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                 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         if (lo->ldo_mirrors) {
 698                 OBD_FREE(lo->ldo_mirrors,
 699                          sizeof(*lo->ldo_mirrors) * lo->ldo_mirror_count);
 700                 lo->ldo_mirrors = NULL;
 701                 lo->ldo_mirror_count = 0;
 702         }
 703         lod_free_comp_buffer(lo->ldo_comp_entries,
 704                              lo->ldo_comp_cnt,
 705                              sizeof(*lo->ldo_comp_entries) * lo->ldo_comp_cnt);
 706         lo->ldo_comp_entries = NULL;
 707         lo->ldo_comp_cnt = 0;
 708         lo->ldo_is_composite = 0;
 709 }
 710
 711 int lod_alloc_comp_entries(struct lod_object *lo,
 712                            int mirror_count, int comp_count)
 713 {
 714         LASSERT(comp_count != 0);
 715         LASSERT(lo->ldo_comp_cnt == 0 && lo->ldo_comp_entries == NULL);
 716
 717         if (mirror_count > 0) {
 718                 OBD_ALLOC(lo->ldo_mirrors,
 719                           sizeof(*lo->ldo_mirrors) * mirror_count);
 720                 if (!lo->ldo_mirrors)
 721                         return -ENOMEM;
 722
 723                 lo->ldo_mirror_count = mirror_count;
 724         }
 725
 726         OBD_ALLOC_LARGE(lo->ldo_comp_entries,
 727                         sizeof(*lo->ldo_comp_entries) * comp_count);
 728         if (lo->ldo_comp_entries == NULL) {
 729                 OBD_FREE(lo->ldo_mirrors,
 730                          sizeof(*lo->ldo_mirrors) * mirror_count);
 731                 lo->ldo_mirror_count = 0;
 732                 return -ENOMEM;
 733         }
 734
 735         lo->ldo_comp_cnt = comp_count;
 736         return 0;
 737 }
 738
 739 int lod_fill_mirrors(struct lod_object *lo)
 740 {
 741         struct lod_layout_component *lod_comp;
 742         int mirror_idx = -1;
 743         __u16 mirror_id = 0xffff;
 744         int i;
 745         ENTRY;
 746
 747         LASSERT(equi(!lo->ldo_is_composite, lo->ldo_mirror_count == 0));
 748
 749         if (!lo->ldo_is_composite)
 750                 RETURN(0);
 751
 752         lod_comp = &lo->ldo_comp_entries[0];
 753         for (i = 0; i < lo->ldo_comp_cnt; i++, lod_comp++) {
 754                 int stale = !!(lod_comp->llc_flags & LCME_FL_STALE);
 755                 int preferred = !!(lod_comp->llc_flags & LCME_FL_PREF_WR);
 756
 757                 if (mirror_id_of(lod_comp->llc_id) == mirror_id) {
 758                         lo->ldo_mirrors[mirror_idx].lme_stale |= stale;
 759                         lo->ldo_mirrors[mirror_idx].lme_primary |= preferred;
 760                         lo->ldo_mirrors[mirror_idx].lme_end = i;
 761                         continue;
 762                 }
 763
 764                 /* new mirror */
 765                 ++mirror_idx;
 766                 if (mirror_idx >= lo->ldo_mirror_count)
 767                         RETURN(-EINVAL);
 768
 769                 mirror_id = mirror_id_of(lod_comp->llc_id);
 770
 771                 lo->ldo_mirrors[mirror_idx].lme_id = mirror_id;
 772                 lo->ldo_mirrors[mirror_idx].lme_stale = stale;
 773                 lo->ldo_mirrors[mirror_idx].lme_primary = preferred;
 774                 lo->ldo_mirrors[mirror_idx].lme_start = i;
 775                 lo->ldo_mirrors[mirror_idx].lme_end = i;
 776         }
 777         if (mirror_idx != lo->ldo_mirror_count - 1)
 778                 RETURN(-EINVAL);
 779
 780         RETURN(0);
 781 }
 782
 783 /**
 784  * Generate on-disk lov_mds_md structure for each layout component based on
 785  * the information in lod_object->ldo_comp_entries[i].
 786  *
 787  * \param[in] env               execution environment for this thread
 788  * \param[in] lo                LOD object
 789  * \param[in] comp_idx          index of ldo_comp_entries
 790  * \param[in] lmm               buffer to cotain the on-disk lov_mds_md
 791  * \param[in|out] lmm_size      buffer size/lmm size
 792  * \param[in] is_dir            generate lov ea for dir or file? For dir case,
 793  *                              the stripe info is from the default stripe
 794  *                              template, which is collected in lod_ah_init(),
 795  *                              either from parent object or root object; for
 796  *                              file case, it's from the @lo object
 797  *
 798  * \retval                      0 if on disk structure is created successfully
 799  * \retval                      negative error number on failure
 800  */
 801 static int lod_gen_component_ea(const struct lu_env *env,
 802                                 struct lod_object *lo, int comp_idx,
 803                                 struct lov_mds_md *lmm, int *lmm_size,
 804                                 bool is_dir)
 805 {
 806         struct lod_thread_info  *info = lod_env_info(env);
 807         const struct lu_fid     *fid  = lu_object_fid(&lo->ldo_obj.do_lu);
 808         struct lod_device       *lod;
 809         struct lov_ost_data_v1  *objs;
 810         struct lod_layout_component *lod_comp;
 811         __u32   magic;
 812         __u16 stripe_count;
 813         int     i, rc = 0;
 814         ENTRY;
 815
 816         LASSERT(lo);
 817         if (is_dir)
 818                 lod_comp =
 819                         &lo->ldo_def_striping->lds_def_comp_entries[comp_idx];
 820         else
 821                 lod_comp = &lo->ldo_comp_entries[comp_idx];
 822
 823         magic = lod_comp->llc_pool != NULL ? LOV_MAGIC_V3 : LOV_MAGIC_V1;
 824         if (lod_comp->llc_pattern == 0) /* default striping */
 825                 lod_comp->llc_pattern = LOV_PATTERN_RAID0;
 826
 827         lmm->lmm_magic = cpu_to_le32(magic);
 828         lmm->lmm_pattern = cpu_to_le32(lod_comp->llc_pattern);
 829         fid_to_lmm_oi(fid, &lmm->lmm_oi);
 830         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_LMMOI))
 831                 lmm->lmm_oi.oi.oi_id++;
 832         lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
 833
 834         lmm->lmm_stripe_size = cpu_to_le32(lod_comp->llc_stripe_size);
 835         lmm->lmm_stripe_count = cpu_to_le16(lod_comp->llc_stripe_count);
 836         /**
 837          * for dir and uninstantiated component, lmm_layout_gen stores
 838          * default stripe offset.
 839          */
 840         lmm->lmm_layout_gen =
 841                 (is_dir || !lod_comp_inited(lod_comp)) ?
 842                         cpu_to_le16(lod_comp->llc_stripe_offset) :
 843                         cpu_to_le16(lod_comp->llc_layout_gen);
 844
 845         if (magic == LOV_MAGIC_V1) {
 846                 objs = &lmm->lmm_objects[0];
 847         } else {
 848                 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
 849                 size_t cplen = strlcpy(v3->lmm_pool_name,
 850                                        lod_comp->llc_pool,
 851                                        sizeof(v3->lmm_pool_name));
 852                 if (cplen >= sizeof(v3->lmm_pool_name))
 853                         RETURN(-E2BIG);
 854                 objs = &v3->lmm_objects[0];
 855         }
 856         stripe_count = lod_comp_entry_stripe_count(lo, lod_comp, is_dir);
 857         if (!is_dir && lo->ldo_is_composite)
 858                 lod_comp_shrink_stripe_count(lod_comp, &stripe_count);
 859
 860         if (is_dir || lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)
 861                 GOTO(done, rc = 0);
 862
 863         /* generate ost_idx of this component stripe */
 864         lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
 865         for (i = 0; i < stripe_count; i++) {
 866                 struct dt_object *object;
 867                 __u32 ost_idx = (__u32)-1UL;
 868                 int type = LU_SEQ_RANGE_OST;
 869
 870                 if (lod_comp->llc_stripe && lod_comp->llc_stripe[i]) {
 871                         object = lod_comp->llc_stripe[i];
 872                         /* instantiated component */
 873                         info->lti_fid = *lu_object_fid(&object->do_lu);
 874
 875                         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_MULTIPLE_REF) &&
 876                             comp_idx == 0) {
 877                                 if (cfs_fail_val == 0)
 878                                         cfs_fail_val = info->lti_fid.f_oid;
 879                                 else if (i == 0)
 880                                         info->lti_fid.f_oid = cfs_fail_val;
 881                         }
 882
 883                         rc = fid_to_ostid(&info->lti_fid, &info->lti_ostid);
 884                         LASSERT(rc == 0);
 885
 886                         ostid_cpu_to_le(&info->lti_ostid, &objs[i].l_ost_oi);
 887                         objs[i].l_ost_gen = cpu_to_le32(0);
 888                         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FLD_LOOKUP))
 889                                 rc = -ENOENT;
 890                         else
 891                                 rc = lod_fld_lookup(env, lod, &info->lti_fid,
 892                                                     &ost_idx, &type);
 893                         if (rc < 0) {
 894                                 CERROR("%s: Can not locate "DFID": rc = %d\n",
 895                                        lod2obd(lod)->obd_name,
 896                                        PFID(&info->lti_fid), rc);
 897                                 RETURN(rc);
 898                         }
 899                 } else if (lod_comp->llc_ostlist.op_array) {
 900                         /* user specified ost list */
 901                         ost_idx = lod_comp->llc_ostlist.op_array[i];
 902                 }
 903                 /*
 904                  * with un-instantiated or with no specified ost list
 905                  * component, its l_ost_idx does not matter.
 906                  */
 907                 objs[i].l_ost_idx = cpu_to_le32(ost_idx);
 908         }
 909 done:
 910         if (lmm_size != NULL)
 911                 *lmm_size = lov_mds_md_size(stripe_count, magic);
 912         RETURN(rc);
 913 }
 914
 915 /**
 916  * Generate on-disk lov_mds_md structure based on the information in
 917  * the lod_object->ldo_comp_entries.
 918  *
 919  * \param[in] env               execution environment for this thread
 920  * \param[in] lo                LOD object
 921  * \param[in] lmm               buffer to cotain the on-disk lov_mds_md
 922  * \param[in|out] lmm_size      buffer size/lmm size
 923  * \param[in] is_dir            generate lov ea for dir or file? For dir case,
 924  *                              the stripe info is from the default stripe
 925  *                              template, which is collected in lod_ah_init(),
 926  *                              either from parent object or root object; for
 927  *                              file case, it's from the @lo object
 928  *
 929  * \retval                      0 if on disk structure is created successfully
 930  * \retval                      negative error number on failure
 931  */
 932 int lod_generate_lovea(const struct lu_env *env, struct lod_object *lo,
 933                        struct lov_mds_md *lmm, int *lmm_size, bool is_dir)
 934 {
 935         struct lov_comp_md_entry_v1 *lcme;
 936         struct lov_comp_md_v1 *lcm;
 937         struct lod_layout_component *comp_entries;
 938         __u16 comp_cnt, mirror_cnt;
 939         bool is_composite;
 940         int i, rc = 0, offset;
 941         ENTRY;
 942
 943         if (is_dir) {
 944                 comp_cnt = lo->ldo_def_striping->lds_def_comp_cnt;
 945                 mirror_cnt = lo->ldo_def_striping->lds_def_mirror_cnt;
 946                 comp_entries = lo->ldo_def_striping->lds_def_comp_entries;
 947                 is_composite =
 948                         lo->ldo_def_striping->lds_def_striping_is_composite;
 949         } else {
 950                 comp_cnt = lo->ldo_comp_cnt;
 951                 mirror_cnt = lo->ldo_mirror_count;
 952                 comp_entries = lo->ldo_comp_entries;
 953                 is_composite = lo->ldo_is_composite;
 954         }
 955
 956         LASSERT(lmm_size != NULL);
 957         LASSERT(comp_cnt != 0 && comp_entries != NULL);
 958
 959         if (!is_composite) {
 960                 rc = lod_gen_component_ea(env, lo, 0, lmm, lmm_size, is_dir);
 961                 RETURN(rc);
 962         }
 963
 964         lcm = (struct lov_comp_md_v1 *)lmm;
 965         lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
 966         lcm->lcm_entry_count = cpu_to_le16(comp_cnt);
 967         lcm->lcm_mirror_count = cpu_to_le16(mirror_cnt - 1);
 968         lcm->lcm_flags = cpu_to_le16(lo->ldo_flr_state);
 969
 970         offset = sizeof(*lcm) + sizeof(*lcme) * comp_cnt;
 971         LASSERT(offset % sizeof(__u64) == 0);
 972
 973         for (i = 0; i < comp_cnt; i++) {
 974                 struct lod_layout_component *lod_comp;
 975                 struct lov_mds_md *sub_md;
 976                 int size;
 977
 978                 lod_comp = &comp_entries[i];
 979                 lcme = &lcm->lcm_entries[i];
 980
 981                 LASSERT(ergo(!is_dir, lod_comp->llc_id != LCME_ID_INVAL));
 982                 lcme->lcme_id = cpu_to_le32(lod_comp->llc_id);
 983
 984                 /* component could be un-inistantiated */
 985                 lcme->lcme_flags = cpu_to_le32(lod_comp->llc_flags);
 986                 lcme->lcme_extent.e_start =
 987                         cpu_to_le64(lod_comp->llc_extent.e_start);
 988                 lcme->lcme_extent.e_end =
 989                         cpu_to_le64(lod_comp->llc_extent.e_end);
 990                 lcme->lcme_offset = cpu_to_le32(offset);
 991
 992                 sub_md = (struct lov_mds_md *)((char *)lcm + offset);
 993                 rc = lod_gen_component_ea(env, lo, i, sub_md, &size, is_dir);
 994                 if (rc)
 995                         GOTO(out, rc);
 996                 lcme->lcme_size = cpu_to_le32(size);
 997                 offset += size;
 998                 LASSERTF((offset <= *lmm_size) && (offset % sizeof(__u64) == 0),
 999                          "offset:%d lmm_size:%d\n", offset, *lmm_size);
1000         }
1001         lcm->lcm_size = cpu_to_le32(offset);
1002         lcm->lcm_layout_gen = cpu_to_le32(is_dir ? 0 : lo->ldo_layout_gen);
1003
1004         lustre_print_user_md(D_LAYOUT, (struct lov_user_md *)lmm,
1005                              "generate lum");
1006 out:
1007         if (rc == 0)
1008                 *lmm_size = offset;
1009         RETURN(rc);
1010 }
1011
1012 /**
1013  * Get LOV EA.
1014  *
1015  * Fill lti_ea_store buffer in the environment with a value for the given
1016  * EA. The buffer is reallocated if the value doesn't fit.
1017  *
1018  * \param[in,out] env           execution environment for this thread
1019  *                              .lti_ea_store buffer is filled with EA's value
1020  * \param[in] lo                LOD object
1021  * \param[in] name              name of the EA
1022  *
1023  * \retval                      > 0 if EA is fetched successfully
1024  * \retval                      0 if EA is empty
1025  * \retval                      negative error number on failure
1026  */
1027 int lod_get_ea(const struct lu_env *env, struct lod_object *lo,
1028                const char *name)
1029 {
1030         struct lod_thread_info  *info = lod_env_info(env);
1031         struct dt_object        *next = dt_object_child(&lo->ldo_obj);
1032         int                     rc;
1033         ENTRY;
1034
1035         LASSERT(info);
1036
1037         if (unlikely(info->lti_ea_store == NULL)) {
1038                 /* just to enter in allocation block below */
1039                 rc = -ERANGE;
1040         } else {
1041 repeat:
1042                 info->lti_buf.lb_buf = info->lti_ea_store;
1043                 info->lti_buf.lb_len = info->lti_ea_store_size;
1044                 rc = dt_xattr_get(env, next, &info->lti_buf, name);
1045         }
1046
1047         /* if object is not striped or inaccessible */
1048         if (rc == -ENODATA || rc == -ENOENT)
1049                 RETURN(0);
1050
1051         if (rc == -ERANGE) {
1052                 /* EA doesn't fit, reallocate new buffer */
1053                 rc = dt_xattr_get(env, next, &LU_BUF_NULL, name);
1054                 if (rc == -ENODATA || rc == -ENOENT)
1055                         RETURN(0);
1056                 else if (rc < 0)
1057                         RETURN(rc);
1058
1059                 LASSERT(rc > 0);
1060                 rc = lod_ea_store_resize(info, rc);
1061                 if (rc)
1062                         RETURN(rc);
1063                 goto repeat;
1064         }
1065
1066         RETURN(rc);
1067 }
1068
1069 /**
1070  * Verify the target index is present in the current configuration.
1071  *
1072  * \param[in] md                LOD device where the target table is stored
1073  * \param[in] idx               target's index
1074  *
1075  * \retval                      0 if the index is present
1076  * \retval                      -EINVAL if not
1077  */
1078 static int validate_lod_and_idx(struct lod_device *md, __u32 idx)
1079 {
1080         if (unlikely(idx >= md->lod_ost_descs.ltd_tgts_size ||
1081                      !cfs_bitmap_check(md->lod_ost_bitmap, idx))) {
1082                 CERROR("%s: bad idx: %d of %d\n", lod2obd(md)->obd_name, idx,
1083                        md->lod_ost_descs.ltd_tgts_size);
1084                 return -EINVAL;
1085         }
1086
1087         if (unlikely(OST_TGT(md, idx) == NULL)) {
1088                 CERROR("%s: bad lod_tgt_desc for idx: %d\n",
1089                        lod2obd(md)->obd_name, idx);
1090                 return -EINVAL;
1091         }
1092
1093         if (unlikely(OST_TGT(md, idx)->ltd_ost == NULL)) {
1094                 CERROR("%s: invalid lod device, for idx: %d\n",
1095                        lod2obd(md)->obd_name , idx);
1096                 return -EINVAL;
1097         }
1098
1099         return 0;
1100 }
1101
1102 /**
1103  * Instantiate objects for stripes.
1104  *
1105  * Allocate and initialize LU-objects representing the stripes. The number
1106  * of the stripes (ldo_stripe_count) must be initialized already. The caller
1107  * must ensure nobody else is calling the function on the object at the same
1108  * time. FLDB service must be running to be able to map a FID to the targets
1109  * and find appropriate device representing that target.
1110  *
1111  * \param[in] env               execution environment for this thread
1112  * \param[in,out] lo            LOD object
1113  * \param[in] objs              an array of IDs to creates the objects from
1114  * \param[in] comp_idx          index of ldo_comp_entries
1115  *
1116  * \retval                      0 if the objects are instantiated successfully
1117  * \retval                      negative error number on failure
1118  */
1119 int lod_initialize_objects(const struct lu_env *env, struct lod_object *lo,
1120                            struct lov_ost_data_v1 *objs, int comp_idx)
1121 {
1122         struct lod_layout_component     *lod_comp;
1123         struct lod_thread_info  *info = lod_env_info(env);
1124         struct lod_device       *md;
1125         struct lu_object        *o, *n;
1126         struct lu_device        *nd;
1127         struct dt_object       **stripe;
1128         int                      stripe_len;
1129         int                      i, rc = 0;
1130         __u32                   idx;
1131         ENTRY;
1132
1133         LASSERT(lo != NULL);
1134         md = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1135
1136         LASSERT(lo->ldo_comp_cnt != 0 && lo->ldo_comp_entries != NULL);
1137         lod_comp = &lo->ldo_comp_entries[comp_idx];
1138
1139         LASSERT(lod_comp->llc_stripe == NULL);
1140         LASSERT(lod_comp->llc_stripe_count > 0);
1141         LASSERT(lod_comp->llc_stripe_size > 0);
1142
1143         stripe_len = lod_comp->llc_stripe_count;
1144         OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
1145         if (stripe == NULL)
1146                 RETURN(-ENOMEM);
1147
1148         for (i = 0; i < lod_comp->llc_stripe_count; i++) {
1149                 if (unlikely(lovea_slot_is_dummy(&objs[i])))
1150                         continue;
1151
1152                 ostid_le_to_cpu(&objs[i].l_ost_oi, &info->lti_ostid);
1153                 idx = le32_to_cpu(objs[i].l_ost_idx);
1154                 rc = ostid_to_fid(&info->lti_fid, &info->lti_ostid, idx);
1155                 if (rc != 0)
1156                         GOTO(out, rc);
1157                 LASSERTF(fid_is_sane(&info->lti_fid), ""DFID" insane!\n",
1158                          PFID(&info->lti_fid));
1159                 lod_getref(&md->lod_ost_descs);
1160
1161                 rc = validate_lod_and_idx(md, idx);
1162                 if (unlikely(rc != 0)) {
1163                         lod_putref(md, &md->lod_ost_descs);
1164                         GOTO(out, rc);
1165                 }
1166
1167                 nd = &OST_TGT(md,idx)->ltd_ost->dd_lu_dev;
1168                 lod_putref(md, &md->lod_ost_descs);
1169
1170                 /* In the function below, .hs_keycmp resolves to
1171                  * u_obj_hop_keycmp() */
1172                 /* coverity[overrun-buffer-val] */
1173                 o = lu_object_find_at(env, nd, &info->lti_fid, NULL);
1174                 if (IS_ERR(o))
1175                         GOTO(out, rc = PTR_ERR(o));
1176
1177                 n = lu_object_locate(o->lo_header, nd->ld_type);
1178                 LASSERT(n);
1179
1180                 stripe[i] = container_of(n, struct dt_object, do_lu);
1181         }
1182
1183 out:
1184         if (rc != 0) {
1185                 for (i = 0; i < stripe_len; i++)
1186                         if (stripe[i] != NULL)
1187                                 dt_object_put(env, stripe[i]);
1188
1189                 OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
1190                 lod_comp->llc_stripe_count = 0;
1191         } else {
1192                 lod_comp->llc_stripe = stripe;
1193                 lod_comp->llc_stripes_allocated = stripe_len;
1194         }
1195
1196         RETURN(rc);
1197 }
1198
1199 /**
1200  * Instantiate objects for striping.
1201  *
1202  * Parse striping information in \a buf and instantiate the objects
1203  * representing the stripes.
1204  *
1205  * \param[in] env               execution environment for this thread
1206  * \param[in] lo                LOD object
1207  * \param[in] buf               buffer storing LOV EA to parse
1208  *
1209  * \retval                      0 if parsing and objects creation succeed
1210  * \retval                      negative error number on failure
1211  */
1212 int lod_parse_striping(const struct lu_env *env, struct lod_object *lo,
1213                        const struct lu_buf *buf)
1214 {
1215         struct lov_mds_md_v1    *lmm;
1216         struct lov_comp_md_v1   *comp_v1 = NULL;
1217         struct lov_ost_data_v1  *objs;
1218         __u32   magic, pattern;
1219         int     i, j, rc = 0;
1220         __u16   comp_cnt;
1221         __u16   mirror_cnt = 0;
1222         ENTRY;
1223
1224         LASSERT(buf);
1225         LASSERT(buf->lb_buf);
1226         LASSERT(buf->lb_len);
1227
1228         lmm = (struct lov_mds_md_v1 *)buf->lb_buf;
1229         magic = le32_to_cpu(lmm->lmm_magic);
1230
1231         if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 &&
1232             magic != LOV_MAGIC_COMP_V1)
1233                 GOTO(out, rc = -EINVAL);
1234
1235         lod_free_comp_entries(lo);
1236
1237         if (magic == LOV_MAGIC_COMP_V1) {
1238                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1239                 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1240                 if (comp_cnt == 0)
1241                         GOTO(out, rc = -EINVAL);
1242                 lo->ldo_layout_gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1243                 lo->ldo_is_composite = 1;
1244                 lo->ldo_flr_state = le16_to_cpu(comp_v1->lcm_flags) &
1245                                         LCM_FL_FLR_MASK;
1246                 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1247         } else {
1248                 comp_cnt = 1;
1249                 lo->ldo_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1250                 lo->ldo_is_composite = 0;
1251         }
1252
1253         rc = lod_alloc_comp_entries(lo, mirror_cnt, comp_cnt);
1254         if (rc)
1255                 GOTO(out, rc);
1256
1257         for (i = 0; i < comp_cnt; i++) {
1258                 struct lod_layout_component     *lod_comp;
1259                 struct lu_extent        *ext;
1260                 __u32   offs;
1261
1262                 lod_comp = &lo->ldo_comp_entries[i];
1263                 if (lo->ldo_is_composite) {
1264                         offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1265                         lmm = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs);
1266                         magic = le32_to_cpu(lmm->lmm_magic);
1267
1268                         ext = &comp_v1->lcm_entries[i].lcme_extent;
1269                         lod_comp->llc_extent.e_start =
1270                                 le64_to_cpu(ext->e_start);
1271                         lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end);
1272                         lod_comp->llc_flags =
1273                                 le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags);
1274                         lod_comp->llc_id =
1275                                 le32_to_cpu(comp_v1->lcm_entries[i].lcme_id);
1276                         if (lod_comp->llc_id == LCME_ID_INVAL)
1277                                 GOTO(out, rc = -EINVAL);
1278                 } else {
1279                         lod_comp_set_init(lod_comp);
1280                 }
1281
1282                 pattern = le32_to_cpu(lmm->lmm_pattern);
1283                 if (lov_pattern(pattern) != LOV_PATTERN_RAID0 &&
1284                     lov_pattern(pattern) != LOV_PATTERN_MDT)
1285                         GOTO(out, rc = -EINVAL);
1286
1287                 lod_comp->llc_pattern = pattern;
1288                 lod_comp->llc_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
1289                 lod_comp->llc_stripe_count = le16_to_cpu(lmm->lmm_stripe_count);
1290                 lod_comp->llc_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1291
1292                 if (magic == LOV_MAGIC_V3) {
1293                         struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
1294                         lod_set_pool(&lod_comp->llc_pool, v3->lmm_pool_name);
1295                         objs = &v3->lmm_objects[0];
1296                 } else {
1297                         lod_set_pool(&lod_comp->llc_pool, NULL);
1298                         objs = &lmm->lmm_objects[0];
1299                 }
1300
1301                 /**
1302                  * If uninstantiated template component has valid l_ost_idx,
1303                  * then user has specified ost list for this component.
1304                  */
1305                 if (!lod_comp_inited(lod_comp)) {
1306                         if (objs[0].l_ost_idx != (__u32)-1UL) {
1307                                 /**
1308                                  * load the user specified ost list, when this
1309                                  * component is instantiated later, it will be
1310                                  * used in lod_alloc_ost_list().
1311                                  */
1312                                 lod_comp->llc_ostlist.op_count =
1313                                         lod_comp->llc_stripe_count;
1314                                 lod_comp->llc_ostlist.op_size =
1315                                         lod_comp->llc_stripe_count *
1316                                         sizeof(__u32);
1317                                 OBD_ALLOC(lod_comp->llc_ostlist.op_array,
1318                                           lod_comp->llc_ostlist.op_size);
1319                                 if (!lod_comp->llc_ostlist.op_array)
1320                                         GOTO(out, rc = -ENOMEM);
1321
1322                                 for (j = 0; j < lod_comp->llc_stripe_count; j++)
1323                                         lod_comp->llc_ostlist.op_array[j] =
1324                                                 le32_to_cpu(objs[j].l_ost_idx);
1325
1326                                 /**
1327                                  * this component OST objects starts from the
1328                                  * first ost_idx, lod_alloc_ost_list() will
1329                                  * check this.
1330                                  */
1331                                 lod_comp->llc_stripe_offset = objs[0].l_ost_idx;
1332                         } else {
1333                                 /**
1334                                  * for uninstantiated component,
1335                                  * lmm_layout_gen stores default stripe offset.
1336                                  */
1337                                 lod_comp->llc_stripe_offset =
1338                                                         lmm->lmm_layout_gen;
1339                         }
1340                 }
1341
1342                 /* skip un-instantiated component object initialization */
1343                 if (!lod_comp_inited(lod_comp))
1344                         continue;
1345
1346                 if (!(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1347                     !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) {
1348                         rc = lod_initialize_objects(env, lo, objs, i);
1349                         if (rc)
1350                                 GOTO(out, rc);
1351                 }
1352         }
1353
1354         rc = lod_fill_mirrors(lo);
1355         if (rc)
1356                 GOTO(out, rc);
1357
1358 out:
1359         if (rc)
1360                 lod_object_free_striping(env, lo);
1361         RETURN(rc);
1362 }
1363
1364 /**
1365  * Check whether the striping (LOVEA for regular file, LMVEA for directory)
1366  * is already cached.
1367  *
1368  * \param[in] lo        LOD object
1369  *
1370  * \retval              True if the striping is cached, otherwise
1371  *                      return false.
1372  */
1373 static bool lod_striping_loaded(struct lod_object *lo)
1374 {
1375         if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr) &&
1376             lo->ldo_comp_cached)
1377                 return true;
1378
1379         if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1380                 if (lo->ldo_stripe != NULL || lo->ldo_dir_stripe_loaded)
1381                         return true;
1382
1383                 /* Never load LMV stripe for slaves of striped dir */
1384                 if (lo->ldo_dir_slave_stripe)
1385                         return true;
1386         }
1387
1388         return false;
1389 }
1390
1391 /**
1392  * Initialize the object representing the stripes.
1393  *
1394  * Unless the stripes are initialized already, fetch LOV (for regular
1395  * objects) or LMV (for directory objects) EA and call lod_parse_striping()
1396  * to instantiate the objects representing the stripes. Caller should
1397  * hold the dt_write_lock(next).
1398  *
1399  * \param[in] env               execution environment for this thread
1400  * \param[in,out] lo            LOD object
1401  *
1402  * \retval                      0 if parsing and object creation succeed
1403  * \retval                      negative error number on failure
1404  */
1405 int lod_load_striping_locked(const struct lu_env *env, struct lod_object *lo)
1406 {
1407         struct lod_thread_info  *info = lod_env_info(env);
1408         struct lu_buf           *buf  = &info->lti_buf;
1409         struct dt_object        *next = dt_object_child(&lo->ldo_obj);
1410         int                      rc = 0;
1411         ENTRY;
1412
1413         if (!dt_object_exists(next))
1414                 GOTO(out, rc = 0);
1415
1416         if (lod_striping_loaded(lo))
1417                 GOTO(out, rc = 0);
1418
1419         if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr)) {
1420                 rc = lod_get_lov_ea(env, lo);
1421                 if (rc <= 0)
1422                         GOTO(out, rc);
1423                 /*
1424                  * there is LOV EA (striping information) in this object
1425                  * let's parse it and create in-core objects for the stripes
1426                  */
1427                 buf->lb_buf = info->lti_ea_store;
1428                 buf->lb_len = info->lti_ea_store_size;
1429                 rc = lod_parse_striping(env, lo, buf);
1430                 if (rc == 0)
1431                         lo->ldo_comp_cached = 1;
1432         } else if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1433                 rc = lod_get_lmv_ea(env, lo);
1434                 if (rc < (typeof(rc))sizeof(struct lmv_mds_md_v1)) {
1435                         /* Let's set stripe_loaded to avoid further
1436                          * stripe loading especially for non-stripe directory,
1437                          * which can hurt performance. (See LU-9840)
1438                          */
1439                         if (rc == 0)
1440                                 lo->ldo_dir_stripe_loaded = 1;
1441                         GOTO(out, rc = rc > 0 ? -EINVAL : rc);
1442                 }
1443                 buf->lb_buf = info->lti_ea_store;
1444                 buf->lb_len = info->lti_ea_store_size;
1445                 if (rc == sizeof(struct lmv_mds_md_v1)) {
1446                         rc = lod_load_lmv_shards(env, lo, buf, true);
1447                         if (buf->lb_buf != info->lti_ea_store) {
1448                                 OBD_FREE_LARGE(info->lti_ea_store,
1449                                                info->lti_ea_store_size);
1450                                 info->lti_ea_store = buf->lb_buf;
1451                                 info->lti_ea_store_size = buf->lb_len;
1452                         }
1453
1454                         if (rc < 0)
1455                                 GOTO(out, rc);
1456                 }
1457
1458                 /*
1459                  * there is LMV EA (striping information) in this object
1460                  * let's parse it and create in-core objects for the stripes
1461                  */
1462                 rc = lod_parse_dir_striping(env, lo, buf);
1463                 if (rc == 0)
1464                         lo->ldo_dir_stripe_loaded = 1;
1465         }
1466 out:
1467         RETURN(rc);
1468 }
1469
1470 /**
1471  * A generic function to initialize the stripe objects.
1472  *
1473  * A protected version of lod_load_striping_locked() - load the striping
1474  * information from storage, parse that and instantiate LU objects to
1475  * represent the stripes.  The LOD object \a lo supplies a pointer to the
1476  * next sub-object in the LU stack so we can lock it. Also use \a lo to
1477  * return an array of references to the newly instantiated objects.
1478  *
1479  * \param[in] env               execution environment for this thread
1480  * \param[in,out] lo            LOD object, where striping is stored and
1481  *                              which gets an array of references
1482  *
1483  * \retval                      0 if parsing and object creation succeed
1484  * \retval                      negative error number on failure
1485  **/
1486 int lod_load_striping(const struct lu_env *env, struct lod_object *lo)
1487 {
1488         struct dt_object        *next = dt_object_child(&lo->ldo_obj);
1489         int                     rc;
1490
1491         if (!dt_object_exists(next))
1492                 return 0;
1493
1494         /* Check without locking first */
1495         if (lod_striping_loaded(lo))
1496                 return 0;
1497
1498         /* currently this code is supposed to be called from declaration
1499          * phase only, thus the object is not expected to be locked by caller */
1500         dt_write_lock(env, next, 0);
1501         rc = lod_load_striping_locked(env, lo);
1502         dt_write_unlock(env, next);
1503         return rc;
1504 }
1505
1506 /**
1507  * Verify lov_user_md_v1/v3 striping.
1508  *
1509  * Check the validity of all fields including the magic, stripe size,
1510  * stripe count, stripe offset and that the pool is present.  Also check
1511  * that each target index points to an existing target. The additional
1512  * \a is_from_disk turns additional checks. In some cases zero fields
1513  * are allowed (like pattern=0).
1514  *
1515  * \param[in] d                 LOD device
1516  * \param[in] buf               buffer with LOV EA to verify
1517  * \param[in] is_from_disk      0 - from user, allow some fields to be 0
1518  *                              1 - from disk, do not allow
1519  *
1520  * \retval                      0 if the striping is valid
1521  * \retval                      -EINVAL if striping is invalid
1522  */
1523 static int lod_verify_v1v3(struct lod_device *d, const struct lu_buf *buf,
1524                            bool is_from_disk)
1525 {
1526         struct lov_user_md_v1   *lum;
1527         struct lov_user_md_v3   *lum3;
1528         struct pool_desc        *pool = NULL;
1529         __u32                    magic;
1530         __u32                    stripe_size;
1531         __u16                    stripe_count;
1532         __u16                    stripe_offset;
1533         size_t                   lum_size;
1534         int                      rc = 0;
1535         ENTRY;
1536
1537         lum = buf->lb_buf;
1538
1539         if (buf->lb_len < sizeof(*lum)) {
1540                 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
1541                        buf->lb_len);
1542                 GOTO(out, rc = -EINVAL);
1543         }
1544
1545         magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
1546         if (magic != LOV_USER_MAGIC_V1 &&
1547             magic != LOV_USER_MAGIC_V3 &&
1548             magic != LOV_USER_MAGIC_SPECIFIC) {
1549                 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
1550                        le32_to_cpu(lum->lmm_magic));
1551                 GOTO(out, rc = -EINVAL);
1552         }
1553
1554         /* the user uses "0" for default stripe pattern normally. */
1555         if (!is_from_disk && lum->lmm_pattern == LOV_PATTERN_NONE)
1556                 lum->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1557
1558         if (!lov_pattern_supported(le32_to_cpu(lum->lmm_pattern))) {
1559                 CDEBUG(D_LAYOUT, "bad userland stripe pattern: %#x\n",
1560                        le32_to_cpu(lum->lmm_pattern));
1561                 GOTO(out, rc = -EINVAL);
1562         }
1563
1564         /* a released lum comes from creating orphan on hsm release,
1565          * doesn't make sense to verify it. */
1566         if (le32_to_cpu(lum->lmm_pattern) & LOV_PATTERN_F_RELEASED)
1567                 GOTO(out, rc = 0);
1568
1569         /* 64kB is the largest common page size we see (ia64), and matches the
1570          * check in lfs */
1571         stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1572         if (stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
1573                 CDEBUG(D_LAYOUT, "stripe size %u not a multiple of %u\n",
1574                        stripe_size, LOV_MIN_STRIPE_SIZE);
1575                 GOTO(out, rc = -EINVAL);
1576         }
1577
1578         stripe_offset = le16_to_cpu(lum->lmm_stripe_offset);
1579         if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT &&
1580             lov_pattern(le32_to_cpu(lum->lmm_pattern)) != LOV_PATTERN_MDT) {
1581                 /* if offset is not within valid range [0, osts_size) */
1582                 if (stripe_offset >= d->lod_osts_size) {
1583                         CDEBUG(D_LAYOUT, "stripe offset %u >= bitmap size %u\n",
1584                                stripe_offset, d->lod_osts_size);
1585                         GOTO(out, rc = -EINVAL);
1586                 }
1587
1588                 /* if lmm_stripe_offset is *not* in bitmap */
1589                 if (!cfs_bitmap_check(d->lod_ost_bitmap, stripe_offset)) {
1590                         CDEBUG(D_LAYOUT, "stripe offset %u not in bitmap\n",
1591                                stripe_offset);
1592                         GOTO(out, rc = -EINVAL);
1593                 }
1594         }
1595
1596         if (magic == LOV_USER_MAGIC_V1)
1597                 lum_size = offsetof(struct lov_user_md_v1,
1598                                     lmm_objects[0]);
1599         else if (magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC)
1600                 lum_size = offsetof(struct lov_user_md_v3,
1601                                     lmm_objects[0]);
1602         else
1603                 GOTO(out, rc = -EINVAL);
1604
1605         stripe_count = le16_to_cpu(lum->lmm_stripe_count);
1606         if (buf->lb_len < lum_size) {
1607                 CDEBUG(D_LAYOUT, "invalid buf len %zu/%zu for lov_user_md with "
1608                        "magic %#x and stripe_count %u\n",
1609                        buf->lb_len, lum_size, magic, stripe_count);
1610                 GOTO(out, rc = -EINVAL);
1611         }
1612
1613         if (!(magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC))
1614                 goto out;
1615
1616         lum3 = buf->lb_buf;
1617         /* In the function below, .hs_keycmp resolves to
1618          * pool_hashkey_keycmp() */
1619         /* coverity[overrun-buffer-val] */
1620         pool = lod_find_pool(d, lum3->lmm_pool_name);
1621         if (pool == NULL)
1622                 goto out;
1623
1624         if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT) {
1625                 rc = lod_check_index_in_pool(stripe_offset, pool);
1626                 if (rc < 0)
1627                         GOTO(out, rc = -EINVAL);
1628         }
1629
1630         if (is_from_disk && stripe_count > pool_tgt_count(pool)) {
1631                 CDEBUG(D_LAYOUT, "stripe count %u > # OSTs %u in the pool\n",
1632                        stripe_count, pool_tgt_count(pool));
1633                 GOTO(out, rc = -EINVAL);
1634         }
1635
1636 out:
1637         if (pool != NULL)
1638                 lod_pool_putref(pool);
1639
1640         RETURN(rc);
1641 }
1642
1643 /**
1644  * Verify LOV striping.
1645  *
1646  * \param[in] d                 LOD device
1647  * \param[in] buf               buffer with LOV EA to verify
1648  * \param[in] is_from_disk      0 - from user, allow some fields to be 0
1649  *                              1 - from disk, do not allow
1650  * \param[in] start             extent start for composite layout
1651  *
1652  * \retval                      0 if the striping is valid
1653  * \retval                      -EINVAL if striping is invalid
1654  */
1655 int lod_verify_striping(struct lod_device *d, struct lod_object *lo,
1656                         const struct lu_buf *buf, bool is_from_disk)
1657 {
1658         struct lov_desc *desc = &d->lod_desc;
1659         struct lov_user_md_v1   *lum;
1660         struct lov_comp_md_v1   *comp_v1;
1661         struct lov_comp_md_entry_v1     *ent;
1662         struct lu_extent        *ext;
1663         struct lu_buf   tmp;
1664         __u64   prev_end = 0;
1665         __u32   stripe_size = 0;
1666         __u16   prev_mid = -1, mirror_id = -1;
1667         __u32   mirror_count = 0;
1668         __u32   magic;
1669         int     rc = 0, i;
1670         ENTRY;
1671
1672         lum = buf->lb_buf;
1673
1674         if (buf->lb_len < sizeof(*lum)) {
1675                 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
1676                        buf->lb_len);
1677                 RETURN(-EINVAL);
1678         }
1679
1680         magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
1681         if (magic != LOV_USER_MAGIC_V1 &&
1682             magic != LOV_USER_MAGIC_V3 &&
1683             magic != LOV_USER_MAGIC_SPECIFIC &&
1684             magic != LOV_USER_MAGIC_COMP_V1) {
1685                 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
1686                        le32_to_cpu(lum->lmm_magic));
1687                 RETURN(-EINVAL);
1688         }
1689
1690         if (magic != LOV_USER_MAGIC_COMP_V1)
1691                 RETURN(lod_verify_v1v3(d, buf, is_from_disk));
1692
1693         /* magic == LOV_USER_MAGIC_COMP_V1 */
1694         comp_v1 = buf->lb_buf;
1695         if (buf->lb_len < le32_to_cpu(comp_v1->lcm_size)) {
1696                 CDEBUG(D_LAYOUT, "buf len %zu is less than %u\n",
1697                        buf->lb_len, le32_to_cpu(comp_v1->lcm_size));
1698                 RETURN(-EINVAL);
1699         }
1700
1701         if (le16_to_cpu(comp_v1->lcm_entry_count) == 0) {
1702                 CDEBUG(D_LAYOUT, "entry count is zero\n");
1703                 RETURN(-EINVAL);
1704         }
1705
1706         if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr) &&
1707             lo->ldo_comp_cnt > 0) {
1708                 /* could be called from lustre.lov.add */
1709                 __u32 cnt = lo->ldo_comp_cnt;
1710
1711                 ext = &lo->ldo_comp_entries[cnt - 1].llc_extent;
1712                 prev_end = ext->e_end;
1713
1714                 ++mirror_count;
1715         }
1716
1717         for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1718                 ent = &comp_v1->lcm_entries[i];
1719                 ext = &ent->lcme_extent;
1720
1721                 if (le64_to_cpu(ext->e_start) >= le64_to_cpu(ext->e_end)) {
1722                         CDEBUG(D_LAYOUT, "invalid extent "DEXT"\n",
1723                                le64_to_cpu(ext->e_start),
1724                                le64_to_cpu(ext->e_end));
1725                         RETURN(-EINVAL);
1726                 }
1727
1728                 if (is_from_disk) {
1729                         /* lcme_id contains valid value */
1730                         if (le32_to_cpu(ent->lcme_id) == 0 ||
1731                             le32_to_cpu(ent->lcme_id) > LCME_ID_MAX) {
1732                                 CDEBUG(D_LAYOUT, "invalid id %u\n",
1733                                        le32_to_cpu(ent->lcme_id));
1734                                 RETURN(-EINVAL);
1735                         }
1736
1737                         if (le16_to_cpu(comp_v1->lcm_mirror_count) > 0) {
1738                                 mirror_id = mirror_id_of(
1739                                                 le32_to_cpu(ent->lcme_id));
1740
1741                                 /* first component must start with 0 */
1742                                 if (mirror_id != prev_mid &&
1743                                     le64_to_cpu(ext->e_start) != 0) {
1744                                         CDEBUG(D_LAYOUT,
1745                                                "invalid start:%llu, expect:0\n",
1746                                                le64_to_cpu(ext->e_start));
1747                                         RETURN(-EINVAL);
1748                                 }
1749
1750                                 prev_mid = mirror_id;
1751                         }
1752                 }
1753
1754                 if (le64_to_cpu(ext->e_start) == 0) {
1755                         ++mirror_count;
1756                         prev_end = 0;
1757                 }
1758
1759                 /* the next must be adjacent with the previous one */
1760                 if (le64_to_cpu(ext->e_start) != prev_end) {
1761                         CDEBUG(D_LAYOUT,
1762                                "invalid start actual:%llu, expect:%llu\n",
1763                                le64_to_cpu(ext->e_start), prev_end);
1764                         RETURN(-EINVAL);
1765                 }
1766
1767                 prev_end = le64_to_cpu(ext->e_end);
1768
1769                 tmp.lb_buf = (char *)comp_v1 + le32_to_cpu(ent->lcme_offset);
1770                 tmp.lb_len = le32_to_cpu(ent->lcme_size);
1771
1772                 /* Check DoM entry is always the first one */
1773                 lum = tmp.lb_buf;
1774                 if (lov_pattern(le32_to_cpu(lum->lmm_pattern)) ==
1775                     LOV_PATTERN_MDT) {
1776                         /* DoM component can be only the first entry */
1777                         if (i > 0) {
1778                                 CDEBUG(D_LAYOUT, "invalid DoM layout "
1779                                        "entry found at %i index\n", i);
1780                                 RETURN(-EINVAL);
1781                         }
1782                         stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1783                         /* There is just one stripe on MDT and it must
1784                          * cover whole component size. */
1785                         if (stripe_size != prev_end) {
1786                                 CDEBUG(D_LAYOUT, "invalid DoM layout "
1787                                        "stripe size %u != %llu "
1788                                        "(component size)\n",
1789                                        stripe_size, prev_end);
1790                                 RETURN(-EINVAL);
1791                         }
1792                         /* Check stripe size againts per-MDT limit */
1793                         if (stripe_size > d->lod_dom_max_stripesize) {
1794                                 CDEBUG(D_LAYOUT, "DoM component size "
1795                                        "%u is bigger than MDT limit %u, check "
1796                                        "dom_max_stripesize parameter\n",
1797                                        stripe_size, d->lod_dom_max_stripesize);
1798                                 RETURN(-EINVAL);
1799                         }
1800                 }
1801
1802                 rc = lod_verify_v1v3(d, &tmp, is_from_disk);
1803                 if (rc)
1804                         RETURN(rc);
1805
1806                 if (prev_end == LUSTRE_EOF)
1807                         continue;
1808
1809                 /* extent end must be aligned with the stripe_size */
1810                 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1811                 if (stripe_size == 0)
1812                         stripe_size = desc->ld_default_stripe_size;
1813                 if (stripe_size == 0 || (prev_end & (stripe_size - 1))) {
1814                         CDEBUG(D_LAYOUT, "stripe size isn't aligned, "
1815                                "stripe_sz: %u, [%llu, %llu)\n",
1816                                stripe_size, ext->e_start, prev_end);
1817                         RETURN(-EINVAL);
1818                 }
1819         }
1820
1821         /* make sure that the mirror_count is telling the truth */
1822         if (mirror_count != le16_to_cpu(comp_v1->lcm_mirror_count) + 1)
1823                 RETURN(-EINVAL);
1824
1825         RETURN(0);
1826 }
1827
1828 /**
1829  * set the default stripe size, if unset.
1830  *
1831  * \param[in,out] val   number of bytes per OST stripe
1832  *
1833  * The minimum stripe size is 64KB to ensure that a single stripe is an
1834  * even multiple of a client PAGE_SIZE (IA64, PPC, etc).  Otherwise, it
1835  * is difficult to split dirty pages across OSCs during writes.
1836  */
1837 void lod_fix_desc_stripe_size(__u64 *val)
1838 {
1839         if (*val < LOV_MIN_STRIPE_SIZE) {
1840                 if (*val != 0)
1841                         LCONSOLE_INFO("Increasing default stripe size to "
1842                                       "minimum value %u\n",
1843                                       LOV_DESC_STRIPE_SIZE_DEFAULT);
1844                 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
1845         } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
1846                 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
1847                 LCONSOLE_WARN("Changing default stripe size to %llu (a "
1848                               "multiple of %u)\n",
1849                               *val, LOV_MIN_STRIPE_SIZE);
1850         }
1851 }
1852
1853 /**
1854  * set the filesystem default number of stripes, if unset.
1855  *
1856  * \param[in,out] val   number of stripes
1857  *
1858  * A value of "0" means "use the system-wide default stripe count", which
1859  * has either been inherited by now, or falls back to 1 stripe per file.
1860  * A value of "-1" (0xffffffff) means "stripe over all available OSTs",
1861  * and is a valid value, so is left unchanged here.
1862  */
1863 void lod_fix_desc_stripe_count(__u32 *val)
1864 {
1865         if (*val == 0)
1866                 *val = 1;
1867 }
1868
1869 /**
1870  * set the filesystem default layout pattern
1871  *
1872  * \param[in,out] val   LOV_PATTERN_* layout
1873  *
1874  * A value of "0" means "use the system-wide default layout type", which
1875  * has either been inherited by now, or falls back to plain RAID0 striping.
1876  */
1877 void lod_fix_desc_pattern(__u32 *val)
1878 {
1879         /* from lov_setstripe */
1880         if ((*val != 0) && (*val != LOV_PATTERN_RAID0) &&
1881             (*val != LOV_PATTERN_MDT)) {
1882                 LCONSOLE_WARN("Unknown stripe pattern: %#x\n", *val);
1883                 *val = 0;
1884         }
1885 }
1886
1887 void lod_fix_desc_qos_maxage(__u32 *val)
1888 {
1889         /* fix qos_maxage */
1890         if (*val == 0)
1891                 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
1892 }
1893
1894 /**
1895  * Used to fix insane default striping.
1896  *
1897  * \param[in] desc      striping description
1898  */
1899 void lod_fix_desc(struct lov_desc *desc)
1900 {
1901         lod_fix_desc_stripe_size(&desc->ld_default_stripe_size);
1902         lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
1903         lod_fix_desc_pattern(&desc->ld_pattern);
1904         lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
1905 }
1906
1907 /**
1908  * Initialize the structures used to store pools and default striping.
1909  *
1910  * \param[in] lod       LOD device
1911  * \param[in] lcfg      configuration structure storing default striping.
1912  *
1913  * \retval              0 if initialization succeeds
1914  * \retval              negative error number on failure
1915  */
1916 int lod_pools_init(struct lod_device *lod, struct lustre_cfg *lcfg)
1917 {
1918         struct obd_device          *obd;
1919         struct lov_desc            *desc;
1920         int                         rc;
1921         ENTRY;
1922
1923         obd = class_name2obd(lustre_cfg_string(lcfg, 0));
1924         LASSERT(obd != NULL);
1925         obd->obd_lu_dev = &lod->lod_dt_dev.dd_lu_dev;
1926
1927         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1928                 CERROR("LOD setup requires a descriptor\n");
1929                 RETURN(-EINVAL);
1930         }
1931
1932         desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
1933
1934         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1935                 CERROR("descriptor size wrong: %d > %d\n",
1936                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1937                 RETURN(-EINVAL);
1938         }
1939
1940         if (desc->ld_magic != LOV_DESC_MAGIC) {
1941                 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
1942                         CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
1943                                obd->obd_name, desc);
1944                         lustre_swab_lov_desc(desc);
1945                 } else {
1946                         CERROR("%s: Bad lov desc magic: %#x\n",
1947                                obd->obd_name, desc->ld_magic);
1948                         RETURN(-EINVAL);
1949                 }
1950         }
1951
1952         lod_fix_desc(desc);
1953
1954         desc->ld_active_tgt_count = 0;
1955         lod->lod_desc = *desc;
1956
1957         lod->lod_sp_me = LUSTRE_SP_CLI;
1958
1959         /* Set up allocation policy (QoS and RR) */
1960         INIT_LIST_HEAD(&lod->lod_qos.lq_oss_list);
1961         init_rwsem(&lod->lod_qos.lq_rw_sem);
1962         lod->lod_qos.lq_dirty = 1;
1963         lod->lod_qos.lq_rr.lqr_dirty = 1;
1964         lod->lod_qos.lq_reset = 1;
1965         /* Default priority is toward free space balance */
1966         lod->lod_qos.lq_prio_free = 232;
1967         /* Default threshold for rr (roughly 17%) */
1968         lod->lod_qos.lq_threshold_rr = 43;
1969
1970         /* Set up OST pool environment */
1971         lod->lod_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,
1972                                                    HASH_POOLS_MAX_BITS,
1973                                                    HASH_POOLS_BKT_BITS, 0,
1974                                                    CFS_HASH_MIN_THETA,
1975                                                    CFS_HASH_MAX_THETA,
1976                                                    &pool_hash_operations,
1977                                                    CFS_HASH_DEFAULT);
1978         if (lod->lod_pools_hash_body == NULL)
1979                 RETURN(-ENOMEM);
1980
1981         INIT_LIST_HEAD(&lod->lod_pool_list);
1982         lod->lod_pool_count = 0;
1983         rc = lod_ost_pool_init(&lod->lod_pool_info, 0);
1984         if (rc)
1985                 GOTO(out_hash, rc);
1986         lod_qos_rr_init(&lod->lod_qos.lq_rr);
1987         rc = lod_ost_pool_init(&lod->lod_qos.lq_rr.lqr_pool, 0);
1988         if (rc)
1989                 GOTO(out_pool_info, rc);
1990
1991         RETURN(0);
1992
1993 out_pool_info:
1994         lod_ost_pool_free(&lod->lod_pool_info);
1995 out_hash:
1996         cfs_hash_putref(lod->lod_pools_hash_body);
1997
1998         return rc;
1999 }
2000
2001 /**
2002  * Release the structures describing the pools.
2003  *
2004  * \param[in] lod       LOD device from which we release the structures
2005  *
2006  * \retval              0 always
2007  */
2008 int lod_pools_fini(struct lod_device *lod)
2009 {
2010         struct obd_device   *obd = lod2obd(lod);
2011         struct pool_desc    *pool, *tmp;
2012         ENTRY;
2013
2014         list_for_each_entry_safe(pool, tmp, &lod->lod_pool_list, pool_list) {
2015                 /* free pool structs */
2016                 CDEBUG(D_INFO, "delete pool %p\n", pool);
2017                 /* In the function below, .hs_keycmp resolves to
2018                  * pool_hashkey_keycmp() */
2019                 /* coverity[overrun-buffer-val] */
2020                 lod_pool_del(obd, pool->pool_name);
2021         }
2022
2023         cfs_hash_putref(lod->lod_pools_hash_body);
2024         lod_ost_pool_free(&(lod->lod_qos.lq_rr.lqr_pool));
2025         lod_ost_pool_free(&lod->lod_pool_info);
2026
2027         RETURN(0);
2028 }