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                         cfs_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         cfs_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 (!cfs_atomic_dec_and_test(&ls->ls_refcount))
 180                 return;
 181
 182         mutex_lock(&ls_list_mutex);
 183         if (cfs_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 dt_object        *dto;
 302         struct thandle          *th;
 303         int                      rc;
 304
 305         dto = ls_locate(env, ls, fid);
 306         if (unlikely(IS_ERR(dto)))
 307                 RETURN(dto);
 308
 309         LASSERT(dto != NULL);
 310         if (dt_object_exists(dto))
 311                 GOTO(out, rc = -EEXIST);
 312
 313         th = dt_trans_create(env, ls->ls_osd);
 314         if (IS_ERR(th))
 315                 GOTO(out, rc = PTR_ERR(th));
 316
 317         rc = local_object_declare_create(env, los, dto, attr, dof, th);
 318         if (rc)
 319                 GOTO(trans_stop, rc);
 320
 321         if (dti->dti_dof.dof_type == DFT_DIR) {
 322                 dt_declare_ref_add(env, dto, th);
 323                 dt_declare_ref_add(env, parent, th);
 324         }
 325
 326         rc = dt_declare_insert(env, parent, (void *)fid, (void *)name, th);
 327         if (rc)
 328                 GOTO(trans_stop, rc);
 329
 330         rc = dt_trans_start_local(env, ls->ls_osd, th);
 331         if (rc)
 332                 GOTO(trans_stop, rc);
 333
 334         dt_write_lock(env, dto, 0);
 335         if (dt_object_exists(dto))
 336                 GOTO(unlock, rc = 0);
 337
 338         CDEBUG(D_OTHER, "create new object "DFID"\n",
 339                PFID(lu_object_fid(&dto->do_lu)));
 340         rc = local_object_create(env, los, dto, attr, dof, th);
 341         if (rc)
 342                 GOTO(unlock, rc);
 343         LASSERT(dt_object_exists(dto));
 344
 345         if (dti->dti_dof.dof_type == DFT_DIR) {
 346                 if (!dt_try_as_dir(env, dto))
 347                         GOTO(destroy, rc = -ENOTDIR);
 348                 /* Add "." and ".." for newly created dir */
 349                 rc = dt_insert(env, dto, (void *)fid, (void *)".", th,
 350                                BYPASS_CAPA, 1);
 351                 if (rc)
 352                         GOTO(destroy, rc);
 353                 dt_ref_add(env, dto, th);
 354                 rc = dt_insert(env, dto, (void *)lu_object_fid(&parent->do_lu),
 355                                (void *)"..", th, BYPASS_CAPA, 1);
 356                 if (rc)
 357                         GOTO(destroy, rc);
 358         }
 359
 360         dt_write_lock(env, parent, 0);
 361         rc = dt_insert(env, parent, (const struct dt_rec *)fid,
 362                        (const struct dt_key *)name, th, BYPASS_CAPA, 1);
 363         if (dti->dti_dof.dof_type == DFT_DIR)
 364                 dt_ref_add(env, parent, th);
 365         dt_write_unlock(env, parent);
 366         if (rc)
 367                 GOTO(destroy, rc);
 368 destroy:
 369         if (rc)
 370                 dt_destroy(env, dto, th);
 371 unlock:
 372         dt_write_unlock(env, dto);
 373 trans_stop:
 374         dt_trans_stop(env, ls->ls_osd, th);
 375 out:
 376         if (rc) {
 377                 lu_object_put_nocache(env, &dto->do_lu);
 378                 dto = ERR_PTR(rc);
 379         }
 380         RETURN(dto);
 381 }
 382
 383 /*
 384  * Look up and create (if it does not exist) a local named file or directory in
 385  * parent directory.
 386  */
 387 struct dt_object *local_file_find_or_create(const struct lu_env *env,
 388                                             struct local_oid_storage *los,
 389                                             struct dt_object *parent,
 390                                             const char *name, __u32 mode)
 391 {
 392         struct dt_thread_info   *dti = dt_info(env);
 393         struct dt_object        *dto;
 394         int                      rc;
 395
 396         LASSERT(parent);
 397
 398         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 399         if (rc == 0)
 400                 /* name is found, get the object */
 401                 dto = ls_locate(env, dt2ls_dev(los->los_dev), &dti->dti_fid);
 402         else if (rc != -ENOENT)
 403                 dto = ERR_PTR(rc);
 404         else {
 405                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 406                 if (rc < 0) {
 407                         dto = ERR_PTR(rc);
 408                 } else {
 409                         /* create the object */
 410                         dti->dti_attr.la_valid  = LA_MODE;
 411                         dti->dti_attr.la_mode   = mode;
 412                         dti->dti_dof.dof_type   = dt_mode_to_dft(mode & S_IFMT);
 413                         dto = __local_file_create(env, &dti->dti_fid, los,
 414                                                   dt2ls_dev(los->los_dev),
 415                                                   parent, name, &dti->dti_attr,
 416                                                   &dti->dti_dof);
 417                 }
 418         }
 419         return dto;
 420 }
 421 EXPORT_SYMBOL(local_file_find_or_create);
 422
 423 struct dt_object *local_file_find_or_create_with_fid(const struct lu_env *env,
 424                                                      struct dt_device *dt,
 425                                                      const struct lu_fid *fid,
 426                                                      struct dt_object *parent,
 427                                                      const char *name,
 428                                                      __u32 mode)
 429 {
 430         struct dt_thread_info   *dti = dt_info(env);
 431         struct dt_object        *dto;
 432         int                      rc;
 433
 434         LASSERT(parent);
 435
 436         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 437         if (rc == 0) {
 438                 dto = dt_locate(env, dt, &dti->dti_fid);
 439         } else if (rc != -ENOENT) {
 440                 dto = ERR_PTR(rc);
 441         } else {
 442                 struct ls_device *ls;
 443
 444                 ls = ls_device_get(dt);
 445                 if (IS_ERR(ls)) {
 446                         dto = ERR_PTR(PTR_ERR(ls));
 447                 } else {
 448                         /* create the object */
 449                         dti->dti_attr.la_valid  = LA_MODE;
 450                         dti->dti_attr.la_mode   = mode;
 451                         dti->dti_dof.dof_type   = dt_mode_to_dft(mode & S_IFMT);
 452                         dto = __local_file_create(env, fid, NULL, ls, parent,
 453                                                   name, &dti->dti_attr,
 454                                                   &dti->dti_dof);
 455                         /* ls_device_put() will finalize the ls device, we
 456                          * have to open the object in other device stack */
 457                         if (!IS_ERR(dto)) {
 458                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 459                                 lu_object_put_nocache(env, &dto->do_lu);
 460                                 dto = dt_locate(env, dt, &dti->dti_fid);
 461                         }
 462                         ls_device_put(env, ls);
 463                 }
 464         }
 465         return dto;
 466 }
 467 EXPORT_SYMBOL(local_file_find_or_create_with_fid);
 468
 469 /*
 470  * Look up and create (if it does not exist) a local named index file in parent
 471  * directory.
 472  */
 473 struct dt_object *local_index_find_or_create(const struct lu_env *env,
 474                                              struct local_oid_storage *los,
 475                                              struct dt_object *parent,
 476                                              const char *name, __u32 mode,
 477                                              const struct dt_index_features *ft)
 478 {
 479         struct dt_thread_info   *dti = dt_info(env);
 480         struct dt_object        *dto;
 481         int                      rc;
 482
 483         LASSERT(parent);
 484
 485         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 486         if (rc == 0) {
 487                 /* name is found, get the object */
 488                 dto = ls_locate(env, dt2ls_dev(los->los_dev), &dti->dti_fid);
 489         } else if (rc != -ENOENT) {
 490                 dto = ERR_PTR(rc);
 491         } else {
 492                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 493                 if (rc < 0) {
 494                         dto = ERR_PTR(rc);
 495                 } else {
 496                         /* create the object */
 497                         dti->dti_attr.la_valid          = LA_MODE;
 498                         dti->dti_attr.la_mode           = mode;
 499                         dti->dti_dof.dof_type           = DFT_INDEX;
 500                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 501                         dto = __local_file_create(env, &dti->dti_fid, los,
 502                                                   dt2ls_dev(los->los_dev),
 503                                                   parent, name, &dti->dti_attr,
 504                                                   &dti->dti_dof);
 505                 }
 506         }
 507         return dto;
 508
 509 }
 510 EXPORT_SYMBOL(local_index_find_or_create);
 511
 512 struct dt_object *
 513 local_index_find_or_create_with_fid(const struct lu_env *env,
 514                                     struct dt_device *dt,
 515                                     const struct lu_fid *fid,
 516                                     struct dt_object *parent,
 517                                     const char *name, __u32 mode,
 518                                     const struct dt_index_features *ft)
 519 {
 520         struct dt_thread_info   *dti = dt_info(env);
 521         struct dt_object        *dto;
 522         int                      rc;
 523
 524         LASSERT(parent);
 525
 526         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 527         if (rc == 0) {
 528                 /* name is found, get the object */
 529                 if (!lu_fid_eq(fid, &dti->dti_fid))
 530                         dto = ERR_PTR(-EINVAL);
 531                 else
 532                         dto = dt_locate(env, dt, fid);
 533         } else if (rc != -ENOENT) {
 534                 dto = ERR_PTR(rc);
 535         } else {
 536                 struct ls_device *ls;
 537
 538                 ls = ls_device_get(dt);
 539                 if (IS_ERR(ls)) {
 540                         dto = ERR_PTR(PTR_ERR(ls));
 541                 } else {
 542                         /* create the object */
 543                         dti->dti_attr.la_valid          = LA_MODE;
 544                         dti->dti_attr.la_mode           = mode;
 545                         dti->dti_dof.dof_type           = DFT_INDEX;
 546                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 547                         dto = __local_file_create(env, fid, NULL, ls, parent,
 548                                                   name, &dti->dti_attr,
 549                                                   &dti->dti_dof);
 550                         /* ls_device_put() will finalize the ls device, we
 551                          * have to open the object in other device stack */
 552                         if (!IS_ERR(dto)) {
 553                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 554                                 lu_object_put_nocache(env, &dto->do_lu);
 555                                 dto = dt_locate(env, dt, &dti->dti_fid);
 556                         }
 557                         ls_device_put(env, ls);
 558                 }
 559         }
 560         return dto;
 561 }
 562 EXPORT_SYMBOL(local_index_find_or_create_with_fid);
 563
 564 static int local_object_declare_unlink(const struct lu_env *env,
 565                                        struct dt_device *dt,
 566                                        struct dt_object *p,
 567                                        struct dt_object *c, const char *name,
 568                                        struct thandle *th)
 569 {
 570         int rc;
 571
 572         rc = dt_declare_delete(env, p, (const struct dt_key *)name, th);
 573         if (rc < 0)
 574                 return rc;
 575
 576         rc = dt_declare_ref_del(env, c, th);
 577         if (rc < 0)
 578                 return rc;
 579
 580         return dt_declare_destroy(env, c, th);
 581 }
 582
 583 int local_object_unlink(const struct lu_env *env, struct dt_device *dt,
 584                         struct dt_object *parent, const char *name)
 585 {
 586         struct dt_thread_info   *dti = dt_info(env);
 587         struct dt_object        *dto;
 588         struct thandle          *th;
 589         int                      rc;
 590
 591         ENTRY;
 592
 593         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 594         if (rc == -ENOENT)
 595                 RETURN(0);
 596         else if (rc < 0)
 597                 RETURN(rc);
 598
 599         dto = dt_locate(env, dt, &dti->dti_fid);
 600         if (unlikely(IS_ERR(dto)))
 601                 RETURN(PTR_ERR(dto));
 602
 603         th = dt_trans_create(env, dt);
 604         if (IS_ERR(th))
 605                 GOTO(out, rc = PTR_ERR(th));
 606
 607         rc = local_object_declare_unlink(env, dt, parent, dto, name, th);
 608         if (rc < 0)
 609                 GOTO(stop, rc);
 610
 611         rc = dt_trans_start_local(env, dt, th);
 612         if (rc < 0)
 613                 GOTO(stop, rc);
 614
 615         dt_write_lock(env, dto, 0);
 616         rc = dt_delete(env, parent, (struct dt_key *)name, th, BYPASS_CAPA);
 617         if (rc < 0)
 618                 GOTO(unlock, rc);
 619
 620         rc = dt_ref_del(env, dto, th);
 621         if (rc < 0) {
 622                 rc = dt_insert(env, parent,
 623                                (const struct dt_rec *)&dti->dti_fid,
 624                                (const struct dt_key *)name, th, BYPASS_CAPA, 1);
 625                 GOTO(unlock, rc);
 626         }
 627
 628         rc = dt_destroy(env, dto, th);
 629 unlock:
 630         dt_write_unlock(env, dto);
 631 stop:
 632         dt_trans_stop(env, dt, th);
 633 out:
 634         lu_object_put_nocache(env, &dto->do_lu);
 635         return rc;
 636 }
 637 EXPORT_SYMBOL(local_object_unlink);
 638
 639 struct local_oid_storage *dt_los_find(struct ls_device *ls, __u64 seq)
 640 {
 641         struct local_oid_storage *los, *ret = NULL;
 642
 643         cfs_list_for_each_entry(los, &ls->ls_los_list, los_list) {
 644                 if (los->los_seq == seq) {
 645                         cfs_atomic_inc(&los->los_refcount);
 646                         ret = los;
 647                         break;
 648                 }
 649         }
 650         return ret;
 651 }
 652
 653 void dt_los_put(struct local_oid_storage *los)
 654 {
 655         if (cfs_atomic_dec_and_test(&los->los_refcount))
 656                 /* should never happen, only local_oid_storage_fini should
 657                  * drop refcount to zero */
 658                 LBUG();
 659         return;
 660 }
 661
 662 /* after Lustre 2.3 release there may be old file to store last generated FID
 663  * If such file exists then we have to read its content
 664  */
 665 int lastid_compat_check(const struct lu_env *env, struct dt_device *dev,
 666                         __u64 lastid_seq, __u32 *first_oid, struct ls_device *ls)
 667 {
 668         struct dt_thread_info   *dti = dt_info(env);
 669         struct dt_object        *root = NULL;
 670         struct los_ondisk        losd;
 671         struct dt_object        *o = NULL;
 672         int                      rc = 0;
 673
 674         rc = dt_root_get(env, dev, &dti->dti_fid);
 675         if (rc)
 676                 return rc;
 677
 678         root = ls_locate(env, ls, &dti->dti_fid);
 679         if (IS_ERR(root))
 680                 return PTR_ERR(root);
 681
 682         /* find old last_id file */
 683         snprintf(dti->dti_buf, sizeof(dti->dti_buf), "seq-"LPX64"-lastid",
 684                  lastid_seq);
 685         rc = dt_lookup_dir(env, root, dti->dti_buf, &dti->dti_fid);
 686         lu_object_put_nocache(env, &root->do_lu);
 687         if (rc == -ENOENT) {
 688                 /* old llog lastid accessed by FID only */
 689                 if (lastid_seq != FID_SEQ_LLOG)
 690                         return 0;
 691                 dti->dti_fid.f_seq = FID_SEQ_LLOG;
 692                 dti->dti_fid.f_oid = 1;
 693                 dti->dti_fid.f_ver = 0;
 694                 o = ls_locate(env, ls, &dti->dti_fid);
 695                 if (IS_ERR(o))
 696                         return PTR_ERR(o);
 697
 698                 if (!dt_object_exists(o)) {
 699                         lu_object_put_nocache(env, &o->do_lu);
 700                         return 0;
 701                 }
 702                 CDEBUG(D_INFO, "Found old llog lastid file\n");
 703         } else if (rc < 0) {
 704                 return rc;
 705         } else {
 706                 CDEBUG(D_INFO, "Found old lastid file for sequence "LPX64"\n",
 707                        lastid_seq);
 708                 o = ls_locate(env, ls, &dti->dti_fid);
 709                 if (IS_ERR(o))
 710                         return PTR_ERR(o);
 711         }
 712         /* let's read seq-NNNNNN-lastid file value */
 713         LASSERT(dt_object_exists(o));
 714         dti->dti_off = 0;
 715         dti->dti_lb.lb_buf = &losd;
 716         dti->dti_lb.lb_len = sizeof(losd);
 717         dt_read_lock(env, o, 0);
 718         rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 719         dt_read_unlock(env, o);
 720         lu_object_put_nocache(env, &o->do_lu);
 721         if (rc == 0 && le32_to_cpu(losd.lso_magic) != LOS_MAGIC) {
 722                 CERROR("%s: wrong content of seq-"LPX64"-lastid file, magic %x\n",
 723                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq,
 724                        le32_to_cpu(losd.lso_magic));
 725                 return -EINVAL;
 726         } else if (rc < 0) {
 727                 CERROR("%s: failed to read seq-"LPX64"-lastid: rc = %d\n",
 728                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq, rc);
 729                 return rc;
 730         }
 731         *first_oid = le32_to_cpu(losd.lso_next_oid);
 732         return rc;
 733 }
 734
 735 /**
 736  * Initialize local OID storage for required sequence.
 737  * That may be needed for services that uses local files and requires
 738  * dynamic OID allocation for them.
 739  *
 740  * Per each sequence we have an object with 'first_fid' identificator
 741  * containing the counter for OIDs of locally created files with that
 742  * sequence.
 743  *
 744  * It is used now by llog subsystem and MGS for NID tables
 745  *
 746  * Function gets first_fid to create counter object.
 747  * All dynamic fids will be generated with the same sequence and incremented
 748  * OIDs
 749  *
 750  * Returned local_oid_storage is in-memory representaion of OID storage
 751  */
 752 int local_oid_storage_init(const struct lu_env *env, struct dt_device *dev,
 753                            const struct lu_fid *first_fid,
 754                            struct local_oid_storage **los)
 755 {
 756         struct dt_thread_info   *dti = dt_info(env);
 757         struct ls_device        *ls;
 758         obd_id                   lastid;
 759         struct dt_object        *o = NULL;
 760         struct thandle          *th;
 761         __u32                    first_oid = fid_oid(first_fid);
 762         int                      rc = 0;
 763
 764         ENTRY;
 765
 766         ls = ls_device_get(dev);
 767         if (IS_ERR(ls))
 768                 RETURN(PTR_ERR(ls));
 769
 770         mutex_lock(&ls->ls_los_mutex);
 771         *los = dt_los_find(ls, fid_seq(first_fid));
 772         if (*los != NULL)
 773                 GOTO(out, rc = 0);
 774
 775         /* not found, then create */
 776         OBD_ALLOC_PTR(*los);
 777         if (*los == NULL)
 778                 GOTO(out, rc = -ENOMEM);
 779
 780         cfs_atomic_set(&(*los)->los_refcount, 1);
 781         mutex_init(&(*los)->los_id_lock);
 782         (*los)->los_dev = &ls->ls_top_dev;
 783         cfs_atomic_inc(&ls->ls_refcount);
 784         cfs_list_add(&(*los)->los_list, &ls->ls_los_list);
 785
 786         /* Use {seq, 0, 0} to create the LAST_ID file for every
 787          * sequence.  OIDs start at LUSTRE_FID_INIT_OID.
 788          */
 789         dti->dti_fid.f_seq = fid_seq(first_fid);
 790         dti->dti_fid.f_oid = LUSTRE_FID_LASTID_OID;
 791         dti->dti_fid.f_ver = 0;
 792         o = ls_locate(env, ls, &dti->dti_fid);
 793         if (IS_ERR(o))
 794                 GOTO(out_los, rc = PTR_ERR(o));
 795
 796         if (!dt_object_exists(o)) {
 797                 rc = lastid_compat_check(env, dev, fid_seq(first_fid),
 798                                          &first_oid, ls);
 799                 if (rc < 0)
 800                         GOTO(out_los, rc);
 801
 802                 th = dt_trans_create(env, dev);
 803                 if (IS_ERR(th))
 804                         GOTO(out_los, rc = PTR_ERR(th));
 805
 806                 dti->dti_attr.la_valid = LA_MODE | LA_TYPE;
 807                 dti->dti_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
 808                 dti->dti_dof.dof_type = dt_mode_to_dft(S_IFREG);
 809
 810                 rc = dt_declare_create(env, o, &dti->dti_attr, NULL,
 811                                        &dti->dti_dof, th);
 812                 if (rc)
 813                         GOTO(out_trans, rc);
 814
 815                 rc = dt_declare_record_write(env, o, sizeof(lastid), 0, th);
 816                 if (rc)
 817                         GOTO(out_trans, rc);
 818
 819                 rc = dt_trans_start_local(env, dev, th);
 820                 if (rc)
 821                         GOTO(out_trans, rc);
 822
 823                 dt_write_lock(env, o, 0);
 824                 if (dt_object_exists(o))
 825                         GOTO(out_lock, rc = 0);
 826
 827                 rc = dt_create(env, o, &dti->dti_attr, NULL, &dti->dti_dof,
 828                                th);
 829                 if (rc)
 830                         GOTO(out_lock, rc);
 831
 832                 lastid = cpu_to_le64(first_oid);
 833
 834                 dti->dti_off = 0;
 835                 dti->dti_lb.lb_buf = &lastid;
 836                 dti->dti_lb.lb_len = sizeof(lastid);
 837                 rc = dt_record_write(env, o, &dti->dti_lb, &dti->dti_off, th);
 838                 if (rc)
 839                         GOTO(out_lock, rc);
 840 out_lock:
 841                 dt_write_unlock(env, o);
 842 out_trans:
 843                 dt_trans_stop(env, dev, th);
 844         } else {
 845                 dti->dti_off = 0;
 846                 dti->dti_lb.lb_buf = &lastid;
 847                 dti->dti_lb.lb_len = sizeof(lastid);
 848                 dt_read_lock(env, o, 0);
 849                 rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 850                 dt_read_unlock(env, o);
 851                 if (rc == 0 && le64_to_cpu(lastid) > OBIF_MAX_OID) {
 852                         CERROR("%s: bad oid "LPU64" is read from LAST_ID\n",
 853                                o->do_lu.lo_dev->ld_obd->obd_name,
 854                                le64_to_cpu(lastid));
 855                         rc = -EINVAL;
 856                 }
 857         }
 858 out_los:
 859         if (rc != 0) {
 860                 cfs_list_del(&(*los)->los_list);
 861                 cfs_atomic_dec(&ls->ls_refcount);
 862                 OBD_FREE_PTR(*los);
 863                 *los = NULL;
 864                 if (o != NULL && !IS_ERR(o))
 865                         lu_object_put_nocache(env, &o->do_lu);
 866         } else {
 867                 (*los)->los_seq = fid_seq(first_fid);
 868                 (*los)->los_last_oid = le64_to_cpu(lastid);
 869                 (*los)->los_obj = o;
 870                 /* read value should not be less than initial one */
 871                 LASSERTF((*los)->los_last_oid >= first_oid, "%u < %u\n",
 872                          (*los)->los_last_oid, first_oid);
 873         }
 874 out:
 875         mutex_unlock(&ls->ls_los_mutex);
 876         ls_device_put(env, ls);
 877         return rc;
 878 }
 879 EXPORT_SYMBOL(local_oid_storage_init);
 880
 881 void local_oid_storage_fini(const struct lu_env *env,
 882                             struct local_oid_storage *los)
 883 {
 884         struct ls_device *ls;
 885
 886         if (!cfs_atomic_dec_and_test(&los->los_refcount))
 887                 return;
 888
 889         LASSERT(env);
 890         LASSERT(los->los_dev);
 891         ls = dt2ls_dev(los->los_dev);
 892
 893         mutex_lock(&ls->ls_los_mutex);
 894         if (cfs_atomic_read(&los->los_refcount) == 0) {
 895                 if (los->los_obj)
 896                         lu_object_put_nocache(env, &los->los_obj->do_lu);
 897                 cfs_list_del(&los->los_list);
 898                 OBD_FREE_PTR(los);
 899         }
 900         mutex_unlock(&ls->ls_los_mutex);
 901         ls_device_put(env, ls);
 902 }
 903 EXPORT_SYMBOL(local_oid_storage_fini);