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, 2017, 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 LIST_HEAD(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_of(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_PRE(obj, sizeof(*obj), "kfreed");
  69         kfree_rcu(obj, ls_header.loh_rcu);
  70 }
  71
  72 static struct lu_object_operations ls_lu_obj_ops = {
  73         .loo_object_init  = ls_object_init,
  74         .loo_object_free  = ls_object_free,
  75 };
  76
  77 static struct lu_object *ls_object_alloc(const struct lu_env *env,
  78                                          const struct lu_object_header *_h,
  79                                          struct lu_device *d)
  80 {
  81         struct lu_object_header *h;
  82         struct ls_object        *o;
  83         struct lu_object        *l;
  84
  85         LASSERT(_h == NULL);
  86
  87         OBD_ALLOC_PTR(o);
  88         if (o != NULL) {
  89                 l = &o->ls_obj.do_lu;
  90                 h = &o->ls_header;
  91
  92                 lu_object_header_init(h);
  93                 dt_object_init(&o->ls_obj, h, d);
  94                 lu_object_add_top(h, l);
  95
  96                 l->lo_ops = &ls_lu_obj_ops;
  97
  98                 return l;
  99         } else {
 100                 return NULL;
 101         }
 102 }
 103
 104 static struct lu_device_operations ls_lu_dev_ops = {
 105         .ldo_object_alloc =     ls_object_alloc
 106 };
 107
 108 static struct ls_device *__ls_find_dev(struct dt_device *dev)
 109 {
 110         struct ls_device *ls, *ret = NULL;
 111
 112         list_for_each_entry(ls, &ls_list_head, ls_linkage) {
 113                 if (ls->ls_osd == dev) {
 114                         atomic_inc(&ls->ls_refcount);
 115                         ret = ls;
 116                         break;
 117                 }
 118         }
 119         return ret;
 120 }
 121
 122 struct ls_device *ls_find_dev(struct dt_device *dev)
 123 {
 124         struct ls_device *ls;
 125
 126         mutex_lock(&ls_list_mutex);
 127         ls = __ls_find_dev(dev);
 128         mutex_unlock(&ls_list_mutex);
 129
 130         return ls;
 131 }
 132
 133 static struct lu_device_type_operations ls_device_type_ops = {
 134         .ldto_start = NULL,
 135         .ldto_stop  = NULL,
 136 };
 137
 138 static struct lu_device_type ls_lu_type = {
 139         .ldt_name = "local_storage",
 140         .ldt_ops  = &ls_device_type_ops,
 141 };
 142
 143 struct ls_device *ls_device_get(struct dt_device *dev)
 144 {
 145         struct ls_device *ls;
 146
 147         ENTRY;
 148
 149         mutex_lock(&ls_list_mutex);
 150         ls = __ls_find_dev(dev);
 151         if (ls)
 152                 GOTO(out_ls, ls);
 153
 154         /* not found, then create */
 155         OBD_ALLOC_PTR(ls);
 156         if (ls == NULL)
 157                 GOTO(out_ls, ls = ERR_PTR(-ENOMEM));
 158
 159         atomic_set(&ls->ls_refcount, 1);
 160         INIT_LIST_HEAD(&ls->ls_los_list);
 161         mutex_init(&ls->ls_los_mutex);
 162
 163         ls->ls_osd = dev;
 164
 165         LASSERT(dev->dd_lu_dev.ld_site);
 166         lu_device_init(&ls->ls_top_dev.dd_lu_dev, &ls_lu_type);
 167         ls->ls_top_dev.dd_lu_dev.ld_ops = &ls_lu_dev_ops;
 168         ls->ls_top_dev.dd_lu_dev.ld_site = dev->dd_lu_dev.ld_site;
 169
 170         /* finally add ls to the list */
 171         list_add(&ls->ls_linkage, &ls_list_head);
 172 out_ls:
 173         mutex_unlock(&ls_list_mutex);
 174         RETURN(ls);
 175 }
 176
 177 void ls_device_put(const struct lu_env *env, struct ls_device *ls)
 178 {
 179         LASSERT(env);
 180         if (!atomic_dec_and_test(&ls->ls_refcount))
 181                 return;
 182
 183         mutex_lock(&ls_list_mutex);
 184         if (atomic_read(&ls->ls_refcount) == 0) {
 185                 LASSERT(list_empty(&ls->ls_los_list));
 186                 list_del(&ls->ls_linkage);
 187                 lu_site_purge(env, ls->ls_top_dev.dd_lu_dev.ld_site, ~0);
 188                 lu_device_fini(&ls->ls_top_dev.dd_lu_dev);
 189                 OBD_FREE_PTR(ls);
 190         }
 191         mutex_unlock(&ls_list_mutex);
 192 }
 193
 194 /**
 195  * local file fid generation
 196  */
 197 int local_object_fid_generate(const struct lu_env *env,
 198                               struct local_oid_storage *los,
 199                               struct lu_fid *fid)
 200 {
 201         LASSERT(los->los_dev);
 202         LASSERT(los->los_obj);
 203
 204         /* take next OID */
 205
 206         /* to make it unique after reboot we store
 207          * the latest generated fid atomically with
 208          * object creation see local_object_create() */
 209
 210         mutex_lock(&los->los_id_lock);
 211         fid->f_seq = los->los_seq;
 212         fid->f_oid = ++los->los_last_oid;
 213         fid->f_ver = 0;
 214         mutex_unlock(&los->los_id_lock);
 215
 216         return 0;
 217 }
 218
 219 int local_object_declare_create(const struct lu_env *env,
 220                                 struct local_oid_storage *los,
 221                                 struct dt_object *o, struct lu_attr *attr,
 222                                 struct dt_object_format *dof,
 223                                 struct thandle *th)
 224 {
 225         struct dt_thread_info   *dti = dt_info(env);
 226         int                      rc;
 227
 228         ENTRY;
 229
 230         /* update fid generation file */
 231         if (los != NULL) {
 232                 LASSERT(dt_object_exists(los->los_obj));
 233                 dti->dti_lb.lb_buf = NULL;
 234                 dti->dti_lb.lb_len = sizeof(struct los_ondisk);
 235                 rc = dt_declare_record_write(env, los->los_obj,
 236                                              &dti->dti_lb, 0, th);
 237                 if (rc)
 238                         RETURN(rc);
 239         }
 240
 241         rc = dt_declare_create(env, o, attr, NULL, dof, th);
 242         if (rc)
 243                 RETURN(rc);
 244
 245         dti->dti_lb.lb_buf = NULL;
 246         dti->dti_lb.lb_len = sizeof(dti->dti_lma);
 247         rc = dt_declare_xattr_set(env, o, &dti->dti_lb, XATTR_NAME_LMA, 0, th);
 248
 249         RETURN(rc);
 250 }
 251
 252 int local_object_create(const struct lu_env *env,
 253                         struct local_oid_storage *los,
 254                         struct dt_object *o, struct lu_attr *attr,
 255                         struct dt_object_format *dof, struct thandle *th)
 256 {
 257         struct dt_thread_info   *dti = dt_info(env);
 258         u64                      lastid;
 259         int                      rc;
 260
 261         ENTRY;
 262
 263         rc = dt_create(env, o, attr, NULL, dof, th);
 264         if (rc)
 265                 RETURN(rc);
 266
 267         if (los == NULL)
 268                 RETURN(rc);
 269
 270         LASSERT(los->los_obj);
 271         LASSERT(dt_object_exists(los->los_obj));
 272
 273         /* many threads can be updated this, serialize
 274          * them here to avoid the race where one thread
 275          * takes the value first, but writes it last */
 276         mutex_lock(&los->los_id_lock);
 277
 278         /* update local oid number on disk so that
 279          * we know the last one used after reboot */
 280         lastid = cpu_to_le64(los->los_last_oid);
 281
 282         dti->dti_off = 0;
 283         dti->dti_lb.lb_buf = &lastid;
 284         dti->dti_lb.lb_len = sizeof(lastid);
 285         rc = dt_record_write(env, los->los_obj, &dti->dti_lb, &dti->dti_off,
 286                              th);
 287         mutex_unlock(&los->los_id_lock);
 288
 289         RETURN(rc);
 290 }
 291
 292 /*
 293  * Create local named object (file, directory or index) in parent directory.
 294  */
 295 static struct dt_object *__local_file_create(const struct lu_env *env,
 296                                              const struct lu_fid *fid,
 297                                              struct local_oid_storage *los,
 298                                              struct ls_device *ls,
 299                                              struct dt_object *parent,
 300                                              const char *name,
 301                                              struct lu_attr *attr,
 302                                              struct dt_object_format *dof)
 303 {
 304         struct dt_thread_info   *dti    = dt_info(env);
 305         struct lu_object_conf   *conf   = &dti->dti_conf;
 306         struct dt_insert_rec    *rec    = &dti->dti_dt_rec;
 307         struct dt_object        *dto;
 308         struct thandle          *th;
 309         int                      rc;
 310
 311         /* We know that the target object does not exist, to be created,
 312          * then give some hints - LOC_F_NEW to help low layer to handle
 313          * that efficiently and properly. */
 314         memset(conf, 0, sizeof(*conf));
 315         conf->loc_flags = LOC_F_NEW;
 316         dto = ls_locate(env, ls, fid, conf);
 317         if (unlikely(IS_ERR(dto)))
 318                 RETURN(dto);
 319
 320         LASSERT(dto != NULL);
 321         if (dt_object_exists(dto))
 322                 GOTO(out, rc = -EEXIST);
 323
 324         th = dt_trans_create(env, ls->ls_osd);
 325         if (IS_ERR(th))
 326                 GOTO(out, rc = PTR_ERR(th));
 327
 328         rc = local_object_declare_create(env, los, dto, attr, dof, th);
 329         if (rc)
 330                 GOTO(trans_stop, rc);
 331
 332         if (dti->dti_dof.dof_type == DFT_DIR) {
 333                 rc = dt_declare_ref_add(env, dto, th);
 334                 if (rc < 0)
 335                         GOTO(trans_stop, rc);
 336
 337                 rc = dt_declare_ref_add(env, parent, th);
 338                 if (rc < 0)
 339                         GOTO(trans_stop, rc);
 340         }
 341
 342         rec->rec_fid = fid;
 343         rec->rec_type = attr->la_mode & S_IFMT;
 344         rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
 345                                (const struct dt_key *)name, th);
 346         if (rc)
 347                 GOTO(trans_stop, rc);
 348
 349         if (dti->dti_dof.dof_type == DFT_DIR) {
 350                 if (!dt_try_as_dir(env, dto))
 351                         GOTO(trans_stop, rc = -ENOTDIR);
 352
 353                 rec->rec_type = S_IFDIR;
 354                 rec->rec_fid = fid;
 355                 rc = dt_declare_insert(env, dto, (const struct dt_rec *)rec,
 356                                 (const struct dt_key *)".", th);
 357                 if (rc != 0)
 358                         GOTO(trans_stop, rc);
 359
 360                 rec->rec_fid = lu_object_fid(&parent->do_lu);
 361                 rc = dt_declare_insert(env, dto, (const struct dt_rec *)rec,
 362                                 (const struct dt_key *)"..", th);
 363                 if (rc != 0)
 364                         GOTO(trans_stop, rc);
 365
 366                 rc = dt_declare_ref_add(env, dto, th);
 367                 if (rc != 0)
 368                         GOTO(trans_stop, rc);
 369         }
 370
 371         rc = dt_trans_start_local(env, ls->ls_osd, th);
 372         if (rc)
 373                 GOTO(trans_stop, rc);
 374
 375         dt_write_lock(env, dto, DT_SRC_CHILD);
 376         if (dt_object_exists(dto))
 377                 GOTO(unlock, rc = 0);
 378
 379         CDEBUG(D_OTHER, "create new object "DFID"\n",
 380                PFID(lu_object_fid(&dto->do_lu)));
 381         rc = local_object_create(env, los, dto, attr, dof, th);
 382         if (rc)
 383                 GOTO(unlock, rc);
 384         LASSERT(dt_object_exists(dto));
 385
 386         if (dti->dti_dof.dof_type == DFT_DIR) {
 387
 388                 rec->rec_type = S_IFDIR;
 389                 rec->rec_fid = fid;
 390                 /* Add "." and ".." for newly created dir */
 391                 rc = dt_insert(env, dto, (const struct dt_rec *)rec,
 392                                (const struct dt_key *)".", th);
 393                 if (rc != 0)
 394                         GOTO(destroy, rc);
 395
 396                 dt_ref_add(env, dto, th);
 397                 rec->rec_fid = lu_object_fid(&parent->do_lu);
 398                 rc = dt_insert(env, dto, (const struct dt_rec *)rec,
 399                                (const struct dt_key *)"..", th);
 400                 if (rc != 0)
 401                         GOTO(destroy, rc);
 402         }
 403
 404         rec->rec_fid = fid;
 405         rec->rec_type = dto->do_lu.lo_header->loh_attr;
 406         dt_write_lock(env, parent, DT_SRC_PARENT);
 407         rc = dt_insert(env, parent, (const struct dt_rec *)rec,
 408                        (const struct dt_key *)name, th);
 409         if (dti->dti_dof.dof_type == DFT_DIR)
 410                 dt_ref_add(env, parent, th);
 411         dt_write_unlock(env, parent);
 412         if (rc)
 413                 GOTO(destroy, rc);
 414 destroy:
 415         if (rc)
 416                 dt_destroy(env, dto, th);
 417 unlock:
 418         dt_write_unlock(env, dto);
 419 trans_stop:
 420         dt_trans_stop(env, ls->ls_osd, th);
 421 out:
 422         if (rc) {
 423                 dt_object_put_nocache(env, dto);
 424                 dto = ERR_PTR(rc);
 425         }
 426         RETURN(dto);
 427 }
 428
 429 struct dt_object *local_file_find(const struct lu_env *env,
 430                                   struct local_oid_storage *los,
 431                                   struct dt_object *parent,
 432                                   const char *name)
 433 {
 434         struct dt_thread_info   *dti = dt_info(env);
 435         struct dt_object        *dto;
 436         int                      rc;
 437
 438         LASSERT(parent);
 439
 440         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 441         if (!rc)
 442                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 443                                 &dti->dti_fid, NULL);
 444         else
 445                 dto = ERR_PTR(rc);
 446
 447         return dto;
 448 }
 449 EXPORT_SYMBOL(local_file_find);
 450
 451 /*
 452  * Look up and create (if it does not exist) a local named file or directory in
 453  * parent directory.
 454  */
 455 struct dt_object *local_file_find_or_create(const struct lu_env *env,
 456                                             struct local_oid_storage *los,
 457                                             struct dt_object *parent,
 458                                             const char *name, __u32 mode)
 459 {
 460         struct dt_thread_info   *dti = dt_info(env);
 461         struct dt_object        *dto;
 462         int                      rc;
 463
 464         dto = local_file_find(env, los, parent, name);
 465         if (!IS_ERR(dto) || PTR_ERR(dto) != -ENOENT)
 466                 return dto;
 467
 468         rc = local_object_fid_generate(env, los, &dti->dti_fid);
 469         if (rc)
 470                 return ERR_PTR(rc);
 471
 472         /* create the object */
 473         dti->dti_attr.la_valid = LA_MODE;
 474         dti->dti_attr.la_mode = mode;
 475         dti->dti_dof.dof_type = dt_mode_to_dft(mode & S_IFMT);
 476         dto = __local_file_create(env, &dti->dti_fid, los,
 477                                   dt2ls_dev(los->los_dev), parent, name,
 478                                   &dti->dti_attr, &dti->dti_dof);
 479         return dto;
 480 }
 481 EXPORT_SYMBOL(local_file_find_or_create);
 482
 483 struct dt_object *local_file_find_or_create_with_fid(const struct lu_env *env,
 484                                                      struct dt_device *dt,
 485                                                      const struct lu_fid *fid,
 486                                                      struct dt_object *parent,
 487                                                      const char *name,
 488                                                      __u32 mode)
 489 {
 490         struct dt_thread_info   *dti = dt_info(env);
 491         struct dt_object        *dto;
 492         int                      rc;
 493
 494         LASSERT(parent);
 495
 496         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 497         if (rc == 0) {
 498                 dto = dt_locate(env, dt, &dti->dti_fid);
 499         } else if (rc != -ENOENT) {
 500                 dto = ERR_PTR(rc);
 501         } else {
 502                 struct ls_device *ls;
 503
 504                 ls = ls_device_get(dt);
 505                 if (IS_ERR(ls)) {
 506                         dto = ERR_CAST(ls);
 507                 } else {
 508                         /* create the object */
 509                         dti->dti_attr.la_valid  = LA_MODE;
 510                         dti->dti_attr.la_mode   = mode;
 511                         dti->dti_dof.dof_type   = dt_mode_to_dft(mode & S_IFMT);
 512                         dto = __local_file_create(env, fid, NULL, ls, parent,
 513                                                   name, &dti->dti_attr,
 514                                                   &dti->dti_dof);
 515                         /* ls_device_put() will finalize the ls device, we
 516                          * have to open the object in other device stack */
 517                         if (!IS_ERR(dto)) {
 518                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 519                                 dt_object_put_nocache(env, dto);
 520                                 dto = dt_locate(env, dt, &dti->dti_fid);
 521                         }
 522                         ls_device_put(env, ls);
 523                 }
 524         }
 525         return dto;
 526 }
 527 EXPORT_SYMBOL(local_file_find_or_create_with_fid);
 528
 529 /*
 530  * Look up and create (if it does not exist) a local named index file in parent
 531  * directory.
 532  */
 533 struct dt_object *local_index_find_or_create(const struct lu_env *env,
 534                                              struct local_oid_storage *los,
 535                                              struct dt_object *parent,
 536                                              const char *name, __u32 mode,
 537                                              const struct dt_index_features *ft)
 538 {
 539         struct dt_thread_info   *dti = dt_info(env);
 540         struct dt_object        *dto;
 541         int                      rc;
 542
 543         LASSERT(parent);
 544
 545         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 546         if (rc == 0) {
 547                 /* name is found, get the object */
 548                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 549                                 &dti->dti_fid, NULL);
 550         } else if (rc != -ENOENT) {
 551                 dto = ERR_PTR(rc);
 552         } else {
 553                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 554                 if (rc < 0) {
 555                         dto = ERR_PTR(rc);
 556                 } else {
 557                         /* create the object */
 558                         dti->dti_attr.la_valid          = LA_MODE;
 559                         dti->dti_attr.la_mode           = mode;
 560                         dti->dti_dof.dof_type           = DFT_INDEX;
 561                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 562                         dto = __local_file_create(env, &dti->dti_fid, los,
 563                                                   dt2ls_dev(los->los_dev),
 564                                                   parent, name, &dti->dti_attr,
 565                                                   &dti->dti_dof);
 566                 }
 567         }
 568         return dto;
 569
 570 }
 571 EXPORT_SYMBOL(local_index_find_or_create);
 572
 573 struct dt_object *
 574 local_index_find_or_create_with_fid(const struct lu_env *env,
 575                                     struct dt_device *dt,
 576                                     const struct lu_fid *fid,
 577                                     struct dt_object *parent,
 578                                     const char *name, __u32 mode,
 579                                     const struct dt_index_features *ft)
 580 {
 581         struct dt_thread_info   *dti = dt_info(env);
 582         struct dt_object        *dto;
 583         int                      rc;
 584
 585         LASSERT(parent);
 586
 587         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 588         if (rc == 0) {
 589                 /* name is found, get the object */
 590                 if (!lu_fid_eq(fid, &dti->dti_fid))
 591                         dto = ERR_PTR(-EINVAL);
 592                 else
 593                         dto = dt_locate(env, dt, fid);
 594         } else if (rc != -ENOENT) {
 595                 dto = ERR_PTR(rc);
 596         } else {
 597                 struct ls_device *ls;
 598
 599                 ls = ls_device_get(dt);
 600                 if (IS_ERR(ls)) {
 601                         dto = ERR_CAST(ls);
 602                 } else {
 603                         /* create the object */
 604                         dti->dti_attr.la_valid          = LA_MODE;
 605                         dti->dti_attr.la_mode           = mode;
 606                         dti->dti_dof.dof_type           = DFT_INDEX;
 607                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 608                         dto = __local_file_create(env, fid, NULL, ls, parent,
 609                                                   name, &dti->dti_attr,
 610                                                   &dti->dti_dof);
 611                         /* ls_device_put() will finalize the ls device, we
 612                          * have to open the object in other device stack */
 613                         if (!IS_ERR(dto)) {
 614                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 615                                 dt_object_put_nocache(env, dto);
 616                                 dto = dt_locate(env, dt, &dti->dti_fid);
 617                         }
 618                         ls_device_put(env, ls);
 619                 }
 620         }
 621         return dto;
 622 }
 623 EXPORT_SYMBOL(local_index_find_or_create_with_fid);
 624
 625 static int local_object_declare_unlink(const struct lu_env *env,
 626                                        struct dt_device *dt,
 627                                        struct dt_object *p,
 628                                        struct dt_object *c, const char *name,
 629                                        struct thandle *th)
 630 {
 631         int rc;
 632
 633         rc = dt_declare_delete(env, p, (const struct dt_key *)name, th);
 634         if (rc < 0)
 635                 return rc;
 636
 637         rc = dt_declare_ref_del(env, c, th);
 638         if (rc < 0)
 639                 return rc;
 640
 641         return dt_declare_destroy(env, c, th);
 642 }
 643
 644 int local_object_unlink(const struct lu_env *env, struct dt_device *dt,
 645                         struct dt_object *parent, const char *name)
 646 {
 647         struct dt_thread_info   *dti = dt_info(env);
 648         struct dt_object        *dto;
 649         struct thandle          *th;
 650         int                      rc;
 651
 652         ENTRY;
 653
 654         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 655         if (rc == -ENOENT)
 656                 RETURN(0);
 657         else if (rc < 0)
 658                 RETURN(rc);
 659
 660         dto = dt_locate(env, dt, &dti->dti_fid);
 661         if (unlikely(IS_ERR(dto)))
 662                 RETURN(PTR_ERR(dto));
 663
 664         th = dt_trans_create(env, dt);
 665         if (IS_ERR(th))
 666                 GOTO(out, rc = PTR_ERR(th));
 667
 668         rc = local_object_declare_unlink(env, dt, parent, dto, name, th);
 669         if (rc < 0)
 670                 GOTO(stop, rc);
 671
 672         rc = dt_trans_start_local(env, dt, th);
 673         if (rc < 0)
 674                 GOTO(stop, rc);
 675
 676         dt_write_lock(env, dto, 0);
 677         rc = dt_delete(env, parent, (struct dt_key *)name, th);
 678         if (rc < 0)
 679                 GOTO(unlock, rc);
 680
 681         rc = dt_ref_del(env, dto, th);
 682         if (rc < 0) {
 683                 struct dt_insert_rec *rec = &dti->dti_dt_rec;
 684
 685                 rec->rec_fid = &dti->dti_fid;
 686                 rec->rec_type = dto->do_lu.lo_header->loh_attr;
 687                 rc = dt_insert(env, parent, (const struct dt_rec *)rec,
 688                                (const struct dt_key *)name, th);
 689                 GOTO(unlock, rc);
 690         }
 691
 692         rc = dt_destroy(env, dto, th);
 693 unlock:
 694         dt_write_unlock(env, dto);
 695 stop:
 696         dt_trans_stop(env, dt, th);
 697 out:
 698         dt_object_put_nocache(env, dto);
 699         return rc;
 700 }
 701 EXPORT_SYMBOL(local_object_unlink);
 702
 703 struct local_oid_storage *dt_los_find(struct ls_device *ls, __u64 seq)
 704 {
 705         struct local_oid_storage *los, *ret = NULL;
 706
 707         list_for_each_entry(los, &ls->ls_los_list, los_list) {
 708                 if (los->los_seq == seq) {
 709                         atomic_inc(&los->los_refcount);
 710                         ret = los;
 711                         break;
 712                 }
 713         }
 714         return ret;
 715 }
 716
 717 void dt_los_put(struct local_oid_storage *los)
 718 {
 719         if (atomic_dec_and_test(&los->los_refcount))
 720                 /* should never happen, only local_oid_storage_fini should
 721                  * drop refcount to zero */
 722                 LBUG();
 723 }
 724
 725 /* after Lustre 2.3 release there may be old file to store last generated FID
 726  * If such file exists then we have to read its content
 727  */
 728 static int lastid_compat_check(const struct lu_env *env, struct dt_device *dev,
 729                                __u64 lastid_seq, __u32 *first_oid,
 730                                struct ls_device *ls)
 731 {
 732         struct dt_thread_info   *dti = dt_info(env);
 733         struct dt_object        *root = NULL;
 734         struct los_ondisk        losd;
 735         struct dt_object        *o = NULL;
 736         int                      rc = 0;
 737
 738         rc = dt_root_get(env, dev, &dti->dti_fid);
 739         if (rc)
 740                 return rc;
 741
 742         root = ls_locate(env, ls, &dti->dti_fid, NULL);
 743         if (IS_ERR(root))
 744                 return PTR_ERR(root);
 745
 746         /* find old last_id file */
 747         snprintf(dti->dti_buf, sizeof(dti->dti_buf), "seq-%#llx-lastid",
 748                  lastid_seq);
 749         rc = dt_lookup_dir(env, root, dti->dti_buf, &dti->dti_fid);
 750         dt_object_put_nocache(env, root);
 751         if (rc == -ENOENT) {
 752                 /* old llog lastid accessed by FID only */
 753                 if (lastid_seq != FID_SEQ_LLOG)
 754                         return 0;
 755                 dti->dti_fid.f_seq = FID_SEQ_LLOG;
 756                 dti->dti_fid.f_oid = 1;
 757                 dti->dti_fid.f_ver = 0;
 758                 o = ls_locate(env, ls, &dti->dti_fid, NULL);
 759                 if (IS_ERR(o))
 760                         return PTR_ERR(o);
 761
 762                 if (!dt_object_exists(o)) {
 763                         dt_object_put_nocache(env, o);
 764                         return 0;
 765                 }
 766                 CDEBUG(D_INFO, "Found old llog lastid file\n");
 767         } else if (rc < 0) {
 768                 return rc;
 769         } else {
 770                 CDEBUG(D_INFO, "Found old lastid file for sequence %#llx\n",
 771                        lastid_seq);
 772                 o = ls_locate(env, ls, &dti->dti_fid, NULL);
 773                 if (IS_ERR(o))
 774                         return PTR_ERR(o);
 775         }
 776         /* let's read seq-NNNNNN-lastid file value */
 777         LASSERT(dt_object_exists(o));
 778         dti->dti_off = 0;
 779         dti->dti_lb.lb_buf = &losd;
 780         dti->dti_lb.lb_len = sizeof(losd);
 781         dt_read_lock(env, o, 0);
 782         rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 783         dt_read_unlock(env, o);
 784         if (rc == 0 && le32_to_cpu(losd.lso_magic) != LOS_MAGIC) {
 785                 CERROR("%s: wrong content of seq-%#llx-lastid file, magic %x\n",
 786                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq,
 787                        le32_to_cpu(losd.lso_magic));
 788                 rc = -EINVAL;
 789         } else if (rc < 0) {
 790                 CERROR("%s: failed to read seq-%#llx-lastid: rc = %d\n",
 791                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq, rc);
 792         }
 793         dt_object_put_nocache(env, o);
 794         if (rc == 0)
 795                 *first_oid = le32_to_cpu(losd.lso_next_oid);
 796         return rc;
 797 }
 798
 799 /**
 800  * Initialize local OID storage for required sequence.
 801  * That may be needed for services that uses local files and requires
 802  * dynamic OID allocation for them.
 803  *
 804  * Per each sequence we have an object with 'first_fid' identificator
 805  * containing the counter for OIDs of locally created files with that
 806  * sequence.
 807  *
 808  * It is used now by llog subsystem and MGS for NID tables
 809  *
 810  * Function gets first_fid to create counter object.
 811  * All dynamic fids will be generated with the same sequence and incremented
 812  * OIDs
 813  *
 814  * Returned local_oid_storage is in-memory representaion of OID storage
 815  */
 816 int local_oid_storage_init(const struct lu_env *env, struct dt_device *dev,
 817                            const struct lu_fid *first_fid,
 818                            struct local_oid_storage **los)
 819 {
 820         struct dt_thread_info   *dti = dt_info(env);
 821         struct ls_device        *ls;
 822         u64                      lastid;
 823         struct dt_object        *o = NULL;
 824         struct thandle          *th;
 825         __u32                    first_oid = fid_oid(first_fid);
 826         int                      rc = 0;
 827
 828         ENTRY;
 829
 830         ls = ls_device_get(dev);
 831         if (IS_ERR(ls))
 832                 RETURN(PTR_ERR(ls));
 833
 834         mutex_lock(&ls->ls_los_mutex);
 835         *los = dt_los_find(ls, fid_seq(first_fid));
 836         if (*los != NULL)
 837                 GOTO(out, rc = 0);
 838
 839         /* not found, then create */
 840         OBD_ALLOC_PTR(*los);
 841         if (*los == NULL)
 842                 GOTO(out, rc = -ENOMEM);
 843
 844         atomic_set(&(*los)->los_refcount, 1);
 845         mutex_init(&(*los)->los_id_lock);
 846         (*los)->los_dev = &ls->ls_top_dev;
 847         atomic_inc(&ls->ls_refcount);
 848         list_add(&(*los)->los_list, &ls->ls_los_list);
 849
 850         /* Use {seq, 0, 0} to create the LAST_ID file for every
 851          * sequence.  OIDs start at LUSTRE_FID_INIT_OID.
 852          */
 853         dti->dti_fid.f_seq = fid_seq(first_fid);
 854         dti->dti_fid.f_oid = LUSTRE_FID_LASTID_OID;
 855         dti->dti_fid.f_ver = 0;
 856         o = ls_locate(env, ls, &dti->dti_fid, NULL);
 857         if (IS_ERR(o))
 858                 GOTO(out_los, rc = PTR_ERR(o));
 859
 860         if (!dt_object_exists(o)) {
 861                 rc = lastid_compat_check(env, dev, fid_seq(first_fid),
 862                                          &first_oid, ls);
 863                 if (rc < 0)
 864                         GOTO(out_los, rc);
 865
 866                 th = dt_trans_create(env, dev);
 867                 if (IS_ERR(th))
 868                         GOTO(out_los, rc = PTR_ERR(th));
 869
 870                 dti->dti_attr.la_valid = LA_MODE | LA_TYPE;
 871                 dti->dti_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
 872                 dti->dti_dof.dof_type = dt_mode_to_dft(S_IFREG);
 873
 874                 rc = dt_declare_create(env, o, &dti->dti_attr, NULL,
 875                                        &dti->dti_dof, th);
 876                 if (rc)
 877                         GOTO(out_trans, rc);
 878
 879                 lastid = cpu_to_le64(first_oid);
 880
 881                 dti->dti_off = 0;
 882                 dti->dti_lb.lb_buf = &lastid;
 883                 dti->dti_lb.lb_len = sizeof(lastid);
 884                 rc = dt_declare_record_write(env, o, &dti->dti_lb, dti->dti_off,
 885                                              th);
 886                 if (rc)
 887                         GOTO(out_trans, rc);
 888
 889                 rc = dt_trans_start_local(env, dev, th);
 890                 if (rc)
 891                         GOTO(out_trans, rc);
 892
 893                 dt_write_lock(env, o, 0);
 894                 if (dt_object_exists(o))
 895                         GOTO(out_lock, rc = 0);
 896
 897                 rc = dt_create(env, o, &dti->dti_attr, NULL, &dti->dti_dof,
 898                                th);
 899                 if (rc)
 900                         GOTO(out_lock, rc);
 901
 902                 rc = dt_record_write(env, o, &dti->dti_lb, &dti->dti_off, th);
 903                 if (rc)
 904                         GOTO(out_lock, rc);
 905 out_lock:
 906                 dt_write_unlock(env, o);
 907 out_trans:
 908                 dt_trans_stop(env, dev, th);
 909         } else {
 910                 dti->dti_off = 0;
 911                 dti->dti_lb.lb_buf = &lastid;
 912                 dti->dti_lb.lb_len = sizeof(lastid);
 913                 dt_read_lock(env, o, 0);
 914                 rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 915                 dt_read_unlock(env, o);
 916                 if (rc == 0 && le64_to_cpu(lastid) > OBIF_MAX_OID) {
 917                         CERROR("%s: bad oid %llu is read from LAST_ID\n",
 918                                o->do_lu.lo_dev->ld_obd->obd_name,
 919                                le64_to_cpu(lastid));
 920                         rc = -EINVAL;
 921                 }
 922         }
 923 out_los:
 924         if (rc != 0) {
 925                 list_del(&(*los)->los_list);
 926                 atomic_dec(&ls->ls_refcount);
 927                 OBD_FREE_PTR(*los);
 928                 *los = NULL;
 929                 if (o != NULL && !IS_ERR(o))
 930                         dt_object_put_nocache(env, o);
 931         } else {
 932                 (*los)->los_seq = fid_seq(first_fid);
 933                 (*los)->los_last_oid = le64_to_cpu(lastid);
 934                 (*los)->los_obj = o;
 935                 /* Read value should not be less than initial one
 936                  * but possible after upgrade from older fs.
 937                  * In this case just switch to the first_oid in memory and
 938                  * it will be updated on disk with first object generated */
 939                 if ((*los)->los_last_oid < first_oid)
 940                         (*los)->los_last_oid = first_oid;
 941         }
 942 out:
 943         mutex_unlock(&ls->ls_los_mutex);
 944         ls_device_put(env, ls);
 945         return rc;
 946 }
 947 EXPORT_SYMBOL(local_oid_storage_init);
 948
 949 void local_oid_storage_fini(const struct lu_env *env,
 950                             struct local_oid_storage *los)
 951 {
 952         struct ls_device *ls;
 953
 954         LASSERT(env);
 955         LASSERT(los->los_dev);
 956         ls = dt2ls_dev(los->los_dev);
 957
 958         /* Take the mutex before decreasing the reference to avoid race
 959          * conditions as described in LU-4721. */
 960         mutex_lock(&ls->ls_los_mutex);
 961         if (!atomic_dec_and_test(&los->los_refcount)) {
 962                 mutex_unlock(&ls->ls_los_mutex);
 963                 return;
 964         }
 965
 966         if (los->los_obj)
 967                 dt_object_put_nocache(env, los->los_obj);
 968         list_del(&los->los_list);
 969         OBD_FREE_PTR(los);
 970         mutex_unlock(&ls->ls_los_mutex);
 971         ls_device_put(env, ls);
 972 }
 973 EXPORT_SYMBOL(local_oid_storage_fini);