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                 /* copy UUID */
 752                 if (cfs_copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 753                                      min((int) data->ioc_plen2,
 754                                          (int) sizeof(struct obd_uuid))))
 755                         RETURN(-EFAULT);
 756
 757                 rc = obd_statfs(mdc_obd, &stat_buf,
 758                                 cfs_time_current_64() - CFS_HZ, 0);
 759                 if (rc)
 760                         RETURN(rc);
 761                 if (cfs_copy_to_user(data->ioc_pbuf1, &stat_buf,
 762                                      min((int) data->ioc_plen1,
 763                                          (int) sizeof(stat_buf))))
 764                         RETURN(-EFAULT);
 765                 break;
 766         }
 767         case OBD_IOC_QUOTACTL: {
 768                 struct if_quotactl *qctl = karg;
 769                 struct lmv_tgt_desc *tgt = NULL;
 770                 struct obd_quotactl *oqctl;
 771
 772                 if (qctl->qc_valid == QC_MDTIDX) {
 773                         if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
 774                                 RETURN(-EINVAL);
 775
 776                         tgt = &lmv->tgts[qctl->qc_idx];
 777                         if (!tgt->ltd_exp)
 778                                 RETURN(-EINVAL);
 779                 } else if (qctl->qc_valid == QC_UUID) {
 780                         for (i = 0; i < count; i++) {
 781                                 tgt = &lmv->tgts[i];
 782                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 783                                                      &qctl->obd_uuid))
 784                                         continue;
 785
 786                                 if (tgt->ltd_exp == NULL)
 787                                         RETURN(-EINVAL);
 788
 789                                 break;
 790                         }
 791                 } else {
 792                         RETURN(-EINVAL);
 793                 }
 794
 795                 if (i >= count)
 796                         RETURN(-EAGAIN);
 797
 798                 LASSERT(tgt && tgt->ltd_exp);
 799                 OBD_ALLOC_PTR(oqctl);
 800                 if (!oqctl)
 801                         RETURN(-ENOMEM);
 802
 803                 QCTL_COPY(oqctl, qctl);
 804                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 805                 if (rc == 0) {
 806                         QCTL_COPY(qctl, oqctl);
 807                         qctl->qc_valid = QC_MDTIDX;
 808                         qctl->obd_uuid = tgt->ltd_uuid;
 809                 }
 810                 OBD_FREE_PTR(oqctl);
 811                 break;
 812         }
 813         case OBD_IOC_CHANGELOG_SEND:
 814         case OBD_IOC_CHANGELOG_CLEAR: {
 815                 struct ioc_changelog *icc = karg;
 816
 817                 if (icc->icc_mdtindex >= count)
 818                         RETURN(-ENODEV);
 819
 820                 rc = obd_iocontrol(cmd, lmv->tgts[icc->icc_mdtindex].ltd_exp,
 821                                    sizeof(*icc), icc, NULL);
 822                 break;
 823         }
 824
 825         default : {
 826                 for (i = 0; i < count; i++) {
 827                         int err;
 828
 829                         if (lmv->tgts[i].ltd_exp == NULL)
 830                                 continue;
 831
 832                         err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
 833                                             karg, uarg);
 834                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
 835                                 RETURN(err);
 836                         } else if (err) {
 837                                 if (lmv->tgts[i].ltd_active) {
 838                                         CERROR("error: iocontrol MDC %s on MDT"
 839                                                "idx %d cmd %x: err = %d\n",
 840                                                 lmv->tgts[i].ltd_uuid.uuid,
 841                                                 i, cmd, err);
 842                                         if (!rc)
 843                                                 rc = err;
 844                                 }
 845                         } else
 846                                 set = 1;
 847                 }
 848                 if (!set && !rc)
 849                         rc = -EIO;
 850         }
 851         }
 852         RETURN(rc);
 853 }
 854
 855 static int lmv_all_chars_policy(int count, const char *name,
 856                                 int len)
 857 {
 858         unsigned int c = 0;
 859
 860         while (len > 0)
 861                 c += name[--len];
 862         c = c % count;
 863         return c;
 864 }
 865
 866 static int lmv_nid_policy(struct lmv_obd *lmv)
 867 {
 868         struct obd_import *imp;
 869         __u32              id;
 870
 871         /*
 872          * XXX: To get nid we assume that underlying obd device is mdc.
 873          */
 874         imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
 875         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
 876         return id % lmv->desc.ld_tgt_count;
 877 }
 878
 879 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
 880                           placement_policy_t placement)
 881 {
 882         switch (placement) {
 883         case PLACEMENT_CHAR_POLICY:
 884                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
 885                                             op_data->op_name,
 886                                             op_data->op_namelen);
 887         case PLACEMENT_NID_POLICY:
 888                 return lmv_nid_policy(lmv);
 889
 890         default:
 891                 break;
 892         }
 893
 894         CERROR("Unsupported placement policy %x\n", placement);
 895         return -EINVAL;
 896 }
 897
 898 /**
 899  * This is _inode_ placement policy function (not name).
 900  */
 901 static int lmv_placement_policy(struct obd_device *obd,
 902                                 struct md_op_data *op_data,
 903                                 mdsno_t *mds)
 904 {
 905         struct lmv_obd          *lmv = &obd->u.lmv;
 906         struct lmv_object       *obj;
 907         int                      rc;
 908         ENTRY;
 909
 910         LASSERT(mds != NULL);
 911
 912         if (lmv->desc.ld_tgt_count == 1) {
 913                 *mds = 0;
 914                 RETURN(0);
 915         }
 916
 917         /*
 918          * Allocate new fid on target according to operation type and parent
 919          * home mds.
 920          */
 921         obj = lmv_object_find(obd, &op_data->op_fid1);
 922         if (obj != NULL || op_data->op_name == NULL ||
 923             op_data->op_opc != LUSTRE_OPC_MKDIR) {
 924                 /*
 925                  * Allocate fid for non-dir or for null name or for case parent
 926                  * dir is split.
 927                  */
 928                 if (obj) {
 929                         lmv_object_put(obj);
 930
 931                         /*
 932                          * If we have this flag turned on, and we see that
 933                          * parent dir is split, this means, that caller did not
 934                          * notice split yet. This is race and we would like to
 935                          * let caller know that.
 936                          */
 937                         if (op_data->op_bias & MDS_CHECK_SPLIT)
 938                                 RETURN(-ERESTART);
 939                 }
 940
 941                 /*
 942                  * Allocate new fid on same mds where parent fid is located and
 943                  * where operation will be sent. In case of split dir, ->op_fid1
 944                  * and ->op_mds here will contain fid and mds of slave directory
 945                  * object (assigned by caller).
 946                  */
 947                 *mds = op_data->op_mds;
 948                 rc = 0;
 949         } else {
 950                 /*
 951                  * Parent directory is not split and we want to create a
 952                  * directory in it. Let's calculate where to place it according
 953                  * to operation data @op_data.
 954                  */
 955                 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
 956                 rc = 0;
 957         }
 958
 959         if (rc) {
 960                 CERROR("Can't choose MDS, err = %d\n", rc);
 961         } else {
 962                 LASSERT(*mds < lmv->desc.ld_tgt_count);
 963         }
 964
 965         RETURN(rc);
 966 }
 967
 968 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
 969                     mdsno_t mds)
 970 {
 971         struct lmv_tgt_desc *tgt;
 972         int                  rc;
 973         ENTRY;
 974
 975         tgt = lmv_get_target(lmv, mds);
 976
 977         /*
 978          * New seq alloc and FLD setup should be atomic. Otherwise we may find
 979          * on server that seq in new allocated fid is not yet known.
 980          */
 981         cfs_down(&tgt->ltd_fid_sem);
 982
 983         if (!tgt->ltd_active)
 984                 GOTO(out, rc = -ENODEV);
 985
 986         /*
 987          * Asking underlaying tgt layer to allocate new fid.
 988          */
 989         rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
 990         if (rc > 0) {
 991                 LASSERT(fid_is_sane(fid));
 992                 rc = 0;
 993         }
 994
 995         EXIT;
 996 out:
 997         cfs_up(&tgt->ltd_fid_sem);
 998         return rc;
 999 }
1000
1001 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
1002                   struct md_op_data *op_data)
1003 {
1004         struct obd_device     *obd = class_exp2obd(exp);
1005         struct lmv_obd        *lmv = &obd->u.lmv;
1006         mdsno_t                mds;
1007         int                    rc;
1008         ENTRY;
1009
1010         LASSERT(op_data != NULL);
1011         LASSERT(fid != NULL);
1012
1013         rc = lmv_placement_policy(obd, op_data, &mds);
1014         if (rc) {
1015                 CERROR("Can't get target for allocating fid, "
1016                        "rc %d\n", rc);
1017                 RETURN(rc);
1018         }
1019
1020         rc = __lmv_fid_alloc(lmv, fid, mds);
1021         if (rc) {
1022                 CERROR("Can't alloc new fid, rc %d\n", rc);
1023                 RETURN(rc);
1024         }
1025
1026         RETURN(rc);
1027 }
1028
1029 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1030 {
1031         ENTRY;
1032         LASSERT(exp != NULL && fid != NULL);
1033         if (lmv_object_delete(exp, fid)) {
1034                 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1035                        PFID(fid));
1036         }
1037         RETURN(0);
1038 }
1039
1040 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1041 {
1042         struct lmv_obd             *lmv = &obd->u.lmv;
1043         struct lprocfs_static_vars  lvars;
1044         struct lmv_desc            *desc;
1045         int                         rc;
1046         int                         i = 0;
1047         ENTRY;
1048
1049         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1050                 CERROR("LMV setup requires a descriptor\n");
1051                 RETURN(-EINVAL);
1052         }
1053
1054         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1055         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1056                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1057                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1058                 RETURN(-EINVAL);
1059         }
1060
1061         lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1062
1063         OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1064         if (lmv->tgts == NULL)
1065                 RETURN(-ENOMEM);
1066
1067         for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1068                 cfs_sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
1069                 lmv->tgts[i].ltd_idx = i;
1070         }
1071
1072         lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1073
1074         OBD_ALLOC(lmv->datas, lmv->datas_size);
1075         if (lmv->datas == NULL)
1076                 GOTO(out_free_tgts, rc = -ENOMEM);
1077
1078         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1079         lmv->desc.ld_tgt_count = 0;
1080         lmv->desc.ld_active_tgt_count = 0;
1081         lmv->max_cookiesize = 0;
1082         lmv->max_def_easize = 0;
1083         lmv->max_easize = 0;
1084         lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1085
1086         cfs_spin_lock_init(&lmv->lmv_lock);
1087         cfs_sema_init(&lmv->init_sem, 1);
1088
1089         rc = lmv_object_setup(obd);
1090         if (rc) {
1091                 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1092                 GOTO(out_free_datas, rc);
1093         }
1094
1095         lprocfs_lmv_init_vars(&lvars);
1096         lprocfs_obd_setup(obd, lvars.obd_vars);
1097 #ifdef LPROCFS
1098         {
1099                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1100                                         0444, &lmv_proc_target_fops, obd);
1101                 if (rc)
1102                         CWARN("Error adding target_obd_stats file (%d)\n", rc);
1103        }
1104 #endif
1105         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1106                              LUSTRE_CLI_FLD_HASH_DHT);
1107         if (rc) {
1108                 CERROR("Can't init FLD, err %d\n", rc);
1109                 GOTO(out_free_datas, rc);
1110         }
1111
1112         RETURN(0);
1113
1114 out_free_datas:
1115         OBD_FREE(lmv->datas, lmv->datas_size);
1116         lmv->datas = NULL;
1117 out_free_tgts:
1118         OBD_FREE(lmv->tgts, lmv->tgts_size);
1119         lmv->tgts = NULL;
1120         return rc;
1121 }
1122
1123 static int lmv_cleanup(struct obd_device *obd)
1124 {
1125         struct lmv_obd   *lmv = &obd->u.lmv;
1126         ENTRY;
1127
1128         fld_client_fini(&lmv->lmv_fld);
1129         lprocfs_obd_cleanup(obd);
1130         lmv_object_cleanup(obd);
1131         OBD_FREE(lmv->datas, lmv->datas_size);
1132         OBD_FREE(lmv->tgts, lmv->tgts_size);
1133
1134         RETURN(0);
1135 }
1136
1137 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1138 {
1139         struct lustre_cfg     *lcfg = buf;
1140         struct obd_uuid        tgt_uuid;
1141         int                    rc;
1142         ENTRY;
1143
1144         switch(lcfg->lcfg_command) {
1145         case LCFG_ADD_MDC:
1146                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1147                         GOTO(out, rc = -EINVAL);
1148
1149                 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1150                 rc = lmv_add_target(obd, &tgt_uuid);
1151                 GOTO(out, rc);
1152         default: {
1153                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1154                 GOTO(out, rc = -EINVAL);
1155         }
1156         }
1157 out:
1158         RETURN(rc);
1159 }
1160
1161 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1162                       __u64 max_age, __u32 flags)
1163 {
1164         struct lmv_obd        *lmv = &obd->u.lmv;
1165         struct obd_statfs     *temp;
1166         int                    rc = 0;
1167         int                    i;
1168         ENTRY;
1169
1170         rc = lmv_check_connect(obd);
1171         if (rc)
1172                 RETURN(rc);
1173
1174         OBD_ALLOC(temp, sizeof(*temp));
1175         if (temp == NULL)
1176                 RETURN(-ENOMEM);
1177
1178         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1179                 if (lmv->tgts[i].ltd_exp == NULL)
1180                         continue;
1181
1182                 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1183                                 max_age, flags);
1184                 if (rc) {
1185                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1186                                lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1187                                rc);
1188                         GOTO(out_free_temp, rc);
1189                 }
1190                 if (i == 0) {
1191                         *osfs = *temp;
1192                 } else {
1193                         osfs->os_bavail += temp->os_bavail;
1194                         osfs->os_blocks += temp->os_blocks;
1195                         osfs->os_ffree += temp->os_ffree;
1196                         osfs->os_files += temp->os_files;
1197                 }
1198         }
1199
1200         EXIT;
1201 out_free_temp:
1202         OBD_FREE(temp, sizeof(*temp));
1203         return rc;
1204 }
1205
1206 static int lmv_getstatus(struct obd_export *exp,
1207                          struct lu_fid *fid,
1208                          struct obd_capa **pc)
1209 {
1210         struct obd_device    *obd = exp->exp_obd;
1211         struct lmv_obd       *lmv = &obd->u.lmv;
1212         int                   rc;
1213         ENTRY;
1214
1215         rc = lmv_check_connect(obd);
1216         if (rc)
1217                 RETURN(rc);
1218
1219         rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1220         RETURN(rc);
1221 }
1222
1223 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1224                         struct obd_capa *oc, obd_valid valid, const char *name,
1225                         const char *input, int input_size, int output_size,
1226                         int flags, struct ptlrpc_request **request)
1227 {
1228         struct obd_device      *obd = exp->exp_obd;
1229         struct lmv_obd         *lmv = &obd->u.lmv;
1230         struct lmv_tgt_desc    *tgt;
1231         int                     rc;
1232         ENTRY;
1233
1234         rc = lmv_check_connect(obd);
1235         if (rc)
1236                 RETURN(rc);
1237
1238         tgt = lmv_find_target(lmv, fid);
1239         if (IS_ERR(tgt))
1240                 RETURN(PTR_ERR(tgt));
1241
1242         rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1243                          input_size, output_size, flags, request);
1244
1245         RETURN(rc);
1246 }
1247
1248 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1249                         struct obd_capa *oc, obd_valid valid, const char *name,
1250                         const char *input, int input_size, int output_size,
1251                         int flags, __u32 suppgid,
1252                         struct ptlrpc_request **request)
1253 {
1254         struct obd_device      *obd = exp->exp_obd;
1255         struct lmv_obd         *lmv = &obd->u.lmv;
1256         struct lmv_tgt_desc    *tgt;
1257         int                     rc;
1258         ENTRY;
1259
1260         rc = lmv_check_connect(obd);
1261         if (rc)
1262                 RETURN(rc);
1263
1264         tgt = lmv_find_target(lmv, fid);
1265         if (IS_ERR(tgt))
1266                 RETURN(PTR_ERR(tgt));
1267
1268         rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1269                          input_size, output_size, flags, suppgid,
1270                          request);
1271
1272         RETURN(rc);
1273 }
1274
1275 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1276                        struct ptlrpc_request **request)
1277 {
1278         struct obd_device       *obd = exp->exp_obd;
1279         struct lmv_obd          *lmv = &obd->u.lmv;
1280         struct lmv_tgt_desc     *tgt;
1281         struct lmv_object       *obj;
1282         int                      rc;
1283         int                      i;
1284         ENTRY;
1285
1286         rc = lmv_check_connect(obd);
1287         if (rc)
1288                 RETURN(rc);
1289
1290         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1291         if (IS_ERR(tgt))
1292                 RETURN(PTR_ERR(tgt));
1293
1294         if (op_data->op_valid & OBD_MD_MDTIDX) {
1295                 op_data->op_mds = tgt->ltd_idx;
1296                 RETURN(0);
1297         }
1298
1299         rc = md_getattr(tgt->ltd_exp, op_data, request);
1300         if (rc)
1301                 RETURN(rc);
1302
1303         obj = lmv_object_find_lock(obd, &op_data->op_fid1);
1304
1305         CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(&op_data->op_fid1),
1306                obj ? "(split)" : "");
1307
1308         /*
1309          * If object is split, then we loop over all the slaves and gather size
1310          * attribute. In ideal world we would have to gather also mds field from
1311          * all slaves, as object is spread over the cluster and this is
1312          * definitely interesting information and it is not good to loss it,
1313          * but...
1314          */
1315         if (obj) {
1316                 struct mdt_body *body;
1317
1318                 if (*request == NULL) {
1319                         lmv_object_put(obj);
1320                         RETURN(rc);
1321                 }
1322
1323                 body = req_capsule_server_get(&(*request)->rq_pill,
1324                                               &RMF_MDT_BODY);
1325                 LASSERT(body != NULL);
1326
1327                 for (i = 0; i < obj->lo_objcount; i++) {
1328                         if (lmv->tgts[i].ltd_exp == NULL) {
1329                                 CWARN("%s: NULL export for %d\n",
1330                                       obd->obd_name, i);
1331                                 continue;
1332                         }
1333
1334                         /*
1335                          * Skip master object.
1336                          */
1337                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1338                                 continue;
1339
1340                         body->size += obj->lo_stripes[i].ls_size;
1341                 }
1342
1343                 lmv_object_put_unlock(obj);
1344         }
1345
1346         RETURN(rc);
1347 }
1348
1349 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1350                              ldlm_iterator_t it, void *data)
1351 {
1352         struct obd_device   *obd = exp->exp_obd;
1353         struct lmv_obd      *lmv = &obd->u.lmv;
1354         int                  i;
1355         int                  rc;
1356         ENTRY;
1357
1358         rc = lmv_check_connect(obd);
1359         if (rc)
1360                 RETURN(rc);
1361
1362         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1363
1364         /*
1365          * With CMD every object can have two locks in different namespaces:
1366          * lookup lock in space of mds storing direntry and update/open lock in
1367          * space of mds storing inode.
1368          */
1369         for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1370                 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1371
1372         RETURN(0);
1373 }
1374
1375 static int lmv_find_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1376                            ldlm_iterator_t it, void *data)
1377 {
1378         struct obd_device   *obd = exp->exp_obd;
1379         struct lmv_obd      *lmv = &obd->u.lmv;
1380         int                  i;
1381         int                  rc;
1382         ENTRY;
1383
1384         rc = lmv_check_connect(obd);
1385         if (rc)
1386                 RETURN(rc);
1387
1388         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1389
1390         /*
1391          * With CMD every object can have two locks in different namespaces:
1392          * lookup lock in space of mds storing direntry and update/open lock in
1393          * space of mds storing inode.
1394          */
1395         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1396                 rc = md_find_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1397                 if (rc)
1398                         RETURN(rc);
1399         }
1400
1401         RETURN(rc);
1402 }
1403
1404
1405 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1406                      struct md_open_data *mod, struct ptlrpc_request **request)
1407 {
1408         struct obd_device     *obd = exp->exp_obd;
1409         struct lmv_obd        *lmv = &obd->u.lmv;
1410         struct lmv_tgt_desc   *tgt;
1411         int                    rc;
1412         ENTRY;
1413
1414         rc = lmv_check_connect(obd);
1415         if (rc)
1416                 RETURN(rc);
1417
1418         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1419         if (IS_ERR(tgt))
1420                 RETURN(PTR_ERR(tgt));
1421
1422         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1423         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1424         RETURN(rc);
1425 }
1426
1427 /**
1428  * Called in the case MDS returns -ERESTART on create on open, what means that
1429  * directory is split and its LMV presentation object has to be updated.
1430  */
1431 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1432 {
1433         struct obd_device       *obd = exp->exp_obd;
1434         struct lmv_obd          *lmv = &obd->u.lmv;
1435         struct ptlrpc_request   *req = NULL;
1436         struct lmv_tgt_desc     *tgt;
1437         struct lmv_object       *obj;
1438         struct lustre_md         md;
1439         struct md_op_data       *op_data;
1440         int                      mealen;
1441         int                      rc;
1442         __u64                    valid;
1443         ENTRY;
1444
1445         md.mea = NULL;
1446         mealen = lmv_get_easize(lmv);
1447
1448         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1449
1450         tgt = lmv_find_target(lmv, fid);
1451         if (IS_ERR(tgt))
1452                 RETURN(PTR_ERR(tgt));
1453
1454         /*
1455          * Time to update mea of parent fid.
1456          */
1457
1458         OBD_ALLOC_PTR(op_data);
1459         if (op_data == NULL)
1460                 RETURN(-ENOMEM);
1461
1462         op_data->op_fid1 = *fid;
1463         op_data->op_mode = mealen;
1464         op_data->op_valid = valid;
1465
1466         rc = md_getattr(tgt->ltd_exp, op_data, &req);
1467         OBD_FREE_PTR(op_data);
1468         if (rc) {
1469                 CERROR("md_getattr() failed, error %d\n", rc);
1470                 GOTO(cleanup, rc);
1471         }
1472
1473         rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1474         if (rc) {
1475                 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1476                 GOTO(cleanup, rc);
1477         }
1478
1479         if (md.mea == NULL)
1480                 GOTO(cleanup, rc = -ENODATA);
1481
1482         obj = lmv_object_create(exp, fid, md.mea);
1483         if (IS_ERR(obj))
1484                 rc = PTR_ERR(obj);
1485         else
1486                 lmv_object_put(obj);
1487
1488         obd_free_memmd(exp, (void *)&md.mea);
1489         EXIT;
1490 cleanup:
1491         if (req)
1492                 ptlrpc_req_finished(req);
1493         return rc;
1494 }
1495
1496 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1497                const void *data, int datalen, int mode, __u32 uid,
1498                __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1499                struct ptlrpc_request **request)
1500 {
1501         struct obd_device       *obd = exp->exp_obd;
1502         struct lmv_obd          *lmv = &obd->u.lmv;
1503         struct lmv_tgt_desc     *tgt;
1504         struct lmv_object       *obj;
1505         int                      rc;
1506         int                      loop = 0;
1507         int                      sidx;
1508         ENTRY;
1509
1510         rc = lmv_check_connect(obd);
1511         if (rc)
1512                 RETURN(rc);
1513
1514         if (!lmv->desc.ld_active_tgt_count)
1515                 RETURN(-EIO);
1516 repeat:
1517         ++loop;
1518         LASSERT(loop <= 2);
1519
1520         obj = lmv_object_find(obd, &op_data->op_fid1);
1521         if (obj) {
1522                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1523                                        op_data->op_name, op_data->op_namelen);
1524                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1525                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1526                 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1527                 tgt = lmv_get_target(lmv, op_data->op_mds);
1528                 lmv_object_put(obj);
1529         } else {
1530                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1531                 op_data->op_bias |= MDS_CHECK_SPLIT;
1532                 op_data->op_mds = tgt->ltd_idx;
1533         }
1534
1535         if (IS_ERR(tgt))
1536                 RETURN(PTR_ERR(tgt));
1537
1538         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1539         if (rc == -ERESTART)
1540                 goto repeat;
1541         else if (rc)
1542                 RETURN(rc);
1543
1544         CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #%x\n",
1545                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1546                op_data->op_mds);
1547
1548         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1549         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1550                        cap_effective, rdev, request);
1551         if (rc == 0) {
1552                 if (*request == NULL)
1553                         RETURN(rc);
1554                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1555         } else if (rc == -ERESTART) {
1556                 LASSERT(*request != NULL);
1557                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1558                           "Got -ERESTART during create!\n");
1559                 ptlrpc_req_finished(*request);
1560                 *request = NULL;
1561
1562                 /*
1563                  * Directory got split. Time to update local object and repeat
1564                  * the request with proper MDS.
1565                  */
1566                 rc = lmv_handle_split(exp, &op_data->op_fid1);
1567                 if (rc == 0) {
1568                         rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1569                                                  op_data, &op_data->op_fid2);
1570                         if (rc)
1571                                 RETURN(rc);
1572                         goto repeat;
1573                 }
1574         }
1575         RETURN(rc);
1576 }
1577
1578 static int lmv_done_writing(struct obd_export *exp,
1579                             struct md_op_data *op_data,
1580                             struct md_open_data *mod)
1581 {
1582         struct obd_device     *obd = exp->exp_obd;
1583         struct lmv_obd        *lmv = &obd->u.lmv;
1584         struct lmv_tgt_desc   *tgt;
1585         int                    rc;
1586         ENTRY;
1587
1588         rc = lmv_check_connect(obd);
1589         if (rc)
1590                 RETURN(rc);
1591
1592         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1593         if (IS_ERR(tgt))
1594                 RETURN(PTR_ERR(tgt));
1595
1596         rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1597         RETURN(rc);
1598 }
1599
1600 static int
1601 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1602                    struct lookup_intent *it, struct md_op_data *op_data,
1603                    struct lustre_handle *lockh, void *lmm, int lmmsize)
1604 {
1605         struct obd_device     *obd = exp->exp_obd;
1606         struct lmv_obd        *lmv = &obd->u.lmv;
1607         struct lmv_stripe_md  *mea = op_data->op_mea1;
1608         struct md_op_data     *op_data2;
1609         struct lmv_tgt_desc   *tgt;
1610         int                    i;
1611         int                    rc = 0;
1612         ENTRY;
1613
1614         OBD_ALLOC_PTR(op_data2);
1615         if (op_data2 == NULL)
1616                 RETURN(-ENOMEM);
1617
1618         LASSERT(mea != NULL);
1619         for (i = 0; i < mea->mea_count; i++) {
1620                 memset(op_data2, 0, sizeof(*op_data2));
1621                 op_data2->op_fid1 = mea->mea_ids[i];
1622                 op_data2->op_bias = 0;
1623
1624                 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1625                 if (IS_ERR(tgt))
1626                         GOTO(cleanup, rc = PTR_ERR(tgt));
1627
1628                 if (tgt->ltd_exp == NULL)
1629                         continue;
1630
1631                 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1632                                 lockh + i, lmm, lmmsize, NULL, 0);
1633
1634                 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1635                        PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1636
1637                 if (rc)
1638                         GOTO(cleanup, rc);
1639
1640                 if (it->d.lustre.it_data) {
1641                         struct ptlrpc_request *req;
1642                         req = (struct ptlrpc_request *)it->d.lustre.it_data;
1643                         ptlrpc_req_finished(req);
1644                 }
1645
1646                 if (it->d.lustre.it_status)
1647                         GOTO(cleanup, rc = it->d.lustre.it_status);
1648         }
1649
1650         EXIT;
1651 cleanup:
1652         OBD_FREE_PTR(op_data2);
1653
1654         if (rc != 0) {
1655                 /*
1656                  * Drop all taken locks.
1657                  */
1658                 while (--i >= 0) {
1659                         if (lockh[i].cookie)
1660                                 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1661                         lockh[i].cookie = 0;
1662                 }
1663         }
1664         return rc;
1665 }
1666
1667 static int
1668 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1669                    struct lookup_intent *it, struct md_op_data *op_data,
1670                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1671                    int extra_lock_flags)
1672 {
1673         struct ptlrpc_request      *req = it->d.lustre.it_data;
1674         struct obd_device          *obd = exp->exp_obd;
1675         struct lmv_obd             *lmv = &obd->u.lmv;
1676         struct lustre_handle        plock;
1677         struct lmv_tgt_desc        *tgt;
1678         struct md_op_data          *rdata;
1679         struct lu_fid               fid1;
1680         struct mdt_body            *body;
1681         int                         rc = 0;
1682         int                         pmode;
1683         ENTRY;
1684
1685         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1686         LASSERT(body != NULL);
1687
1688         if (!(body->valid & OBD_MD_MDS))
1689                 RETURN(0);
1690
1691         CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1692                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1693
1694         /*
1695          * We got LOOKUP lock, but we really need attrs.
1696          */
1697         pmode = it->d.lustre.it_lock_mode;
1698         LASSERT(pmode != 0);
1699         memcpy(&plock, lockh, sizeof(plock));
1700         it->d.lustre.it_lock_mode = 0;
1701         it->d.lustre.it_data = NULL;
1702         fid1 = body->fid1;
1703
1704         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1705         ptlrpc_req_finished(req);
1706
1707         tgt = lmv_find_target(lmv, &fid1);
1708         if (IS_ERR(tgt))
1709                 GOTO(out, rc = PTR_ERR(tgt));
1710
1711         OBD_ALLOC_PTR(rdata);
1712         if (rdata == NULL)
1713                 GOTO(out, rc = -ENOMEM);
1714
1715         rdata->op_fid1 = fid1;
1716         rdata->op_bias = MDS_CROSS_REF;
1717
1718         rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1719                         lmm, lmmsize, NULL, extra_lock_flags);
1720         OBD_FREE_PTR(rdata);
1721         EXIT;
1722 out:
1723         ldlm_lock_decref(&plock, pmode);
1724         return rc;
1725 }
1726
1727 static int
1728 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1729             struct lookup_intent *it, struct md_op_data *op_data,
1730             struct lustre_handle *lockh, void *lmm, int lmmsize,
1731             struct ptlrpc_request **req, int extra_lock_flags)
1732 {
1733         struct obd_device        *obd = exp->exp_obd;
1734         struct lmv_obd           *lmv = &obd->u.lmv;
1735         struct lmv_tgt_desc      *tgt;
1736         struct lmv_object        *obj;
1737         int                       sidx;
1738         int                       rc;
1739         ENTRY;
1740
1741         rc = lmv_check_connect(obd);
1742         if (rc)
1743                 RETURN(rc);
1744
1745         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1746                LL_IT2STR(it), PFID(&op_data->op_fid1));
1747
1748         if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1749                 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1750                                         lockh, lmm, lmmsize);
1751                 RETURN(rc);
1752         }
1753
1754         obj = lmv_object_find(obd, &op_data->op_fid1);
1755         if (obj && op_data->op_namelen) {
1756                 sidx = raw_name2idx(obj->lo_hashtype,
1757                                        obj->lo_objcount,
1758                                        (char *)op_data->op_name,
1759                                        op_data->op_namelen);
1760                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1761                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1762         } else {
1763                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1764         }
1765         if (obj)
1766                 lmv_object_put(obj);
1767
1768         if (IS_ERR(tgt))
1769                 RETURN(PTR_ERR(tgt));
1770
1771         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1772                LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1773
1774         rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1775                         lmm, lmmsize, req, extra_lock_flags);
1776
1777         if (rc == 0 && it && it->it_op == IT_OPEN) {
1778                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1779                                         lmm, lmmsize, extra_lock_flags);
1780         }
1781         RETURN(rc);
1782 }
1783
1784 static int
1785 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1786                  struct ptlrpc_request **request)
1787 {
1788         struct ptlrpc_request   *req = NULL;
1789         struct obd_device       *obd = exp->exp_obd;
1790         struct lmv_obd          *lmv = &obd->u.lmv;
1791         struct lu_fid            rid = op_data->op_fid1;
1792         struct lmv_tgt_desc     *tgt;
1793         struct mdt_body         *body;
1794         struct lmv_object       *obj;
1795         obd_valid                valid = op_data->op_valid;
1796         int                      rc;
1797         int                      loop = 0;
1798         int                      sidx;
1799         ENTRY;
1800
1801         rc = lmv_check_connect(obd);
1802         if (rc)
1803                 RETURN(rc);
1804
1805 repeat:
1806         ++loop;
1807         LASSERT(loop <= 2);
1808         obj = lmv_object_find(obd, &rid);
1809         if (obj) {
1810                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1811                                     op_data->op_name, op_data->op_namelen);
1812                 rid = obj->lo_stripes[sidx].ls_fid;
1813                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1814                 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1815                 valid &= ~OBD_MD_FLCKSPLIT;
1816                 lmv_object_put(obj);
1817         } else {
1818                 tgt = lmv_find_target(lmv, &rid);
1819                 valid |= OBD_MD_FLCKSPLIT;
1820                 op_data->op_mds = tgt->ltd_idx;
1821         }
1822         if (IS_ERR(tgt))
1823                 RETURN(PTR_ERR(tgt));
1824
1825         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1826                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1827                PFID(&rid), tgt->ltd_idx);
1828
1829         op_data->op_valid = valid;
1830         op_data->op_fid1 = rid;
1831         rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1832         if (rc == 0) {
1833                 body = req_capsule_server_get(&(*request)->rq_pill,
1834                                               &RMF_MDT_BODY);
1835                 LASSERT(body != NULL);
1836
1837                 if (body->valid & OBD_MD_MDS) {
1838                         rid = body->fid1;
1839                         CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1840                                PFID(&rid));
1841
1842                         tgt = lmv_find_target(lmv, &rid);
1843                         if (IS_ERR(tgt)) {
1844                                 ptlrpc_req_finished(*request);
1845                                 RETURN(PTR_ERR(tgt));
1846                         }
1847
1848                         op_data->op_fid1 = rid;
1849                         op_data->op_valid |= OBD_MD_FLCROSSREF;
1850                         op_data->op_namelen = 0;
1851                         op_data->op_name = NULL;
1852                         rc = md_getattr_name(tgt->ltd_exp, op_data, &req);
1853                         ptlrpc_req_finished(*request);
1854                         *request = req;
1855                 }
1856         } else if (rc == -ERESTART) {
1857                 LASSERT(*request != NULL);
1858                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1859                           "Got -ERESTART during getattr!\n");
1860                 ptlrpc_req_finished(*request);
1861                 *request = NULL;
1862
1863                 /*
1864                  * Directory got split. Time to update local object and repeat
1865                  * the request with proper MDS.
1866                  */
1867                 rc = lmv_handle_split(exp, &rid);
1868                 if (rc == 0)
1869                         goto repeat;
1870         }
1871         RETURN(rc);
1872 }
1873
1874 #define md_op_data_fid(op_data, fl)                     \
1875         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1876          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1877          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1878          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1879          NULL)
1880
1881 static int lmv_early_cancel_slaves(struct obd_export *exp,
1882                                    struct md_op_data *op_data, int op_tgt,
1883                                    ldlm_mode_t mode, int bits, int flag)
1884 {
1885         struct obd_device      *obd = exp->exp_obd;
1886         struct lmv_obd         *lmv = &obd->u.lmv;
1887         ldlm_policy_data_t      policy = {{0}};
1888         struct lu_fid          *op_fid;
1889         struct lu_fid          *st_fid;
1890         struct lmv_tgt_desc    *tgt;
1891         struct lmv_object      *obj;
1892         int                     rc = 0;
1893         int                     i;
1894         ENTRY;
1895
1896         op_fid = md_op_data_fid(op_data, flag);
1897         if (!fid_is_sane(op_fid))
1898                 RETURN(0);
1899
1900         obj = lmv_object_find(obd, op_fid);
1901         if (obj == NULL)
1902                 RETURN(-EALREADY);
1903
1904         policy.l_inodebits.bits = bits;
1905         for (i = 0; i < obj->lo_objcount; i++) {
1906                 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1907                 st_fid = &obj->lo_stripes[i].ls_fid;
1908                 if (op_tgt != tgt->ltd_idx) {
1909                         CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1910                                PFID(st_fid), tgt->ltd_idx);
1911                         rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1912                                               mode, LDLM_FL_ASYNC, NULL);
1913                         if (rc)
1914                                 GOTO(out_put_obj, rc);
1915                 } else {
1916                         CDEBUG(D_INODE,
1917                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1918                                op_tgt, PFID(st_fid));
1919                         /*
1920                          * Do not cancel locks for operation target, they will
1921                          * be handled later in underlaying layer when calling
1922                          * function we run on behalf of.
1923                          */
1924                         *op_fid = *st_fid;
1925                         op_data->op_flags |= flag;
1926                 }
1927         }
1928         EXIT;
1929 out_put_obj:
1930         lmv_object_put(obj);
1931         return rc;
1932 }
1933
1934 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1935                             int op_tgt, ldlm_mode_t mode, int bits, int flag)
1936 {
1937         struct lu_fid          *fid = md_op_data_fid(op_data, flag);
1938         struct obd_device      *obd = exp->exp_obd;
1939         struct lmv_obd         *lmv = &obd->u.lmv;
1940         struct lmv_tgt_desc    *tgt;
1941         ldlm_policy_data_t      policy = {{0}};
1942         struct lmv_object      *obj;
1943         int                     rc = 0;
1944         ENTRY;
1945
1946         if (!fid_is_sane(fid))
1947                 RETURN(0);
1948
1949         obj = lmv_object_find(obd, fid);
1950         if (obj) {
1951                 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1952                                              bits, flag);
1953                 lmv_object_put(obj);
1954         } else {
1955                 tgt = lmv_find_target(lmv, fid);
1956                 if (IS_ERR(tgt))
1957                         RETURN(PTR_ERR(tgt));
1958
1959                 if (tgt->ltd_idx != op_tgt) {
1960                         CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1961                         policy.l_inodebits.bits = bits;
1962                         rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1963                                               mode, LDLM_FL_ASYNC, NULL);
1964                 } else {
1965                         CDEBUG(D_INODE,
1966                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1967                                op_tgt, PFID(fid));
1968                         op_data->op_flags |= flag;
1969                         rc = 0;
1970                 }
1971
1972         }
1973         RETURN(rc);
1974 }
1975
1976 /*
1977  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1978  * op_data->op_fid2
1979  */
1980 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1981                     struct ptlrpc_request **request)
1982 {
1983         struct obd_device       *obd = exp->exp_obd;
1984         struct lmv_obd          *lmv = &obd->u.lmv;
1985         struct lmv_tgt_desc     *tgt;
1986         struct lmv_object       *obj;
1987         int                      rc;
1988         int                      loop = 0;
1989         mdsno_t                  mds;
1990         int                      sidx;
1991         ENTRY;
1992
1993         rc = lmv_check_connect(obd);
1994         if (rc)
1995                 RETURN(rc);
1996
1997 repeat:
1998         ++loop;
1999         LASSERT(loop <= 2);
2000         LASSERT(op_data->op_namelen != 0);
2001
2002         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2003                PFID(&op_data->op_fid2), op_data->op_namelen,
2004                op_data->op_name, PFID(&op_data->op_fid1));
2005
2006         obj = lmv_object_find(obd, &op_data->op_fid2);
2007         if (obj) {
2008                 sidx = raw_name2idx(obj->lo_hashtype,
2009                                     obj->lo_objcount,
2010                                     op_data->op_name,
2011                                     op_data->op_namelen);
2012                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2013                 mds = obj->lo_stripes[sidx].ls_mds;
2014                 lmv_object_put(obj);
2015         } else {
2016                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
2017                 if (rc)
2018                         RETURN(rc);
2019         }
2020
2021         CDEBUG(D_INODE, "Forward to mds #%x ("DFID")\n",
2022                mds, PFID(&op_data->op_fid1));
2023
2024         op_data->op_fsuid = cfs_curproc_fsuid();
2025         op_data->op_fsgid = cfs_curproc_fsgid();
2026         op_data->op_cap = cfs_curproc_cap_pack();
2027         tgt = lmv_get_target(lmv, mds);
2028
2029         /*
2030          * Cancel UPDATE lock on child (fid1).
2031          */
2032         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2033         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2034                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2035         if (rc == 0)
2036                 rc = md_link(tgt->ltd_exp, op_data, request);
2037         if (rc == -ERESTART) {
2038                 LASSERT(*request != NULL);
2039                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2040                           "Got -ERESTART during link!\n");
2041                 ptlrpc_req_finished(*request);
2042                 *request = NULL;
2043
2044                 /*
2045                  * Directory got split. Time to update local object and repeat
2046                  * the request with proper MDS.
2047                  */
2048                 rc = lmv_handle_split(exp, &op_data->op_fid2);
2049                 if (rc == 0)
2050                         goto repeat;
2051         }
2052
2053         RETURN(rc);
2054 }
2055
2056 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2057                       const char *old, int oldlen, const char *new, int newlen,
2058                       struct ptlrpc_request **request)
2059 {
2060         struct obd_device       *obd = exp->exp_obd;
2061         struct lmv_obd          *lmv = &obd->u.lmv;
2062         struct lmv_tgt_desc     *src_tgt;
2063         struct lmv_tgt_desc     *tgt_tgt;
2064         int                      rc;
2065         int                      sidx;
2066         int                      loop = 0;
2067         struct lmv_object       *obj;
2068         mdsno_t                  mds1;
2069         mdsno_t                  mds2;
2070         ENTRY;
2071
2072         LASSERT(oldlen != 0);
2073
2074         CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2075                oldlen, old, PFID(&op_data->op_fid1),
2076                newlen, new, PFID(&op_data->op_fid2));
2077
2078         rc = lmv_check_connect(obd);
2079         if (rc)
2080                 RETURN(rc);
2081
2082 repeat:
2083         ++loop;
2084         LASSERT(loop <= 2);
2085         obj = lmv_object_find(obd, &op_data->op_fid1);
2086         if (obj) {
2087                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2088                                     (char *)old, oldlen);
2089                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2090                 mds1 = obj->lo_stripes[sidx].ls_mds;
2091                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2092                 lmv_object_put(obj);
2093         } else {
2094                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2095                 if (rc)
2096                         RETURN(rc);
2097         }
2098
2099         obj = lmv_object_find(obd, &op_data->op_fid2);
2100         if (obj) {
2101                 /*
2102                  * Directory is already split, so we have to forward request to
2103                  * the right MDS.
2104                  */
2105                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2106                                        (char *)new, newlen);
2107
2108                 mds2 = obj->lo_stripes[sidx].ls_mds;
2109                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2110                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2111                 lmv_object_put(obj);
2112         } else {
2113                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2114                 if (rc)
2115                         RETURN(rc);
2116         }
2117
2118         op_data->op_fsuid = cfs_curproc_fsuid();
2119         op_data->op_fsgid = cfs_curproc_fsgid();
2120         op_data->op_cap = cfs_curproc_cap_pack();
2121
2122         src_tgt = lmv_get_target(lmv, mds1);
2123         tgt_tgt = lmv_get_target(lmv, mds2);
2124
2125         /*
2126          * LOOKUP lock on src child (fid3) should also be cancelled for
2127          * src_tgt in mdc_rename.
2128          */
2129         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2130
2131         /*
2132          * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2133          * own target.
2134          */
2135         rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2136                               LCK_EX, MDS_INODELOCK_UPDATE,
2137                               MF_MDC_CANCEL_FID2);
2138
2139         /*
2140          * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2141          */
2142         if (rc == 0) {
2143                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2144                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2145                                       MF_MDC_CANCEL_FID4);
2146         }
2147
2148         /*
2149          * Cancel all the locks on tgt child (fid4).
2150          */
2151         if (rc == 0)
2152                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2153                                       LCK_EX, MDS_INODELOCK_FULL,
2154                                       MF_MDC_CANCEL_FID4);
2155
2156         if (rc == 0)
2157                 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2158                                new, newlen, request);
2159
2160         if (rc == -ERESTART) {
2161                 LASSERT(*request != NULL);
2162                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2163                           "Got -ERESTART during rename!\n");
2164                 ptlrpc_req_finished(*request);
2165                 *request = NULL;
2166
2167                 /*
2168                  * Directory got split. Time to update local object and repeat
2169                  * the request with proper MDS.
2170                  */
2171                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2172                 if (rc == 0)
2173                         goto repeat;
2174         }
2175         RETURN(rc);
2176 }
2177
2178 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2179                        void *ea, int ealen, void *ea2, int ea2len,
2180                        struct ptlrpc_request **request,
2181                        struct md_open_data **mod)
2182 {
2183         struct obd_device       *obd = exp->exp_obd;
2184         struct lmv_obd          *lmv = &obd->u.lmv;
2185         struct ptlrpc_request   *req;
2186         struct lmv_tgt_desc     *tgt;
2187         struct lmv_object       *obj;
2188         int                      rc = 0;
2189         int                      i;
2190         ENTRY;
2191
2192         rc = lmv_check_connect(obd);
2193         if (rc)
2194                 RETURN(rc);
2195
2196         obj = lmv_object_find(obd, &op_data->op_fid1);
2197
2198         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2199                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2200                obj ? ", split" : "");
2201
2202         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2203         if (obj) {
2204                 for (i = 0; i < obj->lo_objcount; i++) {
2205                         op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2206
2207                         tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2208                         if (IS_ERR(tgt)) {
2209                                 rc = PTR_ERR(tgt);
2210                                 break;
2211                         }
2212
2213                         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2214                                         ea2, ea2len, &req, mod);
2215
2216                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2217                                 /*
2218                                  * This is master object and this request should
2219                                  * be returned back to llite.
2220                                  */
2221                                 *request = req;
2222                         } else {
2223                                 ptlrpc_req_finished(req);
2224                         }
2225
2226                         if (rc)
2227                                 break;
2228                 }
2229                 lmv_object_put(obj);
2230         } else {
2231                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2232                 if (IS_ERR(tgt))
2233                         RETURN(PTR_ERR(tgt));
2234
2235                 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2236                                 ea2len, request, mod);
2237         }
2238         RETURN(rc);
2239 }
2240
2241 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2242                     struct obd_capa *oc, struct ptlrpc_request **request)
2243 {
2244         struct obd_device         *obd = exp->exp_obd;
2245         struct lmv_obd            *lmv = &obd->u.lmv;
2246         struct lmv_tgt_desc       *tgt;
2247         int                        rc;
2248         ENTRY;
2249
2250         rc = lmv_check_connect(obd);
2251         if (rc)
2252                 RETURN(rc);
2253
2254         tgt = lmv_find_target(lmv, fid);
2255         if (IS_ERR(tgt))
2256                 RETURN(PTR_ERR(tgt));
2257
2258         rc = md_sync(tgt->ltd_exp, fid, oc, request);
2259         RETURN(rc);
2260 }
2261
2262 /**
2263  * Main purpose of LMV blocking ast is to remove split directory LMV
2264  * presentation object (struct lmv_object) attached to the lock being revoked.
2265  */
2266 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2267                      void *data, int flag)
2268 {
2269         struct lustre_handle    lockh;
2270         struct lmv_object      *obj;
2271         int                     rc;
2272         ENTRY;
2273
2274         switch (flag) {
2275         case LDLM_CB_BLOCKING:
2276                 ldlm_lock2handle(lock, &lockh);
2277                 rc = ldlm_cli_cancel(&lockh);
2278                 if (rc < 0) {
2279                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2280                         RETURN(rc);
2281                 }
2282                 break;
2283         case LDLM_CB_CANCELING:
2284                 /*
2285                  * Time to drop cached attrs for split directory object
2286                  */
2287                 obj = lock->l_ast_data;
2288                 if (obj) {
2289                         CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2290                                ", master "DFID"\n",
2291                                lock->l_resource->lr_name.name[3] == 1 ?
2292                                "LOOKUP" : "UPDATE",
2293                                lock->l_resource->lr_name.name[0],
2294                                lock->l_resource->lr_name.name[1],
2295                                PFID(&obj->lo_fid));
2296                         lmv_object_put(obj);
2297                 }
2298                 break;
2299         default:
2300                 LBUG();
2301         }
2302         RETURN(0);
2303 }
2304
2305 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2306 {
2307         __u64         val;
2308
2309         val = le64_to_cpu(*hash);
2310         if (val < hash_adj)
2311                 val += MAX_HASH_SIZE;
2312         if (val != DIR_END_OFF)
2313                 *hash = cpu_to_le64(val - hash_adj);
2314 }
2315
2316 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2317 {
2318         __u64              id;
2319         struct obd_import *imp;
2320
2321         /*
2322          * XXX: to get nid we assume that underlying obd device is mdc.
2323          */
2324         imp  = class_exp2cliimp(exp);
2325         id   = imp->imp_connection->c_self + fid_flatten(fid);
2326
2327         CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2328                imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2329
2330         return id ^ (id >> 32);
2331 }
2332
2333 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2334                         struct obd_capa *oc, __u64 offset64, struct page *page,
2335                         struct ptlrpc_request **request)
2336 {
2337         struct obd_device       *obd = exp->exp_obd;
2338         struct lmv_obd          *lmv = &obd->u.lmv;
2339         struct lu_fid            rid = *fid;
2340         struct lmv_object       *obj;
2341         __u64                    offset;
2342         __u64                    hash_adj = 0;
2343         __u32                    rank = 0;
2344         __u64                    seg_size = 0;
2345         __u64                    tgt_tmp = 0;
2346         int                      tgt_idx = 0;
2347         int                      tgt0_idx = 0;
2348         int                      rc;
2349         int                      nr = 0;
2350         struct lmv_stripe       *los;
2351         struct lmv_tgt_desc     *tgt;
2352         struct lu_dirpage       *dp;
2353         struct lu_dirent        *ent;
2354         ENTRY;
2355
2356         offset = offset64;
2357
2358         rc = lmv_check_connect(obd);
2359         if (rc)
2360                 RETURN(rc);
2361
2362         CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2363
2364         /*
2365          * This case handle directory lookup in clustered metadata case (i.e.
2366          * split directory is located on multiple md servers.)
2367          * each server keeps directory entries for certain range of hashes.
2368          * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2369          * first server will keep records with hashes [ 0 ... MAX_HASH / N  - 1],
2370          * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2371          * so on....
2372          *      readdir can simply start reading entries from 0 - N server in
2373          * order but that will not scale well as all client will request dir in
2374          * to server in same order.
2375          * Following algorithm does optimization:
2376          * Instead of doing readdir in 1, 2, ...., N order, client with a
2377          * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2378          * (every client has rank R)
2379          *      But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2380          * since client ask dir from MDS{R} client has pages with offsets
2381          * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2382          * on hash  values that we get.
2383          */
2384         obj = lmv_object_find_lock(obd, fid);
2385         if (obj) {
2386                 nr       = obj->lo_objcount;
2387                 LASSERT(nr > 0);
2388                 seg_size = MAX_HASH_SIZE;
2389                 do_div(seg_size, nr);
2390                 los      = obj->lo_stripes;
2391                 tgt      = lmv_get_target(lmv, los[0].ls_mds);
2392                 rank     = lmv_node_rank(tgt->ltd_exp, fid) % nr;
2393                 tgt_tmp  = offset;
2394                 do_div(tgt_tmp, seg_size);
2395                 tgt0_idx = do_div(tgt_tmp,  nr);
2396                 tgt_idx  = (tgt0_idx + rank) % nr;
2397
2398                 if (tgt_idx < tgt0_idx)
2399                         /*
2400                          * Wrap around.
2401                          *
2402                          * Last segment has unusual length due to division
2403                          * rounding.
2404                          */
2405                         hash_adj = MAX_HASH_SIZE - seg_size * nr;
2406                 else
2407                         hash_adj = 0;
2408
2409                 hash_adj += rank * seg_size;
2410
2411                 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2412                        LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2413                        offset, tgt0_idx, offset + hash_adj, tgt_idx);
2414
2415                 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2416                 rid = obj->lo_stripes[tgt_idx].ls_fid;
2417                 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2418
2419                 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2420                        PFID(&rid), (unsigned long)offset, tgt_idx);
2421         } else
2422                 tgt = lmv_find_target(lmv, &rid);
2423
2424         if (IS_ERR(tgt))
2425                 GOTO(cleanup, rc = PTR_ERR(tgt));
2426
2427         rc = md_readpage(tgt->ltd_exp, &rid, oc, offset, page, request);
2428         if (rc)
2429                 GOTO(cleanup, rc);
2430         if (obj) {
2431                 dp = cfs_kmap(page);
2432
2433                 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2434                 lmv_hash_adjust(&dp->ldp_hash_end,   hash_adj);
2435                 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2436
2437                 for (ent = lu_dirent_start(dp); ent != NULL;
2438                      ent = lu_dirent_next(ent))
2439                         lmv_hash_adjust(&ent->lde_hash, hash_adj);
2440
2441                 if (tgt0_idx != nr - 1) {
2442                         __u64 end;
2443
2444                         end = le64_to_cpu(dp->ldp_hash_end);
2445                         if (end == DIR_END_OFF) {
2446                                 dp->ldp_hash_end = cpu_to_le32(seg_size *
2447                                                                (tgt0_idx + 1));
2448                                 CDEBUG(D_INODE,
2449                                        ""DFID" reset end "LPX64" tgt %d\n",
2450                                        PFID(&rid),
2451                                        (__u64)le64_to_cpu(dp->ldp_hash_end), tgt_idx);
2452                         }
2453                 }
2454                 cfs_kunmap(page);
2455         }
2456         EXIT;
2457 cleanup:
2458         if (obj)
2459                 lmv_object_put_unlock(obj);
2460         return rc;
2461 }
2462
2463 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2464                       struct ptlrpc_request **request)
2465 {
2466         struct obd_device       *obd = exp->exp_obd;
2467         struct lmv_obd          *lmv = &obd->u.lmv;
2468         struct lmv_tgt_desc     *tgt = NULL;
2469         struct lmv_object       *obj;
2470         int                      rc;
2471         int                      sidx;
2472         int                      loop = 0;
2473         ENTRY;
2474
2475         rc = lmv_check_connect(obd);
2476         if (rc)
2477                 RETURN(rc);
2478
2479 repeat:
2480         ++loop;
2481         LASSERT(loop <= 2);
2482         LASSERT(op_data->op_namelen != 0);
2483
2484         obj = lmv_object_find(obd, &op_data->op_fid1);
2485         if (obj) {
2486                 sidx = raw_name2idx(obj->lo_hashtype,
2487                                     obj->lo_objcount,
2488                                     op_data->op_name,
2489                                     op_data->op_namelen);
2490                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2491                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2492                 tgt = lmv_get_target(lmv,
2493                                      obj->lo_stripes[sidx].ls_mds);
2494                 lmv_object_put(obj);
2495                 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2496                        op_data->op_namelen, op_data->op_name,
2497                        PFID(&op_data->op_fid1), sidx);
2498         }
2499
2500         if (tgt == NULL) {
2501                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2502                 if (IS_ERR(tgt))
2503                         RETURN(PTR_ERR(tgt));
2504                 op_data->op_bias |= MDS_CHECK_SPLIT;
2505         }
2506
2507         op_data->op_fsuid = cfs_curproc_fsuid();
2508         op_data->op_fsgid = cfs_curproc_fsgid();
2509         op_data->op_cap = cfs_curproc_cap_pack();
2510
2511         /*
2512          * If child's fid is given, cancel unused locks for it if it is from
2513          * another export than parent.
2514          *
2515          * LOOKUP lock for child (fid3) should also be cancelled on parent
2516          * tgt_tgt in mdc_unlink().
2517          */
2518         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2519
2520         /*
2521          * Cancel FULL locks on child (fid3).
2522          */
2523         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2524                               MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2525
2526         if (rc == 0)
2527                 rc = md_unlink(tgt->ltd_exp, op_data, request);
2528
2529         if (rc == -ERESTART) {
2530                 LASSERT(*request != NULL);
2531                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2532                           "Got -ERESTART during unlink!\n");
2533                 ptlrpc_req_finished(*request);
2534                 *request = NULL;
2535
2536                 /*
2537                  * Directory got split. Time to update local object and repeat
2538                  * the request with proper MDS.
2539                  */
2540                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2541                 if (rc == 0)
2542                         goto repeat;
2543         }
2544         RETURN(rc);
2545 }
2546
2547 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2548 {
2549         int        rc = 0;
2550
2551         switch (stage) {
2552         case OBD_CLEANUP_EARLY:
2553                 /* XXX: here should be calling obd_precleanup() down to
2554                  * stack. */
2555                 break;
2556         case OBD_CLEANUP_EXPORTS:
2557                 rc = obd_llog_finish(obd, 0);
2558                 if (rc != 0)
2559                         CERROR("failed to cleanup llogging subsystems\n");
2560                 break;
2561         default:
2562                 break;
2563         }
2564         RETURN(rc);
2565 }
2566
2567 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2568                         void *key, __u32 *vallen, void *val,
2569                         struct lov_stripe_md *lsm)
2570 {
2571         struct obd_device       *obd;
2572         struct lmv_obd          *lmv;
2573         int                      rc = 0;
2574         ENTRY;
2575
2576         obd = class_exp2obd(exp);
2577         if (obd == NULL) {
2578                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2579                        exp->exp_handle.h_cookie);
2580                 RETURN(-EINVAL);
2581         }
2582
2583         lmv = &obd->u.lmv;
2584         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2585                 struct lmv_tgt_desc *tgts;
2586                 int i;
2587
2588                 rc = lmv_check_connect(obd);
2589                 if (rc)
2590                         RETURN(rc);
2591
2592                 LASSERT(*vallen == sizeof(__u32));
2593                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2594                      i++, tgts++) {
2595
2596                         /*
2597                          * All tgts should be connected when this gets called.
2598                          */
2599                         if (!tgts || !tgts->ltd_exp) {
2600                                 CERROR("target not setup?\n");
2601                                 continue;
2602                         }
2603
2604                         if (!obd_get_info(tgts->ltd_exp, keylen, key,
2605                                           vallen, val, NULL))
2606                                 RETURN(0);
2607                 }
2608                 RETURN(-EINVAL);
2609         } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2610                 rc = lmv_check_connect(obd);
2611                 if (rc)
2612                         RETURN(rc);
2613
2614                 /*
2615                  * Forwarding this request to first MDS, it should know LOV
2616                  * desc.
2617                  */
2618                 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2619                                   vallen, val, NULL);
2620                 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2621                         exp->exp_connect_flags =
2622                         ((struct obd_connect_data *)val)->ocd_connect_flags;
2623                 }
2624                 RETURN(rc);
2625         }
2626
2627         CDEBUG(D_IOCTL, "Invalid key\n");
2628         RETURN(-EINVAL);
2629 }
2630
2631 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2632                        void *key, obd_count vallen, void *val,
2633                        struct ptlrpc_request_set *set)
2634 {
2635         struct lmv_tgt_desc    *tgt;
2636         struct obd_device      *obd;
2637         struct lmv_obd         *lmv;
2638         int rc = 0;
2639         ENTRY;
2640
2641         obd = class_exp2obd(exp);
2642         if (obd == NULL) {
2643                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2644                        exp->exp_handle.h_cookie);
2645                 RETURN(-EINVAL);
2646         }
2647         lmv = &obd->u.lmv;
2648
2649         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2650             KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2651                 int i, err = 0;
2652
2653                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2654                         tgt = &lmv->tgts[i];
2655
2656                         if (!tgt->ltd_exp)
2657                                 continue;
2658
2659                         err = obd_set_info_async(tgt->ltd_exp,
2660                                                  keylen, key, vallen, val, set);
2661                         if (err && rc == 0)
2662                                 rc = err;
2663                 }
2664
2665                 RETURN(rc);
2666         }
2667
2668         RETURN(-EINVAL);
2669 }
2670
2671 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2672                struct lov_stripe_md *lsm)
2673 {
2674         struct obd_device         *obd = class_exp2obd(exp);
2675         struct lmv_obd            *lmv = &obd->u.lmv;
2676         struct lmv_stripe_md      *meap;
2677         struct lmv_stripe_md      *lsmp;
2678         int                        mea_size;
2679         int                        i;
2680         ENTRY;
2681
2682         mea_size = lmv_get_easize(lmv);
2683         if (!lmmp)
2684                 RETURN(mea_size);
2685
2686         if (*lmmp && !lsm) {
2687                 OBD_FREE(*lmmp, mea_size);
2688                 *lmmp = NULL;
2689                 RETURN(0);
2690         }
2691
2692         if (*lmmp == NULL) {
2693                 OBD_ALLOC(*lmmp, mea_size);
2694                 if (*lmmp == NULL)
2695                         RETURN(-ENOMEM);
2696         }
2697
2698         if (!lsm)
2699                 RETURN(mea_size);
2700
2701         lsmp = (struct lmv_stripe_md *)lsm;
2702         meap = (struct lmv_stripe_md *)*lmmp;
2703
2704         if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2705             lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2706                 RETURN(-EINVAL);
2707
2708         meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2709         meap->mea_count = cpu_to_le32(lsmp->mea_count);
2710         meap->mea_master = cpu_to_le32(lsmp->mea_master);
2711
2712         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2713                 meap->mea_ids[i] = meap->mea_ids[i];
2714                 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2715         }
2716
2717         RETURN(mea_size);
2718 }
2719
2720 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2721                  struct lov_mds_md *lmm, int lmm_size)
2722 {
2723         struct obd_device          *obd = class_exp2obd(exp);
2724         struct lmv_stripe_md      **tmea = (struct lmv_stripe_md **)lsmp;
2725         struct lmv_stripe_md       *mea = (struct lmv_stripe_md *)lmm;
2726         struct lmv_obd             *lmv = &obd->u.lmv;
2727         int                         mea_size;
2728         int                         i;
2729         __u32                       magic;
2730         ENTRY;
2731
2732         mea_size = lmv_get_easize(lmv);
2733         if (lsmp == NULL)
2734                 return mea_size;
2735
2736         if (*lsmp != NULL && lmm == NULL) {
2737                 OBD_FREE(*tmea, mea_size);
2738                 *lsmp = NULL;
2739                 RETURN(0);
2740         }
2741
2742         LASSERT(mea_size == lmm_size);
2743
2744         OBD_ALLOC(*tmea, mea_size);
2745         if (*tmea == NULL)
2746                 RETURN(-ENOMEM);
2747
2748         if (!lmm)
2749                 RETURN(mea_size);
2750
2751         if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2752             mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2753             mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2754         {
2755                 magic = le32_to_cpu(mea->mea_magic);
2756         } else {
2757                 /*
2758                  * Old mea is not handled here.
2759                  */
2760                 CERROR("Old not supportable EA is found\n");
2761                 LBUG();
2762         }
2763
2764         (*tmea)->mea_magic = magic;
2765         (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2766         (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2767
2768         for (i = 0; i < (*tmea)->mea_count; i++) {
2769                 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2770                 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2771         }
2772         RETURN(mea_size);
2773 }
2774
2775 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2776                              ldlm_policy_data_t *policy, ldlm_mode_t mode,
2777                              int flags, void *opaque)
2778 {
2779         struct obd_device       *obd = exp->exp_obd;
2780         struct lmv_obd          *lmv = &obd->u.lmv;
2781         int                      rc = 0;
2782         int                      err;
2783         int                      i;
2784         ENTRY;
2785
2786         LASSERT(fid != NULL);
2787
2788         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2789                 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2790                         continue;
2791
2792                 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2793                                        policy, mode, flags, opaque);
2794                 if (!rc)
2795                         rc = err;
2796         }
2797         RETURN(rc);
2798 }
2799
2800 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
2801                       __u32 *bits)
2802 {
2803         struct obd_device       *obd = exp->exp_obd;
2804         struct lmv_obd          *lmv = &obd->u.lmv;
2805         int                      rc;
2806         ENTRY;
2807
2808         rc =  md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data, bits);
2809         RETURN(rc);
2810 }
2811
2812 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2813                            const struct lu_fid *fid, ldlm_type_t type,
2814                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
2815                            struct lustre_handle *lockh)
2816 {
2817         struct obd_device       *obd = exp->exp_obd;
2818         struct lmv_obd          *lmv = &obd->u.lmv;
2819         ldlm_mode_t              rc;
2820         int                      i;
2821         ENTRY;
2822
2823         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2824
2825         /*
2826          * With CMD every object can have two locks in different namespaces:
2827          * lookup lock in space of mds storing direntry and update/open lock in
2828          * space of mds storing inode. Thus we check all targets, not only that
2829          * one fid was created in.
2830          */
2831         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2832                 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2833                                    type, policy, mode, lockh);
2834                 if (rc)
2835                         RETURN(rc);
2836         }
2837
2838         RETURN(0);
2839 }
2840
2841 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2842                       struct obd_export *dt_exp, struct obd_export *md_exp,
2843                       struct lustre_md *md)
2844 {
2845         struct obd_device       *obd = exp->exp_obd;
2846         struct lmv_obd          *lmv = &obd->u.lmv;
2847         int                      rc;
2848         ENTRY;
2849         rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2850         RETURN(rc);
2851 }
2852
2853 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2854 {
2855         struct obd_device       *obd = exp->exp_obd;
2856         struct lmv_obd          *lmv = &obd->u.lmv;
2857         ENTRY;
2858
2859         if (md->mea)
2860                 obd_free_memmd(exp, (void *)&md->mea);
2861         RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2862 }
2863
2864 int lmv_set_open_replay_data(struct obd_export *exp,
2865                              struct obd_client_handle *och,
2866                              struct ptlrpc_request *open_req)
2867 {
2868         struct obd_device       *obd = exp->exp_obd;
2869         struct lmv_obd          *lmv = &obd->u.lmv;
2870         struct lmv_tgt_desc     *tgt;
2871         ENTRY;
2872
2873         tgt = lmv_find_target(lmv, &och->och_fid);
2874         if (IS_ERR(tgt))
2875                 RETURN(PTR_ERR(tgt));
2876
2877         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2878 }
2879
2880 int lmv_clear_open_replay_data(struct obd_export *exp,
2881                                struct obd_client_handle *och)
2882 {
2883         struct obd_device       *obd = exp->exp_obd;
2884         struct lmv_obd          *lmv = &obd->u.lmv;
2885         struct lmv_tgt_desc     *tgt;
2886         ENTRY;
2887
2888         tgt = lmv_find_target(lmv, &och->och_fid);
2889         if (IS_ERR(tgt))
2890                 RETURN(PTR_ERR(tgt));
2891
2892         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2893 }
2894
2895 static int lmv_get_remote_perm(struct obd_export *exp,
2896                                const struct lu_fid *fid,
2897                                struct obd_capa *oc, __u32 suppgid,
2898                                struct ptlrpc_request **request)
2899 {
2900         struct obd_device       *obd = exp->exp_obd;
2901         struct lmv_obd          *lmv = &obd->u.lmv;
2902         struct lmv_tgt_desc     *tgt;
2903         int                      rc;
2904         ENTRY;
2905
2906         rc = lmv_check_connect(obd);
2907         if (rc)
2908                 RETURN(rc);
2909
2910         tgt = lmv_find_target(lmv, fid);
2911         if (IS_ERR(tgt))
2912                 RETURN(PTR_ERR(tgt));
2913
2914         rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
2915         RETURN(rc);
2916 }
2917
2918 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2919                           renew_capa_cb_t cb)
2920 {
2921         struct obd_device       *obd = exp->exp_obd;
2922         struct lmv_obd          *lmv = &obd->u.lmv;
2923         struct lmv_tgt_desc     *tgt;
2924         int                      rc;
2925         ENTRY;
2926
2927         rc = lmv_check_connect(obd);
2928         if (rc)
2929                 RETURN(rc);
2930
2931         tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
2932         if (IS_ERR(tgt))
2933                 RETURN(PTR_ERR(tgt));
2934
2935         rc = md_renew_capa(tgt->ltd_exp, oc, cb);
2936         RETURN(rc);
2937 }
2938
2939 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
2940                     const struct req_msg_field *field, struct obd_capa **oc)
2941 {
2942         struct obd_device *obd = exp->exp_obd;
2943         struct lmv_obd *lmv = &obd->u.lmv;
2944         int rc;
2945
2946         ENTRY;
2947         rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
2948         RETURN(rc);
2949 }
2950
2951 int lmv_intent_getattr_async(struct obd_export *exp,
2952                              struct md_enqueue_info *minfo,
2953                              struct ldlm_enqueue_info *einfo)
2954 {
2955         struct md_op_data       *op_data = &minfo->mi_data;
2956         struct obd_device       *obd = exp->exp_obd;
2957         struct lmv_obd          *lmv = &obd->u.lmv;
2958         struct lmv_object       *obj;
2959         struct lmv_tgt_desc     *tgt;
2960         int                      rc;
2961         int                      sidx;
2962         ENTRY;
2963
2964         rc = lmv_check_connect(obd);
2965         if (rc)
2966                 RETURN(rc);
2967
2968         if (!fid_is_sane(&op_data->op_fid2)) {
2969                 obj = lmv_object_find(obd, &op_data->op_fid1);
2970                 if (obj && op_data->op_namelen) {
2971                         sidx = raw_name2idx(obj->lo_hashtype,
2972                                             obj->lo_objcount,
2973                                             (char *)op_data->op_name,
2974                                             op_data->op_namelen);
2975                         op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2976                         tgt = lmv_get_target(lmv,
2977                                              obj->lo_stripes[sidx].ls_mds);
2978                         CDEBUG(D_INODE,
2979                                "Choose slave dir ("DFID") -> mds #%d\n",
2980                                PFID(&op_data->op_fid1), tgt->ltd_idx);
2981                 } else {
2982                         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2983                 }
2984                 if (obj)
2985                         lmv_object_put(obj);
2986         } else {
2987                 op_data->op_fid1 = op_data->op_fid2;
2988                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2989                 op_data->op_bias = MDS_CROSS_REF;
2990                 /*
2991                  * Unfortunately, we have to lie to MDC/MDS to retrieve
2992                  * attributes llite needs.
2993                 */
2994                 if (minfo->mi_it.it_op & IT_LOOKUP)
2995                         minfo->mi_it.it_op = IT_GETATTR;
2996         }
2997
2998         if (IS_ERR(tgt))
2999                 RETURN(PTR_ERR(tgt));
3000
3001         rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
3002         RETURN(rc);
3003 }
3004
3005 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3006                         struct lu_fid *fid, __u32 *bits)
3007 {
3008         struct obd_device       *obd = exp->exp_obd;
3009         struct lmv_obd          *lmv = &obd->u.lmv;
3010         struct lmv_tgt_desc     *tgt;
3011         int                      rc;
3012         ENTRY;
3013
3014         rc = lmv_check_connect(obd);
3015         if (rc)
3016                 RETURN(rc);
3017
3018         tgt = lmv_find_target(lmv, fid);
3019         if (IS_ERR(tgt))
3020                 RETURN(PTR_ERR(tgt));
3021
3022         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3023         RETURN(rc);
3024 }
3025
3026
3027 struct obd_ops lmv_obd_ops = {
3028         .o_owner                = THIS_MODULE,
3029         .o_setup                = lmv_setup,
3030         .o_cleanup              = lmv_cleanup,
3031         .o_precleanup           = lmv_precleanup,
3032         .o_process_config       = lmv_process_config,
3033         .o_connect              = lmv_connect,
3034         .o_disconnect           = lmv_disconnect,
3035         .o_statfs               = lmv_statfs,
3036         .o_get_info             = lmv_get_info,
3037         .o_set_info_async       = lmv_set_info_async,
3038         .o_packmd               = lmv_packmd,
3039         .o_unpackmd             = lmv_unpackmd,
3040         .o_notify               = lmv_notify,
3041         .o_get_uuid             = lmv_get_uuid,
3042         .o_iocontrol            = lmv_iocontrol,
3043         .o_fid_delete           = lmv_fid_delete
3044 };
3045
3046 struct md_ops lmv_md_ops = {
3047         .m_getstatus            = lmv_getstatus,
3048         .m_change_cbdata        = lmv_change_cbdata,
3049         .m_find_cbdata          = lmv_find_cbdata,
3050         .m_close                = lmv_close,
3051         .m_create               = lmv_create,
3052         .m_done_writing         = lmv_done_writing,
3053         .m_enqueue              = lmv_enqueue,
3054         .m_getattr              = lmv_getattr,
3055         .m_getxattr             = lmv_getxattr,
3056         .m_getattr_name         = lmv_getattr_name,
3057         .m_intent_lock          = lmv_intent_lock,
3058         .m_link                 = lmv_link,
3059         .m_rename               = lmv_rename,
3060         .m_setattr              = lmv_setattr,
3061         .m_setxattr             = lmv_setxattr,
3062         .m_sync                 = lmv_sync,
3063         .m_readpage             = lmv_readpage,
3064         .m_unlink               = lmv_unlink,
3065         .m_init_ea_size         = lmv_init_ea_size,
3066         .m_cancel_unused        = lmv_cancel_unused,
3067         .m_set_lock_data        = lmv_set_lock_data,
3068         .m_lock_match           = lmv_lock_match,
3069         .m_get_lustre_md        = lmv_get_lustre_md,
3070         .m_free_lustre_md       = lmv_free_lustre_md,
3071         .m_set_open_replay_data = lmv_set_open_replay_data,
3072         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3073         .m_renew_capa           = lmv_renew_capa,
3074         .m_unpack_capa          = lmv_unpack_capa,
3075         .m_get_remote_perm      = lmv_get_remote_perm,
3076         .m_intent_getattr_async = lmv_intent_getattr_async,
3077         .m_revalidate_lock      = lmv_revalidate_lock
3078 };
3079
3080 static quota_interface_t *quota_interface;
3081 extern quota_interface_t lmv_quota_interface;
3082
3083 int __init lmv_init(void)
3084 {
3085         struct lprocfs_static_vars lvars;
3086         int                        rc;
3087
3088         lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3089                                                 sizeof(struct lmv_object),
3090                                                 0, 0);
3091         if (!lmv_object_cache) {
3092                 CERROR("Error allocating lmv objects cache\n");
3093                 return -ENOMEM;
3094         }
3095
3096         lprocfs_lmv_init_vars(&lvars);
3097
3098         cfs_request_module("lquota");
3099         quota_interface = PORTAL_SYMBOL_GET(lmv_quota_interface);
3100         init_obd_quota_ops(quota_interface, &lmv_obd_ops);
3101
3102         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3103                                  lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3104         if (rc) {
3105                 if (quota_interface)
3106                         PORTAL_SYMBOL_PUT(lmv_quota_interface);
3107                 cfs_mem_cache_destroy(lmv_object_cache);
3108         }
3109
3110         return rc;
3111 }
3112
3113 #ifdef __KERNEL__
3114 static void lmv_exit(void)
3115 {
3116         if (quota_interface)
3117                 PORTAL_SYMBOL_PUT(lmv_quota_interface);
3118
3119         class_unregister_type(LUSTRE_LMV_NAME);
3120
3121         LASSERTF(cfs_atomic_read(&lmv_object_count) == 0,
3122                  "Can't free lmv objects cache, %d object(s) busy\n",
3123                  cfs_atomic_read(&lmv_object_count));
3124         cfs_mem_cache_destroy(lmv_object_cache);
3125 }
3126
3127 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3128 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3129 MODULE_LICENSE("GPL");
3130
3131 module_init(lmv_init);
3132 module_exit(lmv_exit);
3133 #endif