lustre/obdclass/local_storage.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License version 2 for more details.  A copy is
  14  * included in the COPYING file that accompanied this code.
  15
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2012, 2013, Intel Corporation.
  24  */
  25 /*
  26  * lustre/obdclass/local_storage.c
  27  *
  28  * Local storage for file/objects with fid generation. Works on top of OSD.
  29  *
  30  * Author: Mikhail Pershin <mike.pershin@intel.com>
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_CLASS
  34
  35 #include "local_storage.h"
  36
  37 /* all initialized local storages on this node are linked on this */
  38 static CFS_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 struct lu_object_operations ls_lu_obj_ops = {
  72         .loo_object_init  = ls_object_init,
  73         .loo_object_free  = ls_object_free,
  74 };
  75
  76 struct lu_object *ls_object_alloc(const struct lu_env *env,
  77                                   const struct lu_object_header *_h,
  78                                   struct lu_device *d)
  79 {
  80         struct lu_object_header *h;
  81         struct ls_object        *o;
  82         struct lu_object        *l;
  83
  84         LASSERT(_h == NULL);
  85
  86         OBD_ALLOC_PTR(o);
  87         if (o != NULL) {
  88                 l = &o->ls_obj.do_lu;
  89                 h = &o->ls_header;
  90
  91                 lu_object_header_init(h);
  92                 dt_object_init(&o->ls_obj, h, d);
  93                 lu_object_add_top(h, l);
  94
  95                 l->lo_ops = &ls_lu_obj_ops;
  96
  97                 return l;
  98         } else {
  99                 return NULL;
 100         }
 101 }
 102
 103 static struct lu_device_operations ls_lu_dev_ops = {
 104         .ldo_object_alloc =     ls_object_alloc
 105 };
 106
 107 static struct ls_device *__ls_find_dev(struct dt_device *dev)
 108 {
 109         struct ls_device *ls, *ret = NULL;
 110
 111         cfs_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         CFS_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         cfs_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(cfs_list_empty(&ls->ls_los_list));
 185                 cfs_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                 rc = dt_declare_record_write(env, los->los_obj,
 233                                              sizeof(struct los_ondisk), 0, th);
 234                 if (rc)
 235                         RETURN(rc);
 236         }
 237
 238         rc = dt_declare_create(env, o, attr, NULL, dof, th);
 239         if (rc)
 240                 RETURN(rc);
 241
 242         dti->dti_lb.lb_buf = NULL;
 243         dti->dti_lb.lb_len = sizeof(dti->dti_lma);
 244         rc = dt_declare_xattr_set(env, o, &dti->dti_lb, XATTR_NAME_LMA, 0, th);
 245
 246         RETURN(rc);
 247 }
 248
 249 int local_object_create(const struct lu_env *env,
 250                         struct local_oid_storage *los,
 251                         struct dt_object *o, struct lu_attr *attr,
 252                         struct dt_object_format *dof, struct thandle *th)
 253 {
 254         struct dt_thread_info   *dti = dt_info(env);
 255         obd_id                   lastid;
 256         int                      rc;
 257
 258         ENTRY;
 259
 260         rc = dt_create(env, o, attr, NULL, dof, th);
 261         if (rc)
 262                 RETURN(rc);
 263
 264         if (los == NULL)
 265                 RETURN(rc);
 266
 267         LASSERT(los->los_obj);
 268         LASSERT(dt_object_exists(los->los_obj));
 269
 270         /* many threads can be updated this, serialize
 271          * them here to avoid the race where one thread
 272          * takes the value first, but writes it last */
 273         mutex_lock(&los->los_id_lock);
 274
 275         /* update local oid number on disk so that
 276          * we know the last one used after reboot */
 277         lastid = cpu_to_le64(los->los_last_oid);
 278
 279         dti->dti_off = 0;
 280         dti->dti_lb.lb_buf = &lastid;
 281         dti->dti_lb.lb_len = sizeof(lastid);
 282         rc = dt_record_write(env, los->los_obj, &dti->dti_lb, &dti->dti_off,
 283                              th);
 284         mutex_unlock(&los->los_id_lock);
 285
 286         RETURN(rc);
 287 }
 288
 289 /*
 290  * Create local named object (file, directory or index) in parent directory.
 291  */
 292 struct dt_object *__local_file_create(const struct lu_env *env,
 293                                       const struct lu_fid *fid,
 294                                       struct local_oid_storage *los,
 295                                       struct ls_device *ls,
 296                                       struct dt_object *parent,
 297                                       const char *name, struct lu_attr *attr,
 298                                       struct dt_object_format *dof)
 299 {
 300         struct dt_thread_info   *dti    = dt_info(env);
 301         struct lu_object_conf   *conf   = &dti->dti_conf;
 302         struct dt_object        *dto;
 303         struct thandle          *th;
 304         int                      rc;
 305
 306         /* We know that the target object does not exist, to be created,
 307          * then give some hints - LOC_F_NEW to help low layer to handle
 308          * that efficiently and properly. */
 309         memset(conf, 0, sizeof(*conf));
 310         conf->loc_flags = LOC_F_NEW;
 311         dto = ls_locate(env, ls, fid, conf);
 312         if (unlikely(IS_ERR(dto)))
 313                 RETURN(dto);
 314
 315         LASSERT(dto != NULL);
 316         if (dt_object_exists(dto))
 317                 GOTO(out, rc = -EEXIST);
 318
 319         th = dt_trans_create(env, ls->ls_osd);
 320         if (IS_ERR(th))
 321                 GOTO(out, rc = PTR_ERR(th));
 322
 323         rc = local_object_declare_create(env, los, dto, attr, dof, th);
 324         if (rc)
 325                 GOTO(trans_stop, rc);
 326
 327         if (dti->dti_dof.dof_type == DFT_DIR) {
 328                 dt_declare_ref_add(env, dto, th);
 329                 dt_declare_ref_add(env, parent, th);
 330         }
 331
 332         rc = dt_declare_insert(env, parent, (void *)fid, (void *)name, th);
 333         if (rc)
 334                 GOTO(trans_stop, rc);
 335
 336         rc = dt_trans_start_local(env, ls->ls_osd, th);
 337         if (rc)
 338                 GOTO(trans_stop, rc);
 339
 340         dt_write_lock(env, dto, 0);
 341         if (dt_object_exists(dto))
 342                 GOTO(unlock, rc = 0);
 343
 344         CDEBUG(D_OTHER, "create new object "DFID"\n",
 345                PFID(lu_object_fid(&dto->do_lu)));
 346         rc = local_object_create(env, los, dto, attr, dof, th);
 347         if (rc)
 348                 GOTO(unlock, rc);
 349         LASSERT(dt_object_exists(dto));
 350
 351         if (dti->dti_dof.dof_type == DFT_DIR) {
 352                 if (!dt_try_as_dir(env, dto))
 353                         GOTO(destroy, rc = -ENOTDIR);
 354                 /* Add "." and ".." for newly created dir */
 355                 rc = dt_insert(env, dto, (void *)fid, (void *)".", th,
 356                                BYPASS_CAPA, 1);
 357                 if (rc)
 358                         GOTO(destroy, rc);
 359                 dt_ref_add(env, dto, th);
 360                 rc = dt_insert(env, dto, (void *)lu_object_fid(&parent->do_lu),
 361                                (void *)"..", th, BYPASS_CAPA, 1);
 362                 if (rc)
 363                         GOTO(destroy, rc);
 364         }
 365
 366         dt_write_lock(env, parent, 0);
 367         rc = dt_insert(env, parent, (const struct dt_rec *)fid,
 368                        (const struct dt_key *)name, th, BYPASS_CAPA, 1);
 369         if (dti->dti_dof.dof_type == DFT_DIR)
 370                 dt_ref_add(env, parent, th);
 371         dt_write_unlock(env, parent);
 372         if (rc)
 373                 GOTO(destroy, rc);
 374 destroy:
 375         if (rc)
 376                 dt_destroy(env, dto, th);
 377 unlock:
 378         dt_write_unlock(env, dto);
 379 trans_stop:
 380         dt_trans_stop(env, ls->ls_osd, th);
 381 out:
 382         if (rc) {
 383                 lu_object_put_nocache(env, &dto->do_lu);
 384                 dto = ERR_PTR(rc);
 385         }
 386         RETURN(dto);
 387 }
 388
 389 /*
 390  * Look up and create (if it does not exist) a local named file or directory in
 391  * parent directory.
 392  */
 393 struct dt_object *local_file_find_or_create(const struct lu_env *env,
 394                                             struct local_oid_storage *los,
 395                                             struct dt_object *parent,
 396                                             const char *name, __u32 mode)
 397 {
 398         struct dt_thread_info   *dti = dt_info(env);
 399         struct dt_object        *dto;
 400         int                      rc;
 401
 402         LASSERT(parent);
 403
 404         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 405         if (rc == 0)
 406                 /* name is found, get the object */
 407                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 408                                 &dti->dti_fid, NULL);
 409         else if (rc != -ENOENT)
 410                 dto = ERR_PTR(rc);
 411         else {
 412                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 413                 if (rc < 0) {
 414                         dto = ERR_PTR(rc);
 415                 } else {
 416                         /* create the object */
 417                         dti->dti_attr.la_valid  = LA_MODE;
 418                         dti->dti_attr.la_mode   = mode;
 419                         dti->dti_dof.dof_type   = dt_mode_to_dft(mode & S_IFMT);
 420                         dto = __local_file_create(env, &dti->dti_fid, los,
 421                                                   dt2ls_dev(los->los_dev),
 422                                                   parent, name, &dti->dti_attr,
 423                                                   &dti->dti_dof);
 424                 }
 425         }
 426         return dto;
 427 }
 428 EXPORT_SYMBOL(local_file_find_or_create);
 429
 430 struct dt_object *local_file_find_or_create_with_fid(const struct lu_env *env,
 431                                                      struct dt_device *dt,
 432                                                      const struct lu_fid *fid,
 433                                                      struct dt_object *parent,
 434                                                      const char *name,
 435                                                      __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                 dto = dt_locate(env, dt, &dti->dti_fid);
 446         } else if (rc != -ENOENT) {
 447                 dto = ERR_PTR(rc);
 448         } else {
 449                 struct ls_device *ls;
 450
 451                 ls = ls_device_get(dt);
 452                 if (IS_ERR(ls)) {
 453                         dto = ERR_PTR(PTR_ERR(ls));
 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, fid, NULL, ls, parent,
 460                                                   name, &dti->dti_attr,
 461                                                   &dti->dti_dof);
 462                         /* ls_device_put() will finalize the ls device, we
 463                          * have to open the object in other device stack */
 464                         if (!IS_ERR(dto)) {
 465                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 466                                 lu_object_put_nocache(env, &dto->do_lu);
 467                                 dto = dt_locate(env, dt, &dti->dti_fid);
 468                         }
 469                         ls_device_put(env, ls);
 470                 }
 471         }
 472         return dto;
 473 }
 474 EXPORT_SYMBOL(local_file_find_or_create_with_fid);
 475
 476 /*
 477  * Look up and create (if it does not exist) a local named index file in parent
 478  * directory.
 479  */
 480 struct dt_object *local_index_find_or_create(const struct lu_env *env,
 481                                              struct local_oid_storage *los,
 482                                              struct dt_object *parent,
 483                                              const char *name, __u32 mode,
 484                                              const struct dt_index_features *ft)
 485 {
 486         struct dt_thread_info   *dti = dt_info(env);
 487         struct dt_object        *dto;
 488         int                      rc;
 489
 490         LASSERT(parent);
 491
 492         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 493         if (rc == 0) {
 494                 /* name is found, get the object */
 495                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 496                                 &dti->dti_fid, NULL);
 497         } else if (rc != -ENOENT) {
 498                 dto = ERR_PTR(rc);
 499         } else {
 500                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 501                 if (rc < 0) {
 502                         dto = ERR_PTR(rc);
 503                 } else {
 504                         /* create the object */
 505                         dti->dti_attr.la_valid          = LA_MODE;
 506                         dti->dti_attr.la_mode           = mode;
 507                         dti->dti_dof.dof_type           = DFT_INDEX;
 508                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 509                         dto = __local_file_create(env, &dti->dti_fid, los,
 510                                                   dt2ls_dev(los->los_dev),
 511                                                   parent, name, &dti->dti_attr,
 512                                                   &dti->dti_dof);
 513                 }
 514         }
 515         return dto;
 516
 517 }
 518 EXPORT_SYMBOL(local_index_find_or_create);
 519
 520 struct dt_object *
 521 local_index_find_or_create_with_fid(const struct lu_env *env,
 522                                     struct dt_device *dt,
 523                                     const struct lu_fid *fid,
 524                                     struct dt_object *parent,
 525                                     const char *name, __u32 mode,
 526                                     const struct dt_index_features *ft)
 527 {
 528         struct dt_thread_info   *dti = dt_info(env);
 529         struct dt_object        *dto;
 530         int                      rc;
 531
 532         LASSERT(parent);
 533
 534         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 535         if (rc == 0) {
 536                 /* name is found, get the object */
 537                 if (!lu_fid_eq(fid, &dti->dti_fid))
 538                         dto = ERR_PTR(-EINVAL);
 539                 else
 540                         dto = dt_locate(env, dt, fid);
 541         } else if (rc != -ENOENT) {
 542                 dto = ERR_PTR(rc);
 543         } else {
 544                 struct ls_device *ls;
 545
 546                 ls = ls_device_get(dt);
 547                 if (IS_ERR(ls)) {
 548                         dto = ERR_PTR(PTR_ERR(ls));
 549                 } else {
 550                         /* create the object */
 551                         dti->dti_attr.la_valid          = LA_MODE;
 552                         dti->dti_attr.la_mode           = mode;
 553                         dti->dti_dof.dof_type           = DFT_INDEX;
 554                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 555                         dto = __local_file_create(env, fid, NULL, ls, parent,
 556                                                   name, &dti->dti_attr,
 557                                                   &dti->dti_dof);
 558                         /* ls_device_put() will finalize the ls device, we
 559                          * have to open the object in other device stack */
 560                         if (!IS_ERR(dto)) {
 561                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 562                                 lu_object_put_nocache(env, &dto->do_lu);
 563                                 dto = dt_locate(env, dt, &dti->dti_fid);
 564                         }
 565                         ls_device_put(env, ls);
 566                 }
 567         }
 568         return dto;
 569 }
 570 EXPORT_SYMBOL(local_index_find_or_create_with_fid);
 571
 572 static int local_object_declare_unlink(const struct lu_env *env,
 573                                        struct dt_device *dt,
 574                                        struct dt_object *p,
 575                                        struct dt_object *c, const char *name,
 576                                        struct thandle *th)
 577 {
 578         int rc;
 579
 580         rc = dt_declare_delete(env, p, (const struct dt_key *)name, th);
 581         if (rc < 0)
 582                 return rc;
 583
 584         rc = dt_declare_ref_del(env, c, th);
 585         if (rc < 0)
 586                 return rc;
 587
 588         return dt_declare_destroy(env, c, th);
 589 }
 590
 591 int local_object_unlink(const struct lu_env *env, struct dt_device *dt,
 592                         struct dt_object *parent, const char *name)
 593 {
 594         struct dt_thread_info   *dti = dt_info(env);
 595         struct dt_object        *dto;
 596         struct thandle          *th;
 597         int                      rc;
 598
 599         ENTRY;
 600
 601         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 602         if (rc == -ENOENT)
 603                 RETURN(0);
 604         else if (rc < 0)
 605                 RETURN(rc);
 606
 607         dto = dt_locate(env, dt, &dti->dti_fid);
 608         if (unlikely(IS_ERR(dto)))
 609                 RETURN(PTR_ERR(dto));
 610
 611         th = dt_trans_create(env, dt);
 612         if (IS_ERR(th))
 613                 GOTO(out, rc = PTR_ERR(th));
 614
 615         rc = local_object_declare_unlink(env, dt, parent, dto, name, th);
 616         if (rc < 0)
 617                 GOTO(stop, rc);
 618
 619         rc = dt_trans_start_local(env, dt, th);
 620         if (rc < 0)
 621                 GOTO(stop, rc);
 622
 623         dt_write_lock(env, dto, 0);
 624         rc = dt_delete(env, parent, (struct dt_key *)name, th, BYPASS_CAPA);
 625         if (rc < 0)
 626                 GOTO(unlock, rc);
 627
 628         rc = dt_ref_del(env, dto, th);
 629         if (rc < 0) {
 630                 rc = dt_insert(env, parent,
 631                                (const struct dt_rec *)&dti->dti_fid,
 632                                (const struct dt_key *)name, th, BYPASS_CAPA, 1);
 633                 GOTO(unlock, rc);
 634         }
 635
 636         rc = dt_destroy(env, dto, th);
 637 unlock:
 638         dt_write_unlock(env, dto);
 639 stop:
 640         dt_trans_stop(env, dt, th);
 641 out:
 642         lu_object_put_nocache(env, &dto->do_lu);
 643         return rc;
 644 }
 645 EXPORT_SYMBOL(local_object_unlink);
 646
 647 struct local_oid_storage *dt_los_find(struct ls_device *ls, __u64 seq)
 648 {
 649         struct local_oid_storage *los, *ret = NULL;
 650
 651         cfs_list_for_each_entry(los, &ls->ls_los_list, los_list) {
 652                 if (los->los_seq == seq) {
 653                         atomic_inc(&los->los_refcount);
 654                         ret = los;
 655                         break;
 656                 }
 657         }
 658         return ret;
 659 }
 660
 661 void dt_los_put(struct local_oid_storage *los)
 662 {
 663         if (atomic_dec_and_test(&los->los_refcount))
 664                 /* should never happen, only local_oid_storage_fini should
 665                  * drop refcount to zero */
 666                 LBUG();
 667         return;
 668 }
 669
 670 /* after Lustre 2.3 release there may be old file to store last generated FID
 671  * If such file exists then we have to read its content
 672  */
 673 int lastid_compat_check(const struct lu_env *env, struct dt_device *dev,
 674                         __u64 lastid_seq, __u32 *first_oid, struct ls_device *ls)
 675 {
 676         struct dt_thread_info   *dti = dt_info(env);
 677         struct dt_object        *root = NULL;
 678         struct los_ondisk        losd;
 679         struct dt_object        *o = NULL;
 680         int                      rc = 0;
 681
 682         rc = dt_root_get(env, dev, &dti->dti_fid);
 683         if (rc)
 684                 return rc;
 685
 686         root = ls_locate(env, ls, &dti->dti_fid, NULL);
 687         if (IS_ERR(root))
 688                 return PTR_ERR(root);
 689
 690         /* find old last_id file */
 691         snprintf(dti->dti_buf, sizeof(dti->dti_buf), "seq-"LPX64"-lastid",
 692                  lastid_seq);
 693         rc = dt_lookup_dir(env, root, dti->dti_buf, &dti->dti_fid);
 694         lu_object_put_nocache(env, &root->do_lu);
 695         if (rc == -ENOENT) {
 696                 struct lu_object_conf *conf = &dti->dti_conf;
 697
 698                 /* old llog lastid accessed by FID only */
 699                 if (lastid_seq != FID_SEQ_LLOG)
 700                         return 0;
 701                 dti->dti_fid.f_seq = FID_SEQ_LLOG;
 702                 dti->dti_fid.f_oid = 1;
 703                 dti->dti_fid.f_ver = 0;
 704                 memset(conf, 0, sizeof(*conf));
 705                 conf->loc_flags = LOC_F_NEW;
 706                 o = ls_locate(env, ls, &dti->dti_fid, conf);
 707                 if (IS_ERR(o))
 708                         return PTR_ERR(o);
 709
 710                 if (!dt_object_exists(o)) {
 711                         lu_object_put_nocache(env, &o->do_lu);
 712                         return 0;
 713                 }
 714                 CDEBUG(D_INFO, "Found old llog lastid file\n");
 715         } else if (rc < 0) {
 716                 return rc;
 717         } else {
 718                 CDEBUG(D_INFO, "Found old lastid file for sequence "LPX64"\n",
 719                        lastid_seq);
 720                 o = ls_locate(env, ls, &dti->dti_fid, NULL);
 721                 if (IS_ERR(o))
 722                         return PTR_ERR(o);
 723         }
 724         /* let's read seq-NNNNNN-lastid file value */
 725         LASSERT(dt_object_exists(o));
 726         dti->dti_off = 0;
 727         dti->dti_lb.lb_buf = &losd;
 728         dti->dti_lb.lb_len = sizeof(losd);
 729         dt_read_lock(env, o, 0);
 730         rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 731         dt_read_unlock(env, o);
 732         lu_object_put_nocache(env, &o->do_lu);
 733         if (rc == 0 && le32_to_cpu(losd.lso_magic) != LOS_MAGIC) {
 734                 CERROR("%s: wrong content of seq-"LPX64"-lastid file, magic %x\n",
 735                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq,
 736                        le32_to_cpu(losd.lso_magic));
 737                 return -EINVAL;
 738         } else if (rc < 0) {
 739                 CERROR("%s: failed to read seq-"LPX64"-lastid: rc = %d\n",
 740                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq, rc);
 741                 return rc;
 742         }
 743         *first_oid = le32_to_cpu(losd.lso_next_oid);
 744         return rc;
 745 }
 746
 747 /**
 748  * Initialize local OID storage for required sequence.
 749  * That may be needed for services that uses local files and requires
 750  * dynamic OID allocation for them.
 751  *
 752  * Per each sequence we have an object with 'first_fid' identificator
 753  * containing the counter for OIDs of locally created files with that
 754  * sequence.
 755  *
 756  * It is used now by llog subsystem and MGS for NID tables
 757  *
 758  * Function gets first_fid to create counter object.
 759  * All dynamic fids will be generated with the same sequence and incremented
 760  * OIDs
 761  *
 762  * Returned local_oid_storage is in-memory representaion of OID storage
 763  */
 764 int local_oid_storage_init(const struct lu_env *env, struct dt_device *dev,
 765                            const struct lu_fid *first_fid,
 766                            struct local_oid_storage **los)
 767 {
 768         struct dt_thread_info   *dti = dt_info(env);
 769         struct ls_device        *ls;
 770         obd_id                   lastid;
 771         struct dt_object        *o = NULL;
 772         struct thandle          *th;
 773         __u32                    first_oid = fid_oid(first_fid);
 774         int                      rc = 0;
 775
 776         ENTRY;
 777
 778         ls = ls_device_get(dev);
 779         if (IS_ERR(ls))
 780                 RETURN(PTR_ERR(ls));
 781
 782         mutex_lock(&ls->ls_los_mutex);
 783         *los = dt_los_find(ls, fid_seq(first_fid));
 784         if (*los != NULL)
 785                 GOTO(out, rc = 0);
 786
 787         /* not found, then create */
 788         OBD_ALLOC_PTR(*los);
 789         if (*los == NULL)
 790                 GOTO(out, rc = -ENOMEM);
 791
 792         atomic_set(&(*los)->los_refcount, 1);
 793         mutex_init(&(*los)->los_id_lock);
 794         (*los)->los_dev = &ls->ls_top_dev;
 795         atomic_inc(&ls->ls_refcount);
 796         cfs_list_add(&(*los)->los_list, &ls->ls_los_list);
 797
 798         /* Use {seq, 0, 0} to create the LAST_ID file for every
 799          * sequence.  OIDs start at LUSTRE_FID_INIT_OID.
 800          */
 801         dti->dti_fid.f_seq = fid_seq(first_fid);
 802         dti->dti_fid.f_oid = LUSTRE_FID_LASTID_OID;
 803         dti->dti_fid.f_ver = 0;
 804         o = ls_locate(env, ls, &dti->dti_fid, NULL);
 805         if (IS_ERR(o))
 806                 GOTO(out_los, rc = PTR_ERR(o));
 807
 808         if (!dt_object_exists(o)) {
 809                 rc = lastid_compat_check(env, dev, fid_seq(first_fid),
 810                                          &first_oid, ls);
 811                 if (rc < 0)
 812                         GOTO(out_los, rc);
 813
 814                 th = dt_trans_create(env, dev);
 815                 if (IS_ERR(th))
 816                         GOTO(out_los, rc = PTR_ERR(th));
 817
 818                 dti->dti_attr.la_valid = LA_MODE | LA_TYPE;
 819                 dti->dti_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
 820                 dti->dti_dof.dof_type = dt_mode_to_dft(S_IFREG);
 821
 822                 rc = dt_declare_create(env, o, &dti->dti_attr, NULL,
 823                                        &dti->dti_dof, th);
 824                 if (rc)
 825                         GOTO(out_trans, rc);
 826
 827                 rc = dt_declare_record_write(env, o, sizeof(lastid), 0, th);
 828                 if (rc)
 829                         GOTO(out_trans, rc);
 830
 831                 rc = dt_trans_start_local(env, dev, th);
 832                 if (rc)
 833                         GOTO(out_trans, rc);
 834
 835                 dt_write_lock(env, o, 0);
 836                 if (dt_object_exists(o))
 837                         GOTO(out_lock, rc = 0);
 838
 839                 rc = dt_create(env, o, &dti->dti_attr, NULL, &dti->dti_dof,
 840                                th);
 841                 if (rc)
 842                         GOTO(out_lock, rc);
 843
 844                 lastid = cpu_to_le64(first_oid);
 845
 846                 dti->dti_off = 0;
 847                 dti->dti_lb.lb_buf = &lastid;
 848                 dti->dti_lb.lb_len = sizeof(lastid);
 849                 rc = dt_record_write(env, o, &dti->dti_lb, &dti->dti_off, th);
 850                 if (rc)
 851                         GOTO(out_lock, rc);
 852 out_lock:
 853                 dt_write_unlock(env, o);
 854 out_trans:
 855                 dt_trans_stop(env, dev, th);
 856         } else {
 857                 dti->dti_off = 0;
 858                 dti->dti_lb.lb_buf = &lastid;
 859                 dti->dti_lb.lb_len = sizeof(lastid);
 860                 dt_read_lock(env, o, 0);
 861                 rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 862                 dt_read_unlock(env, o);
 863                 if (rc == 0 && le64_to_cpu(lastid) > OBIF_MAX_OID) {
 864                         CERROR("%s: bad oid "LPU64" is read from LAST_ID\n",
 865                                o->do_lu.lo_dev->ld_obd->obd_name,
 866                                le64_to_cpu(lastid));
 867                         rc = -EINVAL;
 868                 }
 869         }
 870 out_los:
 871         if (rc != 0) {
 872                 cfs_list_del(&(*los)->los_list);
 873                 atomic_dec(&ls->ls_refcount);
 874                 OBD_FREE_PTR(*los);
 875                 *los = NULL;
 876                 if (o != NULL && !IS_ERR(o))
 877                         lu_object_put_nocache(env, &o->do_lu);
 878         } else {
 879                 (*los)->los_seq = fid_seq(first_fid);
 880                 (*los)->los_last_oid = le64_to_cpu(lastid);
 881                 (*los)->los_obj = o;
 882                 /* Read value should not be less than initial one
 883                  * but possible after upgrade from older fs.
 884                  * In this case just switch to the first_oid in memory and
 885                  * it will be updated on disk with first object generated */
 886                 if ((*los)->los_last_oid < first_oid)
 887                         (*los)->los_last_oid = first_oid;
 888         }
 889 out:
 890         mutex_unlock(&ls->ls_los_mutex);
 891         ls_device_put(env, ls);
 892         return rc;
 893 }
 894 EXPORT_SYMBOL(local_oid_storage_init);
 895
 896 void local_oid_storage_fini(const struct lu_env *env,
 897                             struct local_oid_storage *los)
 898 {
 899         struct ls_device *ls;
 900
 901         if (!atomic_dec_and_test(&los->los_refcount))
 902                 return;
 903
 904         LASSERT(env);
 905         LASSERT(los->los_dev);
 906         ls = dt2ls_dev(los->los_dev);
 907
 908         mutex_lock(&ls->ls_los_mutex);
 909         if (atomic_read(&los->los_refcount) > 0) {
 910                 mutex_unlock(&ls->ls_los_mutex);
 911                 return;
 912         }
 913
 914         if (los->los_obj)
 915                 lu_object_put_nocache(env, &los->los_obj->do_lu);
 916         cfs_list_del(&los->los_list);
 917         OBD_FREE_PTR(los);
 918         mutex_unlock(&ls->ls_los_mutex);
 919         ls_device_put(env, ls);
 920 }
 921 EXPORT_SYMBOL(local_oid_storage_fini);