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_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, DT_SRC_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);
 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);
 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, DT_SRC_PARENT);
 406         rc = dt_insert(env, parent, (const struct dt_rec *)rec,
 407                        (const struct dt_key *)name, th);
 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                 dt_object_put_nocache(env, dto);
 423                 dto = ERR_PTR(rc);
 424         }
 425         RETURN(dto);
 426 }
 427
 428 struct dt_object *local_file_find(const struct lu_env *env,
 429                                   struct local_oid_storage *los,
 430                                   struct dt_object *parent,
 431                                   const char *name)
 432 {
 433         struct dt_thread_info   *dti = dt_info(env);
 434         struct dt_object        *dto;
 435         int                      rc;
 436
 437         LASSERT(parent);
 438
 439         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 440         if (!rc)
 441                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 442                                 &dti->dti_fid, NULL);
 443         else
 444                 dto = ERR_PTR(rc);
 445
 446         return dto;
 447 }
 448 EXPORT_SYMBOL(local_file_find);
 449
 450 /*
 451  * Look up and create (if it does not exist) a local named file or directory in
 452  * parent directory.
 453  */
 454 struct dt_object *local_file_find_or_create(const struct lu_env *env,
 455                                             struct local_oid_storage *los,
 456                                             struct dt_object *parent,
 457                                             const char *name, __u32 mode)
 458 {
 459         struct dt_thread_info   *dti = dt_info(env);
 460         struct dt_object        *dto;
 461         int                      rc;
 462
 463         dto = local_file_find(env, los, parent, name);
 464         if (!IS_ERR(dto) || PTR_ERR(dto) != -ENOENT)
 465                 return dto;
 466
 467         rc = local_object_fid_generate(env, los, &dti->dti_fid);
 468         if (rc)
 469                 return ERR_PTR(rc);
 470
 471         /* create the object */
 472         dti->dti_attr.la_valid = LA_MODE;
 473         dti->dti_attr.la_mode = mode;
 474         dti->dti_dof.dof_type = dt_mode_to_dft(mode & S_IFMT);
 475         dto = __local_file_create(env, &dti->dti_fid, los,
 476                                   dt2ls_dev(los->los_dev), parent, name,
 477                                   &dti->dti_attr, &dti->dti_dof);
 478         return dto;
 479 }
 480 EXPORT_SYMBOL(local_file_find_or_create);
 481
 482 struct dt_object *local_file_find_or_create_with_fid(const struct lu_env *env,
 483                                                      struct dt_device *dt,
 484                                                      const struct lu_fid *fid,
 485                                                      struct dt_object *parent,
 486                                                      const char *name,
 487                                                      __u32 mode)
 488 {
 489         struct dt_thread_info   *dti = dt_info(env);
 490         struct dt_object        *dto;
 491         int                      rc;
 492
 493         LASSERT(parent);
 494
 495         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 496         if (rc == 0) {
 497                 dto = dt_locate(env, dt, &dti->dti_fid);
 498         } else if (rc != -ENOENT) {
 499                 dto = ERR_PTR(rc);
 500         } else {
 501                 struct ls_device *ls;
 502
 503                 ls = ls_device_get(dt);
 504                 if (IS_ERR(ls)) {
 505                         dto = ERR_PTR(PTR_ERR(ls));
 506                 } else {
 507                         /* create the object */
 508                         dti->dti_attr.la_valid  = LA_MODE;
 509                         dti->dti_attr.la_mode   = mode;
 510                         dti->dti_dof.dof_type   = dt_mode_to_dft(mode & S_IFMT);
 511                         dto = __local_file_create(env, fid, NULL, ls, parent,
 512                                                   name, &dti->dti_attr,
 513                                                   &dti->dti_dof);
 514                         /* ls_device_put() will finalize the ls device, we
 515                          * have to open the object in other device stack */
 516                         if (!IS_ERR(dto)) {
 517                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 518                                 dt_object_put_nocache(env, dto);
 519                                 dto = dt_locate(env, dt, &dti->dti_fid);
 520                         }
 521                         ls_device_put(env, ls);
 522                 }
 523         }
 524         return dto;
 525 }
 526 EXPORT_SYMBOL(local_file_find_or_create_with_fid);
 527
 528 /*
 529  * Look up and create (if it does not exist) a local named index file in parent
 530  * directory.
 531  */
 532 struct dt_object *local_index_find_or_create(const struct lu_env *env,
 533                                              struct local_oid_storage *los,
 534                                              struct dt_object *parent,
 535                                              const char *name, __u32 mode,
 536                                              const struct dt_index_features *ft)
 537 {
 538         struct dt_thread_info   *dti = dt_info(env);
 539         struct dt_object        *dto;
 540         int                      rc;
 541
 542         LASSERT(parent);
 543
 544         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 545         if (rc == 0) {
 546                 /* name is found, get the object */
 547                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 548                                 &dti->dti_fid, NULL);
 549         } else if (rc != -ENOENT) {
 550                 dto = ERR_PTR(rc);
 551         } else {
 552                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 553                 if (rc < 0) {
 554                         dto = ERR_PTR(rc);
 555                 } else {
 556                         /* create the object */
 557                         dti->dti_attr.la_valid          = LA_MODE;
 558                         dti->dti_attr.la_mode           = mode;
 559                         dti->dti_dof.dof_type           = DFT_INDEX;
 560                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 561                         dto = __local_file_create(env, &dti->dti_fid, los,
 562                                                   dt2ls_dev(los->los_dev),
 563                                                   parent, name, &dti->dti_attr,
 564                                                   &dti->dti_dof);
 565                 }
 566         }
 567         return dto;
 568
 569 }
 570 EXPORT_SYMBOL(local_index_find_or_create);
 571
 572 struct dt_object *
 573 local_index_find_or_create_with_fid(const struct lu_env *env,
 574                                     struct dt_device *dt,
 575                                     const struct lu_fid *fid,
 576                                     struct dt_object *parent,
 577                                     const char *name, __u32 mode,
 578                                     const struct dt_index_features *ft)
 579 {
 580         struct dt_thread_info   *dti = dt_info(env);
 581         struct dt_object        *dto;
 582         int                      rc;
 583
 584         LASSERT(parent);
 585
 586         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 587         if (rc == 0) {
 588                 /* name is found, get the object */
 589                 if (!lu_fid_eq(fid, &dti->dti_fid))
 590                         dto = ERR_PTR(-EINVAL);
 591                 else
 592                         dto = dt_locate(env, dt, fid);
 593         } else if (rc != -ENOENT) {
 594                 dto = ERR_PTR(rc);
 595         } else {
 596                 struct ls_device *ls;
 597
 598                 ls = ls_device_get(dt);
 599                 if (IS_ERR(ls)) {
 600                         dto = ERR_PTR(PTR_ERR(ls));
 601                 } else {
 602                         /* create the object */
 603                         dti->dti_attr.la_valid          = LA_MODE;
 604                         dti->dti_attr.la_mode           = mode;
 605                         dti->dti_dof.dof_type           = DFT_INDEX;
 606                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 607                         dto = __local_file_create(env, fid, NULL, ls, parent,
 608                                                   name, &dti->dti_attr,
 609                                                   &dti->dti_dof);
 610                         /* ls_device_put() will finalize the ls device, we
 611                          * have to open the object in other device stack */
 612                         if (!IS_ERR(dto)) {
 613                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 614                                 dt_object_put_nocache(env, dto);
 615                                 dto = dt_locate(env, dt, &dti->dti_fid);
 616                         }
 617                         ls_device_put(env, ls);
 618                 }
 619         }
 620         return dto;
 621 }
 622 EXPORT_SYMBOL(local_index_find_or_create_with_fid);
 623
 624 static int local_object_declare_unlink(const struct lu_env *env,
 625                                        struct dt_device *dt,
 626                                        struct dt_object *p,
 627                                        struct dt_object *c, const char *name,
 628                                        struct thandle *th)
 629 {
 630         int rc;
 631
 632         rc = dt_declare_delete(env, p, (const struct dt_key *)name, th);
 633         if (rc < 0)
 634                 return rc;
 635
 636         rc = dt_declare_ref_del(env, c, th);
 637         if (rc < 0)
 638                 return rc;
 639
 640         return dt_declare_destroy(env, c, th);
 641 }
 642
 643 int local_object_unlink(const struct lu_env *env, struct dt_device *dt,
 644                         struct dt_object *parent, const char *name)
 645 {
 646         struct dt_thread_info   *dti = dt_info(env);
 647         struct dt_object        *dto;
 648         struct thandle          *th;
 649         int                      rc;
 650
 651         ENTRY;
 652
 653         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 654         if (rc == -ENOENT)
 655                 RETURN(0);
 656         else if (rc < 0)
 657                 RETURN(rc);
 658
 659         dto = dt_locate(env, dt, &dti->dti_fid);
 660         if (unlikely(IS_ERR(dto)))
 661                 RETURN(PTR_ERR(dto));
 662
 663         th = dt_trans_create(env, dt);
 664         if (IS_ERR(th))
 665                 GOTO(out, rc = PTR_ERR(th));
 666
 667         rc = local_object_declare_unlink(env, dt, parent, dto, name, th);
 668         if (rc < 0)
 669                 GOTO(stop, rc);
 670
 671         rc = dt_trans_start_local(env, dt, th);
 672         if (rc < 0)
 673                 GOTO(stop, rc);
 674
 675         dt_write_lock(env, dto, 0);
 676         rc = dt_delete(env, parent, (struct dt_key *)name, th);
 677         if (rc < 0)
 678                 GOTO(unlock, rc);
 679
 680         rc = dt_ref_del(env, dto, th);
 681         if (rc < 0) {
 682                 struct dt_insert_rec *rec = &dti->dti_dt_rec;
 683
 684                 rec->rec_fid = &dti->dti_fid;
 685                 rec->rec_type = dto->do_lu.lo_header->loh_attr;
 686                 rc = dt_insert(env, parent, (const struct dt_rec *)rec,
 687                                (const struct dt_key *)name, th);
 688                 GOTO(unlock, rc);
 689         }
 690
 691         rc = dt_destroy(env, dto, th);
 692 unlock:
 693         dt_write_unlock(env, dto);
 694 stop:
 695         dt_trans_stop(env, dt, th);
 696 out:
 697         dt_object_put_nocache(env, dto);
 698         return rc;
 699 }
 700 EXPORT_SYMBOL(local_object_unlink);
 701
 702 struct local_oid_storage *dt_los_find(struct ls_device *ls, __u64 seq)
 703 {
 704         struct local_oid_storage *los, *ret = NULL;
 705
 706         list_for_each_entry(los, &ls->ls_los_list, los_list) {
 707                 if (los->los_seq == seq) {
 708                         atomic_inc(&los->los_refcount);
 709                         ret = los;
 710                         break;
 711                 }
 712         }
 713         return ret;
 714 }
 715
 716 void dt_los_put(struct local_oid_storage *los)
 717 {
 718         if (atomic_dec_and_test(&los->los_refcount))
 719                 /* should never happen, only local_oid_storage_fini should
 720                  * drop refcount to zero */
 721                 LBUG();
 722         return;
 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);