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, 2015, 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 = attr->la_mode & S_IFMT;
 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         if (dti->dti_dof.dof_type == DFT_DIR) {
 349                 if (!dt_try_as_dir(env, dto))
 350                         GOTO(trans_stop, rc = -ENOTDIR);
 351
 352                 rec->rec_type = S_IFDIR;
 353                 rec->rec_fid = fid;
 354                 rc = dt_declare_insert(env, dto, (const struct dt_rec *)rec,
 355                                 (const struct dt_key *)".", th);
 356                 if (rc != 0)
 357                         GOTO(trans_stop, rc);
 358
 359                 rec->rec_fid = lu_object_fid(&parent->do_lu);
 360                 rc = dt_declare_insert(env, dto, (const struct dt_rec *)rec,
 361                                 (const struct dt_key *)"..", th);
 362                 if (rc != 0)
 363                         GOTO(trans_stop, rc);
 364
 365                 rc = dt_declare_ref_add(env, dto, th);
 366                 if (rc != 0)
 367                         GOTO(trans_stop, rc);
 368         }
 369
 370         rc = dt_trans_start_local(env, ls->ls_osd, th);
 371         if (rc)
 372                 GOTO(trans_stop, rc);
 373
 374         dt_write_lock(env, dto, LOS_CHILD);
 375         if (dt_object_exists(dto))
 376                 GOTO(unlock, rc = 0);
 377
 378         CDEBUG(D_OTHER, "create new object "DFID"\n",
 379                PFID(lu_object_fid(&dto->do_lu)));
 380         rc = local_object_create(env, los, dto, attr, dof, th);
 381         if (rc)
 382                 GOTO(unlock, rc);
 383         LASSERT(dt_object_exists(dto));
 384
 385         if (dti->dti_dof.dof_type == DFT_DIR) {
 386
 387                 rec->rec_type = S_IFDIR;
 388                 rec->rec_fid = fid;
 389                 /* Add "." and ".." for newly created dir */
 390                 rc = dt_insert(env, dto, (const struct dt_rec *)rec,
 391                                (const struct dt_key *)".", th, 1);
 392                 if (rc != 0)
 393                         GOTO(destroy, rc);
 394
 395                 dt_ref_add(env, dto, th);
 396                 rec->rec_fid = lu_object_fid(&parent->do_lu);
 397                 rc = dt_insert(env, dto, (const struct dt_rec *)rec,
 398                                (const struct dt_key *)"..", th, 1);
 399                 if (rc != 0)
 400                         GOTO(destroy, rc);
 401         }
 402
 403         rec->rec_fid = fid;
 404         rec->rec_type = dto->do_lu.lo_header->loh_attr;
 405         dt_write_lock(env, parent, LOS_PARENT);
 406         rc = dt_insert(env, parent, (const struct dt_rec *)rec,
 407                        (const struct dt_key *)name, th, 1);
 408         if (dti->dti_dof.dof_type == DFT_DIR)
 409                 dt_ref_add(env, parent, th);
 410         dt_write_unlock(env, parent);
 411         if (rc)
 412                 GOTO(destroy, rc);
 413 destroy:
 414         if (rc)
 415                 dt_destroy(env, dto, th);
 416 unlock:
 417         dt_write_unlock(env, dto);
 418 trans_stop:
 419         dt_trans_stop(env, ls->ls_osd, th);
 420 out:
 421         if (rc) {
 422                 lu_object_put_nocache(env, &dto->do_lu);
 423                 dto = ERR_PTR(rc);
 424         }
 425         RETURN(dto);
 426 }
 427
 428 /*
 429  * Look up and create (if it does not exist) a local named file or directory in
 430  * parent directory.
 431  */
 432 struct dt_object *local_file_find_or_create(const struct lu_env *env,
 433                                             struct local_oid_storage *los,
 434                                             struct dt_object *parent,
 435                                             const char *name, __u32 mode)
 436 {
 437         struct dt_thread_info   *dti = dt_info(env);
 438         struct dt_object        *dto;
 439         int                      rc;
 440
 441         LASSERT(parent);
 442
 443         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 444         if (rc == 0)
 445                 /* name is found, get the object */
 446                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 447                                 &dti->dti_fid, NULL);
 448         else if (rc != -ENOENT)
 449                 dto = ERR_PTR(rc);
 450         else {
 451                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 452                 if (rc < 0) {
 453                         dto = ERR_PTR(rc);
 454                 } else {
 455                         /* create the object */
 456                         dti->dti_attr.la_valid  = LA_MODE;
 457                         dti->dti_attr.la_mode   = mode;
 458                         dti->dti_dof.dof_type   = dt_mode_to_dft(mode & S_IFMT);
 459                         dto = __local_file_create(env, &dti->dti_fid, los,
 460                                                   dt2ls_dev(los->los_dev),
 461                                                   parent, name, &dti->dti_attr,
 462                                                   &dti->dti_dof);
 463                 }
 464         }
 465         return dto;
 466 }
 467 EXPORT_SYMBOL(local_file_find_or_create);
 468
 469 struct dt_object *local_file_find_or_create_with_fid(const struct lu_env *env,
 470                                                      struct dt_device *dt,
 471                                                      const struct lu_fid *fid,
 472                                                      struct dt_object *parent,
 473                                                      const char *name,
 474                                                      __u32 mode)
 475 {
 476         struct dt_thread_info   *dti = dt_info(env);
 477         struct dt_object        *dto;
 478         int                      rc;
 479
 480         LASSERT(parent);
 481
 482         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 483         if (rc == 0) {
 484                 dto = dt_locate(env, dt, &dti->dti_fid);
 485         } else if (rc != -ENOENT) {
 486                 dto = ERR_PTR(rc);
 487         } else {
 488                 struct ls_device *ls;
 489
 490                 ls = ls_device_get(dt);
 491                 if (IS_ERR(ls)) {
 492                         dto = ERR_PTR(PTR_ERR(ls));
 493                 } else {
 494                         /* create the object */
 495                         dti->dti_attr.la_valid  = LA_MODE;
 496                         dti->dti_attr.la_mode   = mode;
 497                         dti->dti_dof.dof_type   = dt_mode_to_dft(mode & S_IFMT);
 498                         dto = __local_file_create(env, fid, NULL, ls, parent,
 499                                                   name, &dti->dti_attr,
 500                                                   &dti->dti_dof);
 501                         /* ls_device_put() will finalize the ls device, we
 502                          * have to open the object in other device stack */
 503                         if (!IS_ERR(dto)) {
 504                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 505                                 lu_object_put_nocache(env, &dto->do_lu);
 506                                 dto = dt_locate(env, dt, &dti->dti_fid);
 507                         }
 508                         ls_device_put(env, ls);
 509                 }
 510         }
 511         return dto;
 512 }
 513 EXPORT_SYMBOL(local_file_find_or_create_with_fid);
 514
 515 /*
 516  * Look up and create (if it does not exist) a local named index file in parent
 517  * directory.
 518  */
 519 struct dt_object *local_index_find_or_create(const struct lu_env *env,
 520                                              struct local_oid_storage *los,
 521                                              struct dt_object *parent,
 522                                              const char *name, __u32 mode,
 523                                              const struct dt_index_features *ft)
 524 {
 525         struct dt_thread_info   *dti = dt_info(env);
 526         struct dt_object        *dto;
 527         int                      rc;
 528
 529         LASSERT(parent);
 530
 531         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 532         if (rc == 0) {
 533                 /* name is found, get the object */
 534                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 535                                 &dti->dti_fid, NULL);
 536         } else if (rc != -ENOENT) {
 537                 dto = ERR_PTR(rc);
 538         } else {
 539                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 540                 if (rc < 0) {
 541                         dto = ERR_PTR(rc);
 542                 } else {
 543                         /* create the object */
 544                         dti->dti_attr.la_valid          = LA_MODE;
 545                         dti->dti_attr.la_mode           = mode;
 546                         dti->dti_dof.dof_type           = DFT_INDEX;
 547                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 548                         dto = __local_file_create(env, &dti->dti_fid, los,
 549                                                   dt2ls_dev(los->los_dev),
 550                                                   parent, name, &dti->dti_attr,
 551                                                   &dti->dti_dof);
 552                 }
 553         }
 554         return dto;
 555
 556 }
 557 EXPORT_SYMBOL(local_index_find_or_create);
 558
 559 struct dt_object *
 560 local_index_find_or_create_with_fid(const struct lu_env *env,
 561                                     struct dt_device *dt,
 562                                     const struct lu_fid *fid,
 563                                     struct dt_object *parent,
 564                                     const char *name, __u32 mode,
 565                                     const struct dt_index_features *ft)
 566 {
 567         struct dt_thread_info   *dti = dt_info(env);
 568         struct dt_object        *dto;
 569         int                      rc;
 570
 571         LASSERT(parent);
 572
 573         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 574         if (rc == 0) {
 575                 /* name is found, get the object */
 576                 if (!lu_fid_eq(fid, &dti->dti_fid))
 577                         dto = ERR_PTR(-EINVAL);
 578                 else
 579                         dto = dt_locate(env, dt, fid);
 580         } else if (rc != -ENOENT) {
 581                 dto = ERR_PTR(rc);
 582         } else {
 583                 struct ls_device *ls;
 584
 585                 ls = ls_device_get(dt);
 586                 if (IS_ERR(ls)) {
 587                         dto = ERR_PTR(PTR_ERR(ls));
 588                 } else {
 589                         /* create the object */
 590                         dti->dti_attr.la_valid          = LA_MODE;
 591                         dti->dti_attr.la_mode           = mode;
 592                         dti->dti_dof.dof_type           = DFT_INDEX;
 593                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 594                         dto = __local_file_create(env, fid, NULL, ls, parent,
 595                                                   name, &dti->dti_attr,
 596                                                   &dti->dti_dof);
 597                         /* ls_device_put() will finalize the ls device, we
 598                          * have to open the object in other device stack */
 599                         if (!IS_ERR(dto)) {
 600                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 601                                 lu_object_put_nocache(env, &dto->do_lu);
 602                                 dto = dt_locate(env, dt, &dti->dti_fid);
 603                         }
 604                         ls_device_put(env, ls);
 605                 }
 606         }
 607         return dto;
 608 }
 609 EXPORT_SYMBOL(local_index_find_or_create_with_fid);
 610
 611 static int local_object_declare_unlink(const struct lu_env *env,
 612                                        struct dt_device *dt,
 613                                        struct dt_object *p,
 614                                        struct dt_object *c, const char *name,
 615                                        struct thandle *th)
 616 {
 617         int rc;
 618
 619         rc = dt_declare_delete(env, p, (const struct dt_key *)name, th);
 620         if (rc < 0)
 621                 return rc;
 622
 623         rc = dt_declare_ref_del(env, c, th);
 624         if (rc < 0)
 625                 return rc;
 626
 627         return dt_declare_destroy(env, c, th);
 628 }
 629
 630 int local_object_unlink(const struct lu_env *env, struct dt_device *dt,
 631                         struct dt_object *parent, const char *name)
 632 {
 633         struct dt_thread_info   *dti = dt_info(env);
 634         struct dt_object        *dto;
 635         struct thandle          *th;
 636         int                      rc;
 637
 638         ENTRY;
 639
 640         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 641         if (rc == -ENOENT)
 642                 RETURN(0);
 643         else if (rc < 0)
 644                 RETURN(rc);
 645
 646         dto = dt_locate(env, dt, &dti->dti_fid);
 647         if (unlikely(IS_ERR(dto)))
 648                 RETURN(PTR_ERR(dto));
 649
 650         th = dt_trans_create(env, dt);
 651         if (IS_ERR(th))
 652                 GOTO(out, rc = PTR_ERR(th));
 653
 654         rc = local_object_declare_unlink(env, dt, parent, dto, name, th);
 655         if (rc < 0)
 656                 GOTO(stop, rc);
 657
 658         rc = dt_trans_start_local(env, dt, th);
 659         if (rc < 0)
 660                 GOTO(stop, rc);
 661
 662         dt_write_lock(env, dto, 0);
 663         rc = dt_delete(env, parent, (struct dt_key *)name, th);
 664         if (rc < 0)
 665                 GOTO(unlock, rc);
 666
 667         rc = dt_ref_del(env, dto, th);
 668         if (rc < 0) {
 669                 struct dt_insert_rec *rec = &dti->dti_dt_rec;
 670
 671                 rec->rec_fid = &dti->dti_fid;
 672                 rec->rec_type = dto->do_lu.lo_header->loh_attr;
 673                 rc = dt_insert(env, parent, (const struct dt_rec *)rec,
 674                                (const struct dt_key *)name, th, 1);
 675                 GOTO(unlock, rc);
 676         }
 677
 678         rc = dt_destroy(env, dto, th);
 679 unlock:
 680         dt_write_unlock(env, dto);
 681 stop:
 682         dt_trans_stop(env, dt, th);
 683 out:
 684         lu_object_put_nocache(env, &dto->do_lu);
 685         return rc;
 686 }
 687 EXPORT_SYMBOL(local_object_unlink);
 688
 689 struct local_oid_storage *dt_los_find(struct ls_device *ls, __u64 seq)
 690 {
 691         struct local_oid_storage *los, *ret = NULL;
 692
 693         list_for_each_entry(los, &ls->ls_los_list, los_list) {
 694                 if (los->los_seq == seq) {
 695                         atomic_inc(&los->los_refcount);
 696                         ret = los;
 697                         break;
 698                 }
 699         }
 700         return ret;
 701 }
 702
 703 void dt_los_put(struct local_oid_storage *los)
 704 {
 705         if (atomic_dec_and_test(&los->los_refcount))
 706                 /* should never happen, only local_oid_storage_fini should
 707                  * drop refcount to zero */
 708                 LBUG();
 709         return;
 710 }
 711
 712 /* after Lustre 2.3 release there may be old file to store last generated FID
 713  * If such file exists then we have to read its content
 714  */
 715 static int lastid_compat_check(const struct lu_env *env, struct dt_device *dev,
 716                                __u64 lastid_seq, __u32 *first_oid,
 717                                struct ls_device *ls)
 718 {
 719         struct dt_thread_info   *dti = dt_info(env);
 720         struct dt_object        *root = NULL;
 721         struct los_ondisk        losd;
 722         struct dt_object        *o = NULL;
 723         int                      rc = 0;
 724
 725         rc = dt_root_get(env, dev, &dti->dti_fid);
 726         if (rc)
 727                 return rc;
 728
 729         root = ls_locate(env, ls, &dti->dti_fid, NULL);
 730         if (IS_ERR(root))
 731                 return PTR_ERR(root);
 732
 733         /* find old last_id file */
 734         snprintf(dti->dti_buf, sizeof(dti->dti_buf), "seq-"LPX64"-lastid",
 735                  lastid_seq);
 736         rc = dt_lookup_dir(env, root, dti->dti_buf, &dti->dti_fid);
 737         lu_object_put_nocache(env, &root->do_lu);
 738         if (rc == -ENOENT) {
 739                 /* old llog lastid accessed by FID only */
 740                 if (lastid_seq != FID_SEQ_LLOG)
 741                         return 0;
 742                 dti->dti_fid.f_seq = FID_SEQ_LLOG;
 743                 dti->dti_fid.f_oid = 1;
 744                 dti->dti_fid.f_ver = 0;
 745                 o = ls_locate(env, ls, &dti->dti_fid, NULL);
 746                 if (IS_ERR(o))
 747                         return PTR_ERR(o);
 748
 749                 if (!dt_object_exists(o)) {
 750                         lu_object_put_nocache(env, &o->do_lu);
 751                         return 0;
 752                 }
 753                 CDEBUG(D_INFO, "Found old llog lastid file\n");
 754         } else if (rc < 0) {
 755                 return rc;
 756         } else {
 757                 CDEBUG(D_INFO, "Found old lastid file for sequence "LPX64"\n",
 758                        lastid_seq);
 759                 o = ls_locate(env, ls, &dti->dti_fid, NULL);
 760                 if (IS_ERR(o))
 761                         return PTR_ERR(o);
 762         }
 763         /* let's read seq-NNNNNN-lastid file value */
 764         LASSERT(dt_object_exists(o));
 765         dti->dti_off = 0;
 766         dti->dti_lb.lb_buf = &losd;
 767         dti->dti_lb.lb_len = sizeof(losd);
 768         dt_read_lock(env, o, 0);
 769         rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 770         dt_read_unlock(env, o);
 771         if (rc == 0 && le32_to_cpu(losd.lso_magic) != LOS_MAGIC) {
 772                 CERROR("%s: wrong content of seq-"LPX64"-lastid file, magic %x\n",
 773                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq,
 774                        le32_to_cpu(losd.lso_magic));
 775                 rc = -EINVAL;
 776         } else if (rc < 0) {
 777                 CERROR("%s: failed to read seq-"LPX64"-lastid: rc = %d\n",
 778                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq, rc);
 779         }
 780         lu_object_put_nocache(env, &o->do_lu);
 781         if (rc == 0)
 782                 *first_oid = le32_to_cpu(losd.lso_next_oid);
 783         return rc;
 784 }
 785
 786 /**
 787  * Initialize local OID storage for required sequence.
 788  * That may be needed for services that uses local files and requires
 789  * dynamic OID allocation for them.
 790  *
 791  * Per each sequence we have an object with 'first_fid' identificator
 792  * containing the counter for OIDs of locally created files with that
 793  * sequence.
 794  *
 795  * It is used now by llog subsystem and MGS for NID tables
 796  *
 797  * Function gets first_fid to create counter object.
 798  * All dynamic fids will be generated with the same sequence and incremented
 799  * OIDs
 800  *
 801  * Returned local_oid_storage is in-memory representaion of OID storage
 802  */
 803 int local_oid_storage_init(const struct lu_env *env, struct dt_device *dev,
 804                            const struct lu_fid *first_fid,
 805                            struct local_oid_storage **los)
 806 {
 807         struct dt_thread_info   *dti = dt_info(env);
 808         struct ls_device        *ls;
 809         u64                      lastid;
 810         struct dt_object        *o = NULL;
 811         struct thandle          *th;
 812         __u32                    first_oid = fid_oid(first_fid);
 813         int                      rc = 0;
 814
 815         ENTRY;
 816
 817         ls = ls_device_get(dev);
 818         if (IS_ERR(ls))
 819                 RETURN(PTR_ERR(ls));
 820
 821         mutex_lock(&ls->ls_los_mutex);
 822         *los = dt_los_find(ls, fid_seq(first_fid));
 823         if (*los != NULL)
 824                 GOTO(out, rc = 0);
 825
 826         /* not found, then create */
 827         OBD_ALLOC_PTR(*los);
 828         if (*los == NULL)
 829                 GOTO(out, rc = -ENOMEM);
 830
 831         atomic_set(&(*los)->los_refcount, 1);
 832         mutex_init(&(*los)->los_id_lock);
 833         (*los)->los_dev = &ls->ls_top_dev;
 834         atomic_inc(&ls->ls_refcount);
 835         list_add(&(*los)->los_list, &ls->ls_los_list);
 836
 837         /* Use {seq, 0, 0} to create the LAST_ID file for every
 838          * sequence.  OIDs start at LUSTRE_FID_INIT_OID.
 839          */
 840         dti->dti_fid.f_seq = fid_seq(first_fid);
 841         dti->dti_fid.f_oid = LUSTRE_FID_LASTID_OID;
 842         dti->dti_fid.f_ver = 0;
 843         o = ls_locate(env, ls, &dti->dti_fid, NULL);
 844         if (IS_ERR(o))
 845                 GOTO(out_los, rc = PTR_ERR(o));
 846
 847         if (!dt_object_exists(o)) {
 848                 rc = lastid_compat_check(env, dev, fid_seq(first_fid),
 849                                          &first_oid, ls);
 850                 if (rc < 0)
 851                         GOTO(out_los, rc);
 852
 853                 th = dt_trans_create(env, dev);
 854                 if (IS_ERR(th))
 855                         GOTO(out_los, rc = PTR_ERR(th));
 856
 857                 dti->dti_attr.la_valid = LA_MODE | LA_TYPE;
 858                 dti->dti_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
 859                 dti->dti_dof.dof_type = dt_mode_to_dft(S_IFREG);
 860
 861                 rc = dt_declare_create(env, o, &dti->dti_attr, NULL,
 862                                        &dti->dti_dof, th);
 863                 if (rc)
 864                         GOTO(out_trans, rc);
 865
 866                 lastid = cpu_to_le64(first_oid);
 867
 868                 dti->dti_off = 0;
 869                 dti->dti_lb.lb_buf = &lastid;
 870                 dti->dti_lb.lb_len = sizeof(lastid);
 871                 rc = dt_declare_record_write(env, o, &dti->dti_lb, dti->dti_off,
 872                                              th);
 873                 if (rc)
 874                         GOTO(out_trans, rc);
 875
 876                 rc = dt_trans_start_local(env, dev, th);
 877                 if (rc)
 878                         GOTO(out_trans, rc);
 879
 880                 dt_write_lock(env, o, 0);
 881                 if (dt_object_exists(o))
 882                         GOTO(out_lock, rc = 0);
 883
 884                 rc = dt_create(env, o, &dti->dti_attr, NULL, &dti->dti_dof,
 885                                th);
 886                 if (rc)
 887                         GOTO(out_lock, rc);
 888
 889                 rc = dt_record_write(env, o, &dti->dti_lb, &dti->dti_off, th);
 890                 if (rc)
 891                         GOTO(out_lock, rc);
 892 out_lock:
 893                 dt_write_unlock(env, o);
 894 out_trans:
 895                 dt_trans_stop(env, dev, th);
 896         } else {
 897                 dti->dti_off = 0;
 898                 dti->dti_lb.lb_buf = &lastid;
 899                 dti->dti_lb.lb_len = sizeof(lastid);
 900                 dt_read_lock(env, o, 0);
 901                 rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 902                 dt_read_unlock(env, o);
 903                 if (rc == 0 && le64_to_cpu(lastid) > OBIF_MAX_OID) {
 904                         CERROR("%s: bad oid "LPU64" is read from LAST_ID\n",
 905                                o->do_lu.lo_dev->ld_obd->obd_name,
 906                                le64_to_cpu(lastid));
 907                         rc = -EINVAL;
 908                 }
 909         }
 910 out_los:
 911         if (rc != 0) {
 912                 list_del(&(*los)->los_list);
 913                 atomic_dec(&ls->ls_refcount);
 914                 OBD_FREE_PTR(*los);
 915                 *los = NULL;
 916                 if (o != NULL && !IS_ERR(o))
 917                         lu_object_put_nocache(env, &o->do_lu);
 918         } else {
 919                 (*los)->los_seq = fid_seq(first_fid);
 920                 (*los)->los_last_oid = le64_to_cpu(lastid);
 921                 (*los)->los_obj = o;
 922                 /* Read value should not be less than initial one
 923                  * but possible after upgrade from older fs.
 924                  * In this case just switch to the first_oid in memory and
 925                  * it will be updated on disk with first object generated */
 926                 if ((*los)->los_last_oid < first_oid)
 927                         (*los)->los_last_oid = first_oid;
 928         }
 929 out:
 930         mutex_unlock(&ls->ls_los_mutex);
 931         ls_device_put(env, ls);
 932         return rc;
 933 }
 934 EXPORT_SYMBOL(local_oid_storage_init);
 935
 936 void local_oid_storage_fini(const struct lu_env *env,
 937                             struct local_oid_storage *los)
 938 {
 939         struct ls_device *ls;
 940
 941         LASSERT(env);
 942         LASSERT(los->los_dev);
 943         ls = dt2ls_dev(los->los_dev);
 944
 945         /* Take the mutex before decreasing the reference to avoid race
 946          * conditions as described in LU-4721. */
 947         mutex_lock(&ls->ls_los_mutex);
 948         if (!atomic_dec_and_test(&los->los_refcount)) {
 949                 mutex_unlock(&ls->ls_los_mutex);
 950                 return;
 951         }
 952
 953         if (los->los_obj)
 954                 lu_object_put_nocache(env, &los->los_obj->do_lu);
 955         list_del(&los->los_list);
 956         OBD_FREE_PTR(los);
 957         mutex_unlock(&ls->ls_los_mutex);
 958         ls_device_put(env, ls);
 959 }
 960 EXPORT_SYMBOL(local_oid_storage_fini);