lustre/lod/lod_dev.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  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  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, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/lod/lod_dev.c
  33  *
  34  * Lustre Logical Object Device
  35  *
  36  * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
  37  * Author: Mikhail Pershin <mike.pershin@intel.com>
  38  */
  39 /**
  40  * The Logical Object Device (LOD) layer manages access to striped
  41  * objects (both regular files and directories). It implements the DT
  42  * device and object APIs and is responsible for creating, storing,
  43  * and loading striping information as an extended attribute of the
  44  * underlying OSD object. LOD is the server side analog of the LOV and
  45  * LMV layers on the client side.
  46  *
  47  * Metadata LU object stack (layers of the same compound LU object,
  48  * all have the same FID):
  49  *
  50  *        MDT
  51  *         |      MD API
  52  *        MDD
  53  *         |      DT API
  54  *        LOD
  55  *       /   \    DT API
  56  *     OSD   OSP
  57  *
  58  * During LOD object initialization the localness or remoteness of the
  59  * object FID dictates the choice between OSD and OSP.
  60  *
  61  * An LOD object (file or directory) with N stripes (each has a
  62  * different FID):
  63  *
  64  *          LOD
  65  *           |
  66  *   +---+---+---+...+
  67  *   |   |   |   |   |
  68  *   S0  S1  S2  S3  S(N-1)  OS[DP] objects, seen as DT objects by LOD
  69  *
  70  * When upper layers must access an object's stripes (which are
  71  * themselves OST or MDT LU objects) LOD finds these objects by their
  72  * FIDs and stores them as an array of DT object pointers on the
  73  * object. Declarations and operations on LOD objects are received by
  74  * LOD (as DT object operations) and performed on the underlying
  75  * OS[DP] object and (as needed) on the stripes. From the perspective
  76  * of LOD, a stripe-less file (created by mknod() or open with
  77  * O_LOV_DELAY_CREATE) is an object which does not yet have stripes,
  78  * while a non-striped directory (created by mkdir()) is an object
  79  * which will never have stripes.
  80  *
  81  * The LOD layer also implements a small subset of the OBD device API
  82  * to support MDT stack initialization and finalization (an MDD device
  83  * connects and disconnects itself to and from the underlying LOD
  84  * device), and pool management. In turn LOD uses the OBD device API
  85  * to connect it self to the underlying OSD, and to connect itself to
  86  * OSP devices representing the MDTs and OSTs that bear the stripes of
  87  * its objects.
  88  */
  89
  90 #define DEBUG_SUBSYSTEM S_MDS
  91
  92 #include <linux/kthread.h>
  93 #include <obd_class.h>
  94 #include <md_object.h>
  95 #include <lustre_fid.h>
  96 #include <lustre_param.h>
  97 #include <lustre_update.h>
  98 #include <lustre_log.h>
  99
 100 #include "lod_internal.h"
 101
 102 static const char lod_update_log_name[] = "update_log";
 103 static const char lod_update_log_dir_name[] = "update_log_dir";
 104
 105 /*
 106  * Lookup target by FID.
 107  *
 108  * Lookup MDT/OST target index by FID. Type of the target can be
 109  * specific or any.
 110  *
 111  * \param[in] env               LU environment provided by the caller
 112  * \param[in] lod               lod device
 113  * \param[in] fid               FID
 114  * \param[out] tgt              result target index
 115  * \param[in] type              expected type of the target:
 116  *                              LU_SEQ_RANGE_{MDT,OST,ANY}
 117  *
 118  * \retval 0                    on success
 119  * \retval negative             negated errno on error
 120  **/
 121 int lod_fld_lookup(const struct lu_env *env, struct lod_device *lod,
 122                    const struct lu_fid *fid, __u32 *tgt, int *type)
 123 {
 124         struct lu_seq_range     range = { 0 };
 125         struct lu_server_fld    *server_fld;
 126         int rc;
 127         ENTRY;
 128
 129         if (!fid_is_sane(fid)) {
 130                 CERROR("%s: invalid FID "DFID"\n", lod2obd(lod)->obd_name,
 131                        PFID(fid));
 132                 RETURN(-EIO);
 133         }
 134
 135         if (fid_is_idif(fid)) {
 136                 *tgt = fid_idif_ost_idx(fid);
 137                 *type = LU_SEQ_RANGE_OST;
 138                 RETURN(0);
 139         }
 140
 141         if (fid_is_update_log(fid) || fid_is_update_log_dir(fid)) {
 142                 *tgt = fid_oid(fid);
 143                 *type = LU_SEQ_RANGE_MDT;
 144                 RETURN(0);
 145         }
 146
 147         if (!lod->lod_initialized || (!fid_seq_in_fldb(fid_seq(fid)))) {
 148                 LASSERT(lu_site2seq(lod2lu_dev(lod)->ld_site) != NULL);
 149
 150                 *tgt = lu_site2seq(lod2lu_dev(lod)->ld_site)->ss_node_id;
 151                 *type = LU_SEQ_RANGE_MDT;
 152                 RETURN(0);
 153         }
 154
 155         server_fld = lu_site2seq(lod2lu_dev(lod)->ld_site)->ss_server_fld;
 156         if (server_fld == NULL)
 157                 RETURN(-EIO);
 158
 159         fld_range_set_type(&range, *type);
 160         rc = fld_server_lookup(env, server_fld, fid_seq(fid), &range);
 161         if (rc != 0)
 162                 RETURN(rc);
 163
 164         *tgt = range.lsr_index;
 165         *type = range.lsr_flags;
 166
 167         CDEBUG(D_INFO, "%s: got tgt %x for sequence: "LPX64"\n",
 168                lod2obd(lod)->obd_name, *tgt, fid_seq(fid));
 169
 170         RETURN(0);
 171 }
 172
 173 /* Slab for OSD object allocation */
 174 struct kmem_cache *lod_object_kmem;
 175
 176 /* Slab for dt_txn_callback */
 177 struct kmem_cache *lod_txn_callback_kmem;
 178 static struct lu_kmem_descr lod_caches[] = {
 179         {
 180                 .ckd_cache = &lod_object_kmem,
 181                 .ckd_name  = "lod_obj",
 182                 .ckd_size  = sizeof(struct lod_object)
 183         },
 184         {
 185                 .ckd_cache = &lod_txn_callback_kmem,
 186                 .ckd_name  = "lod_txn_callback",
 187                 .ckd_size  = sizeof(struct dt_txn_callback)
 188         },
 189         {
 190                 .ckd_cache = NULL
 191         }
 192 };
 193
 194 static struct lu_device *lod_device_fini(const struct lu_env *env,
 195                                          struct lu_device *d);
 196
 197 /**
 198  * Implementation of lu_device_operations::ldo_object_alloc() for LOD
 199  *
 200  * Allocates and initializes LOD's slice in the given object.
 201  *
 202  * see include/lu_object.h for the details.
 203  */
 204 static struct lu_object *lod_object_alloc(const struct lu_env *env,
 205                                           const struct lu_object_header *hdr,
 206                                           struct lu_device *dev)
 207 {
 208         struct lod_object       *lod_obj;
 209         struct lu_object        *lu_obj;
 210         ENTRY;
 211
 212         OBD_SLAB_ALLOC_PTR_GFP(lod_obj, lod_object_kmem, GFP_NOFS);
 213         if (lod_obj == NULL)
 214                 RETURN(ERR_PTR(-ENOMEM));
 215
 216         lu_obj = lod2lu_obj(lod_obj);
 217         dt_object_init(&lod_obj->ldo_obj, NULL, dev);
 218         lod_obj->ldo_obj.do_ops = &lod_obj_ops;
 219         lu_obj->lo_ops = &lod_lu_obj_ops;
 220
 221         RETURN(lu_obj);
 222 }
 223
 224 /**
 225  * Process the config log for all sub device.
 226  *
 227  * The function goes through all the targets in the given table
 228  * and apply given configuration command on to the targets.
 229  * Used to cleanup the targets at unmount.
 230  *
 231  * \param[in] env               LU environment provided by the caller
 232  * \param[in] lod               lod device
 233  * \param[in] ltd               target's table to go through
 234  * \param[in] lcfg              configuration command to apply
 235  *
 236  * \retval 0                    on success
 237  * \retval negative             negated errno on error
 238  **/
 239 static int lod_sub_process_config(const struct lu_env *env,
 240                                  struct lod_device *lod,
 241                                  struct lod_tgt_descs *ltd,
 242                                  struct lustre_cfg *lcfg)
 243 {
 244         struct lu_device  *next;
 245         int rc = 0;
 246         unsigned int i;
 247
 248         lod_getref(ltd);
 249         if (ltd->ltd_tgts_size <= 0) {
 250                 lod_putref(lod, ltd);
 251                 return 0;
 252         }
 253         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
 254                 struct lod_tgt_desc *tgt;
 255                 int rc1;
 256
 257                 tgt = LTD_TGT(ltd, i);
 258                 LASSERT(tgt && tgt->ltd_tgt);
 259                 next = &tgt->ltd_tgt->dd_lu_dev;
 260                 rc1 = next->ld_ops->ldo_process_config(env, next, lcfg);
 261                 if (rc1) {
 262                         CERROR("%s: error cleaning up LOD index %u: cmd %#x"
 263                                ": rc = %d\n", lod2obd(lod)->obd_name, i,
 264                                lcfg->lcfg_command, rc1);
 265                         rc = rc1;
 266                 }
 267         }
 268         lod_putref(lod, ltd);
 269         return rc;
 270 }
 271
 272 struct lod_recovery_data {
 273         struct lod_device       *lrd_lod;
 274         struct lod_tgt_desc     *lrd_ltd;
 275         struct ptlrpc_thread    *lrd_thread;
 276         __u32                   lrd_idx;
 277 };
 278
 279
 280 /**
 281  * process update recovery record
 282  *
 283  * Add the update recovery recode to the update recovery list in
 284  * lod_recovery_data. Then the recovery thread (target_recovery_thread)
 285  * will redo these updates.
 286  *
 287  * \param[in]env        execution environment
 288  * \param[in]llh        log handle of update record
 289  * \param[in]rec        update record to be replayed
 290  * \param[in]data       update recovery data which holds the necessary
 291  *                      arguments for recovery (see struct lod_recovery_data)
 292  *
 293  * \retval              0 if the record is processed successfully.
 294  * \retval              negative errno if the record processing fails.
 295  */
 296 static int lod_process_recovery_updates(const struct lu_env *env,
 297                                         struct llog_handle *llh,
 298                                         struct llog_rec_hdr *rec,
 299                                         void *data)
 300 {
 301         struct lod_recovery_data        *lrd = data;
 302         struct llog_cookie      *cookie = &lod_env_info(env)->lti_cookie;
 303         struct lu_target                *lut;
 304         __u32                           index = 0;
 305         ENTRY;
 306
 307         if (lrd->lrd_ltd == NULL) {
 308                 int rc;
 309
 310                 rc = lodname2mdt_index(lod2obd(lrd->lrd_lod)->obd_name, &index);
 311                 if (rc != 0)
 312                         return rc;
 313         } else {
 314                 index = lrd->lrd_ltd->ltd_index;
 315         }
 316
 317         if (rec->lrh_len !=
 318                 llog_update_record_size((struct llog_update_record *)rec)) {
 319                 CERROR("%s broken update record! index %u "DOSTID":%u :"
 320                        " rc = %d\n", lod2obd(lrd->lrd_lod)->obd_name, index,
 321                        POSTID(&llh->lgh_id.lgl_oi), rec->lrh_index, -EIO);
 322                 return -EIO;
 323         }
 324
 325         cookie->lgc_lgl = llh->lgh_id;
 326         cookie->lgc_index = rec->lrh_index;
 327         cookie->lgc_subsys = LLOG_UPDATELOG_ORIG_CTXT;
 328
 329         CDEBUG(D_HA, "%s: process recovery updates "DOSTID":%u\n",
 330                lod2obd(lrd->lrd_lod)->obd_name,
 331                POSTID(&llh->lgh_id.lgl_oi), rec->lrh_index);
 332         lut = lod2lu_dev(lrd->lrd_lod)->ld_site->ls_tgt;
 333
 334         return insert_update_records_to_replay_list(lut->lut_tdtd,
 335                                         (struct llog_update_record *)rec,
 336                                         cookie, index);
 337 }
 338
 339 /**
 340  * recovery thread for update log
 341  *
 342  * Start recovery thread and prepare the sub llog, then it will retrieve
 343  * the update records from the correpondent MDT and do recovery.
 344  *
 345  * \param[in] arg       pointer to the recovery data
 346  *
 347  * \retval              0 if recovery succeeds
 348  * \retval              negative errno if recovery failed.
 349  */
 350 static int lod_sub_recovery_thread(void *arg)
 351 {
 352         struct lod_recovery_data        *lrd = arg;
 353         struct lod_device               *lod = lrd->lrd_lod;
 354         struct dt_device                *dt;
 355         struct ptlrpc_thread            *thread = lrd->lrd_thread;
 356         struct llog_ctxt                *ctxt = NULL;
 357         struct lu_env                   env;
 358         int                             rc;
 359         ENTRY;
 360
 361         thread->t_flags = SVC_RUNNING;
 362         wake_up(&thread->t_ctl_waitq);
 363
 364         rc = lu_env_init(&env, LCT_LOCAL | LCT_MD_THREAD);
 365         if (rc != 0) {
 366                 OBD_FREE_PTR(lrd);
 367                 CERROR("%s: can't initialize env: rc = %d\n",
 368                        lod2obd(lod)->obd_name, rc);
 369                 RETURN(rc);
 370         }
 371
 372         if (lrd->lrd_ltd == NULL)
 373                 dt = lod->lod_child;
 374         else
 375                 dt = lrd->lrd_ltd->ltd_tgt;
 376
 377 again:
 378         rc = lod_sub_prep_llog(&env, lod, dt, lrd->lrd_idx);
 379         if (rc == 0) {
 380                 /* Process the recovery record */
 381                 ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 382                                         LLOG_UPDATELOG_ORIG_CTXT);
 383                 LASSERT(ctxt != NULL);
 384                 LASSERT(ctxt->loc_handle != NULL);
 385
 386                 rc = llog_cat_process(&env, ctxt->loc_handle,
 387                                       lod_process_recovery_updates, lrd, 0, 0);
 388         }
 389
 390         if (rc < 0) {
 391                 struct lu_device *top_device;
 392
 393                 top_device = lod->lod_dt_dev.dd_lu_dev.ld_site->ls_top_dev;
 394                 /* Because the remote target might failover at the same time,
 395                  * let's retry here */
 396                 if ((rc == -ETIMEDOUT || rc == -EAGAIN || rc == -EIO) &&
 397                      dt != lod->lod_child &&
 398                     !top_device->ld_obd->obd_force_abort_recovery &&
 399                     !top_device->ld_obd->obd_stopping) {
 400                         if (ctxt != NULL) {
 401                                 if (ctxt->loc_handle != NULL)
 402                                         llog_cat_close(&env,
 403                                                        ctxt->loc_handle);
 404                                 llog_ctxt_put(ctxt);
 405                         }
 406                         goto again;
 407                 }
 408
 409                 CERROR("%s getting update log failed: rc = %d\n",
 410                        dt->dd_lu_dev.ld_obd->obd_name, rc);
 411                 llog_ctxt_put(ctxt);
 412                 GOTO(out, rc);
 413         }
 414         llog_ctxt_put(ctxt);
 415
 416         CDEBUG(D_HA, "%s retrieve update log: rc = %d\n",
 417                dt->dd_lu_dev.ld_obd->obd_name, rc);
 418
 419         if (lrd->lrd_ltd == NULL)
 420                 lod->lod_child_got_update_log = 1;
 421         else
 422                 lrd->lrd_ltd->ltd_got_update_log = 1;
 423
 424         if (lod->lod_child_got_update_log) {
 425                 struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
 426                 struct lod_tgt_desc     *tgt = NULL;
 427                 bool                    all_got_log = true;
 428                 int                     i;
 429
 430                 cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
 431                         tgt = LTD_TGT(ltd, i);
 432                         if (!tgt->ltd_got_update_log) {
 433                                 all_got_log = false;
 434                                 break;
 435                         }
 436                 }
 437
 438                 if (all_got_log) {
 439                         struct lu_target *lut;
 440
 441                         lut = lod2lu_dev(lod)->ld_site->ls_tgt;
 442                         CDEBUG(D_HA, "%s got update logs from all MDTs.\n",
 443                                lut->lut_obd->obd_name);
 444                         lut->lut_tdtd->tdtd_replay_ready = 1;
 445                         wake_up(&lut->lut_obd->obd_next_transno_waitq);
 446                 }
 447         }
 448
 449 out:
 450         OBD_FREE_PTR(lrd);
 451         thread->t_flags = SVC_STOPPED;
 452         wake_up(&thread->t_ctl_waitq);
 453         lu_env_fini(&env);
 454         RETURN(rc);
 455 }
 456
 457 /**
 458  * finish sub llog context
 459  *
 460  * Stop update recovery thread for the sub device, then cleanup the
 461  * correspondent llog ctxt.
 462  *
 463  * \param[in] env      execution environment
 464  * \param[in] lod      lod device to do update recovery
 465  * \param[in] thread   recovery thread on this sub device
 466  */
 467 void lod_sub_fini_llog(const struct lu_env *env,
 468                        struct dt_device *dt, struct ptlrpc_thread *thread)
 469 {
 470         struct obd_device       *obd;
 471         struct llog_ctxt        *ctxt;
 472         ENTRY;
 473
 474         obd = dt->dd_lu_dev.ld_obd;
 475         CDEBUG(D_INFO, "%s: finish sub llog\n", obd->obd_name);
 476         /* Stop recovery thread first */
 477         if (thread != NULL && thread->t_flags & SVC_RUNNING) {
 478                 thread->t_flags = SVC_STOPPING;
 479                 wake_up(&thread->t_ctl_waitq);
 480                 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
 481         }
 482
 483         ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
 484         if (ctxt == NULL)
 485                 RETURN_EXIT;
 486
 487         if (ctxt->loc_handle != NULL)
 488                 llog_cat_close(env, ctxt->loc_handle);
 489
 490         llog_cleanup(env, ctxt);
 491
 492         RETURN_EXIT;
 493 }
 494
 495 /**
 496  * Extract MDT target index from a device name.
 497  *
 498  * a helper function to extract index from the given device name
 499  * like "fsname-MDTxxxx-mdtlov"
 500  *
 501  * \param[in] lodname           device name
 502  * \param[out] mdt_index        extracted index
 503  *
 504  * \retval 0            on success
 505  * \retval -EINVAL      if the name is invalid
 506  */
 507 int lodname2mdt_index(char *lodname, __u32 *mdt_index)
 508 {
 509         unsigned long index;
 510         char *ptr, *tmp;
 511
 512         /* 1.8 configs don't have "-MDT0000" at the end */
 513         ptr = strstr(lodname, "-MDT");
 514         if (ptr == NULL) {
 515                 *mdt_index = 0;
 516                 return 0;
 517         }
 518
 519         ptr = strrchr(lodname, '-');
 520         if (ptr == NULL) {
 521                 CERROR("invalid MDT index in '%s'\n", lodname);
 522                 return -EINVAL;
 523         }
 524
 525         if (strncmp(ptr, "-mdtlov", 7) != 0) {
 526                 CERROR("invalid MDT index in '%s'\n", lodname);
 527                 return -EINVAL;
 528         }
 529
 530         if ((unsigned long)ptr - (unsigned long)lodname <= 8) {
 531                 CERROR("invalid MDT index in '%s'\n", lodname);
 532                 return -EINVAL;
 533         }
 534
 535         if (strncmp(ptr - 8, "-MDT", 4) != 0) {
 536                 CERROR("invalid MDT index in '%s'\n", lodname);
 537                 return -EINVAL;
 538         }
 539
 540         index = simple_strtol(ptr - 4, &tmp, 16);
 541         if (*tmp != '-' || index > INT_MAX) {
 542                 CERROR("invalid MDT index in '%s'\n", lodname);
 543                 return -EINVAL;
 544         }
 545         *mdt_index = index;
 546         return 0;
 547 }
 548
 549 /**
 550  * Init sub llog context
 551  *
 552  * Setup update llog ctxt for update recovery threads, then start the
 553  * recovery thread (lod_sub_recovery_thread) to read update llog from
 554  * the correspondent MDT to do update recovery.
 555  *
 556  * \param[in] env       execution environment
 557  * \param[in] lod       lod device to do update recovery
 558  * \param[in] dt        sub dt device for which the recovery thread is
 559  *
 560  * \retval              0 if initialization succeeds.
 561  * \retval              negative errno if initialization fails.
 562  */
 563 int lod_sub_init_llog(const struct lu_env *env, struct lod_device *lod,
 564                       struct dt_device *dt)
 565 {
 566         struct obd_device               *obd;
 567         struct lod_recovery_data        *lrd = NULL;
 568         struct ptlrpc_thread            *thread;
 569         struct task_struct              *task;
 570         struct l_wait_info              lwi = { 0 };
 571         struct lod_tgt_desc             *sub_ltd = NULL;
 572         __u32                           index;
 573         __u32                           master_index;
 574         int                             rc;
 575         ENTRY;
 576
 577         rc = lodname2mdt_index(lod2obd(lod)->obd_name, &master_index);
 578         if (rc != 0)
 579                 RETURN(rc);
 580
 581         OBD_ALLOC_PTR(lrd);
 582         if (lrd == NULL)
 583                 RETURN(-ENOMEM);
 584
 585         if (lod->lod_child == dt) {
 586                 thread = &lod->lod_child_recovery_thread;
 587                 index = master_index;
 588         } else {
 589                 struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
 590                 struct lod_tgt_desc     *tgt = NULL;
 591                 unsigned int            i;
 592
 593                 cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
 594                         tgt = LTD_TGT(ltd, i);
 595                         if (tgt->ltd_tgt == dt) {
 596                                 index = tgt->ltd_index;
 597                                 sub_ltd = tgt;
 598                                 break;
 599                         }
 600                 }
 601                 LASSERT(sub_ltd != NULL);
 602                 OBD_ALLOC_PTR(sub_ltd->ltd_recovery_thread);
 603                 if (sub_ltd->ltd_recovery_thread == NULL)
 604                         GOTO(free_lrd, rc = -ENOMEM);
 605
 606                 thread = sub_ltd->ltd_recovery_thread;
 607         }
 608
 609         CDEBUG(D_INFO, "%s init sub log %s\n", lod2obd(lod)->obd_name,
 610                dt->dd_lu_dev.ld_obd->obd_name);
 611         lrd->lrd_lod = lod;
 612         lrd->lrd_ltd = sub_ltd;
 613         lrd->lrd_thread = thread;
 614         lrd->lrd_idx = index;
 615         init_waitqueue_head(&thread->t_ctl_waitq);
 616
 617         obd = dt->dd_lu_dev.ld_obd;
 618         obd->obd_lvfs_ctxt.dt = dt;
 619         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_UPDATELOG_ORIG_CTXT,
 620                         NULL, &llog_common_cat_ops);
 621         if (rc < 0) {
 622                 CERROR("%s: cannot setup updatelog llog: rc = %d\n",
 623                        obd->obd_name, rc);
 624                 GOTO(free_thread, rc);
 625         }
 626
 627         /* Start the recovery thread */
 628         task = kthread_run(lod_sub_recovery_thread, lrd, "lod%04x_rec%04x",
 629                            master_index, index);
 630         if (IS_ERR(task)) {
 631                 rc = PTR_ERR(task);
 632                 CERROR("%s: cannot start recovery thread: rc = %d\n",
 633                        obd->obd_name, rc);
 634                 GOTO(out_llog, rc);
 635         }
 636
 637         l_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING ||
 638                                           thread->t_flags & SVC_STOPPED, &lwi);
 639
 640         RETURN(0);
 641 out_llog:
 642         lod_sub_fini_llog(env, dt, thread);
 643 free_thread:
 644         if (lod->lod_child != dt) {
 645                 OBD_FREE_PTR(sub_ltd->ltd_recovery_thread);
 646                 sub_ltd->ltd_recovery_thread = NULL;
 647         }
 648 free_lrd:
 649         OBD_FREE_PTR(lrd);
 650         RETURN(rc);
 651 }
 652
 653 /**
 654  * Stop sub recovery thread
 655  *
 656  * Stop sub recovery thread on all subs.
 657  *
 658  * \param[in] env       execution environment
 659  * \param[in] lod       lod device to do update recovery
 660  */
 661 static void lod_sub_stop_recovery_threads(const struct lu_env *env,
 662                                           struct lod_device *lod)
 663 {
 664         struct lod_tgt_descs *ltd = &lod->lod_mdt_descs;
 665         struct ptlrpc_thread    *thread;
 666         unsigned int i;
 667
 668         /* Stop the update log commit cancel threads and finish master
 669          * llog ctxt */
 670         thread = &lod->lod_child_recovery_thread;
 671         /* Stop recovery thread first */
 672         if (thread != NULL && thread->t_flags & SVC_RUNNING) {
 673                 thread->t_flags = SVC_STOPPING;
 674                 wake_up(&thread->t_ctl_waitq);
 675                 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
 676         }
 677
 678         lod_getref(ltd);
 679         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
 680                 struct lod_tgt_desc     *tgt;
 681
 682                 tgt = LTD_TGT(ltd, i);
 683                 thread = tgt->ltd_recovery_thread;
 684                 if (thread != NULL && thread->t_flags & SVC_RUNNING) {
 685                         thread->t_flags = SVC_STOPPING;
 686                         wake_up(&thread->t_ctl_waitq);
 687                         wait_event(thread->t_ctl_waitq,
 688                                    thread->t_flags & SVC_STOPPED);
 689                         OBD_FREE_PTR(tgt->ltd_recovery_thread);
 690                         tgt->ltd_recovery_thread = NULL;
 691                 }
 692         }
 693
 694         lod_putref(lod, ltd);
 695 }
 696
 697 /**
 698  * finish all sub llog
 699  *
 700  * cleanup all of sub llog ctxt on the LOD.
 701  *
 702  * \param[in] env       execution environment
 703  * \param[in] lod       lod device to do update recovery
 704  */
 705 static void lod_sub_fini_all_llogs(const struct lu_env *env,
 706                                    struct lod_device *lod)
 707 {
 708         struct lod_tgt_descs *ltd = &lod->lod_mdt_descs;
 709         unsigned int i;
 710
 711         /* Stop the update log commit cancel threads and finish master
 712          * llog ctxt */
 713         lod_sub_fini_llog(env, lod->lod_child,
 714                           &lod->lod_child_recovery_thread);
 715         lod_getref(ltd);
 716         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
 717                 struct lod_tgt_desc     *tgt;
 718
 719                 tgt = LTD_TGT(ltd, i);
 720                 lod_sub_fini_llog(env, tgt->ltd_tgt,
 721                                   tgt->ltd_recovery_thread);
 722         }
 723
 724         lod_putref(lod, ltd);
 725 }
 726
 727 /**
 728  * Prepare distribute txn
 729  *
 730  * Prepare distribute txn structure for LOD
 731  *
 732  * \param[in] env       execution environment
 733  * \param[in] lod_device  LOD device
 734  *
 735  * \retval              0 if preparation succeeds.
 736  * \retval              negative errno if preparation fails.
 737  */
 738 static int lod_prepare_distribute_txn(const struct lu_env *env,
 739                                       struct lod_device *lod)
 740 {
 741         struct target_distribute_txn_data *tdtd;
 742         struct lu_target                  *lut;
 743         int                               rc;
 744         ENTRY;
 745
 746         /* Init update recovery data */
 747         OBD_ALLOC_PTR(tdtd);
 748         if (tdtd == NULL)
 749                 RETURN(-ENOMEM);
 750
 751         lut = lod2lu_dev(lod)->ld_site->ls_tgt;
 752         tdtd->tdtd_dt = &lod->lod_dt_dev;
 753         rc = distribute_txn_init(env, lut, tdtd,
 754                 lu_site2seq(lod2lu_dev(lod)->ld_site)->ss_node_id);
 755
 756         if (rc < 0) {
 757                 CERROR("%s: cannot init distribute txn: rc = %d\n",
 758                        lod2obd(lod)->obd_name, rc);
 759                 OBD_FREE_PTR(tdtd);
 760                 RETURN(rc);
 761         }
 762
 763         lut->lut_tdtd = tdtd;
 764
 765         RETURN(0);
 766 }
 767
 768 /**
 769  * Finish distribute txn
 770  *
 771  * Release the resource holding by distribute txn, i.e. stop distribute
 772  * txn thread.
 773  *
 774  * \param[in] env       execution environment
 775  * \param[in] lod       lod device
 776  */
 777 static void lod_fini_distribute_txn(const struct lu_env *env,
 778                                     struct lod_device *lod)
 779 {
 780         struct lu_target                  *lut;
 781
 782         lut = lod2lu_dev(lod)->ld_site->ls_tgt;
 783         if (lut->lut_tdtd == NULL)
 784                 return;
 785
 786         distribute_txn_fini(env, lut->lut_tdtd);
 787
 788         OBD_FREE_PTR(lut->lut_tdtd);
 789         lut->lut_tdtd = NULL;
 790 }
 791
 792 /**
 793  * Implementation of lu_device_operations::ldo_process_config() for LOD
 794  *
 795  * The method is called by the configuration subsystem during setup,
 796  * cleanup and when the configuration changes. The method processes
 797  * few specific commands like adding/removing the targets, changing
 798  * the runtime parameters.
 799
 800  * \param[in] env               LU environment provided by the caller
 801  * \param[in] dev               lod device
 802  * \param[in] lcfg              configuration command to apply
 803  *
 804  * \retval 0                    on success
 805  * \retval negative             negated errno on error
 806  *
 807  * The examples are below.
 808  *
 809  * Add osc config log:
 810  * marker  20 (flags=0x01, v2.2.49.56) lustre-OST0001  'add osc'
 811  * add_uuid  nid=192.168.122.162@tcp(0x20000c0a87aa2)  0:  1:nidxxx
 812  * attach    0:lustre-OST0001-osc-MDT0001  1:osc  2:lustre-MDT0001-mdtlov_UUID
 813  * setup     0:lustre-OST0001-osc-MDT0001  1:lustre-OST0001_UUID  2:nid
 814  * lov_modify_tgts add 0:lustre-MDT0001-mdtlov  1:lustre-OST0001_UUID  2:1  3:1
 815  * marker  20 (flags=0x02, v2.2.49.56) lustre-OST0001  'add osc'
 816  *
 817  * Add mdc config log:
 818  * marker  10 (flags=0x01, v2.2.49.56) lustre-MDT0000  'add osp'
 819  * add_uuid  nid=192.168.122.162@tcp(0x20000c0a87aa2)  0:  1:nid
 820  * attach 0:lustre-MDT0000-osp-MDT0001  1:osp  2:lustre-MDT0001-mdtlov_UUID
 821  * setup     0:lustre-MDT0000-osp-MDT0001  1:lustre-MDT0000_UUID  2:nid
 822  * modify_mdc_tgts add 0:lustre-MDT0001  1:lustre-MDT0000_UUID  2:0  3:1
 823  * marker  10 (flags=0x02, v2.2.49.56) lustre-MDT0000_UUID  'add osp'
 824  */
 825 static int lod_process_config(const struct lu_env *env,
 826                               struct lu_device *dev,
 827                               struct lustre_cfg *lcfg)
 828 {
 829         struct lod_device *lod = lu2lod_dev(dev);
 830         struct lu_device  *next = &lod->lod_child->dd_lu_dev;
 831         char              *arg1;
 832         int                rc = 0;
 833         ENTRY;
 834
 835         switch(lcfg->lcfg_command) {
 836         case LCFG_LOV_DEL_OBD:
 837         case LCFG_LOV_ADD_INA:
 838         case LCFG_LOV_ADD_OBD:
 839         case LCFG_ADD_MDC: {
 840                 __u32 index;
 841                 __u32 mdt_index;
 842                 int gen;
 843                 /* lov_modify_tgts add  0:lov_mdsA  1:osp  2:0  3:1
 844                  * modify_mdc_tgts add  0:lustre-MDT0001
 845                  *                    1:lustre-MDT0001-mdc0002
 846                  *                    2:2  3:1*/
 847                 arg1 = lustre_cfg_string(lcfg, 1);
 848
 849                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%d", &index) != 1)
 850                         GOTO(out, rc = -EINVAL);
 851                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
 852                         GOTO(out, rc = -EINVAL);
 853
 854                 if (lcfg->lcfg_command == LCFG_LOV_ADD_OBD) {
 855                         __u32 mdt_index;
 856
 857                         rc = lodname2mdt_index(lustre_cfg_string(lcfg, 0),
 858                                                &mdt_index);
 859                         if (rc != 0)
 860                                 GOTO(out, rc);
 861
 862                         rc = lod_add_device(env, lod, arg1, index, gen,
 863                                             mdt_index, LUSTRE_OSC_NAME, 1);
 864                 } else if (lcfg->lcfg_command == LCFG_ADD_MDC) {
 865                         mdt_index = index;
 866                         rc = lod_add_device(env, lod, arg1, index, gen,
 867                                             mdt_index, LUSTRE_MDC_NAME, 1);
 868                 } else if (lcfg->lcfg_command == LCFG_LOV_ADD_INA) {
 869                         /*FIXME: Add mdt_index for LCFG_LOV_ADD_INA*/
 870                         mdt_index = 0;
 871                         rc = lod_add_device(env, lod, arg1, index, gen,
 872                                             mdt_index, LUSTRE_OSC_NAME, 0);
 873                 } else {
 874                         rc = lod_del_device(env, lod,
 875                                             &lod->lod_ost_descs,
 876                                             arg1, index, gen, true);
 877                 }
 878
 879                 break;
 880         }
 881
 882         case LCFG_PARAM: {
 883                 struct obd_device *obd;
 884                 char *param;
 885
 886                 /* Check if it is activate/deactivate mdc
 887                  * lustre-MDTXXXX-osp-MDTXXXX.active=1 */
 888                 param = lustre_cfg_buf(lcfg, 1);
 889                 if (strstr(param, "osp") != NULL &&
 890                     strstr(param, ".active=") != NULL) {
 891                         struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
 892                         struct lod_tgt_desc     *sub_tgt = NULL;
 893                         char *ptr;
 894                         char *tmp;
 895                         int i;
 896
 897                         ptr = strstr(param, ".");
 898                         *ptr = '\0';
 899                         obd = class_name2obd(param);
 900                         if (obd == NULL) {
 901                                 CERROR("%s: can not find %s: rc = %d\n",
 902                                        lod2obd(lod)->obd_name, param, -EINVAL);
 903                                 *ptr = '.';
 904                                 GOTO(out, rc);
 905                         }
 906
 907                         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
 908                                 struct lod_tgt_desc *tgt;
 909
 910                                 tgt = LTD_TGT(ltd, i);
 911                                 if (tgt->ltd_tgt->dd_lu_dev.ld_obd == obd) {
 912                                         sub_tgt = tgt;
 913                                         break;
 914                                 }
 915                         }
 916
 917                         if (sub_tgt == NULL) {
 918                                 CERROR("%s: can not find %s: rc = %d\n",
 919                                        lod2obd(lod)->obd_name, param, -EINVAL);
 920                                 *ptr = '.';
 921                                 GOTO(out, rc);
 922                         }
 923
 924                         *ptr = '.';
 925                         tmp = strstr(param, "=");
 926                         tmp++;
 927                         if (*tmp == '1') {
 928                                 struct llog_ctxt *ctxt;
 929
 930                                 obd = sub_tgt->ltd_tgt->dd_lu_dev.ld_obd;
 931                                 ctxt = llog_get_context(obd,
 932                                                 LLOG_UPDATELOG_ORIG_CTXT);
 933                                 if (ctxt == NULL) {
 934                                         rc = llog_setup(env, obd, &obd->obd_olg,
 935                                                        LLOG_UPDATELOG_ORIG_CTXT,
 936                                                     NULL, &llog_common_cat_ops);
 937                                         if (rc < 0)
 938                                                 GOTO(out, rc);
 939                                 } else {
 940                                         llog_ctxt_put(ctxt);
 941                                 }
 942                                 rc = lod_sub_prep_llog(env, lod,
 943                                                        sub_tgt->ltd_tgt,
 944                                                        sub_tgt->ltd_index);
 945                                 if (rc == 0)
 946                                         sub_tgt->ltd_active = 1;
 947                         } else {
 948                                 lod_sub_fini_llog(env, sub_tgt->ltd_tgt,
 949                                                   NULL);
 950                                 sub_tgt->ltd_active = 0;
 951                         }
 952                         GOTO(out, rc);
 953                 }
 954
 955                 obd = lod2obd(lod);
 956                 rc = class_process_proc_param(PARAM_LOV, obd->obd_vars,
 957                                               lcfg, obd);
 958                 if (rc > 0)
 959                         rc = 0;
 960                 GOTO(out, rc);
 961         }
 962         case LCFG_PRE_CLEANUP: {
 963                 lod_sub_process_config(env, lod, &lod->lod_mdt_descs, lcfg);
 964                 lod_sub_process_config(env, lod, &lod->lod_ost_descs, lcfg);
 965                 next = &lod->lod_child->dd_lu_dev;
 966                 rc = next->ld_ops->ldo_process_config(env, next, lcfg);
 967                 if (rc != 0)
 968                         CDEBUG(D_HA, "%s: can't process %u: %d\n",
 969                                lod2obd(lod)->obd_name, lcfg->lcfg_command, rc);
 970
 971                 lod_sub_stop_recovery_threads(env, lod);
 972                 lod_fini_distribute_txn(env, lod);
 973                 lod_sub_fini_all_llogs(env, lod);
 974                 break;
 975         }
 976         case LCFG_CLEANUP: {
 977                 /*
 978                  * do cleanup on underlying storage only when
 979                  * all OSPs are cleaned up, as they use that OSD as well
 980                  */
 981                 lu_dev_del_linkage(dev->ld_site, dev);
 982                 lod_sub_process_config(env, lod, &lod->lod_mdt_descs, lcfg);
 983                 lod_sub_process_config(env, lod, &lod->lod_ost_descs, lcfg);
 984                 next = &lod->lod_child->dd_lu_dev;
 985                 rc = next->ld_ops->ldo_process_config(env, next, lcfg);
 986                 if (rc)
 987                         CERROR("%s: can't process %u: %d\n",
 988                                lod2obd(lod)->obd_name, lcfg->lcfg_command, rc);
 989
 990                 rc = obd_disconnect(lod->lod_child_exp);
 991                 if (rc)
 992                         CERROR("error in disconnect from storage: %d\n", rc);
 993                 break;
 994         }
 995         default:
 996                CERROR("%s: unknown command %u\n", lod2obd(lod)->obd_name,
 997                       lcfg->lcfg_command);
 998                rc = -EINVAL;
 999                break;
1000         }
1001
1002 out:
1003         RETURN(rc);
1004 }
1005
1006 /**
1007  * Implementation of lu_device_operations::ldo_recovery_complete() for LOD
1008  *
1009  * The method is called once the recovery is complete. This implementation
1010  * distributes the notification to all the known targets.
1011  *
1012  * see include/lu_object.h for the details
1013  */
1014 static int lod_recovery_complete(const struct lu_env *env,
1015                                  struct lu_device *dev)
1016 {
1017         struct lod_device   *lod = lu2lod_dev(dev);
1018         struct lu_device    *next = &lod->lod_child->dd_lu_dev;
1019         unsigned int         i;
1020         int                  rc;
1021         ENTRY;
1022
1023         LASSERT(lod->lod_recovery_completed == 0);
1024         lod->lod_recovery_completed = 1;
1025
1026         rc = next->ld_ops->ldo_recovery_complete(env, next);
1027
1028         lod_getref(&lod->lod_ost_descs);
1029         if (lod->lod_osts_size > 0) {
1030                 cfs_foreach_bit(lod->lod_ost_bitmap, i) {
1031                         struct lod_tgt_desc *tgt;
1032                         tgt = OST_TGT(lod, i);
1033                         LASSERT(tgt && tgt->ltd_tgt);
1034                         next = &tgt->ltd_ost->dd_lu_dev;
1035                         rc = next->ld_ops->ldo_recovery_complete(env, next);
1036                         if (rc)
1037                                 CERROR("%s: can't complete recovery on #%d:"
1038                                         "%d\n", lod2obd(lod)->obd_name, i, rc);
1039                 }
1040         }
1041         lod_putref(lod, &lod->lod_ost_descs);
1042         RETURN(rc);
1043 }
1044
1045 /**
1046  * Init update logs on all sub device
1047  *
1048  * LOD initialize update logs on all of sub devices. Because the initialization
1049  * process might need FLD lookup, see llog_osd_open()->dt_locate()->...->
1050  * lod_object_init(), this API has to be called after LOD is initialized.
1051  * \param[in] env       execution environment
1052  * \param[in] lod       lod device
1053  *
1054  * \retval              0 if update log is initialized successfully.
1055  * \retval              negative errno if initialization fails.
1056  */
1057 static int lod_sub_init_llogs(const struct lu_env *env, struct lod_device *lod)
1058 {
1059         struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
1060         int                     rc;
1061         unsigned int            i;
1062         ENTRY;
1063
1064         /* llog must be setup after LOD is initialized, because llog
1065          * initialization include FLD lookup */
1066         LASSERT(lod->lod_initialized);
1067
1068         /* Init the llog in its own stack */
1069         rc = lod_sub_init_llog(env, lod, lod->lod_child);
1070         if (rc < 0)
1071                 RETURN(rc);
1072
1073         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
1074                 struct lod_tgt_desc     *tgt;
1075
1076                 tgt = LTD_TGT(ltd, i);
1077                 rc = lod_sub_init_llog(env, lod, tgt->ltd_tgt);
1078                 if (rc != 0)
1079                         break;
1080         }
1081
1082         RETURN(rc);
1083 }
1084
1085 /**
1086  * Implementation of lu_device_operations::ldo_prepare() for LOD
1087  *
1088  * see include/lu_object.h for the details.
1089  */
1090 static int lod_prepare(const struct lu_env *env, struct lu_device *pdev,
1091                        struct lu_device *cdev)
1092 {
1093         struct lod_device       *lod = lu2lod_dev(cdev);
1094         struct lu_device        *next = &lod->lod_child->dd_lu_dev;
1095         struct lu_fid           *fid = &lod_env_info(env)->lti_fid;
1096         int                     rc;
1097         struct dt_object        *root;
1098         struct dt_object        *dto;
1099         __u32                   index;
1100         ENTRY;
1101
1102         rc = next->ld_ops->ldo_prepare(env, pdev, next);
1103         if (rc != 0) {
1104                 CERROR("%s: prepare bottom error: rc = %d\n",
1105                        lod2obd(lod)->obd_name, rc);
1106                 RETURN(rc);
1107         }
1108
1109         lod->lod_initialized = 1;
1110
1111         rc = dt_root_get(env, lod->lod_child, fid);
1112         if (rc < 0)
1113                 RETURN(rc);
1114
1115         root = dt_locate(env, lod->lod_child, fid);
1116         if (IS_ERR(root))
1117                 RETURN(PTR_ERR(root));
1118
1119         /* Create update log object */
1120         index = lu_site2seq(lod2lu_dev(lod)->ld_site)->ss_node_id;
1121         lu_update_log_fid(fid, index);
1122
1123         dto = local_file_find_or_create_with_fid(env, lod->lod_child,
1124                                                  fid, root,
1125                                                  lod_update_log_name,
1126                                                  S_IFREG | S_IRUGO | S_IWUSR);
1127         if (IS_ERR(dto))
1128                 GOTO(out_put, rc = PTR_ERR(dto));
1129
1130         lu_object_put(env, &dto->do_lu);
1131
1132         /* Create update log dir */
1133         lu_update_log_dir_fid(fid, index);
1134         dto = local_file_find_or_create_with_fid(env, lod->lod_child,
1135                                                  fid, root,
1136                                                  lod_update_log_dir_name,
1137                                                  S_IFDIR | S_IRUGO | S_IWUSR);
1138         if (IS_ERR(dto))
1139                 GOTO(out_put, rc = PTR_ERR(dto));
1140
1141         lu_object_put(env, &dto->do_lu);
1142
1143         rc = lod_prepare_distribute_txn(env, lod);
1144         if (rc != 0)
1145                 GOTO(out_put, rc);
1146
1147         rc = lod_sub_init_llogs(env, lod);
1148         if (rc != 0)
1149                 GOTO(out_put, rc);
1150
1151 out_put:
1152         lu_object_put(env, &root->do_lu);
1153
1154         RETURN(rc);
1155 }
1156
1157 const struct lu_device_operations lod_lu_ops = {
1158         .ldo_object_alloc       = lod_object_alloc,
1159         .ldo_process_config     = lod_process_config,
1160         .ldo_recovery_complete  = lod_recovery_complete,
1161         .ldo_prepare            = lod_prepare,
1162 };
1163
1164 /**
1165  * Implementation of dt_device_operations::dt_root_get() for LOD
1166  *
1167  * see include/dt_object.h for the details.
1168  */
1169 static int lod_root_get(const struct lu_env *env,
1170                         struct dt_device *dev, struct lu_fid *f)
1171 {
1172         return dt_root_get(env, dt2lod_dev(dev)->lod_child, f);
1173 }
1174
1175 /**
1176  * Implementation of dt_device_operations::dt_statfs() for LOD
1177  *
1178  * see include/dt_object.h for the details.
1179  */
1180 static int lod_statfs(const struct lu_env *env,
1181                       struct dt_device *dev, struct obd_statfs *sfs)
1182 {
1183         return dt_statfs(env, dt2lod_dev(dev)->lod_child, sfs);
1184 }
1185
1186 /**
1187  * Implementation of dt_device_operations::dt_trans_create() for LOD
1188  *
1189  * Creates a transaction using local (to this node) OSD.
1190  *
1191  * see include/dt_object.h for the details.
1192  */
1193 static struct thandle *lod_trans_create(const struct lu_env *env,
1194                                         struct dt_device *dt)
1195 {
1196         struct thandle *th;
1197
1198         th = top_trans_create(env, dt2lod_dev(dt)->lod_child);
1199         if (IS_ERR(th))
1200                 return th;
1201
1202         th->th_dev = dt;
1203
1204         return th;
1205 }
1206
1207 /**
1208  * Implementation of dt_device_operations::dt_trans_start() for LOD
1209  *
1210  * Starts the set of local transactions using the targets involved
1211  * in declare phase. Initial support for the distributed transactions.
1212  *
1213  * see include/dt_object.h for the details.
1214  */
1215 static int lod_trans_start(const struct lu_env *env, struct dt_device *dt,
1216                            struct thandle *th)
1217 {
1218         return top_trans_start(env, dt2lod_dev(dt)->lod_child, th);
1219 }
1220
1221 static int lod_trans_cb_add(struct thandle *th,
1222                             struct dt_txn_commit_cb *dcb)
1223 {
1224         struct top_thandle      *top_th = container_of(th, struct top_thandle,
1225                                                        tt_super);
1226         return dt_trans_cb_add(top_th->tt_master_sub_thandle, dcb);
1227 }
1228
1229 /**
1230  * add noop update to the update records
1231  *
1232  * Add noop updates to the update records, which is only used in
1233  * test right now.
1234  *
1235  * \param[in] env       execution environment
1236  * \param[in] dt        dt device of lod
1237  * \param[in] th        thandle
1238  * \param[in] count     the count of update records to be added.
1239  *
1240  * \retval              0 if adding succeeds.
1241  * \retval              negative errno if adding fails.
1242  */
1243 static int lod_add_noop_records(const struct lu_env *env,
1244                                 struct dt_device *dt, struct thandle *th,
1245                                 int count)
1246 {
1247         struct top_thandle *top_th;
1248         struct lu_fid *fid = &lod_env_info(env)->lti_fid;
1249         int i;
1250         int rc = 0;
1251
1252         top_th = container_of(th, struct top_thandle, tt_super);
1253         if (top_th->tt_multiple_thandle == NULL)
1254                 return 0;
1255
1256         fid_zero(fid);
1257         for (i = 0; i < count; i++) {
1258                 rc = update_record_pack(noop, th, fid);
1259                 if (rc < 0)
1260                         return rc;
1261         }
1262         return rc;
1263 }
1264
1265 /**
1266  * Implementation of dt_device_operations::dt_trans_stop() for LOD
1267  *
1268  * Stops the set of local transactions using the targets involved
1269  * in declare phase. Initial support for the distributed transactions.
1270  *
1271  * see include/dt_object.h for the details.
1272  */
1273 static int lod_trans_stop(const struct lu_env *env, struct dt_device *dt,
1274                           struct thandle *th)
1275 {
1276         if (OBD_FAIL_CHECK(OBD_FAIL_SPLIT_UPDATE_REC)) {
1277                 int rc;
1278
1279                 rc = lod_add_noop_records(env, dt, th, 5000);
1280                 if (rc < 0)
1281                         RETURN(rc);
1282         }
1283         return top_trans_stop(env, dt2lod_dev(dt)->lod_child, th);
1284 }
1285
1286 /**
1287  * Implementation of dt_device_operations::dt_conf_get() for LOD
1288  *
1289  * Currently returns the configuration provided by the local OSD.
1290  *
1291  * see include/dt_object.h for the details.
1292  */
1293 static void lod_conf_get(const struct lu_env *env,
1294                          const struct dt_device *dev,
1295                          struct dt_device_param *param)
1296 {
1297         dt_conf_get(env, dt2lod_dev((struct dt_device *)dev)->lod_child, param);
1298 }
1299
1300 /**
1301  * Implementation of dt_device_operations::dt_sync() for LOD
1302  *
1303  * Syncs all known OST targets. Very very expensive and used
1304  * rarely by LFSCK now. Should not be used in general.
1305  *
1306  * see include/dt_object.h for the details.
1307  */
1308 static int lod_sync(const struct lu_env *env, struct dt_device *dev)
1309 {
1310         struct lod_device   *lod = dt2lod_dev(dev);
1311         struct lod_ost_desc *ost;
1312         unsigned int         i;
1313         int                  rc = 0;
1314         ENTRY;
1315
1316         lod_getref(&lod->lod_ost_descs);
1317         lod_foreach_ost(lod, i) {
1318                 ost = OST_TGT(lod, i);
1319                 LASSERT(ost && ost->ltd_ost);
1320                 rc = dt_sync(env, ost->ltd_ost);
1321                 if (rc) {
1322                         CERROR("%s: can't sync %u: %d\n",
1323                                lod2obd(lod)->obd_name, i, rc);
1324                         break;
1325                 }
1326         }
1327         lod_putref(lod, &lod->lod_ost_descs);
1328         if (rc == 0)
1329                 rc = dt_sync(env, lod->lod_child);
1330
1331         RETURN(rc);
1332 }
1333
1334 /**
1335  * Implementation of dt_device_operations::dt_ro() for LOD
1336  *
1337  * Turns local OSD read-only, used for the testing only.
1338  *
1339  * see include/dt_object.h for the details.
1340  */
1341 static int lod_ro(const struct lu_env *env, struct dt_device *dev)
1342 {
1343         return dt_ro(env, dt2lod_dev(dev)->lod_child);
1344 }
1345
1346 /**
1347  * Implementation of dt_device_operations::dt_commit_async() for LOD
1348  *
1349  * Asks local OSD to commit sooner.
1350  *
1351  * see include/dt_object.h for the details.
1352  */
1353 static int lod_commit_async(const struct lu_env *env, struct dt_device *dev)
1354 {
1355         return dt_commit_async(env, dt2lod_dev(dev)->lod_child);
1356 }
1357
1358 static const struct dt_device_operations lod_dt_ops = {
1359         .dt_root_get         = lod_root_get,
1360         .dt_statfs           = lod_statfs,
1361         .dt_trans_create     = lod_trans_create,
1362         .dt_trans_start      = lod_trans_start,
1363         .dt_trans_stop       = lod_trans_stop,
1364         .dt_conf_get         = lod_conf_get,
1365         .dt_sync             = lod_sync,
1366         .dt_ro               = lod_ro,
1367         .dt_commit_async     = lod_commit_async,
1368         .dt_trans_cb_add     = lod_trans_cb_add,
1369 };
1370
1371 /**
1372  * Connect to a local OSD.
1373  *
1374  * Used to connect to the local OSD at mount. OSD name is taken from the
1375  * configuration command passed. This connection is used to identify LU
1376  * site and pin the OSD from early removal.
1377  *
1378  * \param[in] env               LU environment provided by the caller
1379  * \param[in] lod               lod device
1380  * \param[in] cfg               configuration command to apply
1381  *
1382  * \retval 0                    on success
1383  * \retval negative             negated errno on error
1384  **/
1385 static int lod_connect_to_osd(const struct lu_env *env, struct lod_device *lod,
1386                               struct lustre_cfg *cfg)
1387 {
1388         struct obd_connect_data *data = NULL;
1389         struct obd_device       *obd;
1390         char                    *nextdev = NULL, *p, *s;
1391         int                      rc, len = 0;
1392         ENTRY;
1393
1394         LASSERT(cfg);
1395         LASSERT(lod->lod_child_exp == NULL);
1396
1397         /* compatibility hack: we still use old config logs
1398          * which specify LOV, but we need to learn underlying
1399          * OSD device, which is supposed to be:
1400          *  <fsname>-MDTxxxx-osd
1401          *
1402          * 2.x MGS generates lines like the following:
1403          *   #03 (176)lov_setup 0:lustre-MDT0000-mdtlov  1:(struct lov_desc)
1404          * 1.8 MGS generates lines like the following:
1405          *   #03 (168)lov_setup 0:lustre-mdtlov  1:(struct lov_desc)
1406          *
1407          * we use "-MDT" to differentiate 2.x from 1.8 */
1408
1409         if ((p = lustre_cfg_string(cfg, 0)) && strstr(p, "-mdtlov")) {
1410                 len = strlen(p) + 6;
1411                 OBD_ALLOC(nextdev, len);
1412                 if (nextdev == NULL)
1413                         GOTO(out, rc = -ENOMEM);
1414
1415                 strcpy(nextdev, p);
1416                 s = strstr(nextdev, "-mdtlov");
1417                 if (unlikely(s == NULL)) {
1418                         CERROR("unable to parse device name %s\n",
1419                                lustre_cfg_string(cfg, 0));
1420                         GOTO(out, rc = -EINVAL);
1421                 }
1422
1423                 if (strstr(nextdev, "-MDT")) {
1424                         /* 2.x config */
1425                         strcpy(s, "-osd");
1426                 } else {
1427                         /* 1.8 config */
1428                         strcpy(s, "-MDT0000-osd");
1429                 }
1430         } else {
1431                 CERROR("unable to parse device name %s\n",
1432                        lustre_cfg_string(cfg, 0));
1433                 GOTO(out, rc = -EINVAL);
1434         }
1435
1436         OBD_ALLOC_PTR(data);
1437         if (data == NULL)
1438                 GOTO(out, rc = -ENOMEM);
1439
1440         obd = class_name2obd(nextdev);
1441         if (obd == NULL) {
1442                 CERROR("can not locate next device: %s\n", nextdev);
1443                 GOTO(out, rc = -ENOTCONN);
1444         }
1445
1446         data->ocd_connect_flags = OBD_CONNECT_VERSION;
1447         data->ocd_version = LUSTRE_VERSION_CODE;
1448
1449         rc = obd_connect(env, &lod->lod_child_exp, obd, &obd->obd_uuid,
1450                          data, NULL);
1451         if (rc) {
1452                 CERROR("cannot connect to next dev %s (%d)\n", nextdev, rc);
1453                 GOTO(out, rc);
1454         }
1455
1456         lod->lod_dt_dev.dd_lu_dev.ld_site =
1457                 lod->lod_child_exp->exp_obd->obd_lu_dev->ld_site;
1458         LASSERT(lod->lod_dt_dev.dd_lu_dev.ld_site);
1459         lod->lod_child = lu2dt_dev(lod->lod_child_exp->exp_obd->obd_lu_dev);
1460
1461 out:
1462         if (data)
1463                 OBD_FREE_PTR(data);
1464         if (nextdev)
1465                 OBD_FREE(nextdev, len);
1466         RETURN(rc);
1467 }
1468
1469 /**
1470  * Allocate and initialize target table.
1471  *
1472  * A helper function to initialize the target table and allocate
1473  * a bitmap of the available targets.
1474  *
1475  * \param[in] ltd               target's table to initialize
1476  *
1477  * \retval 0                    on success
1478  * \retval negative             negated errno on error
1479  **/
1480 static int lod_tgt_desc_init(struct lod_tgt_descs *ltd)
1481 {
1482         mutex_init(&ltd->ltd_mutex);
1483         init_rwsem(&ltd->ltd_rw_sem);
1484
1485         /* the OST array and bitmap are allocated/grown dynamically as OSTs are
1486          * added to the LOD, see lod_add_device() */
1487         ltd->ltd_tgt_bitmap = CFS_ALLOCATE_BITMAP(32);
1488         if (ltd->ltd_tgt_bitmap == NULL)
1489                 RETURN(-ENOMEM);
1490
1491         ltd->ltd_tgts_size  = 32;
1492         ltd->ltd_tgtnr      = 0;
1493
1494         ltd->ltd_death_row = 0;
1495         ltd->ltd_refcount  = 0;
1496         return 0;
1497 }
1498
1499 /**
1500  * Initialize LOD device at setup.
1501  *
1502  * Initializes the given LOD device using the original configuration command.
1503  * The function initiates a connection to the local OSD and initializes few
1504  * internal structures like pools, target tables, etc.
1505  *
1506  * \param[in] env               LU environment provided by the caller
1507  * \param[in] lod               lod device
1508  * \param[in] ldt               not used
1509  * \param[in] cfg               configuration command
1510  *
1511  * \retval 0                    on success
1512  * \retval negative             negated errno on error
1513  **/
1514 static int lod_init0(const struct lu_env *env, struct lod_device *lod,
1515                      struct lu_device_type *ldt, struct lustre_cfg *cfg)
1516 {
1517         struct dt_device_param ddp;
1518         struct obd_device     *obd;
1519         int                    rc;
1520         ENTRY;
1521
1522         obd = class_name2obd(lustre_cfg_string(cfg, 0));
1523         if (obd == NULL) {
1524                 CERROR("Cannot find obd with name %s\n",
1525                        lustre_cfg_string(cfg, 0));
1526                 RETURN(-ENODEV);
1527         }
1528
1529         obd->obd_lu_dev = &lod->lod_dt_dev.dd_lu_dev;
1530         lod->lod_dt_dev.dd_lu_dev.ld_obd = obd;
1531         lod->lod_dt_dev.dd_lu_dev.ld_ops = &lod_lu_ops;
1532         lod->lod_dt_dev.dd_ops = &lod_dt_ops;
1533
1534         rc = lod_connect_to_osd(env, lod, cfg);
1535         if (rc)
1536                 RETURN(rc);
1537
1538         dt_conf_get(env, &lod->lod_dt_dev, &ddp);
1539         lod->lod_osd_max_easize = ddp.ddp_max_ea_size;
1540
1541         /* setup obd to be used with old lov code */
1542         rc = lod_pools_init(lod, cfg);
1543         if (rc)
1544                 GOTO(out_disconnect, rc);
1545
1546         rc = lod_procfs_init(lod);
1547         if (rc)
1548                 GOTO(out_pools, rc);
1549
1550         spin_lock_init(&lod->lod_desc_lock);
1551         spin_lock_init(&lod->lod_connects_lock);
1552         lod_tgt_desc_init(&lod->lod_mdt_descs);
1553         lod_tgt_desc_init(&lod->lod_ost_descs);
1554
1555         RETURN(0);
1556
1557 out_pools:
1558         lod_pools_fini(lod);
1559 out_disconnect:
1560         obd_disconnect(lod->lod_child_exp);
1561         RETURN(rc);
1562 }
1563
1564 /**
1565  * Implementation of lu_device_type_operations::ldto_device_free() for LOD
1566  *
1567  * Releases the memory allocated for LOD device.
1568  *
1569  * see include/lu_object.h for the details.
1570  */
1571 static struct lu_device *lod_device_free(const struct lu_env *env,
1572                                          struct lu_device *lu)
1573 {
1574         struct lod_device *lod = lu2lod_dev(lu);
1575         struct lu_device  *next = &lod->lod_child->dd_lu_dev;
1576         ENTRY;
1577
1578         LASSERTF(atomic_read(&lu->ld_ref) == 0, "lu is %p\n", lu);
1579         dt_device_fini(&lod->lod_dt_dev);
1580         OBD_FREE_PTR(lod);
1581         RETURN(next);
1582 }
1583
1584 /**
1585  * Implementation of lu_device_type_operations::ldto_device_alloc() for LOD
1586  *
1587  * Allocates LOD device and calls the helpers to initialize it.
1588  *
1589  * see include/lu_object.h for the details.
1590  */
1591 static struct lu_device *lod_device_alloc(const struct lu_env *env,
1592                                           struct lu_device_type *type,
1593                                           struct lustre_cfg *lcfg)
1594 {
1595         struct lod_device *lod;
1596         struct lu_device  *lu_dev;
1597
1598         OBD_ALLOC_PTR(lod);
1599         if (lod == NULL) {
1600                 lu_dev = ERR_PTR(-ENOMEM);
1601         } else {
1602                 int rc;
1603
1604                 lu_dev = lod2lu_dev(lod);
1605                 dt_device_init(&lod->lod_dt_dev, type);
1606                 rc = lod_init0(env, lod, type, lcfg);
1607                 if (rc != 0) {
1608                         lod_device_free(env, lu_dev);
1609                         lu_dev = ERR_PTR(rc);
1610                 }
1611         }
1612
1613         return lu_dev;
1614 }
1615
1616 /**
1617  * Implementation of lu_device_type_operations::ldto_device_fini() for LOD
1618  *
1619  * Releases the internal resources used by LOD device.
1620  *
1621  * see include/lu_object.h for the details.
1622  */
1623 static struct lu_device *lod_device_fini(const struct lu_env *env,
1624                                          struct lu_device *d)
1625 {
1626         struct lod_device *lod = lu2lod_dev(d);
1627         int                rc;
1628         ENTRY;
1629
1630         lod_pools_fini(lod);
1631
1632         lod_procfs_fini(lod);
1633
1634         rc = lod_fini_tgt(env, lod, &lod->lod_ost_descs, true);
1635         if (rc)
1636                 CERROR("%s:can not fini ost descs %d\n",
1637                         lod2obd(lod)->obd_name, rc);
1638
1639         rc = lod_fini_tgt(env, lod, &lod->lod_mdt_descs, false);
1640         if (rc)
1641                 CERROR("%s:can not fini mdt descs %d\n",
1642                         lod2obd(lod)->obd_name, rc);
1643
1644         RETURN(NULL);
1645 }
1646
1647 /**
1648  * Implementation of obd_ops::o_connect() for LOD
1649  *
1650  * Used to track all the users of this specific LOD device,
1651  * so the device stays up until the last user disconnected.
1652  *
1653  * \param[in] env               LU environment provided by the caller
1654  * \param[out] exp              export the caller will be using to access LOD
1655  * \param[in] obd               OBD device representing LOD device
1656  * \param[in] cluuid            unique identifier of the caller
1657  * \param[in] data              not used
1658  * \param[in] localdata         not used
1659  *
1660  * \retval 0                    on success
1661  * \retval negative             negated errno on error
1662  **/
1663 static int lod_obd_connect(const struct lu_env *env, struct obd_export **exp,
1664                            struct obd_device *obd, struct obd_uuid *cluuid,
1665                            struct obd_connect_data *data, void *localdata)
1666 {
1667         struct lod_device    *lod = lu2lod_dev(obd->obd_lu_dev);
1668         struct lustre_handle  conn;
1669         int                   rc;
1670         ENTRY;
1671
1672         CDEBUG(D_CONFIG, "connect #%d\n", lod->lod_connects);
1673
1674         rc = class_connect(&conn, obd, cluuid);
1675         if (rc)
1676                 RETURN(rc);
1677
1678         *exp = class_conn2export(&conn);
1679
1680         spin_lock(&lod->lod_connects_lock);
1681         lod->lod_connects++;
1682         /* at the moment we expect the only user */
1683         LASSERT(lod->lod_connects == 1);
1684         spin_unlock(&lod->lod_connects_lock);
1685
1686         RETURN(0);
1687 }
1688
1689 /**
1690  *
1691  * Implementation of obd_ops::o_disconnect() for LOD
1692  *
1693  * When the caller doesn't need to use this LOD instance, it calls
1694  * obd_disconnect() and LOD releases corresponding export/reference count.
1695  * Once all the users gone, LOD device is released.
1696  *
1697  * \param[in] exp               export provided to the caller in obd_connect()
1698  *
1699  * \retval 0                    on success
1700  * \retval negative             negated errno on error
1701  **/
1702 static int lod_obd_disconnect(struct obd_export *exp)
1703 {
1704         struct obd_device *obd = exp->exp_obd;
1705         struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
1706         int                rc, release = 0;
1707         ENTRY;
1708
1709         /* Only disconnect the underlying layers on the final disconnect. */
1710         spin_lock(&lod->lod_connects_lock);
1711         lod->lod_connects--;
1712         if (lod->lod_connects != 0) {
1713                 /* why should there be more than 1 connect? */
1714                 spin_unlock(&lod->lod_connects_lock);
1715                 CERROR("%s: disconnect #%d\n", exp->exp_obd->obd_name,
1716                        lod->lod_connects);
1717                 goto out;
1718         }
1719         spin_unlock(&lod->lod_connects_lock);
1720
1721         /* the last user of lod has gone, let's release the device */
1722         release = 1;
1723
1724 out:
1725         rc = class_disconnect(exp); /* bz 9811 */
1726
1727         if (rc == 0 && release)
1728                 class_manual_cleanup(obd);
1729         RETURN(rc);
1730 }
1731
1732 LU_KEY_INIT(lod, struct lod_thread_info);
1733
1734 static void lod_key_fini(const struct lu_context *ctx,
1735                 struct lu_context_key *key, void *data)
1736 {
1737         struct lod_thread_info *info = data;
1738         /* allocated in lod_get_lov_ea
1739          * XXX: this is overload, a tread may have such store but used only
1740          * once. Probably better would be pool of such stores per LOD.
1741          */
1742         if (info->lti_ea_store) {
1743                 OBD_FREE_LARGE(info->lti_ea_store, info->lti_ea_store_size);
1744                 info->lti_ea_store = NULL;
1745                 info->lti_ea_store_size = 0;
1746         }
1747         lu_buf_free(&info->lti_linkea_buf);
1748         OBD_FREE_PTR(info);
1749 }
1750
1751 /* context key: lod_thread_key */
1752 LU_CONTEXT_KEY_DEFINE(lod, LCT_MD_THREAD);
1753
1754 LU_TYPE_INIT_FINI(lod, &lod_thread_key);
1755
1756 static struct lu_device_type_operations lod_device_type_ops = {
1757         .ldto_init           = lod_type_init,
1758         .ldto_fini           = lod_type_fini,
1759
1760         .ldto_start          = lod_type_start,
1761         .ldto_stop           = lod_type_stop,
1762
1763         .ldto_device_alloc   = lod_device_alloc,
1764         .ldto_device_free    = lod_device_free,
1765
1766         .ldto_device_fini    = lod_device_fini
1767 };
1768
1769 static struct lu_device_type lod_device_type = {
1770         .ldt_tags     = LU_DEVICE_DT,
1771         .ldt_name     = LUSTRE_LOD_NAME,
1772         .ldt_ops      = &lod_device_type_ops,
1773         .ldt_ctx_tags = LCT_MD_THREAD,
1774 };
1775
1776 /**
1777  * Implementation of obd_ops::o_get_info() for LOD
1778  *
1779  * Currently, there is only one supported key: KEY_OSP_CONNECTED , to provide
1780  * the caller binary status whether LOD has seen connection to any OST target.
1781  * It will also check if the MDT update log context being initialized (if
1782  * needed).
1783  *
1784  * \param[in] env               LU environment provided by the caller
1785  * \param[in] exp               export of the caller
1786  * \param[in] keylen            len of the key
1787  * \param[in] key               the key
1788  * \param[in] vallen            not used
1789  * \param[in] val               not used
1790  *
1791  * \retval                      0 if a connection was seen
1792  * \retval                      -EAGAIN if LOD isn't running yet or no
1793  *                              connection has been seen yet
1794  * \retval                      -EINVAL if not supported key is requested
1795  **/
1796 static int lod_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1797                             __u32 keylen, void *key, __u32 *vallen, void *val)
1798 {
1799         int rc = -EINVAL;
1800
1801         if (KEY_IS(KEY_OSP_CONNECTED)) {
1802                 struct obd_device       *obd = exp->exp_obd;
1803                 struct lod_device       *d;
1804                 struct lod_tgt_desc     *tgt;
1805                 unsigned int            i;
1806                 int                     rc = 1;
1807
1808                 if (!obd->obd_set_up || obd->obd_stopping)
1809                         RETURN(-EAGAIN);
1810
1811                 d = lu2lod_dev(obd->obd_lu_dev);
1812                 lod_getref(&d->lod_ost_descs);
1813                 lod_foreach_ost(d, i) {
1814                         tgt = OST_TGT(d, i);
1815                         LASSERT(tgt && tgt->ltd_tgt);
1816                         rc = obd_get_info(env, tgt->ltd_exp, keylen, key,
1817                                           vallen, val);
1818                         /* one healthy device is enough */
1819                         if (rc == 0)
1820                                 break;
1821                 }
1822                 lod_putref(d, &d->lod_ost_descs);
1823
1824                 lod_getref(&d->lod_mdt_descs);
1825                 lod_foreach_mdt(d, i) {
1826                         struct llog_ctxt *ctxt;
1827
1828                         tgt = MDT_TGT(d, i);
1829                         LASSERT(tgt != NULL);
1830                         LASSERT(tgt->ltd_tgt != NULL);
1831                         if (!tgt->ltd_active)
1832                                 continue;
1833
1834                         ctxt = llog_get_context(tgt->ltd_tgt->dd_lu_dev.ld_obd,
1835                                                 LLOG_UPDATELOG_ORIG_CTXT);
1836                         if (ctxt == NULL) {
1837                                 CDEBUG(D_INFO, "%s: %s is not ready.\n",
1838                                        obd->obd_name,
1839                                       tgt->ltd_tgt->dd_lu_dev.ld_obd->obd_name);
1840                                 rc = -EAGAIN;
1841                                 break;
1842                         }
1843                         if (ctxt->loc_handle == NULL) {
1844                                 CDEBUG(D_INFO, "%s: %s is not ready.\n",
1845                                        obd->obd_name,
1846                                       tgt->ltd_tgt->dd_lu_dev.ld_obd->obd_name);
1847                                 rc = -EAGAIN;
1848                                 llog_ctxt_put(ctxt);
1849                                 break;
1850                         }
1851                         llog_ctxt_put(ctxt);
1852                 }
1853                 lod_putref(d, &d->lod_mdt_descs);
1854
1855                 RETURN(rc);
1856         }
1857
1858         RETURN(rc);
1859 }
1860
1861 static int lod_obd_set_info_async(const struct lu_env *env,
1862                                   struct obd_export *exp,
1863                                   __u32 keylen, void *key,
1864                                   __u32 vallen, void *val,
1865                                   struct ptlrpc_request_set *set)
1866 {
1867         struct obd_device *obd = class_exp2obd(exp);
1868         struct lod_device *d;
1869         struct lod_tgt_desc *tgt;
1870         int no_set = 0;
1871         int i, rc = 0, rc2;
1872         ENTRY;
1873
1874         if (set == NULL) {
1875                 no_set = 1;
1876                 set = ptlrpc_prep_set();
1877                 if (!set)
1878                         RETURN(-ENOMEM);
1879         }
1880
1881         d = lu2lod_dev(obd->obd_lu_dev);
1882         lod_getref(&d->lod_ost_descs);
1883         lod_foreach_ost(d, i) {
1884                 tgt = OST_TGT(d, i);
1885                 LASSERT(tgt && tgt->ltd_tgt);
1886                 if (!tgt->ltd_active)
1887                         continue;
1888
1889                 rc2 = obd_set_info_async(env, tgt->ltd_exp, keylen, key,
1890                                          vallen, val, set);
1891                 if (rc2 != 0 && rc == 0)
1892                         rc = rc2;
1893         }
1894         lod_putref(d, &d->lod_ost_descs);
1895
1896         lod_getref(&d->lod_mdt_descs);
1897         lod_foreach_mdt(d, i) {
1898                 tgt = MDT_TGT(d, i);
1899                 LASSERT(tgt && tgt->ltd_tgt);
1900                 if (!tgt->ltd_active)
1901                         continue;
1902                 rc2 = obd_set_info_async(env, tgt->ltd_exp, keylen, key,
1903                                          vallen, val, set);
1904                 if (rc2 != 0 && rc == 0)
1905                         rc = rc2;
1906         }
1907         lod_putref(d, &d->lod_mdt_descs);
1908
1909
1910         if (no_set) {
1911                 rc2 = ptlrpc_set_wait(set);
1912                 if (rc2 == 0 && rc == 0)
1913                         rc = rc2;
1914                 ptlrpc_set_destroy(set);
1915         }
1916         RETURN(rc);
1917 }
1918
1919 static struct obd_ops lod_obd_device_ops = {
1920         .o_owner        = THIS_MODULE,
1921         .o_connect      = lod_obd_connect,
1922         .o_disconnect   = lod_obd_disconnect,
1923         .o_get_info     = lod_obd_get_info,
1924         .o_set_info_async = lod_obd_set_info_async,
1925         .o_pool_new     = lod_pool_new,
1926         .o_pool_rem     = lod_pool_remove,
1927         .o_pool_add     = lod_pool_add,
1928         .o_pool_del     = lod_pool_del,
1929 };
1930
1931 static int __init lod_init(void)
1932 {
1933         struct obd_type *type;
1934         int rc;
1935
1936         rc = lu_kmem_init(lod_caches);
1937         if (rc)
1938                 return rc;
1939
1940         rc = class_register_type(&lod_obd_device_ops, NULL, true, NULL,
1941                                  LUSTRE_LOD_NAME, &lod_device_type);
1942         if (rc) {
1943                 lu_kmem_fini(lod_caches);
1944                 return rc;
1945         }
1946
1947         /* create "lov" entry in procfs for compatibility purposes */
1948         type = class_search_type(LUSTRE_LOV_NAME);
1949         if (type != NULL && type->typ_procroot != NULL)
1950                 return rc;
1951
1952         type = class_search_type(LUSTRE_LOD_NAME);
1953         type->typ_procsym = lprocfs_register("lov", proc_lustre_root,
1954                                              NULL, NULL);
1955         if (IS_ERR(type->typ_procsym)) {
1956                 CERROR("lod: can't create compat entry \"lov\": %d\n",
1957                        (int)PTR_ERR(type->typ_procsym));
1958                 type->typ_procsym = NULL;
1959         }
1960         return rc;
1961 }
1962
1963 static void __exit lod_exit(void)
1964 {
1965         class_unregister_type(LUSTRE_LOD_NAME);
1966         lu_kmem_fini(lod_caches);
1967 }
1968
1969 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1970 MODULE_DESCRIPTION("Lustre Logical Object Device ("LUSTRE_LOD_NAME")");
1971 MODULE_VERSION(LUSTRE_VERSION_STRING);
1972 MODULE_LICENSE("GPL");
1973
1974 module_init(lod_init);
1975 module_exit(lod_exit);