lustre/lod/lod_lov.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  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright  2009 Sun Microsystems, Inc. All rights reserved
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * lustre/lod/lod_lov.c
  30  *
  31  * A set of helpers to maintain Logical Object Volume (LOV)
  32  * Extended Attribute (EA) and known OST targets
  33  *
  34  * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_MDS
  38
  39 #include <obd_class.h>
  40 #include <lustre_lfsck.h>
  41 #include <lustre_lmv.h>
  42 #include <lustre_swab.h>
  43
  44 #include "lod_internal.h"
  45
  46 /**
  47  * Increase reference count on the target table.
  48  *
  49  * Increase reference count on the target table usage to prevent racing with
  50  * addition/deletion. Any function that expects the table to remain
  51  * stationary must take a ref.
  52  *
  53  * \param[in] ltd       target table (lod_ost_descs or lod_mdt_descs)
  54  */
  55 void lod_getref(struct lod_tgt_descs *ltd)
  56 {
  57         down_read(&ltd->ltd_rw_sem);
  58         mutex_lock(&ltd->ltd_mutex);
  59         ltd->ltd_refcount++;
  60         mutex_unlock(&ltd->ltd_mutex);
  61 }
  62
  63 /**
  64  * Decrease reference count on the target table.
  65  *
  66  * Companion of lod_getref() to release a reference on the target table.
  67  * If this is the last reference and the OST entry was scheduled for deletion,
  68  * the descriptor is removed from the table.
  69  *
  70  * \param[in] lod       LOD device from which we release a reference
  71  * \param[in] ltd       target table (lod_ost_descs or lod_mdt_descs)
  72  */
  73 void lod_putref(struct lod_device *lod, struct lod_tgt_descs *ltd)
  74 {
  75         mutex_lock(&ltd->ltd_mutex);
  76         ltd->ltd_refcount--;
  77         if (ltd->ltd_refcount == 0 && ltd->ltd_death_row) {
  78                 struct lod_tgt_desc *tgt_desc, *tmp;
  79                 LIST_HEAD(kill);
  80
  81                 CDEBUG(D_CONFIG, "destroying %d ltd desc\n",
  82                        ltd->ltd_death_row);
  83
  84                 ltd_foreach_tgt_safe(ltd, tgt_desc, tmp) {
  85                         LASSERT(tgt_desc);
  86                         if (!tgt_desc->ltd_reap)
  87                                 continue;
  88
  89                         list_add(&tgt_desc->ltd_kill, &kill);
  90                         tgt_pool_remove(&ltd->ltd_tgt_pool,
  91                                         tgt_desc->ltd_index);
  92                         ltd_del_tgt(ltd, tgt_desc);
  93                         ltd->ltd_death_row--;
  94                 }
  95                 mutex_unlock(&ltd->ltd_mutex);
  96                 up_read(&ltd->ltd_rw_sem);
  97
  98                 list_for_each_entry_safe(tgt_desc, tmp, &kill, ltd_kill) {
  99                         int rc;
 100
 101                         list_del(&tgt_desc->ltd_kill);
 102                         rc = obd_disconnect(tgt_desc->ltd_exp);
 103                         if (rc)
 104                                 CERROR("%s: failed to disconnect %s: rc = %d\n",
 105                                        lod2obd(lod)->obd_name,
 106                                        obd_uuid2str(&tgt_desc->ltd_uuid), rc);
 107                         OBD_FREE_PTR(tgt_desc);
 108                 }
 109         } else {
 110                 mutex_unlock(&ltd->ltd_mutex);
 111                 up_read(&ltd->ltd_rw_sem);
 112         }
 113 }
 114
 115 /**
 116  * Connect LOD to a new OSP and add it to the target table.
 117  *
 118  * Connect to the OSP device passed, initialize all the internal
 119  * structures related to the device and add it to the target table.
 120  *
 121  * \param[in] env               execution environment for this thread
 122  * \param[in] lod               LOD device to be connected to the new OSP
 123  * \param[in] osp               name of OSP device name to be added
 124  * \param[in] index             index of the new target
 125  * \param[in] gen               target's generation number
 126  * \param[in] tgt_index         OSP's group
 127  * \param[in] type              type of device (mdc or osc)
 128  * \param[in] active            state of OSP: 0 - inactive, 1 - active
 129  *
 130  * \retval                      0 if added successfully
 131  * \retval                      negative error number on failure
 132  */
 133 int lod_add_device(const struct lu_env *env, struct lod_device *lod,
 134                    char *osp, unsigned index, unsigned gen, int tgt_index,
 135                    char *type, int active)
 136 {
 137         struct obd_connect_data *data = NULL;
 138         struct obd_export       *exp = NULL;
 139         struct obd_device       *obd;
 140         struct lu_device        *lu_dev;
 141         struct dt_device        *dt_dev;
 142         int                      rc;
 143         struct lod_tgt_desc     *tgt_desc;
 144         struct lod_tgt_descs    *ltd;
 145         struct lustre_cfg       *lcfg;
 146         struct obd_uuid         obd_uuid;
 147         bool                    for_ost;
 148         bool connected = false;
 149         ENTRY;
 150
 151         CDEBUG(D_CONFIG, "osp:%s idx:%d gen:%d\n", osp, index, gen);
 152
 153         if (gen <= 0) {
 154                 CERROR("request to add OBD %s with invalid generation: %d\n",
 155                        osp, gen);
 156                 RETURN(-EINVAL);
 157         }
 158
 159         obd_str2uuid(&obd_uuid, osp);
 160
 161         obd = class_find_client_obd(&obd_uuid, LUSTRE_OSP_NAME,
 162                                 &lod->lod_dt_dev.dd_lu_dev.ld_obd->obd_uuid);
 163         if (obd == NULL) {
 164                 CERROR("can't find %s device\n", osp);
 165                 RETURN(-EINVAL);
 166         }
 167
 168         LASSERT(obd->obd_lu_dev != NULL);
 169         LASSERT(obd->obd_lu_dev->ld_site == lod->lod_dt_dev.dd_lu_dev.ld_site);
 170
 171         lu_dev = obd->obd_lu_dev;
 172         dt_dev = lu2dt_dev(lu_dev);
 173
 174         OBD_ALLOC_PTR(data);
 175         if (data == NULL)
 176                 GOTO(out_cleanup, rc = -ENOMEM);
 177
 178         data->ocd_connect_flags = OBD_CONNECT_INDEX | OBD_CONNECT_VERSION;
 179         data->ocd_version = LUSTRE_VERSION_CODE;
 180         data->ocd_index = index;
 181
 182         if (strcmp(LUSTRE_OSC_NAME, type) == 0) {
 183                 for_ost = true;
 184                 data->ocd_connect_flags |= OBD_CONNECT_AT |
 185                                            OBD_CONNECT_FULL20 |
 186                                            OBD_CONNECT_INDEX |
 187 #ifdef HAVE_LRU_RESIZE_SUPPORT
 188                                            OBD_CONNECT_LRU_RESIZE |
 189 #endif
 190                                            OBD_CONNECT_MDS |
 191                                            OBD_CONNECT_REQPORTAL |
 192                                            OBD_CONNECT_SKIP_ORPHAN |
 193                                            OBD_CONNECT_FID |
 194                                            OBD_CONNECT_LVB_TYPE |
 195                                            OBD_CONNECT_VERSION |
 196                                            OBD_CONNECT_PINGLESS |
 197                                            OBD_CONNECT_LFSCK |
 198                                            OBD_CONNECT_BULK_MBITS;
 199
 200                 data->ocd_group = tgt_index;
 201                 ltd = &lod->lod_ost_descs;
 202         } else {
 203                 struct obd_import *imp = obd->u.cli.cl_import;
 204
 205                 for_ost = false;
 206                 data->ocd_ibits_known = MDS_INODELOCK_UPDATE;
 207                 data->ocd_connect_flags |= OBD_CONNECT_ACL |
 208                                            OBD_CONNECT_IBITS |
 209                                            OBD_CONNECT_MDS_MDS |
 210                                            OBD_CONNECT_FID |
 211                                            OBD_CONNECT_AT |
 212                                            OBD_CONNECT_FULL20 |
 213                                            OBD_CONNECT_LFSCK |
 214                                            OBD_CONNECT_BULK_MBITS;
 215                 spin_lock(&imp->imp_lock);
 216                 imp->imp_server_timeout = 1;
 217                 spin_unlock(&imp->imp_lock);
 218                 imp->imp_client->cli_request_portal = OUT_PORTAL;
 219                 CDEBUG(D_OTHER, "%s: Set 'mds' portal and timeout\n",
 220                       obd->obd_name);
 221                 ltd = &lod->lod_mdt_descs;
 222         }
 223
 224         rc = obd_connect(env, &exp, obd, &obd->obd_uuid, data, NULL);
 225         OBD_FREE_PTR(data);
 226         if (rc) {
 227                 CERROR("%s: cannot connect to next dev %s (%d)\n",
 228                        obd->obd_name, osp, rc);
 229                 GOTO(out_cleanup, rc);
 230         }
 231         connected = true;
 232
 233         /* Allocate ost descriptor and fill it */
 234         OBD_ALLOC_PTR(tgt_desc);
 235         if (!tgt_desc)
 236                 GOTO(out_cleanup, rc = -ENOMEM);
 237
 238         tgt_desc->ltd_tgt    = dt_dev;
 239         tgt_desc->ltd_exp    = exp;
 240         tgt_desc->ltd_uuid   = obd->u.cli.cl_target_uuid;
 241         tgt_desc->ltd_gen    = gen;
 242         tgt_desc->ltd_index  = index;
 243         tgt_desc->ltd_active = active;
 244
 245         down_write(&ltd->ltd_rw_sem);
 246         mutex_lock(&ltd->ltd_mutex);
 247         rc = ltd_add_tgt(ltd, tgt_desc);
 248         if (rc)
 249                 GOTO(out_mutex, rc);
 250
 251         rc = lu_qos_add_tgt(&ltd->ltd_qos, tgt_desc);
 252         if (rc)
 253                 GOTO(out_del_tgt, rc);
 254
 255         rc = tgt_pool_add(&ltd->ltd_tgt_pool, index,
 256                           ltd->ltd_lov_desc.ld_tgt_count);
 257         if (rc) {
 258                 CERROR("%s: can't set up pool, failed with %d\n",
 259                        obd->obd_name, rc);
 260                 GOTO(out_del_tgt, rc);
 261         }
 262
 263         mutex_unlock(&ltd->ltd_mutex);
 264         up_write(&ltd->ltd_rw_sem);
 265
 266         if (lod->lod_recovery_completed)
 267                 lu_dev->ld_ops->ldo_recovery_complete(env, lu_dev);
 268
 269         if (!for_ost && lod->lod_initialized) {
 270                 rc = lod_sub_init_llog(env, lod, tgt_desc->ltd_tgt);
 271                 if (rc != 0) {
 272                         CERROR("%s: cannot start llog on %s:rc = %d\n",
 273                                lod2obd(lod)->obd_name, osp, rc);
 274                         GOTO(out_ltd, rc);
 275                 }
 276         }
 277
 278         rc = lfsck_add_target(env, lod->lod_child, dt_dev, exp, index, for_ost);
 279         if (rc != 0) {
 280                 CERROR("Fail to add LFSCK target: name = %s, type = %s, "
 281                        "index = %u, rc = %d\n", osp, type, index, rc);
 282                 GOTO(out_fini_llog, rc);
 283         }
 284         RETURN(rc);
 285 out_fini_llog:
 286         lod_sub_fini_llog(env, tgt_desc->ltd_tgt,
 287                           &tgt_desc->ltd_recovery_task);
 288 out_ltd:
 289         down_write(&ltd->ltd_rw_sem);
 290         mutex_lock(&ltd->ltd_mutex);
 291         tgt_pool_remove(&ltd->ltd_tgt_pool, index);
 292 out_del_tgt:
 293         ltd_del_tgt(ltd, tgt_desc);
 294 out_mutex:
 295         mutex_unlock(&ltd->ltd_mutex);
 296         up_write(&ltd->ltd_rw_sem);
 297         OBD_FREE_PTR(tgt_desc);
 298 out_cleanup:
 299         /* XXX OSP needs us to send down LCFG_CLEANUP because it uses
 300          * objects from the MDT stack. See LU-7184. */
 301         lcfg = &lod_env_info(env)->lti_lustre_cfg;
 302         memset(lcfg, 0, sizeof(*lcfg));
 303         lcfg->lcfg_version = LUSTRE_CFG_VERSION;
 304         lcfg->lcfg_command = LCFG_CLEANUP;
 305         lu_dev->ld_ops->ldo_process_config(env, lu_dev, lcfg);
 306
 307         if (connected)
 308                 obd_disconnect(exp);
 309
 310         return rc;
 311 }
 312
 313 /**
 314  * Schedule target removal from the target table.
 315  *
 316  * Mark the device as dead. The device is not removed here because it may
 317  * still be in use. The device will be removed in lod_putref() when the
 318  * last reference is released.
 319  *
 320  * \param[in] env               execution environment for this thread
 321  * \param[in] lod               LOD device the target table belongs to
 322  * \param[in] ltd               target table
 323  * \param[in] tgt               target
 324  */
 325 static void __lod_del_device(const struct lu_env *env, struct lod_device *lod,
 326                              struct lod_tgt_descs *ltd, struct lu_tgt_desc *tgt)
 327 {
 328         lfsck_del_target(env, lod->lod_child, tgt->ltd_tgt, tgt->ltd_index,
 329                          !ltd->ltd_is_mdt);
 330
 331         if (!tgt->ltd_reap) {
 332                 tgt->ltd_reap = 1;
 333                 ltd->ltd_death_row++;
 334         }
 335 }
 336
 337 /**
 338  * Schedule removal of all the targets from the given target table.
 339  *
 340  * See more details in the description for __lod_del_device()
 341  *
 342  * \param[in] env               execution environment for this thread
 343  * \param[in] lod               LOD device the target table belongs to
 344  * \param[in] ltd               target table
 345  *
 346  * \retval                      0 always
 347  */
 348 int lod_fini_tgt(const struct lu_env *env, struct lod_device *lod,
 349                  struct lod_tgt_descs *ltd)
 350 {
 351         struct lu_tgt_desc *tgt;
 352
 353         if (ltd->ltd_tgts_size <= 0)
 354                 return 0;
 355
 356         lod_getref(ltd);
 357         mutex_lock(&ltd->ltd_mutex);
 358         ltd_foreach_tgt(ltd, tgt)
 359                 __lod_del_device(env, lod, ltd, tgt);
 360         mutex_unlock(&ltd->ltd_mutex);
 361         lod_putref(lod, ltd);
 362
 363         lu_tgt_descs_fini(ltd);
 364
 365         return 0;
 366 }
 367
 368 /**
 369  * Remove device by name.
 370  *
 371  * Remove a device identified by \a osp from the target table. Given
 372  * the device can be in use, the real deletion happens in lod_putref().
 373  *
 374  * \param[in] env               execution environment for this thread
 375  * \param[in] lod               LOD device to be connected to the new OSP
 376  * \param[in] ltd               target table
 377  * \param[in] osp               name of OSP device to be removed
 378  * \param[in] idx               index of the target
 379  * \param[in] gen               generation number, not used currently
 380  *
 381  * \retval                      0 if the device was scheduled for removal
 382  * \retval                      -EINVAL if no device was found
 383  */
 384 int lod_del_device(const struct lu_env *env, struct lod_device *lod,
 385                    struct lod_tgt_descs *ltd, char *osp, unsigned int idx,
 386                    unsigned int gen)
 387 {
 388         struct obd_device *obd;
 389         struct lu_tgt_desc *tgt;
 390         struct obd_uuid uuid;
 391         int rc = 0;
 392
 393         ENTRY;
 394
 395         CDEBUG(D_CONFIG, "osp:%s idx:%d gen:%d\n", osp, idx, gen);
 396
 397         obd_str2uuid(&uuid, osp);
 398
 399         obd = class_find_client_obd(&uuid, LUSTRE_OSP_NAME,
 400                                    &lod->lod_dt_dev.dd_lu_dev.ld_obd->obd_uuid);
 401         if (obd == NULL) {
 402                 CERROR("can't find %s device\n", osp);
 403                 RETURN(-EINVAL);
 404         }
 405
 406         if (gen <= 0) {
 407                 CERROR("%s: request to remove OBD %s with invalid generation %d"
 408                        "\n", obd->obd_name, osp, gen);
 409                 RETURN(-EINVAL);
 410         }
 411
 412         obd_str2uuid(&uuid,  osp);
 413
 414         lod_getref(ltd);
 415         mutex_lock(&ltd->ltd_mutex);
 416         tgt = LTD_TGT(ltd, idx);
 417         /* check that the index is allocated in the bitmap */
 418         if (!test_bit(idx, ltd->ltd_tgt_bitmap) || !tgt) {
 419                 CERROR("%s: device %d is not set up\n", obd->obd_name, idx);
 420                 GOTO(out, rc = -EINVAL);
 421         }
 422
 423         /* check that the UUID matches */
 424         if (!obd_uuid_equals(&uuid, &tgt->ltd_uuid)) {
 425                 CERROR("%s: LOD target UUID %s at index %d does not match %s\n",
 426                        obd->obd_name, obd_uuid2str(&tgt->ltd_uuid), idx, osp);
 427                 GOTO(out, rc = -EINVAL);
 428         }
 429
 430         __lod_del_device(env, lod, ltd, tgt);
 431         EXIT;
 432 out:
 433         mutex_unlock(&ltd->ltd_mutex);
 434         lod_putref(lod, ltd);
 435         return(rc);
 436 }
 437
 438 /**
 439  * Resize per-thread storage to hold specified size.
 440  *
 441  * A helper function to resize per-thread temporary storage. This storage
 442  * is used to process LOV/LVM EAs and may be quite large. We do not want to
 443  * allocate/release it every time, so instead we put it into the env and
 444  * reallocate on demand. The memory is released when the correspondent thread
 445  * is finished.
 446  *
 447  * \param[in] info              LOD-specific storage in the environment
 448  * \param[in] size              new size to grow the buffer to
 449
 450  * \retval                      0 on success, -ENOMEM if reallocation failed
 451  */
 452 int lod_ea_store_resize(struct lod_thread_info *info, size_t size)
 453 {
 454         __u32 round = size_roundup_power2(size);
 455
 456         if (info->lti_ea_store) {
 457                 LASSERT(info->lti_ea_store_size);
 458                 LASSERT(info->lti_ea_store_size < round);
 459                 CDEBUG(D_INFO, "EA store size %d is not enough, need %d\n",
 460                        info->lti_ea_store_size, round);
 461                 OBD_FREE_LARGE(info->lti_ea_store, info->lti_ea_store_size);
 462                 info->lti_ea_store = NULL;
 463                 info->lti_ea_store_size = 0;
 464         }
 465
 466         OBD_ALLOC_LARGE(info->lti_ea_store, round);
 467         if (info->lti_ea_store == NULL)
 468                 RETURN(-ENOMEM);
 469         info->lti_ea_store_size = round;
 470
 471         RETURN(0);
 472 }
 473
 474 static void lod_free_comp_buffer(struct lod_layout_component *entries,
 475                                  __u16 count, __u32 bufsize)
 476 {
 477         struct lod_layout_component *entry;
 478         int i;
 479
 480         for (i = 0; i < count; i++) {
 481                 entry = &entries[i];
 482                 if (entry->llc_pool != NULL)
 483                         lod_set_pool(&entry->llc_pool, NULL);
 484                 if (entry->llc_ostlist.op_array)
 485                         OBD_FREE(entry->llc_ostlist.op_array,
 486                                  entry->llc_ostlist.op_size);
 487                 LASSERT(entry->llc_stripe == NULL);
 488                 LASSERT(entry->llc_stripes_allocated == 0);
 489         }
 490
 491         if (bufsize != 0)
 492                 OBD_FREE_LARGE(entries, bufsize);
 493 }
 494
 495 void lod_free_def_comp_entries(struct lod_default_striping *lds)
 496 {
 497         lod_free_comp_buffer(lds->lds_def_comp_entries,
 498                              lds->lds_def_comp_size_cnt,
 499                              size_roundup_power2(
 500                                      sizeof(*lds->lds_def_comp_entries) *
 501                                      lds->lds_def_comp_size_cnt));
 502         lds->lds_def_comp_entries = NULL;
 503         lds->lds_def_comp_cnt = 0;
 504         lds->lds_def_striping_is_composite = 0;
 505         lds->lds_def_comp_size_cnt = 0;
 506 }
 507
 508 /**
 509  * Resize per-thread storage to hold default striping component entries
 510  *
 511  * A helper function to resize per-thread temporary storage. This storage
 512  * is used to hold default LOV/LVM EAs and may be quite large. We do not want
 513  * to allocate/release it every time, so instead we put it into the env and
 514  * reallocate it on demand. The memory is released when the correspondent
 515  * thread is finished.
 516  *
 517  * \param[in,out] lds           default striping
 518  * \param[in] count             new component count to grow the buffer to
 519
 520  * \retval                      0 on success, -ENOMEM if reallocation failed
 521  */
 522 int lod_def_striping_comp_resize(struct lod_default_striping *lds, __u16 count)
 523 {
 524         struct lod_layout_component *entries;
 525         __u32 new = size_roundup_power2(sizeof(*lds->lds_def_comp_entries) *
 526                                         count);
 527         __u32 old = size_roundup_power2(sizeof(*lds->lds_def_comp_entries) *
 528                                         lds->lds_def_comp_size_cnt);
 529
 530         if (new <= old)
 531                 return 0;
 532
 533         OBD_ALLOC_LARGE(entries, new);
 534         if (entries == NULL)
 535                 return -ENOMEM;
 536
 537         if (lds->lds_def_comp_entries != NULL) {
 538                 CDEBUG(D_INFO, "default striping component size %d is not "
 539                        "enough, need %d\n", old, new);
 540                 lod_free_def_comp_entries(lds);
 541         }
 542
 543         lds->lds_def_comp_entries = entries;
 544         lds->lds_def_comp_size_cnt = count;
 545
 546         RETURN(0);
 547 }
 548
 549 void lod_free_comp_entries(struct lod_object *lo)
 550 {
 551         if (lo->ldo_mirrors) {
 552                 OBD_FREE_PTR_ARRAY(lo->ldo_mirrors, lo->ldo_mirror_count);
 553                 lo->ldo_mirrors = NULL;
 554                 lo->ldo_mirror_count = 0;
 555         }
 556         lod_free_comp_buffer(lo->ldo_comp_entries,
 557                              lo->ldo_comp_cnt,
 558                              sizeof(*lo->ldo_comp_entries) * lo->ldo_comp_cnt);
 559         lo->ldo_comp_entries = NULL;
 560         lo->ldo_comp_cnt = 0;
 561         lo->ldo_is_composite = 0;
 562 }
 563
 564 int lod_alloc_comp_entries(struct lod_object *lo,
 565                            int mirror_count, int comp_count)
 566 {
 567         LASSERT(comp_count != 0);
 568         LASSERT(lo->ldo_comp_cnt == 0 && lo->ldo_comp_entries == NULL);
 569
 570         if (mirror_count > 0) {
 571                 OBD_ALLOC_PTR_ARRAY(lo->ldo_mirrors, mirror_count);
 572                 if (!lo->ldo_mirrors)
 573                         return -ENOMEM;
 574
 575                 lo->ldo_mirror_count = mirror_count;
 576         }
 577
 578         OBD_ALLOC_LARGE(lo->ldo_comp_entries,
 579                         sizeof(*lo->ldo_comp_entries) * comp_count);
 580         if (lo->ldo_comp_entries == NULL) {
 581                 OBD_FREE_PTR_ARRAY(lo->ldo_mirrors, mirror_count);
 582                 lo->ldo_mirror_count = 0;
 583                 return -ENOMEM;
 584         }
 585
 586         lo->ldo_comp_cnt = comp_count;
 587         return 0;
 588 }
 589
 590 int lod_fill_mirrors(struct lod_object *lo)
 591 {
 592         struct lod_layout_component *lod_comp;
 593         int mirror_idx = -1;
 594         __u16 mirror_id = 0xffff;
 595         int i;
 596         ENTRY;
 597
 598         LASSERT(equi(!lo->ldo_is_composite, lo->ldo_mirror_count == 0));
 599
 600         if (!lo->ldo_is_composite)
 601                 RETURN(0);
 602
 603         lod_comp = &lo->ldo_comp_entries[0];
 604         for (i = 0; i < lo->ldo_comp_cnt; i++, lod_comp++) {
 605                 int stale = !!(lod_comp->llc_flags & LCME_FL_STALE);
 606                 int preferred = !!(lod_comp->llc_flags & LCME_FL_PREF_WR);
 607
 608                 if (mirror_id_of(lod_comp->llc_id) == mirror_id) {
 609                         lo->ldo_mirrors[mirror_idx].lme_stale |= stale;
 610                         lo->ldo_mirrors[mirror_idx].lme_primary |= preferred;
 611                         lo->ldo_mirrors[mirror_idx].lme_end = i;
 612                         continue;
 613                 }
 614
 615                 /* new mirror */
 616                 ++mirror_idx;
 617                 if (mirror_idx >= lo->ldo_mirror_count)
 618                         RETURN(-EINVAL);
 619
 620                 mirror_id = mirror_id_of(lod_comp->llc_id);
 621
 622                 lo->ldo_mirrors[mirror_idx].lme_id = mirror_id;
 623                 lo->ldo_mirrors[mirror_idx].lme_stale = stale;
 624                 lo->ldo_mirrors[mirror_idx].lme_primary = preferred;
 625                 lo->ldo_mirrors[mirror_idx].lme_start = i;
 626                 lo->ldo_mirrors[mirror_idx].lme_end = i;
 627         }
 628         if (mirror_idx != lo->ldo_mirror_count - 1)
 629                 RETURN(-EINVAL);
 630
 631         RETURN(0);
 632 }
 633
 634 /**
 635  * Generate on-disk lov_mds_md structure for each layout component based on
 636  * the information in lod_object->ldo_comp_entries[i].
 637  *
 638  * \param[in] env               execution environment for this thread
 639  * \param[in] lo                LOD object
 640  * \param[in] comp_idx          index of ldo_comp_entries
 641  * \param[in] lmm               buffer to cotain the on-disk lov_mds_md
 642  * \param[in|out] lmm_size      buffer size/lmm size
 643  * \param[in] is_dir            generate lov ea for dir or file? For dir case,
 644  *                              the stripe info is from the default stripe
 645  *                              template, which is collected in lod_ah_init(),
 646  *                              either from parent object or root object; for
 647  *                              file case, it's from the @lo object
 648  *
 649  * \retval                      0 if on disk structure is created successfully
 650  * \retval                      negative error number on failure
 651  */
 652 static int lod_gen_component_ea(const struct lu_env *env,
 653                                 struct lod_object *lo, int comp_idx,
 654                                 struct lov_mds_md *lmm, int *lmm_size,
 655                                 bool is_dir)
 656 {
 657         struct lod_thread_info  *info = lod_env_info(env);
 658         const struct lu_fid     *fid  = lu_object_fid(&lo->ldo_obj.do_lu);
 659         struct lod_device       *lod;
 660         struct lov_ost_data_v1  *objs;
 661         struct lod_layout_component *lod_comp;
 662         __u32   magic;
 663         __u16 stripe_count;
 664         int     i, rc = 0;
 665         ENTRY;
 666
 667         LASSERT(lo);
 668         if (is_dir)
 669                 lod_comp =
 670                         &lo->ldo_def_striping->lds_def_comp_entries[comp_idx];
 671         else
 672                 lod_comp = &lo->ldo_comp_entries[comp_idx];
 673
 674         magic = lod_comp->llc_pool != NULL ? LOV_MAGIC_V3 : LOV_MAGIC_V1;
 675         if (lod_comp->llc_pattern == 0) /* default striping */
 676                 lod_comp->llc_pattern = LOV_PATTERN_RAID0;
 677
 678         lmm->lmm_magic = cpu_to_le32(magic);
 679         lmm->lmm_pattern = cpu_to_le32(lod_comp->llc_pattern);
 680         fid_to_lmm_oi(fid, &lmm->lmm_oi);
 681         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_LMMOI))
 682                 lmm->lmm_oi.oi.oi_id++;
 683         lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
 684
 685         lmm->lmm_stripe_size = cpu_to_le32(lod_comp->llc_stripe_size);
 686         lmm->lmm_stripe_count = cpu_to_le16(lod_comp->llc_stripe_count);
 687         /**
 688          * for dir and uninstantiated component, lmm_layout_gen stores
 689          * default stripe offset.
 690          */
 691         lmm->lmm_layout_gen =
 692                 (is_dir || !lod_comp_inited(lod_comp)) ?
 693                         cpu_to_le16(lod_comp->llc_stripe_offset) :
 694                         cpu_to_le16(lod_comp->llc_layout_gen);
 695
 696         if (magic == LOV_MAGIC_V1) {
 697                 objs = &lmm->lmm_objects[0];
 698         } else {
 699                 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
 700                 size_t cplen = strlcpy(v3->lmm_pool_name,
 701                                        lod_comp->llc_pool,
 702                                        sizeof(v3->lmm_pool_name));
 703                 if (cplen >= sizeof(v3->lmm_pool_name))
 704                         RETURN(-E2BIG);
 705                 objs = &v3->lmm_objects[0];
 706         }
 707         stripe_count = lod_comp_entry_stripe_count(lo, lod_comp, is_dir);
 708         if (stripe_count == 0 && !is_dir &&
 709             !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
 710             !(lod_comp->llc_pattern & LOV_PATTERN_MDT))
 711                 RETURN(-E2BIG);
 712
 713         if (!is_dir && lo->ldo_is_composite)
 714                 lod_comp_shrink_stripe_count(lod_comp, &stripe_count);
 715
 716         if (is_dir || lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)
 717                 GOTO(done, rc = 0);
 718
 719         /* generate ost_idx of this component stripe */
 720         lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
 721         for (i = 0; i < stripe_count; i++) {
 722                 struct dt_object *object;
 723                 __u32 ost_idx = (__u32)-1UL;
 724                 int type = LU_SEQ_RANGE_OST;
 725
 726                 if (lod_comp->llc_stripe && lod_comp->llc_stripe[i]) {
 727                         object = lod_comp->llc_stripe[i];
 728                         /* instantiated component */
 729                         info->lti_fid = *lu_object_fid(&object->do_lu);
 730
 731                         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_MULTIPLE_REF) &&
 732                             comp_idx == 0) {
 733                                 if (cfs_fail_val == 0)
 734                                         cfs_fail_val = info->lti_fid.f_oid;
 735                                 else if (i == 0)
 736                                         info->lti_fid.f_oid = cfs_fail_val;
 737                         }
 738
 739                         rc = fid_to_ostid(&info->lti_fid, &info->lti_ostid);
 740                         LASSERT(rc == 0);
 741
 742                         ostid_cpu_to_le(&info->lti_ostid, &objs[i].l_ost_oi);
 743                         objs[i].l_ost_gen = cpu_to_le32(0);
 744                         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FLD_LOOKUP))
 745                                 rc = -ENOENT;
 746                         else
 747                                 rc = lod_fld_lookup(env, lod, &info->lti_fid,
 748                                                     &ost_idx, &type);
 749                         if (rc < 0) {
 750                                 CERROR("%s: Can not locate "DFID": rc = %d\n",
 751                                        lod2obd(lod)->obd_name,
 752                                        PFID(&info->lti_fid), rc);
 753                                 RETURN(rc);
 754                         }
 755                 } else if (lod_comp->llc_ostlist.op_array &&
 756                            lod_comp->llc_ostlist.op_count) {
 757                         /* user specified ost list */
 758                         ost_idx = lod_comp->llc_ostlist.op_array[i];
 759                 }
 760                 /*
 761                  * with un-instantiated or with no specified ost list
 762                  * component, its l_ost_idx does not matter.
 763                  */
 764                 objs[i].l_ost_idx = cpu_to_le32(ost_idx);
 765         }
 766 done:
 767         if (lmm_size != NULL)
 768                 *lmm_size = lov_mds_md_size(stripe_count, magic);
 769         RETURN(rc);
 770 }
 771
 772 /**
 773  * Generate on-disk lov_mds_md structure based on the information in
 774  * the lod_object->ldo_comp_entries.
 775  *
 776  * \param[in] env               execution environment for this thread
 777  * \param[in] lo                LOD object
 778  * \param[in] lmm               buffer to cotain the on-disk lov_mds_md
 779  * \param[in|out] lmm_size      buffer size/lmm size
 780  * \param[in] is_dir            generate lov ea for dir or file? For dir case,
 781  *                              the stripe info is from the default stripe
 782  *                              template, which is collected in lod_ah_init(),
 783  *                              either from parent object or root object; for
 784  *                              file case, it's from the @lo object
 785  *
 786  * \retval                      0 if on disk structure is created successfully
 787  * \retval                      negative error number on failure
 788  */
 789 int lod_generate_lovea(const struct lu_env *env, struct lod_object *lo,
 790                        struct lov_mds_md *lmm, int *lmm_size, bool is_dir)
 791 {
 792         struct lov_comp_md_entry_v1 *lcme;
 793         struct lov_comp_md_v1 *lcm;
 794         struct lod_layout_component *comp_entries;
 795         __u16 comp_cnt, mirror_cnt;
 796         bool is_composite, is_foreign = false;
 797         int i, rc = 0, offset;
 798         ENTRY;
 799
 800         if (is_dir) {
 801                 comp_cnt = lo->ldo_def_striping->lds_def_comp_cnt;
 802                 mirror_cnt = lo->ldo_def_striping->lds_def_mirror_cnt;
 803                 comp_entries = lo->ldo_def_striping->lds_def_comp_entries;
 804                 is_composite =
 805                         lo->ldo_def_striping->lds_def_striping_is_composite;
 806         } else {
 807                 comp_cnt = lo->ldo_comp_cnt;
 808                 mirror_cnt = lo->ldo_mirror_count;
 809                 comp_entries = lo->ldo_comp_entries;
 810                 is_composite = lo->ldo_is_composite;
 811                 is_foreign = lo->ldo_is_foreign;
 812         }
 813
 814         LASSERT(lmm_size != NULL);
 815
 816         if (is_foreign) {
 817                 struct lov_foreign_md *lfm;
 818
 819                 lfm = (struct lov_foreign_md *)lmm;
 820                 memcpy(lfm, lo->ldo_foreign_lov, lo->ldo_foreign_lov_size);
 821                 /* need to store little-endian */
 822                 if (cpu_to_le32(LOV_MAGIC_FOREIGN) != LOV_MAGIC_FOREIGN) {
 823                         __swab32s(&lfm->lfm_magic);
 824                         __swab32s(&lfm->lfm_length);
 825                         __swab32s(&lfm->lfm_type);
 826                         __swab32s(&lfm->lfm_flags);
 827                 }
 828                 *lmm_size = lo->ldo_foreign_lov_size;
 829                 RETURN(0);
 830         }
 831
 832         LASSERT(comp_cnt != 0 && comp_entries != NULL);
 833
 834         if (!is_composite) {
 835                 rc = lod_gen_component_ea(env, lo, 0, lmm, lmm_size, is_dir);
 836                 RETURN(rc);
 837         }
 838
 839         lcm = (struct lov_comp_md_v1 *)lmm;
 840         memset(lcm, 0, sizeof(*lcm));
 841
 842         lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
 843         lcm->lcm_entry_count = cpu_to_le16(comp_cnt);
 844         lcm->lcm_mirror_count = cpu_to_le16(mirror_cnt - 1);
 845         lcm->lcm_flags = cpu_to_le16(lo->ldo_flr_state);
 846
 847         offset = sizeof(*lcm) + sizeof(*lcme) * comp_cnt;
 848         LASSERT(offset % sizeof(__u64) == 0);
 849
 850         for (i = 0; i < comp_cnt; i++) {
 851                 struct lod_layout_component *lod_comp;
 852                 struct lov_mds_md *sub_md;
 853                 int size;
 854
 855                 lod_comp = &comp_entries[i];
 856                 lcme = &lcm->lcm_entries[i];
 857
 858                 LASSERT(ergo(!is_dir, lod_comp->llc_id != LCME_ID_INVAL));
 859                 lcme->lcme_id = cpu_to_le32(lod_comp->llc_id);
 860
 861                 /* component could be un-inistantiated */
 862                 lcme->lcme_flags = cpu_to_le32(lod_comp->llc_flags);
 863                 if (lod_comp->llc_flags & LCME_FL_NOSYNC)
 864                         lcme->lcme_timestamp =
 865                                 cpu_to_le64(lod_comp->llc_timestamp);
 866                 if (lod_comp->llc_flags & LCME_FL_EXTENSION && !is_dir)
 867                         lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_SEL);
 868
 869                 lcme->lcme_extent.e_start =
 870                         cpu_to_le64(lod_comp->llc_extent.e_start);
 871                 lcme->lcme_extent.e_end =
 872                         cpu_to_le64(lod_comp->llc_extent.e_end);
 873                 lcme->lcme_offset = cpu_to_le32(offset);
 874
 875                 sub_md = (struct lov_mds_md *)((char *)lcm + offset);
 876                 rc = lod_gen_component_ea(env, lo, i, sub_md, &size, is_dir);
 877                 if (rc)
 878                         GOTO(out, rc);
 879                 lcme->lcme_size = cpu_to_le32(size);
 880                 offset += size;
 881                 LASSERTF((offset <= *lmm_size) && (offset % sizeof(__u64) == 0),
 882                          "offset:%d lmm_size:%d\n", offset, *lmm_size);
 883         }
 884         lcm->lcm_size = cpu_to_le32(offset);
 885         lcm->lcm_layout_gen = cpu_to_le32(is_dir ? 0 : lo->ldo_layout_gen);
 886
 887         lustre_print_user_md(D_LAYOUT, (struct lov_user_md *)lmm,
 888                              "generate lum");
 889 out:
 890         if (rc == 0)
 891                 *lmm_size = offset;
 892         RETURN(rc);
 893 }
 894
 895 /**
 896  * Get LOV EA.
 897  *
 898  * Fill lti_ea_store buffer in the environment with a value for the given
 899  * EA. The buffer is reallocated if the value doesn't fit.
 900  *
 901  * \param[in,out] env           execution environment for this thread
 902  *                              .lti_ea_store buffer is filled with EA's value
 903  * \param[in] lo                LOD object
 904  * \param[in] name              name of the EA
 905  *
 906  * \retval                      > 0 if EA is fetched successfully
 907  * \retval                      0 if EA is empty
 908  * \retval                      negative error number on failure
 909  */
 910 int lod_get_ea(const struct lu_env *env, struct lod_object *lo,
 911                const char *name)
 912 {
 913         struct lod_thread_info  *info = lod_env_info(env);
 914         struct dt_object        *next = dt_object_child(&lo->ldo_obj);
 915         int                     rc;
 916         ENTRY;
 917
 918         LASSERT(info);
 919
 920         if (unlikely(info->lti_ea_store == NULL)) {
 921                 /* just to enter in allocation block below */
 922                 rc = -ERANGE;
 923         } else {
 924 repeat:
 925                 info->lti_buf.lb_buf = info->lti_ea_store;
 926                 info->lti_buf.lb_len = info->lti_ea_store_size;
 927                 rc = dt_xattr_get(env, next, &info->lti_buf, name);
 928         }
 929
 930         /* if object is not striped or inaccessible */
 931         if (rc == -ENODATA || rc == -ENOENT)
 932                 RETURN(0);
 933
 934         if (rc == -ERANGE) {
 935                 /* EA doesn't fit, reallocate new buffer */
 936                 rc = dt_xattr_get(env, next, &LU_BUF_NULL, name);
 937                 if (rc == -ENODATA || rc == -ENOENT)
 938                         RETURN(0);
 939                 else if (rc < 0)
 940                         RETURN(rc);
 941
 942                 LASSERT(rc > 0);
 943                 rc = lod_ea_store_resize(info, rc);
 944                 if (rc)
 945                         RETURN(rc);
 946                 goto repeat;
 947         }
 948
 949         RETURN(rc);
 950 }
 951
 952 /**
 953  * Verify the target index is present in the current configuration.
 954  *
 955  * \param[in] md                LOD device where the target table is stored
 956  * \param[in] idx               target's index
 957  *
 958  * \retval                      0 if the index is present
 959  * \retval                      -EINVAL if not
 960  */
 961 static int validate_lod_and_idx(struct lod_device *md, __u32 idx)
 962 {
 963         if (unlikely(idx >= md->lod_ost_descs.ltd_tgts_size ||
 964                      !test_bit(idx, md->lod_ost_bitmap))) {
 965                 CERROR("%s: bad idx: %d of %d\n", lod2obd(md)->obd_name, idx,
 966                        md->lod_ost_descs.ltd_tgts_size);
 967                 return -EINVAL;
 968         }
 969
 970         if (unlikely(OST_TGT(md, idx) == NULL)) {
 971                 CERROR("%s: bad lod_tgt_desc for idx: %d\n",
 972                        lod2obd(md)->obd_name, idx);
 973                 return -EINVAL;
 974         }
 975
 976         if (unlikely(OST_TGT(md, idx)->ltd_tgt == NULL)) {
 977                 CERROR("%s: invalid lod device, for idx: %d\n",
 978                        lod2obd(md)->obd_name , idx);
 979                 return -EINVAL;
 980         }
 981
 982         return 0;
 983 }
 984
 985 /**
 986  * Instantiate objects for stripes.
 987  *
 988  * Allocate and initialize LU-objects representing the stripes. The number
 989  * of the stripes (ldo_stripe_count) must be initialized already. The caller
 990  * must ensure nobody else is calling the function on the object at the same
 991  * time. FLDB service must be running to be able to map a FID to the targets
 992  * and find appropriate device representing that target.
 993  *
 994  * \param[in] env               execution environment for this thread
 995  * \param[in,out] lo            LOD object
 996  * \param[in] objs              an array of IDs to creates the objects from
 997  * \param[in] comp_idx          index of ldo_comp_entries
 998  *
 999  * \retval                      0 if the objects are instantiated successfully
1000  * \retval                      negative error number on failure
1001  */
1002 int lod_initialize_objects(const struct lu_env *env, struct lod_object *lo,
1003                            struct lov_ost_data_v1 *objs, int comp_idx)
1004 {
1005         struct lod_layout_component *lod_comp;
1006         struct lod_thread_info *info = lod_env_info(env);
1007         struct lod_device *md;
1008         struct lu_object *o, *n;
1009         struct lu_device *nd;
1010         struct dt_object **stripe = NULL;
1011         __u32 *ost_indices = NULL;
1012         int stripe_len;
1013         int i, rc = 0;
1014         __u32 idx;
1015         ENTRY;
1016
1017         LASSERT(lo != NULL);
1018         md = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1019
1020         LASSERT(lo->ldo_comp_cnt != 0 && lo->ldo_comp_entries != NULL);
1021         lod_comp = &lo->ldo_comp_entries[comp_idx];
1022
1023         LASSERT(lod_comp->llc_stripe == NULL);
1024         LASSERT(lod_comp->llc_stripe_count > 0);
1025         LASSERT(lod_comp->llc_stripe_size > 0);
1026
1027         stripe_len = lod_comp->llc_stripe_count;
1028         OBD_ALLOC_PTR_ARRAY(stripe, stripe_len);
1029         if (stripe == NULL)
1030                 RETURN(-ENOMEM);
1031         OBD_ALLOC_PTR_ARRAY(ost_indices, stripe_len);
1032         if (!ost_indices)
1033                 GOTO(out, rc = -ENOMEM);
1034
1035         for (i = 0; i < lod_comp->llc_stripe_count; i++) {
1036                 if (unlikely(lovea_slot_is_dummy(&objs[i])))
1037                         continue;
1038
1039                 ostid_le_to_cpu(&objs[i].l_ost_oi, &info->lti_ostid);
1040                 idx = le32_to_cpu(objs[i].l_ost_idx);
1041                 rc = ostid_to_fid(&info->lti_fid, &info->lti_ostid, idx);
1042                 if (rc != 0)
1043                         GOTO(out, rc);
1044                 LASSERTF(fid_is_sane(&info->lti_fid), ""DFID" insane!\n",
1045                          PFID(&info->lti_fid));
1046                 lod_getref(&md->lod_ost_descs);
1047
1048                 rc = validate_lod_and_idx(md, idx);
1049                 if (unlikely(rc != 0)) {
1050                         lod_putref(md, &md->lod_ost_descs);
1051                         GOTO(out, rc);
1052                 }
1053
1054                 nd = &OST_TGT(md, idx)->ltd_tgt->dd_lu_dev;
1055                 lod_putref(md, &md->lod_ost_descs);
1056
1057                 /* In the function below, .hs_keycmp resolves to
1058                  * u_obj_hop_keycmp() */
1059                 /* coverity[overrun-buffer-val] */
1060                 o = lu_object_find_at(env, nd, &info->lti_fid, NULL);
1061                 if (IS_ERR(o))
1062                         GOTO(out, rc = PTR_ERR(o));
1063
1064                 n = lu_object_locate(o->lo_header, nd->ld_type);
1065                 LASSERT(n);
1066
1067                 stripe[i] = container_of(n, struct dt_object, do_lu);
1068                 ost_indices[i] = idx;
1069         }
1070
1071 out:
1072         if (rc != 0) {
1073                 for (i = 0; i < stripe_len; i++)
1074                         if (stripe[i] != NULL)
1075                                 dt_object_put(env, stripe[i]);
1076
1077                 OBD_FREE_PTR_ARRAY(stripe, stripe_len);
1078                 lod_comp->llc_stripe_count = 0;
1079                 if (ost_indices)
1080                         OBD_FREE_PTR_ARRAY(ost_indices, stripe_len);
1081         } else {
1082                 lod_comp->llc_stripe = stripe;
1083                 lod_comp->llc_ost_indices = ost_indices;
1084                 lod_comp->llc_stripes_allocated = stripe_len;
1085         }
1086
1087         RETURN(rc);
1088 }
1089
1090 /**
1091  * Instantiate objects for striping.
1092  *
1093  * Parse striping information in \a buf and instantiate the objects
1094  * representing the stripes.
1095  *
1096  * \param[in] env               execution environment for this thread
1097  * \param[in] lo                LOD object
1098  * \param[in] buf               buffer storing LOV EA to parse
1099  *
1100  * \retval                      0 if parsing and objects creation succeed
1101  * \retval                      negative error number on failure
1102  */
1103 int lod_parse_striping(const struct lu_env *env, struct lod_object *lo,
1104                        const struct lu_buf *buf)
1105 {
1106         struct lov_mds_md_v1 *lmm;
1107         struct lov_comp_md_v1 *comp_v1 = NULL;
1108         struct lov_foreign_md *foreign = NULL;
1109         struct lov_ost_data_v1 *objs;
1110         __u32 magic, pattern;
1111         __u16 mirror_cnt = 0;
1112         __u16 comp_cnt;
1113         int i, rc;
1114         ENTRY;
1115
1116         LASSERT(buf);
1117         LASSERT(buf->lb_buf);
1118         LASSERT(buf->lb_len);
1119         LASSERT(mutex_is_locked(&lo->ldo_layout_mutex));
1120
1121         lmm = (struct lov_mds_md_v1 *)buf->lb_buf;
1122         magic = le32_to_cpu(lmm->lmm_magic);
1123
1124         if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 &&
1125             magic != LOV_MAGIC_COMP_V1 && magic != LOV_MAGIC_FOREIGN &&
1126             magic != LOV_MAGIC_SEL)
1127                 GOTO(out, rc = -EINVAL);
1128
1129         if (lo->ldo_is_foreign)
1130                 lod_free_foreign_lov(lo);
1131         else
1132                 lod_free_comp_entries(lo);
1133
1134         if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) {
1135                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1136                 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1137                 if (comp_cnt == 0)
1138                         GOTO(out, rc = -EINVAL);
1139                 lo->ldo_layout_gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1140                 lo->ldo_is_composite = 1;
1141                 lo->ldo_flr_state = le16_to_cpu(comp_v1->lcm_flags) &
1142                                         LCM_FL_FLR_MASK;
1143                 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1144         } else if (magic == LOV_MAGIC_FOREIGN) {
1145                 size_t length;
1146
1147                 foreign = (struct lov_foreign_md *)buf->lb_buf;
1148                 length = offsetof(typeof(*foreign), lfm_value);
1149                 if (buf->lb_len < length ||
1150                     buf->lb_len < (length + le32_to_cpu(foreign->lfm_length))) {
1151                         CDEBUG(D_LAYOUT,
1152                                "buf len %zu too small for lov_foreign_md\n",
1153                                buf->lb_len);
1154                         GOTO(out, rc = -EINVAL);
1155                 }
1156
1157                 /* just cache foreign LOV EA raw */
1158                 rc = lod_alloc_foreign_lov(lo, length);
1159                 if (rc)
1160                         GOTO(out, rc);
1161                 memcpy(lo->ldo_foreign_lov, buf->lb_buf, length);
1162                 GOTO(out, rc);
1163         } else {
1164                 comp_cnt = 1;
1165                 lo->ldo_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1166                 lo->ldo_is_composite = 0;
1167         }
1168
1169         rc = lod_alloc_comp_entries(lo, mirror_cnt, comp_cnt);
1170         if (rc)
1171                 GOTO(out, rc);
1172
1173         for (i = 0; i < comp_cnt; i++) {
1174                 struct lod_layout_component *lod_comp;
1175                 struct lu_extent *ext;
1176                 __u32 offs;
1177
1178                 lod_comp = &lo->ldo_comp_entries[i];
1179                 if (lo->ldo_is_composite) {
1180                         offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1181                         lmm = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs);
1182
1183                         ext = &comp_v1->lcm_entries[i].lcme_extent;
1184                         lod_comp->llc_extent.e_start =
1185                                 le64_to_cpu(ext->e_start);
1186                         lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end);
1187                         lod_comp->llc_flags =
1188                                 le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags);
1189                         if (lod_comp->llc_flags & LCME_FL_NOSYNC)
1190                                 lod_comp->llc_timestamp = le64_to_cpu(
1191                                         comp_v1->lcm_entries[i].lcme_timestamp);
1192                         lod_comp->llc_id =
1193                                 le32_to_cpu(comp_v1->lcm_entries[i].lcme_id);
1194                         if (lod_comp->llc_id == LCME_ID_INVAL)
1195                                 GOTO(out, rc = -EINVAL);
1196
1197                         if ((lod_comp->llc_flags & LCME_FL_EXTENSION) &&
1198                             comp_v1->lcm_magic != cpu_to_le32(LOV_MAGIC_SEL)) {
1199                                 struct lod_device *d =
1200                                         lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1201
1202                                 CWARN("%s: EXTENSION flags=%x set on component[%u]=%x of non-SEL file "DFID" with magic=%#08x\n",
1203                                       lod2obd(d)->obd_name,
1204                                       lod_comp->llc_flags, lod_comp->llc_id, i,
1205                                       PFID(lod_object_fid(lo)),
1206                                       le32_to_cpu(comp_v1->lcm_magic));
1207                         }
1208                 } else {
1209                         lod_comp_set_init(lod_comp);
1210                 }
1211
1212                 pattern = le32_to_cpu(lmm->lmm_pattern);
1213                 if (!lov_pattern_supported(lov_pattern(pattern)))
1214                         GOTO(out, rc = -EINVAL);
1215
1216                 lod_comp->llc_pattern = pattern;
1217                 lod_comp->llc_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
1218                 lod_comp->llc_stripe_count = le16_to_cpu(lmm->lmm_stripe_count);
1219                 lod_comp->llc_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1220
1221                 if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)) {
1222                         struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
1223
1224                         lod_set_pool(&lod_comp->llc_pool, v3->lmm_pool_name);
1225                         objs = &v3->lmm_objects[0];
1226                 } else {
1227                         lod_set_pool(&lod_comp->llc_pool, NULL);
1228                         objs = &lmm->lmm_objects[0];
1229                 }
1230
1231                 /**
1232                  * If uninstantiated template component has valid l_ost_idx,
1233                  * then user has specified ost list for this component.
1234                  */
1235                 if (!lod_comp_inited(lod_comp)) {
1236                         __u16 stripe_count;
1237
1238                         if (objs[0].l_ost_idx != (__u32)-1UL) {
1239                                 int j;
1240
1241                                 stripe_count = lod_comp_entry_stripe_count(
1242                                                         lo, lod_comp, false);
1243                                 if (stripe_count == 0 &&
1244                                     !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1245                                     !(lod_comp->llc_pattern & LOV_PATTERN_MDT))
1246                                         GOTO(out, rc = -E2BIG);
1247                                 /**
1248                                  * load the user specified ost list, when this
1249                                  * component is instantiated later, it will be
1250                                  * used in lod_alloc_ost_list().
1251                                  */
1252                                 lod_comp->llc_ostlist.op_count = stripe_count;
1253                                 lod_comp->llc_ostlist.op_size =
1254                                         stripe_count * sizeof(__u32);
1255                                 OBD_ALLOC(lod_comp->llc_ostlist.op_array,
1256                                           lod_comp->llc_ostlist.op_size);
1257                                 if (!lod_comp->llc_ostlist.op_array)
1258                                         GOTO(out, rc = -ENOMEM);
1259
1260                                 for (j = 0; j < stripe_count; j++)
1261                                         lod_comp->llc_ostlist.op_array[j] =
1262                                                 le32_to_cpu(objs[j].l_ost_idx);
1263
1264                                 /**
1265                                  * this component OST objects starts from the
1266                                  * first ost_idx, lod_alloc_ost_list() will
1267                                  * check this.
1268                                  */
1269                                 lod_comp->llc_stripe_offset = objs[0].l_ost_idx;
1270                         } else {
1271                                 /**
1272                                  * for uninstantiated component,
1273                                  * lmm_layout_gen stores default stripe offset.
1274                                  */
1275                                 lod_comp->llc_stripe_offset =
1276                                                         lmm->lmm_layout_gen;
1277                         }
1278                 }
1279
1280                 /* skip un-instantiated component object initialization */
1281                 if (!lod_comp_inited(lod_comp))
1282                         continue;
1283
1284                 if (!(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1285                     !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) {
1286                         rc = lod_initialize_objects(env, lo, objs, i);
1287                         if (rc)
1288                                 GOTO(out, rc);
1289                 }
1290         }
1291
1292         rc = lod_fill_mirrors(lo);
1293         if (rc)
1294                 GOTO(out, rc);
1295
1296 out:
1297         if (rc)
1298                 lod_striping_free_nolock(env, lo);
1299         RETURN(rc);
1300 }
1301
1302 /**
1303  * Check whether the striping (LOVEA for regular file, LMVEA for directory)
1304  * is already cached.
1305  *
1306  * \param[in] lo        LOD object
1307  *
1308  * \retval              True if the striping is cached, otherwise
1309  *                      return false.
1310  */
1311 static bool lod_striping_loaded(struct lod_object *lo)
1312 {
1313         if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr) &&
1314             lo->ldo_comp_cached)
1315                 return true;
1316
1317         if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1318                 if (lo->ldo_dir_stripe_loaded)
1319                         return true;
1320
1321                 /* Never load LMV stripe for slaves of striped dir */
1322                 if (lo->ldo_dir_slave_stripe)
1323                         return true;
1324         }
1325
1326         return false;
1327 }
1328
1329 /**
1330  * A generic function to initialize the stripe objects.
1331  *
1332  * A protected version of lod_striping_load_locked() - load the striping
1333  * information from storage, parse that and instantiate LU objects to
1334  * represent the stripes.  The LOD object \a lo supplies a pointer to the
1335  * next sub-object in the LU stack so we can lock it. Also use \a lo to
1336  * return an array of references to the newly instantiated objects.
1337  *
1338  * \param[in] env               execution environment for this thread
1339  * \param[in,out] lo            LOD object, where striping is stored and
1340  *                              which gets an array of references
1341  *
1342  * \retval                      0 if parsing and object creation succeed
1343  * \retval                      negative error number on failure
1344  **/
1345 int lod_striping_load(const struct lu_env *env, struct lod_object *lo)
1346 {
1347         struct lod_thread_info *info = lod_env_info(env);
1348         struct dt_object *next = dt_object_child(&lo->ldo_obj);
1349         struct lu_buf *buf = &info->lti_buf;
1350         int rc = 0;
1351
1352         ENTRY;
1353
1354         if (!dt_object_exists(next))
1355                 RETURN(0);
1356
1357         if (lod_striping_loaded(lo))
1358                 RETURN(0);
1359
1360         mutex_lock(&lo->ldo_layout_mutex);
1361         if (lod_striping_loaded(lo))
1362                 GOTO(unlock, rc = 0);
1363
1364         if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr)) {
1365                 rc = lod_get_lov_ea(env, lo);
1366                 if (rc <= 0)
1367                         GOTO(unlock, rc);
1368
1369                 /*
1370                  * there is LOV EA (striping information) in this object
1371                  * let's parse it and create in-core objects for the stripes
1372                  */
1373                 buf->lb_buf = info->lti_ea_store;
1374                 buf->lb_len = info->lti_ea_store_size;
1375                 rc = lod_parse_striping(env, lo, buf);
1376                 if (rc == 0)
1377                         lo->ldo_comp_cached = 1;
1378         } else if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1379                 rc = lod_get_lmv_ea(env, lo);
1380                 if (rc > sizeof(struct lmv_foreign_md)) {
1381                         struct lmv_foreign_md *lfm = info->lti_ea_store;
1382
1383                         if (le32_to_cpu(lfm->lfm_magic) == LMV_MAGIC_FOREIGN) {
1384                                 lo->ldo_foreign_lmv = info->lti_ea_store;
1385                                 lo->ldo_foreign_lmv_size =
1386                                         info->lti_ea_store_size;
1387                                 info->lti_ea_store = NULL;
1388                                 info->lti_ea_store_size = 0;
1389
1390                                 lo->ldo_dir_stripe_loaded = 1;
1391                                 lo->ldo_dir_is_foreign = 1;
1392                                 GOTO(unlock, rc = 0);
1393                         }
1394                 }
1395
1396                 if (rc < (int)sizeof(struct lmv_mds_md_v1)) {
1397                         /* Let's set stripe_loaded to avoid further
1398                          * stripe loading especially for non-stripe directory,
1399                          * which can hurt performance. (See LU-9840)
1400                          */
1401                         if (rc == 0)
1402                                 lo->ldo_dir_stripe_loaded = 1;
1403                         GOTO(unlock, rc = rc > 0 ? -EINVAL : rc);
1404                 }
1405                 buf->lb_buf = info->lti_ea_store;
1406                 buf->lb_len = info->lti_ea_store_size;
1407                 if (rc == sizeof(struct lmv_mds_md_v1)) {
1408                         rc = lod_load_lmv_shards(env, lo, buf, true);
1409                         if (buf->lb_buf != info->lti_ea_store) {
1410                                 OBD_FREE_LARGE(info->lti_ea_store,
1411                                                info->lti_ea_store_size);
1412                                 info->lti_ea_store = buf->lb_buf;
1413                                 info->lti_ea_store_size = buf->lb_len;
1414                         }
1415
1416                         if (rc < 0)
1417                                 GOTO(unlock, rc);
1418                 }
1419
1420                 /*
1421                  * there is LMV EA (striping information) in this object
1422                  * let's parse it and create in-core objects for the stripes
1423                  */
1424                 rc = lod_parse_dir_striping(env, lo, buf);
1425                 if (rc == 0)
1426                         lo->ldo_dir_stripe_loaded = 1;
1427         }
1428         EXIT;
1429 unlock:
1430         mutex_unlock(&lo->ldo_layout_mutex);
1431
1432         return rc;
1433 }
1434
1435 int lod_striping_reload(const struct lu_env *env, struct lod_object *lo,
1436                          const struct lu_buf *buf)
1437 {
1438         int rc;
1439
1440         ENTRY;
1441
1442         mutex_lock(&lo->ldo_layout_mutex);
1443         lod_striping_free_nolock(env, lo);
1444         rc = lod_parse_striping(env, lo, buf);
1445         mutex_unlock(&lo->ldo_layout_mutex);
1446
1447         RETURN(rc);
1448 }
1449
1450 /**
1451  * Verify lov_user_md_v1/v3 striping.
1452  *
1453  * Check the validity of all fields including the magic, stripe size,
1454  * stripe count, stripe offset and that the pool is present.  Also check
1455  * that each target index points to an existing target. The additional
1456  * \a is_from_disk turns additional checks. In some cases zero fields
1457  * are allowed (like pattern=0).
1458  *
1459  * \param[in] d                 LOD device
1460  * \param[in] buf               buffer with LOV EA to verify
1461  * \param[in] is_from_disk      0 - from user, allow some fields to be 0
1462  *                              1 - from disk, do not allow
1463  *
1464  * \retval                      0 if the striping is valid
1465  * \retval                      -EINVAL if striping is invalid
1466  */
1467 static int lod_verify_v1v3(struct lod_device *d, const struct lu_buf *buf,
1468                            bool is_from_disk)
1469 {
1470         struct lov_user_md_v1   *lum;
1471         struct lov_user_md_v3   *lum3;
1472         struct pool_desc        *pool = NULL;
1473         __u32                    magic;
1474         __u32                    stripe_size;
1475         __u16                    stripe_count;
1476         __u16                    stripe_offset;
1477         size_t                   lum_size;
1478         int                      rc = 0;
1479         ENTRY;
1480
1481         lum = buf->lb_buf;
1482
1483         if (buf->lb_len < sizeof(*lum)) {
1484                 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
1485                        buf->lb_len);
1486                 GOTO(out, rc = -EINVAL);
1487         }
1488
1489         magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
1490         if (magic != LOV_USER_MAGIC_V1 &&
1491             magic != LOV_USER_MAGIC_V3 &&
1492             magic != LOV_USER_MAGIC_SPECIFIC) {
1493                 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
1494                        le32_to_cpu(lum->lmm_magic));
1495                 GOTO(out, rc = -EINVAL);
1496         }
1497
1498         /* the user uses "0" for default stripe pattern normally. */
1499         if (!is_from_disk && lum->lmm_pattern == LOV_PATTERN_NONE)
1500                 lum->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1501
1502         if (!lov_pattern_supported(le32_to_cpu(lum->lmm_pattern))) {
1503                 CDEBUG(D_LAYOUT, "bad userland stripe pattern: %#x\n",
1504                        le32_to_cpu(lum->lmm_pattern));
1505                 GOTO(out, rc = -EINVAL);
1506         }
1507
1508         /* a released lum comes from creating orphan on hsm release,
1509          * doesn't make sense to verify it. */
1510         if (le32_to_cpu(lum->lmm_pattern) & LOV_PATTERN_F_RELEASED)
1511                 GOTO(out, rc = 0);
1512
1513         /* 64kB is the largest common page size we see (ia64), and matches the
1514          * check in lfs */
1515         stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1516         if (stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
1517                 CDEBUG(D_LAYOUT, "stripe size %u not a multiple of %u\n",
1518                        stripe_size, LOV_MIN_STRIPE_SIZE);
1519                 GOTO(out, rc = -EINVAL);
1520         }
1521
1522         stripe_offset = le16_to_cpu(lum->lmm_stripe_offset);
1523         if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT &&
1524             lov_pattern(le32_to_cpu(lum->lmm_pattern)) != LOV_PATTERN_MDT) {
1525                 /* if offset is not within valid range [0, osts_size) */
1526                 if (stripe_offset >= d->lod_ost_descs.ltd_tgts_size) {
1527                         CDEBUG(D_LAYOUT, "stripe offset %u >= bitmap size %u\n",
1528                                stripe_offset, d->lod_ost_descs.ltd_tgts_size);
1529                         GOTO(out, rc = -EINVAL);
1530                 }
1531
1532                 /* if lmm_stripe_offset is *not* in bitmap */
1533                 if (!test_bit(stripe_offset, d->lod_ost_bitmap)) {
1534                         CDEBUG(D_LAYOUT, "stripe offset %u not in bitmap\n",
1535                                stripe_offset);
1536                         GOTO(out, rc = -EINVAL);
1537                 }
1538         }
1539
1540         if (magic == LOV_USER_MAGIC_V1)
1541                 lum_size = offsetof(struct lov_user_md_v1,
1542                                     lmm_objects[0]);
1543         else if (magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC)
1544                 lum_size = offsetof(struct lov_user_md_v3,
1545                                     lmm_objects[0]);
1546         else
1547                 GOTO(out, rc = -EINVAL);
1548
1549         stripe_count = le16_to_cpu(lum->lmm_stripe_count);
1550         if (buf->lb_len < lum_size) {
1551                 CDEBUG(D_LAYOUT, "invalid buf len %zu/%zu for lov_user_md with "
1552                        "magic %#x and stripe_count %u\n",
1553                        buf->lb_len, lum_size, magic, stripe_count);
1554                 GOTO(out, rc = -EINVAL);
1555         }
1556
1557         if (!(magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC))
1558                 goto out;
1559
1560         lum3 = buf->lb_buf;
1561         /* In the function below, .hs_keycmp resolves to
1562          * pool_hashkey_keycmp() */
1563         /* coverity[overrun-buffer-val] */
1564         pool = lod_find_pool(d, lum3->lmm_pool_name);
1565         if (pool == NULL)
1566                 goto out;
1567
1568         if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT) {
1569                 rc = lod_check_index_in_pool(stripe_offset, pool);
1570                 if (rc < 0)
1571                         GOTO(out, rc = -EINVAL);
1572         }
1573
1574         if (is_from_disk && stripe_count > pool_tgt_count(pool)) {
1575                 CDEBUG(D_LAYOUT, "stripe count %u > # OSTs %u in the pool\n",
1576                        stripe_count, pool_tgt_count(pool));
1577                 GOTO(out, rc = -EINVAL);
1578         }
1579
1580 out:
1581         if (pool != NULL)
1582                 lod_pool_putref(pool);
1583
1584         RETURN(rc);
1585 }
1586
1587 static inline
1588 struct lov_comp_md_entry_v1 *comp_entry_v1(struct lov_comp_md_v1 *comp, int i)
1589 {
1590         LASSERTF((le32_to_cpu(comp->lcm_magic) & ~LOV_MAGIC_DEFINED) ==
1591                  LOV_USER_MAGIC_COMP_V1, "Wrong magic %x\n",
1592                  le32_to_cpu(comp->lcm_magic));
1593         LASSERTF(i >= 0 && i < le16_to_cpu(comp->lcm_entry_count),
1594                  "bad index %d, max = %d\n",
1595                  i, le16_to_cpu(comp->lcm_entry_count));
1596
1597         return &comp->lcm_entries[i];
1598 }
1599
1600 #define for_each_comp_entry_v1(comp, entry) \
1601         for (entry = comp_entry_v1(comp, 0); \
1602              entry <= comp_entry_v1(comp, \
1603                                    le16_to_cpu(comp->lcm_entry_count) - 1); \
1604              entry++)
1605
1606 int lod_erase_dom_stripe(struct lov_comp_md_v1 *comp_v1,
1607                          struct lov_comp_md_entry_v1 *dom_ent)
1608 {
1609         struct lov_comp_md_entry_v1 *ent;
1610         __u16 entries;
1611         __u32 dom_off, dom_size, comp_size, off;
1612         void *src, *dst;
1613         unsigned int size, shift;
1614
1615         entries = le16_to_cpu(comp_v1->lcm_entry_count) - 1;
1616         LASSERT(entries > 0);
1617         comp_v1->lcm_entry_count = cpu_to_le16(entries);
1618
1619         comp_size = le32_to_cpu(comp_v1->lcm_size);
1620         dom_off = le32_to_cpu(dom_ent->lcme_offset);
1621         dom_size = le32_to_cpu(dom_ent->lcme_size);
1622
1623         /* all entries offsets are shifted by entry size at least */
1624         shift = sizeof(*dom_ent);
1625         for_each_comp_entry_v1(comp_v1, ent) {
1626                 off = le32_to_cpu(ent->lcme_offset);
1627                 if (off == dom_off) {
1628                         /* Entry deletion creates two holes in layout data:
1629                          * - hole in entries array
1630                          * - hole in layout data at dom_off with dom_size
1631                          *
1632                          * First memmove is one entry shift from next entry
1633                          * start with size up to dom_off in blob
1634                          */
1635                         dst = (void *)ent;
1636                         src = (void *)(ent + 1);
1637                         size = (unsigned long)((void *)comp_v1 + dom_off - src);
1638                         memmove(dst, src, size);
1639                         /* take 'off' from just moved entry */
1640                         off = le32_to_cpu(ent->lcme_offset);
1641                         /* second memmove is blob tail after 'off' up to
1642                          * component end
1643                          */
1644                         dst = (void *)comp_v1 + dom_off - sizeof(*ent);
1645                         src = (void *)comp_v1 + off;
1646                         size = (unsigned long)(comp_size - off);
1647                         memmove(dst, src, size);
1648                         /* all entries offsets after DoM entry are shifted by
1649                          * dom_size additionally
1650                          */
1651                         shift += dom_size;
1652                 }
1653                 ent->lcme_offset = cpu_to_le32(off - shift);
1654         }
1655         comp_v1->lcm_size = cpu_to_le32(comp_size - shift);
1656
1657         /* notify a caller to re-check entry */
1658         return -ERESTART;
1659 }
1660
1661 void lod_dom_stripesize_recalc(struct lod_device *d)
1662 {
1663         __u64 threshold_mb = d->lod_dom_threshold_free_mb;
1664         __u32 max_size = d->lod_dom_stripesize_max_kb;
1665         __u32 def_size = d->lod_dom_stripesize_cur_kb;
1666
1667         /* use maximum allowed value if free space is above threshold */
1668         if (d->lod_lsfs_free_mb >= threshold_mb) {
1669                 def_size = max_size;
1670         } else if (!d->lod_lsfs_free_mb || max_size <= LOD_DOM_MIN_SIZE_KB) {
1671                 def_size = 0;
1672         } else {
1673                 /* recalc threshold like it would be with def_size as max */
1674                 threshold_mb = mult_frac(threshold_mb, def_size, max_size);
1675                 if (d->lod_lsfs_free_mb < threshold_mb)
1676                         def_size = rounddown(def_size / 2, LOD_DOM_MIN_SIZE_KB);
1677                 else if (d->lod_lsfs_free_mb > threshold_mb * 2)
1678                         def_size = max_t(unsigned int, def_size * 2,
1679                                          LOD_DOM_MIN_SIZE_KB);
1680         }
1681
1682         if (d->lod_dom_stripesize_cur_kb != def_size) {
1683                 CDEBUG(D_LAYOUT, "Change default DOM stripe size %d->%d\n",
1684                        d->lod_dom_stripesize_cur_kb, def_size);
1685                 d->lod_dom_stripesize_cur_kb = def_size;
1686         }
1687 }
1688
1689 static __u32 lod_dom_stripesize_limit(const struct lu_env *env,
1690                                       struct lod_device *d)
1691 {
1692         int rc;
1693
1694         /* set bfree as fraction of total space */
1695         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_SPOOF)) {
1696                 spin_lock(&d->lod_lsfs_lock);
1697                 d->lod_lsfs_free_mb = mult_frac(d->lod_lsfs_total_mb,
1698                                         min_t(int, cfs_fail_val, 100), 100);
1699                 GOTO(recalc, rc = 0);
1700         }
1701
1702         if (d->lod_lsfs_age < ktime_get_seconds() - LOD_DOM_SFS_MAX_AGE) {
1703                 struct obd_statfs sfs;
1704
1705                 spin_lock(&d->lod_lsfs_lock);
1706                 if (d->lod_lsfs_age > ktime_get_seconds() - LOD_DOM_SFS_MAX_AGE)
1707                         GOTO(unlock, rc = 0);
1708
1709                 d->lod_lsfs_age = ktime_get_seconds();
1710                 spin_unlock(&d->lod_lsfs_lock);
1711                 rc = dt_statfs(env, d->lod_child, &sfs);
1712                 if (rc) {
1713                         CDEBUG(D_LAYOUT,
1714                                "%s: failed to get OSD statfs: rc = %d\n",
1715                                lod2obd(d)->obd_name, rc);
1716                         GOTO(out, rc);
1717                 }
1718                 /* udpate local OSD cached statfs data */
1719                 spin_lock(&d->lod_lsfs_lock);
1720                 d->lod_lsfs_total_mb = (sfs.os_blocks * sfs.os_bsize) >> 20;
1721                 d->lod_lsfs_free_mb = (sfs.os_bfree * sfs.os_bsize) >> 20;
1722 recalc:
1723                 lod_dom_stripesize_recalc(d);
1724 unlock:
1725                 spin_unlock(&d->lod_lsfs_lock);
1726         }
1727 out:
1728         return d->lod_dom_stripesize_cur_kb << 10;
1729 }
1730
1731 int lod_dom_stripesize_choose(const struct lu_env *env, struct lod_device *d,
1732                               struct lov_comp_md_v1 *comp_v1,
1733                               struct lov_comp_md_entry_v1 *dom_ent,
1734                               __u32 stripe_size)
1735 {
1736         struct lov_comp_md_entry_v1 *ent;
1737         struct lu_extent *dom_ext, *ext;
1738         struct lov_user_md_v1 *lum;
1739         __u32 max_stripe_size;
1740         __u16 mid, dom_mid;
1741         int rc = 0;
1742         bool dom_next_entry = false;
1743
1744         dom_ext = &dom_ent->lcme_extent;
1745         dom_mid = mirror_id_of(le32_to_cpu(dom_ent->lcme_id));
1746         max_stripe_size = lod_dom_stripesize_limit(env, d);
1747
1748         /* Check stripe size againts current per-MDT limit */
1749         if (stripe_size <= max_stripe_size)
1750                 return 0;
1751
1752         lum = (void *)comp_v1 + le32_to_cpu(dom_ent->lcme_offset);
1753         CDEBUG(D_LAYOUT, "overwrite DoM component size %u with MDT limit %u\n",
1754                stripe_size, max_stripe_size);
1755         lum->lmm_stripe_size = cpu_to_le32(max_stripe_size);
1756
1757         /* In common case the DoM stripe is first entry in a mirror and
1758          * can be deleted only if it is not single entry in layout or
1759          * mirror, otherwise error should be returned.
1760          */
1761         for_each_comp_entry_v1(comp_v1, ent) {
1762                 if (ent == dom_ent)
1763                         continue;
1764
1765                 mid = mirror_id_of(le32_to_cpu(ent->lcme_id));
1766                 if (mid != dom_mid)
1767                         continue;
1768
1769                 ext = &ent->lcme_extent;
1770                 if (ext->e_start != dom_ext->e_end)
1771                         continue;
1772
1773                 /* Found next component after the DoM one with the same
1774                  * mirror_id and adjust its start with DoM component end.
1775                  *
1776                  * NOTE: we are considering here that there can be only one
1777                  * DoM component in a file, all replicas are located on OSTs
1778                  * always and don't need adjustment since use own layouts.
1779                  */
1780                 ext->e_start = cpu_to_le64(max_stripe_size);
1781                 dom_next_entry = true;
1782                 break;
1783         }
1784
1785         if (max_stripe_size == 0) {
1786                 /* DoM component size is zero due to server setting, remove
1787                  * it from the layout but only if next component exists in
1788                  * the same mirror. That must be checked prior calling the
1789                  * lod_erase_dom_stripe().
1790                  */
1791                 if (!dom_next_entry)
1792                         return -EFBIG;
1793
1794                 rc = lod_erase_dom_stripe(comp_v1, dom_ent);
1795         } else {
1796                 /* Update DoM extent end finally */
1797                 dom_ext->e_end = cpu_to_le64(max_stripe_size);
1798         }
1799
1800         return rc;
1801 }
1802
1803 /**
1804  * Verify LOV striping.
1805  *
1806  * \param[in] d                 LOD device
1807  * \param[in] buf               buffer with LOV EA to verify
1808  * \param[in] is_from_disk      0 - from user, allow some fields to be 0
1809  *                              1 - from disk, do not allow
1810  * \param[in] start             extent start for composite layout
1811  *
1812  * \retval                      0 if the striping is valid
1813  * \retval                      -EINVAL if striping is invalid
1814  */
1815 int lod_verify_striping(const struct lu_env *env, struct lod_device *d,
1816                         struct lod_object *lo, const struct lu_buf *buf,
1817                         bool is_from_disk)
1818 {
1819         struct lov_user_md_v1   *lum;
1820         struct lov_comp_md_v1   *comp_v1;
1821         struct lov_comp_md_entry_v1     *ent;
1822         struct lu_extent        *ext;
1823         struct lu_buf   tmp;
1824         __u64   prev_end = 0;
1825         __u32   stripe_size = 0;
1826         __u16   prev_mid = -1, mirror_id = -1;
1827         __u32   mirror_count;
1828         __u32   magic;
1829         int     rc = 0;
1830         ENTRY;
1831
1832         if (buf->lb_len < sizeof(lum->lmm_magic)) {
1833                 CDEBUG(D_LAYOUT, "invalid buf len %zu\n", buf->lb_len);
1834                 RETURN(-EINVAL);
1835         }
1836
1837         lum = buf->lb_buf;
1838
1839         magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
1840         /* treat foreign LOV EA/object case first
1841          * XXX is it expected to try setting again a foreign?
1842          * XXX should we care about different current vs new layouts ?
1843          */
1844         if (unlikely(magic == LOV_USER_MAGIC_FOREIGN)) {
1845                 struct lov_foreign_md *lfm = buf->lb_buf;
1846
1847                 if (buf->lb_len < offsetof(typeof(*lfm), lfm_value)) {
1848                         CDEBUG(D_LAYOUT,
1849                                "buf len %zu < min lov_foreign_md size (%zu)\n",
1850                                buf->lb_len, offsetof(typeof(*lfm),
1851                                lfm_value));
1852                         RETURN(-EINVAL);
1853                 }
1854
1855                 if (foreign_size_le(lfm) > buf->lb_len) {
1856                         CDEBUG(D_LAYOUT,
1857                                "buf len %zu < this lov_foreign_md size (%zu)\n",
1858                                buf->lb_len, foreign_size_le(lfm));
1859                         RETURN(-EINVAL);
1860                 }
1861                 /* Don't do anything with foreign layouts */
1862                 RETURN(0);
1863         }
1864
1865         /* normal LOV/layout cases */
1866
1867         if (buf->lb_len < sizeof(*lum)) {
1868                 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
1869                        buf->lb_len);
1870                 RETURN(-EINVAL);
1871         }
1872
1873         if (magic != LOV_USER_MAGIC_V1 &&
1874             magic != LOV_USER_MAGIC_V3 &&
1875             magic != LOV_USER_MAGIC_SPECIFIC &&
1876             magic != LOV_USER_MAGIC_COMP_V1) {
1877                 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
1878                        le32_to_cpu(lum->lmm_magic));
1879                 RETURN(-EINVAL);
1880         }
1881
1882         if (magic != LOV_USER_MAGIC_COMP_V1)
1883                 RETURN(lod_verify_v1v3(d, buf, is_from_disk));
1884
1885         /* magic == LOV_USER_MAGIC_COMP_V1 */
1886         comp_v1 = buf->lb_buf;
1887         if (buf->lb_len < le32_to_cpu(comp_v1->lcm_size)) {
1888                 CDEBUG(D_LAYOUT, "buf len %zu is less than %u\n",
1889                        buf->lb_len, le32_to_cpu(comp_v1->lcm_size));
1890                 RETURN(-EINVAL);
1891         }
1892
1893 recheck:
1894         mirror_count = 0;
1895         if (le16_to_cpu(comp_v1->lcm_entry_count) == 0) {
1896                 CDEBUG(D_LAYOUT, "entry count is zero\n");
1897                 RETURN(-EINVAL);
1898         }
1899
1900         if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr) &&
1901             lo->ldo_comp_cnt > 0) {
1902                 /* could be called from lustre.lov.add */
1903                 __u32 cnt = lo->ldo_comp_cnt;
1904
1905                 ext = &lo->ldo_comp_entries[cnt - 1].llc_extent;
1906                 prev_end = ext->e_end;
1907
1908                 ++mirror_count;
1909         }
1910
1911         for_each_comp_entry_v1(comp_v1, ent) {
1912                 ext = &ent->lcme_extent;
1913
1914                 if (le64_to_cpu(ext->e_start) > le64_to_cpu(ext->e_end)) {
1915                         CDEBUG(D_LAYOUT, "invalid extent "DEXT"\n",
1916                                le64_to_cpu(ext->e_start),
1917                                le64_to_cpu(ext->e_end));
1918                         RETURN(-EINVAL);
1919                 }
1920
1921                 if (is_from_disk) {
1922                         /* lcme_id contains valid value */
1923                         if (le32_to_cpu(ent->lcme_id) == 0 ||
1924                             le32_to_cpu(ent->lcme_id) > LCME_ID_MAX) {
1925                                 CDEBUG(D_LAYOUT, "invalid id %u\n",
1926                                        le32_to_cpu(ent->lcme_id));
1927                                 RETURN(-EINVAL);
1928                         }
1929
1930                         if (le16_to_cpu(comp_v1->lcm_mirror_count) > 0) {
1931                                 mirror_id = mirror_id_of(
1932                                                 le32_to_cpu(ent->lcme_id));
1933
1934                                 /* first component must start with 0 */
1935                                 if (mirror_id != prev_mid &&
1936                                     le64_to_cpu(ext->e_start) != 0) {
1937                                         CDEBUG(D_LAYOUT,
1938                                                "invalid start:%llu, expect:0\n",
1939                                                le64_to_cpu(ext->e_start));
1940                                         RETURN(-EINVAL);
1941                                 }
1942
1943                                 prev_mid = mirror_id;
1944                         }
1945                 }
1946
1947                 if (le64_to_cpu(ext->e_start) == 0) {
1948                         ++mirror_count;
1949                         prev_end = 0;
1950                 }
1951
1952                 /* the next must be adjacent with the previous one */
1953                 if (le64_to_cpu(ext->e_start) != prev_end) {
1954                         CDEBUG(D_LAYOUT,
1955                                "invalid start actual:%llu, expect:%llu\n",
1956                                le64_to_cpu(ext->e_start), prev_end);
1957                         RETURN(-EINVAL);
1958                 }
1959
1960                 tmp.lb_buf = (char *)comp_v1 + le32_to_cpu(ent->lcme_offset);
1961                 tmp.lb_len = le32_to_cpu(ent->lcme_size);
1962
1963                 /* Check DoM entry is always the first one */
1964                 lum = tmp.lb_buf;
1965                 if (lov_pattern(le32_to_cpu(lum->lmm_pattern)) ==
1966                     LOV_PATTERN_MDT) {
1967                         /* DoM component must be the first in a mirror */
1968                         if (le64_to_cpu(ext->e_start) > 0) {
1969                                 CDEBUG(D_LAYOUT, "invalid DoM component "
1970                                        "with %llu extent start\n",
1971                                        le64_to_cpu(ext->e_start));
1972                                 RETURN(-EINVAL);
1973                         }
1974                         stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1975                         /* There is just one stripe on MDT and it must
1976                          * cover whole component size. */
1977                         if (stripe_size != le64_to_cpu(ext->e_end)) {
1978                                 CDEBUG(D_LAYOUT, "invalid DoM layout "
1979                                        "stripe size %u != %llu "
1980                                        "(component size)\n",
1981                                        stripe_size, prev_end);
1982                                 RETURN(-EINVAL);
1983                         }
1984                         /* Check and adjust stripe size by per-MDT limit */
1985                         rc = lod_dom_stripesize_choose(env, d, comp_v1, ent,
1986                                                        stripe_size);
1987                         /* DoM entry was removed, re-check layout from start */
1988                         if (rc == -ERESTART)
1989                                 goto recheck;
1990                         else if (rc)
1991                                 RETURN(rc);
1992
1993                         /* Any stripe count is forbidden on DoM component */
1994                         if (lum->lmm_stripe_count) {
1995                                 CDEBUG(D_LAYOUT,
1996                                        "invalid DoM layout stripe count %u, must be 0\n",
1997                                        le16_to_cpu(lum->lmm_stripe_count));
1998                                 RETURN(-EINVAL);
1999                         }
2000
2001                         /* Any pool is forbidden on DoM component */
2002                         if (lum->lmm_magic == LOV_USER_MAGIC_V3) {
2003                                 struct lov_user_md_v3 *v3 = (void *)lum;
2004
2005                                 if (v3->lmm_pool_name[0] != '\0') {
2006                                         CDEBUG(D_LAYOUT,
2007                                                "DoM component cannot have pool assigned\n");
2008                                         RETURN(-EINVAL);
2009                                 }
2010                         }
2011                 }
2012
2013                 prev_end = le64_to_cpu(ext->e_end);
2014
2015                 rc = lod_verify_v1v3(d, &tmp, is_from_disk);
2016                 if (rc)
2017                         RETURN(rc);
2018
2019                 if (prev_end == LUSTRE_EOF)
2020                         continue;
2021
2022                 /* extent end must be aligned with the stripe_size */
2023                 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
2024                 if (stripe_size && prev_end % stripe_size) {
2025                         CDEBUG(D_LAYOUT, "stripe size isn't aligned, "
2026                                "stripe_sz: %u, [%llu, %llu)\n",
2027                                stripe_size, ext->e_start, prev_end);
2028                         RETURN(-EINVAL);
2029                 }
2030         }
2031
2032         /* make sure that the mirror_count is telling the truth */
2033         if (mirror_count != le16_to_cpu(comp_v1->lcm_mirror_count) + 1)
2034                 RETURN(-EINVAL);
2035
2036         RETURN(0);
2037 }
2038
2039 /**
2040  * set the default stripe size, if unset.
2041  *
2042  * \param[in,out] val   number of bytes per OST stripe
2043  *
2044  * The minimum stripe size is 64KB to ensure that a single stripe is an
2045  * even multiple of a client PAGE_SIZE (IA64, PPC, etc).  Otherwise, it
2046  * is difficult to split dirty pages across OSCs during writes.
2047  */
2048 void lod_fix_desc_stripe_size(__u64 *val)
2049 {
2050         if (*val < LOV_MIN_STRIPE_SIZE) {
2051                 if (*val != 0)
2052                         LCONSOLE_INFO("Increasing default stripe size to "
2053                                       "minimum value %u\n",
2054                                       LOV_DESC_STRIPE_SIZE_DEFAULT);
2055                 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
2056         } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
2057                 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
2058                 LCONSOLE_WARN("Changing default stripe size to %llu (a "
2059                               "multiple of %u)\n",
2060                               *val, LOV_MIN_STRIPE_SIZE);
2061         }
2062 }
2063
2064 /**
2065  * set the filesystem default number of stripes, if unset.
2066  *
2067  * \param[in,out] val   number of stripes
2068  *
2069  * A value of "0" means "use the system-wide default stripe count", which
2070  * has either been inherited by now, or falls back to 1 stripe per file.
2071  * A value of "-1" (0xffffffff) means "stripe over all available OSTs",
2072  * and is a valid value, so is left unchanged here.
2073  */
2074 void lod_fix_desc_stripe_count(__u32 *val)
2075 {
2076         if (*val == 0)
2077                 *val = 1;
2078 }
2079
2080 /**
2081  * set the filesystem default layout pattern
2082  *
2083  * \param[in,out] val   LOV_PATTERN_* layout
2084  *
2085  * A value of "0" means "use the system-wide default layout type", which
2086  * has either been inherited by now, or falls back to plain RAID0 striping.
2087  */
2088 void lod_fix_desc_pattern(__u32 *val)
2089 {
2090         /* from lov_setstripe */
2091         if ((*val != 0) && !lov_pattern_supported_normal_comp(*val)) {
2092                 LCONSOLE_WARN("lod: Unknown stripe pattern: %#x\n", *val);
2093                 *val = 0;
2094         }
2095 }
2096
2097 void lod_fix_lmv_desc_pattern(__u32 *val)
2098 {
2099         if ((*val) && !lmv_is_known_hash_type(*val)) {
2100                 LCONSOLE_WARN("lod: Unknown md stripe pattern: %#x\n", *val);
2101                 *val = 0;
2102         }
2103 }
2104
2105 void lod_fix_desc_qos_maxage(__u32 *val)
2106 {
2107         /* fix qos_maxage */
2108         if (*val == 0)
2109                 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
2110 }
2111
2112 /**
2113  * Used to fix insane default striping.
2114  *
2115  * \param[in] desc      striping description
2116  */
2117 void lod_fix_desc(struct lov_desc *desc)
2118 {
2119         lod_fix_desc_stripe_size(&desc->ld_default_stripe_size);
2120         lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
2121         lod_fix_desc_pattern(&desc->ld_pattern);
2122         lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
2123 }
2124
2125 static void lod_fix_lmv_desc(struct lmv_desc *desc)
2126 {
2127         desc->ld_active_tgt_count = 0;
2128         lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
2129         lod_fix_lmv_desc_pattern(&desc->ld_pattern);
2130         lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
2131 }
2132
2133 /**
2134  * Initialize the structures used to store pools and default striping.
2135  *
2136  * \param[in] lod       LOD device
2137  * \param[in] lcfg      configuration structure storing default striping.
2138  *
2139  * \retval              0 if initialization succeeds
2140  * \retval              negative error number on failure
2141  */
2142 int lod_pools_init(struct lod_device *lod, struct lustre_cfg *lcfg)
2143 {
2144         struct obd_device          *obd;
2145         struct lov_desc            *desc;
2146         int                         rc;
2147         ENTRY;
2148
2149         obd = class_name2obd(lustre_cfg_string(lcfg, 0));
2150         LASSERT(obd != NULL);
2151         obd->obd_lu_dev = &lod->lod_dt_dev.dd_lu_dev;
2152
2153         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
2154                 CERROR("LOD setup requires a descriptor\n");
2155                 RETURN(-EINVAL);
2156         }
2157
2158         desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
2159
2160         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
2161                 CERROR("descriptor size wrong: %d > %d\n",
2162                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
2163                 RETURN(-EINVAL);
2164         }
2165
2166         if (desc->ld_magic != LOV_DESC_MAGIC) {
2167                 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
2168                         CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
2169                                obd->obd_name, desc);
2170                         lustre_swab_lov_desc(desc);
2171                 } else {
2172                         CERROR("%s: Bad lov desc magic: %#x\n",
2173                                obd->obd_name, desc->ld_magic);
2174                         RETURN(-EINVAL);
2175                 }
2176         }
2177
2178         lod_fix_desc(desc);
2179
2180         desc->ld_active_tgt_count = 0;
2181         lod->lod_ost_descs.ltd_lov_desc = *desc;
2182
2183         /* NB: config doesn't contain lmv_desc, alter it via sysfs. */
2184         lod_fix_lmv_desc(&lod->lod_mdt_descs.ltd_lmv_desc);
2185
2186         lod->lod_sp_me = LUSTRE_SP_CLI;
2187
2188         /* Set up OST pool environment */
2189         lod->lod_pool_count = 0;
2190         rc = lod_pool_hash_init(&lod->lod_pools_hash_body);
2191         if (rc)
2192                 RETURN(-ENOMEM);
2193
2194         INIT_LIST_HEAD(&lod->lod_pool_list);
2195         lod->lod_pool_count = 0;
2196         rc = tgt_pool_init(&lod->lod_mdt_descs.ltd_tgt_pool, 0);
2197         if (rc)
2198                 GOTO(out_hash, rc);
2199
2200         rc = tgt_pool_init(&lod->lod_mdt_descs.ltd_qos.lq_rr.lqr_pool, 0);
2201         if (rc)
2202                 GOTO(out_mdt_pool, rc);
2203
2204         rc = tgt_pool_init(&lod->lod_ost_descs.ltd_tgt_pool, 0);
2205         if (rc)
2206                 GOTO(out_mdt_rr_pool, rc);
2207
2208         rc = tgt_pool_init(&lod->lod_ost_descs.ltd_qos.lq_rr.lqr_pool, 0);
2209         if (rc)
2210                 GOTO(out_ost_pool, rc);
2211
2212         RETURN(0);
2213
2214 out_ost_pool:
2215         tgt_pool_free(&lod->lod_ost_descs.ltd_tgt_pool);
2216 out_mdt_rr_pool:
2217         tgt_pool_free(&lod->lod_mdt_descs.ltd_qos.lq_rr.lqr_pool);
2218 out_mdt_pool:
2219         tgt_pool_free(&lod->lod_mdt_descs.ltd_tgt_pool);
2220 out_hash:
2221         lod_pool_hash_destroy(&lod->lod_pools_hash_body);
2222
2223         return rc;
2224 }
2225
2226 /**
2227  * Release the structures describing the pools.
2228  *
2229  * \param[in] lod       LOD device from which we release the structures
2230  *
2231  * \retval              0 always
2232  */
2233 int lod_pools_fini(struct lod_device *lod)
2234 {
2235         struct obd_device   *obd = lod2obd(lod);
2236         struct pool_desc    *pool, *tmp;
2237         ENTRY;
2238
2239         list_for_each_entry_safe(pool, tmp, &lod->lod_pool_list, pool_list) {
2240                 /* free pool structs */
2241                 CDEBUG(D_INFO, "delete pool %p\n", pool);
2242                 /* In the function below, .hs_keycmp resolves to
2243                  * pool_hashkey_keycmp() */
2244                 /* coverity[overrun-buffer-val] */
2245                 lod_pool_del(obd, pool->pool_name);
2246         }
2247
2248         lod_pool_hash_destroy(&lod->lod_pools_hash_body);
2249         tgt_pool_free(&lod->lod_ost_descs.ltd_qos.lq_rr.lqr_pool);
2250         tgt_pool_free(&lod->lod_ost_descs.ltd_tgt_pool);
2251         tgt_pool_free(&lod->lod_mdt_descs.ltd_qos.lq_rr.lqr_pool);
2252         tgt_pool_free(&lod->lod_mdt_descs.ltd_tgt_pool);
2253
2254         RETURN(0);
2255 }