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