if (rc == 0 && le64_to_cpu(lastid) > OBIF_MAX_OID) {
 884                         CERROR("%s: bad oid "LPU64" is read from LAST_ID\n",
 885                                o->do_lu.lo_dev->ld_obd->obd_name,
 886                                le64_to_cpu(lastid));
 887                         rc = -EINVAL;
 888                 }
 889         }
 890 out_los:
 891         if (rc != 0) {
 892                 list_del(&(*los)->los_list);
 893                 atomic_dec(&ls->ls_refcount);
 894                 OBD_FREE_PTR(*los);
 895                 *los = NULL;
 896                 if (o != NULL && !IS_ERR(o))
 897                         lu_object_put_nocache(env, &o->do_lu);
 898         } else {
 899                 (*los)->los_seq = fid_seq(first_fid);
 900                 (*los)->los_last_oid = le64_to_cpu(lastid);
 901                 (*los)->los_obj = o;
 902                 /* Read value should not be less than initial one
 903                  * but possible after upgrade from older fs.
 904                  * In this case just switch to the first_oid in memory and
 905                  * it will be updated on disk with first object generated */
 906                 if ((*los)->los_last_oid < first_oid)
 907                         (*los)->los_last_oid = first_oid;
 908         }
 909 out:
 910         mutex_unlock(&ls->ls_los_mutex);
 911         ls_device_put(env, ls);
 912         return rc;
 913 }
 914 EXPORT_SYMBOL(local_oid_storage_init);
 915
 916 void local_oid_storage_fini(const struct lu_env *env,
 917                             struct local_oid_storage *los)
 918 {
 919         struct ls_device *ls;
 920
 921         LASSERT(env);
 922         LASSERT(los->los_dev);
 923         ls = dt2ls_dev(los->los_dev);
 924
 925         /* Take the mutex before decreasing the reference to avoid race
 926          * conditions as described in LU-4721. */
 927         mutex_lock(&ls->ls_los_mutex);
 928         if (!atomic_dec_and_test(&los->los_refcount)) {
 929                 mutex_unlock(&ls->ls_los_mutex);
 930                 return;
 931         }
 932
 933         if (los->los_obj)
 934                 lu_object_put_nocache(env, &los->los_obj->do_lu);
 935         list_del(&los->los_list);
 936         OBD_FREE_PTR(los);
 937         mutex_unlock(&ls->ls_los_mutex);
 938         ls_device_put(env, ls);
 939 }
 940 EXPORT_SYMBOL(local_oid_storage_fini);