lustre/lmv/lmv_obd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LMV
  41 #ifdef __KERNEL__
  42 #include <linux/slab.h>
  43 #include <linux/module.h>
  44 #include <linux/init.h>
  45 #include <linux/slab.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/mm.h>
  48 #include <asm/div64.h>
  49 #include <linux/seq_file.h>
  50 #include <linux/namei.h>
  51 #else
  52 #include <liblustre.h>
  53 #endif
  54
  55 #include <lustre/lustre_idl.h>
  56 #include <lustre_log.h>
  57 #include <obd_support.h>
  58 #include <lustre_lib.h>
  59 #include <lustre_net.h>
  60 #include <obd_class.h>
  61 #include <lprocfs_status.h>
  62 #include <lustre_lite.h>
  63 #include <lustre_fid.h>
  64 #include "lmv_internal.h"
  65
  66 /* object cache. */
  67 cfs_mem_cache_t *lmv_object_cache;
  68 cfs_atomic_t lmv_object_count = CFS_ATOMIC_INIT(0);
  69
  70 static void lmv_activate_target(struct lmv_obd *lmv,
  71                                 struct lmv_tgt_desc *tgt,
  72                                 int activate)
  73 {
  74         if (tgt->ltd_active == activate)
  75                 return;
  76
  77         tgt->ltd_active = activate;
  78         lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
  79 }
  80
  81 /**
  82  * Error codes:
  83  *
  84  *  -EINVAL  : UUID can't be found in the LMV's target list
  85  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  86  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  87  */
  88 static int lmv_set_mdc_active(struct lmv_obd *lmv, struct obd_uuid *uuid,
  89                               int activate)
  90 {
  91         struct lmv_tgt_desc    *tgt;
  92         struct obd_device      *obd;
  93         int                     i;
  94         int                     rc = 0;
  95         ENTRY;
  96
  97         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  98                lmv, uuid->uuid, activate);
  99
 100         cfs_spin_lock(&lmv->lmv_lock);
 101         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 102                 if (tgt->ltd_exp == NULL)
 103                         continue;
 104
 105                 CDEBUG(D_INFO, "Target idx %d is %s conn "LPX64"\n",
 106                        i, tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
 107
 108                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 109                         break;
 110         }
 111
 112         if (i == lmv->desc.ld_tgt_count)
 113                 GOTO(out_lmv_lock, rc = -EINVAL);
 114
 115         obd = class_exp2obd(tgt->ltd_exp);
 116         if (obd == NULL)
 117                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 118
 119         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 120                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 121                obd->obd_type->typ_name, i);
 122         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 123
 124         if (tgt->ltd_active == activate) {
 125                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 126                        activate ? "" : "in");
 127                 GOTO(out_lmv_lock, rc);
 128         }
 129
 130         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 131                activate ? "" : "in");
 132         lmv_activate_target(lmv, tgt, activate);
 133         EXIT;
 134
 135  out_lmv_lock:
 136         cfs_spin_unlock(&lmv->lmv_lock);
 137         return rc;
 138 }
 139
 140 static int lmv_set_mdc_data(struct lmv_obd *lmv, struct obd_uuid *uuid,
 141                             struct obd_connect_data *data)
 142 {
 143         struct lmv_tgt_desc    *tgt;
 144         int                     i;
 145         ENTRY;
 146
 147         LASSERT(data != NULL);
 148
 149         cfs_spin_lock(&lmv->lmv_lock);
 150         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 151                 if (tgt->ltd_exp == NULL)
 152                         continue;
 153
 154                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
 155                         lmv->datas[tgt->ltd_idx] = *data;
 156                         break;
 157                 }
 158         }
 159         cfs_spin_unlock(&lmv->lmv_lock);
 160         RETURN(0);
 161 }
 162
 163 struct obd_uuid *lmv_get_uuid(struct obd_export *exp) {
 164         struct obd_device *obd = exp->exp_obd;
 165         struct lmv_obd *lmv = &obd->u.lmv;
 166         return obd_get_uuid(lmv->tgts[0].ltd_exp);
 167 }
 168
 169 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 170                       enum obd_notify_event ev, void *data)
 171 {
 172         struct obd_connect_data *conn_data;
 173         struct lmv_obd          *lmv = &obd->u.lmv;
 174         struct obd_uuid         *uuid;
 175         int                      rc = 0;
 176         ENTRY;
 177
 178         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 179                 CERROR("unexpected notification of %s %s!\n",
 180                        watched->obd_type->typ_name,
 181                        watched->obd_name);
 182                 RETURN(-EINVAL);
 183         }
 184
 185         uuid = &watched->u.cli.cl_target_uuid;
 186         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 187                 /*
 188                  * Set MDC as active before notifying the observer, so the
 189                  * observer can use the MDC normally.
 190                  */
 191                 rc = lmv_set_mdc_active(lmv, uuid,
 192                                         ev == OBD_NOTIFY_ACTIVE);
 193                 if (rc) {
 194                         CERROR("%sactivation of %s failed: %d\n",
 195                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 196                                uuid->uuid, rc);
 197                         RETURN(rc);
 198                 }
 199         } else if (ev == OBD_NOTIFY_OCD) {
 200                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 201
 202                 /*
 203                  * Set connect data to desired target, update exp_connect_flags.
 204                  */
 205                 rc = lmv_set_mdc_data(lmv, uuid, conn_data);
 206                 if (rc) {
 207                         CERROR("can't set connect data to target %s, rc %d\n",
 208                                uuid->uuid, rc);
 209                         RETURN(rc);
 210                 }
 211
 212                 /*
 213                  * XXX: Make sure that ocd_connect_flags from all targets are
 214                  * the same. Otherwise one of MDTs runs wrong version or
 215                  * something like this.  --umka
 216                  */
 217                 obd->obd_self_export->exp_connect_flags =
 218                         conn_data->ocd_connect_flags;
 219         }
 220 #if 0
 221         else if (ev == OBD_NOTIFY_DISCON) {
 222                 /*
 223                  * For disconnect event, flush fld cache for failout MDS case.
 224                  */
 225                 fld_client_flush(&lmv->lmv_fld);
 226         }
 227 #endif
 228         /*
 229          * Pass the notification up the chain.
 230          */
 231         if (obd->obd_observer)
 232                 rc = obd_notify(obd->obd_observer, watched, ev, data);
 233
 234         RETURN(rc);
 235 }
 236
 237 /**
 238  * This is fake connect function. Its purpose is to initialize lmv and say
 239  * caller that everything is okay. Real connection will be performed later.
 240  */
 241 static int lmv_connect(const struct lu_env *env,
 242                        struct obd_export **exp, struct obd_device *obd,
 243                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 244                        void *localdata)
 245 {
 246 #ifdef __KERNEL__
 247         struct proc_dir_entry *lmv_proc_dir;
 248 #endif
 249         struct lmv_obd        *lmv = &obd->u.lmv;
 250         struct lustre_handle  conn = { 0 };
 251         int                    rc = 0;
 252         ENTRY;
 253
 254         /*
 255          * We don't want to actually do the underlying connections more than
 256          * once, so keep track.
 257          */
 258         lmv->refcount++;
 259         if (lmv->refcount > 1) {
 260                 *exp = NULL;
 261                 RETURN(0);
 262         }
 263
 264         rc = class_connect(&conn, obd, cluuid);
 265         if (rc) {
 266                 CERROR("class_connection() returned %d\n", rc);
 267                 RETURN(rc);
 268         }
 269
 270         *exp = class_conn2export(&conn);
 271         class_export_get(*exp);
 272
 273         lmv->exp = *exp;
 274         lmv->connected = 0;
 275         lmv->cluuid = *cluuid;
 276
 277         if (data)
 278                 lmv->conn_data = *data;
 279
 280 #ifdef __KERNEL__
 281         lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
 282                                         NULL, NULL);
 283         if (IS_ERR(lmv_proc_dir)) {
 284                 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
 285                        obd->obd_type->typ_name, obd->obd_name);
 286                 lmv_proc_dir = NULL;
 287         }
 288 #endif
 289
 290         /*
 291          * All real clients should perform actual connection right away, because
 292          * it is possible, that LMV will not have opportunity to connect targets
 293          * and MDC stuff will be called directly, for instance while reading
 294          * ../mdc/../kbytesfree procfs file, etc.
 295          */
 296         if (data->ocd_connect_flags & OBD_CONNECT_REAL)
 297                 rc = lmv_check_connect(obd);
 298
 299 #ifdef __KERNEL__
 300         if (rc) {
 301                 if (lmv_proc_dir)
 302                         lprocfs_remove(&lmv_proc_dir);
 303         }
 304 #endif
 305
 306         RETURN(rc);
 307 }
 308
 309 static void lmv_set_timeouts(struct obd_device *obd)
 310 {
 311         struct lmv_tgt_desc   *tgts;
 312         struct lmv_obd        *lmv;
 313         int                    i;
 314
 315         lmv = &obd->u.lmv;
 316         if (lmv->server_timeout == 0)
 317                 return;
 318
 319         if (lmv->connected == 0)
 320                 return;
 321
 322         for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
 323                 if (tgts->ltd_exp == NULL)
 324                         continue;
 325
 326                 obd_set_info_async(tgts->ltd_exp, sizeof(KEY_INTERMDS),
 327                                    KEY_INTERMDS, 0, NULL, NULL);
 328         }
 329 }
 330
 331 static int lmv_init_ea_size(struct obd_export *exp, int easize,
 332                             int def_easize, int cookiesize)
 333 {
 334         struct obd_device   *obd = exp->exp_obd;
 335         struct lmv_obd      *lmv = &obd->u.lmv;
 336         int                  i;
 337         int                  rc = 0;
 338         int                  change = 0;
 339         ENTRY;
 340
 341         if (lmv->max_easize < easize) {
 342                 lmv->max_easize = easize;
 343                 change = 1;
 344         }
 345         if (lmv->max_def_easize < def_easize) {
 346                 lmv->max_def_easize = def_easize;
 347                 change = 1;
 348         }
 349         if (lmv->max_cookiesize < cookiesize) {
 350                 lmv->max_cookiesize = cookiesize;
 351                 change = 1;
 352         }
 353         if (change == 0)
 354                 RETURN(0);
 355
 356         if (lmv->connected == 0)
 357                 RETURN(0);
 358
 359         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 360                 if (lmv->tgts[i].ltd_exp == NULL) {
 361                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
 362                         continue;
 363                 }
 364
 365                 rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize,
 366                                      cookiesize);
 367                 if (rc) {
 368                         CERROR("obd_init_ea_size() failed on MDT target %d, "
 369                                "error %d.\n", i, rc);
 370                         break;
 371                 }
 372         }
 373         RETURN(rc);
 374 }
 375
 376 #define MAX_STRING_SIZE 128
 377
 378 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 379 {
 380 #ifdef __KERNEL__
 381         struct proc_dir_entry   *lmv_proc_dir;
 382 #endif
 383         struct lmv_obd          *lmv = &obd->u.lmv;
 384         struct obd_uuid         *cluuid = &lmv->cluuid;
 385         struct obd_connect_data *mdc_data = NULL;
 386         struct obd_uuid          lmv_mdc_uuid = { "LMV_MDC_UUID" };
 387         struct obd_device       *mdc_obd;
 388         struct obd_export       *mdc_exp;
 389         struct lu_fld_target     target;
 390         int                      rc;
 391         ENTRY;
 392
 393         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 394                                         &obd->obd_uuid);
 395         if (!mdc_obd) {
 396                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 397                 RETURN(-EINVAL);
 398         }
 399
 400         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
 401                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 402                 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
 403                 cluuid->uuid);
 404
 405         if (!mdc_obd->obd_set_up) {
 406                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 407                 RETURN(-EINVAL);
 408         }
 409
 410         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &lmv_mdc_uuid,
 411                          &lmv->conn_data, NULL);
 412         if (rc) {
 413                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 414                 RETURN(rc);
 415         }
 416
 417         /*
 418          * Init fid sequence client for this mdc and add new fld target.
 419          */
 420         rc = obd_fid_init(mdc_exp);
 421         if (rc)
 422                 RETURN(rc);
 423
 424         target.ft_srv = NULL;
 425         target.ft_exp = mdc_exp;
 426         target.ft_idx = tgt->ltd_idx;
 427
 428         fld_client_add_target(&lmv->lmv_fld, &target);
 429
 430         mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data;
 431
 432         rc = obd_register_observer(mdc_obd, obd);
 433         if (rc) {
 434                 obd_disconnect(mdc_exp);
 435                 CERROR("target %s register_observer error %d\n",
 436                        tgt->ltd_uuid.uuid, rc);
 437                 RETURN(rc);
 438         }
 439
 440         if (obd->obd_observer) {
 441                 /*
 442                  * Tell the observer about the new target.
 443                  */
 444                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 445                                 OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts));
 446                 if (rc) {
 447                         obd_disconnect(mdc_exp);
 448                         RETURN(rc);
 449                 }
 450         }
 451
 452         tgt->ltd_active = 1;
 453         tgt->ltd_exp = mdc_exp;
 454         lmv->desc.ld_active_tgt_count++;
 455
 456         /*
 457          * Copy connect data, it may be used later.
 458          */
 459         lmv->datas[tgt->ltd_idx] = *mdc_data;
 460
 461         md_init_ea_size(tgt->ltd_exp, lmv->max_easize,
 462                         lmv->max_def_easize, lmv->max_cookiesize);
 463
 464         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 465                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 466                 cfs_atomic_read(&obd->obd_refcount));
 467
 468 #ifdef __KERNEL__
 469         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 470         if (lmv_proc_dir) {
 471                 struct proc_dir_entry *mdc_symlink;
 472
 473                 LASSERT(mdc_obd->obd_type != NULL);
 474                 LASSERT(mdc_obd->obd_type->typ_name != NULL);
 475                 mdc_symlink = lprocfs_add_symlink(mdc_obd->obd_name,
 476                                                   lmv_proc_dir,
 477                                                   "../../../%s/%s",
 478                                                   mdc_obd->obd_type->typ_name,
 479                                                   mdc_obd->obd_name);
 480                 if (mdc_symlink == NULL) {
 481                         CERROR("Could not register LMV target "
 482                                "/proc/fs/lustre/%s/%s/target_obds/%s.",
 483                                obd->obd_type->typ_name, obd->obd_name,
 484                                mdc_obd->obd_name);
 485                         lprocfs_remove(&lmv_proc_dir);
 486                         lmv_proc_dir = NULL;
 487                 }
 488         }
 489 #endif
 490         RETURN(0);
 491 }
 492
 493 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
 494 {
 495         struct lmv_obd      *lmv = &obd->u.lmv;
 496         struct lmv_tgt_desc *tgt;
 497         int                  rc = 0;
 498         ENTRY;
 499
 500         CDEBUG(D_CONFIG, "Target uuid: %s.\n", tgt_uuid->uuid);
 501
 502         lmv_init_lock(lmv);
 503
 504         if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
 505                 lmv_init_unlock(lmv);
 506                 CERROR("Can't add %s, LMV module compiled for %d MDCs. "
 507                        "That many MDCs already configured.\n",
 508                        tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
 509                 RETURN(-EINVAL);
 510         }
 511         if (lmv->desc.ld_tgt_count == 0) {
 512                 struct obd_device *mdc_obd;
 513
 514                 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
 515                                                 &obd->obd_uuid);
 516                 if (!mdc_obd) {
 517                         lmv_init_unlock(lmv);
 518                         CERROR("Target %s not attached\n", tgt_uuid->uuid);
 519                         RETURN(-EINVAL);
 520                 }
 521
 522                 rc = obd_llog_init(obd, &obd->obd_olg, mdc_obd, NULL);
 523                 if (rc) {
 524                         lmv_init_unlock(lmv);
 525                         CERROR("lmv failed to setup llogging subsystems\n");
 526                 }
 527         }
 528         cfs_spin_lock(&lmv->lmv_lock);
 529         tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
 530         tgt->ltd_uuid = *tgt_uuid;
 531         cfs_spin_unlock(&lmv->lmv_lock);
 532
 533         if (lmv->connected) {
 534                 rc = lmv_connect_mdc(obd, tgt);
 535                 if (rc) {
 536                         cfs_spin_lock(&lmv->lmv_lock);
 537                         lmv->desc.ld_tgt_count--;
 538                         memset(tgt, 0, sizeof(*tgt));
 539                         cfs_spin_unlock(&lmv->lmv_lock);
 540                 } else {
 541                         int easize = sizeof(struct lmv_stripe_md) +
 542                                      lmv->desc.ld_tgt_count *
 543                                      sizeof(struct lu_fid);
 544                         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 545                 }
 546         }
 547
 548         lmv_init_unlock(lmv);
 549         RETURN(rc);
 550 }
 551
 552 int lmv_check_connect(struct obd_device *obd)
 553 {
 554         struct lmv_obd       *lmv = &obd->u.lmv;
 555         struct lmv_tgt_desc  *tgt;
 556         int                   i;
 557         int                   rc;
 558         int                   easize;
 559         ENTRY;
 560
 561         if (lmv->connected)
 562                 RETURN(0);
 563
 564         lmv_init_lock(lmv);
 565         if (lmv->connected) {
 566                 lmv_init_unlock(lmv);
 567                 RETURN(0);
 568         }
 569
 570         if (lmv->desc.ld_tgt_count == 0) {
 571                 CERROR("%s: no targets configured.\n", obd->obd_name);
 572                 RETURN(-EINVAL);
 573         }
 574
 575         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 576                lmv->cluuid.uuid, obd->obd_name);
 577
 578         LASSERT(lmv->tgts != NULL);
 579
 580         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 581                 rc = lmv_connect_mdc(obd, tgt);
 582                 if (rc)
 583                         GOTO(out_disc, rc);
 584         }
 585
 586         lmv_set_timeouts(obd);
 587         class_export_put(lmv->exp);
 588         lmv->connected = 1;
 589         easize = lmv_get_easize(lmv);
 590         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 591         lmv_init_unlock(lmv);
 592         RETURN(0);
 593
 594  out_disc:
 595         while (i-- > 0) {
 596                 int rc2;
 597                 --tgt;
 598                 tgt->ltd_active = 0;
 599                 if (tgt->ltd_exp) {
 600                         --lmv->desc.ld_active_tgt_count;
 601                         rc2 = obd_disconnect(tgt->ltd_exp);
 602                         if (rc2) {
 603                                 CERROR("LMV target %s disconnect on "
 604                                        "MDC idx %d: error %d\n",
 605                                        tgt->ltd_uuid.uuid, i, rc2);
 606                         }
 607                 }
 608         }
 609         class_disconnect(lmv->exp);
 610         lmv_init_unlock(lmv);
 611         RETURN(rc);
 612 }
 613
 614 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 615 {
 616 #ifdef __KERNEL__
 617         struct proc_dir_entry  *lmv_proc_dir;
 618 #endif
 619         struct lmv_obd         *lmv = &obd->u.lmv;
 620         struct obd_device      *mdc_obd;
 621         int                     rc;
 622         ENTRY;
 623
 624         LASSERT(tgt != NULL);
 625         LASSERT(obd != NULL);
 626
 627         mdc_obd = class_exp2obd(tgt->ltd_exp);
 628
 629         if (mdc_obd)
 630                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 631
 632 #ifdef __KERNEL__
 633         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 634         if (lmv_proc_dir) {
 635                 struct proc_dir_entry *mdc_symlink;
 636
 637                 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
 638                 if (mdc_symlink) {
 639                         lprocfs_remove(&mdc_symlink);
 640                 } else {
 641                         CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
 642                                obd->obd_type->typ_name, obd->obd_name,
 643                                mdc_obd->obd_name);
 644                 }
 645         }
 646 #endif
 647         rc = obd_fid_fini(tgt->ltd_exp);
 648         if (rc)
 649                 CERROR("Can't finanize fids factory\n");
 650
 651         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 652                tgt->ltd_exp->exp_obd->obd_name,
 653                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 654
 655         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 656         rc = obd_disconnect(tgt->ltd_exp);
 657         if (rc) {
 658                 if (tgt->ltd_active) {
 659                         CERROR("Target %s disconnect error %d\n",
 660                                tgt->ltd_uuid.uuid, rc);
 661                 }
 662         }
 663
 664         lmv_activate_target(lmv, tgt, 0);
 665         tgt->ltd_exp = NULL;
 666         RETURN(0);
 667 }
 668
 669 static int lmv_disconnect(struct obd_export *exp)
 670 {
 671         struct obd_device     *obd = class_exp2obd(exp);
 672 #ifdef __KERNEL__
 673         struct proc_dir_entry *lmv_proc_dir;
 674 #endif
 675         struct lmv_obd        *lmv = &obd->u.lmv;
 676         int                    rc;
 677         int                    i;
 678         ENTRY;
 679
 680         if (!lmv->tgts)
 681                 goto out_local;
 682
 683         /*
 684          * Only disconnect the underlying layers on the final disconnect.
 685          */
 686         lmv->refcount--;
 687         if (lmv->refcount != 0)
 688                 goto out_local;
 689
 690         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 691                 if (lmv->tgts[i].ltd_exp == NULL)
 692                         continue;
 693                 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
 694         }
 695
 696 #ifdef __KERNEL__
 697         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 698         if (lmv_proc_dir) {
 699                 lprocfs_remove(&lmv_proc_dir);
 700         } else {
 701                 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
 702                        obd->obd_type->typ_name, obd->obd_name);
 703         }
 704 #endif
 705
 706 out_local:
 707         /*
 708          * This is the case when no real connection is established by
 709          * lmv_check_connect().
 710          */
 711         if (!lmv->connected)
 712                 class_export_put(exp);
 713         rc = class_disconnect(exp);
 714         if (lmv->refcount == 0)
 715                 lmv->connected = 0;
 716         RETURN(rc);
 717 }
 718
 719 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 720                          int len, void *karg, void *uarg)
 721 {
 722         struct obd_device    *obddev = class_exp2obd(exp);
 723         struct lmv_obd       *lmv = &obddev->u.lmv;
 724         int                   i = 0;
 725         int                   rc = 0;
 726         int                   set = 0;
 727         int                   count = lmv->desc.ld_tgt_count;
 728         ENTRY;
 729
 730         if (count == 0)
 731                 RETURN(-ENOTTY);
 732
 733         switch (cmd) {
 734         case IOC_OBD_STATFS: {
 735                 struct obd_ioctl_data *data = karg;
 736                 struct obd_device *mdc_obd;
 737                 struct obd_statfs stat_buf = {0};
 738                 __u32 index;
 739
 740                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 741                 if ((index >= count))
 742                         RETURN(-ENODEV);
 743
 744                 if (!lmv->tgts[index].ltd_active)
 745                         RETURN(-ENODATA);
 746
 747                 mdc_obd = class_exp2obd(lmv->tgts[index].ltd_exp);
 748                 if (!mdc_obd)
 749                         RETURN(-EINVAL);
 750
 751                 rc = obd_statfs(mdc_obd, &stat_buf,
 752                                 cfs_time_current_64() - CFS_HZ, 0);
 753                 if (rc)
 754                         RETURN(rc);
 755                 if (cfs_copy_to_user(data->ioc_pbuf1, &stat_buf,
 756                                      data->ioc_plen1))
 757                         RETURN(-EFAULT);
 758                 if (cfs_copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 759                                      data->ioc_plen2))
 760                         RETURN(-EFAULT);
 761                 break;
 762         }
 763         case OBD_IOC_QUOTACTL: {
 764                 struct if_quotactl *qctl = karg;
 765                 struct lmv_tgt_desc *tgt = NULL;
 766                 struct obd_quotactl *oqctl;
 767
 768                 if (qctl->qc_valid == QC_MDTIDX) {
 769                         if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
 770                                 RETURN(-EINVAL);
 771
 772                         tgt = &lmv->tgts[qctl->qc_idx];
 773                         if (!tgt->ltd_exp)
 774                                 RETURN(-EINVAL);
 775                 } else if (qctl->qc_valid == QC_UUID) {
 776                         for (i = 0; i < count; i++) {
 777                                 tgt = &lmv->tgts[i];
 778                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 779                                                      &qctl->obd_uuid))
 780                                         continue;
 781
 782                                 if (tgt->ltd_exp == NULL)
 783                                         RETURN(-EINVAL);
 784
 785                                 break;
 786                         }
 787                 } else {
 788                         RETURN(-EINVAL);
 789                 }
 790
 791                 if (i >= count)
 792                         RETURN(-EAGAIN);
 793
 794                 LASSERT(tgt && tgt->ltd_exp);
 795                 OBD_ALLOC_PTR(oqctl);
 796                 if (!oqctl)
 797                         RETURN(-ENOMEM);
 798
 799                 QCTL_COPY(oqctl, qctl);
 800                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 801                 if (rc == 0) {
 802                         QCTL_COPY(qctl, oqctl);
 803                         qctl->qc_valid = QC_MDTIDX;
 804                         qctl->obd_uuid = tgt->ltd_uuid;
 805                 }
 806                 OBD_FREE_PTR(oqctl);
 807                 break;
 808         }
 809         case OBD_IOC_CHANGELOG_SEND:
 810         case OBD_IOC_CHANGELOG_CLEAR: {
 811                 struct ioc_changelog *icc = karg;
 812
 813                 if (icc->icc_mdtindex >= count)
 814                         RETURN(-ENODEV);
 815
 816                 rc = obd_iocontrol(cmd, lmv->tgts[icc->icc_mdtindex].ltd_exp,
 817                                    sizeof(*icc), icc, NULL);
 818                 break;
 819         }
 820
 821         default : {
 822                 for (i = 0; i < count; i++) {
 823                         int err;
 824
 825                         if (lmv->tgts[i].ltd_exp == NULL)
 826                                 continue;
 827
 828                         err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
 829                                             karg, uarg);
 830                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
 831                                 RETURN(err);
 832                         } else if (err) {
 833                                 if (lmv->tgts[i].ltd_active) {
 834                                         CERROR("error: iocontrol MDC %s on MDT"
 835                                                "idx %d cmd %x: err = %d\n",
 836                                                 lmv->tgts[i].ltd_uuid.uuid,
 837                                                 i, cmd, err);
 838                                         if (!rc)
 839                                                 rc = err;
 840                                 }
 841                         } else
 842                                 set = 1;
 843                 }
 844                 if (!set && !rc)
 845                         rc = -EIO;
 846         }
 847         }
 848         RETURN(rc);
 849 }
 850
 851 static int lmv_all_chars_policy(int count, const char *name,
 852                                 int len)
 853 {
 854         unsigned int c = 0;
 855
 856         while (len > 0)
 857                 c += name[--len];
 858         c = c % count;
 859         return c;
 860 }
 861
 862 static int lmv_nid_policy(struct lmv_obd *lmv)
 863 {
 864         struct obd_import *imp;
 865         __u32              id;
 866
 867         /*
 868          * XXX: To get nid we assume that underlying obd device is mdc.
 869          */
 870         imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
 871         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
 872         return id % lmv->desc.ld_tgt_count;
 873 }
 874
 875 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
 876                           placement_policy_t placement)
 877 {
 878         switch (placement) {
 879         case PLACEMENT_CHAR_POLICY:
 880                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
 881                                             op_data->op_name,
 882                                             op_data->op_namelen);
 883         case PLACEMENT_NID_POLICY:
 884                 return lmv_nid_policy(lmv);
 885
 886         default:
 887                 break;
 888         }
 889
 890         CERROR("Unsupported placement policy %x\n", placement);
 891         return -EINVAL;
 892 }
 893
 894 /**
 895  * This is _inode_ placement policy function (not name).
 896  */
 897 static int lmv_placement_policy(struct obd_device *obd,
 898                                 struct md_op_data *op_data,
 899                                 mdsno_t *mds)
 900 {
 901         struct lmv_obd          *lmv = &obd->u.lmv;
 902         struct lmv_object       *obj;
 903         int                      rc;
 904         ENTRY;
 905
 906         LASSERT(mds != NULL);
 907
 908         if (lmv->desc.ld_tgt_count == 1) {
 909                 *mds = 0;
 910                 RETURN(0);
 911         }
 912
 913         /*
 914          * Allocate new fid on target according to operation type and parent
 915          * home mds.
 916          */
 917         obj = lmv_object_find(obd, &op_data->op_fid1);
 918         if (obj != NULL || op_data->op_name == NULL ||
 919             op_data->op_opc != LUSTRE_OPC_MKDIR) {
 920                 /*
 921                  * Allocate fid for non-dir or for null name or for case parent
 922                  * dir is split.
 923                  */
 924                 if (obj) {
 925                         lmv_object_put(obj);
 926
 927                         /*
 928                          * If we have this flag turned on, and we see that
 929                          * parent dir is split, this means, that caller did not
 930                          * notice split yet. This is race and we would like to
 931                          * let caller know that.
 932                          */
 933                         if (op_data->op_bias & MDS_CHECK_SPLIT)
 934                                 RETURN(-ERESTART);
 935                 }
 936
 937                 /*
 938                  * Allocate new fid on same mds where parent fid is located and
 939                  * where operation will be sent. In case of split dir, ->op_fid1
 940                  * and ->op_mds here will contain fid and mds of slave directory
 941                  * object (assigned by caller).
 942                  */
 943                 *mds = op_data->op_mds;
 944                 rc = 0;
 945         } else {
 946                 /*
 947                  * Parent directory is not split and we want to create a
 948                  * directory in it. Let's calculate where to place it according
 949                  * to operation data @op_data.
 950                  */
 951                 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
 952                 rc = 0;
 953         }
 954
 955         if (rc) {
 956                 CERROR("Can't choose MDS, err = %d\n", rc);
 957         } else {
 958                 LASSERT(*mds < lmv->desc.ld_tgt_count);
 959         }
 960
 961         RETURN(rc);
 962 }
 963
 964 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
 965                     mdsno_t mds)
 966 {
 967         struct lmv_tgt_desc *tgt;
 968         int                  rc;
 969         ENTRY;
 970
 971         tgt = lmv_get_target(lmv, mds);
 972
 973         /*
 974          * New seq alloc and FLD setup should be atomic. Otherwise we may find
 975          * on server that seq in new allocated fid is not yet known.
 976          */
 977         cfs_down(&tgt->ltd_fid_sem);
 978
 979         if (!tgt->ltd_active)
 980                 GOTO(out, rc = -ENODEV);
 981
 982         /*
 983          * Asking underlaying tgt layer to allocate new fid.
 984          */
 985         rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
 986         if (rc > 0) {
 987                 LASSERT(fid_is_sane(fid));
 988                 rc = 0;
 989         }
 990
 991         EXIT;
 992 out:
 993         cfs_up(&tgt->ltd_fid_sem);
 994         return rc;
 995 }
 996
 997 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
 998                   struct md_op_data *op_data)
 999 {
1000         struct obd_device     *obd = class_exp2obd(exp);
1001         struct lmv_obd        *lmv = &obd->u.lmv;
1002         mdsno_t                mds;
1003         int                    rc;
1004         ENTRY;
1005
1006         LASSERT(op_data != NULL);
1007         LASSERT(fid != NULL);
1008
1009         rc = lmv_placement_policy(obd, op_data, &mds);
1010         if (rc) {
1011                 CERROR("Can't get target for allocating fid, "
1012                        "rc %d\n", rc);
1013                 RETURN(rc);
1014         }
1015
1016         rc = __lmv_fid_alloc(lmv, fid, mds);
1017         if (rc) {
1018                 CERROR("Can't alloc new fid, rc %d\n", rc);
1019                 RETURN(rc);
1020         }
1021
1022         RETURN(rc);
1023 }
1024
1025 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1026 {
1027         ENTRY;
1028         LASSERT(exp != NULL && fid != NULL);
1029         if (lmv_object_delete(exp, fid)) {
1030                 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1031                        PFID(fid));
1032         }
1033         RETURN(0);
1034 }
1035
1036 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1037 {
1038         struct lmv_obd             *lmv = &obd->u.lmv;
1039         struct lprocfs_static_vars  lvars;
1040         struct lmv_desc            *desc;
1041         int                         rc;
1042         int                         i = 0;
1043         ENTRY;
1044
1045         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1046                 CERROR("LMV setup requires a descriptor\n");
1047                 RETURN(-EINVAL);
1048         }
1049
1050         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1051         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1052                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1053                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1054                 RETURN(-EINVAL);
1055         }
1056
1057         lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1058
1059         OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1060         if (lmv->tgts == NULL)
1061                 RETURN(-ENOMEM);
1062
1063         for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1064                 cfs_sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
1065                 lmv->tgts[i].ltd_idx = i;
1066         }
1067
1068         lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1069
1070         OBD_ALLOC(lmv->datas, lmv->datas_size);
1071         if (lmv->datas == NULL)
1072                 GOTO(out_free_tgts, rc = -ENOMEM);
1073
1074         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1075         lmv->desc.ld_tgt_count = 0;
1076         lmv->desc.ld_active_tgt_count = 0;
1077         lmv->max_cookiesize = 0;
1078         lmv->max_def_easize = 0;
1079         lmv->max_easize = 0;
1080         lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1081
1082         cfs_spin_lock_init(&lmv->lmv_lock);
1083         cfs_sema_init(&lmv->init_sem, 1);
1084
1085         rc = lmv_object_setup(obd);
1086         if (rc) {
1087                 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1088                 GOTO(out_free_datas, rc);
1089         }
1090
1091         lprocfs_lmv_init_vars(&lvars);
1092         lprocfs_obd_setup(obd, lvars.obd_vars);
1093 #ifdef LPROCFS
1094         {
1095                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1096                                         0444, &lmv_proc_target_fops, obd);
1097                 if (rc)
1098                         CWARN("Error adding target_obd_stats file (%d)\n", rc);
1099        }
1100 #endif
1101         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1102                              LUSTRE_CLI_FLD_HASH_DHT);
1103         if (rc) {
1104                 CERROR("Can't init FLD, err %d\n", rc);
1105                 GOTO(out_free_datas, rc);
1106         }
1107
1108         RETURN(0);
1109
1110 out_free_datas:
1111         OBD_FREE(lmv->datas, lmv->datas_size);
1112         lmv->datas = NULL;
1113 out_free_tgts:
1114         OBD_FREE(lmv->tgts, lmv->tgts_size);
1115         lmv->tgts = NULL;
1116         return rc;
1117 }
1118
1119 static int lmv_cleanup(struct obd_device *obd)
1120 {
1121         struct lmv_obd   *lmv = &obd->u.lmv;
1122         ENTRY;
1123
1124         fld_client_fini(&lmv->lmv_fld);
1125         lprocfs_obd_cleanup(obd);
1126         lmv_object_cleanup(obd);
1127         OBD_FREE(lmv->datas, lmv->datas_size);
1128         OBD_FREE(lmv->tgts, lmv->tgts_size);
1129
1130         RETURN(0);
1131 }
1132
1133 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1134 {
1135         struct lustre_cfg     *lcfg = buf;
1136         struct obd_uuid        tgt_uuid;
1137         int                    rc;
1138         ENTRY;
1139
1140         switch(lcfg->lcfg_command) {
1141         case LCFG_ADD_MDC:
1142                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1143                         GOTO(out, rc = -EINVAL);
1144
1145                 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1146                 rc = lmv_add_target(obd, &tgt_uuid);
1147                 GOTO(out, rc);
1148         default: {
1149                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1150                 GOTO(out, rc = -EINVAL);
1151         }
1152         }
1153 out:
1154         RETURN(rc);
1155 }
1156
1157 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1158                       __u64 max_age, __u32 flags)
1159 {
1160         struct lmv_obd        *lmv = &obd->u.lmv;
1161         struct obd_statfs     *temp;
1162         int                    rc = 0;
1163         int                    i;
1164         ENTRY;
1165
1166         rc = lmv_check_connect(obd);
1167         if (rc)
1168                 RETURN(rc);
1169
1170         OBD_ALLOC(temp, sizeof(*temp));
1171         if (temp == NULL)
1172                 RETURN(-ENOMEM);
1173
1174         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1175                 if (lmv->tgts[i].ltd_exp == NULL)
1176                         continue;
1177
1178                 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1179                                 max_age, flags);
1180                 if (rc) {
1181                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1182                                lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1183                                rc);
1184                         GOTO(out_free_temp, rc);
1185                 }
1186                 if (i == 0) {
1187                         *osfs = *temp;
1188                 } else {
1189                         osfs->os_bavail += temp->os_bavail;
1190                         osfs->os_blocks += temp->os_blocks;
1191                         osfs->os_ffree += temp->os_ffree;
1192                         osfs->os_files += temp->os_files;
1193                 }
1194         }
1195
1196         EXIT;
1197 out_free_temp:
1198         OBD_FREE(temp, sizeof(*temp));
1199         return rc;
1200 }
1201
1202 static int lmv_getstatus(struct obd_export *exp,
1203                          struct lu_fid *fid,
1204                          struct obd_capa **pc)
1205 {
1206         struct obd_device    *obd = exp->exp_obd;
1207         struct lmv_obd       *lmv = &obd->u.lmv;
1208         int                   rc;
1209         ENTRY;
1210
1211         rc = lmv_check_connect(obd);
1212         if (rc)
1213                 RETURN(rc);
1214
1215         rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1216         RETURN(rc);
1217 }
1218
1219 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1220                         struct obd_capa *oc, obd_valid valid, const char *name,
1221                         const char *input, int input_size, int output_size,
1222                         int flags, struct ptlrpc_request **request)
1223 {
1224         struct obd_device      *obd = exp->exp_obd;
1225         struct lmv_obd         *lmv = &obd->u.lmv;
1226         struct lmv_tgt_desc    *tgt;
1227         int                     rc;
1228         ENTRY;
1229
1230         rc = lmv_check_connect(obd);
1231         if (rc)
1232                 RETURN(rc);
1233
1234         tgt = lmv_find_target(lmv, fid);
1235         if (IS_ERR(tgt))
1236                 RETURN(PTR_ERR(tgt));
1237
1238         rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1239                          input_size, output_size, flags, request);
1240
1241         RETURN(rc);
1242 }
1243
1244 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1245                         struct obd_capa *oc, obd_valid valid, const char *name,
1246                         const char *input, int input_size, int output_size,
1247                         int flags, __u32 suppgid,
1248                         struct ptlrpc_request **request)
1249 {
1250         struct obd_device      *obd = exp->exp_obd;
1251         struct lmv_obd         *lmv = &obd->u.lmv;
1252         struct lmv_tgt_desc    *tgt;
1253         int                     rc;
1254         ENTRY;
1255
1256         rc = lmv_check_connect(obd);
1257         if (rc)
1258                 RETURN(rc);
1259
1260         tgt = lmv_find_target(lmv, fid);
1261         if (IS_ERR(tgt))
1262                 RETURN(PTR_ERR(tgt));
1263
1264         rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1265                          input_size, output_size, flags, suppgid,
1266                          request);
1267
1268         RETURN(rc);
1269 }
1270
1271 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1272                        struct ptlrpc_request **request)
1273 {
1274         struct obd_device       *obd = exp->exp_obd;
1275         struct lmv_obd          *lmv = &obd->u.lmv;
1276         struct lmv_tgt_desc     *tgt;
1277         struct lmv_object       *obj;
1278         int                      rc;
1279         int                      i;
1280         ENTRY;
1281
1282         rc = lmv_check_connect(obd);
1283         if (rc)
1284                 RETURN(rc);
1285
1286         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1287         if (IS_ERR(tgt))
1288                 RETURN(PTR_ERR(tgt));
1289
1290         if (op_data->op_valid & OBD_MD_MDTIDX) {
1291                 op_data->op_mds = tgt->ltd_idx;
1292                 RETURN(0);
1293         }
1294
1295         rc = md_getattr(tgt->ltd_exp, op_data, request);
1296         if (rc)
1297                 RETURN(rc);
1298
1299         obj = lmv_object_find_lock(obd, &op_data->op_fid1);
1300
1301         CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(&op_data->op_fid1),
1302                obj ? "(split)" : "");
1303
1304         /*
1305          * If object is split, then we loop over all the slaves and gather size
1306          * attribute. In ideal world we would have to gather also mds field from
1307          * all slaves, as object is spread over the cluster and this is
1308          * definitely interesting information and it is not good to loss it,
1309          * but...
1310          */
1311         if (obj) {
1312                 struct mdt_body *body;
1313
1314                 if (*request == NULL) {
1315                         lmv_object_put(obj);
1316                         RETURN(rc);
1317                 }
1318
1319                 body = req_capsule_server_get(&(*request)->rq_pill,
1320                                               &RMF_MDT_BODY);
1321                 LASSERT(body != NULL);
1322
1323                 for (i = 0; i < obj->lo_objcount; i++) {
1324                         if (lmv->tgts[i].ltd_exp == NULL) {
1325                                 CWARN("%s: NULL export for %d\n",
1326                                       obd->obd_name, i);
1327                                 continue;
1328                         }
1329
1330                         /*
1331                          * Skip master object.
1332                          */
1333                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1334                                 continue;
1335
1336                         body->size += obj->lo_stripes[i].ls_size;
1337                 }
1338
1339                 lmv_object_put_unlock(obj);
1340         }
1341
1342         RETURN(rc);
1343 }
1344
1345 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1346                              ldlm_iterator_t it, void *data)
1347 {
1348         struct obd_device   *obd = exp->exp_obd;
1349         struct lmv_obd      *lmv = &obd->u.lmv;
1350         int                  i;
1351         int                  rc;
1352         ENTRY;
1353
1354         rc = lmv_check_connect(obd);
1355         if (rc)
1356                 RETURN(rc);
1357
1358         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1359
1360         /*
1361          * With CMD every object can have two locks in different namespaces:
1362          * lookup lock in space of mds storing direntry and update/open lock in
1363          * space of mds storing inode.
1364          */
1365         for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1366                 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1367
1368         RETURN(0);
1369 }
1370
1371 static int lmv_find_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1372                            ldlm_iterator_t it, void *data)
1373 {
1374         struct obd_device   *obd = exp->exp_obd;
1375         struct lmv_obd      *lmv = &obd->u.lmv;
1376         int                  i;
1377         int                  rc;
1378         ENTRY;
1379
1380         rc = lmv_check_connect(obd);
1381         if (rc)
1382                 RETURN(rc);
1383
1384         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1385
1386         /*
1387          * With CMD every object can have two locks in different namespaces:
1388          * lookup lock in space of mds storing direntry and update/open lock in
1389          * space of mds storing inode.
1390          */
1391         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1392                 rc = md_find_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1393                 if (rc)
1394                         RETURN(rc);
1395         }
1396
1397         RETURN(rc);
1398 }
1399
1400
1401 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1402                      struct md_open_data *mod, struct ptlrpc_request **request)
1403 {
1404         struct obd_device     *obd = exp->exp_obd;
1405         struct lmv_obd        *lmv = &obd->u.lmv;
1406         struct lmv_tgt_desc   *tgt;
1407         int                    rc;
1408         ENTRY;
1409
1410         rc = lmv_check_connect(obd);
1411         if (rc)
1412                 RETURN(rc);
1413
1414         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1415         if (IS_ERR(tgt))
1416                 RETURN(PTR_ERR(tgt));
1417
1418         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1419         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1420         RETURN(rc);
1421 }
1422
1423 /**
1424  * Called in the case MDS returns -ERESTART on create on open, what means that
1425  * directory is split and its LMV presentation object has to be updated.
1426  */
1427 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1428 {
1429         struct obd_device       *obd = exp->exp_obd;
1430         struct lmv_obd          *lmv = &obd->u.lmv;
1431         struct ptlrpc_request   *req = NULL;
1432         struct lmv_tgt_desc     *tgt;
1433         struct lmv_object       *obj;
1434         struct lustre_md         md;
1435         struct md_op_data       *op_data;
1436         int                      mealen;
1437         int                      rc;
1438         __u64                    valid;
1439         ENTRY;
1440
1441         md.mea = NULL;
1442         mealen = lmv_get_easize(lmv);
1443
1444         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1445
1446         tgt = lmv_find_target(lmv, fid);
1447         if (IS_ERR(tgt))
1448                 RETURN(PTR_ERR(tgt));
1449
1450         /*
1451          * Time to update mea of parent fid.
1452          */
1453
1454         OBD_ALLOC_PTR(op_data);
1455         if (op_data == NULL)
1456                 RETURN(-ENOMEM);
1457
1458         op_data->op_fid1 = *fid;
1459         op_data->op_mode = mealen;
1460         op_data->op_valid = valid;
1461
1462         rc = md_getattr(tgt->ltd_exp, op_data, &req);
1463         OBD_FREE_PTR(op_data);
1464         if (rc) {
1465                 CERROR("md_getattr() failed, error %d\n", rc);
1466                 GOTO(cleanup, rc);
1467         }
1468
1469         rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1470         if (rc) {
1471                 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1472                 GOTO(cleanup, rc);
1473         }
1474
1475         if (md.mea == NULL)
1476                 GOTO(cleanup, rc = -ENODATA);
1477
1478         obj = lmv_object_create(exp, fid, md.mea);
1479         if (IS_ERR(obj))
1480                 rc = PTR_ERR(obj);
1481         else
1482                 lmv_object_put(obj);
1483
1484         obd_free_memmd(exp, (void *)&md.mea);
1485         EXIT;
1486 cleanup:
1487         if (req)
1488                 ptlrpc_req_finished(req);
1489         return rc;
1490 }
1491
1492 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1493                const void *data, int datalen, int mode, __u32 uid,
1494                __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1495                struct ptlrpc_request **request)
1496 {
1497         struct obd_device       *obd = exp->exp_obd;
1498         struct lmv_obd          *lmv = &obd->u.lmv;
1499         struct lmv_tgt_desc     *tgt;
1500         struct lmv_object       *obj;
1501         int                      rc;
1502         int                      loop = 0;
1503         int                      sidx;
1504         ENTRY;
1505
1506         rc = lmv_check_connect(obd);
1507         if (rc)
1508                 RETURN(rc);
1509
1510         if (!lmv->desc.ld_active_tgt_count)
1511                 RETURN(-EIO);
1512 repeat:
1513         ++loop;
1514         LASSERT(loop <= 2);
1515
1516         obj = lmv_object_find(obd, &op_data->op_fid1);
1517         if (obj) {
1518                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1519                                        op_data->op_name, op_data->op_namelen);
1520                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1521                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1522                 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1523                 tgt = lmv_get_target(lmv, op_data->op_mds);
1524                 lmv_object_put(obj);
1525         } else {
1526                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1527                 op_data->op_bias |= MDS_CHECK_SPLIT;
1528                 op_data->op_mds = tgt->ltd_idx;
1529         }
1530
1531         if (IS_ERR(tgt))
1532                 RETURN(PTR_ERR(tgt));
1533
1534         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1535         if (rc == -ERESTART)
1536                 goto repeat;
1537         else if (rc)
1538                 RETURN(rc);
1539
1540         CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #%x\n",
1541                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1542                op_data->op_mds);
1543
1544         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1545         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1546                        cap_effective, rdev, request);
1547         if (rc == 0) {
1548                 if (*request == NULL)
1549                         RETURN(rc);
1550                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1551         } else if (rc == -ERESTART) {
1552                 LASSERT(*request != NULL);
1553                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1554                           "Got -ERESTART during create!\n");
1555                 ptlrpc_req_finished(*request);
1556                 *request = NULL;
1557
1558                 /*
1559                  * Directory got split. Time to update local object and repeat
1560                  * the request with proper MDS.
1561                  */
1562                 rc = lmv_handle_split(exp, &op_data->op_fid1);
1563                 if (rc == 0) {
1564                         rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1565                                                  op_data, &op_data->op_fid2);
1566                         if (rc)
1567                                 RETURN(rc);
1568                         goto repeat;
1569                 }
1570         }
1571         RETURN(rc);
1572 }
1573
1574 static int lmv_done_writing(struct obd_export *exp,
1575                             struct md_op_data *op_data,
1576                             struct md_open_data *mod)
1577 {
1578         struct obd_device     *obd = exp->exp_obd;
1579         struct lmv_obd        *lmv = &obd->u.lmv;
1580         struct lmv_tgt_desc   *tgt;
1581         int                    rc;
1582         ENTRY;
1583
1584         rc = lmv_check_connect(obd);
1585         if (rc)
1586                 RETURN(rc);
1587
1588         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1589         if (IS_ERR(tgt))
1590                 RETURN(PTR_ERR(tgt));
1591
1592         rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1593         RETURN(rc);
1594 }
1595
1596 static int
1597 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1598                    struct lookup_intent *it, struct md_op_data *op_data,
1599                    struct lustre_handle *lockh, void *lmm, int lmmsize)
1600 {
1601         struct obd_device     *obd = exp->exp_obd;
1602         struct lmv_obd        *lmv = &obd->u.lmv;
1603         struct lmv_stripe_md  *mea = op_data->op_mea1;
1604         struct md_op_data     *op_data2;
1605         struct lmv_tgt_desc   *tgt;
1606         int                    i;
1607         int                    rc = 0;
1608         ENTRY;
1609
1610         OBD_ALLOC_PTR(op_data2);
1611         if (op_data2 == NULL)
1612                 RETURN(-ENOMEM);
1613
1614         LASSERT(mea != NULL);
1615         for (i = 0; i < mea->mea_count; i++) {
1616                 memset(op_data2, 0, sizeof(*op_data2));
1617                 op_data2->op_fid1 = mea->mea_ids[i];
1618                 op_data2->op_bias = 0;
1619
1620                 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1621                 if (IS_ERR(tgt))
1622                         GOTO(cleanup, rc = PTR_ERR(tgt));
1623
1624                 if (tgt->ltd_exp == NULL)
1625                         continue;
1626
1627                 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1628                                 lockh + i, lmm, lmmsize, NULL, 0);
1629
1630                 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1631                        PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1632
1633                 if (rc)
1634                         GOTO(cleanup, rc);
1635
1636                 if (it->d.lustre.it_data) {
1637                         struct ptlrpc_request *req;
1638                         req = (struct ptlrpc_request *)it->d.lustre.it_data;
1639                         ptlrpc_req_finished(req);
1640                 }
1641
1642                 if (it->d.lustre.it_status)
1643                         GOTO(cleanup, rc = it->d.lustre.it_status);
1644         }
1645
1646         EXIT;
1647 cleanup:
1648         OBD_FREE_PTR(op_data2);
1649
1650         if (rc != 0) {
1651                 /*
1652                  * Drop all taken locks.
1653                  */
1654                 while (--i >= 0) {
1655                         if (lockh[i].cookie)
1656                                 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1657                         lockh[i].cookie = 0;
1658                 }
1659         }
1660         return rc;
1661 }
1662
1663 static int
1664 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1665                    struct lookup_intent *it, struct md_op_data *op_data,
1666                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1667                    int extra_lock_flags)
1668 {
1669         struct ptlrpc_request      *req = it->d.lustre.it_data;
1670         struct obd_device          *obd = exp->exp_obd;
1671         struct lmv_obd             *lmv = &obd->u.lmv;
1672         struct lustre_handle        plock;
1673         struct lmv_tgt_desc        *tgt;
1674         struct md_op_data          *rdata;
1675         struct lu_fid               fid1;
1676         struct mdt_body            *body;
1677         int                         rc = 0;
1678         int                         pmode;
1679         ENTRY;
1680
1681         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1682         LASSERT(body != NULL);
1683
1684         if (!(body->valid & OBD_MD_MDS))
1685                 RETURN(0);
1686
1687         CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1688                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1689
1690         /*
1691          * We got LOOKUP lock, but we really need attrs.
1692          */
1693         pmode = it->d.lustre.it_lock_mode;
1694         LASSERT(pmode != 0);
1695         memcpy(&plock, lockh, sizeof(plock));
1696         it->d.lustre.it_lock_mode = 0;
1697         it->d.lustre.it_data = NULL;
1698         fid1 = body->fid1;
1699
1700         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1701         ptlrpc_req_finished(req);
1702
1703         tgt = lmv_find_target(lmv, &fid1);
1704         if (IS_ERR(tgt))
1705                 GOTO(out, rc = PTR_ERR(tgt));
1706
1707         OBD_ALLOC_PTR(rdata);
1708         if (rdata == NULL)
1709                 GOTO(out, rc = -ENOMEM);
1710
1711         rdata->op_fid1 = fid1;
1712         rdata->op_bias = MDS_CROSS_REF;
1713
1714         rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1715                         lmm, lmmsize, NULL, extra_lock_flags);
1716         OBD_FREE_PTR(rdata);
1717         EXIT;
1718 out:
1719         ldlm_lock_decref(&plock, pmode);
1720         return rc;
1721 }
1722
1723 static int
1724 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1725             struct lookup_intent *it, struct md_op_data *op_data,
1726             struct lustre_handle *lockh, void *lmm, int lmmsize,
1727             struct ptlrpc_request **req, int extra_lock_flags)
1728 {
1729         struct obd_device        *obd = exp->exp_obd;
1730         struct lmv_obd           *lmv = &obd->u.lmv;
1731         struct lmv_tgt_desc      *tgt;
1732         struct lmv_object        *obj;
1733         int                       sidx;
1734         int                       rc;
1735         ENTRY;
1736
1737         rc = lmv_check_connect(obd);
1738         if (rc)
1739                 RETURN(rc);
1740
1741         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1742                LL_IT2STR(it), PFID(&op_data->op_fid1));
1743
1744         if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1745                 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1746                                         lockh, lmm, lmmsize);
1747                 RETURN(rc);
1748         }
1749
1750         obj = lmv_object_find(obd, &op_data->op_fid1);
1751         if (obj && op_data->op_namelen) {
1752                 sidx = raw_name2idx(obj->lo_hashtype,
1753                                        obj->lo_objcount,
1754                                        (char *)op_data->op_name,
1755                                        op_data->op_namelen);
1756                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1757                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1758         } else {
1759                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1760         }
1761         if (obj)
1762                 lmv_object_put(obj);
1763
1764         if (IS_ERR(tgt))
1765                 RETURN(PTR_ERR(tgt));
1766
1767         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1768                LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1769
1770         rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1771                         lmm, lmmsize, req, extra_lock_flags);
1772
1773         if (rc == 0 && it && it->it_op == IT_OPEN) {
1774                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1775                                         lmm, lmmsize, extra_lock_flags);
1776         }
1777         RETURN(rc);
1778 }
1779
1780 static int
1781 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1782                  struct ptlrpc_request **request)
1783 {
1784         struct ptlrpc_request   *req = NULL;
1785         struct obd_device       *obd = exp->exp_obd;
1786         struct lmv_obd          *lmv = &obd->u.lmv;
1787         struct lu_fid            rid = op_data->op_fid1;
1788         struct lmv_tgt_desc     *tgt;
1789         struct mdt_body         *body;
1790         struct lmv_object       *obj;
1791         obd_valid                valid = op_data->op_valid;
1792         int                      rc;
1793         int                      loop = 0;
1794         int                      sidx;
1795         ENTRY;
1796
1797         rc = lmv_check_connect(obd);
1798         if (rc)
1799                 RETURN(rc);
1800
1801 repeat:
1802         ++loop;
1803         LASSERT(loop <= 2);
1804         obj = lmv_object_find(obd, &rid);
1805         if (obj) {
1806                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1807                                     op_data->op_name, op_data->op_namelen);
1808                 rid = obj->lo_stripes[sidx].ls_fid;
1809                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1810                 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1811                 valid &= ~OBD_MD_FLCKSPLIT;
1812                 lmv_object_put(obj);
1813         } else {
1814                 tgt = lmv_find_target(lmv, &rid);
1815                 valid |= OBD_MD_FLCKSPLIT;
1816                 op_data->op_mds = tgt->ltd_idx;
1817         }
1818         if (IS_ERR(tgt))
1819                 RETURN(PTR_ERR(tgt));
1820
1821         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1822                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1823                PFID(&rid), tgt->ltd_idx);
1824
1825         op_data->op_valid = valid;
1826         op_data->op_fid1 = rid;
1827         rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1828         if (rc == 0) {
1829                 body = req_capsule_server_get(&(*request)->rq_pill,
1830                                               &RMF_MDT_BODY);
1831                 LASSERT(body != NULL);
1832
1833                 if (body->valid & OBD_MD_MDS) {
1834                         rid = body->fid1;
1835                         CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1836                                PFID(&rid));
1837
1838                         tgt = lmv_find_target(lmv, &rid);
1839                         if (IS_ERR(tgt)) {
1840                                 ptlrpc_req_finished(*request);
1841                                 RETURN(PTR_ERR(tgt));
1842                         }
1843
1844                         op_data->op_fid1 = rid;
1845                         op_data->op_valid |= OBD_MD_FLCROSSREF;
1846                         op_data->op_namelen = 0;
1847                         op_data->op_name = NULL;
1848                         rc = md_getattr_name(tgt->ltd_exp, op_data, &req);
1849                         ptlrpc_req_finished(*request);
1850                         *request = req;
1851                 }
1852         } else if (rc == -ERESTART) {
1853                 LASSERT(*request != NULL);
1854                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1855                           "Got -ERESTART during getattr!\n");
1856                 ptlrpc_req_finished(*request);
1857                 *request = NULL;
1858
1859                 /*
1860                  * Directory got split. Time to update local object and repeat
1861                  * the request with proper MDS.
1862                  */
1863                 rc = lmv_handle_split(exp, &rid);
1864                 if (rc == 0)
1865                         goto repeat;
1866         }
1867         RETURN(rc);
1868 }
1869
1870 #define md_op_data_fid(op_data, fl)                     \
1871         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1872          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1873          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1874          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1875          NULL)
1876
1877 static int lmv_early_cancel_slaves(struct obd_export *exp,
1878                                    struct md_op_data *op_data, int op_tgt,
1879                                    ldlm_mode_t mode, int bits, int flag)
1880 {
1881         struct obd_device      *obd = exp->exp_obd;
1882         struct lmv_obd         *lmv = &obd->u.lmv;
1883         ldlm_policy_data_t      policy = {{0}};
1884         struct lu_fid          *op_fid;
1885         struct lu_fid          *st_fid;
1886         struct lmv_tgt_desc    *tgt;
1887         struct lmv_object      *obj;
1888         int                     rc = 0;
1889         int                     i;
1890         ENTRY;
1891
1892         op_fid = md_op_data_fid(op_data, flag);
1893         if (!fid_is_sane(op_fid))
1894                 RETURN(0);
1895
1896         obj = lmv_object_find(obd, op_fid);
1897         if (obj == NULL)
1898                 RETURN(-EALREADY);
1899
1900         policy.l_inodebits.bits = bits;
1901         for (i = 0; i < obj->lo_objcount; i++) {
1902                 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1903                 st_fid = &obj->lo_stripes[i].ls_fid;
1904                 if (op_tgt != tgt->ltd_idx) {
1905                         CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1906                                PFID(st_fid), tgt->ltd_idx);
1907                         rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1908                                               mode, LDLM_FL_ASYNC, NULL);
1909                         if (rc)
1910                                 GOTO(out_put_obj, rc);
1911                 } else {
1912                         CDEBUG(D_INODE,
1913                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1914                                op_tgt, PFID(st_fid));
1915                         /*
1916                          * Do not cancel locks for operation target, they will
1917                          * be handled later in underlaying layer when calling
1918                          * function we run on behalf of.
1919                          */
1920                         *op_fid = *st_fid;
1921                         op_data->op_flags |= flag;
1922                 }
1923         }
1924         EXIT;
1925 out_put_obj:
1926         lmv_object_put(obj);
1927         return rc;
1928 }
1929
1930 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1931                             int op_tgt, ldlm_mode_t mode, int bits, int flag)
1932 {
1933         struct lu_fid          *fid = md_op_data_fid(op_data, flag);
1934         struct obd_device      *obd = exp->exp_obd;
1935         struct lmv_obd         *lmv = &obd->u.lmv;
1936         struct lmv_tgt_desc    *tgt;
1937         ldlm_policy_data_t      policy = {{0}};
1938         struct lmv_object      *obj;
1939         int                     rc = 0;
1940         ENTRY;
1941
1942         if (!fid_is_sane(fid))
1943                 RETURN(0);
1944
1945         obj = lmv_object_find(obd, fid);
1946         if (obj) {
1947                 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1948                                              bits, flag);
1949                 lmv_object_put(obj);
1950         } else {
1951                 tgt = lmv_find_target(lmv, fid);
1952                 if (IS_ERR(tgt))
1953                         RETURN(PTR_ERR(tgt));
1954
1955                 if (tgt->ltd_idx != op_tgt) {
1956                         CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1957                         policy.l_inodebits.bits = bits;
1958                         rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1959                                               mode, LDLM_FL_ASYNC, NULL);
1960                 } else {
1961                         CDEBUG(D_INODE,
1962                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1963                                op_tgt, PFID(fid));
1964                         op_data->op_flags |= flag;
1965                         rc = 0;
1966                 }
1967
1968         }
1969         RETURN(rc);
1970 }
1971
1972 /*
1973  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1974  * op_data->op_fid2
1975  */
1976 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1977                     struct ptlrpc_request **request)
1978 {
1979         struct obd_device       *obd = exp->exp_obd;
1980         struct lmv_obd          *lmv = &obd->u.lmv;
1981         struct lmv_tgt_desc     *tgt;
1982         struct lmv_object       *obj;
1983         int                      rc;
1984         int                      loop = 0;
1985         mdsno_t                  mds;
1986         int                      sidx;
1987         ENTRY;
1988
1989         rc = lmv_check_connect(obd);
1990         if (rc)
1991                 RETURN(rc);
1992
1993 repeat:
1994         ++loop;
1995         LASSERT(loop <= 2);
1996         LASSERT(op_data->op_namelen != 0);
1997
1998         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
1999                PFID(&op_data->op_fid2), op_data->op_namelen,
2000                op_data->op_name, PFID(&op_data->op_fid1));
2001
2002         obj = lmv_object_find(obd, &op_data->op_fid2);
2003         if (obj) {
2004                 sidx = raw_name2idx(obj->lo_hashtype,
2005                                     obj->lo_objcount,
2006                                     op_data->op_name,
2007                                     op_data->op_namelen);
2008                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2009                 mds = obj->lo_stripes[sidx].ls_mds;
2010                 lmv_object_put(obj);
2011         } else {
2012                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
2013                 if (rc)
2014                         RETURN(rc);
2015         }
2016
2017         CDEBUG(D_INODE, "Forward to mds #%x ("DFID")\n",
2018                mds, PFID(&op_data->op_fid1));
2019
2020         op_data->op_fsuid = cfs_curproc_fsuid();
2021         op_data->op_fsgid = cfs_curproc_fsgid();
2022         op_data->op_cap = cfs_curproc_cap_pack();
2023         tgt = lmv_get_target(lmv, mds);
2024
2025         /*
2026          * Cancel UPDATE lock on child (fid1).
2027          */
2028         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2029         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2030                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2031         if (rc == 0)
2032                 rc = md_link(tgt->ltd_exp, op_data, request);
2033         if (rc == -ERESTART) {
2034                 LASSERT(*request != NULL);
2035                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2036                           "Got -ERESTART during link!\n");
2037                 ptlrpc_req_finished(*request);
2038                 *request = NULL;
2039
2040                 /*
2041                  * Directory got split. Time to update local object and repeat
2042                  * the request with proper MDS.
2043                  */
2044                 rc = lmv_handle_split(exp, &op_data->op_fid2);
2045                 if (rc == 0)
2046                         goto repeat;
2047         }
2048
2049         RETURN(rc);
2050 }
2051
2052 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2053                       const char *old, int oldlen, const char *new, int newlen,
2054                       struct ptlrpc_request **request)
2055 {
2056         struct obd_device       *obd = exp->exp_obd;
2057         struct lmv_obd          *lmv = &obd->u.lmv;
2058         struct lmv_tgt_desc     *src_tgt;
2059         struct lmv_tgt_desc     *tgt_tgt;
2060         int                      rc;
2061         int                      sidx;
2062         int                      loop = 0;
2063         struct lmv_object       *obj;
2064         mdsno_t                  mds1;
2065         mdsno_t                  mds2;
2066         ENTRY;
2067
2068         LASSERT(oldlen != 0);
2069
2070         CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2071                oldlen, old, PFID(&op_data->op_fid1),
2072                newlen, new, PFID(&op_data->op_fid2));
2073
2074         rc = lmv_check_connect(obd);
2075         if (rc)
2076                 RETURN(rc);
2077
2078 repeat:
2079         ++loop;
2080         LASSERT(loop <= 2);
2081         obj = lmv_object_find(obd, &op_data->op_fid1);
2082         if (obj) {
2083                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2084                                     (char *)old, oldlen);
2085                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2086                 mds1 = obj->lo_stripes[sidx].ls_mds;
2087                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2088                 lmv_object_put(obj);
2089         } else {
2090                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2091                 if (rc)
2092                         RETURN(rc);
2093         }
2094
2095         obj = lmv_object_find(obd, &op_data->op_fid2);
2096         if (obj) {
2097                 /*
2098                  * Directory is already split, so we have to forward request to
2099                  * the right MDS.
2100                  */
2101                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2102                                        (char *)new, newlen);
2103
2104                 mds2 = obj->lo_stripes[sidx].ls_mds;
2105                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2106                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2107                 lmv_object_put(obj);
2108         } else {
2109                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2110                 if (rc)
2111                         RETURN(rc);
2112         }
2113
2114         op_data->op_fsuid = cfs_curproc_fsuid();
2115         op_data->op_fsgid = cfs_curproc_fsgid();
2116         op_data->op_cap = cfs_curproc_cap_pack();
2117
2118         src_tgt = lmv_get_target(lmv, mds1);
2119         tgt_tgt = lmv_get_target(lmv, mds2);
2120
2121         /*
2122          * LOOKUP lock on src child (fid3) should also be cancelled for
2123          * src_tgt in mdc_rename.
2124          */
2125         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2126
2127         /*
2128          * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2129          * own target.
2130          */
2131         rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2132                               LCK_EX, MDS_INODELOCK_UPDATE,
2133                               MF_MDC_CANCEL_FID2);
2134
2135         /*
2136          * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2137          */
2138         if (rc == 0) {
2139                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2140                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2141                                       MF_MDC_CANCEL_FID4);
2142         }
2143
2144         /*
2145          * Cancel all the locks on tgt child (fid4).
2146          */
2147         if (rc == 0)
2148                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2149                                       LCK_EX, MDS_INODELOCK_FULL,
2150                                       MF_MDC_CANCEL_FID4);
2151
2152         if (rc == 0)
2153                 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2154                                new, newlen, request);
2155
2156         if (rc == -ERESTART) {
2157                 LASSERT(*request != NULL);
2158                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2159                           "Got -ERESTART during rename!\n");
2160                 ptlrpc_req_finished(*request);
2161                 *request = NULL;
2162
2163                 /*
2164                  * Directory got split. Time to update local object and repeat
2165                  * the request with proper MDS.
2166                  */
2167                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2168                 if (rc == 0)
2169                         goto repeat;
2170         }
2171         RETURN(rc);
2172 }
2173
2174 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2175                        void *ea, int ealen, void *ea2, int ea2len,
2176                        struct ptlrpc_request **request,
2177                        struct md_open_data **mod)
2178 {
2179         struct obd_device       *obd = exp->exp_obd;
2180         struct lmv_obd          *lmv = &obd->u.lmv;
2181         struct ptlrpc_request   *req;
2182         struct lmv_tgt_desc     *tgt;
2183         struct lmv_object       *obj;
2184         int                      rc = 0;
2185         int                      i;
2186         ENTRY;
2187
2188         rc = lmv_check_connect(obd);
2189         if (rc)
2190                 RETURN(rc);
2191
2192         obj = lmv_object_find(obd, &op_data->op_fid1);
2193
2194         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2195                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2196                obj ? ", split" : "");
2197
2198         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2199         if (obj) {
2200                 for (i = 0; i < obj->lo_objcount; i++) {
2201                         op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2202
2203                         tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2204                         if (IS_ERR(tgt)) {
2205                                 rc = PTR_ERR(tgt);
2206                                 break;
2207                         }
2208
2209                         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2210                                         ea2, ea2len, &req, mod);
2211
2212                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2213                                 /*
2214                                  * This is master object and this request should
2215                                  * be returned back to llite.
2216                                  */
2217                                 *request = req;
2218                         } else {
2219                                 ptlrpc_req_finished(req);
2220                         }
2221
2222                         if (rc)
2223                                 break;
2224                 }
2225                 lmv_object_put(obj);
2226         } else {
2227                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2228                 if (IS_ERR(tgt))
2229                         RETURN(PTR_ERR(tgt));
2230
2231                 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2232                                 ea2len, request, mod);
2233         }
2234         RETURN(rc);
2235 }
2236
2237 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2238                     struct obd_capa *oc, struct ptlrpc_request **request)
2239 {
2240         struct obd_device         *obd = exp->exp_obd;
2241         struct lmv_obd            *lmv = &obd->u.lmv;
2242         struct lmv_tgt_desc       *tgt;
2243         int                        rc;
2244         ENTRY;
2245
2246         rc = lmv_check_connect(obd);
2247         if (rc)
2248                 RETURN(rc);
2249
2250         tgt = lmv_find_target(lmv, fid);
2251         if (IS_ERR(tgt))
2252                 RETURN(PTR_ERR(tgt));
2253
2254         rc = md_sync(tgt->ltd_exp, fid, oc, request);
2255         RETURN(rc);
2256 }
2257
2258 /**
2259  * Main purpose of LMV blocking ast is to remove split directory LMV
2260  * presentation object (struct lmv_object) attached to the lock being revoked.
2261  */
2262 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2263                      void *data, int flag)
2264 {
2265         struct lustre_handle    lockh;
2266         struct lmv_object      *obj;
2267         int                     rc;
2268         ENTRY;
2269
2270         switch (flag) {
2271         case LDLM_CB_BLOCKING:
2272                 ldlm_lock2handle(lock, &lockh);
2273                 rc = ldlm_cli_cancel(&lockh);
2274                 if (rc < 0) {
2275                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2276                         RETURN(rc);
2277                 }
2278                 break;
2279         case LDLM_CB_CANCELING:
2280                 /*
2281                  * Time to drop cached attrs for split directory object
2282                  */
2283                 obj = lock->l_ast_data;
2284                 if (obj) {
2285                         CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2286                                ", master "DFID"\n",
2287                                lock->l_resource->lr_name.name[3] == 1 ?
2288                                "LOOKUP" : "UPDATE",
2289                                lock->l_resource->lr_name.name[0],
2290                                lock->l_resource->lr_name.name[1],
2291                                PFID(&obj->lo_fid));
2292                         lmv_object_put(obj);
2293                 }
2294                 break;
2295         default:
2296                 LBUG();
2297         }
2298         RETURN(0);
2299 }
2300
2301 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2302 {
2303         __u64         val;
2304
2305         val = le64_to_cpu(*hash);
2306         if (val < hash_adj)
2307                 val += MAX_HASH_SIZE;
2308         if (val != DIR_END_OFF)
2309                 *hash = cpu_to_le64(val - hash_adj);
2310 }
2311
2312 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2313 {
2314         __u64              id;
2315         struct obd_import *imp;
2316
2317         /*
2318          * XXX: to get nid we assume that underlying obd device is mdc.
2319          */
2320         imp  = class_exp2cliimp(exp);
2321         id   = imp->imp_connection->c_self + fid_flatten(fid);
2322
2323         CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2324                imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2325
2326         return id ^ (id >> 32);
2327 }
2328
2329 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2330                         struct obd_capa *oc, __u64 offset64, struct page *page,
2331                         struct ptlrpc_request **request)
2332 {
2333         struct obd_device       *obd = exp->exp_obd;
2334         struct lmv_obd          *lmv = &obd->u.lmv;
2335         struct lu_fid            rid = *fid;
2336         struct lmv_object       *obj;
2337         __u64                    offset;
2338         __u64                    hash_adj = 0;
2339         __u32                    rank = 0;
2340         __u64                    seg_size = 0;
2341         __u64                    tgt_tmp = 0;
2342         int                      tgt_idx = 0;
2343         int                      tgt0_idx = 0;
2344         int                      rc;
2345         int                      nr = 0;
2346         struct lmv_stripe       *los;
2347         struct lmv_tgt_desc     *tgt;
2348         struct lu_dirpage       *dp;
2349         struct lu_dirent        *ent;
2350         ENTRY;
2351
2352         offset = offset64;
2353
2354         rc = lmv_check_connect(obd);
2355         if (rc)
2356                 RETURN(rc);
2357
2358         CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2359
2360         /*
2361          * This case handle directory lookup in clustered metadata case (i.e.
2362          * split directory is located on multiple md servers.)
2363          * each server keeps directory entries for certain range of hashes.
2364          * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2365          * first server will keep records with hashes [ 0 ... MAX_HASH / N  - 1],
2366          * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2367          * so on....
2368          *      readdir can simply start reading entries from 0 - N server in
2369          * order but that will not scale well as all client will request dir in
2370          * to server in same order.
2371          * Following algorithm does optimization:
2372          * Instead of doing readdir in 1, 2, ...., N order, client with a
2373          * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2374          * (every client has rank R)
2375          *      But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2376          * since client ask dir from MDS{R} client has pages with offsets
2377          * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2378          * on hash  values that we get.
2379          */
2380         obj = lmv_object_find_lock(obd, fid);
2381         if (obj) {
2382                 nr       = obj->lo_objcount;
2383                 LASSERT(nr > 0);
2384                 seg_size = MAX_HASH_SIZE;
2385                 do_div(seg_size, nr);
2386                 los      = obj->lo_stripes;
2387                 tgt      = lmv_get_target(lmv, los[0].ls_mds);
2388                 rank     = lmv_node_rank(tgt->ltd_exp, fid) % nr;
2389                 tgt_tmp  = offset;
2390                 do_div(tgt_tmp, seg_size);
2391                 tgt0_idx = do_div(tgt_tmp,  nr);
2392                 tgt_idx  = (tgt0_idx + rank) % nr;
2393
2394                 if (tgt_idx < tgt0_idx)
2395                         /*
2396                          * Wrap around.
2397                          *
2398                          * Last segment has unusual length due to division
2399                          * rounding.
2400                          */
2401                         hash_adj = MAX_HASH_SIZE - seg_size * nr;
2402                 else
2403                         hash_adj = 0;
2404
2405                 hash_adj += rank * seg_size;
2406
2407                 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2408                        LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2409                        offset, tgt0_idx, offset + hash_adj, tgt_idx);
2410
2411                 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2412                 rid = obj->lo_stripes[tgt_idx].ls_fid;
2413                 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2414
2415                 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2416                        PFID(&rid), (unsigned long)offset, tgt_idx);
2417         } else
2418                 tgt = lmv_find_target(lmv, &rid);
2419
2420         if (IS_ERR(tgt))
2421                 GOTO(cleanup, rc = PTR_ERR(tgt));
2422
2423         rc = md_readpage(tgt->ltd_exp, &rid, oc, offset, page, request);
2424         if (rc)
2425                 GOTO(cleanup, rc);
2426         if (obj) {
2427                 dp = cfs_kmap(page);
2428
2429                 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2430                 lmv_hash_adjust(&dp->ldp_hash_end,   hash_adj);
2431                 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2432
2433                 for (ent = lu_dirent_start(dp); ent != NULL;
2434                      ent = lu_dirent_next(ent))
2435                         lmv_hash_adjust(&ent->lde_hash, hash_adj);
2436
2437                 if (tgt0_idx != nr - 1) {
2438                         __u64 end;
2439
2440                         end = le64_to_cpu(dp->ldp_hash_end);
2441                         if (end == DIR_END_OFF) {
2442                                 dp->ldp_hash_end = cpu_to_le32(seg_size *
2443                                                                (tgt0_idx + 1));
2444                                 CDEBUG(D_INODE,
2445                                        ""DFID" reset end "LPX64" tgt %d\n",
2446                                        PFID(&rid),
2447                                        (__u64)le64_to_cpu(dp->ldp_hash_end), tgt_idx);
2448                         }
2449                 }
2450                 cfs_kunmap(page);
2451         }
2452         EXIT;
2453 cleanup:
2454         if (obj)
2455                 lmv_object_put_unlock(obj);
2456         return rc;
2457 }
2458
2459 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2460                       struct ptlrpc_request **request)
2461 {
2462         struct obd_device       *obd = exp->exp_obd;
2463         struct lmv_obd          *lmv = &obd->u.lmv;
2464         struct lmv_tgt_desc     *tgt = NULL;
2465         struct lmv_object       *obj;
2466         int                      rc;
2467         int                      sidx;
2468         int                      loop = 0;
2469         ENTRY;
2470
2471         rc = lmv_check_connect(obd);
2472         if (rc)
2473                 RETURN(rc);
2474
2475 repeat:
2476         ++loop;
2477         LASSERT(loop <= 2);
2478         LASSERT(op_data->op_namelen != 0);
2479
2480         obj = lmv_object_find(obd, &op_data->op_fid1);
2481         if (obj) {
2482                 sidx = raw_name2idx(obj->lo_hashtype,
2483                                     obj->lo_objcount,
2484                                     op_data->op_name,
2485                                     op_data->op_namelen);
2486                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2487                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2488                 tgt = lmv_get_target(lmv,
2489                                      obj->lo_stripes[sidx].ls_mds);
2490                 lmv_object_put(obj);
2491                 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2492                        op_data->op_namelen, op_data->op_name,
2493                        PFID(&op_data->op_fid1), sidx);
2494         }
2495
2496         if (tgt == NULL) {
2497                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2498                 if (IS_ERR(tgt))
2499                         RETURN(PTR_ERR(tgt));
2500                 op_data->op_bias |= MDS_CHECK_SPLIT;
2501         }
2502
2503         op_data->op_fsuid = cfs_curproc_fsuid();
2504         op_data->op_fsgid = cfs_curproc_fsgid();
2505         op_data->op_cap = cfs_curproc_cap_pack();
2506
2507         /*
2508          * If child's fid is given, cancel unused locks for it if it is from
2509          * another export than parent.
2510          *
2511          * LOOKUP lock for child (fid3) should also be cancelled on parent
2512          * tgt_tgt in mdc_unlink().
2513          */
2514         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2515
2516         /*
2517          * Cancel FULL locks on child (fid3).
2518          */
2519         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2520                               MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2521
2522         if (rc == 0)
2523                 rc = md_unlink(tgt->ltd_exp, op_data, request);
2524
2525         if (rc == -ERESTART) {
2526                 LASSERT(*request != NULL);
2527                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2528                           "Got -ERESTART during unlink!\n");
2529                 ptlrpc_req_finished(*request);
2530                 *request = NULL;
2531
2532                 /*
2533                  * Directory got split. Time to update local object and repeat
2534                  * the request with proper MDS.
2535                  */
2536                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2537                 if (rc == 0)
2538                         goto repeat;
2539         }
2540         RETURN(rc);
2541 }
2542
2543 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2544 {
2545         int        rc = 0;
2546
2547         switch (stage) {
2548         case OBD_CLEANUP_EARLY:
2549                 /* XXX: here should be calling obd_precleanup() down to
2550                  * stack. */
2551                 break;
2552         case OBD_CLEANUP_EXPORTS:
2553                 rc = obd_llog_finish(obd, 0);
2554                 if (rc != 0)
2555                         CERROR("failed to cleanup llogging subsystems\n");
2556                 break;
2557         default:
2558                 break;
2559         }
2560         RETURN(rc);
2561 }
2562
2563 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2564                         void *key, __u32 *vallen, void *val,
2565                         struct lov_stripe_md *lsm)
2566 {
2567         struct obd_device       *obd;
2568         struct lmv_obd          *lmv;
2569         int                      rc = 0;
2570         ENTRY;
2571
2572         obd = class_exp2obd(exp);
2573         if (obd == NULL) {
2574                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2575                        exp->exp_handle.h_cookie);
2576                 RETURN(-EINVAL);
2577         }
2578
2579         lmv = &obd->u.lmv;
2580         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2581                 struct lmv_tgt_desc *tgts;
2582                 int i;
2583
2584                 rc = lmv_check_connect(obd);
2585                 if (rc)
2586                         RETURN(rc);
2587
2588                 LASSERT(*vallen == sizeof(__u32));
2589                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2590                      i++, tgts++) {
2591
2592                         /*
2593                          * All tgts should be connected when this gets called.
2594                          */
2595                         if (!tgts || !tgts->ltd_exp) {
2596                                 CERROR("target not setup?\n");
2597                                 continue;
2598                         }
2599
2600                         if (!obd_get_info(tgts->ltd_exp, keylen, key,
2601                                           vallen, val, NULL))
2602                                 RETURN(0);
2603                 }
2604                 RETURN(-EINVAL);
2605         } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2606                 rc = lmv_check_connect(obd);
2607                 if (rc)
2608                         RETURN(rc);
2609
2610                 /*
2611                  * Forwarding this request to first MDS, it should know LOV
2612                  * desc.
2613                  */
2614                 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2615                                   vallen, val, NULL);
2616                 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2617                         exp->exp_connect_flags =
2618                         ((struct obd_connect_data *)val)->ocd_connect_flags;
2619                 }
2620                 RETURN(rc);
2621         }
2622
2623         CDEBUG(D_IOCTL, "Invalid key\n");
2624         RETURN(-EINVAL);
2625 }
2626
2627 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2628                        void *key, obd_count vallen, void *val,
2629                        struct ptlrpc_request_set *set)
2630 {
2631         struct lmv_tgt_desc    *tgt;
2632         struct obd_device      *obd;
2633         struct lmv_obd         *lmv;
2634         int rc = 0;
2635         ENTRY;
2636
2637         obd = class_exp2obd(exp);
2638         if (obd == NULL) {
2639                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2640                        exp->exp_handle.h_cookie);
2641                 RETURN(-EINVAL);
2642         }
2643         lmv = &obd->u.lmv;
2644
2645         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2646             KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2647                 int i, err = 0;
2648
2649                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2650                         tgt = &lmv->tgts[i];
2651
2652                         if (!tgt->ltd_exp)
2653                                 continue;
2654
2655                         err = obd_set_info_async(tgt->ltd_exp,
2656                                                  keylen, key, vallen, val, set);
2657                         if (err && rc == 0)
2658                                 rc = err;
2659                 }
2660
2661                 RETURN(rc);
2662         }
2663
2664         RETURN(-EINVAL);
2665 }
2666
2667 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2668                struct lov_stripe_md *lsm)
2669 {
2670         struct obd_device         *obd = class_exp2obd(exp);
2671         struct lmv_obd            *lmv = &obd->u.lmv;
2672         struct lmv_stripe_md      *meap;
2673         struct lmv_stripe_md      *lsmp;
2674         int                        mea_size;
2675         int                        i;
2676         ENTRY;
2677
2678         mea_size = lmv_get_easize(lmv);
2679         if (!lmmp)
2680                 RETURN(mea_size);
2681
2682         if (*lmmp && !lsm) {
2683                 OBD_FREE(*lmmp, mea_size);
2684                 *lmmp = NULL;
2685                 RETURN(0);
2686         }
2687
2688         if (*lmmp == NULL) {
2689                 OBD_ALLOC(*lmmp, mea_size);
2690                 if (*lmmp == NULL)
2691                         RETURN(-ENOMEM);
2692         }
2693
2694         if (!lsm)
2695                 RETURN(mea_size);
2696
2697         lsmp = (struct lmv_stripe_md *)lsm;
2698         meap = (struct lmv_stripe_md *)*lmmp;
2699
2700         if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2701             lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2702                 RETURN(-EINVAL);
2703
2704         meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2705         meap->mea_count = cpu_to_le32(lsmp->mea_count);
2706         meap->mea_master = cpu_to_le32(lsmp->mea_master);
2707
2708         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2709                 meap->mea_ids[i] = meap->mea_ids[i];
2710                 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2711         }
2712
2713         RETURN(mea_size);
2714 }
2715
2716 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2717                  struct lov_mds_md *lmm, int lmm_size)
2718 {
2719         struct obd_device          *obd = class_exp2obd(exp);
2720         struct lmv_stripe_md      **tmea = (struct lmv_stripe_md **)lsmp;
2721         struct lmv_stripe_md       *mea = (struct lmv_stripe_md *)lmm;
2722         struct lmv_obd             *lmv = &obd->u.lmv;
2723         int                         mea_size;
2724         int                         i;
2725         __u32                       magic;
2726         ENTRY;
2727
2728         mea_size = lmv_get_easize(lmv);
2729         if (lsmp == NULL)
2730                 return mea_size;
2731
2732         if (*lsmp != NULL && lmm == NULL) {
2733                 OBD_FREE(*tmea, mea_size);
2734                 *lsmp = NULL;
2735                 RETURN(0);
2736         }
2737
2738         LASSERT(mea_size == lmm_size);
2739
2740         OBD_ALLOC(*tmea, mea_size);
2741         if (*tmea == NULL)
2742                 RETURN(-ENOMEM);
2743
2744         if (!lmm)
2745                 RETURN(mea_size);
2746
2747         if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2748             mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2749             mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2750         {
2751                 magic = le32_to_cpu(mea->mea_magic);
2752         } else {
2753                 /*
2754                  * Old mea is not handled here.
2755                  */
2756                 CERROR("Old not supportable EA is found\n");
2757                 LBUG();
2758         }
2759
2760         (*tmea)->mea_magic = magic;
2761         (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2762         (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2763
2764         for (i = 0; i < (*tmea)->mea_count; i++) {
2765                 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2766                 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2767         }
2768         RETURN(mea_size);
2769 }
2770
2771 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2772                              ldlm_policy_data_t *policy, ldlm_mode_t mode,
2773                              int flags, void *opaque)
2774 {
2775         struct obd_device       *obd = exp->exp_obd;
2776         struct lmv_obd          *lmv = &obd->u.lmv;
2777         int                      rc = 0;
2778         int                      err;
2779         int                      i;
2780         ENTRY;
2781
2782         LASSERT(fid != NULL);
2783
2784         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2785                 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2786                         continue;
2787
2788                 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2789                                        policy, mode, flags, opaque);
2790                 if (!rc)
2791                         rc = err;
2792         }
2793         RETURN(rc);
2794 }
2795
2796 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
2797                       __u32 *bits)
2798 {
2799         struct obd_device       *obd = exp->exp_obd;
2800         struct lmv_obd          *lmv = &obd->u.lmv;
2801         int                      rc;
2802         ENTRY;
2803
2804         rc =  md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data, bits);
2805         RETURN(rc);
2806 }
2807
2808 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2809                            const struct lu_fid *fid, ldlm_type_t type,
2810                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
2811                            struct lustre_handle *lockh)
2812 {
2813         struct obd_device       *obd = exp->exp_obd;
2814         struct lmv_obd          *lmv = &obd->u.lmv;
2815         ldlm_mode_t              rc;
2816         int                      i;
2817         ENTRY;
2818
2819         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2820
2821         /*
2822          * With CMD every object can have two locks in different namespaces:
2823          * lookup lock in space of mds storing direntry and update/open lock in
2824          * space of mds storing inode. Thus we check all targets, not only that
2825          * one fid was created in.
2826          */
2827         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2828                 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2829                                    type, policy, mode, lockh);
2830                 if (rc)
2831                         RETURN(rc);
2832         }
2833
2834         RETURN(0);
2835 }
2836
2837 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2838                       struct obd_export *dt_exp, struct obd_export *md_exp,
2839                       struct lustre_md *md)
2840 {
2841         struct obd_device       *obd = exp->exp_obd;
2842         struct lmv_obd          *lmv = &obd->u.lmv;
2843         int                      rc;
2844         ENTRY;
2845         rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2846         RETURN(rc);
2847 }
2848
2849 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2850 {
2851         struct obd_device       *obd = exp->exp_obd;
2852         struct lmv_obd          *lmv = &obd->u.lmv;
2853         ENTRY;
2854
2855         if (md->mea)
2856                 obd_free_memmd(exp, (void *)&md->mea);
2857         RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2858 }
2859
2860 int lmv_set_open_replay_data(struct obd_export *exp,
2861                              struct obd_client_handle *och,
2862                              struct ptlrpc_request *open_req)
2863 {
2864         struct obd_device       *obd = exp->exp_obd;
2865         struct lmv_obd          *lmv = &obd->u.lmv;
2866         struct lmv_tgt_desc     *tgt;
2867         ENTRY;
2868
2869         tgt = lmv_find_target(lmv, &och->och_fid);
2870         if (IS_ERR(tgt))
2871                 RETURN(PTR_ERR(tgt));
2872
2873         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2874 }
2875
2876 int lmv_clear_open_replay_data(struct obd_export *exp,
2877                                struct obd_client_handle *och)
2878 {
2879         struct obd_device       *obd = exp->exp_obd;
2880         struct lmv_obd          *lmv = &obd->u.lmv;
2881         struct lmv_tgt_desc     *tgt;
2882         ENTRY;
2883
2884         tgt = lmv_find_target(lmv, &och->och_fid);
2885         if (IS_ERR(tgt))
2886                 RETURN(PTR_ERR(tgt));
2887
2888         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2889 }
2890
2891 static int lmv_get_remote_perm(struct obd_export *exp,
2892                                const struct lu_fid *fid,
2893                                struct obd_capa *oc, __u32 suppgid,
2894                                struct ptlrpc_request **request)
2895 {
2896         struct obd_device       *obd = exp->exp_obd;
2897         struct lmv_obd          *lmv = &obd->u.lmv;
2898         struct lmv_tgt_desc     *tgt;
2899         int                      rc;
2900         ENTRY;
2901
2902         rc = lmv_check_connect(obd);
2903         if (rc)
2904                 RETURN(rc);
2905
2906         tgt = lmv_find_target(lmv, fid);
2907         if (IS_ERR(tgt))
2908                 RETURN(PTR_ERR(tgt));
2909
2910         rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
2911         RETURN(rc);
2912 }
2913
2914 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2915                           renew_capa_cb_t cb)
2916 {
2917         struct obd_device       *obd = exp->exp_obd;
2918         struct lmv_obd          *lmv = &obd->u.lmv;
2919         struct lmv_tgt_desc     *tgt;
2920         int                      rc;
2921         ENTRY;
2922
2923         rc = lmv_check_connect(obd);
2924         if (rc)
2925                 RETURN(rc);
2926
2927         tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
2928         if (IS_ERR(tgt))
2929                 RETURN(PTR_ERR(tgt));
2930
2931         rc = md_renew_capa(tgt->ltd_exp, oc, cb);
2932         RETURN(rc);
2933 }
2934
2935 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
2936                     const struct req_msg_field *field, struct obd_capa **oc)
2937 {
2938         struct obd_device *obd = exp->exp_obd;
2939         struct lmv_obd *lmv = &obd->u.lmv;
2940         int rc;
2941
2942         ENTRY;
2943         rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
2944         RETURN(rc);
2945 }
2946
2947 int lmv_intent_getattr_async(struct obd_export *exp,
2948                              struct md_enqueue_info *minfo,
2949                              struct ldlm_enqueue_info *einfo)
2950 {
2951         struct md_op_data       *op_data = &minfo->mi_data;
2952         struct obd_device       *obd = exp->exp_obd;
2953         struct lmv_obd          *lmv = &obd->u.lmv;
2954         struct lmv_object       *obj;
2955         struct lmv_tgt_desc     *tgt;
2956         int                      rc;
2957         int                      sidx;
2958         ENTRY;
2959
2960         rc = lmv_check_connect(obd);
2961         if (rc)
2962                 RETURN(rc);
2963
2964         if (!fid_is_sane(&op_data->op_fid2)) {
2965                 obj = lmv_object_find(obd, &op_data->op_fid1);
2966                 if (obj && op_data->op_namelen) {
2967                         sidx = raw_name2idx(obj->lo_hashtype,
2968                                             obj->lo_objcount,
2969                                             (char *)op_data->op_name,
2970                                             op_data->op_namelen);
2971                         op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2972                         tgt = lmv_get_target(lmv,
2973                                              obj->lo_stripes[sidx].ls_mds);
2974                         CDEBUG(D_INODE,
2975                                "Choose slave dir ("DFID") -> mds #%d\n",
2976                                PFID(&op_data->op_fid1), tgt->ltd_idx);
2977                 } else {
2978                         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2979                 }
2980                 if (obj)
2981                         lmv_object_put(obj);
2982         } else {
2983                 op_data->op_fid1 = op_data->op_fid2;
2984                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2985                 op_data->op_bias = MDS_CROSS_REF;
2986                 /*
2987                  * Unfortunately, we have to lie to MDC/MDS to retrieve
2988                  * attributes llite needs.
2989                 */
2990                 if (minfo->mi_it.it_op & IT_LOOKUP)
2991                         minfo->mi_it.it_op = IT_GETATTR;
2992         }
2993
2994         if (IS_ERR(tgt))
2995                 RETURN(PTR_ERR(tgt));
2996
2997         rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
2998         RETURN(rc);
2999 }
3000
3001 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3002                         struct lu_fid *fid, __u32 *bits)
3003 {
3004         struct obd_device       *obd = exp->exp_obd;
3005         struct lmv_obd          *lmv = &obd->u.lmv;
3006         struct lmv_tgt_desc     *tgt;
3007         int                      rc;
3008         ENTRY;
3009
3010         rc = lmv_check_connect(obd);
3011         if (rc)
3012                 RETURN(rc);
3013
3014         tgt = lmv_find_target(lmv, fid);
3015         if (IS_ERR(tgt))
3016                 RETURN(PTR_ERR(tgt));
3017
3018         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3019         RETURN(rc);
3020 }
3021
3022
3023 struct obd_ops lmv_obd_ops = {
3024         .o_owner                = THIS_MODULE,
3025         .o_setup                = lmv_setup,
3026         .o_cleanup              = lmv_cleanup,
3027         .o_precleanup           = lmv_precleanup,
3028         .o_process_config       = lmv_process_config,
3029         .o_connect              = lmv_connect,
3030         .o_disconnect           = lmv_disconnect,
3031         .o_statfs               = lmv_statfs,
3032         .o_get_info             = lmv_get_info,
3033         .o_set_info_async       = lmv_set_info_async,
3034         .o_packmd               = lmv_packmd,
3035         .o_unpackmd             = lmv_unpackmd,
3036         .o_notify               = lmv_notify,
3037         .o_get_uuid             = lmv_get_uuid,
3038         .o_iocontrol            = lmv_iocontrol,
3039         .o_fid_delete           = lmv_fid_delete
3040 };
3041
3042 struct md_ops lmv_md_ops = {
3043         .m_getstatus            = lmv_getstatus,
3044         .m_change_cbdata        = lmv_change_cbdata,
3045         .m_find_cbdata          = lmv_find_cbdata,
3046         .m_close                = lmv_close,
3047         .m_create               = lmv_create,
3048         .m_done_writing         = lmv_done_writing,
3049         .m_enqueue              = lmv_enqueue,
3050         .m_getattr              = lmv_getattr,
3051         .m_getxattr             = lmv_getxattr,
3052         .m_getattr_name         = lmv_getattr_name,
3053         .m_intent_lock          = lmv_intent_lock,
3054         .m_link                 = lmv_link,
3055         .m_rename               = lmv_rename,
3056         .m_setattr              = lmv_setattr,
3057         .m_setxattr             = lmv_setxattr,
3058         .m_sync                 = lmv_sync,
3059         .m_readpage             = lmv_readpage,
3060         .m_unlink               = lmv_unlink,
3061         .m_init_ea_size         = lmv_init_ea_size,
3062         .m_cancel_unused        = lmv_cancel_unused,
3063         .m_set_lock_data        = lmv_set_lock_data,
3064         .m_lock_match           = lmv_lock_match,
3065         .m_get_lustre_md        = lmv_get_lustre_md,
3066         .m_free_lustre_md       = lmv_free_lustre_md,
3067         .m_set_open_replay_data = lmv_set_open_replay_data,
3068         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3069         .m_renew_capa           = lmv_renew_capa,
3070         .m_unpack_capa          = lmv_unpack_capa,
3071         .m_get_remote_perm      = lmv_get_remote_perm,
3072         .m_intent_getattr_async = lmv_intent_getattr_async,
3073         .m_revalidate_lock      = lmv_revalidate_lock
3074 };
3075
3076 static quota_interface_t *quota_interface;
3077 extern quota_interface_t lmv_quota_interface;
3078
3079 int __init lmv_init(void)
3080 {
3081         struct lprocfs_static_vars lvars;
3082         int                        rc;
3083
3084         lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3085                                                 sizeof(struct lmv_object),
3086                                                 0, 0);
3087         if (!lmv_object_cache) {
3088                 CERROR("Error allocating lmv objects cache\n");
3089                 return -ENOMEM;
3090         }
3091
3092         lprocfs_lmv_init_vars(&lvars);
3093
3094         cfs_request_module("lquota");
3095         quota_interface = PORTAL_SYMBOL_GET(lmv_quota_interface);
3096         init_obd_quota_ops(quota_interface, &lmv_obd_ops);
3097
3098         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3099                                  lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3100         if (rc) {
3101                 if (quota_interface)
3102                         PORTAL_SYMBOL_PUT(lmv_quota_interface);
3103                 cfs_mem_cache_destroy(lmv_object_cache);
3104         }
3105
3106         return rc;
3107 }
3108
3109 #ifdef __KERNEL__
3110 static void lmv_exit(void)
3111 {
3112         if (quota_interface)
3113                 PORTAL_SYMBOL_PUT(lmv_quota_interface);
3114
3115         class_unregister_type(LUSTRE_LMV_NAME);
3116
3117         LASSERTF(cfs_atomic_read(&lmv_object_count) == 0,
3118                  "Can't free lmv objects cache, %d object(s) busy\n",
3119                  cfs_atomic_read(&lmv_object_count));
3120         cfs_mem_cache_destroy(lmv_object_cache);
3121 }
3122
3123 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3124 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3125 MODULE_LICENSE("GPL");
3126
3127 module_init(lmv_init);
3128 module_exit(lmv_exit);
3129 #endif