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                 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         obd_id                   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 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, struct lu_attr *attr,
 300                                       struct dt_object_format *dof)
 301 {
 302         struct dt_thread_info   *dti    = dt_info(env);
 303         struct lu_object_conf   *conf   = &dti->dti_conf;
 304         struct dt_object        *dto;
 305         struct thandle          *th;
 306         int                      rc;
 307
 308         /* We know that the target object does not exist, to be created,
 309          * then give some hints - LOC_F_NEW to help low layer to handle
 310          * that efficiently and properly. */
 311         memset(conf, 0, sizeof(*conf));
 312         conf->loc_flags = LOC_F_NEW;
 313         dto = ls_locate(env, ls, fid, conf);
 314         if (unlikely(IS_ERR(dto)))
 315                 RETURN(dto);
 316
 317         LASSERT(dto != NULL);
 318         if (dt_object_exists(dto))
 319                 GOTO(out, rc = -EEXIST);
 320
 321         th = dt_trans_create(env, ls->ls_osd);
 322         if (IS_ERR(th))
 323                 GOTO(out, rc = PTR_ERR(th));
 324
 325         rc = local_object_declare_create(env, los, dto, attr, dof, th);
 326         if (rc)
 327                 GOTO(trans_stop, rc);
 328
 329         if (dti->dti_dof.dof_type == DFT_DIR) {
 330                 dt_declare_ref_add(env, dto, th);
 331                 dt_declare_ref_add(env, parent, th);
 332         }
 333
 334         rc = dt_declare_insert(env, parent, (void *)fid, (void *)name, th);
 335         if (rc)
 336                 GOTO(trans_stop, rc);
 337
 338         rc = dt_trans_start_local(env, ls->ls_osd, th);
 339         if (rc)
 340                 GOTO(trans_stop, rc);
 341
 342         dt_write_lock(env, dto, 0);
 343         if (dt_object_exists(dto))
 344                 GOTO(unlock, rc = 0);
 345
 346         CDEBUG(D_OTHER, "create new object "DFID"\n",
 347                PFID(lu_object_fid(&dto->do_lu)));
 348         rc = local_object_create(env, los, dto, attr, dof, th);
 349         if (rc)
 350                 GOTO(unlock, rc);
 351         LASSERT(dt_object_exists(dto));
 352
 353         if (dti->dti_dof.dof_type == DFT_DIR) {
 354                 if (!dt_try_as_dir(env, dto))
 355                         GOTO(destroy, rc = -ENOTDIR);
 356                 /* Add "." and ".." for newly created dir */
 357                 rc = dt_insert(env, dto, (void *)fid, (void *)".", th,
 358                                BYPASS_CAPA, 1);
 359                 if (rc)
 360                         GOTO(destroy, rc);
 361                 dt_ref_add(env, dto, th);
 362                 rc = dt_insert(env, dto, (void *)lu_object_fid(&parent->do_lu),
 363                                (void *)"..", th, BYPASS_CAPA, 1);
 364                 if (rc)
 365                         GOTO(destroy, rc);
 366         }
 367
 368         dt_write_lock(env, parent, 0);
 369         rc = dt_insert(env, parent, (const struct dt_rec *)fid,
 370                        (const struct dt_key *)name, th, BYPASS_CAPA, 1);
 371         if (dti->dti_dof.dof_type == DFT_DIR)
 372                 dt_ref_add(env, parent, th);
 373         dt_write_unlock(env, parent);
 374         if (rc)
 375                 GOTO(destroy, rc);
 376 destroy:
 377         if (rc)
 378                 dt_destroy(env, dto, th);
 379 unlock:
 380         dt_write_unlock(env, dto);
 381 trans_stop:
 382         dt_trans_stop(env, ls->ls_osd, th);
 383 out:
 384         if (rc) {
 385                 lu_object_put_nocache(env, &dto->do_lu);
 386                 dto = ERR_PTR(rc);
 387         }
 388         RETURN(dto);
 389 }
 390
 391 /*
 392  * Look up and create (if it does not exist) a local named file or directory in
 393  * parent directory.
 394  */
 395 struct dt_object *local_file_find_or_create(const struct lu_env *env,
 396                                             struct local_oid_storage *los,
 397                                             struct dt_object *parent,
 398                                             const char *name, __u32 mode)
 399 {
 400         struct dt_thread_info   *dti = dt_info(env);
 401         struct dt_object        *dto;
 402         int                      rc;
 403
 404         LASSERT(parent);
 405
 406         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 407         if (rc == 0)
 408                 /* name is found, get the object */
 409                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 410                                 &dti->dti_fid, NULL);
 411         else if (rc != -ENOENT)
 412                 dto = ERR_PTR(rc);
 413         else {
 414                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 415                 if (rc < 0) {
 416                         dto = ERR_PTR(rc);
 417                 } else {
 418                         /* create the object */
 419                         dti->dti_attr.la_valid  = LA_MODE;
 420                         dti->dti_attr.la_mode   = mode;
 421                         dti->dti_dof.dof_type   = dt_mode_to_dft(mode & S_IFMT);
 422                         dto = __local_file_create(env, &dti->dti_fid, los,
 423                                                   dt2ls_dev(los->los_dev),
 424                                                   parent, name, &dti->dti_attr,
 425                                                   &dti->dti_dof);
 426                 }
 427         }
 428         return dto;
 429 }
 430 EXPORT_SYMBOL(local_file_find_or_create);
 431
 432 struct dt_object *local_file_find_or_create_with_fid(const struct lu_env *env,
 433                                                      struct dt_device *dt,
 434                                                      const struct lu_fid *fid,
 435                                                      struct dt_object *parent,
 436                                                      const char *name,
 437                                                      __u32 mode)
 438 {
 439         struct dt_thread_info   *dti = dt_info(env);
 440         struct dt_object        *dto;
 441         int                      rc;
 442
 443         LASSERT(parent);
 444
 445         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 446         if (rc == 0) {
 447                 dto = dt_locate(env, dt, &dti->dti_fid);
 448         } else if (rc != -ENOENT) {
 449                 dto = ERR_PTR(rc);
 450         } else {
 451                 struct ls_device *ls;
 452
 453                 ls = ls_device_get(dt);
 454                 if (IS_ERR(ls)) {
 455                         dto = ERR_PTR(PTR_ERR(ls));
 456                 } else {
 457                         /* create the object */
 458                         dti->dti_attr.la_valid  = LA_MODE;
 459                         dti->dti_attr.la_mode   = mode;
 460                         dti->dti_dof.dof_type   = dt_mode_to_dft(mode & S_IFMT);
 461                         dto = __local_file_create(env, fid, NULL, ls, parent,
 462                                                   name, &dti->dti_attr,
 463                                                   &dti->dti_dof);
 464                         /* ls_device_put() will finalize the ls device, we
 465                          * have to open the object in other device stack */
 466                         if (!IS_ERR(dto)) {
 467                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 468                                 lu_object_put_nocache(env, &dto->do_lu);
 469                                 dto = dt_locate(env, dt, &dti->dti_fid);
 470                         }
 471                         ls_device_put(env, ls);
 472                 }
 473         }
 474         return dto;
 475 }
 476 EXPORT_SYMBOL(local_file_find_or_create_with_fid);
 477
 478 /*
 479  * Look up and create (if it does not exist) a local named index file in parent
 480  * directory.
 481  */
 482 struct dt_object *local_index_find_or_create(const struct lu_env *env,
 483                                              struct local_oid_storage *los,
 484                                              struct dt_object *parent,
 485                                              const char *name, __u32 mode,
 486                                              const struct dt_index_features *ft)
 487 {
 488         struct dt_thread_info   *dti = dt_info(env);
 489         struct dt_object        *dto;
 490         int                      rc;
 491
 492         LASSERT(parent);
 493
 494         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 495         if (rc == 0) {
 496                 /* name is found, get the object */
 497                 dto = ls_locate(env, dt2ls_dev(los->los_dev),
 498                                 &dti->dti_fid, NULL);
 499         } else if (rc != -ENOENT) {
 500                 dto = ERR_PTR(rc);
 501         } else {
 502                 rc = local_object_fid_generate(env, los, &dti->dti_fid);
 503                 if (rc < 0) {
 504                         dto = ERR_PTR(rc);
 505                 } else {
 506                         /* create the object */
 507                         dti->dti_attr.la_valid          = LA_MODE;
 508                         dti->dti_attr.la_mode           = mode;
 509                         dti->dti_dof.dof_type           = DFT_INDEX;
 510                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 511                         dto = __local_file_create(env, &dti->dti_fid, los,
 512                                                   dt2ls_dev(los->los_dev),
 513                                                   parent, name, &dti->dti_attr,
 514                                                   &dti->dti_dof);
 515                 }
 516         }
 517         return dto;
 518
 519 }
 520 EXPORT_SYMBOL(local_index_find_or_create);
 521
 522 struct dt_object *
 523 local_index_find_or_create_with_fid(const struct lu_env *env,
 524                                     struct dt_device *dt,
 525                                     const struct lu_fid *fid,
 526                                     struct dt_object *parent,
 527                                     const char *name, __u32 mode,
 528                                     const struct dt_index_features *ft)
 529 {
 530         struct dt_thread_info   *dti = dt_info(env);
 531         struct dt_object        *dto;
 532         int                      rc;
 533
 534         LASSERT(parent);
 535
 536         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 537         if (rc == 0) {
 538                 /* name is found, get the object */
 539                 if (!lu_fid_eq(fid, &dti->dti_fid))
 540                         dto = ERR_PTR(-EINVAL);
 541                 else
 542                         dto = dt_locate(env, dt, fid);
 543         } else if (rc != -ENOENT) {
 544                 dto = ERR_PTR(rc);
 545         } else {
 546                 struct ls_device *ls;
 547
 548                 ls = ls_device_get(dt);
 549                 if (IS_ERR(ls)) {
 550                         dto = ERR_PTR(PTR_ERR(ls));
 551                 } else {
 552                         /* create the object */
 553                         dti->dti_attr.la_valid          = LA_MODE;
 554                         dti->dti_attr.la_mode           = mode;
 555                         dti->dti_dof.dof_type           = DFT_INDEX;
 556                         dti->dti_dof.u.dof_idx.di_feat  = ft;
 557                         dto = __local_file_create(env, fid, NULL, ls, parent,
 558                                                   name, &dti->dti_attr,
 559                                                   &dti->dti_dof);
 560                         /* ls_device_put() will finalize the ls device, we
 561                          * have to open the object in other device stack */
 562                         if (!IS_ERR(dto)) {
 563                                 dti->dti_fid = dto->do_lu.lo_header->loh_fid;
 564                                 lu_object_put_nocache(env, &dto->do_lu);
 565                                 dto = dt_locate(env, dt, &dti->dti_fid);
 566                         }
 567                         ls_device_put(env, ls);
 568                 }
 569         }
 570         return dto;
 571 }
 572 EXPORT_SYMBOL(local_index_find_or_create_with_fid);
 573
 574 static int local_object_declare_unlink(const struct lu_env *env,
 575                                        struct dt_device *dt,
 576                                        struct dt_object *p,
 577                                        struct dt_object *c, const char *name,
 578                                        struct thandle *th)
 579 {
 580         int rc;
 581
 582         rc = dt_declare_delete(env, p, (const struct dt_key *)name, th);
 583         if (rc < 0)
 584                 return rc;
 585
 586         rc = dt_declare_ref_del(env, c, th);
 587         if (rc < 0)
 588                 return rc;
 589
 590         return dt_declare_destroy(env, c, th);
 591 }
 592
 593 int local_object_unlink(const struct lu_env *env, struct dt_device *dt,
 594                         struct dt_object *parent, const char *name)
 595 {
 596         struct dt_thread_info   *dti = dt_info(env);
 597         struct dt_object        *dto;
 598         struct thandle          *th;
 599         int                      rc;
 600
 601         ENTRY;
 602
 603         rc = dt_lookup_dir(env, parent, name, &dti->dti_fid);
 604         if (rc == -ENOENT)
 605                 RETURN(0);
 606         else if (rc < 0)
 607                 RETURN(rc);
 608
 609         dto = dt_locate(env, dt, &dti->dti_fid);
 610         if (unlikely(IS_ERR(dto)))
 611                 RETURN(PTR_ERR(dto));
 612
 613         th = dt_trans_create(env, dt);
 614         if (IS_ERR(th))
 615                 GOTO(out, rc = PTR_ERR(th));
 616
 617         rc = local_object_declare_unlink(env, dt, parent, dto, name, th);
 618         if (rc < 0)
 619                 GOTO(stop, rc);
 620
 621         rc = dt_trans_start_local(env, dt, th);
 622         if (rc < 0)
 623                 GOTO(stop, rc);
 624
 625         dt_write_lock(env, dto, 0);
 626         rc = dt_delete(env, parent, (struct dt_key *)name, th, BYPASS_CAPA);
 627         if (rc < 0)
 628                 GOTO(unlock, rc);
 629
 630         rc = dt_ref_del(env, dto, th);
 631         if (rc < 0) {
 632                 rc = dt_insert(env, parent,
 633                                (const struct dt_rec *)&dti->dti_fid,
 634                                (const struct dt_key *)name, th, BYPASS_CAPA, 1);
 635                 GOTO(unlock, rc);
 636         }
 637
 638         rc = dt_destroy(env, dto, th);
 639 unlock:
 640         dt_write_unlock(env, dto);
 641 stop:
 642         dt_trans_stop(env, dt, th);
 643 out:
 644         lu_object_put_nocache(env, &dto->do_lu);
 645         return rc;
 646 }
 647 EXPORT_SYMBOL(local_object_unlink);
 648
 649 struct local_oid_storage *dt_los_find(struct ls_device *ls, __u64 seq)
 650 {
 651         struct local_oid_storage *los, *ret = NULL;
 652
 653         cfs_list_for_each_entry(los, &ls->ls_los_list, los_list) {
 654                 if (los->los_seq == seq) {
 655                         atomic_inc(&los->los_refcount);
 656                         ret = los;
 657                         break;
 658                 }
 659         }
 660         return ret;
 661 }
 662
 663 void dt_los_put(struct local_oid_storage *los)
 664 {
 665         if (atomic_dec_and_test(&los->los_refcount))
 666                 /* should never happen, only local_oid_storage_fini should
 667                  * drop refcount to zero */
 668                 LBUG();
 669         return;
 670 }
 671
 672 /* after Lustre 2.3 release there may be old file to store last generated FID
 673  * If such file exists then we have to read its content
 674  */
 675 int lastid_compat_check(const struct lu_env *env, struct dt_device *dev,
 676                         __u64 lastid_seq, __u32 *first_oid, struct ls_device *ls)
 677 {
 678         struct dt_thread_info   *dti = dt_info(env);
 679         struct dt_object        *root = NULL;
 680         struct los_ondisk        losd;
 681         struct dt_object        *o = NULL;
 682         int                      rc = 0;
 683
 684         rc = dt_root_get(env, dev, &dti->dti_fid);
 685         if (rc)
 686                 return rc;
 687
 688         root = ls_locate(env, ls, &dti->dti_fid, NULL);
 689         if (IS_ERR(root))
 690                 return PTR_ERR(root);
 691
 692         /* find old last_id file */
 693         snprintf(dti->dti_buf, sizeof(dti->dti_buf), "seq-"LPX64"-lastid",
 694                  lastid_seq);
 695         rc = dt_lookup_dir(env, root, dti->dti_buf, &dti->dti_fid);
 696         lu_object_put_nocache(env, &root->do_lu);
 697         if (rc == -ENOENT) {
 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                 o = ls_locate(env, ls, &dti->dti_fid, NULL);
 705                 if (IS_ERR(o))
 706                         return PTR_ERR(o);
 707
 708                 if (!dt_object_exists(o)) {
 709                         lu_object_put_nocache(env, &o->do_lu);
 710                         return 0;
 711                 }
 712                 CDEBUG(D_INFO, "Found old llog lastid file\n");
 713         } else if (rc < 0) {
 714                 return rc;
 715         } else {
 716                 CDEBUG(D_INFO, "Found old lastid file for sequence "LPX64"\n",
 717                        lastid_seq);
 718                 o = ls_locate(env, ls, &dti->dti_fid, NULL);
 719                 if (IS_ERR(o))
 720                         return PTR_ERR(o);
 721         }
 722         /* let's read seq-NNNNNN-lastid file value */
 723         LASSERT(dt_object_exists(o));
 724         dti->dti_off = 0;
 725         dti->dti_lb.lb_buf = &losd;
 726         dti->dti_lb.lb_len = sizeof(losd);
 727         dt_read_lock(env, o, 0);
 728         rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 729         dt_read_unlock(env, o);
 730         if (rc == 0 && le32_to_cpu(losd.lso_magic) != LOS_MAGIC) {
 731                 CERROR("%s: wrong content of seq-"LPX64"-lastid file, magic %x\n",
 732                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq,
 733                        le32_to_cpu(losd.lso_magic));
 734                 rc = -EINVAL;
 735         } else if (rc < 0) {
 736                 CERROR("%s: failed to read seq-"LPX64"-lastid: rc = %d\n",
 737                        o->do_lu.lo_dev->ld_obd->obd_name, lastid_seq, rc);
 738         }
 739         lu_object_put_nocache(env, &o->do_lu);
 740         if (rc == 0)
 741                 *first_oid = le32_to_cpu(losd.lso_next_oid);
 742         return rc;
 743 }
 744
 745 /**
 746  * Initialize local OID storage for required sequence.
 747  * That may be needed for services that uses local files and requires
 748  * dynamic OID allocation for them.
 749  *
 750  * Per each sequence we have an object with 'first_fid' identificator
 751  * containing the counter for OIDs of locally created files with that
 752  * sequence.
 753  *
 754  * It is used now by llog subsystem and MGS for NID tables
 755  *
 756  * Function gets first_fid to create counter object.
 757  * All dynamic fids will be generated with the same sequence and incremented
 758  * OIDs
 759  *
 760  * Returned local_oid_storage is in-memory representaion of OID storage
 761  */
 762 int local_oid_storage_init(const struct lu_env *env, struct dt_device *dev,
 763                            const struct lu_fid *first_fid,
 764                            struct local_oid_storage **los)
 765 {
 766         struct dt_thread_info   *dti = dt_info(env);
 767         struct ls_device        *ls;
 768         obd_id                   lastid;
 769         struct dt_object        *o = NULL;
 770         struct thandle          *th;
 771         __u32                    first_oid = fid_oid(first_fid);
 772         int                      rc = 0;
 773
 774         ENTRY;
 775
 776         ls = ls_device_get(dev);
 777         if (IS_ERR(ls))
 778                 RETURN(PTR_ERR(ls));
 779
 780         mutex_lock(&ls->ls_los_mutex);
 781         *los = dt_los_find(ls, fid_seq(first_fid));
 782         if (*los != NULL)
 783                 GOTO(out, rc = 0);
 784
 785         /* not found, then create */
 786         OBD_ALLOC_PTR(*los);
 787         if (*los == NULL)
 788                 GOTO(out, rc = -ENOMEM);
 789
 790         atomic_set(&(*los)->los_refcount, 1);
 791         mutex_init(&(*los)->los_id_lock);
 792         (*los)->los_dev = &ls->ls_top_dev;
 793         atomic_inc(&ls->ls_refcount);
 794         cfs_list_add(&(*los)->los_list, &ls->ls_los_list);
 795
 796         /* Use {seq, 0, 0} to create the LAST_ID file for every
 797          * sequence.  OIDs start at LUSTRE_FID_INIT_OID.
 798          */
 799         dti->dti_fid.f_seq = fid_seq(first_fid);
 800         dti->dti_fid.f_oid = LUSTRE_FID_LASTID_OID;
 801         dti->dti_fid.f_ver = 0;
 802         o = ls_locate(env, ls, &dti->dti_fid, NULL);
 803         if (IS_ERR(o))
 804                 GOTO(out_los, rc = PTR_ERR(o));
 805
 806         if (!dt_object_exists(o)) {
 807                 rc = lastid_compat_check(env, dev, fid_seq(first_fid),
 808                                          &first_oid, ls);
 809                 if (rc < 0)
 810                         GOTO(out_los, rc);
 811
 812                 th = dt_trans_create(env, dev);
 813                 if (IS_ERR(th))
 814                         GOTO(out_los, rc = PTR_ERR(th));
 815
 816                 dti->dti_attr.la_valid = LA_MODE | LA_TYPE;
 817                 dti->dti_attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
 818                 dti->dti_dof.dof_type = dt_mode_to_dft(S_IFREG);
 819
 820                 rc = dt_declare_create(env, o, &dti->dti_attr, NULL,
 821                                        &dti->dti_dof, th);
 822                 if (rc)
 823                         GOTO(out_trans, rc);
 824
 825                 lastid = cpu_to_le64(first_oid);
 826
 827                 dti->dti_off = 0;
 828                 dti->dti_lb.lb_buf = &lastid;
 829                 dti->dti_lb.lb_len = sizeof(lastid);
 830                 rc = dt_declare_record_write(env, o, &dti->dti_lb, dti->dti_off,
 831                                              th);
 832                 if (rc)
 833                         GOTO(out_trans, rc);
 834
 835                 rc = dt_trans_start_local(env, dev, th);
 836                 if (rc)
 837                         GOTO(out_trans, rc);
 838
 839                 dt_write_lock(env, o, 0);
 840                 if (dt_object_exists(o))
 841                         GOTO(out_lock, rc = 0);
 842
 843                 rc = dt_create(env, o, &dti->dti_attr, NULL, &dti->dti_dof,
 844                                th);
 845                 if (rc)
 846                         GOTO(out_lock, rc);
 847
 848                 rc = dt_record_write(env, o, &dti->dti_lb, &dti->dti_off, th);
 849                 if (rc)
 850                         GOTO(out_lock, rc);
 851 out_lock:
 852                 dt_write_unlock(env, o);
 853 out_trans:
 854                 dt_trans_stop(env, dev, th);
 855         } else {
 856                 dti->dti_off = 0;
 857                 dti->dti_lb.lb_buf = &lastid;
 858                 dti->dti_lb.lb_len = sizeof(lastid);
 859                 dt_read_lock(env, o, 0);
 860                 rc = dt_record_read(env, o, &dti->dti_lb, &dti->dti_off);
 861                 dt_read_unlock(env, o);
 862                 if (rc == 0 && le64_to_cpu(lastid) > OBIF_MAX_OID) {
 863                         CERROR("%s: bad oid "LPU64" is read from LAST_ID\n",
 864                                o->do_lu.lo_dev->ld_obd->obd_name,
 865                                le64_to_cpu(lastid));
 866                         rc = -EINVAL;
 867                 }
 868         }
 869 out_los:
 870         if (rc != 0) {
 871                 cfs_list_del(&(*los)->los_list);
 872                 atomic_dec(&ls->ls_refcount);
 873                 OBD_FREE_PTR(*los);
 874                 *los = NULL;
 875                 if (o != NULL && !IS_ERR(o))
 876                         lu_object_put_nocache(env, &o->do_lu);
 877         } else {
 878                 (*los)->los_seq = fid_seq(first_fid);
 879                 (*los)->los_last_oid = le64_to_cpu(lastid);
 880                 (*los)->los_obj = o;
 881                 /* Read value should not be less than initial one
 882                  * but possible after upgrade from older fs.
 883                  * In this case just switch to the first_oid in memory and
 884                  * it will be updated on disk with first object generated */
 885                 if ((*los)->los_last_oid < first_oid)
 886                         (*los)->los_last_oid = first_oid;
 887         }
 888 out:
 889         mutex_unlock(&ls->ls_los_mutex);
 890         ls_device_put(env, ls);
 891         return rc;
 892 }
 893 EXPORT_SYMBOL(local_oid_storage_init);
 894
 895 void local_oid_storage_fini(const struct lu_env *env,
 896                             struct local_oid_storage *los)
 897 {
 898         struct ls_device *ls;
 899
 900         LASSERT(env);
 901         LASSERT(los->los_dev);
 902         ls = dt2ls_dev(los->los_dev);
 903
 904         /* Take the mutex before decreasing the reference to avoid race
 905          * conditions as described in LU-4721. */
 906         mutex_lock(&ls->ls_los_mutex);
 907         if (!atomic_dec_and_test(&los->los_refcount)) {
 908                 mutex_unlock(&ls->ls_los_mutex);
 909                 return;
 910         }
 911
 912         if (los->los_obj)
 913                 lu_object_put_nocache(env, &los->los_obj->do_lu);
 914         cfs_list_del(&los->los_list);
 915         OBD_FREE_PTR(los);
 916         mutex_unlock(&ls->ls_los_mutex);
 917         ls_device_put(env, ls);
 918 }
 919 EXPORT_SYMBOL(local_oid_storage_fini);