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