lustre/lmv/lmv_obd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  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_log.h>
  56 #include <obd_support.h>
  57 #include <lustre_lib.h>
  58 #include <lustre_net.h>
  59 #include <obd_class.h>
  60 #include <lprocfs_status.h>
  61 #include <lustre_lite.h>
  62 #include <lustre_fid.h>
  63 #include "lmv_internal.h"
  64
  65 /* object cache. */
  66 cfs_mem_cache_t *lmv_object_cache;
  67 cfs_atomic_t lmv_object_count = CFS_ATOMIC_INIT(0);
  68
  69 static void lmv_activate_target(struct lmv_obd *lmv,
  70                                 struct lmv_tgt_desc *tgt,
  71                                 int activate)
  72 {
  73         if (tgt->ltd_active == activate)
  74                 return;
  75
  76         tgt->ltd_active = activate;
  77         lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
  78 }
  79
  80 /**
  81  * Error codes:
  82  *
  83  *  -EINVAL  : UUID can't be found in the LMV's target list
  84  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  85  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  86  */
  87 static int lmv_set_mdc_active(struct lmv_obd *lmv, struct obd_uuid *uuid,
  88                               int activate)
  89 {
  90         struct lmv_tgt_desc    *tgt;
  91         struct obd_device      *obd;
  92         int                     i;
  93         int                     rc = 0;
  94         ENTRY;
  95
  96         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  97                lmv, uuid->uuid, activate);
  98
  99         cfs_spin_lock(&lmv->lmv_lock);
 100         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 101                 if (tgt->ltd_exp == NULL)
 102                         continue;
 103
 104                 CDEBUG(D_INFO, "Target idx %d is %s conn "LPX64"\n",
 105                        i, tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
 106
 107                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 108                         break;
 109         }
 110
 111         if (i == lmv->desc.ld_tgt_count)
 112                 GOTO(out_lmv_lock, rc = -EINVAL);
 113
 114         obd = class_exp2obd(tgt->ltd_exp);
 115         if (obd == NULL)
 116                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 117
 118         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 119                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 120                obd->obd_type->typ_name, i);
 121         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 122
 123         if (tgt->ltd_active == activate) {
 124                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 125                        activate ? "" : "in");
 126                 GOTO(out_lmv_lock, rc);
 127         }
 128
 129         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 130                activate ? "" : "in");
 131         lmv_activate_target(lmv, tgt, activate);
 132         EXIT;
 133
 134  out_lmv_lock:
 135         cfs_spin_unlock(&lmv->lmv_lock);
 136         return rc;
 137 }
 138
 139 static int lmv_set_mdc_data(struct lmv_obd *lmv, struct obd_uuid *uuid,
 140                             struct obd_connect_data *data)
 141 {
 142         struct lmv_tgt_desc    *tgt;
 143         int                     i;
 144         ENTRY;
 145
 146         LASSERT(data != NULL);
 147
 148         cfs_spin_lock(&lmv->lmv_lock);
 149         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 150                 if (tgt->ltd_exp == NULL)
 151                         continue;
 152
 153                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
 154                         lmv->datas[tgt->ltd_idx] = *data;
 155                         break;
 156                 }
 157         }
 158         cfs_spin_unlock(&lmv->lmv_lock);
 159         RETURN(0);
 160 }
 161
 162 struct obd_uuid *lmv_get_uuid(struct obd_export *exp) {
 163         struct obd_device *obd = exp->exp_obd;
 164         struct lmv_obd *lmv = &obd->u.lmv;
 165         return obd_get_uuid(lmv->tgts[0].ltd_exp);
 166 }
 167
 168 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 169                       enum obd_notify_event ev, void *data)
 170 {
 171         struct obd_connect_data *conn_data;
 172         struct lmv_obd          *lmv = &obd->u.lmv;
 173         struct obd_uuid         *uuid;
 174         int                      rc = 0;
 175         ENTRY;
 176
 177         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 178                 CERROR("unexpected notification of %s %s!\n",
 179                        watched->obd_type->typ_name,
 180                        watched->obd_name);
 181                 RETURN(-EINVAL);
 182         }
 183
 184         uuid = &watched->u.cli.cl_target_uuid;
 185         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 186                 /*
 187                  * Set MDC as active before notifying the observer, so the
 188                  * observer can use the MDC normally.
 189                  */
 190                 rc = lmv_set_mdc_active(lmv, uuid,
 191                                         ev == OBD_NOTIFY_ACTIVE);
 192                 if (rc) {
 193                         CERROR("%sactivation of %s failed: %d\n",
 194                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 195                                uuid->uuid, rc);
 196                         RETURN(rc);
 197                 }
 198         } else if (ev == OBD_NOTIFY_OCD) {
 199                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 200
 201                 /*
 202                  * Set connect data to desired target, update exp_connect_flags.
 203                  */
 204                 rc = lmv_set_mdc_data(lmv, uuid, conn_data);
 205                 if (rc) {
 206                         CERROR("can't set connect data to target %s, rc %d\n",
 207                                uuid->uuid, rc);
 208                         RETURN(rc);
 209                 }
 210
 211                 /*
 212                  * XXX: Make sure that ocd_connect_flags from all targets are
 213                  * the same. Otherwise one of MDTs runs wrong version or
 214                  * something like this.  --umka
 215                  */
 216                 obd->obd_self_export->exp_connect_flags =
 217                         conn_data->ocd_connect_flags;
 218         }
 219 #if 0
 220         else if (ev == OBD_NOTIFY_DISCON) {
 221                 /*
 222                  * For disconnect event, flush fld cache for failout MDS case.
 223                  */
 224                 fld_client_flush(&lmv->lmv_fld);
 225         }
 226 #endif
 227         /*
 228          * Pass the notification up the chain.
 229          */
 230         if (obd->obd_observer)
 231                 rc = obd_notify(obd->obd_observer, watched, ev, data);
 232
 233         RETURN(rc);
 234 }
 235
 236 /**
 237  * This is fake connect function. Its purpose is to initialize lmv and say
 238  * caller that everything is okay. Real connection will be performed later.
 239  */
 240 static int lmv_connect(const struct lu_env *env,
 241                        struct obd_export **exp, struct obd_device *obd,
 242                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 243                        void *localdata)
 244 {
 245 #ifdef __KERNEL__
 246         struct proc_dir_entry *lmv_proc_dir;
 247 #endif
 248         struct lmv_obd        *lmv = &obd->u.lmv;
 249         struct lustre_handle  conn = { 0 };
 250         int                    rc = 0;
 251         ENTRY;
 252
 253         /*
 254          * We don't want to actually do the underlying connections more than
 255          * once, so keep track.
 256          */
 257         lmv->refcount++;
 258         if (lmv->refcount > 1) {
 259                 *exp = NULL;
 260                 RETURN(0);
 261         }
 262
 263         rc = class_connect(&conn, obd, cluuid);
 264         if (rc) {
 265                 CERROR("class_connection() returned %d\n", rc);
 266                 RETURN(rc);
 267         }
 268
 269         *exp = class_conn2export(&conn);
 270         class_export_get(*exp);
 271
 272         lmv->exp = *exp;
 273         lmv->connected = 0;
 274         lmv->cluuid = *cluuid;
 275
 276         if (data)
 277                 lmv->conn_data = *data;
 278
 279 #ifdef __KERNEL__
 280         lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
 281                                         NULL, NULL);
 282         if (IS_ERR(lmv_proc_dir)) {
 283                 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
 284                        obd->obd_type->typ_name, obd->obd_name);
 285                 lmv_proc_dir = NULL;
 286         }
 287 #endif
 288
 289         /*
 290          * All real clients should perform actual connection right away, because
 291          * it is possible, that LMV will not have opportunity to connect targets
 292          * and MDC stuff will be called directly, for instance while reading
 293          * ../mdc/../kbytesfree procfs file, etc.
 294          */
 295         if (data->ocd_connect_flags & OBD_CONNECT_REAL)
 296                 rc = lmv_check_connect(obd);
 297
 298 #ifdef __KERNEL__
 299         if (rc) {
 300                 if (lmv_proc_dir)
 301                         lprocfs_remove(&lmv_proc_dir);
 302         }
 303 #endif
 304
 305         RETURN(rc);
 306 }
 307
 308 static void lmv_set_timeouts(struct obd_device *obd)
 309 {
 310         struct lmv_tgt_desc   *tgts;
 311         struct lmv_obd        *lmv;
 312         int                    i;
 313
 314         lmv = &obd->u.lmv;
 315         if (lmv->server_timeout == 0)
 316                 return;
 317
 318         if (lmv->connected == 0)
 319                 return;
 320
 321         for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
 322                 if (tgts->ltd_exp == NULL)
 323                         continue;
 324
 325                 obd_set_info_async(tgts->ltd_exp, sizeof(KEY_INTERMDS),
 326                                    KEY_INTERMDS, 0, NULL, NULL);
 327         }
 328 }
 329
 330 static int lmv_init_ea_size(struct obd_export *exp, int easize,
 331                             int def_easize, int cookiesize)
 332 {
 333         struct obd_device   *obd = exp->exp_obd;
 334         struct lmv_obd      *lmv = &obd->u.lmv;
 335         int                  i;
 336         int                  rc = 0;
 337         int                  change = 0;
 338         ENTRY;
 339
 340         if (lmv->max_easize < easize) {
 341                 lmv->max_easize = easize;
 342                 change = 1;
 343         }
 344         if (lmv->max_def_easize < def_easize) {
 345                 lmv->max_def_easize = def_easize;
 346                 change = 1;
 347         }
 348         if (lmv->max_cookiesize < cookiesize) {
 349                 lmv->max_cookiesize = cookiesize;
 350                 change = 1;
 351         }
 352         if (change == 0)
 353                 RETURN(0);
 354
 355         if (lmv->connected == 0)
 356                 RETURN(0);
 357
 358         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 359                 if (lmv->tgts[i].ltd_exp == NULL) {
 360                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
 361                         continue;
 362                 }
 363
 364                 rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize,
 365                                      cookiesize);
 366                 if (rc) {
 367                         CERROR("obd_init_ea_size() failed on MDT target %d, "
 368                                "error %d.\n", i, rc);
 369                         break;
 370                 }
 371         }
 372         RETURN(rc);
 373 }
 374
 375 #define MAX_STRING_SIZE 128
 376
 377 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 378 {
 379 #ifdef __KERNEL__
 380         struct proc_dir_entry   *lmv_proc_dir;
 381 #endif
 382         struct lmv_obd          *lmv = &obd->u.lmv;
 383         struct obd_uuid         *cluuid = &lmv->cluuid;
 384         struct obd_connect_data *mdc_data = NULL;
 385         struct obd_uuid          lmv_mdc_uuid = { "LMV_MDC_UUID" };
 386         struct obd_device       *mdc_obd;
 387         struct obd_export       *mdc_exp;
 388         struct lu_fld_target     target;
 389         int                      rc;
 390         ENTRY;
 391
 392         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 393                                         &obd->obd_uuid);
 394         if (!mdc_obd) {
 395                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 396                 RETURN(-EINVAL);
 397         }
 398
 399         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
 400                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 401                 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
 402                 cluuid->uuid);
 403
 404         if (!mdc_obd->obd_set_up) {
 405                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 406                 RETURN(-EINVAL);
 407         }
 408
 409         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &lmv_mdc_uuid,
 410                          &lmv->conn_data, NULL);
 411         if (rc) {
 412                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 413                 RETURN(rc);
 414         }
 415
 416         /*
 417          * Init fid sequence client for this mdc and add new fld target.
 418          */
 419         rc = obd_fid_init(mdc_exp);
 420         if (rc)
 421                 RETURN(rc);
 422
 423         target.ft_srv = NULL;
 424         target.ft_exp = mdc_exp;
 425         target.ft_idx = tgt->ltd_idx;
 426
 427         fld_client_add_target(&lmv->lmv_fld, &target);
 428
 429         mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data;
 430
 431         rc = obd_register_observer(mdc_obd, obd);
 432         if (rc) {
 433                 obd_disconnect(mdc_exp);
 434                 CERROR("target %s register_observer error %d\n",
 435                        tgt->ltd_uuid.uuid, rc);
 436                 RETURN(rc);
 437         }
 438
 439         if (obd->obd_observer) {
 440                 /*
 441                  * Tell the observer about the new target.
 442                  */
 443                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 444                                 OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts));
 445                 if (rc) {
 446                         obd_disconnect(mdc_exp);
 447                         RETURN(rc);
 448                 }
 449         }
 450
 451         tgt->ltd_active = 1;
 452         tgt->ltd_exp = mdc_exp;
 453         lmv->desc.ld_active_tgt_count++;
 454
 455         /*
 456          * Copy connect data, it may be used later.
 457          */
 458         lmv->datas[tgt->ltd_idx] = *mdc_data;
 459
 460         md_init_ea_size(tgt->ltd_exp, lmv->max_easize,
 461                         lmv->max_def_easize, lmv->max_cookiesize);
 462
 463         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 464                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 465                 cfs_atomic_read(&obd->obd_refcount));
 466
 467 #ifdef __KERNEL__
 468         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 469         if (lmv_proc_dir) {
 470                 struct proc_dir_entry *mdc_symlink;
 471
 472                 LASSERT(mdc_obd->obd_type != NULL);
 473                 LASSERT(mdc_obd->obd_type->typ_name != NULL);
 474                 mdc_symlink = lprocfs_add_symlink(mdc_obd->obd_name,
 475                                                   lmv_proc_dir,
 476                                                   "../../../%s/%s",
 477                                                   mdc_obd->obd_type->typ_name,
 478                                                   mdc_obd->obd_name);
 479                 if (mdc_symlink == NULL) {
 480                         CERROR("Could not register LMV target "
 481                                "/proc/fs/lustre/%s/%s/target_obds/%s.",
 482                                obd->obd_type->typ_name, obd->obd_name,
 483                                mdc_obd->obd_name);
 484                         lprocfs_remove(&lmv_proc_dir);
 485                         lmv_proc_dir = NULL;
 486                 }
 487         }
 488 #endif
 489         RETURN(0);
 490 }
 491
 492 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
 493 {
 494         struct lmv_obd      *lmv = &obd->u.lmv;
 495         struct lmv_tgt_desc *tgt;
 496         int                  rc = 0;
 497         ENTRY;
 498
 499         CDEBUG(D_CONFIG, "Target uuid: %s.\n", tgt_uuid->uuid);
 500
 501         lmv_init_lock(lmv);
 502
 503         if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
 504                 lmv_init_unlock(lmv);
 505                 CERROR("Can't add %s, LMV module compiled for %d MDCs. "
 506                        "That many MDCs already configured.\n",
 507                        tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
 508                 RETURN(-EINVAL);
 509         }
 510         if (lmv->desc.ld_tgt_count == 0) {
 511                 struct obd_device *mdc_obd;
 512
 513                 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
 514                                                 &obd->obd_uuid);
 515                 if (!mdc_obd) {
 516                         lmv_init_unlock(lmv);
 517                         CERROR("Target %s not attached\n", tgt_uuid->uuid);
 518                         RETURN(-EINVAL);
 519                 }
 520
 521                 rc = obd_llog_init(obd, &obd->obd_olg, mdc_obd, NULL);
 522                 if (rc) {
 523                         lmv_init_unlock(lmv);
 524                         CERROR("lmv failed to setup llogging subsystems\n");
 525                 }
 526         }
 527         cfs_spin_lock(&lmv->lmv_lock);
 528         tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
 529         tgt->ltd_uuid = *tgt_uuid;
 530         cfs_spin_unlock(&lmv->lmv_lock);
 531
 532         if (lmv->connected) {
 533                 rc = lmv_connect_mdc(obd, tgt);
 534                 if (rc) {
 535                         cfs_spin_lock(&lmv->lmv_lock);
 536                         lmv->desc.ld_tgt_count--;
 537                         memset(tgt, 0, sizeof(*tgt));
 538                         cfs_spin_unlock(&lmv->lmv_lock);
 539                 } else {
 540                         int easize = sizeof(struct lmv_stripe_md) +
 541                                      lmv->desc.ld_tgt_count *
 542                                      sizeof(struct lu_fid);
 543                         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 544                 }
 545         }
 546
 547         lmv_init_unlock(lmv);
 548         RETURN(rc);
 549 }
 550
 551 int lmv_check_connect(struct obd_device *obd)
 552 {
 553         struct lmv_obd       *lmv = &obd->u.lmv;
 554         struct lmv_tgt_desc  *tgt;
 555         int                   i;
 556         int                   rc;
 557         int                   easize;
 558         ENTRY;
 559
 560         if (lmv->connected)
 561                 RETURN(0);
 562
 563         lmv_init_lock(lmv);
 564         if (lmv->connected) {
 565                 lmv_init_unlock(lmv);
 566                 RETURN(0);
 567         }
 568
 569         if (lmv->desc.ld_tgt_count == 0) {
 570                 lmv_init_unlock(lmv);
 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_force = obd->obd_force;
 631                 mdc_obd->obd_fail = obd->obd_fail;
 632                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 633         }
 634
 635 #ifdef __KERNEL__
 636         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 637         if (lmv_proc_dir) {
 638                 struct proc_dir_entry *mdc_symlink;
 639
 640                 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
 641                 if (mdc_symlink) {
 642                         lprocfs_remove(&mdc_symlink);
 643                 } else {
 644                         CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
 645                                obd->obd_type->typ_name, obd->obd_name,
 646                                mdc_obd->obd_name);
 647                 }
 648         }
 649 #endif
 650         rc = obd_fid_fini(tgt->ltd_exp);
 651         if (rc)
 652                 CERROR("Can't finanize fids factory\n");
 653
 654         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 655                tgt->ltd_exp->exp_obd->obd_name,
 656                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 657
 658         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 659         rc = obd_disconnect(tgt->ltd_exp);
 660         if (rc) {
 661                 if (tgt->ltd_active) {
 662                         CERROR("Target %s disconnect error %d\n",
 663                                tgt->ltd_uuid.uuid, rc);
 664                 }
 665         }
 666
 667         lmv_activate_target(lmv, tgt, 0);
 668         tgt->ltd_exp = NULL;
 669         RETURN(0);
 670 }
 671
 672 static int lmv_disconnect(struct obd_export *exp)
 673 {
 674         struct obd_device     *obd = class_exp2obd(exp);
 675 #ifdef __KERNEL__
 676         struct proc_dir_entry *lmv_proc_dir;
 677 #endif
 678         struct lmv_obd        *lmv = &obd->u.lmv;
 679         int                    rc;
 680         int                    i;
 681         ENTRY;
 682
 683         if (!lmv->tgts)
 684                 goto out_local;
 685
 686         /*
 687          * Only disconnect the underlying layers on the final disconnect.
 688          */
 689         lmv->refcount--;
 690         if (lmv->refcount != 0)
 691                 goto out_local;
 692
 693         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 694                 if (lmv->tgts[i].ltd_exp == NULL)
 695                         continue;
 696                 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
 697         }
 698
 699 #ifdef __KERNEL__
 700         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 701         if (lmv_proc_dir) {
 702                 lprocfs_remove(&lmv_proc_dir);
 703         } else {
 704                 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
 705                        obd->obd_type->typ_name, obd->obd_name);
 706         }
 707 #endif
 708
 709 out_local:
 710         /*
 711          * This is the case when no real connection is established by
 712          * lmv_check_connect().
 713          */
 714         if (!lmv->connected)
 715                 class_export_put(exp);
 716         rc = class_disconnect(exp);
 717         if (lmv->refcount == 0)
 718                 lmv->connected = 0;
 719         RETURN(rc);
 720 }
 721
 722 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 723                          int len, void *karg, void *uarg)
 724 {
 725         struct obd_device    *obddev = class_exp2obd(exp);
 726         struct lmv_obd       *lmv = &obddev->u.lmv;
 727         int                   i = 0;
 728         int                   rc = 0;
 729         int                   set = 0;
 730         int                   count = lmv->desc.ld_tgt_count;
 731         ENTRY;
 732
 733         if (count == 0)
 734                 RETURN(-ENOTTY);
 735
 736         switch (cmd) {
 737         case IOC_OBD_STATFS: {
 738                 struct obd_ioctl_data *data = karg;
 739                 struct obd_device *mdc_obd;
 740                 struct obd_statfs stat_buf = {0};
 741                 __u32 index;
 742
 743                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 744                 if ((index >= count))
 745                         RETURN(-ENODEV);
 746
 747                 if (!lmv->tgts[index].ltd_active)
 748                         RETURN(-ENODATA);
 749
 750                 mdc_obd = class_exp2obd(lmv->tgts[index].ltd_exp);
 751                 if (!mdc_obd)
 752                         RETURN(-EINVAL);
 753
 754                 /* copy UUID */
 755                 if (cfs_copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 756                                      min((int) data->ioc_plen2,
 757                                          (int) sizeof(struct obd_uuid))))
 758                         RETURN(-EFAULT);
 759
 760                 rc = obd_statfs(mdc_obd, &stat_buf,
 761                                 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
 762                                 0);
 763                 if (rc)
 764                         RETURN(rc);
 765                 if (cfs_copy_to_user(data->ioc_pbuf1, &stat_buf,
 766                                      min((int) data->ioc_plen1,
 767                                          (int) sizeof(stat_buf))))
 768                         RETURN(-EFAULT);
 769                 break;
 770         }
 771         case OBD_IOC_QUOTACTL: {
 772                 struct if_quotactl *qctl = karg;
 773                 struct lmv_tgt_desc *tgt = NULL;
 774                 struct obd_quotactl *oqctl;
 775
 776                 if (qctl->qc_valid == QC_MDTIDX) {
 777                         if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
 778                                 RETURN(-EINVAL);
 779
 780                         tgt = &lmv->tgts[qctl->qc_idx];
 781                         if (!tgt->ltd_exp)
 782                                 RETURN(-EINVAL);
 783                 } else if (qctl->qc_valid == QC_UUID) {
 784                         for (i = 0; i < count; i++) {
 785                                 tgt = &lmv->tgts[i];
 786                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 787                                                      &qctl->obd_uuid))
 788                                         continue;
 789
 790                                 if (tgt->ltd_exp == NULL)
 791                                         RETURN(-EINVAL);
 792
 793                                 break;
 794                         }
 795                 } else {
 796                         RETURN(-EINVAL);
 797                 }
 798
 799                 if (i >= count)
 800                         RETURN(-EAGAIN);
 801
 802                 LASSERT(tgt && tgt->ltd_exp);
 803                 OBD_ALLOC_PTR(oqctl);
 804                 if (!oqctl)
 805                         RETURN(-ENOMEM);
 806
 807                 QCTL_COPY(oqctl, qctl);
 808                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 809                 if (rc == 0) {
 810                         QCTL_COPY(qctl, oqctl);
 811                         qctl->qc_valid = QC_MDTIDX;
 812                         qctl->obd_uuid = tgt->ltd_uuid;
 813                 }
 814                 OBD_FREE_PTR(oqctl);
 815                 break;
 816         }
 817         case OBD_IOC_CHANGELOG_SEND:
 818         case OBD_IOC_CHANGELOG_CLEAR: {
 819                 struct ioc_changelog *icc = karg;
 820
 821                 if (icc->icc_mdtindex >= count)
 822                         RETURN(-ENODEV);
 823
 824                 rc = obd_iocontrol(cmd, lmv->tgts[icc->icc_mdtindex].ltd_exp,
 825                                    sizeof(*icc), icc, NULL);
 826                 break;
 827         }
 828         case LL_IOC_GET_CONNECT_FLAGS: {
 829                 rc = obd_iocontrol(cmd, lmv->tgts[0].ltd_exp, len, karg, uarg);
 830                 break;
 831         }
 832
 833         default : {
 834                 for (i = 0; i < count; i++) {
 835                         int err;
 836                         struct obd_device *mdc_obd;
 837
 838                         if (lmv->tgts[i].ltd_exp == NULL)
 839                                 continue;
 840                         /* ll_umount_begin() sets force flag but for lmv, not
 841                          * mdc. Let's pass it through */
 842                         mdc_obd = class_exp2obd(lmv->tgts[i].ltd_exp);
 843                         mdc_obd->obd_force = obddev->obd_force;
 844                         err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
 845                                             karg, uarg);
 846                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
 847                                 RETURN(err);
 848                         } else if (err) {
 849                                 if (lmv->tgts[i].ltd_active) {
 850                                         CERROR("error: iocontrol MDC %s on MDT"
 851                                                "idx %d cmd %x: err = %d\n",
 852                                                 lmv->tgts[i].ltd_uuid.uuid,
 853                                                 i, cmd, err);
 854                                         if (!rc)
 855                                                 rc = err;
 856                                 }
 857                         } else
 858                                 set = 1;
 859                 }
 860                 if (!set && !rc)
 861                         rc = -EIO;
 862         }
 863         }
 864         RETURN(rc);
 865 }
 866
 867 static int lmv_all_chars_policy(int count, const char *name,
 868                                 int len)
 869 {
 870         unsigned int c = 0;
 871
 872         while (len > 0)
 873                 c += name[--len];
 874         c = c % count;
 875         return c;
 876 }
 877
 878 static int lmv_nid_policy(struct lmv_obd *lmv)
 879 {
 880         struct obd_import *imp;
 881         __u32              id;
 882
 883         /*
 884          * XXX: To get nid we assume that underlying obd device is mdc.
 885          */
 886         imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
 887         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
 888         return id % lmv->desc.ld_tgt_count;
 889 }
 890
 891 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
 892                           placement_policy_t placement)
 893 {
 894         switch (placement) {
 895         case PLACEMENT_CHAR_POLICY:
 896                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
 897                                             op_data->op_name,
 898                                             op_data->op_namelen);
 899         case PLACEMENT_NID_POLICY:
 900                 return lmv_nid_policy(lmv);
 901
 902         default:
 903                 break;
 904         }
 905
 906         CERROR("Unsupported placement policy %x\n", placement);
 907         return -EINVAL;
 908 }
 909
 910 /**
 911  * This is _inode_ placement policy function (not name).
 912  */
 913 static int lmv_placement_policy(struct obd_device *obd,
 914                                 struct md_op_data *op_data,
 915                                 mdsno_t *mds)
 916 {
 917         struct lmv_obd          *lmv = &obd->u.lmv;
 918         struct lmv_object       *obj;
 919         int                      rc;
 920         ENTRY;
 921
 922         LASSERT(mds != NULL);
 923
 924         if (lmv->desc.ld_tgt_count == 1) {
 925                 *mds = 0;
 926                 RETURN(0);
 927         }
 928
 929         /*
 930          * Allocate new fid on target according to operation type and parent
 931          * home mds.
 932          */
 933         obj = lmv_object_find(obd, &op_data->op_fid1);
 934         if (obj != NULL || op_data->op_name == NULL ||
 935             op_data->op_opc != LUSTRE_OPC_MKDIR) {
 936                 /*
 937                  * Allocate fid for non-dir or for null name or for case parent
 938                  * dir is split.
 939                  */
 940                 if (obj) {
 941                         lmv_object_put(obj);
 942
 943                         /*
 944                          * If we have this flag turned on, and we see that
 945                          * parent dir is split, this means, that caller did not
 946                          * notice split yet. This is race and we would like to
 947                          * let caller know that.
 948                          */
 949                         if (op_data->op_bias & MDS_CHECK_SPLIT)
 950                                 RETURN(-ERESTART);
 951                 }
 952
 953                 /*
 954                  * Allocate new fid on same mds where parent fid is located and
 955                  * where operation will be sent. In case of split dir, ->op_fid1
 956                  * and ->op_mds here will contain fid and mds of slave directory
 957                  * object (assigned by caller).
 958                  */
 959                 *mds = op_data->op_mds;
 960                 rc = 0;
 961         } else {
 962                 /*
 963                  * Parent directory is not split and we want to create a
 964                  * directory in it. Let's calculate where to place it according
 965                  * to operation data @op_data.
 966                  */
 967                 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
 968                 rc = 0;
 969         }
 970
 971         if (rc) {
 972                 CERROR("Can't choose MDS, err = %d\n", rc);
 973         } else {
 974                 LASSERT(*mds < lmv->desc.ld_tgt_count);
 975         }
 976
 977         RETURN(rc);
 978 }
 979
 980 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
 981                     mdsno_t mds)
 982 {
 983         struct lmv_tgt_desc *tgt;
 984         int                  rc;
 985         ENTRY;
 986
 987         tgt = lmv_get_target(lmv, mds);
 988
 989         /*
 990          * New seq alloc and FLD setup should be atomic. Otherwise we may find
 991          * on server that seq in new allocated fid is not yet known.
 992          */
 993         cfs_mutex_lock(&tgt->ltd_fid_mutex);
 994
 995         if (!tgt->ltd_active)
 996                 GOTO(out, rc = -ENODEV);
 997
 998         /*
 999          * Asking underlaying tgt layer to allocate new fid.
1000          */
1001         rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
1002         if (rc > 0) {
1003                 LASSERT(fid_is_sane(fid));
1004                 rc = 0;
1005         }
1006
1007         EXIT;
1008 out:
1009         cfs_mutex_unlock(&tgt->ltd_fid_mutex);
1010         return rc;
1011 }
1012
1013 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
1014                   struct md_op_data *op_data)
1015 {
1016         struct obd_device     *obd = class_exp2obd(exp);
1017         struct lmv_obd        *lmv = &obd->u.lmv;
1018         mdsno_t                mds = 0;
1019         int                    rc;
1020         ENTRY;
1021
1022         LASSERT(op_data != NULL);
1023         LASSERT(fid != NULL);
1024
1025         rc = lmv_placement_policy(obd, op_data, &mds);
1026         if (rc) {
1027                 CERROR("Can't get target for allocating fid, "
1028                        "rc %d\n", rc);
1029                 RETURN(rc);
1030         }
1031
1032         rc = __lmv_fid_alloc(lmv, fid, mds);
1033         if (rc) {
1034                 CERROR("Can't alloc new fid, rc %d\n", rc);
1035                 RETURN(rc);
1036         }
1037
1038         RETURN(rc);
1039 }
1040
1041 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1042 {
1043         ENTRY;
1044         LASSERT(exp != NULL && fid != NULL);
1045         if (lmv_object_delete(exp, fid)) {
1046                 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1047                        PFID(fid));
1048         }
1049         RETURN(0);
1050 }
1051
1052 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1053 {
1054         struct lmv_obd             *lmv = &obd->u.lmv;
1055         struct lprocfs_static_vars  lvars;
1056         struct lmv_desc            *desc;
1057         int                         rc;
1058         int                         i = 0;
1059         ENTRY;
1060
1061         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1062                 CERROR("LMV setup requires a descriptor\n");
1063                 RETURN(-EINVAL);
1064         }
1065
1066         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1067         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1068                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1069                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1070                 RETURN(-EINVAL);
1071         }
1072
1073         lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1074
1075         OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1076         if (lmv->tgts == NULL)
1077                 RETURN(-ENOMEM);
1078
1079         for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1080                 cfs_mutex_init(&lmv->tgts[i].ltd_fid_mutex);
1081                 lmv->tgts[i].ltd_idx = i;
1082         }
1083
1084         lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1085
1086         OBD_ALLOC(lmv->datas, lmv->datas_size);
1087         if (lmv->datas == NULL)
1088                 GOTO(out_free_tgts, rc = -ENOMEM);
1089
1090         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1091         lmv->desc.ld_tgt_count = 0;
1092         lmv->desc.ld_active_tgt_count = 0;
1093         lmv->max_cookiesize = 0;
1094         lmv->max_def_easize = 0;
1095         lmv->max_easize = 0;
1096         lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1097
1098         cfs_spin_lock_init(&lmv->lmv_lock);
1099         cfs_mutex_init(&lmv->init_mutex);
1100
1101         rc = lmv_object_setup(obd);
1102         if (rc) {
1103                 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1104                 GOTO(out_free_datas, rc);
1105         }
1106
1107         lprocfs_lmv_init_vars(&lvars);
1108         lprocfs_obd_setup(obd, lvars.obd_vars);
1109 #ifdef LPROCFS
1110         {
1111                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd",
1112                                         0444, &lmv_proc_target_fops, obd);
1113                 if (rc)
1114                         CWARN("%s: error adding LMV target_obd file: rc = %d\n",
1115                                obd->obd_name, rc);
1116        }
1117 #endif
1118         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1119                              LUSTRE_CLI_FLD_HASH_DHT);
1120         if (rc) {
1121                 CERROR("Can't init FLD, err %d\n", rc);
1122                 GOTO(out_free_datas, rc);
1123         }
1124
1125         RETURN(0);
1126
1127 out_free_datas:
1128         OBD_FREE(lmv->datas, lmv->datas_size);
1129         lmv->datas = NULL;
1130 out_free_tgts:
1131         OBD_FREE(lmv->tgts, lmv->tgts_size);
1132         lmv->tgts = NULL;
1133         return rc;
1134 }
1135
1136 static int lmv_cleanup(struct obd_device *obd)
1137 {
1138         struct lmv_obd   *lmv = &obd->u.lmv;
1139         ENTRY;
1140
1141         fld_client_fini(&lmv->lmv_fld);
1142         lmv_object_cleanup(obd);
1143         OBD_FREE(lmv->datas, lmv->datas_size);
1144         OBD_FREE(lmv->tgts, lmv->tgts_size);
1145
1146         RETURN(0);
1147 }
1148
1149 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1150 {
1151         struct lustre_cfg     *lcfg = buf;
1152         struct obd_uuid        tgt_uuid;
1153         int                    rc;
1154         ENTRY;
1155
1156         switch(lcfg->lcfg_command) {
1157         case LCFG_ADD_MDC:
1158                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1159                         GOTO(out, rc = -EINVAL);
1160
1161                 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1162                 rc = lmv_add_target(obd, &tgt_uuid);
1163                 GOTO(out, rc);
1164         default: {
1165                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1166                 GOTO(out, rc = -EINVAL);
1167         }
1168         }
1169 out:
1170         RETURN(rc);
1171 }
1172
1173 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1174                       __u64 max_age, __u32 flags)
1175 {
1176         struct lmv_obd        *lmv = &obd->u.lmv;
1177         struct obd_statfs     *temp;
1178         int                    rc = 0;
1179         int                    i;
1180         ENTRY;
1181
1182         rc = lmv_check_connect(obd);
1183         if (rc)
1184                 RETURN(rc);
1185
1186         OBD_ALLOC(temp, sizeof(*temp));
1187         if (temp == NULL)
1188                 RETURN(-ENOMEM);
1189
1190         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1191                 if (lmv->tgts[i].ltd_exp == NULL)
1192                         continue;
1193
1194                 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1195                                 max_age, flags);
1196                 if (rc) {
1197                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1198                                lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1199                                rc);
1200                         GOTO(out_free_temp, rc);
1201                 }
1202                 if (i == 0) {
1203                         *osfs = *temp;
1204                 } else {
1205                         osfs->os_bavail += temp->os_bavail;
1206                         osfs->os_blocks += temp->os_blocks;
1207                         osfs->os_ffree += temp->os_ffree;
1208                         osfs->os_files += temp->os_files;
1209                 }
1210         }
1211
1212         EXIT;
1213 out_free_temp:
1214         OBD_FREE(temp, sizeof(*temp));
1215         return rc;
1216 }
1217
1218 static int lmv_getstatus(struct obd_export *exp,
1219                          struct lu_fid *fid,
1220                          struct obd_capa **pc)
1221 {
1222         struct obd_device    *obd = exp->exp_obd;
1223         struct lmv_obd       *lmv = &obd->u.lmv;
1224         int                   rc;
1225         ENTRY;
1226
1227         rc = lmv_check_connect(obd);
1228         if (rc)
1229                 RETURN(rc);
1230
1231         rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1232         RETURN(rc);
1233 }
1234
1235 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1236                         struct obd_capa *oc, obd_valid valid, const char *name,
1237                         const char *input, int input_size, int output_size,
1238                         int flags, struct ptlrpc_request **request)
1239 {
1240         struct obd_device      *obd = exp->exp_obd;
1241         struct lmv_obd         *lmv = &obd->u.lmv;
1242         struct lmv_tgt_desc    *tgt;
1243         int                     rc;
1244         ENTRY;
1245
1246         rc = lmv_check_connect(obd);
1247         if (rc)
1248                 RETURN(rc);
1249
1250         tgt = lmv_find_target(lmv, fid);
1251         if (IS_ERR(tgt))
1252                 RETURN(PTR_ERR(tgt));
1253
1254         rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1255                          input_size, output_size, flags, request);
1256
1257         RETURN(rc);
1258 }
1259
1260 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1261                         struct obd_capa *oc, obd_valid valid, const char *name,
1262                         const char *input, int input_size, int output_size,
1263                         int flags, __u32 suppgid,
1264                         struct ptlrpc_request **request)
1265 {
1266         struct obd_device      *obd = exp->exp_obd;
1267         struct lmv_obd         *lmv = &obd->u.lmv;
1268         struct lmv_tgt_desc    *tgt;
1269         int                     rc;
1270         ENTRY;
1271
1272         rc = lmv_check_connect(obd);
1273         if (rc)
1274                 RETURN(rc);
1275
1276         tgt = lmv_find_target(lmv, fid);
1277         if (IS_ERR(tgt))
1278                 RETURN(PTR_ERR(tgt));
1279
1280         rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1281                          input_size, output_size, flags, suppgid,
1282                          request);
1283
1284         RETURN(rc);
1285 }
1286
1287 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1288                        struct ptlrpc_request **request)
1289 {
1290         struct obd_device       *obd = exp->exp_obd;
1291         struct lmv_obd          *lmv = &obd->u.lmv;
1292         struct lmv_tgt_desc     *tgt;
1293         struct lmv_object       *obj;
1294         int                      rc;
1295         int                      i;
1296         ENTRY;
1297
1298         rc = lmv_check_connect(obd);
1299         if (rc)
1300                 RETURN(rc);
1301
1302         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1303         if (IS_ERR(tgt))
1304                 RETURN(PTR_ERR(tgt));
1305
1306         if (op_data->op_valid & OBD_MD_MDTIDX) {
1307                 op_data->op_mds = tgt->ltd_idx;
1308                 RETURN(0);
1309         }
1310
1311         rc = md_getattr(tgt->ltd_exp, op_data, request);
1312         if (rc)
1313                 RETURN(rc);
1314
1315         obj = lmv_object_find_lock(obd, &op_data->op_fid1);
1316
1317         CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(&op_data->op_fid1),
1318                obj ? "(split)" : "");
1319
1320         /*
1321          * If object is split, then we loop over all the slaves and gather size
1322          * attribute. In ideal world we would have to gather also mds field from
1323          * all slaves, as object is spread over the cluster and this is
1324          * definitely interesting information and it is not good to loss it,
1325          * but...
1326          */
1327         if (obj) {
1328                 struct mdt_body *body;
1329
1330                 if (*request == NULL) {
1331                         lmv_object_put(obj);
1332                         RETURN(rc);
1333                 }
1334
1335                 body = req_capsule_server_get(&(*request)->rq_pill,
1336                                               &RMF_MDT_BODY);
1337                 LASSERT(body != NULL);
1338
1339                 for (i = 0; i < obj->lo_objcount; i++) {
1340                         if (lmv->tgts[i].ltd_exp == NULL) {
1341                                 CWARN("%s: NULL export for %d\n",
1342                                       obd->obd_name, i);
1343                                 continue;
1344                         }
1345
1346                         /*
1347                          * Skip master object.
1348                          */
1349                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1350                                 continue;
1351
1352                         body->size += obj->lo_stripes[i].ls_size;
1353                 }
1354
1355                 lmv_object_put_unlock(obj);
1356         }
1357
1358         RETURN(rc);
1359 }
1360
1361 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1362                              ldlm_iterator_t it, void *data)
1363 {
1364         struct obd_device   *obd = exp->exp_obd;
1365         struct lmv_obd      *lmv = &obd->u.lmv;
1366         int                  i;
1367         int                  rc;
1368         ENTRY;
1369
1370         rc = lmv_check_connect(obd);
1371         if (rc)
1372                 RETURN(rc);
1373
1374         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1375
1376         /*
1377          * With CMD every object can have two locks in different namespaces:
1378          * lookup lock in space of mds storing direntry and update/open lock in
1379          * space of mds storing inode.
1380          */
1381         for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1382                 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1383
1384         RETURN(0);
1385 }
1386
1387 static int lmv_find_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1388                            ldlm_iterator_t it, void *data)
1389 {
1390         struct obd_device   *obd = exp->exp_obd;
1391         struct lmv_obd      *lmv = &obd->u.lmv;
1392         int                  i;
1393         int                  rc;
1394         ENTRY;
1395
1396         rc = lmv_check_connect(obd);
1397         if (rc)
1398                 RETURN(rc);
1399
1400         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1401
1402         /*
1403          * With CMD every object can have two locks in different namespaces:
1404          * lookup lock in space of mds storing direntry and update/open lock in
1405          * space of mds storing inode.
1406          */
1407         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1408                 rc = md_find_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1409                 if (rc)
1410                         RETURN(rc);
1411         }
1412
1413         RETURN(rc);
1414 }
1415
1416
1417 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1418                      struct md_open_data *mod, struct ptlrpc_request **request)
1419 {
1420         struct obd_device     *obd = exp->exp_obd;
1421         struct lmv_obd        *lmv = &obd->u.lmv;
1422         struct lmv_tgt_desc   *tgt;
1423         int                    rc;
1424         ENTRY;
1425
1426         rc = lmv_check_connect(obd);
1427         if (rc)
1428                 RETURN(rc);
1429
1430         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1431         if (IS_ERR(tgt))
1432                 RETURN(PTR_ERR(tgt));
1433
1434         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1435         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1436         RETURN(rc);
1437 }
1438
1439 /**
1440  * Called in the case MDS returns -ERESTART on create on open, what means that
1441  * directory is split and its LMV presentation object has to be updated.
1442  */
1443 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1444 {
1445         struct obd_device       *obd = exp->exp_obd;
1446         struct lmv_obd          *lmv = &obd->u.lmv;
1447         struct ptlrpc_request   *req = NULL;
1448         struct lmv_tgt_desc     *tgt;
1449         struct lmv_object       *obj;
1450         struct lustre_md         md;
1451         struct md_op_data       *op_data;
1452         int                      mealen;
1453         int                      rc;
1454         __u64                    valid;
1455         ENTRY;
1456
1457         md.mea = NULL;
1458         mealen = lmv_get_easize(lmv);
1459
1460         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1461
1462         tgt = lmv_find_target(lmv, fid);
1463         if (IS_ERR(tgt))
1464                 RETURN(PTR_ERR(tgt));
1465
1466         /*
1467          * Time to update mea of parent fid.
1468          */
1469
1470         OBD_ALLOC_PTR(op_data);
1471         if (op_data == NULL)
1472                 RETURN(-ENOMEM);
1473
1474         op_data->op_fid1 = *fid;
1475         op_data->op_mode = mealen;
1476         op_data->op_valid = valid;
1477
1478         rc = md_getattr(tgt->ltd_exp, op_data, &req);
1479         OBD_FREE_PTR(op_data);
1480         if (rc) {
1481                 CERROR("md_getattr() failed, error %d\n", rc);
1482                 GOTO(cleanup, rc);
1483         }
1484
1485         rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1486         if (rc) {
1487                 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1488                 GOTO(cleanup, rc);
1489         }
1490
1491         if (md.mea == NULL)
1492                 GOTO(cleanup, rc = -ENODATA);
1493
1494         obj = lmv_object_create(exp, fid, md.mea);
1495         if (IS_ERR(obj))
1496                 rc = PTR_ERR(obj);
1497         else
1498                 lmv_object_put(obj);
1499
1500         obd_free_memmd(exp, (void *)&md.mea);
1501         EXIT;
1502 cleanup:
1503         if (req)
1504                 ptlrpc_req_finished(req);
1505         return rc;
1506 }
1507
1508 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1509                const void *data, int datalen, int mode, __u32 uid,
1510                __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1511                struct ptlrpc_request **request)
1512 {
1513         struct obd_device       *obd = exp->exp_obd;
1514         struct lmv_obd          *lmv = &obd->u.lmv;
1515         struct lmv_tgt_desc     *tgt;
1516         struct lmv_object       *obj;
1517         int                      rc;
1518         int                      loop = 0;
1519         int                      sidx;
1520         ENTRY;
1521
1522         rc = lmv_check_connect(obd);
1523         if (rc)
1524                 RETURN(rc);
1525
1526         if (!lmv->desc.ld_active_tgt_count)
1527                 RETURN(-EIO);
1528 repeat:
1529         ++loop;
1530         LASSERT(loop <= 2);
1531
1532         obj = lmv_object_find(obd, &op_data->op_fid1);
1533         if (obj) {
1534                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1535                                        op_data->op_name, op_data->op_namelen);
1536                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1537                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1538                 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1539                 tgt = lmv_get_target(lmv, op_data->op_mds);
1540                 lmv_object_put(obj);
1541         } else {
1542                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1543                 op_data->op_bias |= MDS_CHECK_SPLIT;
1544                 op_data->op_mds = tgt->ltd_idx;
1545         }
1546
1547         if (IS_ERR(tgt))
1548                 RETURN(PTR_ERR(tgt));
1549
1550         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1551         if (rc == -ERESTART)
1552                 goto repeat;
1553         else if (rc)
1554                 RETURN(rc);
1555
1556         CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #%x\n",
1557                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1558                op_data->op_mds);
1559
1560         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1561         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1562                        cap_effective, rdev, request);
1563         if (rc == 0) {
1564                 if (*request == NULL)
1565                         RETURN(rc);
1566                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1567         } else if (rc == -ERESTART) {
1568                 LASSERT(*request != NULL);
1569                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1570                           "Got -ERESTART during create!\n");
1571                 ptlrpc_req_finished(*request);
1572                 *request = NULL;
1573
1574                 /*
1575                  * Directory got split. Time to update local object and repeat
1576                  * the request with proper MDS.
1577                  */
1578                 rc = lmv_handle_split(exp, &op_data->op_fid1);
1579                 if (rc == 0) {
1580                         rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1581                                                  op_data, &op_data->op_fid2);
1582                         if (rc)
1583                                 RETURN(rc);
1584                         goto repeat;
1585                 }
1586         }
1587         RETURN(rc);
1588 }
1589
1590 static int lmv_done_writing(struct obd_export *exp,
1591                             struct md_op_data *op_data,
1592                             struct md_open_data *mod)
1593 {
1594         struct obd_device     *obd = exp->exp_obd;
1595         struct lmv_obd        *lmv = &obd->u.lmv;
1596         struct lmv_tgt_desc   *tgt;
1597         int                    rc;
1598         ENTRY;
1599
1600         rc = lmv_check_connect(obd);
1601         if (rc)
1602                 RETURN(rc);
1603
1604         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1605         if (IS_ERR(tgt))
1606                 RETURN(PTR_ERR(tgt));
1607
1608         rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1609         RETURN(rc);
1610 }
1611
1612 static int
1613 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1614                    struct lookup_intent *it, struct md_op_data *op_data,
1615                    struct lustre_handle *lockh, void *lmm, int lmmsize)
1616 {
1617         struct obd_device     *obd = exp->exp_obd;
1618         struct lmv_obd        *lmv = &obd->u.lmv;
1619         struct lmv_stripe_md  *mea = op_data->op_mea1;
1620         struct md_op_data     *op_data2;
1621         struct lmv_tgt_desc   *tgt;
1622         int                    i;
1623         int                    rc = 0;
1624         ENTRY;
1625
1626         OBD_ALLOC_PTR(op_data2);
1627         if (op_data2 == NULL)
1628                 RETURN(-ENOMEM);
1629
1630         LASSERT(mea != NULL);
1631         for (i = 0; i < mea->mea_count; i++) {
1632                 memset(op_data2, 0, sizeof(*op_data2));
1633                 op_data2->op_fid1 = mea->mea_ids[i];
1634                 op_data2->op_bias = 0;
1635
1636                 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1637                 if (IS_ERR(tgt))
1638                         GOTO(cleanup, rc = PTR_ERR(tgt));
1639
1640                 if (tgt->ltd_exp == NULL)
1641                         continue;
1642
1643                 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1644                                 lockh + i, lmm, lmmsize, NULL, 0);
1645
1646                 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1647                        PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1648
1649                 if (rc)
1650                         GOTO(cleanup, rc);
1651
1652                 if (it->d.lustre.it_data) {
1653                         struct ptlrpc_request *req;
1654                         req = (struct ptlrpc_request *)it->d.lustre.it_data;
1655                         ptlrpc_req_finished(req);
1656                 }
1657
1658                 if (it->d.lustre.it_status)
1659                         GOTO(cleanup, rc = it->d.lustre.it_status);
1660         }
1661
1662         EXIT;
1663 cleanup:
1664         OBD_FREE_PTR(op_data2);
1665
1666         if (rc != 0) {
1667                 /*
1668                  * Drop all taken locks.
1669                  */
1670                 while (--i >= 0) {
1671                         if (lockh[i].cookie)
1672                                 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1673                         lockh[i].cookie = 0;
1674                 }
1675         }
1676         return rc;
1677 }
1678
1679 static int
1680 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1681                    struct lookup_intent *it, struct md_op_data *op_data,
1682                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1683                    int extra_lock_flags)
1684 {
1685         struct ptlrpc_request      *req = it->d.lustre.it_data;
1686         struct obd_device          *obd = exp->exp_obd;
1687         struct lmv_obd             *lmv = &obd->u.lmv;
1688         struct lustre_handle        plock;
1689         struct lmv_tgt_desc        *tgt;
1690         struct md_op_data          *rdata;
1691         struct lu_fid               fid1;
1692         struct mdt_body            *body;
1693         int                         rc = 0;
1694         int                         pmode;
1695         ENTRY;
1696
1697         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1698         LASSERT(body != NULL);
1699
1700         if (!(body->valid & OBD_MD_MDS))
1701                 RETURN(0);
1702
1703         CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1704                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1705
1706         /*
1707          * We got LOOKUP lock, but we really need attrs.
1708          */
1709         pmode = it->d.lustre.it_lock_mode;
1710         LASSERT(pmode != 0);
1711         memcpy(&plock, lockh, sizeof(plock));
1712         it->d.lustre.it_lock_mode = 0;
1713         it->d.lustre.it_data = NULL;
1714         fid1 = body->fid1;
1715
1716         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1717         ptlrpc_req_finished(req);
1718
1719         tgt = lmv_find_target(lmv, &fid1);
1720         if (IS_ERR(tgt))
1721                 GOTO(out, rc = PTR_ERR(tgt));
1722
1723         OBD_ALLOC_PTR(rdata);
1724         if (rdata == NULL)
1725                 GOTO(out, rc = -ENOMEM);
1726
1727         rdata->op_fid1 = fid1;
1728         rdata->op_bias = MDS_CROSS_REF;
1729
1730         rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1731                         lmm, lmmsize, NULL, extra_lock_flags);
1732         OBD_FREE_PTR(rdata);
1733         EXIT;
1734 out:
1735         ldlm_lock_decref(&plock, pmode);
1736         return rc;
1737 }
1738
1739 static int
1740 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1741             struct lookup_intent *it, struct md_op_data *op_data,
1742             struct lustre_handle *lockh, void *lmm, int lmmsize,
1743             struct ptlrpc_request **req, int extra_lock_flags)
1744 {
1745         struct obd_device        *obd = exp->exp_obd;
1746         struct lmv_obd           *lmv = &obd->u.lmv;
1747         struct lmv_tgt_desc      *tgt;
1748         struct lmv_object        *obj;
1749         int                       sidx;
1750         int                       rc;
1751         ENTRY;
1752
1753         rc = lmv_check_connect(obd);
1754         if (rc)
1755                 RETURN(rc);
1756
1757         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1758                LL_IT2STR(it), PFID(&op_data->op_fid1));
1759
1760         if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1761                 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1762                                         lockh, lmm, lmmsize);
1763                 RETURN(rc);
1764         }
1765
1766         obj = lmv_object_find(obd, &op_data->op_fid1);
1767         if (obj && op_data->op_namelen) {
1768                 sidx = raw_name2idx(obj->lo_hashtype,
1769                                        obj->lo_objcount,
1770                                        (char *)op_data->op_name,
1771                                        op_data->op_namelen);
1772                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1773                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1774         } else {
1775                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1776         }
1777         if (obj)
1778                 lmv_object_put(obj);
1779
1780         if (IS_ERR(tgt))
1781                 RETURN(PTR_ERR(tgt));
1782
1783         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1784                LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1785
1786         rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1787                         lmm, lmmsize, req, extra_lock_flags);
1788
1789         if (rc == 0 && it && it->it_op == IT_OPEN) {
1790                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1791                                         lmm, lmmsize, extra_lock_flags);
1792         }
1793         RETURN(rc);
1794 }
1795
1796 static int
1797 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1798                  struct ptlrpc_request **request)
1799 {
1800         struct ptlrpc_request   *req = NULL;
1801         struct obd_device       *obd = exp->exp_obd;
1802         struct lmv_obd          *lmv = &obd->u.lmv;
1803         struct lu_fid            rid = op_data->op_fid1;
1804         struct lmv_tgt_desc     *tgt;
1805         struct mdt_body         *body;
1806         struct lmv_object       *obj;
1807         obd_valid                valid = op_data->op_valid;
1808         int                      rc;
1809         int                      loop = 0;
1810         int                      sidx;
1811         ENTRY;
1812
1813         rc = lmv_check_connect(obd);
1814         if (rc)
1815                 RETURN(rc);
1816
1817 repeat:
1818         ++loop;
1819         LASSERT(loop <= 2);
1820         obj = lmv_object_find(obd, &rid);
1821         if (obj) {
1822                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1823                                     op_data->op_name, op_data->op_namelen);
1824                 rid = obj->lo_stripes[sidx].ls_fid;
1825                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1826                 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1827                 valid &= ~OBD_MD_FLCKSPLIT;
1828                 lmv_object_put(obj);
1829         } else {
1830                 tgt = lmv_find_target(lmv, &rid);
1831                 valid |= OBD_MD_FLCKSPLIT;
1832                 op_data->op_mds = tgt->ltd_idx;
1833         }
1834         if (IS_ERR(tgt))
1835                 RETURN(PTR_ERR(tgt));
1836
1837         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1838                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1839                PFID(&rid), tgt->ltd_idx);
1840
1841         op_data->op_valid = valid;
1842         op_data->op_fid1 = rid;
1843         rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1844         if (rc == 0) {
1845                 body = req_capsule_server_get(&(*request)->rq_pill,
1846                                               &RMF_MDT_BODY);
1847                 LASSERT(body != NULL);
1848
1849                 if (body->valid & OBD_MD_MDS) {
1850                         rid = body->fid1;
1851                         CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1852                                PFID(&rid));
1853
1854                         tgt = lmv_find_target(lmv, &rid);
1855                         if (IS_ERR(tgt)) {
1856                                 ptlrpc_req_finished(*request);
1857                                 RETURN(PTR_ERR(tgt));
1858                         }
1859
1860                         op_data->op_fid1 = rid;
1861                         op_data->op_valid |= OBD_MD_FLCROSSREF;
1862                         op_data->op_namelen = 0;
1863                         op_data->op_name = NULL;
1864                         rc = md_getattr_name(tgt->ltd_exp, op_data, &req);
1865                         ptlrpc_req_finished(*request);
1866                         *request = req;
1867                 }
1868         } else if (rc == -ERESTART) {
1869                 LASSERT(*request != NULL);
1870                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1871                           "Got -ERESTART during getattr!\n");
1872                 ptlrpc_req_finished(*request);
1873                 *request = NULL;
1874
1875                 /*
1876                  * Directory got split. Time to update local object and repeat
1877                  * the request with proper MDS.
1878                  */
1879                 rc = lmv_handle_split(exp, &rid);
1880                 if (rc == 0)
1881                         goto repeat;
1882         }
1883         RETURN(rc);
1884 }
1885
1886 #define md_op_data_fid(op_data, fl)                     \
1887         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1888          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1889          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1890          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1891          NULL)
1892
1893 static int lmv_early_cancel_slaves(struct obd_export *exp,
1894                                    struct md_op_data *op_data, int op_tgt,
1895                                    ldlm_mode_t mode, int bits, int flag)
1896 {
1897         struct obd_device      *obd = exp->exp_obd;
1898         struct lmv_obd         *lmv = &obd->u.lmv;
1899         ldlm_policy_data_t      policy = {{0}};
1900         struct lu_fid          *op_fid;
1901         struct lu_fid          *st_fid;
1902         struct lmv_tgt_desc    *tgt;
1903         struct lmv_object      *obj;
1904         int                     rc = 0;
1905         int                     i;
1906         ENTRY;
1907
1908         op_fid = md_op_data_fid(op_data, flag);
1909         if (!fid_is_sane(op_fid))
1910                 RETURN(0);
1911
1912         obj = lmv_object_find(obd, op_fid);
1913         if (obj == NULL)
1914                 RETURN(-EALREADY);
1915
1916         policy.l_inodebits.bits = bits;
1917         for (i = 0; i < obj->lo_objcount; i++) {
1918                 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1919                 st_fid = &obj->lo_stripes[i].ls_fid;
1920                 if (op_tgt != tgt->ltd_idx) {
1921                         CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1922                                PFID(st_fid), tgt->ltd_idx);
1923                         rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1924                                               mode, LCF_ASYNC, NULL);
1925                         if (rc)
1926                                 GOTO(out_put_obj, rc);
1927                 } else {
1928                         CDEBUG(D_INODE,
1929                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1930                                op_tgt, PFID(st_fid));
1931                         /*
1932                          * Do not cancel locks for operation target, they will
1933                          * be handled later in underlaying layer when calling
1934                          * function we run on behalf of.
1935                          */
1936                         *op_fid = *st_fid;
1937                         op_data->op_flags |= flag;
1938                 }
1939         }
1940         EXIT;
1941 out_put_obj:
1942         lmv_object_put(obj);
1943         return rc;
1944 }
1945
1946 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1947                             int op_tgt, ldlm_mode_t mode, int bits, int flag)
1948 {
1949         struct lu_fid          *fid = md_op_data_fid(op_data, flag);
1950         struct obd_device      *obd = exp->exp_obd;
1951         struct lmv_obd         *lmv = &obd->u.lmv;
1952         struct lmv_tgt_desc    *tgt;
1953         ldlm_policy_data_t      policy = {{0}};
1954         struct lmv_object      *obj;
1955         int                     rc = 0;
1956         ENTRY;
1957
1958         if (!fid_is_sane(fid))
1959                 RETURN(0);
1960
1961         obj = lmv_object_find(obd, fid);
1962         if (obj) {
1963                 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1964                                              bits, flag);
1965                 lmv_object_put(obj);
1966         } else {
1967                 tgt = lmv_find_target(lmv, fid);
1968                 if (IS_ERR(tgt))
1969                         RETURN(PTR_ERR(tgt));
1970
1971                 if (tgt->ltd_idx != op_tgt) {
1972                         CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1973                         policy.l_inodebits.bits = bits;
1974                         rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1975                                               mode, LCF_ASYNC, NULL);
1976                 } else {
1977                         CDEBUG(D_INODE,
1978                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1979                                op_tgt, PFID(fid));
1980                         op_data->op_flags |= flag;
1981                         rc = 0;
1982                 }
1983
1984         }
1985         RETURN(rc);
1986 }
1987
1988 /*
1989  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1990  * op_data->op_fid2
1991  */
1992 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1993                     struct ptlrpc_request **request)
1994 {
1995         struct obd_device       *obd = exp->exp_obd;
1996         struct lmv_obd          *lmv = &obd->u.lmv;
1997         struct lmv_tgt_desc     *tgt;
1998         struct lmv_object       *obj;
1999         int                      rc;
2000         int                      loop = 0;
2001         mdsno_t                  mds;
2002         int                      sidx;
2003         ENTRY;
2004
2005         rc = lmv_check_connect(obd);
2006         if (rc)
2007                 RETURN(rc);
2008
2009 repeat:
2010         ++loop;
2011         LASSERT(loop <= 2);
2012         LASSERT(op_data->op_namelen != 0);
2013
2014         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2015                PFID(&op_data->op_fid2), op_data->op_namelen,
2016                op_data->op_name, PFID(&op_data->op_fid1));
2017
2018         obj = lmv_object_find(obd, &op_data->op_fid2);
2019         if (obj) {
2020                 sidx = raw_name2idx(obj->lo_hashtype,
2021                                     obj->lo_objcount,
2022                                     op_data->op_name,
2023                                     op_data->op_namelen);
2024                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2025                 mds = obj->lo_stripes[sidx].ls_mds;
2026                 lmv_object_put(obj);
2027         } else {
2028                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
2029                 if (rc)
2030                         RETURN(rc);
2031         }
2032
2033         CDEBUG(D_INODE, "Forward to mds #%x ("DFID")\n",
2034                mds, PFID(&op_data->op_fid1));
2035
2036         op_data->op_fsuid = cfs_curproc_fsuid();
2037         op_data->op_fsgid = cfs_curproc_fsgid();
2038         op_data->op_cap = cfs_curproc_cap_pack();
2039         tgt = lmv_get_target(lmv, mds);
2040
2041         /*
2042          * Cancel UPDATE lock on child (fid1).
2043          */
2044         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2045         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2046                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2047         if (rc == 0)
2048                 rc = md_link(tgt->ltd_exp, op_data, request);
2049         if (rc == -ERESTART) {
2050                 LASSERT(*request != NULL);
2051                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2052                           "Got -ERESTART during link!\n");
2053                 ptlrpc_req_finished(*request);
2054                 *request = NULL;
2055
2056                 /*
2057                  * Directory got split. Time to update local object and repeat
2058                  * the request with proper MDS.
2059                  */
2060                 rc = lmv_handle_split(exp, &op_data->op_fid2);
2061                 if (rc == 0)
2062                         goto repeat;
2063         }
2064
2065         RETURN(rc);
2066 }
2067
2068 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2069                       const char *old, int oldlen, const char *new, int newlen,
2070                       struct ptlrpc_request **request)
2071 {
2072         struct obd_device       *obd = exp->exp_obd;
2073         struct lmv_obd          *lmv = &obd->u.lmv;
2074         struct lmv_tgt_desc     *src_tgt;
2075         int                      rc;
2076         int                      sidx;
2077         int                      loop = 0;
2078         struct lmv_object       *obj;
2079         mdsno_t                  mds1;
2080         mdsno_t                  mds2;
2081         ENTRY;
2082
2083         LASSERT(oldlen != 0);
2084
2085         CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2086                oldlen, old, PFID(&op_data->op_fid1),
2087                newlen, new, PFID(&op_data->op_fid2));
2088
2089         rc = lmv_check_connect(obd);
2090         if (rc)
2091                 RETURN(rc);
2092
2093 repeat:
2094         ++loop;
2095         LASSERT(loop <= 2);
2096         obj = lmv_object_find(obd, &op_data->op_fid1);
2097         if (obj) {
2098                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2099                                     (char *)old, oldlen);
2100                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2101                 mds1 = obj->lo_stripes[sidx].ls_mds;
2102                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2103                 lmv_object_put(obj);
2104         } else {
2105                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2106                 if (rc)
2107                         RETURN(rc);
2108         }
2109
2110         obj = lmv_object_find(obd, &op_data->op_fid2);
2111         if (obj) {
2112                 /*
2113                  * Directory is already split, so we have to forward request to
2114                  * the right MDS.
2115                  */
2116                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2117                                        (char *)new, newlen);
2118
2119                 mds2 = obj->lo_stripes[sidx].ls_mds;
2120                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2121                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2122                 lmv_object_put(obj);
2123         } else {
2124                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2125                 if (rc)
2126                         RETURN(rc);
2127         }
2128
2129         op_data->op_fsuid = cfs_curproc_fsuid();
2130         op_data->op_fsgid = cfs_curproc_fsgid();
2131         op_data->op_cap = cfs_curproc_cap_pack();
2132
2133         src_tgt = lmv_get_target(lmv, mds1);
2134
2135         /*
2136          * LOOKUP lock on src child (fid3) should also be cancelled for
2137          * src_tgt in mdc_rename.
2138          */
2139         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2140
2141         /*
2142          * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2143          * own target.
2144          */
2145         rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2146                               LCK_EX, MDS_INODELOCK_UPDATE,
2147                               MF_MDC_CANCEL_FID2);
2148
2149         /*
2150          * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2151          */
2152         if (rc == 0) {
2153                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2154                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2155                                       MF_MDC_CANCEL_FID4);
2156         }
2157
2158         /*
2159          * Cancel all the locks on tgt child (fid4).
2160          */
2161         if (rc == 0)
2162                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2163                                       LCK_EX, MDS_INODELOCK_FULL,
2164                                       MF_MDC_CANCEL_FID4);
2165
2166         if (rc == 0)
2167                 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2168                                new, newlen, request);
2169
2170         if (rc == -ERESTART) {
2171                 LASSERT(*request != NULL);
2172                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2173                           "Got -ERESTART during rename!\n");
2174                 ptlrpc_req_finished(*request);
2175                 *request = NULL;
2176
2177                 /*
2178                  * Directory got split. Time to update local object and repeat
2179                  * the request with proper MDS.
2180                  */
2181                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2182                 if (rc == 0)
2183                         goto repeat;
2184         }
2185         RETURN(rc);
2186 }
2187
2188 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2189                        void *ea, int ealen, void *ea2, int ea2len,
2190                        struct ptlrpc_request **request,
2191                        struct md_open_data **mod)
2192 {
2193         struct obd_device       *obd = exp->exp_obd;
2194         struct lmv_obd          *lmv = &obd->u.lmv;
2195         struct ptlrpc_request   *req;
2196         struct lmv_tgt_desc     *tgt;
2197         struct lmv_object       *obj;
2198         int                      rc = 0;
2199         int                      i;
2200         ENTRY;
2201
2202         rc = lmv_check_connect(obd);
2203         if (rc)
2204                 RETURN(rc);
2205
2206         obj = lmv_object_find(obd, &op_data->op_fid1);
2207
2208         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2209                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2210                obj ? ", split" : "");
2211
2212         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2213         if (obj) {
2214                 for (i = 0; i < obj->lo_objcount; i++) {
2215                         op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2216
2217                         tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2218                         if (IS_ERR(tgt)) {
2219                                 rc = PTR_ERR(tgt);
2220                                 break;
2221                         }
2222
2223                         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2224                                         ea2, ea2len, &req, mod);
2225
2226                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2227                                 /*
2228                                  * This is master object and this request should
2229                                  * be returned back to llite.
2230                                  */
2231                                 *request = req;
2232                         } else {
2233                                 ptlrpc_req_finished(req);
2234                         }
2235
2236                         if (rc)
2237                                 break;
2238                 }
2239                 lmv_object_put(obj);
2240         } else {
2241                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2242                 if (IS_ERR(tgt))
2243                         RETURN(PTR_ERR(tgt));
2244
2245                 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2246                                 ea2len, request, mod);
2247         }
2248         RETURN(rc);
2249 }
2250
2251 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2252                     struct obd_capa *oc, struct ptlrpc_request **request)
2253 {
2254         struct obd_device         *obd = exp->exp_obd;
2255         struct lmv_obd            *lmv = &obd->u.lmv;
2256         struct lmv_tgt_desc       *tgt;
2257         int                        rc;
2258         ENTRY;
2259
2260         rc = lmv_check_connect(obd);
2261         if (rc)
2262                 RETURN(rc);
2263
2264         tgt = lmv_find_target(lmv, fid);
2265         if (IS_ERR(tgt))
2266                 RETURN(PTR_ERR(tgt));
2267
2268         rc = md_sync(tgt->ltd_exp, fid, oc, request);
2269         RETURN(rc);
2270 }
2271
2272 /**
2273  * Main purpose of LMV blocking ast is to remove split directory LMV
2274  * presentation object (struct lmv_object) attached to the lock being revoked.
2275  */
2276 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2277                      void *data, int flag)
2278 {
2279         struct lustre_handle    lockh;
2280         struct lmv_object      *obj;
2281         int                     rc;
2282         ENTRY;
2283
2284         switch (flag) {
2285         case LDLM_CB_BLOCKING:
2286                 ldlm_lock2handle(lock, &lockh);
2287                 rc = ldlm_cli_cancel(&lockh);
2288                 if (rc < 0) {
2289                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2290                         RETURN(rc);
2291                 }
2292                 break;
2293         case LDLM_CB_CANCELING:
2294                 /*
2295                  * Time to drop cached attrs for split directory object
2296                  */
2297                 obj = lock->l_ast_data;
2298                 if (obj) {
2299                         CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2300                                ", master "DFID"\n",
2301                                lock->l_resource->lr_name.name[3] == 1 ?
2302                                "LOOKUP" : "UPDATE",
2303                                lock->l_resource->lr_name.name[0],
2304                                lock->l_resource->lr_name.name[1],
2305                                PFID(&obj->lo_fid));
2306                         lmv_object_put(obj);
2307                 }
2308                 break;
2309         default:
2310                 LBUG();
2311         }
2312         RETURN(0);
2313 }
2314
2315 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2316 {
2317         __u64         val;
2318
2319         val = le64_to_cpu(*hash);
2320         if (val < hash_adj)
2321                 val += MAX_HASH_SIZE;
2322         if (val != MDS_DIR_END_OFF)
2323                 *hash = cpu_to_le64(val - hash_adj);
2324 }
2325
2326 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2327 {
2328         __u64              id;
2329         struct obd_import *imp;
2330
2331         /*
2332          * XXX: to get nid we assume that underlying obd device is mdc.
2333          */
2334         imp  = class_exp2cliimp(exp);
2335         id   = imp->imp_connection->c_self + fid_flatten(fid);
2336
2337         CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2338                imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2339
2340         return id ^ (id >> 32);
2341 }
2342
2343 static int lmv_readpage(struct obd_export *exp, struct md_op_data *op_data,
2344                         struct page **pages, struct ptlrpc_request **request)
2345 {
2346         struct obd_device       *obd = exp->exp_obd;
2347         struct lmv_obd          *lmv = &obd->u.lmv;
2348         struct lmv_object       *obj;
2349         struct lu_fid            rid = op_data->op_fid1;
2350         __u64                    offset = op_data->op_offset;
2351         __u64                    hash_adj = 0;
2352         __u32                    rank = 0;
2353         __u64                    seg_size = 0;
2354         __u64                    tgt_tmp = 0;
2355         int                      tgt_idx = 0;
2356         int                      tgt0_idx = 0;
2357         int                      rc;
2358         int                      nr = 0;
2359         int                      i;
2360         /* number of pages read, in CFS_PAGE_SIZE */
2361         int                      nrdpgs;
2362         /* number of pages transferred in LU_PAGE_SIZE */
2363         int                      nlupgs;
2364         struct lmv_stripe       *los;
2365         struct lmv_tgt_desc     *tgt;
2366         struct lu_dirpage       *dp;
2367         struct lu_dirent        *ent;
2368         ENTRY;
2369
2370         rc = lmv_check_connect(obd);
2371         if (rc)
2372                 RETURN(rc);
2373
2374         CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2375
2376         /*
2377          * This case handle directory lookup in clustered metadata case (i.e.
2378          * split directory is located on multiple md servers.)
2379          * each server keeps directory entries for certain range of hashes.
2380          * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2381          * first server will keep records with hashes [ 0 ... MAX_HASH / N  - 1],
2382          * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2383          * so on....
2384          *      readdir can simply start reading entries from 0 - N server in
2385          * order but that will not scale well as all client will request dir in
2386          * to server in same order.
2387          * Following algorithm does optimization:
2388          * Instead of doing readdir in 1, 2, ...., N order, client with a
2389          * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2390          * (every client has rank R)
2391          *      But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2392          * since client ask dir from MDS{R} client has pages with offsets
2393          * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2394          * on hash  values that we get.
2395          */
2396         obj = lmv_object_find_lock(obd, &rid);
2397         if (obj) {
2398                 nr       = obj->lo_objcount;
2399                 LASSERT(nr > 0);
2400                 seg_size = MAX_HASH_SIZE;
2401                 do_div(seg_size, nr);
2402                 los      = obj->lo_stripes;
2403                 tgt      = lmv_get_target(lmv, los[0].ls_mds);
2404                 rank     = lmv_node_rank(tgt->ltd_exp, &rid) % nr;
2405                 tgt_tmp  = offset;
2406                 do_div(tgt_tmp, seg_size);
2407                 tgt0_idx = do_div(tgt_tmp,  nr);
2408                 tgt_idx  = (tgt0_idx + rank) % nr;
2409
2410                 if (tgt_idx < tgt0_idx)
2411                         /*
2412                          * Wrap around.
2413                          *
2414                          * Last segment has unusual length due to division
2415                          * rounding.
2416                          */
2417                         hash_adj = MAX_HASH_SIZE - seg_size * nr;
2418                 else
2419                         hash_adj = 0;
2420
2421                 hash_adj += rank * seg_size;
2422
2423                 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2424                        LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2425                        offset, tgt0_idx, offset + hash_adj, tgt_idx);
2426
2427                 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2428                 rid = obj->lo_stripes[tgt_idx].ls_fid;
2429                 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2430
2431                 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2432                        PFID(&rid), (unsigned long)offset, tgt_idx);
2433         } else
2434                 tgt = lmv_find_target(lmv, &rid);
2435
2436         if (IS_ERR(tgt))
2437                 GOTO(cleanup, rc = PTR_ERR(tgt));
2438
2439         op_data->op_fid1 = rid;
2440         rc = md_readpage(tgt->ltd_exp, op_data, pages, request);
2441         if (rc)
2442                 GOTO(cleanup, rc);
2443
2444         nrdpgs = ((*request)->rq_bulk->bd_nob_transferred + CFS_PAGE_SIZE - 1)
2445                  >> CFS_PAGE_SHIFT;
2446         nlupgs = (*request)->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
2447         LASSERT(!((*request)->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
2448         LASSERT(nrdpgs > 0 && nrdpgs <= op_data->op_npages);
2449
2450         CDEBUG(D_INODE, "read %d(%d)/%d pages\n", nrdpgs, nlupgs,
2451                op_data->op_npages);
2452
2453         for (i = 0; i < nrdpgs; i++) {
2454 #if CFS_PAGE_SIZE > LU_PAGE_SIZE
2455                 struct lu_dirpage *first;
2456                 __u64 hash_end = 0;
2457                 __u32 flags = 0;
2458 #endif
2459                 struct lu_dirent *tmp = NULL;
2460
2461                 dp = cfs_kmap(pages[i]);
2462                 if (obj) {
2463                         lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2464                         lmv_hash_adjust(&dp->ldp_hash_end,   hash_adj);
2465                         LASSERT(le64_to_cpu(dp->ldp_hash_start) <=
2466                                 op_data->op_offset);
2467
2468                         if ((tgt0_idx != nr - 1) &&
2469                             (le64_to_cpu(dp->ldp_hash_end) == MDS_DIR_END_OFF))
2470                         {
2471                                 dp->ldp_hash_end = cpu_to_le32(seg_size *
2472                                                                (tgt0_idx + 1));
2473                                 CDEBUG(D_INODE,
2474                                        ""DFID" reset end "LPX64" tgt %d\n",
2475                                        PFID(&rid),
2476                                        (__u64)le64_to_cpu(dp->ldp_hash_end),
2477                                        tgt_idx);
2478                         }
2479                 }
2480
2481                 ent = lu_dirent_start(dp);
2482 #if CFS_PAGE_SIZE > LU_PAGE_SIZE
2483                 first = dp;
2484                 hash_end = dp->ldp_hash_end;
2485 repeat:
2486 #endif
2487                 nlupgs--;
2488                 for (tmp = ent; ent != NULL;
2489                      tmp = ent, ent = lu_dirent_next(ent)) {
2490                         if (obj)
2491                                 lmv_hash_adjust(&ent->lde_hash, hash_adj);
2492                 }
2493
2494 #if CFS_PAGE_SIZE > LU_PAGE_SIZE
2495                 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
2496                 if (((unsigned long)dp & ~CFS_PAGE_MASK) && nlupgs > 0) {
2497                         ent = lu_dirent_start(dp);
2498
2499                         if (obj) {
2500                                 lmv_hash_adjust(&dp->ldp_hash_end, hash_adj);
2501                                 if ((tgt0_idx != nr - 1) &&
2502                                     (le64_to_cpu(dp->ldp_hash_end) ==
2503                                      MDS_DIR_END_OFF)) {
2504                                         hash_end = cpu_to_le32(seg_size *
2505                                                                (tgt0_idx + 1));
2506                                         CDEBUG(D_INODE,
2507                                             ""DFID" reset end "LPX64" tgt %d\n",
2508                                             PFID(&rid),
2509                                             (__u64)le64_to_cpu(hash_end),
2510                                             tgt_idx);
2511                                 }
2512                         }
2513                         hash_end = dp->ldp_hash_end;
2514                         flags = dp->ldp_flags;
2515
2516                         if (tmp) {
2517                                 /* enlarge the end entry lde_reclen from 0 to
2518                                  * first entry of next lu_dirpage, in this way
2519                                  * several lu_dirpages can be stored into one
2520                                  * client page on client. */
2521                                 tmp = ((void *)tmp) +
2522                                       le16_to_cpu(tmp->lde_reclen);
2523                                 tmp->lde_reclen =
2524                                         cpu_to_le16((char *)(dp->ldp_entries) -
2525                                                     (char *)tmp);
2526                                 goto repeat;
2527                         }
2528                 }
2529                 first->ldp_hash_end = hash_end;
2530                 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
2531                 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
2532 #else
2533                 SET_BUT_UNUSED(tmp);
2534 #endif
2535                 cfs_kunmap(pages[i]);
2536         }
2537         EXIT;
2538 cleanup:
2539         if (obj)
2540                 lmv_object_put_unlock(obj);
2541         return rc;
2542 }
2543
2544 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2545                       struct ptlrpc_request **request)
2546 {
2547         struct obd_device       *obd = exp->exp_obd;
2548         struct lmv_obd          *lmv = &obd->u.lmv;
2549         struct lmv_tgt_desc     *tgt = NULL;
2550         struct lmv_object       *obj;
2551         int                      rc;
2552         int                      sidx;
2553         int                      loop = 0;
2554         ENTRY;
2555
2556         rc = lmv_check_connect(obd);
2557         if (rc)
2558                 RETURN(rc);
2559
2560 repeat:
2561         ++loop;
2562         LASSERT(loop <= 2);
2563         LASSERT(op_data->op_namelen != 0);
2564
2565         obj = lmv_object_find(obd, &op_data->op_fid1);
2566         if (obj) {
2567                 sidx = raw_name2idx(obj->lo_hashtype,
2568                                     obj->lo_objcount,
2569                                     op_data->op_name,
2570                                     op_data->op_namelen);
2571                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2572                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2573                 tgt = lmv_get_target(lmv,
2574                                      obj->lo_stripes[sidx].ls_mds);
2575                 lmv_object_put(obj);
2576                 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2577                        op_data->op_namelen, op_data->op_name,
2578                        PFID(&op_data->op_fid1), sidx);
2579         }
2580
2581         if (tgt == NULL) {
2582                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2583                 if (IS_ERR(tgt))
2584                         RETURN(PTR_ERR(tgt));
2585                 op_data->op_bias |= MDS_CHECK_SPLIT;
2586         }
2587
2588         op_data->op_fsuid = cfs_curproc_fsuid();
2589         op_data->op_fsgid = cfs_curproc_fsgid();
2590         op_data->op_cap = cfs_curproc_cap_pack();
2591
2592         /*
2593          * If child's fid is given, cancel unused locks for it if it is from
2594          * another export than parent.
2595          *
2596          * LOOKUP lock for child (fid3) should also be cancelled on parent
2597          * tgt_tgt in mdc_unlink().
2598          */
2599         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2600
2601         /*
2602          * Cancel FULL locks on child (fid3).
2603          */
2604         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2605                               MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2606
2607         if (rc == 0)
2608                 rc = md_unlink(tgt->ltd_exp, op_data, request);
2609
2610         if (rc == -ERESTART) {
2611                 LASSERT(*request != NULL);
2612                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2613                           "Got -ERESTART during unlink!\n");
2614                 ptlrpc_req_finished(*request);
2615                 *request = NULL;
2616
2617                 /*
2618                  * Directory got split. Time to update local object and repeat
2619                  * the request with proper MDS.
2620                  */
2621                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2622                 if (rc == 0)
2623                         goto repeat;
2624         }
2625         RETURN(rc);
2626 }
2627
2628 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2629 {
2630         struct lmv_obd *lmv = &obd->u.lmv;
2631         int rc = 0;
2632
2633         switch (stage) {
2634         case OBD_CLEANUP_EARLY:
2635                 /* XXX: here should be calling obd_precleanup() down to
2636                  * stack. */
2637                 break;
2638         case OBD_CLEANUP_EXPORTS:
2639                 fld_client_proc_fini(&lmv->lmv_fld);
2640                 lprocfs_obd_cleanup(obd);
2641                 rc = obd_llog_finish(obd, 0);
2642                 if (rc != 0)
2643                         CERROR("failed to cleanup llogging subsystems\n");
2644                 break;
2645         default:
2646                 break;
2647         }
2648         RETURN(rc);
2649 }
2650
2651 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2652                         void *key, __u32 *vallen, void *val,
2653                         struct lov_stripe_md *lsm)
2654 {
2655         struct obd_device       *obd;
2656         struct lmv_obd          *lmv;
2657         int                      rc = 0;
2658         ENTRY;
2659
2660         obd = class_exp2obd(exp);
2661         if (obd == NULL) {
2662                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2663                        exp->exp_handle.h_cookie);
2664                 RETURN(-EINVAL);
2665         }
2666
2667         lmv = &obd->u.lmv;
2668         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2669                 struct lmv_tgt_desc *tgts;
2670                 int i;
2671
2672                 rc = lmv_check_connect(obd);
2673                 if (rc)
2674                         RETURN(rc);
2675
2676                 LASSERT(*vallen == sizeof(__u32));
2677                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2678                      i++, tgts++) {
2679
2680                         /*
2681                          * All tgts should be connected when this gets called.
2682                          */
2683                         if (!tgts || !tgts->ltd_exp) {
2684                                 CERROR("target not setup?\n");
2685                                 continue;
2686                         }
2687
2688                         if (!obd_get_info(tgts->ltd_exp, keylen, key,
2689                                           vallen, val, NULL))
2690                                 RETURN(0);
2691                 }
2692                 RETURN(-EINVAL);
2693         } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2694                 rc = lmv_check_connect(obd);
2695                 if (rc)
2696                         RETURN(rc);
2697
2698                 /*
2699                  * Forwarding this request to first MDS, it should know LOV
2700                  * desc.
2701                  */
2702                 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2703                                   vallen, val, NULL);
2704                 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2705                         exp->exp_connect_flags =
2706                         ((struct obd_connect_data *)val)->ocd_connect_flags;
2707                 }
2708                 RETURN(rc);
2709         } else if (KEY_IS(KEY_TGT_COUNT)) {
2710                 *((int *)val) = lmv->desc.ld_tgt_count;
2711                 RETURN(0);
2712         }
2713
2714         CDEBUG(D_IOCTL, "Invalid key\n");
2715         RETURN(-EINVAL);
2716 }
2717
2718 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2719                        void *key, obd_count vallen, void *val,
2720                        struct ptlrpc_request_set *set)
2721 {
2722         struct lmv_tgt_desc    *tgt;
2723         struct obd_device      *obd;
2724         struct lmv_obd         *lmv;
2725         int rc = 0;
2726         ENTRY;
2727
2728         obd = class_exp2obd(exp);
2729         if (obd == NULL) {
2730                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2731                        exp->exp_handle.h_cookie);
2732                 RETURN(-EINVAL);
2733         }
2734         lmv = &obd->u.lmv;
2735
2736         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX)) {
2737                 int i, err = 0;
2738
2739                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2740                         tgt = &lmv->tgts[i];
2741
2742                         if (!tgt->ltd_exp)
2743                                 continue;
2744
2745                         err = obd_set_info_async(tgt->ltd_exp,
2746                                                  keylen, key, vallen, val, set);
2747                         if (err && rc == 0)
2748                                 rc = err;
2749                 }
2750
2751                 RETURN(rc);
2752         }
2753
2754         RETURN(-EINVAL);
2755 }
2756
2757 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2758                struct lov_stripe_md *lsm)
2759 {
2760         struct obd_device         *obd = class_exp2obd(exp);
2761         struct lmv_obd            *lmv = &obd->u.lmv;
2762         struct lmv_stripe_md      *meap;
2763         struct lmv_stripe_md      *lsmp;
2764         int                        mea_size;
2765         int                        i;
2766         ENTRY;
2767
2768         mea_size = lmv_get_easize(lmv);
2769         if (!lmmp)
2770                 RETURN(mea_size);
2771
2772         if (*lmmp && !lsm) {
2773                 OBD_FREE_LARGE(*lmmp, mea_size);
2774                 *lmmp = NULL;
2775                 RETURN(0);
2776         }
2777
2778         if (*lmmp == NULL) {
2779                 OBD_ALLOC_LARGE(*lmmp, mea_size);
2780                 if (*lmmp == NULL)
2781                         RETURN(-ENOMEM);
2782         }
2783
2784         if (!lsm)
2785                 RETURN(mea_size);
2786
2787         lsmp = (struct lmv_stripe_md *)lsm;
2788         meap = (struct lmv_stripe_md *)*lmmp;
2789
2790         if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2791             lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2792                 RETURN(-EINVAL);
2793
2794         meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2795         meap->mea_count = cpu_to_le32(lsmp->mea_count);
2796         meap->mea_master = cpu_to_le32(lsmp->mea_master);
2797
2798         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2799                 meap->mea_ids[i] = meap->mea_ids[i];
2800                 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2801         }
2802
2803         RETURN(mea_size);
2804 }
2805
2806 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2807                  struct lov_mds_md *lmm, int lmm_size)
2808 {
2809         struct obd_device          *obd = class_exp2obd(exp);
2810         struct lmv_stripe_md      **tmea = (struct lmv_stripe_md **)lsmp;
2811         struct lmv_stripe_md       *mea = (struct lmv_stripe_md *)lmm;
2812         struct lmv_obd             *lmv = &obd->u.lmv;
2813         int                         mea_size;
2814         int                         i;
2815         __u32                       magic;
2816         ENTRY;
2817
2818         mea_size = lmv_get_easize(lmv);
2819         if (lsmp == NULL)
2820                 return mea_size;
2821
2822         if (*lsmp != NULL && lmm == NULL) {
2823                 OBD_FREE_LARGE(*tmea, mea_size);
2824                 *lsmp = NULL;
2825                 RETURN(0);
2826         }
2827
2828         LASSERT(mea_size == lmm_size);
2829
2830         OBD_ALLOC_LARGE(*tmea, mea_size);
2831         if (*tmea == NULL)
2832                 RETURN(-ENOMEM);
2833
2834         if (!lmm)
2835                 RETURN(mea_size);
2836
2837         if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2838             mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2839             mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2840         {
2841                 magic = le32_to_cpu(mea->mea_magic);
2842         } else {
2843                 /*
2844                  * Old mea is not handled here.
2845                  */
2846                 CERROR("Old not supportable EA is found\n");
2847                 LBUG();
2848         }
2849
2850         (*tmea)->mea_magic = magic;
2851         (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2852         (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2853
2854         for (i = 0; i < (*tmea)->mea_count; i++) {
2855                 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2856                 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2857         }
2858         RETURN(mea_size);
2859 }
2860
2861 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2862                              ldlm_policy_data_t *policy, ldlm_mode_t mode,
2863                              ldlm_cancel_flags_t flags, void *opaque)
2864 {
2865         struct obd_device       *obd = exp->exp_obd;
2866         struct lmv_obd          *lmv = &obd->u.lmv;
2867         int                      rc = 0;
2868         int                      err;
2869         int                      i;
2870         ENTRY;
2871
2872         LASSERT(fid != NULL);
2873
2874         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2875                 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2876                         continue;
2877
2878                 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2879                                        policy, mode, flags, opaque);
2880                 if (!rc)
2881                         rc = err;
2882         }
2883         RETURN(rc);
2884 }
2885
2886 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
2887                       __u64 *bits)
2888 {
2889         struct obd_device       *obd = exp->exp_obd;
2890         struct lmv_obd          *lmv = &obd->u.lmv;
2891         int                      rc;
2892         ENTRY;
2893
2894         rc =  md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data, bits);
2895         RETURN(rc);
2896 }
2897
2898 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2899                            const struct lu_fid *fid, ldlm_type_t type,
2900                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
2901                            struct lustre_handle *lockh)
2902 {
2903         struct obd_device       *obd = exp->exp_obd;
2904         struct lmv_obd          *lmv = &obd->u.lmv;
2905         ldlm_mode_t              rc;
2906         int                      i;
2907         ENTRY;
2908
2909         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2910
2911         /*
2912          * With CMD every object can have two locks in different namespaces:
2913          * lookup lock in space of mds storing direntry and update/open lock in
2914          * space of mds storing inode. Thus we check all targets, not only that
2915          * one fid was created in.
2916          */
2917         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2918                 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2919                                    type, policy, mode, lockh);
2920                 if (rc)
2921                         RETURN(rc);
2922         }
2923
2924         RETURN(0);
2925 }
2926
2927 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2928                       struct obd_export *dt_exp, struct obd_export *md_exp,
2929                       struct lustre_md *md)
2930 {
2931         struct obd_device       *obd = exp->exp_obd;
2932         struct lmv_obd          *lmv = &obd->u.lmv;
2933         int                      rc;
2934         ENTRY;
2935         rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2936         RETURN(rc);
2937 }
2938
2939 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2940 {
2941         struct obd_device       *obd = exp->exp_obd;
2942         struct lmv_obd          *lmv = &obd->u.lmv;
2943         ENTRY;
2944
2945         if (md->mea)
2946                 obd_free_memmd(exp, (void *)&md->mea);
2947         RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2948 }
2949
2950 int lmv_set_open_replay_data(struct obd_export *exp,
2951                              struct obd_client_handle *och,
2952                              struct ptlrpc_request *open_req)
2953 {
2954         struct obd_device       *obd = exp->exp_obd;
2955         struct lmv_obd          *lmv = &obd->u.lmv;
2956         struct lmv_tgt_desc     *tgt;
2957         ENTRY;
2958
2959         tgt = lmv_find_target(lmv, &och->och_fid);
2960         if (IS_ERR(tgt))
2961                 RETURN(PTR_ERR(tgt));
2962
2963         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2964 }
2965
2966 int lmv_clear_open_replay_data(struct obd_export *exp,
2967                                struct obd_client_handle *och)
2968 {
2969         struct obd_device       *obd = exp->exp_obd;
2970         struct lmv_obd          *lmv = &obd->u.lmv;
2971         struct lmv_tgt_desc     *tgt;
2972         ENTRY;
2973
2974         tgt = lmv_find_target(lmv, &och->och_fid);
2975         if (IS_ERR(tgt))
2976                 RETURN(PTR_ERR(tgt));
2977
2978         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2979 }
2980
2981 static int lmv_get_remote_perm(struct obd_export *exp,
2982                                const struct lu_fid *fid,
2983                                struct obd_capa *oc, __u32 suppgid,
2984                                struct ptlrpc_request **request)
2985 {
2986         struct obd_device       *obd = exp->exp_obd;
2987         struct lmv_obd          *lmv = &obd->u.lmv;
2988         struct lmv_tgt_desc     *tgt;
2989         int                      rc;
2990         ENTRY;
2991
2992         rc = lmv_check_connect(obd);
2993         if (rc)
2994                 RETURN(rc);
2995
2996         tgt = lmv_find_target(lmv, fid);
2997         if (IS_ERR(tgt))
2998                 RETURN(PTR_ERR(tgt));
2999
3000         rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
3001         RETURN(rc);
3002 }
3003
3004 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
3005                           renew_capa_cb_t cb)
3006 {
3007         struct obd_device       *obd = exp->exp_obd;
3008         struct lmv_obd          *lmv = &obd->u.lmv;
3009         struct lmv_tgt_desc     *tgt;
3010         int                      rc;
3011         ENTRY;
3012
3013         rc = lmv_check_connect(obd);
3014         if (rc)
3015                 RETURN(rc);
3016
3017         tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
3018         if (IS_ERR(tgt))
3019                 RETURN(PTR_ERR(tgt));
3020
3021         rc = md_renew_capa(tgt->ltd_exp, oc, cb);
3022         RETURN(rc);
3023 }
3024
3025 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
3026                     const struct req_msg_field *field, struct obd_capa **oc)
3027 {
3028         struct obd_device *obd = exp->exp_obd;
3029         struct lmv_obd *lmv = &obd->u.lmv;
3030         int rc;
3031
3032         ENTRY;
3033         rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
3034         RETURN(rc);
3035 }
3036
3037 int lmv_intent_getattr_async(struct obd_export *exp,
3038                              struct md_enqueue_info *minfo,
3039                              struct ldlm_enqueue_info *einfo)
3040 {
3041         struct md_op_data       *op_data = &minfo->mi_data;
3042         struct obd_device       *obd = exp->exp_obd;
3043         struct lmv_obd          *lmv = &obd->u.lmv;
3044         struct lmv_object       *obj;
3045         struct lmv_tgt_desc     *tgt = NULL;
3046         int                      rc;
3047         int                      sidx;
3048         ENTRY;
3049
3050         rc = lmv_check_connect(obd);
3051         if (rc)
3052                 RETURN(rc);
3053
3054         if (op_data->op_namelen) {
3055                 obj = lmv_object_find(obd, &op_data->op_fid1);
3056                 if (obj) {
3057                         sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
3058                                             (char *)op_data->op_name,
3059                                             op_data->op_namelen);
3060                         op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
3061                         tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
3062                         lmv_object_put(obj);
3063                 }
3064         }
3065
3066         if (tgt == NULL)
3067                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
3068
3069         if (IS_ERR(tgt))
3070                 RETURN(PTR_ERR(tgt));
3071
3072         rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
3073         RETURN(rc);
3074 }
3075
3076 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3077                         struct lu_fid *fid, __u64 *bits)
3078 {
3079         struct obd_device       *obd = exp->exp_obd;
3080         struct lmv_obd          *lmv = &obd->u.lmv;
3081         struct lmv_tgt_desc     *tgt;
3082         int                      rc;
3083         ENTRY;
3084
3085         rc = lmv_check_connect(obd);
3086         if (rc)
3087                 RETURN(rc);
3088
3089         tgt = lmv_find_target(lmv, fid);
3090         if (IS_ERR(tgt))
3091                 RETURN(PTR_ERR(tgt));
3092
3093         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3094         RETURN(rc);
3095 }
3096
3097 /**
3098  * For lmv, only need to send request to master MDT, and the master MDT will
3099  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3100  * we directly fetch data from the slave MDTs.
3101  */
3102 int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3103                  struct obd_quotactl *oqctl)
3104 {
3105         struct obd_device   *obd = class_exp2obd(exp);
3106         struct lmv_obd      *lmv = &obd->u.lmv;
3107         struct lmv_tgt_desc *tgt = &lmv->tgts[0];
3108         int                  rc = 0, i;
3109         __u64                curspace, curinodes;
3110         ENTRY;
3111
3112         if (!lmv->desc.ld_tgt_count || !tgt->ltd_active) {
3113                 CERROR("master lmv inactive\n");
3114                 RETURN(-EIO);
3115         }
3116
3117         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3118                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3119                 RETURN(rc);
3120         }
3121
3122         curspace = curinodes = 0;
3123         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3124                 int err;
3125                 tgt = &lmv->tgts[i];
3126
3127                 if (tgt->ltd_exp == NULL)
3128                         continue;
3129                 if (!tgt->ltd_active) {
3130                         CDEBUG(D_HA, "mdt %d is inactive.\n", i);
3131                         continue;
3132                 }
3133
3134                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3135                 if (err) {
3136                         CERROR("getquota on mdt %d failed. %d\n", i, err);
3137                         if (!rc)
3138                                 rc = err;
3139                 } else {
3140                         curspace += oqctl->qc_dqblk.dqb_curspace;
3141                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3142                 }
3143         }
3144         oqctl->qc_dqblk.dqb_curspace = curspace;
3145         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3146
3147         RETURN(rc);
3148 }
3149
3150 int lmv_quotacheck(struct obd_device *unused, struct obd_export *exp,
3151                    struct obd_quotactl *oqctl)
3152 {
3153         struct obd_device   *obd = class_exp2obd(exp);
3154         struct lmv_obd      *lmv = &obd->u.lmv;
3155         struct lmv_tgt_desc *tgt;
3156         int                  i, rc = 0;
3157         ENTRY;
3158
3159         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
3160                 int err;
3161
3162                 if (!tgt->ltd_active) {
3163                         CERROR("lmv idx %d inactive\n", i);
3164                         RETURN(-EIO);
3165                 }
3166
3167                 err = obd_quotacheck(tgt->ltd_exp, oqctl);
3168                 if (err && !rc)
3169                         rc = err;
3170         }
3171
3172         RETURN(rc);
3173 }
3174
3175 struct obd_ops lmv_obd_ops = {
3176         .o_owner                = THIS_MODULE,
3177         .o_setup                = lmv_setup,
3178         .o_cleanup              = lmv_cleanup,
3179         .o_precleanup           = lmv_precleanup,
3180         .o_process_config       = lmv_process_config,
3181         .o_connect              = lmv_connect,
3182         .o_disconnect           = lmv_disconnect,
3183         .o_statfs               = lmv_statfs,
3184         .o_get_info             = lmv_get_info,
3185         .o_set_info_async       = lmv_set_info_async,
3186         .o_packmd               = lmv_packmd,
3187         .o_unpackmd             = lmv_unpackmd,
3188         .o_notify               = lmv_notify,
3189         .o_get_uuid             = lmv_get_uuid,
3190         .o_iocontrol            = lmv_iocontrol,
3191         .o_fid_delete           = lmv_fid_delete,
3192         .o_quotacheck           = lmv_quotacheck,
3193         .o_quotactl             = lmv_quotactl
3194 };
3195
3196 struct md_ops lmv_md_ops = {
3197         .m_getstatus            = lmv_getstatus,
3198         .m_change_cbdata        = lmv_change_cbdata,
3199         .m_find_cbdata          = lmv_find_cbdata,
3200         .m_close                = lmv_close,
3201         .m_create               = lmv_create,
3202         .m_done_writing         = lmv_done_writing,
3203         .m_enqueue              = lmv_enqueue,
3204         .m_getattr              = lmv_getattr,
3205         .m_getxattr             = lmv_getxattr,
3206         .m_getattr_name         = lmv_getattr_name,
3207         .m_intent_lock          = lmv_intent_lock,
3208         .m_link                 = lmv_link,
3209         .m_rename               = lmv_rename,
3210         .m_setattr              = lmv_setattr,
3211         .m_setxattr             = lmv_setxattr,
3212         .m_sync                 = lmv_sync,
3213         .m_readpage             = lmv_readpage,
3214         .m_unlink               = lmv_unlink,
3215         .m_init_ea_size         = lmv_init_ea_size,
3216         .m_cancel_unused        = lmv_cancel_unused,
3217         .m_set_lock_data        = lmv_set_lock_data,
3218         .m_lock_match           = lmv_lock_match,
3219         .m_get_lustre_md        = lmv_get_lustre_md,
3220         .m_free_lustre_md       = lmv_free_lustre_md,
3221         .m_set_open_replay_data = lmv_set_open_replay_data,
3222         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3223         .m_renew_capa           = lmv_renew_capa,
3224         .m_unpack_capa          = lmv_unpack_capa,
3225         .m_get_remote_perm      = lmv_get_remote_perm,
3226         .m_intent_getattr_async = lmv_intent_getattr_async,
3227         .m_revalidate_lock      = lmv_revalidate_lock
3228 };
3229
3230 int __init lmv_init(void)
3231 {
3232         struct lprocfs_static_vars lvars;
3233         int                        rc;
3234
3235         lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3236                                                 sizeof(struct lmv_object),
3237                                                 0, 0);
3238         if (!lmv_object_cache) {
3239                 CERROR("Error allocating lmv objects cache\n");
3240                 return -ENOMEM;
3241         }
3242
3243         lprocfs_lmv_init_vars(&lvars);
3244
3245         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3246                                  lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3247         if (rc)
3248                 cfs_mem_cache_destroy(lmv_object_cache);
3249
3250         return rc;
3251 }
3252
3253 #ifdef __KERNEL__
3254 static void lmv_exit(void)
3255 {
3256         class_unregister_type(LUSTRE_LMV_NAME);
3257
3258         LASSERTF(cfs_atomic_read(&lmv_object_count) == 0,
3259                  "Can't free lmv objects cache, %d object(s) busy\n",
3260                  cfs_atomic_read(&lmv_object_count));
3261         cfs_mem_cache_destroy(lmv_object_cache);
3262 }
3263
3264 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3265 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3266 MODULE_LICENSE("GPL");
3267
3268 module_init(lmv_init);
3269 module_exit(lmv_exit);
3270 #endif