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                         lu_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                                   OBD_CONNECT_FLAGS2;
 180         data->ocd_version = LUSTRE_VERSION_CODE;
 181         data->ocd_index = index;
 182
 183         if (strcmp(LUSTRE_OSC_NAME, type) == 0) {
 184                 for_ost = true;
 185                 data->ocd_connect_flags |= OBD_CONNECT_AT |
 186                                            OBD_CONNECT_FULL20 |
 187                                            OBD_CONNECT_INDEX |
 188 #ifdef HAVE_LRU_RESIZE_SUPPORT
 189                                            OBD_CONNECT_LRU_RESIZE |
 190 #endif
 191                                            OBD_CONNECT_MDS |
 192                                            OBD_CONNECT_REQPORTAL |
 193                                            OBD_CONNECT_SKIP_ORPHAN |
 194                                            OBD_CONNECT_FID |
 195                                            OBD_CONNECT_LVB_TYPE |
 196                                            OBD_CONNECT_VERSION |
 197                                            OBD_CONNECT_PINGLESS |
 198                                            OBD_CONNECT_LFSCK |
 199                                            OBD_CONNECT_BULK_MBITS;
 200                 data->ocd_connect_flags2 = OBD_CONNECT2_REPLAY_CREATE;
 201
 202                 data->ocd_group = tgt_index;
 203                 ltd = &lod->lod_ost_descs;
 204         } else {
 205                 struct obd_import *imp = obd->u.cli.cl_import;
 206
 207                 for_ost = false;
 208                 data->ocd_ibits_known = MDS_INODELOCK_UPDATE;
 209                 data->ocd_connect_flags |= OBD_CONNECT_ACL |
 210                                            OBD_CONNECT_IBITS |
 211                                            OBD_CONNECT_MDS_MDS |
 212                                            OBD_CONNECT_MULTIMODRPCS |
 213                                            OBD_CONNECT_FID |
 214                                            OBD_CONNECT_AT |
 215                                            OBD_CONNECT_FULL20 |
 216                                            OBD_CONNECT_LFSCK |
 217                                            OBD_CONNECT_BULK_MBITS;
 218                 spin_lock(&imp->imp_lock);
 219                 imp->imp_server_timeout = 1;
 220                 spin_unlock(&imp->imp_lock);
 221                 imp->imp_client->cli_request_portal = OUT_PORTAL;
 222                 CDEBUG(D_OTHER, "%s: Set 'mds' portal and timeout\n",
 223                       obd->obd_name);
 224                 ltd = &lod->lod_mdt_descs;
 225         }
 226
 227         rc = obd_connect(env, &exp, obd, &obd->obd_uuid, data, NULL);
 228         OBD_FREE_PTR(data);
 229         if (rc) {
 230                 CERROR("%s: cannot connect to next dev %s (%d)\n",
 231                        obd->obd_name, osp, rc);
 232                 GOTO(out_cleanup, rc);
 233         }
 234         connected = true;
 235
 236         /* Allocate ost descriptor and fill it */
 237         OBD_ALLOC_PTR(tgt_desc);
 238         if (!tgt_desc)
 239                 GOTO(out_cleanup, rc = -ENOMEM);
 240
 241         tgt_desc->ltd_tgt    = dt_dev;
 242         tgt_desc->ltd_exp    = exp;
 243         tgt_desc->ltd_uuid   = obd->u.cli.cl_target_uuid;
 244         tgt_desc->ltd_gen    = gen;
 245         tgt_desc->ltd_index  = index;
 246         tgt_desc->ltd_active = active;
 247
 248         down_write(&ltd->ltd_rw_sem);
 249         mutex_lock(&ltd->ltd_mutex);
 250         rc = ltd_add_tgt(ltd, tgt_desc);
 251         if (rc)
 252                 GOTO(out_mutex, rc);
 253
 254         rc = lu_qos_add_tgt(&ltd->ltd_qos, tgt_desc);
 255         if (rc)
 256                 GOTO(out_del_tgt, rc);
 257
 258         rc = lu_tgt_pool_add(&ltd->ltd_tgt_pool, index,
 259                           ltd->ltd_lov_desc.ld_tgt_count);
 260         if (rc) {
 261                 CERROR("%s: can't set up pool, failed with %d\n",
 262                        obd->obd_name, rc);
 263                 GOTO(out_del_tgt, rc);
 264         }
 265
 266         mutex_unlock(&ltd->ltd_mutex);
 267         up_write(&ltd->ltd_rw_sem);
 268
 269         if (lod->lod_recovery_completed)
 270                 lu_dev->ld_ops->ldo_recovery_complete(env, lu_dev);
 271
 272         if (!for_ost && lod->lod_initialized) {
 273                 rc = lod_sub_init_llog(env, lod, tgt_desc->ltd_tgt);
 274                 if (rc != 0) {
 275                         CERROR("%s: cannot start llog on %s:rc = %d\n",
 276                                lod2obd(lod)->obd_name, osp, rc);
 277                         GOTO(out_ltd, rc);
 278                 }
 279         }
 280
 281         rc = lfsck_add_target(env, lod->lod_child, dt_dev, exp, index, for_ost);
 282         if (rc != 0) {
 283                 CERROR("Fail to add LFSCK target: name = %s, type = %s, "
 284                        "index = %u, rc = %d\n", osp, type, index, rc);
 285                 GOTO(out_fini_llog, rc);
 286         }
 287         RETURN(rc);
 288 out_fini_llog:
 289         lod_sub_fini_llog(env, tgt_desc->ltd_tgt,
 290                           &tgt_desc->ltd_recovery_task);
 291 out_ltd:
 292         down_write(&ltd->ltd_rw_sem);
 293         mutex_lock(&ltd->ltd_mutex);
 294         lu_tgt_pool_remove(&ltd->ltd_tgt_pool, index);
 295 out_del_tgt:
 296         ltd_del_tgt(ltd, tgt_desc);
 297 out_mutex:
 298         mutex_unlock(&ltd->ltd_mutex);
 299         up_write(&ltd->ltd_rw_sem);
 300         OBD_FREE_PTR(tgt_desc);
 301 out_cleanup:
 302         /* XXX OSP needs us to send down LCFG_CLEANUP because it uses
 303          * objects from the MDT stack. See LU-7184. */
 304         lcfg = &lod_env_info(env)->lti_lustre_cfg;
 305         memset(lcfg, 0, sizeof(*lcfg));
 306         lcfg->lcfg_version = LUSTRE_CFG_VERSION;
 307         lcfg->lcfg_command = LCFG_CLEANUP;
 308         lu_dev->ld_ops->ldo_process_config(env, lu_dev, lcfg);
 309
 310         if (connected)
 311                 obd_disconnect(exp);
 312
 313         return rc;
 314 }
 315
 316 /**
 317  * Schedule target removal from the target table.
 318  *
 319  * Mark the device as dead. The device is not removed here because it may
 320  * still be in use. The device will be removed in lod_putref() when the
 321  * last reference is released.
 322  *
 323  * \param[in] env               execution environment for this thread
 324  * \param[in] lod               LOD device the target table belongs to
 325  * \param[in] ltd               target table
 326  * \param[in] tgt               target
 327  */
 328 static void __lod_del_device(const struct lu_env *env, struct lod_device *lod,
 329                              struct lod_tgt_descs *ltd, struct lu_tgt_desc *tgt)
 330 {
 331         lfsck_del_target(env, lod->lod_child, tgt->ltd_tgt, tgt->ltd_index,
 332                          !ltd->ltd_is_mdt);
 333
 334         if (!tgt->ltd_reap) {
 335                 tgt->ltd_reap = 1;
 336                 ltd->ltd_death_row++;
 337         }
 338 }
 339
 340 /**
 341  * Schedule removal of all the targets from the given target table.
 342  *
 343  * See more details in the description for __lod_del_device()
 344  *
 345  * \param[in] env               execution environment for this thread
 346  * \param[in] lod               LOD device the target table belongs to
 347  * \param[in] ltd               target table
 348  *
 349  * \retval                      0 always
 350  */
 351 int lod_fini_tgt(const struct lu_env *env, struct lod_device *lod,
 352                  struct lod_tgt_descs *ltd)
 353 {
 354         struct lu_tgt_desc *tgt;
 355
 356         if (ltd->ltd_tgts_size <= 0)
 357                 return 0;
 358
 359         lod_getref(ltd);
 360         mutex_lock(&ltd->ltd_mutex);
 361         ltd_foreach_tgt(ltd, tgt)
 362                 __lod_del_device(env, lod, ltd, tgt);
 363         mutex_unlock(&ltd->ltd_mutex);
 364         lod_putref(lod, ltd);
 365
 366         lu_tgt_descs_fini(ltd);
 367
 368         return 0;
 369 }
 370
 371 /**
 372  * Remove device by name.
 373  *
 374  * Remove a device identified by \a osp from the target table. Given
 375  * the device can be in use, the real deletion happens in lod_putref().
 376  *
 377  * \param[in] env               execution environment for this thread
 378  * \param[in] lod               LOD device to be connected to the new OSP
 379  * \param[in] ltd               target table
 380  * \param[in] osp               name of OSP device to be removed
 381  * \param[in] idx               index of the target
 382  * \param[in] gen               generation number, not used currently
 383  *
 384  * \retval                      0 if the device was scheduled for removal
 385  * \retval                      -EINVAL if no device was found
 386  */
 387 int lod_del_device(const struct lu_env *env, struct lod_device *lod,
 388                    struct lod_tgt_descs *ltd, char *osp, unsigned int idx,
 389                    unsigned int gen)
 390 {
 391         struct obd_device *obd;
 392         struct lu_tgt_desc *tgt;
 393         struct obd_uuid uuid;
 394         int rc = 0;
 395
 396         ENTRY;
 397
 398         CDEBUG(D_CONFIG, "osp:%s idx:%d gen:%d\n", osp, idx, gen);
 399
 400         obd_str2uuid(&uuid, osp);
 401
 402         obd = class_find_client_obd(&uuid, LUSTRE_OSP_NAME,
 403                                    &lod->lod_dt_dev.dd_lu_dev.ld_obd->obd_uuid);
 404         if (obd == NULL) {
 405                 CERROR("can't find %s device\n", osp);
 406                 RETURN(-EINVAL);
 407         }
 408
 409         if (gen <= 0) {
 410                 CERROR("%s: request to remove OBD %s with invalid generation %d"
 411                        "\n", obd->obd_name, osp, gen);
 412                 RETURN(-EINVAL);
 413         }
 414
 415         obd_str2uuid(&uuid,  osp);
 416
 417         lod_getref(ltd);
 418         mutex_lock(&ltd->ltd_mutex);
 419         tgt = LTD_TGT(ltd, idx);
 420         /* check that the index is allocated in the bitmap */
 421         if (!test_bit(idx, ltd->ltd_tgt_bitmap) || !tgt) {
 422                 CERROR("%s: device %d is not set up\n", obd->obd_name, idx);
 423                 GOTO(out, rc = -EINVAL);
 424         }
 425
 426         /* check that the UUID matches */
 427         if (!obd_uuid_equals(&uuid, &tgt->ltd_uuid)) {
 428                 CERROR("%s: LOD target UUID %s at index %d does not match %s\n",
 429                        obd->obd_name, obd_uuid2str(&tgt->ltd_uuid), idx, osp);
 430                 GOTO(out, rc = -EINVAL);
 431         }
 432
 433         __lod_del_device(env, lod, ltd, tgt);
 434         EXIT;
 435 out:
 436         mutex_unlock(&ltd->ltd_mutex);
 437         lod_putref(lod, ltd);
 438         return(rc);
 439 }
 440
 441 /**
 442  * Resize per-thread storage to hold specified size.
 443  *
 444  * A helper function to resize per-thread temporary storage. This storage
 445  * is used to process LOV/LVM EAs and may be quite large. We do not want to
 446  * allocate/release it every time, so instead we put it into the env and
 447  * reallocate on demand. The memory is released when the correspondent thread
 448  * is finished.
 449  *
 450  * \param[in] info              LOD-specific storage in the environment
 451  * \param[in] size              new size to grow the buffer to
 452
 453  * \retval                      0 on success, -ENOMEM if reallocation failed
 454  */
 455 int lod_ea_store_resize(struct lod_thread_info *info, size_t size)
 456 {
 457         __u32 round = size_roundup_power2(size);
 458
 459         if (info->lti_ea_store) {
 460                 LASSERT(info->lti_ea_store_size);
 461                 LASSERT(info->lti_ea_store_size < round);
 462                 CDEBUG(D_INFO, "EA store size %d is not enough, need %d\n",
 463                        info->lti_ea_store_size, round);
 464                 OBD_FREE_LARGE(info->lti_ea_store, info->lti_ea_store_size);
 465                 info->lti_ea_store = NULL;
 466                 info->lti_ea_store_size = 0;
 467         }
 468
 469         OBD_ALLOC_LARGE(info->lti_ea_store, round);
 470         if (info->lti_ea_store == NULL)
 471                 RETURN(-ENOMEM);
 472         info->lti_ea_store_size = round;
 473
 474         RETURN(0);
 475 }
 476
 477 static void lod_free_comp_buffer(struct lod_layout_component *entries,
 478                                  __u16 count, __u32 bufsize)
 479 {
 480         struct lod_layout_component *entry;
 481         int i;
 482
 483         for (i = 0; i < count; i++) {
 484                 entry = &entries[i];
 485                 if (entry->llc_pool != NULL)
 486                         lod_set_pool(&entry->llc_pool, NULL);
 487                 if (entry->llc_ostlist.op_array)
 488                         OBD_FREE(entry->llc_ostlist.op_array,
 489                                  entry->llc_ostlist.op_size);
 490                 LASSERT(entry->llc_stripe == NULL);
 491                 LASSERT(entry->llc_stripes_allocated == 0);
 492         }
 493
 494         if (bufsize != 0)
 495                 OBD_FREE_LARGE(entries, bufsize);
 496 }
 497
 498 void lod_free_def_comp_entries(struct lod_default_striping *lds)
 499 {
 500         lod_free_comp_buffer(lds->lds_def_comp_entries,
 501                              lds->lds_def_comp_size_cnt,
 502                              size_roundup_power2(
 503                                      sizeof(*lds->lds_def_comp_entries) *
 504                                      lds->lds_def_comp_size_cnt));
 505         lds->lds_def_comp_entries = NULL;
 506         lds->lds_def_comp_cnt = 0;
 507         lds->lds_def_striping_is_composite = 0;
 508         lds->lds_def_comp_size_cnt = 0;
 509 }
 510
 511 /**
 512  * Resize per-thread storage to hold default striping component entries
 513  *
 514  * A helper function to resize per-thread temporary storage. This storage
 515  * is used to hold default LOV/LVM EAs and may be quite large. We do not want
 516  * to allocate/release it every time, so instead we put it into the env and
 517  * reallocate it on demand. The memory is released when the correspondent
 518  * thread is finished.
 519  *
 520  * \param[in,out] lds           default striping
 521  * \param[in] count             new component count to grow the buffer to
 522
 523  * \retval                      0 on success, -ENOMEM if reallocation failed
 524  */
 525 int lod_def_striping_comp_resize(struct lod_default_striping *lds, __u16 count)
 526 {
 527         struct lod_layout_component *entries;
 528         __u32 new = size_roundup_power2(sizeof(*lds->lds_def_comp_entries) *
 529                                         count);
 530         __u32 old = size_roundup_power2(sizeof(*lds->lds_def_comp_entries) *
 531                                         lds->lds_def_comp_size_cnt);
 532
 533         if (new <= old)
 534                 return 0;
 535
 536         OBD_ALLOC_LARGE(entries, new);
 537         if (entries == NULL)
 538                 return -ENOMEM;
 539
 540         if (lds->lds_def_comp_entries != NULL) {
 541                 CDEBUG(D_INFO, "default striping component size %d is not "
 542                        "enough, need %d\n", old, new);
 543                 lod_free_def_comp_entries(lds);
 544         }
 545
 546         lds->lds_def_comp_entries = entries;
 547         lds->lds_def_comp_size_cnt = count;
 548
 549         RETURN(0);
 550 }
 551
 552 void lod_free_comp_entries(struct lod_object *lo)
 553 {
 554         if (lo->ldo_mirrors) {
 555                 OBD_FREE_PTR_ARRAY(lo->ldo_mirrors, lo->ldo_mirror_count);
 556                 lo->ldo_mirrors = NULL;
 557                 lo->ldo_mirror_count = 0;
 558         }
 559         lod_free_comp_buffer(lo->ldo_comp_entries,
 560                              lo->ldo_comp_cnt,
 561                              sizeof(*lo->ldo_comp_entries) * lo->ldo_comp_cnt);
 562         lo->ldo_comp_entries = NULL;
 563         lo->ldo_comp_cnt = 0;
 564         lo->ldo_is_composite = 0;
 565 }
 566
 567 int lod_alloc_comp_entries(struct lod_object *lo,
 568                            int mirror_count, int comp_count)
 569 {
 570         LASSERT(comp_count != 0);
 571         LASSERT(lo->ldo_comp_cnt == 0 && lo->ldo_comp_entries == NULL);
 572
 573         if (mirror_count > 0) {
 574                 OBD_ALLOC_PTR_ARRAY(lo->ldo_mirrors, mirror_count);
 575                 if (!lo->ldo_mirrors)
 576                         return -ENOMEM;
 577
 578                 lo->ldo_mirror_count = mirror_count;
 579         }
 580
 581         OBD_ALLOC_LARGE(lo->ldo_comp_entries,
 582                         sizeof(*lo->ldo_comp_entries) * comp_count);
 583         if (lo->ldo_comp_entries == NULL) {
 584                 OBD_FREE_PTR_ARRAY(lo->ldo_mirrors, mirror_count);
 585                 lo->ldo_mirrors = NULL;
 586                 lo->ldo_mirror_count = 0;
 587                 return -ENOMEM;
 588         }
 589
 590         lo->ldo_comp_cnt = comp_count;
 591         lo->ldo_is_foreign = 0;
 592         return 0;
 593 }
 594
 595 int lod_fill_mirrors(struct lod_object *lo)
 596 {
 597         struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
 598         struct lod_layout_component *lod_comp;
 599         bool found_preferred = false;
 600         int mirror_idx = -1;
 601         __u16 mirror_id = 0xffff;
 602         int i, pref;
 603         ENTRY;
 604
 605         LASSERT(equi(!lo->ldo_is_composite, lo->ldo_mirror_count == 0));
 606
 607         if (!lo->ldo_is_composite)
 608                 RETURN(0);
 609
 610         lod_comp = &lo->ldo_comp_entries[0];
 611
 612         for (i = 0; i < lo->ldo_comp_cnt; i++, lod_comp++) {
 613                 bool stale = lod_comp->llc_flags & LCME_FL_STALE;
 614                 bool preferred = lod_comp->llc_flags & LCME_FL_PREF_WR;
 615                 bool init = (lod_comp->llc_stripe != NULL) &&
 616                             !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
 617                             !(lod_comp->llc_pattern & LOV_PATTERN_MDT);
 618                 int j;
 619
 620                 pref = 0;
 621                 /* calculate component preference over all used OSTs */
 622                 for (j = 0; init && j < lod_comp->llc_stripes_allocated; j++) {
 623                         __u32 idx = lod_comp->llc_ost_indices[j];
 624                         struct lod_tgt_desc *ltd;
 625
 626                         if (lod_comp->llc_stripe[j] == NULL)
 627                                 continue;
 628
 629                         if (unlikely(idx >= lod->lod_ost_descs.ltd_tgts_size)) {
 630                                 CERROR("%s: "DFID" OST idx %u > max %u\n",
 631                                        lod2obd(lod)->obd_name,
 632                                        PFID(lu_object_fid(&lo->ldo_obj.do_lu)),
 633                                        idx, lod->lod_ost_descs.ltd_tgts_size);
 634                                 continue;
 635                         }
 636                         ltd = OST_TGT(lod, idx);
 637                         if (unlikely(!ltd)) {
 638                                 CERROR("%s: "DFID" OST idx %u is NULL\n",
 639                                        lod2obd(lod)->obd_name,
 640                                        PFID(lu_object_fid(&lo->ldo_obj.do_lu)),
 641                                        idx);
 642                                 continue;
 643                         }
 644
 645                         if (ltd->ltd_statfs.os_state & OS_STATFS_NONROT)
 646                                 pref++;
 647                 }
 648
 649                 if (mirror_id_of(lod_comp->llc_id) == mirror_id) {
 650                         lo->ldo_mirrors[mirror_idx].lme_stale |= stale;
 651                         lo->ldo_mirrors[mirror_idx].lme_prefer |= preferred;
 652                         lo->ldo_mirrors[mirror_idx].lme_preference += pref;
 653                         lo->ldo_mirrors[mirror_idx].lme_end = i;
 654                         continue;
 655                 }
 656
 657                 if (mirror_idx >= 0 && preferred &&
 658                     !lo->ldo_mirrors[mirror_idx].lme_stale)
 659                         found_preferred = true;
 660
 661                 /* new mirror */
 662                 ++mirror_idx;
 663                 if (mirror_idx >= lo->ldo_mirror_count)
 664                         RETURN(-EINVAL);
 665
 666                 mirror_id = mirror_id_of(lod_comp->llc_id);
 667
 668                 lo->ldo_mirrors[mirror_idx].lme_id = mirror_id;
 669                 lo->ldo_mirrors[mirror_idx].lme_stale = stale;
 670                 lo->ldo_mirrors[mirror_idx].lme_prefer = preferred;
 671                 lo->ldo_mirrors[mirror_idx].lme_preference = pref;
 672                 lo->ldo_mirrors[mirror_idx].lme_start = i;
 673                 lo->ldo_mirrors[mirror_idx].lme_end = i;
 674         }
 675         if (mirror_idx != lo->ldo_mirror_count - 1)
 676                 RETURN(-EINVAL);
 677
 678         if (!found_preferred && mirror_idx > 0) {
 679                 int best = -1;
 680
 681                 /*
 682                  * if no explicited preferred found, then find a mirror
 683                  * with higher number of non-rotational OSTs
 684                  * */
 685                 pref = -1;
 686                 for (i = 0; i <= mirror_idx; i++) {
 687                         if (lo->ldo_mirrors[i].lme_stale)
 688                                 continue;
 689                         if (lo->ldo_mirrors[i].lme_preference > pref) {
 690                                 pref = lo->ldo_mirrors[i].lme_preference;
 691                                 best = i;
 692                         }
 693                 }
 694
 695                 LASSERT(best >= 0);
 696                 lo->ldo_mirrors[best].lme_prefer = 1;
 697         }
 698
 699         RETURN(0);
 700 }
 701
 702 /**
 703  * Generate on-disk lov_mds_md structure for each layout component based on
 704  * the information in lod_object->ldo_comp_entries[i].
 705  *
 706  * \param[in] env               execution environment for this thread
 707  * \param[in] lo                LOD object
 708  * \param[in] comp_idx          index of ldo_comp_entries
 709  * \param[in] lmm               buffer to cotain the on-disk lov_mds_md
 710  * \param[in|out] lmm_size      buffer size/lmm size
 711  * \param[in] is_dir            generate lov ea for dir or file? For dir case,
 712  *                              the stripe info is from the default stripe
 713  *                              template, which is collected in lod_ah_init(),
 714  *                              either from parent object or root object; for
 715  *                              file case, it's from the @lo object
 716  *
 717  * \retval                      0 if on disk structure is created successfully
 718  * \retval                      negative error number on failure
 719  */
 720 static int lod_gen_component_ea(const struct lu_env *env,
 721                                 struct lod_object *lo, int comp_idx,
 722                                 struct lov_mds_md *lmm, int *lmm_size,
 723                                 bool is_dir)
 724 {
 725         struct lod_thread_info  *info = lod_env_info(env);
 726         const struct lu_fid     *fid  = lu_object_fid(&lo->ldo_obj.do_lu);
 727         struct lod_device       *lod;
 728         struct lov_ost_data_v1  *objs;
 729         struct lod_layout_component *lod_comp;
 730         __u32   magic;
 731         __u16 stripe_count;
 732         int     i, rc = 0;
 733         ENTRY;
 734
 735         LASSERT(lo);
 736         if (is_dir)
 737                 lod_comp =
 738                         &lo->ldo_def_striping->lds_def_comp_entries[comp_idx];
 739         else
 740                 lod_comp = &lo->ldo_comp_entries[comp_idx];
 741
 742         magic = lod_comp->llc_pool != NULL ? LOV_MAGIC_V3 : LOV_MAGIC_V1;
 743         if (lod_comp->llc_pattern == 0) /* default striping */
 744                 lod_comp->llc_pattern = LOV_PATTERN_RAID0;
 745
 746         lmm->lmm_magic = cpu_to_le32(magic);
 747         lmm->lmm_pattern = cpu_to_le32(lod_comp->llc_pattern);
 748         fid_to_lmm_oi(fid, &lmm->lmm_oi);
 749         if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_LMMOI))
 750                 lmm->lmm_oi.oi.oi_id++;
 751         lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
 752
 753         lmm->lmm_stripe_size = cpu_to_le32(lod_comp->llc_stripe_size);
 754         lmm->lmm_stripe_count = cpu_to_le16(lod_comp->llc_stripe_count);
 755         /**
 756          * for dir and uninstantiated component, lmm_layout_gen stores
 757          * default stripe offset.
 758          */
 759         lmm->lmm_layout_gen =
 760                 (is_dir || !lod_comp_inited(lod_comp)) ?
 761                         cpu_to_le16(lod_comp->llc_stripe_offset) :
 762                         cpu_to_le16(lod_comp->llc_layout_gen);
 763
 764         if (magic == LOV_MAGIC_V1) {
 765                 objs = &lmm->lmm_objects[0];
 766         } else {
 767                 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
 768                 size_t cplen = strlcpy(v3->lmm_pool_name,
 769                                        lod_comp->llc_pool,
 770                                        sizeof(v3->lmm_pool_name));
 771                 if (cplen >= sizeof(v3->lmm_pool_name))
 772                         RETURN(-E2BIG);
 773                 objs = &v3->lmm_objects[0];
 774         }
 775         lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
 776         stripe_count = lod_comp_entry_stripe_count(lo, comp_idx, is_dir);
 777         if (stripe_count == 0 && !is_dir &&
 778             !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
 779             !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) {
 780                 /* Try again if all active targets are disconnected.
 781                  * It is possible when MDS does failover. */
 782                 if (!lod->lod_ost_active_count &&
 783                     lod->lod_ost_count)
 784                         RETURN(-EAGAIN);
 785                 RETURN(-E2BIG);
 786         }
 787
 788         if (!is_dir && lo->ldo_is_composite)
 789                 lod_comp_shrink_stripe_count(lod_comp, &stripe_count);
 790
 791         if (is_dir || lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)
 792                 GOTO(done, rc = 0);
 793
 794         /* generate ost_idx of this component stripe */
 795         for (i = 0; i < stripe_count; i++) {
 796                 struct dt_object *object;
 797                 __u32 ost_idx = (__u32)-1UL;
 798                 int type = LU_SEQ_RANGE_OST;
 799
 800                 if (lod_comp->llc_stripe && lod_comp->llc_stripe[i]) {
 801                         object = lod_comp->llc_stripe[i];
 802                         /* instantiated component */
 803                         info->lti_fid = *lu_object_fid(&object->do_lu);
 804
 805                         if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_MULTIPLE_REF) &&
 806                             comp_idx == 0) {
 807                                 if (cfs_fail_val == 0)
 808                                         cfs_fail_val = info->lti_fid.f_oid;
 809                                 else if (i == 0)
 810                                         info->lti_fid.f_oid = cfs_fail_val;
 811                         }
 812
 813                         rc = fid_to_ostid(&info->lti_fid, &info->lti_ostid);
 814                         LASSERT(rc == 0);
 815
 816                         ostid_cpu_to_le(&info->lti_ostid, &objs[i].l_ost_oi);
 817                         objs[i].l_ost_gen = cpu_to_le32(0);
 818                         if (CFS_FAIL_CHECK(OBD_FAIL_MDS_FLD_LOOKUP))
 819                                 rc = -ENOENT;
 820                         else
 821                                 rc = lod_fld_lookup(env, lod, &info->lti_fid,
 822                                                     &ost_idx, &type);
 823                         if (rc < 0) {
 824                                 CERROR("%s: Can not locate "DFID": rc = %d\n",
 825                                        lod2obd(lod)->obd_name,
 826                                        PFID(&info->lti_fid), rc);
 827                                 RETURN(rc);
 828                         }
 829                 } else if (lod_comp->llc_ostlist.op_array &&
 830                            lod_comp->llc_ostlist.op_count) {
 831                         /* user specified ost list */
 832                         ost_idx = lod_comp->llc_ostlist.op_array[i];
 833                 }
 834                 /*
 835                  * with un-instantiated or with no specified ost list
 836                  * component, its l_ost_idx does not matter.
 837                  */
 838                 objs[i].l_ost_idx = cpu_to_le32(ost_idx);
 839         }
 840 done:
 841         if (lmm_size != NULL)
 842                 *lmm_size = lov_mds_md_size(stripe_count, magic);
 843         RETURN(rc);
 844 }
 845
 846 /**
 847  * Generate on-disk lov_mds_md structure based on the information in
 848  * the lod_object->ldo_comp_entries.
 849  *
 850  * \param[in] env               execution environment for this thread
 851  * \param[in] lo                LOD object
 852  * \param[in] lmm               buffer to cotain the on-disk lov_mds_md
 853  * \param[in|out] lmm_size      buffer size/lmm size
 854  * \param[in] is_dir            generate lov ea for dir or file? For dir case,
 855  *                              the stripe info is from the default stripe
 856  *                              template, which is collected in lod_ah_init(),
 857  *                              either from parent object or root object; for
 858  *                              file case, it's from the @lo object
 859  *
 860  * \retval                      0 if on disk structure is created successfully
 861  * \retval                      negative error number on failure
 862  */
 863 int lod_generate_lovea(const struct lu_env *env, struct lod_object *lo,
 864                        struct lov_mds_md *lmm, int *lmm_size, bool is_dir)
 865 {
 866         struct lov_comp_md_entry_v1 *lcme;
 867         struct lov_comp_md_v1 *lcm;
 868         struct lod_layout_component *comp_entries;
 869         __u16 comp_cnt, mirror_cnt;
 870         bool is_composite, is_foreign = false;
 871         int i, rc = 0, offset;
 872         ENTRY;
 873
 874         if (is_dir) {
 875                 comp_cnt = lo->ldo_def_striping->lds_def_comp_cnt;
 876                 mirror_cnt = lo->ldo_def_striping->lds_def_mirror_cnt;
 877                 comp_entries = lo->ldo_def_striping->lds_def_comp_entries;
 878                 is_composite =
 879                         lo->ldo_def_striping->lds_def_striping_is_composite;
 880         } else {
 881                 comp_cnt = lo->ldo_comp_cnt;
 882                 mirror_cnt = lo->ldo_mirror_count;
 883                 comp_entries = lo->ldo_comp_entries;
 884                 is_composite = lo->ldo_is_composite;
 885                 is_foreign = lo->ldo_is_foreign;
 886         }
 887
 888         LASSERT(lmm_size != NULL);
 889
 890         if (is_foreign) {
 891                 struct lov_foreign_md *lfm;
 892
 893                 lfm = (struct lov_foreign_md *)lmm;
 894                 memcpy(lfm, lo->ldo_foreign_lov, lo->ldo_foreign_lov_size);
 895                 /* need to store little-endian */
 896                 if (cpu_to_le32(LOV_MAGIC_FOREIGN) != LOV_MAGIC_FOREIGN) {
 897                         __swab32s(&lfm->lfm_magic);
 898                         __swab32s(&lfm->lfm_length);
 899                         __swab32s(&lfm->lfm_type);
 900                         __swab32s(&lfm->lfm_flags);
 901                 }
 902                 *lmm_size = lo->ldo_foreign_lov_size;
 903                 RETURN(0);
 904         }
 905
 906         LASSERT(comp_cnt != 0 && comp_entries != NULL);
 907
 908         if (!is_composite) {
 909                 rc = lod_gen_component_ea(env, lo, 0, lmm, lmm_size, is_dir);
 910                 RETURN(rc);
 911         }
 912
 913         lcm = (struct lov_comp_md_v1 *)lmm;
 914         memset(lcm, 0, sizeof(*lcm));
 915
 916         lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
 917         lcm->lcm_entry_count = cpu_to_le16(comp_cnt);
 918         lcm->lcm_mirror_count = cpu_to_le16(mirror_cnt - 1);
 919         if (mirror_cnt > 1)
 920                 lcm->lcm_flags = cpu_to_le16(lo->ldo_flr_state);
 921         else
 922                 lcm->lcm_flags = LCM_FL_NONE;
 923
 924         offset = sizeof(*lcm) + sizeof(*lcme) * comp_cnt;
 925         LASSERT(offset % sizeof(__u64) == 0);
 926
 927         for (i = 0; i < comp_cnt; i++) {
 928                 struct lod_layout_component *lod_comp;
 929                 struct lov_mds_md *sub_md;
 930                 int size;
 931
 932                 lod_comp = &comp_entries[i];
 933                 lcme = &lcm->lcm_entries[i];
 934
 935                 LASSERT(ergo(!is_dir, lod_comp->llc_id != LCME_ID_INVAL));
 936                 lcme->lcme_id = cpu_to_le32(lod_comp->llc_id);
 937
 938                 /* component could be un-inistantiated */
 939                 lcme->lcme_flags = cpu_to_le32(lod_comp->llc_flags);
 940                 if (lod_comp->llc_flags & LCME_FL_NOSYNC)
 941                         lcme->lcme_timestamp =
 942                                 cpu_to_le64(lod_comp->llc_timestamp);
 943                 if (lod_comp->llc_flags & LCME_FL_EXTENSION && !is_dir)
 944                         lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_SEL);
 945
 946                 lcme->lcme_extent.e_start =
 947                         cpu_to_le64(lod_comp->llc_extent.e_start);
 948                 lcme->lcme_extent.e_end =
 949                         cpu_to_le64(lod_comp->llc_extent.e_end);
 950                 lcme->lcme_offset = cpu_to_le32(offset);
 951
 952                 sub_md = (struct lov_mds_md *)((char *)lcm + offset);
 953                 rc = lod_gen_component_ea(env, lo, i, sub_md, &size, is_dir);
 954                 if (rc)
 955                         GOTO(out, rc);
 956                 lcme->lcme_size = cpu_to_le32(size);
 957                 offset += size;
 958                 LASSERTF((offset <= *lmm_size) && (offset % sizeof(__u64) == 0),
 959                          "offset:%d lmm_size:%d\n", offset, *lmm_size);
 960         }
 961         lcm->lcm_size = cpu_to_le32(offset);
 962         lcm->lcm_layout_gen = cpu_to_le32(is_dir ? 0 : lo->ldo_layout_gen);
 963
 964         lustre_print_user_md(D_LAYOUT, (struct lov_user_md *)lmm,
 965                              "generate lum");
 966 out:
 967         if (rc == 0)
 968                 *lmm_size = offset;
 969         RETURN(rc);
 970 }
 971
 972 /**
 973  * Get LOV EA.
 974  *
 975  * Fill lti_ea_store buffer in the environment with a value for the given
 976  * EA. The buffer is reallocated if the value doesn't fit.
 977  *
 978  * \param[in,out] env           execution environment for this thread
 979  *                              .lti_ea_store buffer is filled with EA's value
 980  * \param[in] lo                LOD object
 981  * \param[in] name              name of the EA
 982  *
 983  * \retval                      > 0 if EA is fetched successfully
 984  * \retval                      0 if EA is empty
 985  * \retval                      negative error number on failure
 986  */
 987 int lod_get_ea(const struct lu_env *env, struct lod_object *lo,
 988                const char *name)
 989 {
 990         struct lod_thread_info  *info = lod_env_info(env);
 991         struct dt_object        *next = dt_object_child(&lo->ldo_obj);
 992         int                     rc;
 993         ENTRY;
 994
 995         LASSERT(info);
 996
 997         if (unlikely(info->lti_ea_store == NULL)) {
 998                 /* just to enter in allocation block below */
 999                 rc = -ERANGE;
1000         } else {
1001 repeat:
1002                 info->lti_buf.lb_buf = info->lti_ea_store;
1003                 info->lti_buf.lb_len = info->lti_ea_store_size;
1004                 rc = dt_xattr_get(env, next, &info->lti_buf, name);
1005         }
1006
1007         /* if object is not striped or inaccessible */
1008         if (rc == -ENODATA || rc == -ENOENT)
1009                 RETURN(0);
1010
1011         if (rc == -ERANGE) {
1012                 /* EA doesn't fit, reallocate new buffer */
1013                 rc = dt_xattr_get(env, next, &LU_BUF_NULL, name);
1014                 if (rc == -ENODATA || rc == -ENOENT)
1015                         RETURN(0);
1016                 else if (rc < 0)
1017                         RETURN(rc);
1018
1019                 LASSERT(rc > 0);
1020                 rc = lod_ea_store_resize(info, rc);
1021                 if (rc)
1022                         RETURN(rc);
1023                 goto repeat;
1024         }
1025
1026         RETURN(rc);
1027 }
1028
1029 /**
1030  * Verify the target index is present in the current configuration.
1031  *
1032  * \param[in] md                LOD device where the target table is stored
1033  * \param[in] idx               target's index
1034  *
1035  * \retval                      0 if the index is present
1036  * \retval                      -EINVAL if not
1037  */
1038 int validate_lod_and_idx(struct lod_device *md, __u32 idx)
1039 {
1040         if (unlikely(idx >= md->lod_ost_descs.ltd_tgts_size ||
1041                      !test_bit(idx, md->lod_ost_bitmap))) {
1042                 CERROR("%s: bad idx: %d of %d\n", lod2obd(md)->obd_name, idx,
1043                        md->lod_ost_descs.ltd_tgts_size);
1044                 return -EINVAL;
1045         }
1046
1047         if (unlikely(OST_TGT(md, idx) == NULL)) {
1048                 CERROR("%s: bad lod_tgt_desc for idx: %d\n",
1049                        lod2obd(md)->obd_name, idx);
1050                 return -EINVAL;
1051         }
1052
1053         if (unlikely(OST_TGT(md, idx)->ltd_tgt == NULL)) {
1054                 CERROR("%s: invalid lod device, for idx: %d\n",
1055                        lod2obd(md)->obd_name , idx);
1056                 return -EINVAL;
1057         }
1058
1059         return 0;
1060 }
1061
1062 /**
1063  * Instantiate objects for stripes.
1064  *
1065  * Allocate and initialize LU-objects representing the stripes. The number
1066  * of the stripes (llc_stripe_count) must be initialized already. The caller
1067  * must ensure nobody else is calling the function on the object at the same
1068  * time. FLDB service must be running to be able to map a FID to the targets
1069  * and find appropriate device representing that target.
1070  *
1071  * \param[in] env               execution environment for this thread
1072  * \param[in,out] lo            LOD object
1073  * \param[in] objs              an array of IDs to creates the objects from
1074  * \param[in] comp_idx          index of ldo_comp_entries
1075  *
1076  * \retval                      0 if the objects are instantiated successfully
1077  * \retval                      negative error number on failure
1078  */
1079 int lod_initialize_objects(const struct lu_env *env, struct lod_object *lo,
1080                            struct lov_ost_data_v1 *objs, int comp_idx)
1081 {
1082         struct lod_layout_component *lod_comp;
1083         struct lod_thread_info *info = lod_env_info(env);
1084         struct lod_device *md;
1085         struct lu_object *o, *n;
1086         struct lu_device *nd;
1087         struct dt_object **stripe = NULL;
1088         __u32 *ost_indices = NULL;
1089         int stripe_len;
1090         int i, rc = 0;
1091         __u32 idx;
1092         ENTRY;
1093
1094         LASSERT(lo != NULL);
1095         md = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1096
1097         LASSERT(lo->ldo_comp_cnt != 0 && lo->ldo_comp_entries != NULL);
1098         lod_comp = &lo->ldo_comp_entries[comp_idx];
1099
1100         LASSERT(lod_comp->llc_stripe == NULL);
1101         LASSERT(lod_comp->llc_stripe_count > 0);
1102         LASSERT(lod_comp->llc_stripe_size > 0);
1103
1104         stripe_len = lod_comp->llc_stripe_count;
1105         OBD_ALLOC_PTR_ARRAY(stripe, stripe_len);
1106         if (stripe == NULL)
1107                 RETURN(-ENOMEM);
1108         OBD_ALLOC_PTR_ARRAY(ost_indices, stripe_len);
1109         if (!ost_indices)
1110                 GOTO(out, rc = -ENOMEM);
1111
1112         for (i = 0; i < lod_comp->llc_stripe_count; i++) {
1113                 if (unlikely(lovea_slot_is_dummy(&objs[i])))
1114                         continue;
1115
1116                 ostid_le_to_cpu(&objs[i].l_ost_oi, &info->lti_ostid);
1117                 idx = le32_to_cpu(objs[i].l_ost_idx);
1118                 rc = ostid_to_fid(&info->lti_fid, &info->lti_ostid, idx);
1119                 if (rc != 0)
1120                         GOTO(out, rc);
1121                 LASSERTF(fid_is_sane(&info->lti_fid), ""DFID" insane!\n",
1122                          PFID(&info->lti_fid));
1123                 lod_getref(&md->lod_ost_descs);
1124
1125                 rc = validate_lod_and_idx(md, idx);
1126                 if (unlikely(rc != 0)) {
1127                         lod_putref(md, &md->lod_ost_descs);
1128                         GOTO(out, rc);
1129                 }
1130
1131                 nd = &OST_TGT(md, idx)->ltd_tgt->dd_lu_dev;
1132                 lod_putref(md, &md->lod_ost_descs);
1133
1134                 /* In the function below, .hs_keycmp resolves to
1135                  * u_obj_hop_keycmp() */
1136                 o = lu_object_find_at(env, nd, &info->lti_fid, NULL);
1137                 if (IS_ERR(o))
1138                         GOTO(out, rc = PTR_ERR(o));
1139
1140                 n = lu_object_locate(o->lo_header, nd->ld_type);
1141                 LASSERT(n);
1142
1143                 stripe[i] = container_of(n, struct dt_object, do_lu);
1144                 ost_indices[i] = idx;
1145         }
1146
1147 out:
1148         if (rc != 0) {
1149                 for (i = 0; i < stripe_len; i++)
1150                         if (stripe[i] != NULL)
1151                                 dt_object_put(env, stripe[i]);
1152
1153                 OBD_FREE_PTR_ARRAY(stripe, stripe_len);
1154                 lod_comp->llc_stripe_count = 0;
1155                 if (ost_indices)
1156                         OBD_FREE_PTR_ARRAY(ost_indices, stripe_len);
1157         } else {
1158                 lod_comp->llc_stripe = stripe;
1159                 lod_comp->llc_ost_indices = ost_indices;
1160                 lod_comp->llc_stripes_allocated = stripe_len;
1161         }
1162
1163         RETURN(rc);
1164 }
1165
1166 /**
1167  * Instantiate objects for striping.
1168  *
1169  * Parse striping information in \a buf and instantiate the objects
1170  * representing the stripes.
1171  *
1172  * \param[in] env               execution environment for this thread
1173  * \param[in] lo                LOD object
1174  * \param[in] buf               buffer storing LOV EA to parse
1175  * \param[in] lvf               verify flags when parsing the layout
1176  *
1177  * \retval                      0 if parsing and objects creation succeed
1178  * \retval                      negative error number on failure
1179  */
1180 int lod_parse_striping(const struct lu_env *env, struct lod_object *lo,
1181                        const struct lu_buf *buf, enum layout_verify_flags lvf)
1182 {
1183         struct lod_device *d = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1184         struct lov_mds_md_v1 *lmm;
1185         struct lov_comp_md_v1 *comp_v1 = NULL;
1186         struct lov_foreign_md *foreign = NULL;
1187         struct lov_ost_data_v1 *objs;
1188         __u32 magic, pattern;
1189         __u16 mirror_cnt = 0;
1190         __u16 comp_cnt;
1191         int i, rc;
1192         __u16 mirror_id = MIRROR_ID_NEG;
1193         bool stale = false;
1194         int stale_mirrors = 0;
1195         ENTRY;
1196
1197         LASSERT(buf);
1198         LASSERT(buf->lb_buf);
1199         LASSERT(buf->lb_len);
1200         LASSERT(mutex_is_locked(&lo->ldo_layout_mutex));
1201
1202         lmm = (struct lov_mds_md_v1 *)buf->lb_buf;
1203         magic = le32_to_cpu(lmm->lmm_magic);
1204
1205         if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 &&
1206             magic != LOV_MAGIC_COMP_V1 && magic != LOV_MAGIC_FOREIGN &&
1207             magic != LOV_MAGIC_SEL)
1208                 GOTO(out, rc = -EINVAL);
1209
1210         lod_striping_free_nolock(env, lo);
1211
1212         if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) {
1213                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1214                 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1215                 if (comp_cnt == 0)
1216                         GOTO(out, rc = -EINVAL);
1217                 lo->ldo_layout_gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1218                 lo->ldo_is_composite = 1;
1219                 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1220                 if (mirror_cnt > 1)
1221                         lo->ldo_flr_state = le16_to_cpu(comp_v1->lcm_flags) &
1222                                                         LCM_FL_FLR_MASK;
1223                 else
1224                         lo->ldo_flr_state = LCM_FL_NONE;
1225         } else if (magic == LOV_MAGIC_FOREIGN) {
1226                 size_t length;
1227
1228                 foreign = (struct lov_foreign_md *)buf->lb_buf;
1229                 length = offsetof(typeof(*foreign), lfm_value);
1230                 if (buf->lb_len < length ||
1231                     buf->lb_len < (length + le32_to_cpu(foreign->lfm_length))) {
1232                         CDEBUG(D_LAYOUT,
1233                                "buf len %zu too small for lov_foreign_md\n",
1234                                buf->lb_len);
1235                         GOTO(out, rc = -EINVAL);
1236                 }
1237
1238                 /* just cache foreign LOV EA raw */
1239                 rc = lod_alloc_foreign_lov(lo, length);
1240                 if (rc)
1241                         GOTO(out, rc);
1242                 memcpy(lo->ldo_foreign_lov, buf->lb_buf, length);
1243                 GOTO(out, rc);
1244         } else {
1245                 comp_cnt = 1;
1246                 lo->ldo_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1247                 lo->ldo_is_composite = 0;
1248         }
1249
1250         rc = lod_alloc_comp_entries(lo, mirror_cnt, comp_cnt);
1251         if (rc)
1252                 GOTO(out, rc);
1253
1254         for (i = 0; i < comp_cnt; i++) {
1255                 struct lod_layout_component *lod_comp;
1256                 struct lu_extent *ext;
1257                 __u32 offs;
1258
1259                 lod_comp = &lo->ldo_comp_entries[i];
1260                 if (lo->ldo_is_composite) {
1261                         offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1262                         lmm = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs);
1263
1264                         ext = &comp_v1->lcm_entries[i].lcme_extent;
1265                         lod_comp->llc_extent.e_start =
1266                                 le64_to_cpu(ext->e_start);
1267                         if (lod_comp->llc_extent.e_start &
1268                             (LOV_MIN_STRIPE_SIZE - 1)) {
1269                                 CDEBUG(D_LAYOUT,
1270                                        "extent start %llu is not a multiple of min size %u\n",
1271                                        lod_comp->llc_extent.e_start,
1272                                        LOV_MIN_STRIPE_SIZE);
1273                                 GOTO(out, rc = -EINVAL);
1274                         }
1275
1276                         lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end);
1277                         if (lod_comp->llc_extent.e_end != LUSTRE_EOF &&
1278                             lod_comp->llc_extent.e_end &
1279                             (LOV_MIN_STRIPE_SIZE - 1)) {
1280                                 CDEBUG(D_LAYOUT,
1281                                        "extent end %llu is not a multiple of min size %u\n",
1282                                        lod_comp->llc_extent.e_end,
1283                                        LOV_MIN_STRIPE_SIZE);
1284                                 GOTO(out, rc = -EINVAL);
1285                         }
1286
1287                         lod_comp->llc_flags =
1288                                 le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags);
1289
1290                         if (lod_comp->llc_flags & LCME_FL_NOSYNC)
1291                                 lod_comp->llc_timestamp = le64_to_cpu(
1292                                         comp_v1->lcm_entries[i].lcme_timestamp);
1293                         lod_comp->llc_id =
1294                                 le32_to_cpu(comp_v1->lcm_entries[i].lcme_id);
1295                         if (lod_comp->llc_id == LCME_ID_INVAL)
1296                                 GOTO(out, rc = -EINVAL);
1297
1298                         if (lvf & LVF_ALL_STALE) {
1299                                 if (mirror_id_of(lod_comp->llc_id) ==
1300                                     mirror_id) {
1301                                         /* remaining comps in the mirror */
1302                                         stale |= lod_comp->llc_flags &
1303                                                  LCME_FL_STALE;
1304                                 } else {
1305                                         /*
1306                                          * new mirror, check last mirror's
1307                                          * stale-ness
1308                                          */
1309                                         if (stale)
1310                                                 stale_mirrors++;
1311
1312                                         mirror_id =
1313                                                 mirror_id_of(lod_comp->llc_id);
1314
1315                                         /* the first comp of the new mirror */
1316                                         stale = lod_comp->llc_flags &
1317                                                 LCME_FL_STALE;
1318                                 }
1319                         }
1320
1321                         if ((lod_comp->llc_flags & LCME_FL_EXTENSION) &&
1322                             comp_v1->lcm_magic != cpu_to_le32(LOV_MAGIC_SEL)) {
1323                                 CWARN("%s: EXTENSION flags=%x set on component[%u]=%x of non-SEL file "DFID" with magic=%#08x\n",
1324                                       lod2obd(d)->obd_name,
1325                                       lod_comp->llc_flags, lod_comp->llc_id, i,
1326                                       PFID(lod_object_fid(lo)),
1327                                       le32_to_cpu(comp_v1->lcm_magic));
1328                         }
1329                 } else {
1330                         lod_comp_set_init(lod_comp);
1331                 }
1332
1333                 pattern = le32_to_cpu(lmm->lmm_pattern);
1334                 if (!lov_pattern_supported(lov_pattern(pattern)))
1335                         GOTO(out, rc = -EINVAL);
1336
1337                 lod_comp->llc_pattern = pattern;
1338                 lod_comp->llc_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
1339                 lod_comp->llc_stripe_count = le16_to_cpu(lmm->lmm_stripe_count);
1340                 lod_comp->llc_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1341
1342                 if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)) {
1343                         struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
1344
1345                         lod_set_pool(&lod_comp->llc_pool, v3->lmm_pool_name);
1346                         objs = &v3->lmm_objects[0];
1347                 } else {
1348                         lod_set_pool(&lod_comp->llc_pool, NULL);
1349                         objs = &lmm->lmm_objects[0];
1350                 }
1351
1352                 /**
1353                  * If uninstantiated template component has valid l_ost_idx,
1354                  * then user has specified ost list for this component.
1355                  */
1356                 if (!lod_comp_inited(lod_comp)) {
1357                         __u16 stripe_count;
1358
1359                         if (objs[0].l_ost_idx != (__u32)-1UL) {
1360                                 int j;
1361
1362                                 stripe_count = lod_comp_entry_stripe_count(
1363                                                         lo, i, false);
1364                                 if (stripe_count == 0 &&
1365                                     !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1366                                     !(lod_comp->llc_pattern & LOV_PATTERN_MDT))
1367                                         GOTO(out, rc = -E2BIG);
1368                                 /**
1369                                  * load the user specified ost list, when this
1370                                  * component is instantiated later, it will be
1371                                  * used in lod_alloc_ost_list().
1372                                  */
1373                                 lod_comp->llc_ostlist.op_count = stripe_count;
1374                                 lod_comp->llc_ostlist.op_size =
1375                                         stripe_count * sizeof(__u32);
1376                                 OBD_ALLOC(lod_comp->llc_ostlist.op_array,
1377                                           lod_comp->llc_ostlist.op_size);
1378                                 if (!lod_comp->llc_ostlist.op_array)
1379                                         GOTO(out, rc = -ENOMEM);
1380
1381                                 for (j = 0; j < stripe_count; j++)
1382                                         lod_comp->llc_ostlist.op_array[j] =
1383                                                 le32_to_cpu(objs[j].l_ost_idx);
1384
1385                                 /**
1386                                  * this component OST objects starts from the
1387                                  * first ost_idx, lod_alloc_ost_list() will
1388                                  * check this.
1389                                  */
1390                                 lod_comp->llc_stripe_offset = objs[0].l_ost_idx;
1391                         } else {
1392                                 /**
1393                                  * for uninstantiated component,
1394                                  * lmm_layout_gen stores default stripe offset.
1395                                  */
1396                                 lod_comp->llc_stripe_offset =
1397                                                         lmm->lmm_layout_gen;
1398                         }
1399                 }
1400
1401                 /* skip un-instantiated component object initialization */
1402                 if (!lod_comp_inited(lod_comp))
1403                         continue;
1404
1405                 if (!(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1406                     !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) {
1407                         rc = lod_initialize_objects(env, lo, objs, i);
1408                         if (rc)
1409                                 GOTO(out, rc);
1410                 }
1411         }
1412
1413         if (lo->ldo_is_composite && (lvf & LVF_ALL_STALE)) {
1414                 /* check the last mirror stale-ness */
1415                 if (stale)
1416                         stale_mirrors++;
1417
1418                 if (mirror_cnt == stale_mirrors) {
1419                         rc = -EPERM;
1420                         CERROR("%s: can not set all stale mirrors for "
1421                                DFID": rc = %d\n",
1422                                lod2obd(d)->obd_name, PFID(lod_object_fid(lo)),
1423                                rc);
1424                         GOTO(out, rc);
1425                 }
1426         }
1427
1428         rc = lod_fill_mirrors(lo);
1429         if (rc)
1430                 GOTO(out, rc);
1431
1432 out:
1433         if (rc)
1434                 lod_striping_free_nolock(env, lo);
1435         RETURN(rc);
1436 }
1437
1438 /**
1439  * Check whether the striping (LOVEA for regular file, LMVEA for directory)
1440  * is already cached.
1441  *
1442  * \param[in] lo        LOD object
1443  *
1444  * \retval              True if the striping is cached, otherwise
1445  *                      return false.
1446  */
1447 static bool lod_striping_loaded(struct lod_object *lo)
1448 {
1449         if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr) &&
1450             lo->ldo_comp_cached)
1451                 return true;
1452
1453         if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1454                 if (lo->ldo_dir_stripe_loaded)
1455                         return true;
1456
1457                 /* Never load LMV stripe for slaves of striped dir */
1458                 if (lo->ldo_dir_slave_stripe)
1459                         return true;
1460         }
1461
1462         return false;
1463 }
1464
1465 /**
1466  * A generic function to initialize the stripe objects.
1467  *
1468  * A protected version of lod_striping_load_locked() - load the striping
1469  * information from storage, parse that and instantiate LU objects to
1470  * represent the stripes.  The LOD object \a lo supplies a pointer to the
1471  * next sub-object in the LU stack so we can lock it. Also use \a lo to
1472  * return an array of references to the newly instantiated objects.
1473  *
1474  * \param[in] env               execution environment for this thread
1475  * \param[in,out] lo            LOD object, where striping is stored and
1476  *                              which gets an array of references
1477  *
1478  * \retval                      0 if parsing and object creation succeed
1479  * \retval                      negative error number on failure
1480  **/
1481 int lod_striping_load(const struct lu_env *env, struct lod_object *lo)
1482 {
1483         struct lod_thread_info *info = lod_env_info(env);
1484         struct dt_object *next = dt_object_child(&lo->ldo_obj);
1485         struct lu_buf *buf = &info->lti_buf;
1486         int rc = 0;
1487
1488         ENTRY;
1489
1490         if (!dt_object_exists(next))
1491                 RETURN(0);
1492
1493         if (lod_striping_loaded(lo))
1494                 RETURN(0);
1495
1496         mutex_lock(&lo->ldo_layout_mutex);
1497         if (lod_striping_loaded(lo))
1498                 GOTO(unlock, rc = 0);
1499
1500         if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr)) {
1501                 rc = lod_get_lov_ea(env, lo);
1502                 if (rc <= 0)
1503                         GOTO(unlock, rc);
1504
1505                 /*
1506                  * there is LOV EA (striping information) in this object
1507                  * let's parse it and create in-core objects for the stripes
1508                  */
1509                 buf->lb_buf = info->lti_ea_store;
1510                 buf->lb_len = info->lti_ea_store_size;
1511                 rc = lod_parse_striping(env, lo, buf, 0);
1512                 if (rc == 0)
1513                         lo->ldo_comp_cached = 1;
1514         } else if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1515                 rc = lod_get_lmv_ea(env, lo);
1516                 if (rc > sizeof(struct lmv_foreign_md)) {
1517                         struct lmv_foreign_md *lfm = info->lti_ea_store;
1518
1519                         if (le32_to_cpu(lfm->lfm_magic) == LMV_MAGIC_FOREIGN) {
1520                                 lo->ldo_foreign_lmv = info->lti_ea_store;
1521                                 lo->ldo_foreign_lmv_size =
1522                                         info->lti_ea_store_size;
1523                                 info->lti_ea_store = NULL;
1524                                 info->lti_ea_store_size = 0;
1525
1526                                 lo->ldo_dir_stripe_loaded = 1;
1527                                 lo->ldo_is_foreign = 1;
1528                                 GOTO(unlock, rc = 0);
1529                         }
1530                 }
1531
1532                 if (rc < (int)sizeof(struct lmv_mds_md_v1)) {
1533                         /* Let's set stripe_loaded to avoid further
1534                          * stripe loading especially for non-stripe directory,
1535                          * which can hurt performance. (See LU-9840)
1536                          */
1537                         if (rc == 0)
1538                                 lo->ldo_dir_stripe_loaded = 1;
1539                         GOTO(unlock, rc = rc > 0 ? -EINVAL : rc);
1540                 }
1541                 buf->lb_buf = info->lti_ea_store;
1542                 buf->lb_len = info->lti_ea_store_size;
1543                 if (rc == sizeof(struct lmv_mds_md_v1)) {
1544                         rc = lod_load_lmv_shards(env, lo, buf, true);
1545                         if (buf->lb_buf != info->lti_ea_store) {
1546                                 OBD_FREE_LARGE(info->lti_ea_store,
1547                                                info->lti_ea_store_size);
1548                                 info->lti_ea_store = buf->lb_buf;
1549                                 info->lti_ea_store_size = buf->lb_len;
1550                         }
1551
1552                         if (rc < 0)
1553                                 GOTO(unlock, rc);
1554                 }
1555
1556                 /*
1557                  * there is LMV EA (striping information) in this object
1558                  * let's parse it and create in-core objects for the stripes
1559                  */
1560                 rc = lod_parse_dir_striping(env, lo, buf);
1561                 if (rc == 0)
1562                         lo->ldo_dir_stripe_loaded = 1;
1563         }
1564         EXIT;
1565 unlock:
1566         mutex_unlock(&lo->ldo_layout_mutex);
1567
1568         return rc;
1569 }
1570
1571 int lod_striping_reload(const struct lu_env *env, struct lod_object *lo,
1572                         const struct lu_buf *buf, enum layout_verify_flags lvf)
1573 {
1574         int rc;
1575
1576         ENTRY;
1577
1578         mutex_lock(&lo->ldo_layout_mutex);
1579         rc = lod_parse_striping(env, lo, buf, lvf);
1580         mutex_unlock(&lo->ldo_layout_mutex);
1581
1582         RETURN(rc);
1583 }
1584
1585 /**
1586  * Verify lov_user_md_v1/v3 striping.
1587  *
1588  * Check the validity of all fields including the magic, stripe size,
1589  * stripe count, stripe offset and that the pool is present.  Also check
1590  * that each target index points to an existing target. The additional
1591  * \a is_from_disk turns additional checks. In some cases zero fields
1592  * are allowed (like pattern=0).
1593  *
1594  * \param[in] d                 LOD device
1595  * \param[in] buf               buffer with LOV EA to verify
1596  * \param[in] is_from_disk      0 - from user, allow some fields to be 0
1597  *                              1 - from disk, do not allow
1598  *
1599  * \retval                      0 if the striping is valid
1600  * \retval                      -EINVAL if striping is invalid
1601  */
1602 static int lod_verify_v1v3(struct lod_device *d, const struct lu_buf *buf,
1603                            bool is_from_disk)
1604 {
1605         struct lov_user_md_v1   *lum;
1606         struct lov_user_md_v3   *lum3;
1607         struct pool_desc        *pool = NULL;
1608         __u32                    magic;
1609         __u32                    stripe_size;
1610         __u16                    stripe_count;
1611         __u16                    stripe_offset;
1612         size_t                   lum_size;
1613         int                      rc = 0;
1614         ENTRY;
1615
1616         lum = buf->lb_buf;
1617
1618         if (buf->lb_len < sizeof(*lum)) {
1619                 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
1620                        buf->lb_len);
1621                 GOTO(out, rc = -EINVAL);
1622         }
1623
1624         magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
1625         if (magic != LOV_USER_MAGIC_V1 &&
1626             magic != LOV_USER_MAGIC_V3 &&
1627             magic != LOV_USER_MAGIC_SPECIFIC) {
1628                 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
1629                        le32_to_cpu(lum->lmm_magic));
1630                 GOTO(out, rc = -EINVAL);
1631         }
1632
1633         /* the user uses "0" for default stripe pattern normally. */
1634         if (!is_from_disk && lum->lmm_pattern == LOV_PATTERN_NONE)
1635                 lum->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1636
1637         if (!lov_pattern_supported(le32_to_cpu(lum->lmm_pattern))) {
1638                 CDEBUG(D_LAYOUT, "bad userland stripe pattern: %#x\n",
1639                        le32_to_cpu(lum->lmm_pattern));
1640                 GOTO(out, rc = -EINVAL);
1641         }
1642
1643         /* a released lum comes from creating orphan on hsm release,
1644          * doesn't make sense to verify it. */
1645         if (le32_to_cpu(lum->lmm_pattern) & LOV_PATTERN_F_RELEASED)
1646                 GOTO(out, rc = 0);
1647
1648         /* 64kB is the largest common page size we see (ia64), and matches the
1649          * check in lfs */
1650         stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1651         if (stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
1652                 CDEBUG(D_LAYOUT, "stripe size %u not a multiple of %u\n",
1653                        stripe_size, LOV_MIN_STRIPE_SIZE);
1654                 GOTO(out, rc = -EINVAL);
1655         }
1656
1657         stripe_offset = le16_to_cpu(lum->lmm_stripe_offset);
1658         if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT &&
1659             !(lov_pattern(le32_to_cpu(lum->lmm_pattern)) & LOV_PATTERN_MDT)) {
1660                 /* if offset is not within valid range [0, osts_size) */
1661                 if (stripe_offset >= d->lod_ost_descs.ltd_tgts_size) {
1662                         CDEBUG(D_LAYOUT, "stripe offset %u >= bitmap size %u\n",
1663                                stripe_offset, d->lod_ost_descs.ltd_tgts_size);
1664                         GOTO(out, rc = -EINVAL);
1665                 }
1666
1667                 /* if lmm_stripe_offset is *not* in bitmap */
1668                 if (!test_bit(stripe_offset, d->lod_ost_bitmap)) {
1669                         CDEBUG(D_LAYOUT, "stripe offset %u not in bitmap\n",
1670                                stripe_offset);
1671                         GOTO(out, rc = -EINVAL);
1672                 }
1673         }
1674
1675         if (magic == LOV_USER_MAGIC_V1)
1676                 lum_size = offsetof(struct lov_user_md_v1,
1677                                     lmm_objects[0]);
1678         else if (magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC)
1679                 lum_size = offsetof(struct lov_user_md_v3,
1680                                     lmm_objects[0]);
1681         else
1682                 GOTO(out, rc = -EINVAL);
1683
1684         stripe_count = le16_to_cpu(lum->lmm_stripe_count);
1685         if (buf->lb_len < lum_size) {
1686                 CDEBUG(D_LAYOUT, "invalid buf len %zu/%zu for lov_user_md with "
1687                        "magic %#x and stripe_count %u\n",
1688                        buf->lb_len, lum_size, magic, stripe_count);
1689                 GOTO(out, rc = -EINVAL);
1690         }
1691
1692         if (!(magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC))
1693                 goto out;
1694
1695         lum3 = buf->lb_buf;
1696         /* In the function below, .hs_keycmp resolves to
1697          * pool_hashkey_keycmp() */
1698         pool = lod_find_pool(d, lum3->lmm_pool_name);
1699         if (pool == NULL)
1700                 goto out;
1701
1702         if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT) {
1703                 rc = lod_check_index_in_pool(stripe_offset, pool);
1704                 if (rc < 0)
1705                         GOTO(out, rc = -EINVAL);
1706         }
1707
1708         if (is_from_disk && stripe_count > pool_tgt_count(pool)) {
1709                 CDEBUG(D_LAYOUT, "stripe count %u > # OSTs %u in the pool\n",
1710                        stripe_count, pool_tgt_count(pool));
1711                 GOTO(out, rc = -EINVAL);
1712         }
1713
1714 out:
1715         if (pool != NULL)
1716                 lod_pool_putref(pool);
1717
1718         RETURN(rc);
1719 }
1720
1721 static inline
1722 struct lov_comp_md_entry_v1 *comp_entry_v1(struct lov_comp_md_v1 *comp, int i)
1723 {
1724         LASSERTF((le32_to_cpu(comp->lcm_magic) & ~LOV_MAGIC_DEFINED) ==
1725                  LOV_USER_MAGIC_COMP_V1 ||
1726                  (le32_to_cpu(comp->lcm_magic) & ~LOV_MAGIC_DEFINED) ==
1727                  LOV_USER_MAGIC_SEL, "Wrong magic %x\n",
1728                  le32_to_cpu(comp->lcm_magic));
1729         LASSERTF(i >= 0 && i < le16_to_cpu(comp->lcm_entry_count),
1730                  "bad index %d, max = %d\n",
1731                  i, le16_to_cpu(comp->lcm_entry_count));
1732
1733         return &comp->lcm_entries[i];
1734 }
1735
1736 #define for_each_comp_entry_v1(comp, entry) \
1737         for (entry = comp_entry_v1(comp, 0); \
1738              entry <= comp_entry_v1(comp, \
1739                                    le16_to_cpu(comp->lcm_entry_count) - 1); \
1740              entry++)
1741
1742 int lod_erase_dom_stripe(struct lov_comp_md_v1 *comp_v1,
1743                          struct lov_comp_md_entry_v1 *dom_ent)
1744 {
1745         struct lov_comp_md_entry_v1 *ent;
1746         __u16 entries;
1747         __u32 dom_off, dom_size, comp_size, off;
1748         void *src, *dst;
1749         unsigned int size, shift;
1750
1751         entries = le16_to_cpu(comp_v1->lcm_entry_count) - 1;
1752         LASSERT(entries > 0);
1753         comp_v1->lcm_entry_count = cpu_to_le16(entries);
1754
1755         comp_size = le32_to_cpu(comp_v1->lcm_size);
1756         dom_off = le32_to_cpu(dom_ent->lcme_offset);
1757         dom_size = le32_to_cpu(dom_ent->lcme_size);
1758
1759         /* all entries offsets are shifted by entry size at least */
1760         shift = sizeof(*dom_ent);
1761         for_each_comp_entry_v1(comp_v1, ent) {
1762                 off = le32_to_cpu(ent->lcme_offset);
1763                 if (off == dom_off) {
1764                         /* Entry deletion creates two holes in layout data:
1765                          * - hole in entries array
1766                          * - hole in layout data at dom_off with dom_size
1767                          *
1768                          * First memmove is one entry shift from next entry
1769                          * start with size up to dom_off in blob
1770                          */
1771                         dst = (void *)ent;
1772                         src = (void *)(ent + 1);
1773                         size = (unsigned long)((void *)comp_v1 + dom_off - src);
1774                         memmove(dst, src, size);
1775                         /* take 'off' from just moved entry */
1776                         off = le32_to_cpu(ent->lcme_offset);
1777                         /* second memmove is blob tail after 'off' up to
1778                          * component end
1779                          */
1780                         dst = (void *)comp_v1 + dom_off - sizeof(*ent);
1781                         src = (void *)comp_v1 + off;
1782                         size = (unsigned long)(comp_size - off);
1783                         memmove(dst, src, size);
1784                         /* all entries offsets after DoM entry are shifted by
1785                          * dom_size additionally
1786                          */
1787                         shift += dom_size;
1788                 }
1789                 ent->lcme_offset = cpu_to_le32(off - shift);
1790         }
1791         comp_v1->lcm_size = cpu_to_le32(comp_size - shift);
1792
1793         /* notify a caller to re-check entry */
1794         return -ERESTART;
1795 }
1796
1797 void lod_dom_stripesize_recalc(struct lod_device *d)
1798 {
1799         __u64 threshold_mb = d->lod_dom_threshold_free_mb;
1800         __u32 max_size = d->lod_dom_stripesize_max_kb;
1801         __u32 def_size = d->lod_dom_stripesize_cur_kb;
1802
1803         /* use maximum allowed value if free space is above threshold */
1804         if (d->lod_lsfs_free_mb >= threshold_mb) {
1805                 def_size = max_size;
1806         } else if (!d->lod_lsfs_free_mb || max_size <= LOD_DOM_MIN_SIZE_KB) {
1807                 def_size = 0;
1808         } else {
1809                 /* recalc threshold like it would be with def_size as max */
1810                 threshold_mb = mult_frac(threshold_mb, def_size, max_size);
1811                 if (d->lod_lsfs_free_mb < threshold_mb)
1812                         def_size = rounddown(def_size / 2, LOD_DOM_MIN_SIZE_KB);
1813                 else if (d->lod_lsfs_free_mb > threshold_mb * 2)
1814                         def_size = max_t(unsigned int, def_size * 2,
1815                                          LOD_DOM_MIN_SIZE_KB);
1816         }
1817
1818         if (d->lod_dom_stripesize_cur_kb != def_size) {
1819                 CDEBUG(D_LAYOUT, "Change default DOM stripe size %d->%d\n",
1820                        d->lod_dom_stripesize_cur_kb, def_size);
1821                 d->lod_dom_stripesize_cur_kb = def_size;
1822         }
1823 }
1824
1825 static __u32 lod_dom_stripesize_limit(const struct lu_env *env,
1826                                       struct lod_device *d)
1827 {
1828         int rc;
1829
1830         /* set bfree as fraction of total space */
1831         if (CFS_FAIL_CHECK(OBD_FAIL_MDS_STATFS_SPOOF)) {
1832                 spin_lock(&d->lod_lsfs_lock);
1833                 d->lod_lsfs_free_mb = mult_frac(d->lod_lsfs_total_mb,
1834                                         min_t(int, cfs_fail_val, 100), 100);
1835                 GOTO(recalc, rc = 0);
1836         }
1837
1838         if (d->lod_lsfs_age < ktime_get_seconds() - LOD_DOM_SFS_MAX_AGE) {
1839                 struct obd_statfs sfs;
1840
1841                 spin_lock(&d->lod_lsfs_lock);
1842                 if (d->lod_lsfs_age > ktime_get_seconds() - LOD_DOM_SFS_MAX_AGE)
1843                         GOTO(unlock, rc = 0);
1844
1845                 d->lod_lsfs_age = ktime_get_seconds();
1846                 spin_unlock(&d->lod_lsfs_lock);
1847                 rc = dt_statfs(env, d->lod_child, &sfs);
1848                 if (rc) {
1849                         CDEBUG(D_LAYOUT,
1850                                "%s: failed to get OSD statfs: rc = %d\n",
1851                                lod2obd(d)->obd_name, rc);
1852                         GOTO(out, rc);
1853                 }
1854                 /* udpate local OSD cached statfs data */
1855                 spin_lock(&d->lod_lsfs_lock);
1856                 d->lod_lsfs_total_mb = (sfs.os_blocks * sfs.os_bsize) >> 20;
1857                 d->lod_lsfs_free_mb = (sfs.os_bfree * sfs.os_bsize) >> 20;
1858 recalc:
1859                 lod_dom_stripesize_recalc(d);
1860 unlock:
1861                 spin_unlock(&d->lod_lsfs_lock);
1862         }
1863 out:
1864         return d->lod_dom_stripesize_cur_kb << 10;
1865 }
1866
1867 int lod_dom_stripesize_choose(const struct lu_env *env, struct lod_device *d,
1868                               struct lov_comp_md_v1 *comp_v1,
1869                               struct lov_comp_md_entry_v1 *dom_ent,
1870                               __u32 stripe_size)
1871 {
1872         struct lov_comp_md_entry_v1 *ent;
1873         struct lu_extent *dom_ext, *ext;
1874         struct lov_user_md_v1 *lum;
1875         __u32 max_stripe_size;
1876         __u16 mid, dom_mid;
1877         int rc = 0;
1878         bool dom_next_entry = false;
1879
1880         dom_ext = &dom_ent->lcme_extent;
1881         dom_mid = mirror_id_of(le32_to_cpu(dom_ent->lcme_id));
1882         max_stripe_size = lod_dom_stripesize_limit(env, d);
1883
1884         /* Check stripe size againts current per-MDT limit */
1885         if (stripe_size <= max_stripe_size)
1886                 return 0;
1887
1888         lum = (void *)comp_v1 + le32_to_cpu(dom_ent->lcme_offset);
1889         CDEBUG(D_LAYOUT, "overwrite DoM component size %u with MDT limit %u\n",
1890                stripe_size, max_stripe_size);
1891         lum->lmm_stripe_size = cpu_to_le32(max_stripe_size);
1892
1893         /* In common case the DoM stripe is first entry in a mirror and
1894          * can be deleted only if it is not single entry in layout or
1895          * mirror, otherwise error should be returned.
1896          */
1897         for_each_comp_entry_v1(comp_v1, ent) {
1898                 if (ent == dom_ent)
1899                         continue;
1900
1901                 mid = mirror_id_of(le32_to_cpu(ent->lcme_id));
1902                 if (mid != dom_mid)
1903                         continue;
1904
1905                 ext = &ent->lcme_extent;
1906                 if (ext->e_start != dom_ext->e_end)
1907                         continue;
1908
1909                 /* Found next component after the DoM one with the same
1910                  * mirror_id and adjust its start with DoM component end.
1911                  *
1912                  * NOTE: we are considering here that there can be only one
1913                  * DoM component in a file, all replicas are located on OSTs
1914                  * always and don't need adjustment since use own layouts.
1915                  */
1916                 ext->e_start = cpu_to_le64(max_stripe_size);
1917                 dom_next_entry = true;
1918                 break;
1919         }
1920
1921         if (max_stripe_size == 0) {
1922                 /* DoM component size is zero due to server setting, remove
1923                  * it from the layout but only if next component exists in
1924                  * the same mirror. That must be checked prior calling the
1925                  * lod_erase_dom_stripe().
1926                  */
1927                 if (!dom_next_entry)
1928                         return -EFBIG;
1929
1930                 rc = lod_erase_dom_stripe(comp_v1, dom_ent);
1931         } else {
1932                 /* Update DoM extent end finally */
1933                 dom_ext->e_end = cpu_to_le64(max_stripe_size);
1934         }
1935
1936         return rc;
1937 }
1938
1939 /**
1940  * Verify LOV striping.
1941  *
1942  * \param[in] d                 LOD device
1943  * \param[in] buf               buffer with LOV EA to verify
1944  * \param[in] is_from_disk      0 - from user, allow some fields to be 0
1945  *                              1 - from disk, do not allow
1946  * \param[in] start             extent start for composite layout
1947  *
1948  * \retval                      0 if the striping is valid
1949  * \retval                      -EINVAL if striping is invalid
1950  */
1951 int lod_verify_striping(const struct lu_env *env, struct lod_device *d,
1952                         struct lod_object *lo, const struct lu_buf *buf,
1953                         bool is_from_disk)
1954 {
1955         struct lov_user_md_v1   *lum;
1956         struct lov_comp_md_v1   *comp_v1;
1957         struct lov_comp_md_entry_v1     *ent;
1958         struct lu_extent        *ext;
1959         struct lu_buf   tmp;
1960         __u64   prev_end = 0;
1961         __u32   stripe_size = 0;
1962         __u16   prev_mid = -1, mirror_id = -1;
1963         __u32   mirror_count;
1964         __u32   magic;
1965         int     rc = 0;
1966         ENTRY;
1967
1968         if (buf->lb_len < sizeof(lum->lmm_magic)) {
1969                 CDEBUG(D_LAYOUT, "invalid buf len %zu\n", buf->lb_len);
1970                 RETURN(-EINVAL);
1971         }
1972
1973         lum = buf->lb_buf;
1974
1975         magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
1976         /* treat foreign LOV EA/object case first
1977          * XXX is it expected to try setting again a foreign?
1978          * XXX should we care about different current vs new layouts ?
1979          */
1980         if (unlikely(magic == LOV_USER_MAGIC_FOREIGN)) {
1981                 struct lov_foreign_md *lfm = buf->lb_buf;
1982
1983                 if (buf->lb_len < offsetof(typeof(*lfm), lfm_value)) {
1984                         CDEBUG(D_LAYOUT,
1985                                "buf len %zu < min lov_foreign_md size (%zu)\n",
1986                                buf->lb_len, offsetof(typeof(*lfm),
1987                                lfm_value));
1988                         RETURN(-EINVAL);
1989                 }
1990
1991                 if (foreign_size_le(lfm) > buf->lb_len) {
1992                         CDEBUG(D_LAYOUT,
1993                                "buf len %zu < this lov_foreign_md size (%zu)\n",
1994                                buf->lb_len, foreign_size_le(lfm));
1995                         RETURN(-EINVAL);
1996                 }
1997                 /* Don't do anything with foreign layouts */
1998                 RETURN(0);
1999         }
2000
2001         /* normal LOV/layout cases */
2002
2003         if (buf->lb_len < sizeof(*lum)) {
2004                 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
2005                        buf->lb_len);
2006                 RETURN(-EINVAL);
2007         }
2008
2009         switch (magic) {
2010         case LOV_USER_MAGIC_FOREIGN:
2011                 RETURN(0);
2012         case LOV_USER_MAGIC_V1:
2013         case LOV_USER_MAGIC_V3:
2014         case LOV_USER_MAGIC_SPECIFIC:
2015                 if (lov_pattern(le32_to_cpu(lum->lmm_pattern)) &
2016                     LOV_PATTERN_MDT) {
2017                         /* DoM must use composite layout */
2018                         CDEBUG(D_LAYOUT, "DoM without composite layout\n");
2019                         RETURN(-EINVAL);
2020                 }
2021                 RETURN(lod_verify_v1v3(d, buf, is_from_disk));
2022         case LOV_USER_MAGIC_COMP_V1:
2023         case LOV_USER_MAGIC_SEL:
2024                 break;
2025         default:
2026                 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
2027                        le32_to_cpu(lum->lmm_magic));
2028                 RETURN(-EINVAL);
2029         }
2030
2031         /* magic == LOV_USER_MAGIC_COMP_V1 */
2032         comp_v1 = buf->lb_buf;
2033         if (buf->lb_len < le32_to_cpu(comp_v1->lcm_size)) {
2034                 CDEBUG(D_LAYOUT, "buf len %zu is less than %u\n",
2035                        buf->lb_len, le32_to_cpu(comp_v1->lcm_size));
2036                 RETURN(-EINVAL);
2037         }
2038
2039 recheck:
2040         mirror_count = 0;
2041         if (le16_to_cpu(comp_v1->lcm_entry_count) == 0) {
2042                 CDEBUG(D_LAYOUT, "entry count is zero\n");
2043                 RETURN(-EINVAL);
2044         }
2045
2046         if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr) &&
2047             lo->ldo_comp_cnt > 0) {
2048                 /* could be called from lustre.lov.add */
2049                 __u32 cnt = lo->ldo_comp_cnt;
2050
2051                 ext = &lo->ldo_comp_entries[cnt - 1].llc_extent;
2052                 prev_end = ext->e_end;
2053
2054                 ++mirror_count;
2055         }
2056
2057         for_each_comp_entry_v1(comp_v1, ent) {
2058                 ext = &ent->lcme_extent;
2059
2060                 if (le64_to_cpu(ext->e_start) > le64_to_cpu(ext->e_end) ||
2061                     le64_to_cpu(ext->e_start) & (LOV_MIN_STRIPE_SIZE - 1) ||
2062                     (le64_to_cpu(ext->e_end) != LUSTRE_EOF &&
2063                     le64_to_cpu(ext->e_end) & (LOV_MIN_STRIPE_SIZE - 1))) {
2064                         CDEBUG(D_LAYOUT, "invalid extent "DEXT"\n",
2065                                le64_to_cpu(ext->e_start),
2066                                le64_to_cpu(ext->e_end));
2067                         RETURN(-EINVAL);
2068                 }
2069
2070                 if (is_from_disk) {
2071                         /* lcme_id contains valid value */
2072                         if (le32_to_cpu(ent->lcme_id) == 0 ||
2073                             le32_to_cpu(ent->lcme_id) > LCME_ID_MAX) {
2074                                 CDEBUG(D_LAYOUT, "invalid id %u\n",
2075                                        le32_to_cpu(ent->lcme_id));
2076                                 RETURN(-EINVAL);
2077                         }
2078
2079                         if (le16_to_cpu(comp_v1->lcm_mirror_count) > 0) {
2080                                 mirror_id = mirror_id_of(
2081                                                 le32_to_cpu(ent->lcme_id));
2082
2083                                 /* first component must start with 0 */
2084                                 if (mirror_id != prev_mid &&
2085                                     le64_to_cpu(ext->e_start) != 0) {
2086                                         CDEBUG(D_LAYOUT,
2087                                                "invalid start:%llu, expect:0\n",
2088                                                le64_to_cpu(ext->e_start));
2089                                         RETURN(-EINVAL);
2090                                 }
2091
2092                                 prev_mid = mirror_id;
2093                         }
2094                 }
2095
2096                 if (le64_to_cpu(ext->e_start) == 0) {
2097                         ++mirror_count;
2098                         prev_end = 0;
2099                 }
2100
2101                 /* the next must be adjacent with the previous one */
2102                 if (le64_to_cpu(ext->e_start) != prev_end) {
2103                         CDEBUG(D_LAYOUT,
2104                                "invalid start actual:%llu, expect:%llu\n",
2105                                le64_to_cpu(ext->e_start), prev_end);
2106                         RETURN(-EINVAL);
2107                 }
2108
2109                 tmp.lb_buf = (char *)comp_v1 + le32_to_cpu(ent->lcme_offset);
2110                 tmp.lb_len = le32_to_cpu(ent->lcme_size);
2111
2112                 /* Check DoM entry is always the first one */
2113                 lum = tmp.lb_buf;
2114                 if (lov_pattern(le32_to_cpu(lum->lmm_pattern)) &
2115                     LOV_PATTERN_MDT) {
2116                         /* DoM component must be the first in a mirror */
2117                         if (le64_to_cpu(ext->e_start) > 0) {
2118                                 CDEBUG(D_LAYOUT, "invalid DoM component "
2119                                        "with %llu extent start\n",
2120                                        le64_to_cpu(ext->e_start));
2121                                 RETURN(-EINVAL);
2122                         }
2123                         stripe_size = le32_to_cpu(lum->lmm_stripe_size);
2124                         /* There is just one stripe on MDT and it must
2125                          * cover whole component size. */
2126                         if (stripe_size != le64_to_cpu(ext->e_end)) {
2127                                 CDEBUG(D_LAYOUT, "invalid DoM layout "
2128                                        "stripe size %u != %llu "
2129                                        "(component size)\n",
2130                                        stripe_size, prev_end);
2131                                 RETURN(-EINVAL);
2132                         }
2133                         /* Check and adjust stripe size by per-MDT limit */
2134                         rc = lod_dom_stripesize_choose(env, d, comp_v1, ent,
2135                                                        stripe_size);
2136                         /* DoM entry was removed, re-check layout from start */
2137                         if (rc == -ERESTART)
2138                                 goto recheck;
2139                         else if (rc)
2140                                 RETURN(rc);
2141
2142                         if (le16_to_cpu(lum->lmm_stripe_count) == 1)
2143                                 lum->lmm_stripe_count = 0;
2144                         /* Any stripe count is forbidden on DoM component */
2145                         if (lum->lmm_stripe_count > 0) {
2146                                 CDEBUG(D_LAYOUT,
2147                                        "invalid DoM layout stripe count %u, must be 0\n",
2148                                        le16_to_cpu(lum->lmm_stripe_count));
2149                                 RETURN(-EINVAL);
2150                         }
2151
2152                         /* Any pool is forbidden on DoM component */
2153                         if (lum->lmm_magic == LOV_USER_MAGIC_V3) {
2154                                 struct lov_user_md_v3 *v3 = (void *)lum;
2155
2156                                 if (v3->lmm_pool_name[0] != '\0') {
2157                                         CDEBUG(D_LAYOUT,
2158                                                "DoM component cannot have pool assigned\n");
2159                                         RETURN(-EINVAL);
2160                                 }
2161                         }
2162                 }
2163
2164                 prev_end = le64_to_cpu(ext->e_end);
2165
2166                 rc = lod_verify_v1v3(d, &tmp, is_from_disk);
2167                 if (rc)
2168                         RETURN(rc);
2169
2170                 if (prev_end == LUSTRE_EOF || ext->e_start == prev_end)
2171                         continue;
2172
2173                 /* extent end must be aligned with the stripe_size */
2174                 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
2175                 if (stripe_size && prev_end % stripe_size) {
2176                         CDEBUG(D_LAYOUT, "stripe size isn't aligned, "
2177                                "stripe_sz: %u, [%llu, %llu)\n",
2178                                stripe_size, ext->e_start, prev_end);
2179                         RETURN(-EINVAL);
2180                 }
2181         }
2182
2183         /* make sure that the mirror_count is telling the truth */
2184         if (mirror_count != le16_to_cpu(comp_v1->lcm_mirror_count) + 1)
2185                 RETURN(-EINVAL);
2186
2187         RETURN(0);
2188 }
2189
2190 /**
2191  * set the default stripe size, if unset.
2192  *
2193  * \param[in,out] val   number of bytes per OST stripe
2194  *
2195  * The minimum stripe size is 64KB to ensure that a single stripe is an
2196  * even multiple of a client PAGE_SIZE (IA64, PPC, etc).  Otherwise, it
2197  * is difficult to split dirty pages across OSCs during writes.
2198  */
2199 void lod_fix_desc_stripe_size(__u64 *val)
2200 {
2201         if (*val < LOV_MIN_STRIPE_SIZE) {
2202                 if (*val != 0)
2203                         LCONSOLE_INFO("Increasing default stripe size to "
2204                                       "minimum value %u\n",
2205                                       LOV_DESC_STRIPE_SIZE_DEFAULT);
2206                 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
2207         } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
2208                 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
2209                 LCONSOLE_WARN("Changing default stripe size to %llu (a "
2210                               "multiple of %u)\n",
2211                               *val, LOV_MIN_STRIPE_SIZE);
2212         }
2213 }
2214
2215 /**
2216  * set the filesystem default number of stripes, if unset.
2217  *
2218  * \param[in,out] val   number of stripes
2219  *
2220  * A value of "0" means "use the system-wide default stripe count", which
2221  * has either been inherited by now, or falls back to 1 stripe per file.
2222  * A value of "-1" (0xffffffff) means "stripe over all available OSTs",
2223  * and is a valid value, so is left unchanged here.
2224  */
2225 void lod_fix_desc_stripe_count(__u32 *val)
2226 {
2227         if (*val == 0)
2228                 *val = 1;
2229 }
2230
2231 /**
2232  * set the filesystem default layout pattern
2233  *
2234  * \param[in,out] val   LOV_PATTERN_* layout
2235  *
2236  * A value of "0" means "use the system-wide default layout type", which
2237  * has either been inherited by now, or falls back to plain RAID0 striping.
2238  */
2239 void lod_fix_desc_pattern(__u32 *val)
2240 {
2241         /* from lov_setstripe */
2242         if ((*val != 0) && !lov_pattern_supported_normal_comp(*val)) {
2243                 LCONSOLE_WARN("lod: Unknown stripe pattern: %#x\n", *val);
2244                 *val = 0;
2245         }
2246 }
2247
2248 void lod_fix_lmv_desc_pattern(__u32 *val)
2249 {
2250         if ((*val) && !lmv_is_known_hash_type(*val)) {
2251                 LCONSOLE_WARN("lod: Unknown md stripe pattern: %#x\n", *val);
2252                 *val = 0;
2253         }
2254 }
2255
2256 void lod_fix_desc_qos_maxage(__u32 *val)
2257 {
2258         /* fix qos_maxage */
2259         if (*val == 0)
2260                 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
2261 }
2262
2263 /**
2264  * Used to fix insane default striping.
2265  *
2266  * \param[in] desc      striping description
2267  */
2268 void lod_fix_desc(struct lov_desc *desc)
2269 {
2270         lod_fix_desc_stripe_size(&desc->ld_default_stripe_size);
2271         lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
2272         lod_fix_desc_pattern(&desc->ld_pattern);
2273         lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
2274 }
2275
2276 static void lod_fix_lmv_desc(struct lmv_desc *desc)
2277 {
2278         desc->ld_active_tgt_count = 0;
2279         lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
2280         lod_fix_lmv_desc_pattern(&desc->ld_pattern);
2281         lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
2282 }
2283
2284 /**
2285  * Initialize the structures used to store pools and default striping.
2286  *
2287  * \param[in] lod       LOD device
2288  * \param[in] lcfg      configuration structure storing default striping.
2289  *
2290  * \retval              0 if initialization succeeds
2291  * \retval              negative error number on failure
2292  */
2293 int lod_pools_init(struct lod_device *lod, struct lustre_cfg *lcfg)
2294 {
2295         struct obd_device          *obd;
2296         struct lov_desc            *desc;
2297         int                         rc;
2298         ENTRY;
2299
2300         obd = class_name2obd(lustre_cfg_string(lcfg, 0));
2301         LASSERT(obd != NULL);
2302         obd->obd_lu_dev = &lod->lod_dt_dev.dd_lu_dev;
2303
2304         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
2305                 CERROR("LOD setup requires a descriptor\n");
2306                 RETURN(-EINVAL);
2307         }
2308
2309         desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
2310
2311         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
2312                 CERROR("descriptor size wrong: %d > %d\n",
2313                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
2314                 RETURN(-EINVAL);
2315         }
2316
2317         if (desc->ld_magic != LOV_DESC_MAGIC) {
2318                 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
2319                         CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
2320                                obd->obd_name, desc);
2321                         lustre_swab_lov_desc(desc);
2322                 } else {
2323                         CERROR("%s: Bad lov desc magic: %#x\n",
2324                                obd->obd_name, desc->ld_magic);
2325                         RETURN(-EINVAL);
2326                 }
2327         }
2328
2329         lod_fix_desc(desc);
2330
2331         desc->ld_active_tgt_count = 0;
2332         lod->lod_ost_descs.ltd_lov_desc = *desc;
2333
2334         /* NB: config doesn't contain lmv_desc, alter it via sysfs. */
2335         lod_fix_lmv_desc(&lod->lod_mdt_descs.ltd_lmv_desc);
2336
2337         lod->lod_sp_me = LUSTRE_SP_CLI;
2338
2339         /* Set up OST pool environment */
2340         lod->lod_pool_count = 0;
2341         rc = lod_pool_hash_init(&lod->lod_pools_hash_body);
2342         if (rc)
2343                 RETURN(-ENOMEM);
2344
2345         INIT_LIST_HEAD(&lod->lod_pool_list);
2346         lod->lod_pool_count = 0;
2347         rc = lu_tgt_pool_init(&lod->lod_mdt_descs.ltd_tgt_pool, 0);
2348         if (rc)
2349                 GOTO(out_hash, rc);
2350
2351         rc = lu_tgt_pool_init(&lod->lod_mdt_descs.ltd_qos.lq_rr.lqr_pool, 0);
2352         if (rc)
2353                 GOTO(out_mdt_pool, rc);
2354
2355         rc = lu_tgt_pool_init(&lod->lod_ost_descs.ltd_tgt_pool, 0);
2356         if (rc)
2357                 GOTO(out_mdt_rr_pool, rc);
2358
2359         rc = lu_tgt_pool_init(&lod->lod_ost_descs.ltd_qos.lq_rr.lqr_pool, 0);
2360         if (rc)
2361                 GOTO(out_ost_pool, rc);
2362
2363         RETURN(0);
2364
2365 out_ost_pool:
2366         lu_tgt_pool_free(&lod->lod_ost_descs.ltd_tgt_pool);
2367 out_mdt_rr_pool:
2368         lu_tgt_pool_free(&lod->lod_mdt_descs.ltd_qos.lq_rr.lqr_pool);
2369 out_mdt_pool:
2370         lu_tgt_pool_free(&lod->lod_mdt_descs.ltd_tgt_pool);
2371 out_hash:
2372         lod_pool_hash_destroy(&lod->lod_pools_hash_body);
2373
2374         return rc;
2375 }
2376
2377 /**
2378  * Release the structures describing the pools.
2379  *
2380  * \param[in] lod       LOD device from which we release the structures
2381  *
2382  * \retval              0 always
2383  */
2384 int lod_pools_fini(struct lod_device *lod)
2385 {
2386         struct obd_device   *obd = lod2obd(lod);
2387         struct pool_desc    *pool, *tmp;
2388         ENTRY;
2389
2390         list_for_each_entry_safe(pool, tmp, &lod->lod_pool_list, pool_list) {
2391                 /* free pool structs */
2392                 CDEBUG(D_INFO, "delete pool %p\n", pool);
2393                 /* In the function below, .hs_keycmp resolves to
2394                  * pool_hashkey_keycmp() */
2395                 lod_pool_del(obd, pool->pool_name);
2396         }
2397
2398         lod_pool_hash_destroy(&lod->lod_pools_hash_body);
2399         lu_tgt_pool_free(&lod->lod_ost_descs.ltd_qos.lq_rr.lqr_pool);
2400         lu_tgt_pool_free(&lod->lod_ost_descs.ltd_tgt_pool);
2401         lu_tgt_pool_free(&lod->lod_mdt_descs.ltd_qos.lq_rr.lqr_pool);
2402         lu_tgt_pool_free(&lod->lod_mdt_descs.ltd_tgt_pool);
2403
2404         RETURN(0);
2405 }