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, 2016, 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: %#llx\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 "DFID".%u :"
 320                        " rc = %d\n", lod2obd(lrd->lrd_lod)->obd_name, index,
 321                        PFID(&llh->lgh_id.lgl_oi.oi_fid), rec->lrh_index, -EIO);
 322                 return -EINVAL;
 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 "DFID".%u\n",
 330                lod2obd(lrd->lrd_lod)->obd_name,
 331                PFID(&llh->lgh_id.lgl_oi.oi_fid), rec->lrh_index);
 332         lut = lod2lu_dev(lrd->lrd_lod)->ld_site->ls_tgt;
 333
 334         if (lut->lut_obd->obd_stopping ||
 335             lut->lut_obd->obd_abort_recovery)
 336                 return -ESHUTDOWN;
 337
 338         return insert_update_records_to_replay_list(lut->lut_tdtd,
 339                                         (struct llog_update_record *)rec,
 340                                         cookie, index);
 341 }
 342
 343 /**
 344  * recovery thread for update log
 345  *
 346  * Start recovery thread and prepare the sub llog, then it will retrieve
 347  * the update records from the correpondent MDT and do recovery.
 348  *
 349  * \param[in] arg       pointer to the recovery data
 350  *
 351  * \retval              0 if recovery succeeds
 352  * \retval              negative errno if recovery failed.
 353  */
 354 static int lod_sub_recovery_thread(void *arg)
 355 {
 356         struct lod_recovery_data        *lrd = arg;
 357         struct lod_device               *lod = lrd->lrd_lod;
 358         struct dt_device                *dt;
 359         struct ptlrpc_thread            *thread = lrd->lrd_thread;
 360         struct llog_ctxt                *ctxt = NULL;
 361         struct lu_env                   env;
 362         struct lu_target *lut;
 363
 364
 365         int                             rc;
 366         ENTRY;
 367
 368         thread->t_flags = SVC_RUNNING;
 369         wake_up(&thread->t_ctl_waitq);
 370
 371         rc = lu_env_init(&env, LCT_LOCAL | LCT_MD_THREAD);
 372         if (rc != 0) {
 373                 OBD_FREE_PTR(lrd);
 374                 CERROR("%s: can't initialize env: rc = %d\n",
 375                        lod2obd(lod)->obd_name, rc);
 376                 RETURN(rc);
 377         }
 378
 379         lut = lod2lu_dev(lod)->ld_site->ls_tgt;
 380         atomic_inc(&lut->lut_tdtd->tdtd_recovery_threads_count);
 381         if (lrd->lrd_ltd == NULL)
 382                 dt = lod->lod_child;
 383         else
 384                 dt = lrd->lrd_ltd->ltd_tgt;
 385
 386 again:
 387         rc = lod_sub_prep_llog(&env, lod, dt, lrd->lrd_idx);
 388         if (!rc && !lod->lod_child->dd_rdonly) {
 389                 /* Process the recovery record */
 390                 ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 391                                         LLOG_UPDATELOG_ORIG_CTXT);
 392                 LASSERT(ctxt != NULL);
 393                 LASSERT(ctxt->loc_handle != NULL);
 394
 395                 rc = llog_cat_process(&env, ctxt->loc_handle,
 396                                       lod_process_recovery_updates, lrd, 0, 0);
 397         }
 398
 399         if (rc < 0) {
 400                 struct lu_device *top_device;
 401
 402                 top_device = lod->lod_dt_dev.dd_lu_dev.ld_site->ls_top_dev;
 403                 /* Because the remote target might failover at the same time,
 404                  * let's retry here */
 405                 if ((rc == -ETIMEDOUT || rc == -EAGAIN || rc == -EIO) &&
 406                      dt != lod->lod_child &&
 407                     !top_device->ld_obd->obd_abort_recovery &&
 408                     !top_device->ld_obd->obd_stopping) {
 409                         if (ctxt != NULL) {
 410                                 if (ctxt->loc_handle != NULL)
 411                                         llog_cat_close(&env,
 412                                                        ctxt->loc_handle);
 413                                 llog_ctxt_put(ctxt);
 414                         }
 415                         goto again;
 416                 }
 417
 418                 CERROR("%s getting update log failed: rc = %d\n",
 419                        dt->dd_lu_dev.ld_obd->obd_name, rc);
 420                 llog_ctxt_put(ctxt);
 421
 422                 spin_lock(&top_device->ld_obd->obd_dev_lock);
 423                 if (!top_device->ld_obd->obd_abort_recovery &&
 424                     !top_device->ld_obd->obd_stopping)
 425                         top_device->ld_obd->obd_abort_recovery = 1;
 426                 spin_unlock(&top_device->ld_obd->obd_dev_lock);
 427
 428                 GOTO(out, rc);
 429         }
 430         llog_ctxt_put(ctxt);
 431
 432         CDEBUG(D_HA, "%s retrieve update log: rc = %d\n",
 433                dt->dd_lu_dev.ld_obd->obd_name, rc);
 434
 435         if (lrd->lrd_ltd == NULL)
 436                 lod->lod_child_got_update_log = 1;
 437         else
 438                 lrd->lrd_ltd->ltd_got_update_log = 1;
 439
 440         if (lod->lod_child_got_update_log) {
 441                 struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
 442                 struct lod_tgt_desc     *tgt = NULL;
 443                 bool                    all_got_log = true;
 444                 int                     i;
 445
 446                 cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
 447                         tgt = LTD_TGT(ltd, i);
 448                         if (!tgt->ltd_got_update_log) {
 449                                 all_got_log = false;
 450                                 break;
 451                         }
 452                 }
 453
 454                 if (all_got_log) {
 455                         CDEBUG(D_HA, "%s got update logs from all MDTs.\n",
 456                                lut->lut_obd->obd_name);
 457                         lut->lut_tdtd->tdtd_replay_ready = 1;
 458                         wake_up(&lut->lut_obd->obd_next_transno_waitq);
 459                 }
 460         }
 461
 462 out:
 463         OBD_FREE_PTR(lrd);
 464         thread->t_flags = SVC_STOPPED;
 465         atomic_dec(&lut->lut_tdtd->tdtd_recovery_threads_count);
 466         wake_up(&lut->lut_tdtd->tdtd_recovery_threads_waitq);
 467         wake_up(&thread->t_ctl_waitq);
 468         lu_env_fini(&env);
 469         RETURN(rc);
 470 }
 471
 472 /**
 473  * finish sub llog context
 474  *
 475  * Stop update recovery thread for the sub device, then cleanup the
 476  * correspondent llog ctxt.
 477  *
 478  * \param[in] env      execution environment
 479  * \param[in] lod      lod device to do update recovery
 480  * \param[in] thread   recovery thread on this sub device
 481  */
 482 void lod_sub_fini_llog(const struct lu_env *env,
 483                        struct dt_device *dt, struct ptlrpc_thread *thread)
 484 {
 485         struct obd_device       *obd;
 486         struct llog_ctxt        *ctxt;
 487         ENTRY;
 488
 489         obd = dt->dd_lu_dev.ld_obd;
 490         CDEBUG(D_INFO, "%s: finish sub llog\n", obd->obd_name);
 491         /* Stop recovery thread first */
 492         if (thread != NULL && thread->t_flags & SVC_RUNNING) {
 493                 thread->t_flags = SVC_STOPPING;
 494                 wake_up(&thread->t_ctl_waitq);
 495                 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
 496         }
 497
 498         ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
 499         if (ctxt == NULL)
 500                 RETURN_EXIT;
 501
 502         if (ctxt->loc_handle != NULL)
 503                 llog_cat_close(env, ctxt->loc_handle);
 504
 505         llog_cleanup(env, ctxt);
 506
 507         RETURN_EXIT;
 508 }
 509
 510 /**
 511  * Extract MDT target index from a device name.
 512  *
 513  * a helper function to extract index from the given device name
 514  * like "fsname-MDTxxxx-mdtlov"
 515  *
 516  * \param[in] lodname           device name
 517  * \param[out] mdt_index        extracted index
 518  *
 519  * \retval 0            on success
 520  * \retval -EINVAL      if the name is invalid
 521  */
 522 int lodname2mdt_index(char *lodname, __u32 *mdt_index)
 523 {
 524         unsigned long index;
 525         char *ptr, *tmp;
 526
 527         /* 1.8 configs don't have "-MDT0000" at the end */
 528         ptr = strstr(lodname, "-MDT");
 529         if (ptr == NULL) {
 530                 *mdt_index = 0;
 531                 return 0;
 532         }
 533
 534         ptr = strrchr(lodname, '-');
 535         if (ptr == NULL) {
 536                 CERROR("invalid MDT index in '%s'\n", lodname);
 537                 return -EINVAL;
 538         }
 539
 540         if (strncmp(ptr, "-mdtlov", 7) != 0) {
 541                 CERROR("invalid MDT index in '%s'\n", lodname);
 542                 return -EINVAL;
 543         }
 544
 545         if ((unsigned long)ptr - (unsigned long)lodname <= 8) {
 546                 CERROR("invalid MDT index in '%s'\n", lodname);
 547                 return -EINVAL;
 548         }
 549
 550         if (strncmp(ptr - 8, "-MDT", 4) != 0) {
 551                 CERROR("invalid MDT index in '%s'\n", lodname);
 552                 return -EINVAL;
 553         }
 554
 555         index = simple_strtol(ptr - 4, &tmp, 16);
 556         if (*tmp != '-' || index > INT_MAX) {
 557                 CERROR("invalid MDT index in '%s'\n", lodname);
 558                 return -EINVAL;
 559         }
 560         *mdt_index = index;
 561         return 0;
 562 }
 563
 564 /**
 565  * Init sub llog context
 566  *
 567  * Setup update llog ctxt for update recovery threads, then start the
 568  * recovery thread (lod_sub_recovery_thread) to read update llog from
 569  * the correspondent MDT to do update recovery.
 570  *
 571  * \param[in] env       execution environment
 572  * \param[in] lod       lod device to do update recovery
 573  * \param[in] dt        sub dt device for which the recovery thread is
 574  *
 575  * \retval              0 if initialization succeeds.
 576  * \retval              negative errno if initialization fails.
 577  */
 578 int lod_sub_init_llog(const struct lu_env *env, struct lod_device *lod,
 579                       struct dt_device *dt)
 580 {
 581         struct obd_device               *obd;
 582         struct lod_recovery_data        *lrd = NULL;
 583         struct ptlrpc_thread            *thread;
 584         struct task_struct              *task;
 585         struct l_wait_info              lwi = { 0 };
 586         struct lod_tgt_desc             *sub_ltd = NULL;
 587         __u32                           index;
 588         __u32                           master_index;
 589         int                             rc;
 590         ENTRY;
 591
 592         rc = lodname2mdt_index(lod2obd(lod)->obd_name, &master_index);
 593         if (rc != 0)
 594                 RETURN(rc);
 595
 596         OBD_ALLOC_PTR(lrd);
 597         if (lrd == NULL)
 598                 RETURN(-ENOMEM);
 599
 600         if (lod->lod_child == dt) {
 601                 thread = &lod->lod_child_recovery_thread;
 602                 index = master_index;
 603         } else {
 604                 struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
 605                 struct lod_tgt_desc     *tgt = NULL;
 606                 unsigned int            i;
 607
 608                 cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
 609                         tgt = LTD_TGT(ltd, i);
 610                         if (tgt->ltd_tgt == dt) {
 611                                 index = tgt->ltd_index;
 612                                 sub_ltd = tgt;
 613                                 break;
 614                         }
 615                 }
 616                 LASSERT(sub_ltd != NULL);
 617                 OBD_ALLOC_PTR(sub_ltd->ltd_recovery_thread);
 618                 if (sub_ltd->ltd_recovery_thread == NULL)
 619                         GOTO(free_lrd, rc = -ENOMEM);
 620
 621                 thread = sub_ltd->ltd_recovery_thread;
 622         }
 623
 624         CDEBUG(D_INFO, "%s init sub log %s\n", lod2obd(lod)->obd_name,
 625                dt->dd_lu_dev.ld_obd->obd_name);
 626         lrd->lrd_lod = lod;
 627         lrd->lrd_ltd = sub_ltd;
 628         lrd->lrd_thread = thread;
 629         lrd->lrd_idx = index;
 630         init_waitqueue_head(&thread->t_ctl_waitq);
 631
 632         obd = dt->dd_lu_dev.ld_obd;
 633         obd->obd_lvfs_ctxt.dt = dt;
 634         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_UPDATELOG_ORIG_CTXT,
 635                         NULL, &llog_common_cat_ops);
 636         if (rc < 0) {
 637                 CERROR("%s: cannot setup updatelog llog: rc = %d\n",
 638                        obd->obd_name, rc);
 639                 GOTO(free_thread, rc);
 640         }
 641
 642         /* Start the recovery thread */
 643         task = kthread_run(lod_sub_recovery_thread, lrd, "lod%04x_rec%04x",
 644                            master_index, index);
 645         if (IS_ERR(task)) {
 646                 rc = PTR_ERR(task);
 647                 CERROR("%s: cannot start recovery thread: rc = %d\n",
 648                        obd->obd_name, rc);
 649                 GOTO(out_llog, rc);
 650         }
 651
 652         l_wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_RUNNING ||
 653                                           thread->t_flags & SVC_STOPPED, &lwi);
 654
 655         RETURN(0);
 656 out_llog:
 657         lod_sub_fini_llog(env, dt, thread);
 658 free_thread:
 659         if (lod->lod_child != dt) {
 660                 OBD_FREE_PTR(sub_ltd->ltd_recovery_thread);
 661                 sub_ltd->ltd_recovery_thread = NULL;
 662         }
 663 free_lrd:
 664         OBD_FREE_PTR(lrd);
 665         RETURN(rc);
 666 }
 667
 668 /**
 669  * Stop sub recovery thread
 670  *
 671  * Stop sub recovery thread on all subs.
 672  *
 673  * \param[in] env       execution environment
 674  * \param[in] lod       lod device to do update recovery
 675  */
 676 static void lod_sub_stop_recovery_threads(const struct lu_env *env,
 677                                           struct lod_device *lod)
 678 {
 679         struct lod_tgt_descs *ltd = &lod->lod_mdt_descs;
 680         struct ptlrpc_thread    *thread;
 681         unsigned int i;
 682
 683         /* Stop the update log commit cancel threads and finish master
 684          * llog ctxt */
 685         thread = &lod->lod_child_recovery_thread;
 686         /* Stop recovery thread first */
 687         if (thread != NULL && thread->t_flags & SVC_RUNNING) {
 688                 thread->t_flags = SVC_STOPPING;
 689                 wake_up(&thread->t_ctl_waitq);
 690                 wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
 691         }
 692
 693         lod_getref(ltd);
 694         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
 695                 struct lod_tgt_desc     *tgt;
 696
 697                 tgt = LTD_TGT(ltd, i);
 698                 thread = tgt->ltd_recovery_thread;
 699                 if (thread != NULL && thread->t_flags & SVC_RUNNING) {
 700                         thread->t_flags = SVC_STOPPING;
 701                         wake_up(&thread->t_ctl_waitq);
 702                         wait_event(thread->t_ctl_waitq,
 703                                    thread->t_flags & SVC_STOPPED);
 704                         OBD_FREE_PTR(tgt->ltd_recovery_thread);
 705                         tgt->ltd_recovery_thread = NULL;
 706                 }
 707         }
 708
 709         lod_putref(lod, ltd);
 710 }
 711
 712 /**
 713  * finish all sub llog
 714  *
 715  * cleanup all of sub llog ctxt on the LOD.
 716  *
 717  * \param[in] env       execution environment
 718  * \param[in] lod       lod device to do update recovery
 719  */
 720 static void lod_sub_fini_all_llogs(const struct lu_env *env,
 721                                    struct lod_device *lod)
 722 {
 723         struct lod_tgt_descs *ltd = &lod->lod_mdt_descs;
 724         unsigned int i;
 725
 726         /* Stop the update log commit cancel threads and finish master
 727          * llog ctxt */
 728         lod_sub_fini_llog(env, lod->lod_child,
 729                           &lod->lod_child_recovery_thread);
 730         lod_getref(ltd);
 731         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
 732                 struct lod_tgt_desc     *tgt;
 733
 734                 tgt = LTD_TGT(ltd, i);
 735                 lod_sub_fini_llog(env, tgt->ltd_tgt,
 736                                   tgt->ltd_recovery_thread);
 737         }
 738
 739         lod_putref(lod, ltd);
 740 }
 741
 742 static char *lod_show_update_logs_retrievers(void *data, int *size, int *count)
 743 {
 744         struct lod_device       *lod = (struct lod_device *)data;
 745         struct lu_target        *lut = lod2lu_dev(lod)->ld_site->ls_tgt;
 746         struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
 747         struct lod_tgt_desc     *tgt = NULL;
 748         char                    *buf;
 749         int                      len = 0;
 750         int                      rc;
 751         int                      i;
 752
 753         *count = atomic_read(&lut->lut_tdtd->tdtd_recovery_threads_count);
 754         if (*count == 0) {
 755                 *size = 0;
 756                 return NULL;
 757         }
 758
 759         *size = 5 * *count + 1;
 760         OBD_ALLOC(buf, *size);
 761         if (buf == NULL)
 762                 return NULL;
 763
 764         *count = 0;
 765         memset(buf, 0, *size);
 766
 767         if (!lod->lod_child_got_update_log) {
 768                 rc = lodname2mdt_index(lod2obd(lod)->obd_name, &i);
 769                 LASSERTF(rc == 0, "Fail to parse target index: rc = %d\n", rc);
 770
 771                 rc = snprintf(buf + len, *size - len, " %04x", i);
 772                 LASSERT(rc > 0);
 773
 774                 len += rc;
 775                 (*count)++;
 776         }
 777
 778         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
 779                 tgt = LTD_TGT(ltd, i);
 780                 if (!tgt->ltd_got_update_log) {
 781                         rc = snprintf(buf + len, *size - len, " %04x", i);
 782                         if (unlikely(rc <= 0))
 783                                 break;
 784
 785                         len += rc;
 786                         (*count)++;
 787                 }
 788         }
 789
 790         return buf;
 791 }
 792
 793 /**
 794  * Prepare distribute txn
 795  *
 796  * Prepare distribute txn structure for LOD
 797  *
 798  * \param[in] env       execution environment
 799  * \param[in] lod_device  LOD device
 800  *
 801  * \retval              0 if preparation succeeds.
 802  * \retval              negative errno if preparation fails.
 803  */
 804 static int lod_prepare_distribute_txn(const struct lu_env *env,
 805                                       struct lod_device *lod)
 806 {
 807         struct target_distribute_txn_data *tdtd;
 808         struct lu_target                  *lut;
 809         int                               rc;
 810         ENTRY;
 811
 812         /* Init update recovery data */
 813         OBD_ALLOC_PTR(tdtd);
 814         if (tdtd == NULL)
 815                 RETURN(-ENOMEM);
 816
 817         lut = lod2lu_dev(lod)->ld_site->ls_tgt;
 818         tdtd->tdtd_dt = &lod->lod_dt_dev;
 819         rc = distribute_txn_init(env, lut, tdtd,
 820                 lu_site2seq(lod2lu_dev(lod)->ld_site)->ss_node_id);
 821
 822         if (rc < 0) {
 823                 CERROR("%s: cannot init distribute txn: rc = %d\n",
 824                        lod2obd(lod)->obd_name, rc);
 825                 OBD_FREE_PTR(tdtd);
 826                 RETURN(rc);
 827         }
 828
 829         tdtd->tdtd_show_update_logs_retrievers =
 830                 lod_show_update_logs_retrievers;
 831         tdtd->tdtd_show_retrievers_cbdata = lod;
 832
 833         lut->lut_tdtd = tdtd;
 834
 835         RETURN(0);
 836 }
 837
 838 /**
 839  * Finish distribute txn
 840  *
 841  * Release the resource holding by distribute txn, i.e. stop distribute
 842  * txn thread.
 843  *
 844  * \param[in] env       execution environment
 845  * \param[in] lod       lod device
 846  */
 847 static void lod_fini_distribute_txn(const struct lu_env *env,
 848                                     struct lod_device *lod)
 849 {
 850         struct lu_target                  *lut;
 851
 852         lut = lod2lu_dev(lod)->ld_site->ls_tgt;
 853         target_recovery_fini(lut->lut_obd);
 854         if (lut->lut_tdtd == NULL)
 855                 return;
 856
 857         distribute_txn_fini(env, lut->lut_tdtd);
 858
 859         OBD_FREE_PTR(lut->lut_tdtd);
 860         lut->lut_tdtd = NULL;
 861 }
 862
 863 /**
 864  * Implementation of lu_device_operations::ldo_process_config() for LOD
 865  *
 866  * The method is called by the configuration subsystem during setup,
 867  * cleanup and when the configuration changes. The method processes
 868  * few specific commands like adding/removing the targets, changing
 869  * the runtime parameters.
 870
 871  * \param[in] env               LU environment provided by the caller
 872  * \param[in] dev               lod device
 873  * \param[in] lcfg              configuration command to apply
 874  *
 875  * \retval 0                    on success
 876  * \retval negative             negated errno on error
 877  *
 878  * The examples are below.
 879  *
 880  * Add osc config log:
 881  * marker  20 (flags=0x01, v2.2.49.56) lustre-OST0001  'add osc'
 882  * add_uuid  nid=192.168.122.162@tcp(0x20000c0a87aa2)  0:  1:nidxxx
 883  * attach    0:lustre-OST0001-osc-MDT0001  1:osc  2:lustre-MDT0001-mdtlov_UUID
 884  * setup     0:lustre-OST0001-osc-MDT0001  1:lustre-OST0001_UUID  2:nid
 885  * lov_modify_tgts add 0:lustre-MDT0001-mdtlov  1:lustre-OST0001_UUID  2:1  3:1
 886  * marker  20 (flags=0x02, v2.2.49.56) lustre-OST0001  'add osc'
 887  *
 888  * Add mdc config log:
 889  * marker  10 (flags=0x01, v2.2.49.56) lustre-MDT0000  'add osp'
 890  * add_uuid  nid=192.168.122.162@tcp(0x20000c0a87aa2)  0:  1:nid
 891  * attach 0:lustre-MDT0000-osp-MDT0001  1:osp  2:lustre-MDT0001-mdtlov_UUID
 892  * setup     0:lustre-MDT0000-osp-MDT0001  1:lustre-MDT0000_UUID  2:nid
 893  * modify_mdc_tgts add 0:lustre-MDT0001  1:lustre-MDT0000_UUID  2:0  3:1
 894  * marker  10 (flags=0x02, v2.2.49.56) lustre-MDT0000_UUID  'add osp'
 895  */
 896 static int lod_process_config(const struct lu_env *env,
 897                               struct lu_device *dev,
 898                               struct lustre_cfg *lcfg)
 899 {
 900         struct lod_device *lod = lu2lod_dev(dev);
 901         struct lu_device  *next = &lod->lod_child->dd_lu_dev;
 902         char              *arg1;
 903         int                rc = 0;
 904         ENTRY;
 905
 906         switch(lcfg->lcfg_command) {
 907         case LCFG_LOV_DEL_OBD:
 908         case LCFG_LOV_ADD_INA:
 909         case LCFG_LOV_ADD_OBD:
 910         case LCFG_ADD_MDC: {
 911                 __u32 index;
 912                 __u32 mdt_index;
 913                 int gen;
 914                 /* lov_modify_tgts add  0:lov_mdsA  1:osp  2:0  3:1
 915                  * modify_mdc_tgts add  0:lustre-MDT0001
 916                  *                    1:lustre-MDT0001-mdc0002
 917                  *                    2:2  3:1*/
 918                 arg1 = lustre_cfg_string(lcfg, 1);
 919
 920                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%d", &index) != 1)
 921                         GOTO(out, rc = -EINVAL);
 922                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
 923                         GOTO(out, rc = -EINVAL);
 924
 925                 if (lcfg->lcfg_command == LCFG_LOV_ADD_OBD) {
 926                         __u32 mdt_index;
 927
 928                         rc = lodname2mdt_index(lustre_cfg_string(lcfg, 0),
 929                                                &mdt_index);
 930                         if (rc != 0)
 931                                 GOTO(out, rc);
 932
 933                         rc = lod_add_device(env, lod, arg1, index, gen,
 934                                             mdt_index, LUSTRE_OSC_NAME, 1);
 935                 } else if (lcfg->lcfg_command == LCFG_ADD_MDC) {
 936                         mdt_index = index;
 937                         rc = lod_add_device(env, lod, arg1, index, gen,
 938                                             mdt_index, LUSTRE_MDC_NAME, 1);
 939                 } else if (lcfg->lcfg_command == LCFG_LOV_ADD_INA) {
 940                         /*FIXME: Add mdt_index for LCFG_LOV_ADD_INA*/
 941                         mdt_index = 0;
 942                         rc = lod_add_device(env, lod, arg1, index, gen,
 943                                             mdt_index, LUSTRE_OSC_NAME, 0);
 944                 } else {
 945                         rc = lod_del_device(env, lod,
 946                                             &lod->lod_ost_descs,
 947                                             arg1, index, gen, true);
 948                 }
 949
 950                 break;
 951         }
 952
 953         case LCFG_PARAM: {
 954                 struct obd_device *obd;
 955                 char *param;
 956
 957                 /* Check if it is activate/deactivate mdc
 958                  * lustre-MDTXXXX-osp-MDTXXXX.active=1 */
 959                 param = lustre_cfg_buf(lcfg, 1);
 960                 if (strstr(param, "osp") != NULL &&
 961                     strstr(param, ".active=") != NULL) {
 962                         struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
 963                         struct lod_tgt_desc     *sub_tgt = NULL;
 964                         char *ptr;
 965                         char *tmp;
 966                         int i;
 967
 968                         ptr = strstr(param, ".");
 969                         *ptr = '\0';
 970                         obd = class_name2obd(param);
 971                         if (obd == NULL) {
 972                                 CERROR("%s: can not find %s: rc = %d\n",
 973                                        lod2obd(lod)->obd_name, param, -EINVAL);
 974                                 *ptr = '.';
 975                                 GOTO(out, rc);
 976                         }
 977
 978                         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
 979                                 struct lod_tgt_desc *tgt;
 980
 981                                 tgt = LTD_TGT(ltd, i);
 982                                 if (tgt->ltd_tgt->dd_lu_dev.ld_obd == obd) {
 983                                         sub_tgt = tgt;
 984                                         break;
 985                                 }
 986                         }
 987
 988                         if (sub_tgt == NULL) {
 989                                 CERROR("%s: can not find %s: rc = %d\n",
 990                                        lod2obd(lod)->obd_name, param, -EINVAL);
 991                                 *ptr = '.';
 992                                 GOTO(out, rc);
 993                         }
 994
 995                         *ptr = '.';
 996                         tmp = strstr(param, "=");
 997                         tmp++;
 998                         if (*tmp == '1') {
 999                                 struct llog_ctxt *ctxt;
1000
1001                                 obd = sub_tgt->ltd_tgt->dd_lu_dev.ld_obd;
1002                                 ctxt = llog_get_context(obd,
1003                                                 LLOG_UPDATELOG_ORIG_CTXT);
1004                                 if (ctxt == NULL) {
1005                                         rc = llog_setup(env, obd, &obd->obd_olg,
1006                                                        LLOG_UPDATELOG_ORIG_CTXT,
1007                                                     NULL, &llog_common_cat_ops);
1008                                         if (rc < 0)
1009                                                 GOTO(out, rc);
1010                                 } else {
1011                                         llog_ctxt_put(ctxt);
1012                                 }
1013                                 rc = lod_sub_prep_llog(env, lod,
1014                                                        sub_tgt->ltd_tgt,
1015                                                        sub_tgt->ltd_index);
1016                                 if (rc == 0)
1017                                         sub_tgt->ltd_active = 1;
1018                         } else {
1019                                 lod_sub_fini_llog(env, sub_tgt->ltd_tgt,
1020                                                   NULL);
1021                                 sub_tgt->ltd_active = 0;
1022                         }
1023                         GOTO(out, rc);
1024                 }
1025
1026                 obd = lod2obd(lod);
1027                 rc = class_process_proc_param(PARAM_LOV, obd->obd_vars,
1028                                               lcfg, obd);
1029                 if (rc > 0)
1030                         rc = 0;
1031                 GOTO(out, rc);
1032         }
1033         case LCFG_PRE_CLEANUP: {
1034                 if (lod->lod_md_root != NULL) {
1035                         dt_object_put(env, &lod->lod_md_root->ldo_obj);
1036                         lod->lod_md_root = NULL;
1037                 }
1038
1039                 lod_sub_process_config(env, lod, &lod->lod_mdt_descs, lcfg);
1040                 lod_sub_process_config(env, lod, &lod->lod_ost_descs, lcfg);
1041                 next = &lod->lod_child->dd_lu_dev;
1042                 rc = next->ld_ops->ldo_process_config(env, next, lcfg);
1043                 if (rc != 0)
1044                         CDEBUG(D_HA, "%s: can't process %u: %d\n",
1045                                lod2obd(lod)->obd_name, lcfg->lcfg_command, rc);
1046
1047                 lod_sub_stop_recovery_threads(env, lod);
1048                 lod_fini_distribute_txn(env, lod);
1049                 lod_sub_fini_all_llogs(env, lod);
1050                 break;
1051         }
1052         case LCFG_CLEANUP: {
1053                 /*
1054                  * do cleanup on underlying storage only when
1055                  * all OSPs are cleaned up, as they use that OSD as well
1056                  */
1057                 lu_dev_del_linkage(dev->ld_site, dev);
1058                 lod_sub_process_config(env, lod, &lod->lod_mdt_descs, lcfg);
1059                 lod_sub_process_config(env, lod, &lod->lod_ost_descs, lcfg);
1060                 next = &lod->lod_child->dd_lu_dev;
1061                 rc = next->ld_ops->ldo_process_config(env, next, lcfg);
1062                 if (rc)
1063                         CERROR("%s: can't process %u: %d\n",
1064                                lod2obd(lod)->obd_name, lcfg->lcfg_command, rc);
1065
1066                 rc = obd_disconnect(lod->lod_child_exp);
1067                 if (rc)
1068                         CERROR("error in disconnect from storage: %d\n", rc);
1069                 break;
1070         }
1071         default:
1072                CERROR("%s: unknown command %u\n", lod2obd(lod)->obd_name,
1073                       lcfg->lcfg_command);
1074                rc = -EINVAL;
1075                break;
1076         }
1077
1078 out:
1079         RETURN(rc);
1080 }
1081
1082 /**
1083  * Implementation of lu_device_operations::ldo_recovery_complete() for LOD
1084  *
1085  * The method is called once the recovery is complete. This implementation
1086  * distributes the notification to all the known targets.
1087  *
1088  * see include/lu_object.h for the details
1089  */
1090 static int lod_recovery_complete(const struct lu_env *env,
1091                                  struct lu_device *dev)
1092 {
1093         struct lod_device   *lod = lu2lod_dev(dev);
1094         struct lu_device    *next = &lod->lod_child->dd_lu_dev;
1095         unsigned int         i;
1096         int                  rc;
1097         ENTRY;
1098
1099         LASSERT(lod->lod_recovery_completed == 0);
1100         lod->lod_recovery_completed = 1;
1101
1102         rc = next->ld_ops->ldo_recovery_complete(env, next);
1103
1104         lod_getref(&lod->lod_ost_descs);
1105         if (lod->lod_osts_size > 0) {
1106                 cfs_foreach_bit(lod->lod_ost_bitmap, i) {
1107                         struct lod_tgt_desc *tgt;
1108                         tgt = OST_TGT(lod, i);
1109                         LASSERT(tgt && tgt->ltd_tgt);
1110                         next = &tgt->ltd_ost->dd_lu_dev;
1111                         rc = next->ld_ops->ldo_recovery_complete(env, next);
1112                         if (rc)
1113                                 CERROR("%s: can't complete recovery on #%d:"
1114                                         "%d\n", lod2obd(lod)->obd_name, i, rc);
1115                 }
1116         }
1117         lod_putref(lod, &lod->lod_ost_descs);
1118         RETURN(rc);
1119 }
1120
1121 /**
1122  * Init update logs on all sub device
1123  *
1124  * LOD initialize update logs on all of sub devices. Because the initialization
1125  * process might need FLD lookup, see llog_osd_open()->dt_locate()->...->
1126  * lod_object_init(), this API has to be called after LOD is initialized.
1127  * \param[in] env       execution environment
1128  * \param[in] lod       lod device
1129  *
1130  * \retval              0 if update log is initialized successfully.
1131  * \retval              negative errno if initialization fails.
1132  */
1133 static int lod_sub_init_llogs(const struct lu_env *env, struct lod_device *lod)
1134 {
1135         struct lod_tgt_descs    *ltd = &lod->lod_mdt_descs;
1136         int                     rc;
1137         unsigned int            i;
1138         ENTRY;
1139
1140         /* llog must be setup after LOD is initialized, because llog
1141          * initialization include FLD lookup */
1142         LASSERT(lod->lod_initialized);
1143
1144         /* Init the llog in its own stack */
1145         rc = lod_sub_init_llog(env, lod, lod->lod_child);
1146         if (rc < 0)
1147                 RETURN(rc);
1148
1149         cfs_foreach_bit(ltd->ltd_tgt_bitmap, i) {
1150                 struct lod_tgt_desc     *tgt;
1151
1152                 tgt = LTD_TGT(ltd, i);
1153                 rc = lod_sub_init_llog(env, lod, tgt->ltd_tgt);
1154                 if (rc != 0)
1155                         break;
1156         }
1157
1158         RETURN(rc);
1159 }
1160
1161 /**
1162  * Implementation of lu_device_operations::ldo_prepare() for LOD
1163  *
1164  * see include/lu_object.h for the details.
1165  */
1166 static int lod_prepare(const struct lu_env *env, struct lu_device *pdev,
1167                        struct lu_device *cdev)
1168 {
1169         struct lod_device       *lod = lu2lod_dev(cdev);
1170         struct lu_device        *next = &lod->lod_child->dd_lu_dev;
1171         struct lu_fid           *fid = &lod_env_info(env)->lti_fid;
1172         int                     rc;
1173         struct dt_object        *root;
1174         struct dt_object        *dto;
1175         __u32                   index;
1176         ENTRY;
1177
1178         rc = next->ld_ops->ldo_prepare(env, pdev, next);
1179         if (rc != 0) {
1180                 CERROR("%s: prepare bottom error: rc = %d\n",
1181                        lod2obd(lod)->obd_name, rc);
1182                 RETURN(rc);
1183         }
1184
1185         lod->lod_initialized = 1;
1186
1187         rc = dt_root_get(env, lod->lod_child, fid);
1188         if (rc < 0)
1189                 RETURN(rc);
1190
1191         root = dt_locate(env, lod->lod_child, fid);
1192         if (IS_ERR(root))
1193                 RETURN(PTR_ERR(root));
1194
1195         /* Create update log object */
1196         index = lu_site2seq(lod2lu_dev(lod)->ld_site)->ss_node_id;
1197         lu_update_log_fid(fid, index);
1198
1199         dto = local_file_find_or_create_with_fid(env, lod->lod_child,
1200                                                  fid, root,
1201                                                  lod_update_log_name,
1202                                                  S_IFREG | S_IRUGO | S_IWUSR);
1203         if (IS_ERR(dto))
1204                 GOTO(out_put, rc = PTR_ERR(dto));
1205
1206         dt_object_put(env, dto);
1207
1208         /* Create update log dir */
1209         lu_update_log_dir_fid(fid, index);
1210         dto = local_file_find_or_create_with_fid(env, lod->lod_child,
1211                                                  fid, root,
1212                                                  lod_update_log_dir_name,
1213                                                  S_IFDIR | S_IRUGO | S_IWUSR);
1214         if (IS_ERR(dto))
1215                 GOTO(out_put, rc = PTR_ERR(dto));
1216
1217         dt_object_put(env, dto);
1218
1219         rc = lod_prepare_distribute_txn(env, lod);
1220         if (rc != 0)
1221                 GOTO(out_put, rc);
1222
1223         rc = lod_sub_init_llogs(env, lod);
1224         if (rc != 0)
1225                 GOTO(out_put, rc);
1226
1227 out_put:
1228         dt_object_put(env, root);
1229
1230         RETURN(rc);
1231 }
1232
1233 const struct lu_device_operations lod_lu_ops = {
1234         .ldo_object_alloc       = lod_object_alloc,
1235         .ldo_process_config     = lod_process_config,
1236         .ldo_recovery_complete  = lod_recovery_complete,
1237         .ldo_prepare            = lod_prepare,
1238 };
1239
1240 /**
1241  * Implementation of dt_device_operations::dt_root_get() for LOD
1242  *
1243  * see include/dt_object.h for the details.
1244  */
1245 static int lod_root_get(const struct lu_env *env,
1246                         struct dt_device *dev, struct lu_fid *f)
1247 {
1248         return dt_root_get(env, dt2lod_dev(dev)->lod_child, f);
1249 }
1250
1251 /**
1252  * Implementation of dt_device_operations::dt_statfs() for LOD
1253  *
1254  * see include/dt_object.h for the details.
1255  */
1256 static int lod_statfs(const struct lu_env *env,
1257                       struct dt_device *dev, struct obd_statfs *sfs)
1258 {
1259         return dt_statfs(env, dt2lod_dev(dev)->lod_child, sfs);
1260 }
1261
1262 /**
1263  * Implementation of dt_device_operations::dt_trans_create() for LOD
1264  *
1265  * Creates a transaction using local (to this node) OSD.
1266  *
1267  * see include/dt_object.h for the details.
1268  */
1269 static struct thandle *lod_trans_create(const struct lu_env *env,
1270                                         struct dt_device *dt)
1271 {
1272         struct thandle *th;
1273
1274         th = top_trans_create(env, dt2lod_dev(dt)->lod_child);
1275         if (IS_ERR(th))
1276                 return th;
1277
1278         th->th_dev = dt;
1279
1280         return th;
1281 }
1282
1283 /**
1284  * Implementation of dt_device_operations::dt_trans_start() for LOD
1285  *
1286  * Starts the set of local transactions using the targets involved
1287  * in declare phase. Initial support for the distributed transactions.
1288  *
1289  * see include/dt_object.h for the details.
1290  */
1291 static int lod_trans_start(const struct lu_env *env, struct dt_device *dt,
1292                            struct thandle *th)
1293 {
1294         return top_trans_start(env, dt2lod_dev(dt)->lod_child, th);
1295 }
1296
1297 static int lod_trans_cb_add(struct thandle *th,
1298                             struct dt_txn_commit_cb *dcb)
1299 {
1300         struct top_thandle      *top_th = container_of(th, struct top_thandle,
1301                                                        tt_super);
1302         return dt_trans_cb_add(top_th->tt_master_sub_thandle, dcb);
1303 }
1304
1305 /**
1306  * add noop update to the update records
1307  *
1308  * Add noop updates to the update records, which is only used in
1309  * test right now.
1310  *
1311  * \param[in] env       execution environment
1312  * \param[in] dt        dt device of lod
1313  * \param[in] th        thandle
1314  * \param[in] count     the count of update records to be added.
1315  *
1316  * \retval              0 if adding succeeds.
1317  * \retval              negative errno if adding fails.
1318  */
1319 static int lod_add_noop_records(const struct lu_env *env,
1320                                 struct dt_device *dt, struct thandle *th,
1321                                 int count)
1322 {
1323         struct top_thandle *top_th;
1324         struct lu_fid *fid = &lod_env_info(env)->lti_fid;
1325         int i;
1326         int rc = 0;
1327
1328         top_th = container_of(th, struct top_thandle, tt_super);
1329         if (top_th->tt_multiple_thandle == NULL)
1330                 return 0;
1331
1332         fid_zero(fid);
1333         for (i = 0; i < count; i++) {
1334                 rc = update_record_pack(noop, th, fid);
1335                 if (rc < 0)
1336                         return rc;
1337         }
1338         return rc;
1339 }
1340
1341 /**
1342  * Implementation of dt_device_operations::dt_trans_stop() for LOD
1343  *
1344  * Stops the set of local transactions using the targets involved
1345  * in declare phase. Initial support for the distributed transactions.
1346  *
1347  * see include/dt_object.h for the details.
1348  */
1349 static int lod_trans_stop(const struct lu_env *env, struct dt_device *dt,
1350                           struct thandle *th)
1351 {
1352         if (OBD_FAIL_CHECK(OBD_FAIL_SPLIT_UPDATE_REC)) {
1353                 int rc;
1354
1355                 rc = lod_add_noop_records(env, dt, th, 5000);
1356                 if (rc < 0)
1357                         RETURN(rc);
1358         }
1359         return top_trans_stop(env, dt2lod_dev(dt)->lod_child, th);
1360 }
1361
1362 /**
1363  * Implementation of dt_device_operations::dt_conf_get() for LOD
1364  *
1365  * Currently returns the configuration provided by the local OSD.
1366  *
1367  * see include/dt_object.h for the details.
1368  */
1369 static void lod_conf_get(const struct lu_env *env,
1370                          const struct dt_device *dev,
1371                          struct dt_device_param *param)
1372 {
1373         dt_conf_get(env, dt2lod_dev((struct dt_device *)dev)->lod_child, param);
1374 }
1375
1376 /**
1377  * Implementation of dt_device_operations::dt_sync() for LOD
1378  *
1379  * Syncs all known OST targets. Very very expensive and used
1380  * rarely by LFSCK now. Should not be used in general.
1381  *
1382  * see include/dt_object.h for the details.
1383  */
1384 static int lod_sync(const struct lu_env *env, struct dt_device *dev)
1385 {
1386         struct lod_device   *lod = dt2lod_dev(dev);
1387         struct lod_ost_desc *ost;
1388         struct lod_mdt_desc *mdt;
1389         unsigned int         i;
1390         int                  rc = 0;
1391         ENTRY;
1392
1393         lod_getref(&lod->lod_ost_descs);
1394         lod_foreach_ost(lod, i) {
1395                 ost = OST_TGT(lod, i);
1396                 LASSERT(ost && ost->ltd_ost);
1397                 rc = dt_sync(env, ost->ltd_ost);
1398                 if (rc) {
1399                         CERROR("%s: can't sync ost %u: %d\n",
1400                                lod2obd(lod)->obd_name, i, rc);
1401                         break;
1402                 }
1403         }
1404         lod_putref(lod, &lod->lod_ost_descs);
1405
1406         if (rc)
1407                 RETURN(rc);
1408
1409         lod_getref(&lod->lod_mdt_descs);
1410         lod_foreach_mdt(lod, i) {
1411                 mdt = MDT_TGT(lod, i);
1412                 LASSERT(mdt && mdt->ltd_mdt);
1413                 rc = dt_sync(env, mdt->ltd_mdt);
1414                 if (rc) {
1415                         CERROR("%s: can't sync mdt %u: %d\n",
1416                                lod2obd(lod)->obd_name, i, rc);
1417                         break;
1418                 }
1419         }
1420         lod_putref(lod, &lod->lod_mdt_descs);
1421
1422         if (rc == 0)
1423                 rc = dt_sync(env, lod->lod_child);
1424
1425         RETURN(rc);
1426 }
1427
1428 /**
1429  * Implementation of dt_device_operations::dt_ro() for LOD
1430  *
1431  * Turns local OSD read-only, used for the testing only.
1432  *
1433  * see include/dt_object.h for the details.
1434  */
1435 static int lod_ro(const struct lu_env *env, struct dt_device *dev)
1436 {
1437         return dt_ro(env, dt2lod_dev(dev)->lod_child);
1438 }
1439
1440 /**
1441  * Implementation of dt_device_operations::dt_commit_async() for LOD
1442  *
1443  * Asks local OSD to commit sooner.
1444  *
1445  * see include/dt_object.h for the details.
1446  */
1447 static int lod_commit_async(const struct lu_env *env, struct dt_device *dev)
1448 {
1449         return dt_commit_async(env, dt2lod_dev(dev)->lod_child);
1450 }
1451
1452 static const struct dt_device_operations lod_dt_ops = {
1453         .dt_root_get         = lod_root_get,
1454         .dt_statfs           = lod_statfs,
1455         .dt_trans_create     = lod_trans_create,
1456         .dt_trans_start      = lod_trans_start,
1457         .dt_trans_stop       = lod_trans_stop,
1458         .dt_conf_get         = lod_conf_get,
1459         .dt_sync             = lod_sync,
1460         .dt_ro               = lod_ro,
1461         .dt_commit_async     = lod_commit_async,
1462         .dt_trans_cb_add     = lod_trans_cb_add,
1463 };
1464
1465 /**
1466  * Connect to a local OSD.
1467  *
1468  * Used to connect to the local OSD at mount. OSD name is taken from the
1469  * configuration command passed. This connection is used to identify LU
1470  * site and pin the OSD from early removal.
1471  *
1472  * \param[in] env               LU environment provided by the caller
1473  * \param[in] lod               lod device
1474  * \param[in] cfg               configuration command to apply
1475  *
1476  * \retval 0                    on success
1477  * \retval negative             negated errno on error
1478  **/
1479 static int lod_connect_to_osd(const struct lu_env *env, struct lod_device *lod,
1480                               struct lustre_cfg *cfg)
1481 {
1482         struct obd_connect_data *data = NULL;
1483         struct obd_device       *obd;
1484         char                    *nextdev = NULL, *p, *s;
1485         int                      rc, len = 0;
1486         ENTRY;
1487
1488         LASSERT(cfg);
1489         LASSERT(lod->lod_child_exp == NULL);
1490
1491         /* compatibility hack: we still use old config logs
1492          * which specify LOV, but we need to learn underlying
1493          * OSD device, which is supposed to be:
1494          *  <fsname>-MDTxxxx-osd
1495          *
1496          * 2.x MGS generates lines like the following:
1497          *   #03 (176)lov_setup 0:lustre-MDT0000-mdtlov  1:(struct lov_desc)
1498          * 1.8 MGS generates lines like the following:
1499          *   #03 (168)lov_setup 0:lustre-mdtlov  1:(struct lov_desc)
1500          *
1501          * we use "-MDT" to differentiate 2.x from 1.8 */
1502
1503         if ((p = lustre_cfg_string(cfg, 0)) && strstr(p, "-mdtlov")) {
1504                 len = strlen(p) + 6;
1505                 OBD_ALLOC(nextdev, len);
1506                 if (nextdev == NULL)
1507                         GOTO(out, rc = -ENOMEM);
1508
1509                 strcpy(nextdev, p);
1510                 s = strstr(nextdev, "-mdtlov");
1511                 if (unlikely(s == NULL)) {
1512                         CERROR("unable to parse device name %s\n",
1513                                lustre_cfg_string(cfg, 0));
1514                         GOTO(out, rc = -EINVAL);
1515                 }
1516
1517                 if (strstr(nextdev, "-MDT")) {
1518                         /* 2.x config */
1519                         strcpy(s, "-osd");
1520                 } else {
1521                         /* 1.8 config */
1522                         strcpy(s, "-MDT0000-osd");
1523                 }
1524         } else {
1525                 CERROR("unable to parse device name %s\n",
1526                        lustre_cfg_string(cfg, 0));
1527                 GOTO(out, rc = -EINVAL);
1528         }
1529
1530         OBD_ALLOC_PTR(data);
1531         if (data == NULL)
1532                 GOTO(out, rc = -ENOMEM);
1533
1534         obd = class_name2obd(nextdev);
1535         if (obd == NULL) {
1536                 CERROR("can not locate next device: %s\n", nextdev);
1537                 GOTO(out, rc = -ENOTCONN);
1538         }
1539
1540         data->ocd_connect_flags = OBD_CONNECT_VERSION;
1541         data->ocd_version = LUSTRE_VERSION_CODE;
1542
1543         rc = obd_connect(env, &lod->lod_child_exp, obd, &obd->obd_uuid,
1544                          data, NULL);
1545         if (rc) {
1546                 CERROR("cannot connect to next dev %s (%d)\n", nextdev, rc);
1547                 GOTO(out, rc);
1548         }
1549
1550         lod->lod_dt_dev.dd_lu_dev.ld_site =
1551                 lod->lod_child_exp->exp_obd->obd_lu_dev->ld_site;
1552         LASSERT(lod->lod_dt_dev.dd_lu_dev.ld_site);
1553         lod->lod_child = lu2dt_dev(lod->lod_child_exp->exp_obd->obd_lu_dev);
1554
1555 out:
1556         if (data)
1557                 OBD_FREE_PTR(data);
1558         if (nextdev)
1559                 OBD_FREE(nextdev, len);
1560         RETURN(rc);
1561 }
1562
1563 /**
1564  * Allocate and initialize target table.
1565  *
1566  * A helper function to initialize the target table and allocate
1567  * a bitmap of the available targets.
1568  *
1569  * \param[in] ltd               target's table to initialize
1570  *
1571  * \retval 0                    on success
1572  * \retval negative             negated errno on error
1573  **/
1574 static int lod_tgt_desc_init(struct lod_tgt_descs *ltd)
1575 {
1576         mutex_init(&ltd->ltd_mutex);
1577         init_rwsem(&ltd->ltd_rw_sem);
1578
1579         /* the OST array and bitmap are allocated/grown dynamically as OSTs are
1580          * added to the LOD, see lod_add_device() */
1581         ltd->ltd_tgt_bitmap = CFS_ALLOCATE_BITMAP(32);
1582         if (ltd->ltd_tgt_bitmap == NULL)
1583                 RETURN(-ENOMEM);
1584
1585         ltd->ltd_tgts_size  = 32;
1586         ltd->ltd_tgtnr      = 0;
1587
1588         ltd->ltd_death_row = 0;
1589         ltd->ltd_refcount  = 0;
1590         return 0;
1591 }
1592
1593 /**
1594  * Initialize LOD device at setup.
1595  *
1596  * Initializes the given LOD device using the original configuration command.
1597  * The function initiates a connection to the local OSD and initializes few
1598  * internal structures like pools, target tables, etc.
1599  *
1600  * \param[in] env               LU environment provided by the caller
1601  * \param[in] lod               lod device
1602  * \param[in] ldt               not used
1603  * \param[in] cfg               configuration command
1604  *
1605  * \retval 0                    on success
1606  * \retval negative             negated errno on error
1607  **/
1608 static int lod_init0(const struct lu_env *env, struct lod_device *lod,
1609                      struct lu_device_type *ldt, struct lustre_cfg *cfg)
1610 {
1611         struct dt_device_param ddp;
1612         struct obd_device     *obd;
1613         int                    rc;
1614         ENTRY;
1615
1616         obd = class_name2obd(lustre_cfg_string(cfg, 0));
1617         if (obd == NULL) {
1618                 CERROR("Cannot find obd with name %s\n",
1619                        lustre_cfg_string(cfg, 0));
1620                 RETURN(-ENODEV);
1621         }
1622
1623         obd->obd_lu_dev = &lod->lod_dt_dev.dd_lu_dev;
1624         lod->lod_dt_dev.dd_lu_dev.ld_obd = obd;
1625         lod->lod_dt_dev.dd_lu_dev.ld_ops = &lod_lu_ops;
1626         lod->lod_dt_dev.dd_ops = &lod_dt_ops;
1627
1628         rc = lod_connect_to_osd(env, lod, cfg);
1629         if (rc)
1630                 RETURN(rc);
1631
1632         dt_conf_get(env, &lod->lod_dt_dev, &ddp);
1633         lod->lod_osd_max_easize = ddp.ddp_max_ea_size;
1634
1635         /* setup obd to be used with old lov code */
1636         rc = lod_pools_init(lod, cfg);
1637         if (rc)
1638                 GOTO(out_disconnect, rc);
1639
1640         rc = lod_procfs_init(lod);
1641         if (rc)
1642                 GOTO(out_pools, rc);
1643
1644         spin_lock_init(&lod->lod_lock);
1645         spin_lock_init(&lod->lod_connects_lock);
1646         lod_tgt_desc_init(&lod->lod_mdt_descs);
1647         lod_tgt_desc_init(&lod->lod_ost_descs);
1648
1649         RETURN(0);
1650
1651 out_pools:
1652         lod_pools_fini(lod);
1653 out_disconnect:
1654         obd_disconnect(lod->lod_child_exp);
1655         RETURN(rc);
1656 }
1657
1658 /**
1659  * Implementation of lu_device_type_operations::ldto_device_free() for LOD
1660  *
1661  * Releases the memory allocated for LOD device.
1662  *
1663  * see include/lu_object.h for the details.
1664  */
1665 static struct lu_device *lod_device_free(const struct lu_env *env,
1666                                          struct lu_device *lu)
1667 {
1668         struct lod_device *lod = lu2lod_dev(lu);
1669         struct lu_device  *next = &lod->lod_child->dd_lu_dev;
1670         ENTRY;
1671
1672         LASSERTF(atomic_read(&lu->ld_ref) == 0, "lu is %p\n", lu);
1673         dt_device_fini(&lod->lod_dt_dev);
1674         OBD_FREE_PTR(lod);
1675         RETURN(next);
1676 }
1677
1678 /**
1679  * Implementation of lu_device_type_operations::ldto_device_alloc() for LOD
1680  *
1681  * Allocates LOD device and calls the helpers to initialize it.
1682  *
1683  * see include/lu_object.h for the details.
1684  */
1685 static struct lu_device *lod_device_alloc(const struct lu_env *env,
1686                                           struct lu_device_type *type,
1687                                           struct lustre_cfg *lcfg)
1688 {
1689         struct lod_device *lod;
1690         struct lu_device  *lu_dev;
1691
1692         OBD_ALLOC_PTR(lod);
1693         if (lod == NULL) {
1694                 lu_dev = ERR_PTR(-ENOMEM);
1695         } else {
1696                 int rc;
1697
1698                 lu_dev = lod2lu_dev(lod);
1699                 dt_device_init(&lod->lod_dt_dev, type);
1700                 rc = lod_init0(env, lod, type, lcfg);
1701                 if (rc != 0) {
1702                         lod_device_free(env, lu_dev);
1703                         lu_dev = ERR_PTR(rc);
1704                 }
1705         }
1706
1707         return lu_dev;
1708 }
1709
1710 /**
1711  * Implementation of lu_device_type_operations::ldto_device_fini() for LOD
1712  *
1713  * Releases the internal resources used by LOD device.
1714  *
1715  * see include/lu_object.h for the details.
1716  */
1717 static struct lu_device *lod_device_fini(const struct lu_env *env,
1718                                          struct lu_device *d)
1719 {
1720         struct lod_device *lod = lu2lod_dev(d);
1721         int                rc;
1722         ENTRY;
1723
1724         lod_pools_fini(lod);
1725
1726         lod_procfs_fini(lod);
1727
1728         rc = lod_fini_tgt(env, lod, &lod->lod_ost_descs, true);
1729         if (rc)
1730                 CERROR("%s:can not fini ost descs %d\n",
1731                         lod2obd(lod)->obd_name, rc);
1732
1733         rc = lod_fini_tgt(env, lod, &lod->lod_mdt_descs, false);
1734         if (rc)
1735                 CERROR("%s:can not fini mdt descs %d\n",
1736                         lod2obd(lod)->obd_name, rc);
1737
1738         RETURN(NULL);
1739 }
1740
1741 /**
1742  * Implementation of obd_ops::o_connect() for LOD
1743  *
1744  * Used to track all the users of this specific LOD device,
1745  * so the device stays up until the last user disconnected.
1746  *
1747  * \param[in] env               LU environment provided by the caller
1748  * \param[out] exp              export the caller will be using to access LOD
1749  * \param[in] obd               OBD device representing LOD device
1750  * \param[in] cluuid            unique identifier of the caller
1751  * \param[in] data              not used
1752  * \param[in] localdata         not used
1753  *
1754  * \retval 0                    on success
1755  * \retval negative             negated errno on error
1756  **/
1757 static int lod_obd_connect(const struct lu_env *env, struct obd_export **exp,
1758                            struct obd_device *obd, struct obd_uuid *cluuid,
1759                            struct obd_connect_data *data, void *localdata)
1760 {
1761         struct lod_device    *lod = lu2lod_dev(obd->obd_lu_dev);
1762         struct lustre_handle  conn;
1763         int                   rc;
1764         ENTRY;
1765
1766         CDEBUG(D_CONFIG, "connect #%d\n", lod->lod_connects);
1767
1768         rc = class_connect(&conn, obd, cluuid);
1769         if (rc)
1770                 RETURN(rc);
1771
1772         *exp = class_conn2export(&conn);
1773
1774         spin_lock(&lod->lod_connects_lock);
1775         lod->lod_connects++;
1776         /* at the moment we expect the only user */
1777         LASSERT(lod->lod_connects == 1);
1778         spin_unlock(&lod->lod_connects_lock);
1779
1780         RETURN(0);
1781 }
1782
1783 /**
1784  *
1785  * Implementation of obd_ops::o_disconnect() for LOD
1786  *
1787  * When the caller doesn't need to use this LOD instance, it calls
1788  * obd_disconnect() and LOD releases corresponding export/reference count.
1789  * Once all the users gone, LOD device is released.
1790  *
1791  * \param[in] exp               export provided to the caller in obd_connect()
1792  *
1793  * \retval 0                    on success
1794  * \retval negative             negated errno on error
1795  **/
1796 static int lod_obd_disconnect(struct obd_export *exp)
1797 {
1798         struct obd_device *obd = exp->exp_obd;
1799         struct lod_device *lod = lu2lod_dev(obd->obd_lu_dev);
1800         int                rc, release = 0;
1801         ENTRY;
1802
1803         /* Only disconnect the underlying layers on the final disconnect. */
1804         spin_lock(&lod->lod_connects_lock);
1805         lod->lod_connects--;
1806         if (lod->lod_connects != 0) {
1807                 /* why should there be more than 1 connect? */
1808                 spin_unlock(&lod->lod_connects_lock);
1809                 CERROR("%s: disconnect #%d\n", exp->exp_obd->obd_name,
1810                        lod->lod_connects);
1811                 goto out;
1812         }
1813         spin_unlock(&lod->lod_connects_lock);
1814
1815         /* the last user of lod has gone, let's release the device */
1816         release = 1;
1817
1818 out:
1819         rc = class_disconnect(exp); /* bz 9811 */
1820
1821         if (rc == 0 && release)
1822                 class_manual_cleanup(obd);
1823         RETURN(rc);
1824 }
1825
1826 LU_KEY_INIT(lod, struct lod_thread_info);
1827
1828 static void lod_key_fini(const struct lu_context *ctx,
1829                 struct lu_context_key *key, void *data)
1830 {
1831         struct lod_thread_info *info = data;
1832         struct lod_layout_component *lds =
1833                                 info->lti_def_striping.lds_def_comp_entries;
1834         struct ost_pool *inuse = &info->lti_inuse_osts;
1835
1836         /* allocated in lod_get_lov_ea
1837          * XXX: this is overload, a tread may have such store but used only
1838          * once. Probably better would be pool of such stores per LOD.
1839          */
1840         if (info->lti_ea_store) {
1841                 OBD_FREE_LARGE(info->lti_ea_store, info->lti_ea_store_size);
1842                 info->lti_ea_store = NULL;
1843                 info->lti_ea_store_size = 0;
1844         }
1845         lu_buf_free(&info->lti_linkea_buf);
1846
1847         if (lds != NULL)
1848                 lod_free_def_comp_entries(&info->lti_def_striping);
1849
1850         if (inuse->op_size)
1851                 OBD_FREE(inuse->op_array, inuse->op_size);
1852
1853         OBD_FREE_PTR(info);
1854 }
1855
1856 /* context key: lod_thread_key */
1857 LU_CONTEXT_KEY_DEFINE(lod, LCT_MD_THREAD);
1858
1859 LU_TYPE_INIT_FINI(lod, &lod_thread_key);
1860
1861 static struct lu_device_type_operations lod_device_type_ops = {
1862         .ldto_init           = lod_type_init,
1863         .ldto_fini           = lod_type_fini,
1864
1865         .ldto_start          = lod_type_start,
1866         .ldto_stop           = lod_type_stop,
1867
1868         .ldto_device_alloc   = lod_device_alloc,
1869         .ldto_device_free    = lod_device_free,
1870
1871         .ldto_device_fini    = lod_device_fini
1872 };
1873
1874 static struct lu_device_type lod_device_type = {
1875         .ldt_tags     = LU_DEVICE_DT,
1876         .ldt_name     = LUSTRE_LOD_NAME,
1877         .ldt_ops      = &lod_device_type_ops,
1878         .ldt_ctx_tags = LCT_MD_THREAD,
1879 };
1880
1881 /**
1882  * Implementation of obd_ops::o_get_info() for LOD
1883  *
1884  * Currently, there is only one supported key: KEY_OSP_CONNECTED , to provide
1885  * the caller binary status whether LOD has seen connection to any OST target.
1886  * It will also check if the MDT update log context being initialized (if
1887  * needed).
1888  *
1889  * \param[in] env               LU environment provided by the caller
1890  * \param[in] exp               export of the caller
1891  * \param[in] keylen            len of the key
1892  * \param[in] key               the key
1893  * \param[in] vallen            not used
1894  * \param[in] val               not used
1895  *
1896  * \retval                      0 if a connection was seen
1897  * \retval                      -EAGAIN if LOD isn't running yet or no
1898  *                              connection has been seen yet
1899  * \retval                      -EINVAL if not supported key is requested
1900  **/
1901 static int lod_obd_get_info(const struct lu_env *env, struct obd_export *exp,
1902                             __u32 keylen, void *key, __u32 *vallen, void *val)
1903 {
1904         int rc = -EINVAL;
1905
1906         if (KEY_IS(KEY_OSP_CONNECTED)) {
1907                 struct obd_device       *obd = exp->exp_obd;
1908                 struct lod_device       *d;
1909                 struct lod_tgt_desc     *tgt;
1910                 unsigned int            i;
1911                 int                     rc = 1;
1912
1913                 if (!obd->obd_set_up || obd->obd_stopping)
1914                         RETURN(-EAGAIN);
1915
1916                 d = lu2lod_dev(obd->obd_lu_dev);
1917                 lod_getref(&d->lod_ost_descs);
1918                 lod_foreach_ost(d, i) {
1919                         tgt = OST_TGT(d, i);
1920                         LASSERT(tgt && tgt->ltd_tgt);
1921                         rc = obd_get_info(env, tgt->ltd_exp, keylen, key,
1922                                           vallen, val);
1923                         /* one healthy device is enough */
1924                         if (rc == 0)
1925                                 break;
1926                 }
1927                 lod_putref(d, &d->lod_ost_descs);
1928
1929                 lod_getref(&d->lod_mdt_descs);
1930                 lod_foreach_mdt(d, i) {
1931                         struct llog_ctxt *ctxt;
1932
1933                         tgt = MDT_TGT(d, i);
1934                         LASSERT(tgt != NULL);
1935                         LASSERT(tgt->ltd_tgt != NULL);
1936                         if (!tgt->ltd_active)
1937                                 continue;
1938
1939                         ctxt = llog_get_context(tgt->ltd_tgt->dd_lu_dev.ld_obd,
1940                                                 LLOG_UPDATELOG_ORIG_CTXT);
1941                         if (ctxt == NULL) {
1942                                 CDEBUG(D_INFO, "%s: %s is not ready.\n",
1943                                        obd->obd_name,
1944                                       tgt->ltd_tgt->dd_lu_dev.ld_obd->obd_name);
1945                                 rc = -EAGAIN;
1946                                 break;
1947                         }
1948                         if (ctxt->loc_handle == NULL) {
1949                                 CDEBUG(D_INFO, "%s: %s is not ready.\n",
1950                                        obd->obd_name,
1951                                       tgt->ltd_tgt->dd_lu_dev.ld_obd->obd_name);
1952                                 rc = -EAGAIN;
1953                                 llog_ctxt_put(ctxt);
1954                                 break;
1955                         }
1956                         llog_ctxt_put(ctxt);
1957                 }
1958                 lod_putref(d, &d->lod_mdt_descs);
1959
1960                 RETURN(rc);
1961         }
1962
1963         RETURN(rc);
1964 }
1965
1966 static int lod_obd_set_info_async(const struct lu_env *env,
1967                                   struct obd_export *exp,
1968                                   __u32 keylen, void *key,
1969                                   __u32 vallen, void *val,
1970                                   struct ptlrpc_request_set *set)
1971 {
1972         struct obd_device *obd = class_exp2obd(exp);
1973         struct lod_device *d;
1974         struct lod_tgt_desc *tgt;
1975         int no_set = 0;
1976         int i, rc = 0, rc2;
1977         ENTRY;
1978
1979         if (set == NULL) {
1980                 no_set = 1;
1981                 set = ptlrpc_prep_set();
1982                 if (!set)
1983                         RETURN(-ENOMEM);
1984         }
1985
1986         d = lu2lod_dev(obd->obd_lu_dev);
1987         lod_getref(&d->lod_ost_descs);
1988         lod_foreach_ost(d, i) {
1989                 tgt = OST_TGT(d, i);
1990                 LASSERT(tgt && tgt->ltd_tgt);
1991                 if (!tgt->ltd_active)
1992                         continue;
1993
1994                 rc2 = obd_set_info_async(env, tgt->ltd_exp, keylen, key,
1995                                          vallen, val, set);
1996                 if (rc2 != 0 && rc == 0)
1997                         rc = rc2;
1998         }
1999         lod_putref(d, &d->lod_ost_descs);
2000
2001         lod_getref(&d->lod_mdt_descs);
2002         lod_foreach_mdt(d, i) {
2003                 tgt = MDT_TGT(d, i);
2004                 LASSERT(tgt && tgt->ltd_tgt);
2005                 if (!tgt->ltd_active)
2006                         continue;
2007                 rc2 = obd_set_info_async(env, tgt->ltd_exp, keylen, key,
2008                                          vallen, val, set);
2009                 if (rc2 != 0 && rc == 0)
2010                         rc = rc2;
2011         }
2012         lod_putref(d, &d->lod_mdt_descs);
2013
2014
2015         if (no_set) {
2016                 rc2 = ptlrpc_set_wait(set);
2017                 if (rc2 == 0 && rc == 0)
2018                         rc = rc2;
2019                 ptlrpc_set_destroy(set);
2020         }
2021         RETURN(rc);
2022 }
2023
2024 static struct obd_ops lod_obd_device_ops = {
2025         .o_owner        = THIS_MODULE,
2026         .o_connect      = lod_obd_connect,
2027         .o_disconnect   = lod_obd_disconnect,
2028         .o_get_info     = lod_obd_get_info,
2029         .o_set_info_async = lod_obd_set_info_async,
2030         .o_pool_new     = lod_pool_new,
2031         .o_pool_rem     = lod_pool_remove,
2032         .o_pool_add     = lod_pool_add,
2033         .o_pool_del     = lod_pool_del,
2034 };
2035
2036 static int __init lod_init(void)
2037 {
2038         struct obd_type *type;
2039         int rc;
2040
2041         rc = lu_kmem_init(lod_caches);
2042         if (rc)
2043                 return rc;
2044
2045         rc = class_register_type(&lod_obd_device_ops, NULL, true, NULL,
2046                                  LUSTRE_LOD_NAME, &lod_device_type);
2047         if (rc) {
2048                 lu_kmem_fini(lod_caches);
2049                 return rc;
2050         }
2051
2052         /* create "lov" entry in procfs for compatibility purposes */
2053         type = class_search_type(LUSTRE_LOV_NAME);
2054         if (type != NULL && type->typ_procroot != NULL)
2055                 return rc;
2056
2057         type = class_search_type(LUSTRE_LOD_NAME);
2058         type->typ_procsym = lprocfs_register("lov", proc_lustre_root,
2059                                              NULL, NULL);
2060         if (IS_ERR(type->typ_procsym)) {
2061                 CERROR("lod: can't create compat entry \"lov\": %d\n",
2062                        (int)PTR_ERR(type->typ_procsym));
2063                 type->typ_procsym = NULL;
2064         }
2065         return rc;
2066 }
2067
2068 static void __exit lod_exit(void)
2069 {
2070         class_unregister_type(LUSTRE_LOD_NAME);
2071         lu_kmem_fini(lod_caches);
2072 }
2073
2074 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2075 MODULE_DESCRIPTION("Lustre Logical Object Device ("LUSTRE_LOD_NAME")");
2076 MODULE_VERSION(LUSTRE_VERSION_STRING);
2077 MODULE_LICENSE("GPL");
2078
2079 module_init(lod_init);
2080 module_exit(lod_exit);