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