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, 2016, 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                 struct list_head kill;
  80                 unsigned int idx;
  81
  82                 CDEBUG(D_CONFIG, "destroying %d ltd desc\n",
  83                        ltd->ltd_death_row);
  84
  85                 INIT_LIST_HEAD(&kill);
  86
  87                 cfs_foreach_bit(ltd->ltd_tgt_bitmap, idx) {
  88                         tgt_desc = LTD_TGT(ltd, idx);
  89                         LASSERT(tgt_desc);
  90
  91                         if (!tgt_desc->ltd_reap)
  92                                 continue;
  93
  94                         list_add(&tgt_desc->ltd_kill, &kill);
  95                         LTD_TGT(ltd, idx) = NULL;
  96                         /*FIXME: only support ost pool for now */
  97                         if (ltd == &lod->lod_ost_descs) {
  98                                 lod_ost_pool_remove(&lod->lod_pool_info, idx);
  99                                 if (tgt_desc->ltd_active)
 100                                         lod->lod_desc.ld_active_tgt_count--;
 101                         }
 102                         ltd->ltd_tgtnr--;
 103                         cfs_bitmap_clear(ltd->ltd_tgt_bitmap, idx);
 104                         ltd->ltd_death_row--;
 105                 }
 106                 mutex_unlock(&ltd->ltd_mutex);
 107                 up_read(&ltd->ltd_rw_sem);
 108
 109                 list_for_each_entry_safe(tgt_desc, tmp, &kill, ltd_kill) {
 110                         int rc;
 111                         list_del(&tgt_desc->ltd_kill);
 112                         if (ltd == &lod->lod_ost_descs) {
 113                                 /* remove from QoS structures */
 114                                 rc = qos_del_tgt(lod, tgt_desc);
 115                                 if (rc)
 116                                         CERROR("%s: qos_del_tgt(%s) failed:"
 117                                                "rc = %d\n",
 118                                                lod2obd(lod)->obd_name,
 119                                               obd_uuid2str(&tgt_desc->ltd_uuid),
 120                                                rc);
 121                         }
 122                         rc = obd_disconnect(tgt_desc->ltd_exp);
 123                         if (rc)
 124                                 CERROR("%s: failed to disconnect %s: rc = %d\n",
 125                                        lod2obd(lod)->obd_name,
 126                                        obd_uuid2str(&tgt_desc->ltd_uuid), rc);
 127                         OBD_FREE_PTR(tgt_desc);
 128                 }
 129         } else {
 130                 mutex_unlock(&ltd->ltd_mutex);
 131                 up_read(&ltd->ltd_rw_sem);
 132         }
 133 }
 134
 135 /**
 136  * Expand size of target table.
 137  *
 138  * When the target table is full, we have to extend the table. To do so,
 139  * we allocate new memory with some reserve, move data from the old table
 140  * to the new one and release memory consumed by the old table.
 141  * Notice we take ltd_rw_sem exclusively to ensure atomic switch.
 142  *
 143  * \param[in] ltd               target table
 144  * \param[in] newsize           new size of the table
 145  *
 146  * \retval                      0 on success
 147  * \retval                      -ENOMEM if reallocation failed
 148  */
 149 static int ltd_bitmap_resize(struct lod_tgt_descs *ltd, __u32 newsize)
 150 {
 151         struct cfs_bitmap *new_bitmap, *old_bitmap = NULL;
 152         int           rc = 0;
 153         ENTRY;
 154
 155         /* grab write reference on the lod. Relocating the array requires
 156          * exclusive access */
 157
 158         down_write(&ltd->ltd_rw_sem);
 159         if (newsize <= ltd->ltd_tgts_size)
 160                 /* someone else has already resize the array */
 161                 GOTO(out, rc = 0);
 162
 163         /* allocate new bitmap */
 164         new_bitmap = CFS_ALLOCATE_BITMAP(newsize);
 165         if (!new_bitmap)
 166                 GOTO(out, rc = -ENOMEM);
 167
 168         if (ltd->ltd_tgts_size > 0) {
 169                 /* the bitmap already exists, we need
 170                  * to copy data from old one */
 171                 cfs_bitmap_copy(new_bitmap, ltd->ltd_tgt_bitmap);
 172                 old_bitmap = ltd->ltd_tgt_bitmap;
 173         }
 174
 175         ltd->ltd_tgts_size  = newsize;
 176         ltd->ltd_tgt_bitmap = new_bitmap;
 177
 178         if (old_bitmap)
 179                 CFS_FREE_BITMAP(old_bitmap);
 180
 181         CDEBUG(D_CONFIG, "tgt size: %d\n", ltd->ltd_tgts_size);
 182
 183         EXIT;
 184 out:
 185         up_write(&ltd->ltd_rw_sem);
 186         return rc;
 187 }
 188
 189 /**
 190  * Connect LOD to a new OSP and add it to the target table.
 191  *
 192  * Connect to the OSP device passed, initialize all the internal
 193  * structures related to the device and add it to the target table.
 194  *
 195  * \param[in] env               execution environment for this thread
 196  * \param[in] lod               LOD device to be connected to the new OSP
 197  * \param[in] osp               name of OSP device name to be added
 198  * \param[in] index             index of the new target
 199  * \param[in] gen               target's generation number
 200  * \param[in] tgt_index         OSP's group
 201  * \param[in] type              type of device (mdc or osc)
 202  * \param[in] active            state of OSP: 0 - inactive, 1 - active
 203  *
 204  * \retval                      0 if added successfully
 205  * \retval                      negative error number on failure
 206  */
 207 int lod_add_device(const struct lu_env *env, struct lod_device *lod,
 208                    char *osp, unsigned index, unsigned gen, int tgt_index,
 209                    char *type, int active)
 210 {
 211         struct obd_connect_data *data = NULL;
 212         struct obd_export       *exp = NULL;
 213         struct obd_device       *obd;
 214         struct lu_device        *lu_dev;
 215         struct dt_device        *dt_dev;
 216         int                      rc;
 217         struct lod_tgt_desc     *tgt_desc;
 218         struct lod_tgt_descs    *ltd;
 219         struct lustre_cfg       *lcfg;
 220         struct obd_uuid         obd_uuid;
 221         bool                    for_ost;
 222         bool lock = false;
 223         ENTRY;
 224
 225         CDEBUG(D_CONFIG, "osp:%s idx:%d gen:%d\n", osp, index, gen);
 226
 227         if (gen <= 0) {
 228                 CERROR("request to add OBD %s with invalid generation: %d\n",
 229                        osp, gen);
 230                 RETURN(-EINVAL);
 231         }
 232
 233         obd_str2uuid(&obd_uuid, osp);
 234
 235         obd = class_find_client_obd(&obd_uuid, LUSTRE_OSP_NAME,
 236                                 &lod->lod_dt_dev.dd_lu_dev.ld_obd->obd_uuid);
 237         if (obd == NULL) {
 238                 CERROR("can't find %s device\n", osp);
 239                 RETURN(-EINVAL);
 240         }
 241
 242         LASSERT(obd->obd_lu_dev != NULL);
 243         LASSERT(obd->obd_lu_dev->ld_site == lod->lod_dt_dev.dd_lu_dev.ld_site);
 244
 245         lu_dev = obd->obd_lu_dev;
 246         dt_dev = lu2dt_dev(lu_dev);
 247
 248         OBD_ALLOC_PTR(data);
 249         if (data == NULL)
 250                 GOTO(out_cleanup, rc = -ENOMEM);
 251
 252         data->ocd_connect_flags = OBD_CONNECT_INDEX | OBD_CONNECT_VERSION;
 253         data->ocd_version = LUSTRE_VERSION_CODE;
 254         data->ocd_index = index;
 255
 256         if (strcmp(LUSTRE_OSC_NAME, type) == 0) {
 257                 for_ost = true;
 258                 data->ocd_connect_flags |= OBD_CONNECT_AT |
 259                                            OBD_CONNECT_FULL20 |
 260                                            OBD_CONNECT_INDEX |
 261 #ifdef HAVE_LRU_RESIZE_SUPPORT
 262                                            OBD_CONNECT_LRU_RESIZE |
 263 #endif
 264                                            OBD_CONNECT_MDS |
 265                                            OBD_CONNECT_REQPORTAL |
 266                                            OBD_CONNECT_SKIP_ORPHAN |
 267                                            OBD_CONNECT_FID |
 268                                            OBD_CONNECT_LVB_TYPE |
 269                                            OBD_CONNECT_VERSION |
 270                                            OBD_CONNECT_PINGLESS |
 271                                            OBD_CONNECT_LFSCK |
 272                                            OBD_CONNECT_BULK_MBITS;
 273
 274                 data->ocd_group = tgt_index;
 275                 ltd = &lod->lod_ost_descs;
 276         } else {
 277                 struct obd_import *imp = obd->u.cli.cl_import;
 278
 279                 for_ost = false;
 280                 data->ocd_ibits_known = MDS_INODELOCK_UPDATE;
 281                 data->ocd_connect_flags |= OBD_CONNECT_ACL |
 282                                            OBD_CONNECT_IBITS |
 283                                            OBD_CONNECT_MDS_MDS |
 284                                            OBD_CONNECT_FID |
 285                                            OBD_CONNECT_AT |
 286                                            OBD_CONNECT_FULL20 |
 287                                            OBD_CONNECT_LFSCK |
 288                                            OBD_CONNECT_BULK_MBITS;
 289                 spin_lock(&imp->imp_lock);
 290                 imp->imp_server_timeout = 1;
 291                 spin_unlock(&imp->imp_lock);
 292                 imp->imp_client->cli_request_portal = OUT_PORTAL;
 293                 CDEBUG(D_OTHER, "%s: Set 'mds' portal and timeout\n",
 294                       obd->obd_name);
 295                 ltd = &lod->lod_mdt_descs;
 296         }
 297
 298         rc = obd_connect(env, &exp, obd, &obd->obd_uuid, data, NULL);
 299         OBD_FREE_PTR(data);
 300         if (rc) {
 301                 CERROR("%s: cannot connect to next dev %s (%d)\n",
 302                        obd->obd_name, osp, rc);
 303                 GOTO(out_cleanup, rc);
 304         }
 305
 306         /* Allocate ost descriptor and fill it */
 307         OBD_ALLOC_PTR(tgt_desc);
 308         if (!tgt_desc)
 309                 GOTO(out_conn, rc = -ENOMEM);
 310
 311         tgt_desc->ltd_tgt    = dt_dev;
 312         tgt_desc->ltd_exp    = exp;
 313         tgt_desc->ltd_uuid   = obd->u.cli.cl_target_uuid;
 314         tgt_desc->ltd_gen    = gen;
 315         tgt_desc->ltd_index  = index;
 316         tgt_desc->ltd_active = active;
 317
 318         lod_getref(ltd);
 319         if (index >= ltd->ltd_tgts_size) {
 320                 /* we have to increase the size of the lod_osts array */
 321                 __u32  newsize;
 322
 323                 newsize = max(ltd->ltd_tgts_size, (__u32)2);
 324                 while (newsize < index + 1)
 325                         newsize = newsize << 1;
 326
 327                 /* lod_bitmap_resize() needs lod_rw_sem
 328                  * which we hold with th reference */
 329                 lod_putref(lod, ltd);
 330
 331                 rc = ltd_bitmap_resize(ltd, newsize);
 332                 if (rc)
 333                         GOTO(out_desc, rc);
 334
 335                 lod_getref(ltd);
 336         }
 337
 338         mutex_lock(&ltd->ltd_mutex);
 339         lock = true;
 340         if (cfs_bitmap_check(ltd->ltd_tgt_bitmap, index)) {
 341                 CERROR("%s: device %d is registered already\n", obd->obd_name,
 342                        index);
 343                 GOTO(out_mutex, rc = -EEXIST);
 344         }
 345
 346         if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
 347                 OBD_ALLOC_PTR(ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK]);
 348                 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
 349                         CERROR("can't allocate index to add %s\n",
 350                                obd->obd_name);
 351                         GOTO(out_mutex, rc = -ENOMEM);
 352                 }
 353         }
 354
 355         if (for_ost) {
 356                 /* pool and qos are not supported for MDS stack yet */
 357                 rc = lod_ost_pool_add(&lod->lod_pool_info, index,
 358                                       lod->lod_osts_size);
 359                 if (rc) {
 360                         CERROR("%s: can't set up pool, failed with %d\n",
 361                                obd->obd_name, rc);
 362                         GOTO(out_mutex, rc);
 363                 }
 364
 365                 rc = qos_add_tgt(lod, tgt_desc);
 366                 if (rc) {
 367                         CERROR("%s: qos_add_tgt failed with %d\n",
 368                                 obd->obd_name, rc);
 369                         GOTO(out_pool, rc);
 370                 }
 371
 372                 /* The new OST is now a full citizen */
 373                 if (index >= lod->lod_desc.ld_tgt_count)
 374                         lod->lod_desc.ld_tgt_count = index + 1;
 375                 if (active)
 376                         lod->lod_desc.ld_active_tgt_count++;
 377         }
 378
 379         LTD_TGT(ltd, index) = tgt_desc;
 380         cfs_bitmap_set(ltd->ltd_tgt_bitmap, index);
 381         ltd->ltd_tgtnr++;
 382         mutex_unlock(&ltd->ltd_mutex);
 383         lod_putref(lod, ltd);
 384         lock = false;
 385         if (lod->lod_recovery_completed)
 386                 lu_dev->ld_ops->ldo_recovery_complete(env, lu_dev);
 387
 388         if (!for_ost && lod->lod_initialized) {
 389                 rc = lod_sub_init_llog(env, lod, tgt_desc->ltd_tgt);
 390                 if (rc != 0) {
 391                         CERROR("%s: cannot start llog on %s:rc = %d\n",
 392                                lod2obd(lod)->obd_name, osp, rc);
 393                         GOTO(out_ltd, rc);
 394                 }
 395         }
 396
 397         rc = lfsck_add_target(env, lod->lod_child, dt_dev, exp, index, for_ost);
 398         if (rc != 0) {
 399                 CERROR("Fail to add LFSCK target: name = %s, type = %s, "
 400                        "index = %u, rc = %d\n", osp, type, index, rc);
 401                 GOTO(out_fini_llog, rc);
 402         }
 403         RETURN(rc);
 404 out_fini_llog:
 405         lod_sub_fini_llog(env, tgt_desc->ltd_tgt,
 406                           tgt_desc->ltd_recovery_thread);
 407 out_ltd:
 408         lod_getref(ltd);
 409         mutex_lock(&ltd->ltd_mutex);
 410         lock = true;
 411         if (!for_ost && LTD_TGT(ltd, index)->ltd_recovery_thread != NULL) {
 412                 struct ptlrpc_thread *thread;
 413
 414                 thread = LTD_TGT(ltd, index)->ltd_recovery_thread;
 415                 OBD_FREE_PTR(thread);
 416         }
 417         ltd->ltd_tgtnr--;
 418         cfs_bitmap_clear(ltd->ltd_tgt_bitmap, index);
 419         LTD_TGT(ltd, index) = NULL;
 420 out_pool:
 421         lod_ost_pool_remove(&lod->lod_pool_info, index);
 422 out_mutex:
 423         if (lock) {
 424                 mutex_unlock(&ltd->ltd_mutex);
 425                 lod_putref(lod, ltd);
 426         }
 427 out_desc:
 428         OBD_FREE_PTR(tgt_desc);
 429 out_conn:
 430         obd_disconnect(exp);
 431 out_cleanup:
 432         /* XXX OSP needs us to send down LCFG_CLEANUP because it uses
 433          * objects from the MDT stack. See LU-7184. */
 434         lcfg = &lod_env_info(env)->lti_lustre_cfg;
 435         memset(lcfg, 0, sizeof(*lcfg));
 436         lcfg->lcfg_version = LUSTRE_CFG_VERSION;
 437         lcfg->lcfg_command = LCFG_CLEANUP;
 438         lu_dev->ld_ops->ldo_process_config(env, lu_dev, lcfg);
 439
 440         return rc;
 441 }
 442
 443 /**
 444  * Schedule target removal from the target table.
 445  *
 446  * Mark the device as dead. The device is not removed here because it may
 447  * still be in use. The device will be removed in lod_putref() when the
 448  * last reference is released.
 449  *
 450  * \param[in] env               execution environment for this thread
 451  * \param[in] lod               LOD device the target table belongs to
 452  * \param[in] ltd               target table
 453  * \param[in] idx               index of the target
 454  * \param[in] for_ost           type of the target: 0 - MDT, 1 - OST
 455  */
 456 static void __lod_del_device(const struct lu_env *env, struct lod_device *lod,
 457                              struct lod_tgt_descs *ltd, unsigned idx,
 458                              bool for_ost)
 459 {
 460         LASSERT(LTD_TGT(ltd, idx));
 461
 462         lfsck_del_target(env, lod->lod_child, LTD_TGT(ltd, idx)->ltd_tgt,
 463                          idx, for_ost);
 464
 465         if (!for_ost && LTD_TGT(ltd, idx)->ltd_recovery_thread != NULL) {
 466                 struct ptlrpc_thread *thread;
 467
 468                 thread = LTD_TGT(ltd, idx)->ltd_recovery_thread;
 469                 OBD_FREE_PTR(thread);
 470         }
 471
 472         if (LTD_TGT(ltd, idx)->ltd_reap == 0) {
 473                 LTD_TGT(ltd, idx)->ltd_reap = 1;
 474                 ltd->ltd_death_row++;
 475         }
 476 }
 477
 478 /**
 479  * Schedule removal of all the targets from the given target table.
 480  *
 481  * See more details in the description for __lod_del_device()
 482  *
 483  * \param[in] env               execution environment for this thread
 484  * \param[in] lod               LOD device the target table belongs to
 485  * \param[in] ltd               target table
 486  * \param[in] for_ost           type of the target: MDT or OST
 487  *
 488  * \retval                      0 always
 489  */
 490 int lod_fini_tgt(const struct lu_env *env, struct lod_device *lod,
 491                  struct lod_tgt_descs *ltd, bool for_ost)
 492 {
 493         unsigned int idx;
 494
 495         if (ltd->ltd_tgts_size <= 0)
 496                 return 0;
 497         lod_getref(ltd);
 498         mutex_lock(&ltd->ltd_mutex);
 499         cfs_foreach_bit(ltd->ltd_tgt_bitmap, idx)
 500                 __lod_del_device(env, lod, ltd, idx, for_ost);
 501         mutex_unlock(&ltd->ltd_mutex);
 502         lod_putref(lod, ltd);
 503         CFS_FREE_BITMAP(ltd->ltd_tgt_bitmap);
 504         for (idx = 0; idx < TGT_PTRS; idx++) {
 505                 if (ltd->ltd_tgt_idx[idx])
 506                         OBD_FREE_PTR(ltd->ltd_tgt_idx[idx]);
 507         }
 508         ltd->ltd_tgts_size = 0;
 509         return 0;
 510 }
 511
 512 /**
 513  * Remove device by name.
 514  *
 515  * Remove a device identified by \a osp from the target table. Given
 516  * the device can be in use, the real deletion happens in lod_putref().
 517  *
 518  * \param[in] env               execution environment for this thread
 519  * \param[in] lod               LOD device to be connected to the new OSP
 520  * \param[in] ltd               target table
 521  * \param[in] osp               name of OSP device to be removed
 522  * \param[in] idx               index of the target
 523  * \param[in] gen               generation number, not used currently
 524  * \param[in] for_ost           type of the target: 0 - MDT, 1 - OST
 525  *
 526  * \retval                      0 if the device was scheduled for removal
 527  * \retval                      -EINVAL if no device was found
 528  */
 529 int lod_del_device(const struct lu_env *env, struct lod_device *lod,
 530                    struct lod_tgt_descs *ltd, char *osp, unsigned idx,
 531                    unsigned gen, bool for_ost)
 532 {
 533         struct obd_device *obd;
 534         int                rc = 0;
 535         struct obd_uuid    uuid;
 536         ENTRY;
 537
 538         CDEBUG(D_CONFIG, "osp:%s idx:%d gen:%d\n", osp, idx, gen);
 539
 540         obd_str2uuid(&uuid, osp);
 541
 542         obd = class_find_client_obd(&uuid, LUSTRE_OSP_NAME,
 543                                    &lod->lod_dt_dev.dd_lu_dev.ld_obd->obd_uuid);
 544         if (obd == NULL) {
 545                 CERROR("can't find %s device\n", osp);
 546                 RETURN(-EINVAL);
 547         }
 548
 549         if (gen <= 0) {
 550                 CERROR("%s: request to remove OBD %s with invalid generation %d"
 551                        "\n", obd->obd_name, osp, gen);
 552                 RETURN(-EINVAL);
 553         }
 554
 555         obd_str2uuid(&uuid,  osp);
 556
 557         lod_getref(ltd);
 558         mutex_lock(&ltd->ltd_mutex);
 559         /* check that the index is allocated in the bitmap */
 560         if (!cfs_bitmap_check(ltd->ltd_tgt_bitmap, idx) ||
 561             !LTD_TGT(ltd, idx)) {
 562                 CERROR("%s: device %d is not set up\n", obd->obd_name, idx);
 563                 GOTO(out, rc = -EINVAL);
 564         }
 565
 566         /* check that the UUID matches */
 567         if (!obd_uuid_equals(&uuid, &LTD_TGT(ltd, idx)->ltd_uuid)) {
 568                 CERROR("%s: LOD target UUID %s at index %d does not match %s\n",
 569                        obd->obd_name, obd_uuid2str(&LTD_TGT(ltd,idx)->ltd_uuid),
 570                        idx, osp);
 571                 GOTO(out, rc = -EINVAL);
 572         }
 573
 574         __lod_del_device(env, lod, ltd, idx, for_ost);
 575         EXIT;
 576 out:
 577         mutex_unlock(&ltd->ltd_mutex);
 578         lod_putref(lod, ltd);
 579         return(rc);
 580 }
 581
 582 /**
 583  * Resize per-thread storage to hold specified size.
 584  *
 585  * A helper function to resize per-thread temporary storage. This storage
 586  * is used to process LOV/LVM EAs and may be quite large. We do not want to
 587  * allocate/release it every time, so instead we put it into the env and
 588  * reallocate on demand. The memory is released when the correspondent thread
 589  * is finished.
 590  *
 591  * \param[in] info              LOD-specific storage in the environment
 592  * \param[in] size              new size to grow the buffer to
 593
 594  * \retval                      0 on success, -ENOMEM if reallocation failed
 595  */
 596 int lod_ea_store_resize(struct lod_thread_info *info, size_t size)
 597 {
 598         __u32 round = size_roundup_power2(size);
 599
 600         LASSERT(round <=
 601                 lov_mds_md_size(LOV_MAX_STRIPE_COUNT, LOV_MAGIC_V3));
 602         if (info->lti_ea_store) {
 603                 LASSERT(info->lti_ea_store_size);
 604                 LASSERT(info->lti_ea_store_size < round);
 605                 CDEBUG(D_INFO, "EA store size %d is not enough, need %d\n",
 606                        info->lti_ea_store_size, round);
 607                 OBD_FREE_LARGE(info->lti_ea_store, info->lti_ea_store_size);
 608                 info->lti_ea_store = NULL;
 609                 info->lti_ea_store_size = 0;
 610         }
 611
 612         OBD_ALLOC_LARGE(info->lti_ea_store, round);
 613         if (info->lti_ea_store == NULL)
 614                 RETURN(-ENOMEM);
 615         info->lti_ea_store_size = round;
 616
 617         RETURN(0);
 618 }
 619
 620 static void lod_free_comp_buffer(struct lod_layout_component *entries,
 621                                  __u16 count, __u32 bufsize)
 622 {
 623         struct lod_layout_component *entry;
 624         int i;
 625
 626         for (i = 0; i < count; i++) {
 627                 entry = &entries[i];
 628                 if (entry->llc_pool != NULL)
 629                         lod_set_pool(&entry->llc_pool, NULL);
 630                 if (entry->llc_ostlist.op_array)
 631                         OBD_FREE(entry->llc_ostlist.op_array,
 632                                  entry->llc_ostlist.op_size);
 633                 LASSERT(entry->llc_stripe == NULL);
 634                 LASSERT(entry->llc_stripes_allocated == 0);
 635         }
 636
 637         if (bufsize != 0)
 638                 OBD_FREE_LARGE(entries, bufsize);
 639 }
 640
 641 void lod_free_def_comp_entries(struct lod_default_striping *lds)
 642 {
 643         lod_free_comp_buffer(lds->lds_def_comp_entries,
 644                              lds->lds_def_comp_size_cnt,
 645                              size_roundup_power2(
 646                                      sizeof(*lds->lds_def_comp_entries) *
 647                                      lds->lds_def_comp_size_cnt));
 648         lds->lds_def_comp_entries = NULL;
 649         lds->lds_def_comp_cnt = 0;
 650         lds->lds_def_striping_is_composite = 0;
 651         lds->lds_def_comp_size_cnt = 0;
 652 }
 653
 654 /**
 655  * Resize per-thread storage to hold default striping component entries
 656  *
 657  * A helper function to resize per-thread temporary storage. This storage
 658  * is used to hold default LOV/LVM EAs and may be quite large. We do not want
 659  * to allocate/release it every time, so instead we put it into the env and
 660  * reallocate it on demand. The memory is released when the correspondent
 661  * thread is finished.
 662  *
 663  * \param[in,out] lds           default striping
 664  * \param[in] count             new component count to grow the buffer to
 665
 666  * \retval                      0 on success, -ENOMEM if reallocation failed
 667  */
 668 int lod_def_striping_comp_resize(struct lod_default_striping *lds, __u16 count)
 669 {
 670         struct lod_layout_component *entries;
 671         __u32 new = size_roundup_power2(sizeof(*lds->lds_def_comp_entries) *
 672                                         count);
 673         __u32 old = size_roundup_power2(sizeof(*lds->lds_def_comp_entries) *
 674                                         lds->lds_def_comp_size_cnt);
 675
 676         if (new <= old)
 677                 return 0;
 678
 679         OBD_ALLOC_LARGE(entries, new);
 680         if (entries == NULL)
 681                 return -ENOMEM;
 682
 683         if (lds->lds_def_comp_entries != NULL) {
 684                 CDEBUG(D_INFO, "default striping component size %d is not "
 685                        "enough, need %d\n", old, new);
 686                 lod_free_def_comp_entries(lds);
 687         }
 688
 689         lds->lds_def_comp_entries = entries;
 690         lds->lds_def_comp_size_cnt = count;
 691
 692         RETURN(0);
 693 }
 694
 695 void lod_free_comp_entries(struct lod_object *lo)
 696 {
 697         lod_free_comp_buffer(lo->ldo_comp_entries,
 698                              lo->ldo_comp_cnt,
 699                              sizeof(*lo->ldo_comp_entries) * lo->ldo_comp_cnt);
 700         lo->ldo_comp_entries = NULL;
 701         lo->ldo_comp_cnt = 0;
 702         lo->ldo_is_composite = 0;
 703 }
 704
 705 int lod_alloc_comp_entries(struct lod_object *lo, int cnt)
 706 {
 707         LASSERT(cnt != 0);
 708         LASSERT(lo->ldo_comp_cnt == 0 && lo->ldo_comp_entries == NULL);
 709
 710         OBD_ALLOC_LARGE(lo->ldo_comp_entries,
 711                         sizeof(*lo->ldo_comp_entries) * cnt);
 712         if (lo->ldo_comp_entries == NULL)
 713                 return -ENOMEM;
 714         lo->ldo_comp_cnt = cnt;
 715         return 0;
 716 }
 717
 718 /**
 719  * Generate on-disk lov_mds_md structure for each layout component based on
 720  * the information in lod_object->ldo_comp_entries[i].
 721  *
 722  * \param[in] env               execution environment for this thread
 723  * \param[in] lo                LOD object
 724  * \param[in] comp_idx          index of ldo_comp_entries
 725  * \param[in] lmm               buffer to cotain the on-disk lov_mds_md
 726  * \param[in|out] lmm_size      buffer size/lmm size
 727  * \param[in] is_dir            generate lov ea for dir or file? For dir case,
 728  *                              the stripe info is from the default stripe
 729  *                              template, which is collected in lod_ah_init(),
 730  *                              either from parent object or root object; for
 731  *                              file case, it's from the @lo object
 732  *
 733  * \retval                      0 if on disk structure is created successfully
 734  * \retval                      negative error number on failure
 735  */
 736 static int lod_gen_component_ea(const struct lu_env *env,
 737                                 struct lod_object *lo, int comp_idx,
 738                                 struct lov_mds_md *lmm, int *lmm_size,
 739                                 bool is_dir)
 740 {
 741         struct lod_thread_info  *info = lod_env_info(env);
 742         const struct lu_fid     *fid  = lu_object_fid(&lo->ldo_obj.do_lu);
 743         struct lod_device       *lod;
 744         struct lov_ost_data_v1  *objs;
 745         struct lod_layout_component *lod_comp;
 746         __u32   magic;
 747         __u16 stripe_count;
 748         int     i, rc = 0;
 749         ENTRY;
 750
 751         LASSERT(lo);
 752         if (is_dir)
 753                 lod_comp =
 754                         &lo->ldo_def_striping->lds_def_comp_entries[comp_idx];
 755         else
 756                 lod_comp = &lo->ldo_comp_entries[comp_idx];
 757
 758         magic = lod_comp->llc_pool != NULL ? LOV_MAGIC_V3 : LOV_MAGIC_V1;
 759         if (lod_comp->llc_pattern == 0) /* default striping */
 760                 lod_comp->llc_pattern = LOV_PATTERN_RAID0;
 761
 762         lmm->lmm_magic = cpu_to_le32(magic);
 763         lmm->lmm_pattern = cpu_to_le32(lod_comp->llc_pattern);
 764         fid_to_lmm_oi(fid, &lmm->lmm_oi);
 765         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_LMMOI))
 766                 lmm->lmm_oi.oi.oi_id++;
 767         lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
 768
 769         lmm->lmm_stripe_size = cpu_to_le32(lod_comp->llc_stripe_size);
 770         lmm->lmm_stripe_count = cpu_to_le16(lod_comp->llc_stripe_count);
 771         /**
 772          * for dir and uninstantiated component, lmm_layout_gen stores
 773          * default stripe offset.
 774          */
 775         lmm->lmm_layout_gen =
 776                 (is_dir || !lod_comp_inited(lod_comp)) ?
 777                         cpu_to_le16(lod_comp->llc_stripe_offset) :
 778                         cpu_to_le16(lod_comp->llc_layout_gen);
 779
 780         if (magic == LOV_MAGIC_V1) {
 781                 objs = &lmm->lmm_objects[0];
 782         } else {
 783                 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
 784                 size_t cplen = strlcpy(v3->lmm_pool_name,
 785                                        lod_comp->llc_pool,
 786                                        sizeof(v3->lmm_pool_name));
 787                 if (cplen >= sizeof(v3->lmm_pool_name))
 788                         RETURN(-E2BIG);
 789                 objs = &v3->lmm_objects[0];
 790         }
 791         stripe_count = lod_comp_entry_stripe_count(lo, lod_comp, is_dir);
 792         if (!is_dir && lo->ldo_is_composite)
 793                 lod_comp_shrink_stripe_count(lod_comp, &stripe_count);
 794
 795         if (is_dir || lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)
 796                 GOTO(done, rc = 0);
 797
 798         /* generate ost_idx of this component stripe */
 799         lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
 800         for (i = 0; i < stripe_count; i++) {
 801                 struct dt_object *object;
 802                 __u32 ost_idx = (__u32)-1UL;
 803                 int type = LU_SEQ_RANGE_OST;
 804
 805                 if (lod_comp->llc_stripe && lod_comp->llc_stripe[i]) {
 806                         object = lod_comp->llc_stripe[i];
 807                         /* instantiated component */
 808                         info->lti_fid = *lu_object_fid(&object->do_lu);
 809
 810                         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_MULTIPLE_REF) &&
 811                             comp_idx == 0) {
 812                                 if (cfs_fail_val == 0)
 813                                         cfs_fail_val = info->lti_fid.f_oid;
 814                                 else if (i == 0)
 815                                         info->lti_fid.f_oid = cfs_fail_val;
 816                         }
 817
 818                         rc = fid_to_ostid(&info->lti_fid, &info->lti_ostid);
 819                         LASSERT(rc == 0);
 820
 821                         ostid_cpu_to_le(&info->lti_ostid, &objs[i].l_ost_oi);
 822                         objs[i].l_ost_gen = cpu_to_le32(0);
 823                         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FLD_LOOKUP))
 824                                 rc = -ENOENT;
 825                         else
 826                                 rc = lod_fld_lookup(env, lod, &info->lti_fid,
 827                                                     &ost_idx, &type);
 828                         if (rc < 0) {
 829                                 CERROR("%s: Can not locate "DFID": rc = %d\n",
 830                                        lod2obd(lod)->obd_name,
 831                                        PFID(&info->lti_fid), rc);
 832                                 RETURN(rc);
 833                         }
 834                 } else if (lod_comp->llc_ostlist.op_array) {
 835                         /* user specified ost list */
 836                         ost_idx = lod_comp->llc_ostlist.op_array[i];
 837                 }
 838                 /*
 839                  * with un-instantiated or with no specified ost list
 840                  * component, its l_ost_idx does not matter.
 841                  */
 842                 objs[i].l_ost_idx = cpu_to_le32(ost_idx);
 843         }
 844 done:
 845         if (lmm_size != NULL)
 846                 *lmm_size = lov_mds_md_size(stripe_count, magic);
 847         RETURN(rc);
 848 }
 849
 850 /**
 851  * Generate component ID for new created component.
 852  *
 853  * \param[in] lo                LOD object
 854  * \param[in] comp_idx          index of ldo_comp_entries
 855  *
 856  * \retval                      component ID on success
 857  * \retval                      LCME_ID_INVAL on failure
 858  */
 859 static __u32 lod_gen_component_id(struct lod_object *lo, int comp_idx)
 860 {
 861         struct lod_layout_component *lod_comp;
 862         __u32   id, start, end;
 863         int     i;
 864
 865         LASSERT(lo->ldo_comp_entries[comp_idx].llc_id == LCME_ID_INVAL);
 866
 867         lod_obj_inc_layout_gen(lo);
 868         id = lo->ldo_layout_gen;
 869         if (likely(id <= LCME_ID_MAX))
 870                 return id;
 871
 872         /* Layout generation wraps, need to check collisions. */
 873         start = id & LCME_ID_MASK;
 874         end = LCME_ID_MAX;
 875 again:
 876         for (id = start; id <= end; id++) {
 877                 for (i = 0; i < lo->ldo_comp_cnt; i++) {
 878                         lod_comp = &lo->ldo_comp_entries[i];
 879                         if (id == lod_comp->llc_id)
 880                                 break;
 881                 }
 882                 /* Found the ununsed ID */
 883                 if (i == lo->ldo_comp_cnt)
 884                         return id;
 885         }
 886         if (end == LCME_ID_MAX) {
 887                 start = 1;
 888                 end = min(lo->ldo_layout_gen & LCME_ID_MASK,
 889                           (__u32)(LCME_ID_MAX - 1));
 890                 goto again;
 891         }
 892
 893         return LCME_ID_INVAL;
 894 }
 895
 896 /**
 897  * Generate on-disk lov_mds_md structure based on the information in
 898  * the lod_object->ldo_comp_entries.
 899  *
 900  * \param[in] env               execution environment for this thread
 901  * \param[in] lo                LOD object
 902  * \param[in] lmm               buffer to cotain the on-disk lov_mds_md
 903  * \param[in|out] lmm_size      buffer size/lmm size
 904  * \param[in] is_dir            generate lov ea for dir or file? For dir case,
 905  *                              the stripe info is from the default stripe
 906  *                              template, which is collected in lod_ah_init(),
 907  *                              either from parent object or root object; for
 908  *                              file case, it's from the @lo object
 909  *
 910  * \retval                      0 if on disk structure is created successfully
 911  * \retval                      negative error number on failure
 912  */
 913 int lod_generate_lovea(const struct lu_env *env, struct lod_object *lo,
 914                        struct lov_mds_md *lmm, int *lmm_size, bool is_dir)
 915 {
 916         struct lov_comp_md_entry_v1 *lcme;
 917         struct lov_comp_md_v1 *lcm;
 918         struct lod_layout_component *comp_entries;
 919         __u16 comp_cnt;
 920         bool is_composite;
 921         int i, rc = 0, offset;
 922         ENTRY;
 923
 924         if (is_dir) {
 925                 comp_cnt = lo->ldo_def_striping->lds_def_comp_cnt;
 926                 comp_entries = lo->ldo_def_striping->lds_def_comp_entries;
 927                 is_composite =
 928                         lo->ldo_def_striping->lds_def_striping_is_composite;
 929         } else {
 930                 comp_cnt = lo->ldo_comp_cnt;
 931                 comp_entries = lo->ldo_comp_entries;
 932                 is_composite = lo->ldo_is_composite;
 933         }
 934
 935         LASSERT(lmm_size != NULL);
 936         LASSERT(comp_cnt != 0 && comp_entries != NULL);
 937
 938         if (!is_composite) {
 939                 rc = lod_gen_component_ea(env, lo, 0, lmm, lmm_size, is_dir);
 940                 RETURN(rc);
 941         }
 942
 943         lcm = (struct lov_comp_md_v1 *)lmm;
 944         lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
 945         lcm->lcm_entry_count = cpu_to_le16(comp_cnt);
 946
 947         offset = sizeof(*lcm) + sizeof(*lcme) * comp_cnt;
 948         LASSERT(offset % sizeof(__u64) == 0);
 949
 950         for (i = 0; i < comp_cnt; i++) {
 951                 struct lod_layout_component *lod_comp;
 952                 struct lov_mds_md *sub_md;
 953                 int size;
 954
 955                 lod_comp = &comp_entries[i];
 956                 lcme = &lcm->lcm_entries[i];
 957
 958                 if (lod_comp->llc_id == LCME_ID_INVAL && !is_dir) {
 959                         lod_comp->llc_id = lod_gen_component_id(lo, i);
 960                         if (lod_comp->llc_id == LCME_ID_INVAL)
 961                                 GOTO(out, rc = -ERANGE);
 962                 }
 963                 lcme->lcme_id = cpu_to_le32(lod_comp->llc_id);
 964
 965                 /* component could be un-inistantiated */
 966                 lcme->lcme_flags = cpu_to_le32(lod_comp->llc_flags);
 967                 lcme->lcme_extent.e_start =
 968                         cpu_to_le64(lod_comp->llc_extent.e_start);
 969                 lcme->lcme_extent.e_end =
 970                         cpu_to_le64(lod_comp->llc_extent.e_end);
 971                 lcme->lcme_offset = cpu_to_le32(offset);
 972
 973                 sub_md = (struct lov_mds_md *)((char *)lcm + offset);
 974                 rc = lod_gen_component_ea(env, lo, i, sub_md, &size, is_dir);
 975                 if (rc)
 976                         GOTO(out, rc);
 977                 lcme->lcme_size = cpu_to_le32(size);
 978                 offset += size;
 979                 LASSERTF((offset <= *lmm_size) && (offset % sizeof(__u64) == 0),
 980                          "offset:%d lmm_size:%d\n", offset, *lmm_size);
 981         }
 982         lcm->lcm_size = cpu_to_le32(offset);
 983         lcm->lcm_layout_gen = cpu_to_le32(is_dir ? 0 : lo->ldo_layout_gen);
 984
 985         lustre_print_user_md(D_LAYOUT, (struct lov_user_md *)lmm,
 986                              "generate lum");
 987 out:
 988         if (rc == 0)
 989                 *lmm_size = offset;
 990         RETURN(rc);
 991 }
 992
 993 /**
 994  * Get LOV EA.
 995  *
 996  * Fill lti_ea_store buffer in the environment with a value for the given
 997  * EA. The buffer is reallocated if the value doesn't fit.
 998  *
 999  * \param[in,out] env           execution environment for this thread
1000  *                              .lti_ea_store buffer is filled with EA's value
1001  * \param[in] lo                LOD object
1002  * \param[in] name              name of the EA
1003  *
1004  * \retval                      > 0 if EA is fetched successfully
1005  * \retval                      0 if EA is empty
1006  * \retval                      negative error number on failure
1007  */
1008 int lod_get_ea(const struct lu_env *env, struct lod_object *lo,
1009                const char *name)
1010 {
1011         struct lod_thread_info  *info = lod_env_info(env);
1012         struct dt_object        *next = dt_object_child(&lo->ldo_obj);
1013         int                     rc;
1014         ENTRY;
1015
1016         LASSERT(info);
1017
1018         if (unlikely(info->lti_ea_store == NULL)) {
1019                 /* just to enter in allocation block below */
1020                 rc = -ERANGE;
1021         } else {
1022 repeat:
1023                 info->lti_buf.lb_buf = info->lti_ea_store;
1024                 info->lti_buf.lb_len = info->lti_ea_store_size;
1025                 rc = dt_xattr_get(env, next, &info->lti_buf, name);
1026         }
1027
1028         /* if object is not striped or inaccessible */
1029         if (rc == -ENODATA || rc == -ENOENT)
1030                 RETURN(0);
1031
1032         if (rc == -ERANGE) {
1033                 /* EA doesn't fit, reallocate new buffer */
1034                 rc = dt_xattr_get(env, next, &LU_BUF_NULL, name);
1035                 if (rc == -ENODATA || rc == -ENOENT)
1036                         RETURN(0);
1037                 else if (rc < 0)
1038                         RETURN(rc);
1039
1040                 LASSERT(rc > 0);
1041                 rc = lod_ea_store_resize(info, rc);
1042                 if (rc)
1043                         RETURN(rc);
1044                 goto repeat;
1045         }
1046
1047         RETURN(rc);
1048 }
1049
1050 /**
1051  * Verify the target index is present in the current configuration.
1052  *
1053  * \param[in] md                LOD device where the target table is stored
1054  * \param[in] idx               target's index
1055  *
1056  * \retval                      0 if the index is present
1057  * \retval                      -EINVAL if not
1058  */
1059 static int validate_lod_and_idx(struct lod_device *md, __u32 idx)
1060 {
1061         if (unlikely(idx >= md->lod_ost_descs.ltd_tgts_size ||
1062                      !cfs_bitmap_check(md->lod_ost_bitmap, idx))) {
1063                 CERROR("%s: bad idx: %d of %d\n", lod2obd(md)->obd_name, idx,
1064                        md->lod_ost_descs.ltd_tgts_size);
1065                 return -EINVAL;
1066         }
1067
1068         if (unlikely(OST_TGT(md, idx) == NULL)) {
1069                 CERROR("%s: bad lod_tgt_desc for idx: %d\n",
1070                        lod2obd(md)->obd_name, idx);
1071                 return -EINVAL;
1072         }
1073
1074         if (unlikely(OST_TGT(md, idx)->ltd_ost == NULL)) {
1075                 CERROR("%s: invalid lod device, for idx: %d\n",
1076                        lod2obd(md)->obd_name , idx);
1077                 return -EINVAL;
1078         }
1079
1080         return 0;
1081 }
1082
1083 /**
1084  * Instantiate objects for stripes.
1085  *
1086  * Allocate and initialize LU-objects representing the stripes. The number
1087  * of the stripes (ldo_stripe_count) must be initialized already. The caller
1088  * must ensure nobody else is calling the function on the object at the same
1089  * time. FLDB service must be running to be able to map a FID to the targets
1090  * and find appropriate device representing that target.
1091  *
1092  * \param[in] env               execution environment for this thread
1093  * \param[in,out] lo            LOD object
1094  * \param[in] objs              an array of IDs to creates the objects from
1095  * \param[in] comp_idx          index of ldo_comp_entries
1096  *
1097  * \retval                      0 if the objects are instantiated successfully
1098  * \retval                      negative error number on failure
1099  */
1100 int lod_initialize_objects(const struct lu_env *env, struct lod_object *lo,
1101                            struct lov_ost_data_v1 *objs, int comp_idx)
1102 {
1103         struct lod_layout_component     *lod_comp;
1104         struct lod_thread_info  *info = lod_env_info(env);
1105         struct lod_device       *md;
1106         struct lu_object        *o, *n;
1107         struct lu_device        *nd;
1108         struct dt_object       **stripe;
1109         int                      stripe_len;
1110         int                      i, rc = 0;
1111         __u32                   idx;
1112         ENTRY;
1113
1114         LASSERT(lo != NULL);
1115         md = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1116
1117         LASSERT(lo->ldo_comp_cnt != 0 && lo->ldo_comp_entries != NULL);
1118         lod_comp = &lo->ldo_comp_entries[comp_idx];
1119
1120         LASSERT(lod_comp->llc_stripe == NULL);
1121         LASSERT(lod_comp->llc_stripe_count > 0);
1122         LASSERT(lod_comp->llc_stripe_size > 0);
1123
1124         stripe_len = lod_comp->llc_stripe_count;
1125         OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
1126         if (stripe == NULL)
1127                 RETURN(-ENOMEM);
1128
1129         for (i = 0; i < lod_comp->llc_stripe_count; i++) {
1130                 if (unlikely(lovea_slot_is_dummy(&objs[i])))
1131                         continue;
1132
1133                 ostid_le_to_cpu(&objs[i].l_ost_oi, &info->lti_ostid);
1134                 idx = le32_to_cpu(objs[i].l_ost_idx);
1135                 rc = ostid_to_fid(&info->lti_fid, &info->lti_ostid, idx);
1136                 if (rc != 0)
1137                         GOTO(out, rc);
1138                 LASSERTF(fid_is_sane(&info->lti_fid), ""DFID" insane!\n",
1139                          PFID(&info->lti_fid));
1140                 lod_getref(&md->lod_ost_descs);
1141
1142                 rc = validate_lod_and_idx(md, idx);
1143                 if (unlikely(rc != 0)) {
1144                         lod_putref(md, &md->lod_ost_descs);
1145                         GOTO(out, rc);
1146                 }
1147
1148                 nd = &OST_TGT(md,idx)->ltd_ost->dd_lu_dev;
1149                 lod_putref(md, &md->lod_ost_descs);
1150
1151                 /* In the function below, .hs_keycmp resolves to
1152                  * u_obj_hop_keycmp() */
1153                 /* coverity[overrun-buffer-val] */
1154                 o = lu_object_find_at(env, nd, &info->lti_fid, NULL);
1155                 if (IS_ERR(o))
1156                         GOTO(out, rc = PTR_ERR(o));
1157
1158                 n = lu_object_locate(o->lo_header, nd->ld_type);
1159                 LASSERT(n);
1160
1161                 stripe[i] = container_of(n, struct dt_object, do_lu);
1162         }
1163
1164 out:
1165         if (rc != 0) {
1166                 for (i = 0; i < stripe_len; i++)
1167                         if (stripe[i] != NULL)
1168                                 dt_object_put(env, stripe[i]);
1169
1170                 OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
1171                 lod_comp->llc_stripe_count = 0;
1172         } else {
1173                 lod_comp->llc_stripe = stripe;
1174                 lod_comp->llc_stripes_allocated = stripe_len;
1175         }
1176
1177         RETURN(rc);
1178 }
1179
1180 /**
1181  * Instantiate objects for striping.
1182  *
1183  * Parse striping information in \a buf and instantiate the objects
1184  * representing the stripes.
1185  *
1186  * \param[in] env               execution environment for this thread
1187  * \param[in] lo                LOD object
1188  * \param[in] buf               buffer storing LOV EA to parse
1189  *
1190  * \retval                      0 if parsing and objects creation succeed
1191  * \retval                      negative error number on failure
1192  */
1193 int lod_parse_striping(const struct lu_env *env, struct lod_object *lo,
1194                        const struct lu_buf *buf)
1195 {
1196         struct lov_mds_md_v1    *lmm;
1197         struct lov_comp_md_v1   *comp_v1 = NULL;
1198         struct lov_ost_data_v1  *objs;
1199         __u32   magic, pattern;
1200         int     i, j, rc = 0;
1201         __u16   comp_cnt;
1202         ENTRY;
1203
1204         LASSERT(buf);
1205         LASSERT(buf->lb_buf);
1206         LASSERT(buf->lb_len);
1207
1208         lmm = (struct lov_mds_md_v1 *)buf->lb_buf;
1209         magic = le32_to_cpu(lmm->lmm_magic);
1210
1211         if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 &&
1212             magic != LOV_MAGIC_COMP_V1)
1213                 GOTO(out, rc = -EINVAL);
1214
1215         lod_free_comp_entries(lo);
1216
1217         if (magic == LOV_MAGIC_COMP_V1) {
1218                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1219                 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1220                 if (comp_cnt == 0)
1221                         GOTO(out, rc = -EINVAL);
1222                 lo->ldo_layout_gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1223                 lo->ldo_is_composite = 1;
1224         } else {
1225                 comp_cnt = 1;
1226                 lo->ldo_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1227                 lo->ldo_is_composite = 0;
1228         }
1229
1230         rc = lod_alloc_comp_entries(lo, comp_cnt);
1231         if (rc)
1232                 GOTO(out, rc);
1233
1234         for (i = 0; i < comp_cnt; i++) {
1235                 struct lod_layout_component     *lod_comp;
1236                 struct lu_extent        *ext;
1237                 __u32   offs;
1238
1239                 lod_comp = &lo->ldo_comp_entries[i];
1240                 if (lo->ldo_is_composite) {
1241                         offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1242                         lmm = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs);
1243                         magic = le32_to_cpu(lmm->lmm_magic);
1244
1245                         ext = &comp_v1->lcm_entries[i].lcme_extent;
1246                         lod_comp->llc_extent.e_start =
1247                                 le64_to_cpu(ext->e_start);
1248                         lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end);
1249                         lod_comp->llc_flags =
1250                                 le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags);
1251                         lod_comp->llc_id =
1252                                 le32_to_cpu(comp_v1->lcm_entries[i].lcme_id);
1253                         if (lod_comp->llc_id == LCME_ID_INVAL)
1254                                 GOTO(out, rc = -EINVAL);
1255                 } else {
1256                         lod_comp_set_init(lod_comp);
1257                 }
1258
1259                 pattern = le32_to_cpu(lmm->lmm_pattern);
1260                 if (lov_pattern(pattern) != LOV_PATTERN_RAID0 &&
1261                     lov_pattern(pattern) != LOV_PATTERN_MDT)
1262                         GOTO(out, rc = -EINVAL);
1263
1264                 lod_comp->llc_pattern = pattern;
1265                 lod_comp->llc_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
1266                 lod_comp->llc_stripe_count = le16_to_cpu(lmm->lmm_stripe_count);
1267                 lod_comp->llc_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1268
1269                 if (magic == LOV_MAGIC_V3) {
1270                         struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
1271                         objs = &v3->lmm_objects[0];
1272                         /* no need to set pool, which is used in create only */
1273                 } else {
1274                         objs = &lmm->lmm_objects[0];
1275                 }
1276
1277                 /**
1278                  * If uninstantiated template component has valid l_ost_idx,
1279                  * then user has specified ost list for this component.
1280                  */
1281                 if (!lod_comp_inited(lod_comp)) {
1282                         if (objs[0].l_ost_idx != (__u32)-1UL) {
1283                                 /**
1284                                  * load the user specified ost list, when this
1285                                  * component is instantiated later, it will be
1286                                  * used in lod_alloc_ost_list().
1287                                  */
1288                                 lod_comp->llc_ostlist.op_count =
1289                                         lod_comp->llc_stripe_count;
1290                                 lod_comp->llc_ostlist.op_size =
1291                                         lod_comp->llc_stripe_count *
1292                                         sizeof(__u32);
1293                                 OBD_ALLOC(lod_comp->llc_ostlist.op_array,
1294                                           lod_comp->llc_ostlist.op_size);
1295                                 if (!lod_comp->llc_ostlist.op_array)
1296                                         GOTO(out, rc = -ENOMEM);
1297
1298                                 for (j = 0; j < lod_comp->llc_stripe_count; j++)
1299                                         lod_comp->llc_ostlist.op_array[j] =
1300                                                 le32_to_cpu(objs[j].l_ost_idx);
1301
1302                                 /**
1303                                  * this component OST objects starts from the
1304                                  * first ost_idx, lod_alloc_ost_list() will
1305                                  * check this.
1306                                  */
1307                                 lod_comp->llc_stripe_offset = objs[0].l_ost_idx;
1308                         } else {
1309                                 /**
1310                                  * for uninstantiated component,
1311                                  * lmm_layout_gen stores default stripe offset.
1312                                  */
1313                                 lod_comp->llc_stripe_offset =
1314                                                         lmm->lmm_layout_gen;
1315                         }
1316                 }
1317
1318                 /* skip un-instantiated component object initialization */
1319                 if (!lod_comp_inited(lod_comp))
1320                         continue;
1321
1322                 if (!(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1323                     !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) {
1324                         rc = lod_initialize_objects(env, lo, objs, i);
1325                         if (rc)
1326                                 GOTO(out, rc);
1327                 }
1328         }
1329 out:
1330         if (rc)
1331                 lod_object_free_striping(env, lo);
1332         RETURN(rc);
1333 }
1334
1335 /**
1336  * Check whether the striping (LOVEA for regular file, LMVEA for directory)
1337  * is already cached.
1338  *
1339  * \param[in] lo        LOD object
1340  *
1341  * \retval              True if the striping is cached, otherwise
1342  *                      return false.
1343  */
1344 static bool lod_striping_loaded(struct lod_object *lo)
1345 {
1346         if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr) &&
1347             lo->ldo_comp_cached)
1348                 return true;
1349
1350         if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1351                 if (lo->ldo_stripe != NULL || lo->ldo_dir_stripe_loaded)
1352                         return true;
1353
1354                 /* Never load LMV stripe for slaves of striped dir */
1355                 if (lo->ldo_dir_slave_stripe)
1356                         return true;
1357         }
1358
1359         return false;
1360 }
1361
1362 /**
1363  * Initialize the object representing the stripes.
1364  *
1365  * Unless the stripes are initialized already, fetch LOV (for regular
1366  * objects) or LMV (for directory objects) EA and call lod_parse_striping()
1367  * to instantiate the objects representing the stripes. Caller should
1368  * hold the dt_write_lock(next).
1369  *
1370  * \param[in] env               execution environment for this thread
1371  * \param[in,out] lo            LOD object
1372  *
1373  * \retval                      0 if parsing and object creation succeed
1374  * \retval                      negative error number on failure
1375  */
1376 int lod_load_striping_locked(const struct lu_env *env, struct lod_object *lo)
1377 {
1378         struct lod_thread_info  *info = lod_env_info(env);
1379         struct lu_buf           *buf  = &info->lti_buf;
1380         struct dt_object        *next = dt_object_child(&lo->ldo_obj);
1381         int                      rc = 0;
1382         ENTRY;
1383
1384         if (!dt_object_exists(next))
1385                 GOTO(out, rc = 0);
1386
1387         if (lod_striping_loaded(lo))
1388                 GOTO(out, rc = 0);
1389
1390         if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr)) {
1391                 rc = lod_get_lov_ea(env, lo);
1392                 if (rc <= 0)
1393                         GOTO(out, rc);
1394                 /*
1395                  * there is LOV EA (striping information) in this object
1396                  * let's parse it and create in-core objects for the stripes
1397                  */
1398                 buf->lb_buf = info->lti_ea_store;
1399                 buf->lb_len = info->lti_ea_store_size;
1400                 rc = lod_parse_striping(env, lo, buf);
1401                 if (rc == 0)
1402                         lo->ldo_comp_cached = 1;
1403         } else if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1404                 rc = lod_get_lmv_ea(env, lo);
1405                 if (rc < (typeof(rc))sizeof(struct lmv_mds_md_v1)) {
1406                         /* Let's set stripe_loaded to avoid further
1407                          * stripe loading especially for non-stripe directory,
1408                          * which can hurt performance. (See LU-9840)
1409                          */
1410                         if (rc == 0)
1411                                 lo->ldo_dir_stripe_loaded = 1;
1412                         GOTO(out, rc = rc > 0 ? -EINVAL : rc);
1413                 }
1414                 buf->lb_buf = info->lti_ea_store;
1415                 buf->lb_len = info->lti_ea_store_size;
1416                 if (rc == sizeof(struct lmv_mds_md_v1)) {
1417                         rc = lod_load_lmv_shards(env, lo, buf, true);
1418                         if (buf->lb_buf != info->lti_ea_store) {
1419                                 OBD_FREE_LARGE(info->lti_ea_store,
1420                                                info->lti_ea_store_size);
1421                                 info->lti_ea_store = buf->lb_buf;
1422                                 info->lti_ea_store_size = buf->lb_len;
1423                         }
1424
1425                         if (rc < 0)
1426                                 GOTO(out, rc);
1427                 }
1428
1429                 /*
1430                  * there is LMV EA (striping information) in this object
1431                  * let's parse it and create in-core objects for the stripes
1432                  */
1433                 rc = lod_parse_dir_striping(env, lo, buf);
1434                 if (rc == 0)
1435                         lo->ldo_dir_stripe_loaded = 1;
1436         }
1437 out:
1438         RETURN(rc);
1439 }
1440
1441 /**
1442  * A generic function to initialize the stripe objects.
1443  *
1444  * A protected version of lod_load_striping_locked() - load the striping
1445  * information from storage, parse that and instantiate LU objects to
1446  * represent the stripes.  The LOD object \a lo supplies a pointer to the
1447  * next sub-object in the LU stack so we can lock it. Also use \a lo to
1448  * return an array of references to the newly instantiated objects.
1449  *
1450  * \param[in] env               execution environment for this thread
1451  * \param[in,out] lo            LOD object, where striping is stored and
1452  *                              which gets an array of references
1453  *
1454  * \retval                      0 if parsing and object creation succeed
1455  * \retval                      negative error number on failure
1456  **/
1457 int lod_load_striping(const struct lu_env *env, struct lod_object *lo)
1458 {
1459         struct dt_object        *next = dt_object_child(&lo->ldo_obj);
1460         int                     rc;
1461
1462         if (!dt_object_exists(next))
1463                 return 0;
1464
1465         /* Check without locking first */
1466         if (lod_striping_loaded(lo))
1467                 return 0;
1468
1469         /* currently this code is supposed to be called from declaration
1470          * phase only, thus the object is not expected to be locked by caller */
1471         dt_write_lock(env, next, 0);
1472         rc = lod_load_striping_locked(env, lo);
1473         dt_write_unlock(env, next);
1474         return rc;
1475 }
1476
1477 /**
1478  * Verify lov_user_md_v1/v3 striping.
1479  *
1480  * Check the validity of all fields including the magic, stripe size,
1481  * stripe count, stripe offset and that the pool is present.  Also check
1482  * that each target index points to an existing target. The additional
1483  * \a is_from_disk turns additional checks. In some cases zero fields
1484  * are allowed (like pattern=0).
1485  *
1486  * \param[in] d                 LOD device
1487  * \param[in] buf               buffer with LOV EA to verify
1488  * \param[in] is_from_disk      0 - from user, allow some fields to be 0
1489  *                              1 - from disk, do not allow
1490  *
1491  * \retval                      0 if the striping is valid
1492  * \retval                      -EINVAL if striping is invalid
1493  */
1494 static int lod_verify_v1v3(struct lod_device *d, const struct lu_buf *buf,
1495                            bool is_from_disk)
1496 {
1497         struct lov_user_md_v1   *lum;
1498         struct lov_user_md_v3   *lum3;
1499         struct pool_desc        *pool = NULL;
1500         __u32                    magic;
1501         __u32                    stripe_size;
1502         __u16                    stripe_count;
1503         __u16                    stripe_offset;
1504         size_t                   lum_size;
1505         int                      rc = 0;
1506         ENTRY;
1507
1508         lum = buf->lb_buf;
1509
1510         if (buf->lb_len < sizeof(*lum)) {
1511                 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
1512                        buf->lb_len);
1513                 GOTO(out, rc = -EINVAL);
1514         }
1515
1516         magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
1517         if (magic != LOV_USER_MAGIC_V1 &&
1518             magic != LOV_USER_MAGIC_V3 &&
1519             magic != LOV_USER_MAGIC_SPECIFIC) {
1520                 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
1521                        le32_to_cpu(lum->lmm_magic));
1522                 GOTO(out, rc = -EINVAL);
1523         }
1524
1525         /* the user uses "0" for default stripe pattern normally. */
1526         if (!is_from_disk && lum->lmm_pattern == LOV_PATTERN_NONE)
1527                 lum->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1528
1529         if (!lov_pattern_supported(le32_to_cpu(lum->lmm_pattern))) {
1530                 CDEBUG(D_LAYOUT, "bad userland stripe pattern: %#x\n",
1531                        le32_to_cpu(lum->lmm_pattern));
1532                 GOTO(out, rc = -EINVAL);
1533         }
1534
1535         /* a released lum comes from creating orphan on hsm release,
1536          * doesn't make sense to verify it. */
1537         if (le32_to_cpu(lum->lmm_pattern) & LOV_PATTERN_F_RELEASED)
1538                 GOTO(out, rc = 0);
1539
1540         /* 64kB is the largest common page size we see (ia64), and matches the
1541          * check in lfs */
1542         stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1543         if (stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
1544                 CDEBUG(D_LAYOUT, "stripe size %u not a multiple of %u\n",
1545                        stripe_size, LOV_MIN_STRIPE_SIZE);
1546                 GOTO(out, rc = -EINVAL);
1547         }
1548
1549         stripe_offset = le16_to_cpu(lum->lmm_stripe_offset);
1550         if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT &&
1551             lov_pattern(le32_to_cpu(lum->lmm_pattern)) != LOV_PATTERN_MDT) {
1552                 /* if offset is not within valid range [0, osts_size) */
1553                 if (stripe_offset >= d->lod_osts_size) {
1554                         CDEBUG(D_LAYOUT, "stripe offset %u >= bitmap size %u\n",
1555                                stripe_offset, d->lod_osts_size);
1556                         GOTO(out, rc = -EINVAL);
1557                 }
1558
1559                 /* if lmm_stripe_offset is *not* in bitmap */
1560                 if (!cfs_bitmap_check(d->lod_ost_bitmap, stripe_offset)) {
1561                         CDEBUG(D_LAYOUT, "stripe offset %u not in bitmap\n",
1562                                stripe_offset);
1563                         GOTO(out, rc = -EINVAL);
1564                 }
1565         }
1566
1567         if (magic == LOV_USER_MAGIC_V1)
1568                 lum_size = offsetof(struct lov_user_md_v1,
1569                                     lmm_objects[0]);
1570         else if (magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC)
1571                 lum_size = offsetof(struct lov_user_md_v3,
1572                                     lmm_objects[0]);
1573         else
1574                 GOTO(out, rc = -EINVAL);
1575
1576         stripe_count = le16_to_cpu(lum->lmm_stripe_count);
1577         if (buf->lb_len < lum_size) {
1578                 CDEBUG(D_LAYOUT, "invalid buf len %zu/%zu for lov_user_md with "
1579                        "magic %#x and stripe_count %u\n",
1580                        buf->lb_len, lum_size, magic, stripe_count);
1581                 GOTO(out, rc = -EINVAL);
1582         }
1583
1584         if (!(magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC))
1585                 goto out;
1586
1587         lum3 = buf->lb_buf;
1588         /* In the function below, .hs_keycmp resolves to
1589          * pool_hashkey_keycmp() */
1590         /* coverity[overrun-buffer-val] */
1591         pool = lod_find_pool(d, lum3->lmm_pool_name);
1592         if (pool == NULL)
1593                 goto out;
1594
1595         if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT) {
1596                 rc = lod_check_index_in_pool(stripe_offset, pool);
1597                 if (rc < 0)
1598                         GOTO(out, rc = -EINVAL);
1599         }
1600
1601         if (is_from_disk && stripe_count > pool_tgt_count(pool)) {
1602                 CDEBUG(D_LAYOUT, "stripe count %u > # OSTs %u in the pool\n",
1603                        stripe_count, pool_tgt_count(pool));
1604                 GOTO(out, rc = -EINVAL);
1605         }
1606
1607 out:
1608         if (pool != NULL)
1609                 lod_pool_putref(pool);
1610
1611         RETURN(rc);
1612 }
1613
1614 /**
1615  * Verify LOV striping.
1616  *
1617  * \param[in] d                 LOD device
1618  * \param[in] buf               buffer with LOV EA to verify
1619  * \param[in] is_from_disk      0 - from user, allow some fields to be 0
1620  *                              1 - from disk, do not allow
1621  * \param[in] start             extent start for composite layout
1622  *
1623  * \retval                      0 if the striping is valid
1624  * \retval                      -EINVAL if striping is invalid
1625  */
1626 int lod_verify_striping(struct lod_device *d, const struct lu_buf *buf,
1627                         bool is_from_disk, __u64 start)
1628 {
1629         struct lov_user_md_v1   *lum;
1630         struct lov_comp_md_v1   *comp_v1;
1631         __u32   magic;
1632         int     rc = 0, i;
1633         ENTRY;
1634
1635         lum = buf->lb_buf;
1636
1637         if (buf->lb_len < sizeof(*lum)) {
1638                 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
1639                        buf->lb_len);
1640                 RETURN(-EINVAL);
1641         }
1642
1643         magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
1644         if (magic != LOV_USER_MAGIC_V1 &&
1645             magic != LOV_USER_MAGIC_V3 &&
1646             magic != LOV_USER_MAGIC_SPECIFIC &&
1647             magic != LOV_USER_MAGIC_COMP_V1) {
1648                 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
1649                        le32_to_cpu(lum->lmm_magic));
1650                 RETURN(-EINVAL);
1651         }
1652
1653         if (magic == LOV_USER_MAGIC_COMP_V1) {
1654                 struct lov_comp_md_entry_v1     *ent;
1655                 struct lu_extent        *ext;
1656                 struct lov_desc *desc = &d->lod_desc;
1657                 struct lu_buf   tmp;
1658                 __u32   stripe_size = 0;
1659                 __u64   prev_end = start;
1660
1661                 comp_v1 = buf->lb_buf;
1662                 if (buf->lb_len < le32_to_cpu(comp_v1->lcm_size)) {
1663                         CDEBUG(D_LAYOUT, "buf len %zu is less than %u\n",
1664                                buf->lb_len, le32_to_cpu(comp_v1->lcm_size));
1665                         RETURN(-EINVAL);
1666                 }
1667
1668                 if (le16_to_cpu(comp_v1->lcm_entry_count) == 0) {
1669                         CDEBUG(D_LAYOUT, "entry count is zero\n");
1670                         RETURN(-EINVAL);
1671                 }
1672
1673                 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1674                         ent = &comp_v1->lcm_entries[i];
1675                         ext = &ent->lcme_extent;
1676
1677                         if (is_from_disk &&
1678                             (le32_to_cpu(ent->lcme_id) == 0 ||
1679                              le32_to_cpu(ent->lcme_id) > LCME_ID_MAX)) {
1680                                 CDEBUG(D_LAYOUT, "invalid id %u\n",
1681                                        le32_to_cpu(ent->lcme_id));
1682                                 RETURN(-EINVAL);
1683                         }
1684
1685                         if (le64_to_cpu(ext->e_start) >=
1686                             le64_to_cpu(ext->e_end)) {
1687                                 CDEBUG(D_LAYOUT, "invalid extent "
1688                                        "[%llu, %llu)\n",
1689                                        le64_to_cpu(ext->e_start),
1690                                        le64_to_cpu(ext->e_end));
1691                                 RETURN(-EINVAL);
1692                         }
1693
1694                         /* first component must start with 0, and the next
1695                          * must be adjacent with the previous one */
1696                         if (le64_to_cpu(ext->e_start) != prev_end) {
1697                                 CDEBUG(D_LAYOUT, "invalid start "
1698                                        "actual:%llu, expect:%llu\n",
1699                                        le64_to_cpu(ext->e_start), prev_end);
1700                                 RETURN(-EINVAL);
1701                         }
1702                         prev_end = le64_to_cpu(ext->e_end);
1703
1704                         tmp.lb_buf = (char *)comp_v1 +
1705                                      le32_to_cpu(ent->lcme_offset);
1706                         tmp.lb_len = le32_to_cpu(ent->lcme_size);
1707
1708                         /* Checks for DoM entry in composite layout. */
1709                         lum = tmp.lb_buf;
1710                         if (lov_pattern(le32_to_cpu(lum->lmm_pattern)) ==
1711                             LOV_PATTERN_MDT) {
1712                                 /* DoM component can be only the first entry */
1713                                 if (i > 0) {
1714                                         CDEBUG(D_LAYOUT, "invalid DoM layout "
1715                                                "entry found at %i index\n", i);
1716                                         RETURN(-EINVAL);
1717                                 }
1718                                 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1719                                 /* There is just one stripe on MDT and it must
1720                                  * cover whole component size. */
1721                                 if (stripe_size != prev_end) {
1722                                         CDEBUG(D_LAYOUT, "invalid DoM layout "
1723                                                "stripe size %u != %llu "
1724                                                "(component size)\n",
1725                                                stripe_size, prev_end);
1726                                         RETURN(-EINVAL);
1727                                 }
1728                                 /* Check stripe size againts per-MDT limit */
1729                                 if (stripe_size > d->lod_dom_max_stripesize) {
1730                                         CDEBUG(D_LAYOUT, "DoM component size "
1731                                                "%u is bigger than MDT limit "
1732                                                "%u, check dom_max_stripesize"
1733                                                " parameter\n",
1734                                                stripe_size,
1735                                                d->lod_dom_max_stripesize);
1736                                         RETURN(-EINVAL);
1737                                 }
1738                         }
1739                         rc = lod_verify_v1v3(d, &tmp, is_from_disk);
1740                         if (rc)
1741                                 break;
1742
1743                         lum = tmp.lb_buf;
1744
1745                         /* extent end must be aligned with the stripe_size */
1746                         stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1747                         if (stripe_size == 0)
1748                                 stripe_size = desc->ld_default_stripe_size;
1749                         if (stripe_size == 0 ||
1750                             (prev_end != LUSTRE_EOF &&
1751                              (prev_end & (stripe_size - 1)))) {
1752                                 CDEBUG(D_LAYOUT, "stripe size isn't aligned. "
1753                                        " stripe_sz: %u, [%llu, %llu)\n",
1754                                        stripe_size, ext->e_start, prev_end);
1755                                 RETURN(-EINVAL);
1756                         }
1757                 }
1758         } else {
1759                 rc = lod_verify_v1v3(d, buf, is_from_disk);
1760         }
1761
1762         RETURN(rc);
1763 }
1764
1765 /**
1766  * set the default stripe size, if unset.
1767  *
1768  * \param[in,out] val   number of bytes per OST stripe
1769  *
1770  * The minimum stripe size is 64KB to ensure that a single stripe is an
1771  * even multiple of a client PAGE_SIZE (IA64, PPC, etc).  Otherwise, it
1772  * is difficult to split dirty pages across OSCs during writes.
1773  */
1774 void lod_fix_desc_stripe_size(__u64 *val)
1775 {
1776         if (*val < LOV_MIN_STRIPE_SIZE) {
1777                 if (*val != 0)
1778                         LCONSOLE_INFO("Increasing default stripe size to "
1779                                       "minimum value %u\n",
1780                                       LOV_DESC_STRIPE_SIZE_DEFAULT);
1781                 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
1782         } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
1783                 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
1784                 LCONSOLE_WARN("Changing default stripe size to %llu (a "
1785                               "multiple of %u)\n",
1786                               *val, LOV_MIN_STRIPE_SIZE);
1787         }
1788 }
1789
1790 /**
1791  * set the filesystem default number of stripes, if unset.
1792  *
1793  * \param[in,out] val   number of stripes
1794  *
1795  * A value of "0" means "use the system-wide default stripe count", which
1796  * has either been inherited by now, or falls back to 1 stripe per file.
1797  * A value of "-1" (0xffffffff) means "stripe over all available OSTs",
1798  * and is a valid value, so is left unchanged here.
1799  */
1800 void lod_fix_desc_stripe_count(__u32 *val)
1801 {
1802         if (*val == 0)
1803                 *val = 1;
1804 }
1805
1806 /**
1807  * set the filesystem default layout pattern
1808  *
1809  * \param[in,out] val   LOV_PATTERN_* layout
1810  *
1811  * A value of "0" means "use the system-wide default layout type", which
1812  * has either been inherited by now, or falls back to plain RAID0 striping.
1813  */
1814 void lod_fix_desc_pattern(__u32 *val)
1815 {
1816         /* from lov_setstripe */
1817         if ((*val != 0) && (*val != LOV_PATTERN_RAID0) &&
1818             (*val != LOV_PATTERN_MDT)) {
1819                 LCONSOLE_WARN("Unknown stripe pattern: %#x\n", *val);
1820                 *val = 0;
1821         }
1822 }
1823
1824 void lod_fix_desc_qos_maxage(__u32 *val)
1825 {
1826         /* fix qos_maxage */
1827         if (*val == 0)
1828                 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
1829 }
1830
1831 /**
1832  * Used to fix insane default striping.
1833  *
1834  * \param[in] desc      striping description
1835  */
1836 void lod_fix_desc(struct lov_desc *desc)
1837 {
1838         lod_fix_desc_stripe_size(&desc->ld_default_stripe_size);
1839         lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
1840         lod_fix_desc_pattern(&desc->ld_pattern);
1841         lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
1842 }
1843
1844 /**
1845  * Initialize the structures used to store pools and default striping.
1846  *
1847  * \param[in] lod       LOD device
1848  * \param[in] lcfg      configuration structure storing default striping.
1849  *
1850  * \retval              0 if initialization succeeds
1851  * \retval              negative error number on failure
1852  */
1853 int lod_pools_init(struct lod_device *lod, struct lustre_cfg *lcfg)
1854 {
1855         struct obd_device          *obd;
1856         struct lov_desc            *desc;
1857         int                         rc;
1858         ENTRY;
1859
1860         obd = class_name2obd(lustre_cfg_string(lcfg, 0));
1861         LASSERT(obd != NULL);
1862         obd->obd_lu_dev = &lod->lod_dt_dev.dd_lu_dev;
1863
1864         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1865                 CERROR("LOD setup requires a descriptor\n");
1866                 RETURN(-EINVAL);
1867         }
1868
1869         desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
1870
1871         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1872                 CERROR("descriptor size wrong: %d > %d\n",
1873                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1874                 RETURN(-EINVAL);
1875         }
1876
1877         if (desc->ld_magic != LOV_DESC_MAGIC) {
1878                 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
1879                         CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
1880                                obd->obd_name, desc);
1881                         lustre_swab_lov_desc(desc);
1882                 } else {
1883                         CERROR("%s: Bad lov desc magic: %#x\n",
1884                                obd->obd_name, desc->ld_magic);
1885                         RETURN(-EINVAL);
1886                 }
1887         }
1888
1889         lod_fix_desc(desc);
1890
1891         desc->ld_active_tgt_count = 0;
1892         lod->lod_desc = *desc;
1893
1894         lod->lod_sp_me = LUSTRE_SP_CLI;
1895
1896         /* Set up allocation policy (QoS and RR) */
1897         INIT_LIST_HEAD(&lod->lod_qos.lq_oss_list);
1898         init_rwsem(&lod->lod_qos.lq_rw_sem);
1899         lod->lod_qos.lq_dirty = 1;
1900         lod->lod_qos.lq_rr.lqr_dirty = 1;
1901         lod->lod_qos.lq_reset = 1;
1902         /* Default priority is toward free space balance */
1903         lod->lod_qos.lq_prio_free = 232;
1904         /* Default threshold for rr (roughly 17%) */
1905         lod->lod_qos.lq_threshold_rr = 43;
1906
1907         /* Set up OST pool environment */
1908         lod->lod_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,
1909                                                    HASH_POOLS_MAX_BITS,
1910                                                    HASH_POOLS_BKT_BITS, 0,
1911                                                    CFS_HASH_MIN_THETA,
1912                                                    CFS_HASH_MAX_THETA,
1913                                                    &pool_hash_operations,
1914                                                    CFS_HASH_DEFAULT);
1915         if (lod->lod_pools_hash_body == NULL)
1916                 RETURN(-ENOMEM);
1917
1918         INIT_LIST_HEAD(&lod->lod_pool_list);
1919         lod->lod_pool_count = 0;
1920         rc = lod_ost_pool_init(&lod->lod_pool_info, 0);
1921         if (rc)
1922                 GOTO(out_hash, rc);
1923         lod_qos_rr_init(&lod->lod_qos.lq_rr);
1924         rc = lod_ost_pool_init(&lod->lod_qos.lq_rr.lqr_pool, 0);
1925         if (rc)
1926                 GOTO(out_pool_info, rc);
1927
1928         RETURN(0);
1929
1930 out_pool_info:
1931         lod_ost_pool_free(&lod->lod_pool_info);
1932 out_hash:
1933         cfs_hash_putref(lod->lod_pools_hash_body);
1934
1935         return rc;
1936 }
1937
1938 /**
1939  * Release the structures describing the pools.
1940  *
1941  * \param[in] lod       LOD device from which we release the structures
1942  *
1943  * \retval              0 always
1944  */
1945 int lod_pools_fini(struct lod_device *lod)
1946 {
1947         struct obd_device   *obd = lod2obd(lod);
1948         struct pool_desc    *pool, *tmp;
1949         ENTRY;
1950
1951         list_for_each_entry_safe(pool, tmp, &lod->lod_pool_list, pool_list) {
1952                 /* free pool structs */
1953                 CDEBUG(D_INFO, "delete pool %p\n", pool);
1954                 /* In the function below, .hs_keycmp resolves to
1955                  * pool_hashkey_keycmp() */
1956                 /* coverity[overrun-buffer-val] */
1957                 lod_pool_del(obd, pool->pool_name);
1958         }
1959
1960         cfs_hash_putref(lod->lod_pools_hash_body);
1961         lod_ost_pool_free(&(lod->lod_qos.lq_rr.lqr_pool));
1962         lod_ost_pool_free(&lod->lod_pool_info);
1963
1964         RETURN(0);
1965 }