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 #define DEBUG_SUBSYSTEM S_LMV
  38 #ifdef __KERNEL__
  39 #include <linux/slab.h>
  40 #include <linux/module.h>
  41 #include <linux/init.h>
  42 #include <linux/slab.h>
  43 #include <linux/pagemap.h>
  44 #include <linux/mm.h>
  45 #include <asm/div64.h>
  46 #include <linux/seq_file.h>
  47 #include <linux/namei.h>
  48 #else
  49 #include <liblustre.h>
  50 #endif
  51
  52 #include <obd_support.h>
  53 #include <lustre_lib.h>
  54 #include <lustre_net.h>
  55 #include <obd_class.h>
  56 #include <lprocfs_status.h>
  57 #include <lustre_lite.h>
  58 #include <lustre_fid.h>
  59 #include "lmv_internal.h"
  60
  61 /* object cache. */
  62 cfs_mem_cache_t *lmv_object_cache;
  63 cfs_atomic_t lmv_object_count = CFS_ATOMIC_INIT(0);
  64
  65 static void lmv_activate_target(struct lmv_obd *lmv,
  66                                 struct lmv_tgt_desc *tgt,
  67                                 int activate)
  68 {
  69         if (tgt->ltd_active == activate)
  70                 return;
  71
  72         tgt->ltd_active = activate;
  73         lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
  74 }
  75
  76 /**
  77  * Error codes:
  78  *
  79  *  -EINVAL  : UUID can't be found in the LMV's target list
  80  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  81  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  82  */
  83 static int lmv_set_mdc_active(struct lmv_obd *lmv, struct obd_uuid *uuid,
  84                               int activate)
  85 {
  86         struct lmv_tgt_desc    *tgt;
  87         struct obd_device      *obd;
  88         int                     i;
  89         int                     rc = 0;
  90         ENTRY;
  91
  92         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  93                lmv, uuid->uuid, activate);
  94
  95         cfs_spin_lock(&lmv->lmv_lock);
  96         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
  97                 if (tgt->ltd_exp == NULL)
  98                         continue;
  99
 100                 CDEBUG(D_INFO, "Target idx %d is %s conn "LPX64"\n",
 101                        i, tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
 102
 103                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 104                         break;
 105         }
 106
 107         if (i == lmv->desc.ld_tgt_count)
 108                 GOTO(out_lmv_lock, rc = -EINVAL);
 109
 110         obd = class_exp2obd(tgt->ltd_exp);
 111         if (obd == NULL)
 112                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 113
 114         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 115                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 116                obd->obd_type->typ_name, i);
 117         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 118
 119         if (tgt->ltd_active == activate) {
 120                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 121                        activate ? "" : "in");
 122                 GOTO(out_lmv_lock, rc);
 123         }
 124
 125         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 126                activate ? "" : "in");
 127         lmv_activate_target(lmv, tgt, activate);
 128         EXIT;
 129
 130  out_lmv_lock:
 131         cfs_spin_unlock(&lmv->lmv_lock);
 132         return rc;
 133 }
 134
 135 static int lmv_set_mdc_data(struct lmv_obd *lmv, struct obd_uuid *uuid,
 136                             struct obd_connect_data *data)
 137 {
 138         struct lmv_tgt_desc    *tgt;
 139         int                     i;
 140         ENTRY;
 141
 142         LASSERT(data != NULL);
 143
 144         cfs_spin_lock(&lmv->lmv_lock);
 145         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 146                 if (tgt->ltd_exp == NULL)
 147                         continue;
 148
 149                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
 150                         lmv->datas[tgt->ltd_idx] = *data;
 151                         break;
 152                 }
 153         }
 154         cfs_spin_unlock(&lmv->lmv_lock);
 155         RETURN(0);
 156 }
 157
 158 struct obd_uuid *lmv_get_uuid(struct obd_export *exp) {
 159         struct obd_device *obd = exp->exp_obd;
 160         struct lmv_obd *lmv = &obd->u.lmv;
 161         return obd_get_uuid(lmv->tgts[0].ltd_exp);
 162 }
 163
 164 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 165                       enum obd_notify_event ev, void *data)
 166 {
 167         struct obd_connect_data *conn_data;
 168         struct lmv_obd          *lmv = &obd->u.lmv;
 169         struct obd_uuid         *uuid;
 170         int                      rc = 0;
 171         ENTRY;
 172
 173         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 174                 CERROR("unexpected notification of %s %s!\n",
 175                        watched->obd_type->typ_name,
 176                        watched->obd_name);
 177                 RETURN(-EINVAL);
 178         }
 179
 180         uuid = &watched->u.cli.cl_target_uuid;
 181         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 182                 /*
 183                  * Set MDC as active before notifying the observer, so the
 184                  * observer can use the MDC normally.
 185                  */
 186                 rc = lmv_set_mdc_active(lmv, uuid,
 187                                         ev == OBD_NOTIFY_ACTIVE);
 188                 if (rc) {
 189                         CERROR("%sactivation of %s failed: %d\n",
 190                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 191                                uuid->uuid, rc);
 192                         RETURN(rc);
 193                 }
 194         } else if (ev == OBD_NOTIFY_OCD) {
 195                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 196
 197                 /*
 198                  * Set connect data to desired target, update exp_connect_flags.
 199                  */
 200                 rc = lmv_set_mdc_data(lmv, uuid, conn_data);
 201                 if (rc) {
 202                         CERROR("can't set connect data to target %s, rc %d\n",
 203                                uuid->uuid, rc);
 204                         RETURN(rc);
 205                 }
 206
 207                 /*
 208                  * XXX: Make sure that ocd_connect_flags from all targets are
 209                  * the same. Otherwise one of MDTs runs wrong version or
 210                  * something like this.  --umka
 211                  */
 212                 obd->obd_self_export->exp_connect_flags =
 213                         conn_data->ocd_connect_flags;
 214         }
 215 #if 0
 216         else if (ev == OBD_NOTIFY_DISCON) {
 217                 /*
 218                  * For disconnect event, flush fld cache for failout MDS case.
 219                  */
 220                 fld_client_flush(&lmv->lmv_fld);
 221         }
 222 #endif
 223         /*
 224          * Pass the notification up the chain.
 225          */
 226         if (obd->obd_observer)
 227                 rc = obd_notify(obd->obd_observer, watched, ev, data);
 228
 229         RETURN(rc);
 230 }
 231
 232 /**
 233  * This is fake connect function. Its purpose is to initialize lmv and say
 234  * caller that everything is okay. Real connection will be performed later.
 235  */
 236 static int lmv_connect(const struct lu_env *env,
 237                        struct obd_export **exp, struct obd_device *obd,
 238                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 239                        void *localdata)
 240 {
 241 #ifdef __KERNEL__
 242         struct proc_dir_entry *lmv_proc_dir;
 243 #endif
 244         struct lmv_obd        *lmv = &obd->u.lmv;
 245         struct lustre_handle  conn = { 0 };
 246         int                    rc = 0;
 247         ENTRY;
 248
 249         /*
 250          * We don't want to actually do the underlying connections more than
 251          * once, so keep track.
 252          */
 253         lmv->refcount++;
 254         if (lmv->refcount > 1) {
 255                 *exp = NULL;
 256                 RETURN(0);
 257         }
 258
 259         rc = class_connect(&conn, obd, cluuid);
 260         if (rc) {
 261                 CERROR("class_connection() returned %d\n", rc);
 262                 RETURN(rc);
 263         }
 264
 265         *exp = class_conn2export(&conn);
 266         class_export_get(*exp);
 267
 268         lmv->exp = *exp;
 269         lmv->connected = 0;
 270         lmv->cluuid = *cluuid;
 271
 272         if (data)
 273                 lmv->conn_data = *data;
 274
 275 #ifdef __KERNEL__
 276         lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
 277                                         NULL, NULL);
 278         if (IS_ERR(lmv_proc_dir)) {
 279                 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
 280                        obd->obd_type->typ_name, obd->obd_name);
 281                 lmv_proc_dir = NULL;
 282         }
 283 #endif
 284
 285         /*
 286          * All real clients should perform actual connection right away, because
 287          * it is possible, that LMV will not have opportunity to connect targets
 288          * and MDC stuff will be called directly, for instance while reading
 289          * ../mdc/../kbytesfree procfs file, etc.
 290          */
 291         if (data->ocd_connect_flags & OBD_CONNECT_REAL)
 292                 rc = lmv_check_connect(obd);
 293
 294 #ifdef __KERNEL__
 295         if (rc) {
 296                 if (lmv_proc_dir)
 297                         lprocfs_remove(&lmv_proc_dir);
 298         }
 299 #endif
 300
 301         RETURN(rc);
 302 }
 303
 304 static void lmv_set_timeouts(struct obd_device *obd)
 305 {
 306         struct lmv_tgt_desc   *tgts;
 307         struct lmv_obd        *lmv;
 308         int                    i;
 309
 310         lmv = &obd->u.lmv;
 311         if (lmv->server_timeout == 0)
 312                 return;
 313
 314         if (lmv->connected == 0)
 315                 return;
 316
 317         for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
 318                 if (tgts->ltd_exp == NULL)
 319                         continue;
 320
 321                 obd_set_info_async(NULL, tgts->ltd_exp, sizeof(KEY_INTERMDS),
 322                                    KEY_INTERMDS, 0, NULL, NULL);
 323         }
 324 }
 325
 326 static int lmv_init_ea_size(struct obd_export *exp, int easize,
 327                             int def_easize, int cookiesize)
 328 {
 329         struct obd_device   *obd = exp->exp_obd;
 330         struct lmv_obd      *lmv = &obd->u.lmv;
 331         int                  i;
 332         int                  rc = 0;
 333         int                  change = 0;
 334         ENTRY;
 335
 336         if (lmv->max_easize < easize) {
 337                 lmv->max_easize = easize;
 338                 change = 1;
 339         }
 340         if (lmv->max_def_easize < def_easize) {
 341                 lmv->max_def_easize = def_easize;
 342                 change = 1;
 343         }
 344         if (lmv->max_cookiesize < cookiesize) {
 345                 lmv->max_cookiesize = cookiesize;
 346                 change = 1;
 347         }
 348         if (change == 0)
 349                 RETURN(0);
 350
 351         if (lmv->connected == 0)
 352                 RETURN(0);
 353
 354         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 355                 if (lmv->tgts[i].ltd_exp == NULL) {
 356                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
 357                         continue;
 358                 }
 359
 360                 rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize,
 361                                      cookiesize);
 362                 if (rc) {
 363                         CERROR("obd_init_ea_size() failed on MDT target %d, "
 364                                "error %d.\n", i, rc);
 365                         break;
 366                 }
 367         }
 368         RETURN(rc);
 369 }
 370
 371 #define MAX_STRING_SIZE 128
 372
 373 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 374 {
 375 #ifdef __KERNEL__
 376         struct proc_dir_entry   *lmv_proc_dir;
 377 #endif
 378         struct lmv_obd          *lmv = &obd->u.lmv;
 379         struct obd_uuid         *cluuid = &lmv->cluuid;
 380         struct obd_connect_data *mdc_data = NULL;
 381         struct obd_uuid          lmv_mdc_uuid = { "LMV_MDC_UUID" };
 382         struct obd_device       *mdc_obd;
 383         struct obd_export       *mdc_exp;
 384         struct lu_fld_target     target;
 385         int                      rc;
 386         ENTRY;
 387
 388         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 389                                         &obd->obd_uuid);
 390         if (!mdc_obd) {
 391                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 392                 RETURN(-EINVAL);
 393         }
 394
 395         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
 396                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 397                 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
 398                 cluuid->uuid);
 399
 400         if (!mdc_obd->obd_set_up) {
 401                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 402                 RETURN(-EINVAL);
 403         }
 404
 405         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &lmv_mdc_uuid,
 406                          &lmv->conn_data, NULL);
 407         if (rc) {
 408                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 409                 RETURN(rc);
 410         }
 411
 412         /*
 413          * Init fid sequence client for this mdc and add new fld target.
 414          */
 415         rc = obd_fid_init(mdc_exp);
 416         if (rc)
 417                 RETURN(rc);
 418
 419         target.ft_srv = NULL;
 420         target.ft_exp = mdc_exp;
 421         target.ft_idx = tgt->ltd_idx;
 422
 423         fld_client_add_target(&lmv->lmv_fld, &target);
 424
 425         mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data;
 426
 427         rc = obd_register_observer(mdc_obd, obd);
 428         if (rc) {
 429                 obd_disconnect(mdc_exp);
 430                 CERROR("target %s register_observer error %d\n",
 431                        tgt->ltd_uuid.uuid, rc);
 432                 RETURN(rc);
 433         }
 434
 435         if (obd->obd_observer) {
 436                 /*
 437                  * Tell the observer about the new target.
 438                  */
 439                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 440                                 OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts));
 441                 if (rc) {
 442                         obd_disconnect(mdc_exp);
 443                         RETURN(rc);
 444                 }
 445         }
 446
 447         tgt->ltd_active = 1;
 448         tgt->ltd_exp = mdc_exp;
 449         lmv->desc.ld_active_tgt_count++;
 450
 451         /*
 452          * Copy connect data, it may be used later.
 453          */
 454         lmv->datas[tgt->ltd_idx] = *mdc_data;
 455
 456         md_init_ea_size(tgt->ltd_exp, lmv->max_easize,
 457                         lmv->max_def_easize, lmv->max_cookiesize);
 458
 459         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 460                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 461                 cfs_atomic_read(&obd->obd_refcount));
 462
 463 #ifdef __KERNEL__
 464         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 465         if (lmv_proc_dir) {
 466                 struct proc_dir_entry *mdc_symlink;
 467
 468                 LASSERT(mdc_obd->obd_type != NULL);
 469                 LASSERT(mdc_obd->obd_type->typ_name != NULL);
 470                 mdc_symlink = lprocfs_add_symlink(mdc_obd->obd_name,
 471                                                   lmv_proc_dir,
 472                                                   "../../../%s/%s",
 473                                                   mdc_obd->obd_type->typ_name,
 474                                                   mdc_obd->obd_name);
 475                 if (mdc_symlink == NULL) {
 476                         CERROR("Could not register LMV target "
 477                                "/proc/fs/lustre/%s/%s/target_obds/%s.",
 478                                obd->obd_type->typ_name, obd->obd_name,
 479                                mdc_obd->obd_name);
 480                         lprocfs_remove(&lmv_proc_dir);
 481                         lmv_proc_dir = NULL;
 482                 }
 483         }
 484 #endif
 485         RETURN(0);
 486 }
 487
 488 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
 489 {
 490         struct lmv_obd      *lmv = &obd->u.lmv;
 491         struct lmv_tgt_desc *tgt;
 492         int                  rc = 0;
 493         ENTRY;
 494
 495         CDEBUG(D_CONFIG, "Target uuid: %s.\n", tgt_uuid->uuid);
 496
 497         lmv_init_lock(lmv);
 498
 499         if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
 500                 lmv_init_unlock(lmv);
 501                 CERROR("Can't add %s, LMV module compiled for %d MDCs. "
 502                        "That many MDCs already configured.\n",
 503                        tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
 504                 RETURN(-EINVAL);
 505         }
 506         if (lmv->desc.ld_tgt_count == 0) {
 507                 struct obd_device *mdc_obd;
 508
 509                 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
 510                                                 &obd->obd_uuid);
 511                 if (!mdc_obd) {
 512                         lmv_init_unlock(lmv);
 513                         CERROR("Target %s not attached\n", tgt_uuid->uuid);
 514                         RETURN(-EINVAL);
 515                 }
 516         }
 517         cfs_spin_lock(&lmv->lmv_lock);
 518         tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
 519         tgt->ltd_uuid = *tgt_uuid;
 520         cfs_spin_unlock(&lmv->lmv_lock);
 521
 522         if (lmv->connected) {
 523                 rc = lmv_connect_mdc(obd, tgt);
 524                 if (rc) {
 525                         cfs_spin_lock(&lmv->lmv_lock);
 526                         lmv->desc.ld_tgt_count--;
 527                         memset(tgt, 0, sizeof(*tgt));
 528                         cfs_spin_unlock(&lmv->lmv_lock);
 529                 } else {
 530                         int easize = sizeof(struct lmv_stripe_md) +
 531                                      lmv->desc.ld_tgt_count *
 532                                      sizeof(struct lu_fid);
 533                         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 534                 }
 535         }
 536
 537         lmv_init_unlock(lmv);
 538         RETURN(rc);
 539 }
 540
 541 int lmv_check_connect(struct obd_device *obd)
 542 {
 543         struct lmv_obd       *lmv = &obd->u.lmv;
 544         struct lmv_tgt_desc  *tgt;
 545         int                   i;
 546         int                   rc;
 547         int                   easize;
 548         ENTRY;
 549
 550         if (lmv->connected)
 551                 RETURN(0);
 552
 553         lmv_init_lock(lmv);
 554         if (lmv->connected) {
 555                 lmv_init_unlock(lmv);
 556                 RETURN(0);
 557         }
 558
 559         if (lmv->desc.ld_tgt_count == 0) {
 560                 lmv_init_unlock(lmv);
 561                 CERROR("%s: no targets configured.\n", obd->obd_name);
 562                 RETURN(-EINVAL);
 563         }
 564
 565         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 566                lmv->cluuid.uuid, obd->obd_name);
 567
 568         LASSERT(lmv->tgts != NULL);
 569
 570         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 571                 rc = lmv_connect_mdc(obd, tgt);
 572                 if (rc)
 573                         GOTO(out_disc, rc);
 574         }
 575
 576         lmv_set_timeouts(obd);
 577         class_export_put(lmv->exp);
 578         lmv->connected = 1;
 579         easize = lmv_get_easize(lmv);
 580         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 581         lmv_init_unlock(lmv);
 582         RETURN(0);
 583
 584  out_disc:
 585         while (i-- > 0) {
 586                 int rc2;
 587                 --tgt;
 588                 tgt->ltd_active = 0;
 589                 if (tgt->ltd_exp) {
 590                         --lmv->desc.ld_active_tgt_count;
 591                         rc2 = obd_disconnect(tgt->ltd_exp);
 592                         if (rc2) {
 593                                 CERROR("LMV target %s disconnect on "
 594                                        "MDC idx %d: error %d\n",
 595                                        tgt->ltd_uuid.uuid, i, rc2);
 596                         }
 597                 }
 598         }
 599         class_disconnect(lmv->exp);
 600         lmv_init_unlock(lmv);
 601         RETURN(rc);
 602 }
 603
 604 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 605 {
 606 #ifdef __KERNEL__
 607         struct proc_dir_entry  *lmv_proc_dir;
 608 #endif
 609         struct lmv_obd         *lmv = &obd->u.lmv;
 610         struct obd_device      *mdc_obd;
 611         int                     rc;
 612         ENTRY;
 613
 614         LASSERT(tgt != NULL);
 615         LASSERT(obd != NULL);
 616
 617         mdc_obd = class_exp2obd(tgt->ltd_exp);
 618
 619         if (mdc_obd) {
 620                 mdc_obd->obd_force = obd->obd_force;
 621                 mdc_obd->obd_fail = obd->obd_fail;
 622                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 623         }
 624
 625 #ifdef __KERNEL__
 626         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 627         if (lmv_proc_dir) {
 628                 struct proc_dir_entry *mdc_symlink;
 629
 630                 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
 631                 if (mdc_symlink) {
 632                         lprocfs_remove(&mdc_symlink);
 633                 } else {
 634                         CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
 635                                obd->obd_type->typ_name, obd->obd_name,
 636                                mdc_obd->obd_name);
 637                 }
 638         }
 639 #endif
 640         rc = obd_fid_fini(tgt->ltd_exp);
 641         if (rc)
 642                 CERROR("Can't finanize fids factory\n");
 643
 644         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 645                tgt->ltd_exp->exp_obd->obd_name,
 646                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 647
 648         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 649         rc = obd_disconnect(tgt->ltd_exp);
 650         if (rc) {
 651                 if (tgt->ltd_active) {
 652                         CERROR("Target %s disconnect error %d\n",
 653                                tgt->ltd_uuid.uuid, rc);
 654                 }
 655         }
 656
 657         lmv_activate_target(lmv, tgt, 0);
 658         tgt->ltd_exp = NULL;
 659         RETURN(0);
 660 }
 661
 662 static int lmv_disconnect(struct obd_export *exp)
 663 {
 664         struct obd_device     *obd = class_exp2obd(exp);
 665 #ifdef __KERNEL__
 666         struct proc_dir_entry *lmv_proc_dir;
 667 #endif
 668         struct lmv_obd        *lmv = &obd->u.lmv;
 669         int                    rc;
 670         int                    i;
 671         ENTRY;
 672
 673         if (!lmv->tgts)
 674                 goto out_local;
 675
 676         /*
 677          * Only disconnect the underlying layers on the final disconnect.
 678          */
 679         lmv->refcount--;
 680         if (lmv->refcount != 0)
 681                 goto out_local;
 682
 683         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 684                 if (lmv->tgts[i].ltd_exp == NULL)
 685                         continue;
 686                 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
 687         }
 688
 689 #ifdef __KERNEL__
 690         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 691         if (lmv_proc_dir) {
 692                 lprocfs_remove(&lmv_proc_dir);
 693         } else {
 694                 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
 695                        obd->obd_type->typ_name, obd->obd_name);
 696         }
 697 #endif
 698
 699 out_local:
 700         /*
 701          * This is the case when no real connection is established by
 702          * lmv_check_connect().
 703          */
 704         if (!lmv->connected)
 705                 class_export_put(exp);
 706         rc = class_disconnect(exp);
 707         if (lmv->refcount == 0)
 708                 lmv->connected = 0;
 709         RETURN(rc);
 710 }
 711
 712 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 713                          int len, void *karg, void *uarg)
 714 {
 715         struct obd_device    *obddev = class_exp2obd(exp);
 716         struct lmv_obd       *lmv = &obddev->u.lmv;
 717         int                   i = 0;
 718         int                   rc = 0;
 719         int                   set = 0;
 720         int                   count = lmv->desc.ld_tgt_count;
 721         ENTRY;
 722
 723         if (count == 0)
 724                 RETURN(-ENOTTY);
 725
 726         switch (cmd) {
 727         case IOC_OBD_STATFS: {
 728                 struct obd_ioctl_data *data = karg;
 729                 struct obd_device *mdc_obd;
 730                 struct obd_statfs stat_buf = {0};
 731                 __u32 index;
 732
 733                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 734                 if ((index >= count))
 735                         RETURN(-ENODEV);
 736
 737                 if (!lmv->tgts[index].ltd_active)
 738                         RETURN(-ENODATA);
 739
 740                 mdc_obd = class_exp2obd(lmv->tgts[index].ltd_exp);
 741                 if (!mdc_obd)
 742                         RETURN(-EINVAL);
 743
 744                 /* copy UUID */
 745                 if (cfs_copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 746                                      min((int) data->ioc_plen2,
 747                                          (int) sizeof(struct obd_uuid))))
 748                         RETURN(-EFAULT);
 749
 750                 rc = obd_statfs(NULL, lmv->tgts[index].ltd_exp, &stat_buf,
 751                                 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
 752                                 0);
 753                 if (rc)
 754                         RETURN(rc);
 755                 if (cfs_copy_to_user(data->ioc_pbuf1, &stat_buf,
 756                                      min((int) data->ioc_plen1,
 757                                          (int) sizeof(stat_buf))))
 758                         RETURN(-EFAULT);
 759                 break;
 760         }
 761         case OBD_IOC_QUOTACTL: {
 762                 struct if_quotactl *qctl = karg;
 763                 struct lmv_tgt_desc *tgt = NULL;
 764                 struct obd_quotactl *oqctl;
 765
 766                 if (qctl->qc_valid == QC_MDTIDX) {
 767                         if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
 768                                 RETURN(-EINVAL);
 769
 770                         tgt = &lmv->tgts[qctl->qc_idx];
 771                         if (!tgt->ltd_exp)
 772                                 RETURN(-EINVAL);
 773                 } else if (qctl->qc_valid == QC_UUID) {
 774                         for (i = 0; i < count; i++) {
 775                                 tgt = &lmv->tgts[i];
 776                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 777                                                      &qctl->obd_uuid))
 778                                         continue;
 779
 780                                 if (tgt->ltd_exp == NULL)
 781                                         RETURN(-EINVAL);
 782
 783                                 break;
 784                         }
 785                 } else {
 786                         RETURN(-EINVAL);
 787                 }
 788
 789                 if (i >= count)
 790                         RETURN(-EAGAIN);
 791
 792                 LASSERT(tgt && tgt->ltd_exp);
 793                 OBD_ALLOC_PTR(oqctl);
 794                 if (!oqctl)
 795                         RETURN(-ENOMEM);
 796
 797                 QCTL_COPY(oqctl, qctl);
 798                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 799                 if (rc == 0) {
 800                         QCTL_COPY(qctl, oqctl);
 801                         qctl->qc_valid = QC_MDTIDX;
 802                         qctl->obd_uuid = tgt->ltd_uuid;
 803                 }
 804                 OBD_FREE_PTR(oqctl);
 805                 break;
 806         }
 807         case OBD_IOC_CHANGELOG_SEND:
 808         case OBD_IOC_CHANGELOG_CLEAR: {
 809                 struct ioc_changelog *icc = karg;
 810
 811                 if (icc->icc_mdtindex >= count)
 812                         RETURN(-ENODEV);
 813
 814                 rc = obd_iocontrol(cmd, lmv->tgts[icc->icc_mdtindex].ltd_exp,
 815                                    sizeof(*icc), icc, NULL);
 816                 break;
 817         }
 818         case LL_IOC_GET_CONNECT_FLAGS: {
 819                 rc = obd_iocontrol(cmd, lmv->tgts[0].ltd_exp, len, karg, uarg);
 820                 break;
 821         }
 822
 823         default : {
 824                 for (i = 0; i < count; i++) {
 825                         int err;
 826                         struct obd_device *mdc_obd;
 827
 828                         if (lmv->tgts[i].ltd_exp == NULL)
 829                                 continue;
 830                         /* ll_umount_begin() sets force flag but for lmv, not
 831                          * mdc. Let's pass it through */
 832                         mdc_obd = class_exp2obd(lmv->tgts[i].ltd_exp);
 833                         mdc_obd->obd_force = obddev->obd_force;
 834                         err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
 835                                             karg, uarg);
 836                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
 837                                 RETURN(err);
 838                         } else if (err) {
 839                                 if (lmv->tgts[i].ltd_active) {
 840                                         CERROR("error: iocontrol MDC %s on MDT"
 841                                                "idx %d cmd %x: err = %d\n",
 842                                                 lmv->tgts[i].ltd_uuid.uuid,
 843                                                 i, cmd, err);
 844                                         if (!rc)
 845                                                 rc = err;
 846                                 }
 847                         } else
 848                                 set = 1;
 849                 }
 850                 if (!set && !rc)
 851                         rc = -EIO;
 852         }
 853         }
 854         RETURN(rc);
 855 }
 856
 857 static int lmv_all_chars_policy(int count, const char *name,
 858                                 int len)
 859 {
 860         unsigned int c = 0;
 861
 862         while (len > 0)
 863                 c += name[--len];
 864         c = c % count;
 865         return c;
 866 }
 867
 868 static int lmv_nid_policy(struct lmv_obd *lmv)
 869 {
 870         struct obd_import *imp;
 871         __u32              id;
 872
 873         /*
 874          * XXX: To get nid we assume that underlying obd device is mdc.
 875          */
 876         imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
 877         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
 878         return id % lmv->desc.ld_tgt_count;
 879 }
 880
 881 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
 882                           placement_policy_t placement)
 883 {
 884         switch (placement) {
 885         case PLACEMENT_CHAR_POLICY:
 886                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
 887                                             op_data->op_name,
 888                                             op_data->op_namelen);
 889         case PLACEMENT_NID_POLICY:
 890                 return lmv_nid_policy(lmv);
 891
 892         default:
 893                 break;
 894         }
 895
 896         CERROR("Unsupported placement policy %x\n", placement);
 897         return -EINVAL;
 898 }
 899
 900 /**
 901  * This is _inode_ placement policy function (not name).
 902  */
 903 static int lmv_placement_policy(struct obd_device *obd,
 904                                 struct md_op_data *op_data,
 905                                 mdsno_t *mds)
 906 {
 907         struct lmv_obd          *lmv = &obd->u.lmv;
 908         struct lmv_object       *obj;
 909         int                      rc;
 910         ENTRY;
 911
 912         LASSERT(mds != NULL);
 913
 914         if (lmv->desc.ld_tgt_count == 1) {
 915                 *mds = 0;
 916                 RETURN(0);
 917         }
 918
 919         /*
 920          * Allocate new fid on target according to operation type and parent
 921          * home mds.
 922          */
 923         obj = lmv_object_find(obd, &op_data->op_fid1);
 924         if (obj != NULL || op_data->op_name == NULL ||
 925             op_data->op_opc != LUSTRE_OPC_MKDIR) {
 926                 /*
 927                  * Allocate fid for non-dir or for null name or for case parent
 928                  * dir is split.
 929                  */
 930                 if (obj) {
 931                         lmv_object_put(obj);
 932
 933                         /*
 934                          * If we have this flag turned on, and we see that
 935                          * parent dir is split, this means, that caller did not
 936                          * notice split yet. This is race and we would like to
 937                          * let caller know that.
 938                          */
 939                         if (op_data->op_bias & MDS_CHECK_SPLIT)
 940                                 RETURN(-ERESTART);
 941                 }
 942
 943                 /*
 944                  * Allocate new fid on same mds where parent fid is located and
 945                  * where operation will be sent. In case of split dir, ->op_fid1
 946                  * and ->op_mds here will contain fid and mds of slave directory
 947                  * object (assigned by caller).
 948                  */
 949                 *mds = op_data->op_mds;
 950                 rc = 0;
 951         } else {
 952                 /*
 953                  * Parent directory is not split and we want to create a
 954                  * directory in it. Let's calculate where to place it according
 955                  * to operation data @op_data.
 956                  */
 957                 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
 958                 rc = 0;
 959         }
 960
 961         if (rc) {
 962                 CERROR("Can't choose MDS, err = %d\n", rc);
 963         } else {
 964                 LASSERT(*mds < lmv->desc.ld_tgt_count);
 965         }
 966
 967         RETURN(rc);
 968 }
 969
 970 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
 971                     mdsno_t mds)
 972 {
 973         struct lmv_tgt_desc *tgt;
 974         int                  rc;
 975         ENTRY;
 976
 977         tgt = lmv_get_target(lmv, mds);
 978
 979         /*
 980          * New seq alloc and FLD setup should be atomic. Otherwise we may find
 981          * on server that seq in new allocated fid is not yet known.
 982          */
 983         cfs_mutex_lock(&tgt->ltd_fid_mutex);
 984
 985         if (!tgt->ltd_active)
 986                 GOTO(out, rc = -ENODEV);
 987
 988         /*
 989          * Asking underlaying tgt layer to allocate new fid.
 990          */
 991         rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
 992         if (rc > 0) {
 993                 LASSERT(fid_is_sane(fid));
 994                 rc = 0;
 995         }
 996
 997         EXIT;
 998 out:
 999         cfs_mutex_unlock(&tgt->ltd_fid_mutex);
1000         return rc;
1001 }
1002
1003 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
1004                   struct md_op_data *op_data)
1005 {
1006         struct obd_device     *obd = class_exp2obd(exp);
1007         struct lmv_obd        *lmv = &obd->u.lmv;
1008         mdsno_t                mds = 0;
1009         int                    rc;
1010         ENTRY;
1011
1012         LASSERT(op_data != NULL);
1013         LASSERT(fid != NULL);
1014
1015         rc = lmv_placement_policy(obd, op_data, &mds);
1016         if (rc) {
1017                 CERROR("Can't get target for allocating fid, "
1018                        "rc %d\n", rc);
1019                 RETURN(rc);
1020         }
1021
1022         rc = __lmv_fid_alloc(lmv, fid, mds);
1023         if (rc) {
1024                 CERROR("Can't alloc new fid, rc %d\n", rc);
1025                 RETURN(rc);
1026         }
1027
1028         RETURN(rc);
1029 }
1030
1031 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1032 {
1033         ENTRY;
1034         LASSERT(exp != NULL && fid != NULL);
1035         if (lmv_object_delete(exp, fid)) {
1036                 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1037                        PFID(fid));
1038         }
1039         RETURN(0);
1040 }
1041
1042 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1043 {
1044         struct lmv_obd             *lmv = &obd->u.lmv;
1045         struct lprocfs_static_vars  lvars;
1046         struct lmv_desc            *desc;
1047         int                         rc;
1048         int                         i = 0;
1049         ENTRY;
1050
1051         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1052                 CERROR("LMV setup requires a descriptor\n");
1053                 RETURN(-EINVAL);
1054         }
1055
1056         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1057         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1058                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1059                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1060                 RETURN(-EINVAL);
1061         }
1062
1063         lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1064
1065         OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1066         if (lmv->tgts == NULL)
1067                 RETURN(-ENOMEM);
1068
1069         for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1070                 cfs_mutex_init(&lmv->tgts[i].ltd_fid_mutex);
1071                 lmv->tgts[i].ltd_idx = i;
1072         }
1073
1074         lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1075
1076         OBD_ALLOC(lmv->datas, lmv->datas_size);
1077         if (lmv->datas == NULL)
1078                 GOTO(out_free_tgts, rc = -ENOMEM);
1079
1080         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1081         lmv->desc.ld_tgt_count = 0;
1082         lmv->desc.ld_active_tgt_count = 0;
1083         lmv->max_cookiesize = 0;
1084         lmv->max_def_easize = 0;
1085         lmv->max_easize = 0;
1086         lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1087
1088         cfs_spin_lock_init(&lmv->lmv_lock);
1089         cfs_mutex_init(&lmv->init_mutex);
1090
1091         rc = lmv_object_setup(obd);
1092         if (rc) {
1093                 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1094                 GOTO(out_free_datas, rc);
1095         }
1096
1097         lprocfs_lmv_init_vars(&lvars);
1098         lprocfs_obd_setup(obd, lvars.obd_vars);
1099 #ifdef LPROCFS
1100         {
1101                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd",
1102                                         0444, &lmv_proc_target_fops, obd);
1103                 if (rc)
1104                         CWARN("%s: error adding LMV target_obd file: rc = %d\n",
1105                                obd->obd_name, rc);
1106        }
1107 #endif
1108         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1109                              LUSTRE_CLI_FLD_HASH_DHT);
1110         if (rc) {
1111                 CERROR("Can't init FLD, err %d\n", rc);
1112                 GOTO(out_free_datas, rc);
1113         }
1114
1115         RETURN(0);
1116
1117 out_free_datas:
1118         OBD_FREE(lmv->datas, lmv->datas_size);
1119         lmv->datas = NULL;
1120 out_free_tgts:
1121         OBD_FREE(lmv->tgts, lmv->tgts_size);
1122         lmv->tgts = NULL;
1123         return rc;
1124 }
1125
1126 static int lmv_cleanup(struct obd_device *obd)
1127 {
1128         struct lmv_obd   *lmv = &obd->u.lmv;
1129         ENTRY;
1130
1131         fld_client_fini(&lmv->lmv_fld);
1132         lmv_object_cleanup(obd);
1133         OBD_FREE(lmv->datas, lmv->datas_size);
1134         OBD_FREE(lmv->tgts, lmv->tgts_size);
1135
1136         RETURN(0);
1137 }
1138
1139 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1140 {
1141         struct lustre_cfg     *lcfg = buf;
1142         struct obd_uuid        tgt_uuid;
1143         int                    rc;
1144         ENTRY;
1145
1146         switch(lcfg->lcfg_command) {
1147         case LCFG_ADD_MDC:
1148                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1149                         GOTO(out, rc = -EINVAL);
1150
1151                 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1152                 rc = lmv_add_target(obd, &tgt_uuid);
1153                 GOTO(out, rc);
1154         default: {
1155                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1156                 GOTO(out, rc = -EINVAL);
1157         }
1158         }
1159 out:
1160         RETURN(rc);
1161 }
1162
1163 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1164                       struct obd_statfs *osfs, __u64 max_age, __u32 flags)
1165 {
1166         struct obd_device     *obd = class_exp2obd(exp);
1167         struct lmv_obd        *lmv = &obd->u.lmv;
1168         struct obd_statfs     *temp;
1169         int                    rc = 0;
1170         int                    i;
1171         ENTRY;
1172
1173         rc = lmv_check_connect(obd);
1174         if (rc)
1175                 RETURN(rc);
1176
1177         OBD_ALLOC(temp, sizeof(*temp));
1178         if (temp == NULL)
1179                 RETURN(-ENOMEM);
1180
1181         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1182                 if (lmv->tgts[i].ltd_exp == NULL)
1183                         continue;
1184
1185                 rc = obd_statfs(env, lmv->tgts[i].ltd_exp, temp,
1186                                 max_age, flags);
1187                 if (rc) {
1188                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1189                                lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1190                                rc);
1191                         GOTO(out_free_temp, rc);
1192                 }
1193                 if (i == 0) {
1194                         *osfs = *temp;
1195                 } else {
1196                         osfs->os_bavail += temp->os_bavail;
1197                         osfs->os_blocks += temp->os_blocks;
1198                         osfs->os_ffree += temp->os_ffree;
1199                         osfs->os_files += temp->os_files;
1200                 }
1201         }
1202
1203         EXIT;
1204 out_free_temp:
1205         OBD_FREE(temp, sizeof(*temp));
1206         return rc;
1207 }
1208
1209 static int lmv_getstatus(struct obd_export *exp,
1210                          struct lu_fid *fid,
1211                          struct obd_capa **pc)
1212 {
1213         struct obd_device    *obd = exp->exp_obd;
1214         struct lmv_obd       *lmv = &obd->u.lmv;
1215         int                   rc;
1216         ENTRY;
1217
1218         rc = lmv_check_connect(obd);
1219         if (rc)
1220                 RETURN(rc);
1221
1222         rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1223         RETURN(rc);
1224 }
1225
1226 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1227                         struct obd_capa *oc, obd_valid valid, const char *name,
1228                         const char *input, int input_size, int output_size,
1229                         int flags, struct ptlrpc_request **request)
1230 {
1231         struct obd_device      *obd = exp->exp_obd;
1232         struct lmv_obd         *lmv = &obd->u.lmv;
1233         struct lmv_tgt_desc    *tgt;
1234         int                     rc;
1235         ENTRY;
1236
1237         rc = lmv_check_connect(obd);
1238         if (rc)
1239                 RETURN(rc);
1240
1241         tgt = lmv_find_target(lmv, fid);
1242         if (IS_ERR(tgt))
1243                 RETURN(PTR_ERR(tgt));
1244
1245         rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1246                          input_size, output_size, flags, request);
1247
1248         RETURN(rc);
1249 }
1250
1251 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1252                         struct obd_capa *oc, obd_valid valid, const char *name,
1253                         const char *input, int input_size, int output_size,
1254                         int flags, __u32 suppgid,
1255                         struct ptlrpc_request **request)
1256 {
1257         struct obd_device      *obd = exp->exp_obd;
1258         struct lmv_obd         *lmv = &obd->u.lmv;
1259         struct lmv_tgt_desc    *tgt;
1260         int                     rc;
1261         ENTRY;
1262
1263         rc = lmv_check_connect(obd);
1264         if (rc)
1265                 RETURN(rc);
1266
1267         tgt = lmv_find_target(lmv, fid);
1268         if (IS_ERR(tgt))
1269                 RETURN(PTR_ERR(tgt));
1270
1271         rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1272                          input_size, output_size, flags, suppgid,
1273                          request);
1274
1275         RETURN(rc);
1276 }
1277
1278 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1279                        struct ptlrpc_request **request)
1280 {
1281         struct obd_device       *obd = exp->exp_obd;
1282         struct lmv_obd          *lmv = &obd->u.lmv;
1283         struct lmv_tgt_desc     *tgt;
1284         struct lmv_object       *obj;
1285         int                      rc;
1286         int                      i;
1287         ENTRY;
1288
1289         rc = lmv_check_connect(obd);
1290         if (rc)
1291                 RETURN(rc);
1292
1293         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1294         if (IS_ERR(tgt))
1295                 RETURN(PTR_ERR(tgt));
1296
1297         if (op_data->op_valid & OBD_MD_MDTIDX) {
1298                 op_data->op_mds = tgt->ltd_idx;
1299                 RETURN(0);
1300         }
1301
1302         rc = md_getattr(tgt->ltd_exp, op_data, request);
1303         if (rc)
1304                 RETURN(rc);
1305
1306         obj = lmv_object_find_lock(obd, &op_data->op_fid1);
1307
1308         CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(&op_data->op_fid1),
1309                obj ? "(split)" : "");
1310
1311         /*
1312          * If object is split, then we loop over all the slaves and gather size
1313          * attribute. In ideal world we would have to gather also mds field from
1314          * all slaves, as object is spread over the cluster and this is
1315          * definitely interesting information and it is not good to loss it,
1316          * but...
1317          */
1318         if (obj) {
1319                 struct mdt_body *body;
1320
1321                 if (*request == NULL) {
1322                         lmv_object_put(obj);
1323                         RETURN(rc);
1324                 }
1325
1326                 body = req_capsule_server_get(&(*request)->rq_pill,
1327                                               &RMF_MDT_BODY);
1328                 LASSERT(body != NULL);
1329
1330                 for (i = 0; i < obj->lo_objcount; i++) {
1331                         if (lmv->tgts[i].ltd_exp == NULL) {
1332                                 CWARN("%s: NULL export for %d\n",
1333                                       obd->obd_name, i);
1334                                 continue;
1335                         }
1336
1337                         /*
1338                          * Skip master object.
1339                          */
1340                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1341                                 continue;
1342
1343                         body->size += obj->lo_stripes[i].ls_size;
1344                 }
1345
1346                 lmv_object_put_unlock(obj);
1347         }
1348
1349         RETURN(rc);
1350 }
1351
1352 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1353                              ldlm_iterator_t it, void *data)
1354 {
1355         struct obd_device   *obd = exp->exp_obd;
1356         struct lmv_obd      *lmv = &obd->u.lmv;
1357         int                  i;
1358         int                  rc;
1359         ENTRY;
1360
1361         rc = lmv_check_connect(obd);
1362         if (rc)
1363                 RETURN(rc);
1364
1365         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1366
1367         /*
1368          * With CMD every object can have two locks in different namespaces:
1369          * lookup lock in space of mds storing direntry and update/open lock in
1370          * space of mds storing inode.
1371          */
1372         for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1373                 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1374
1375         RETURN(0);
1376 }
1377
1378 static int lmv_find_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1379                            ldlm_iterator_t it, void *data)
1380 {
1381         struct obd_device   *obd = exp->exp_obd;
1382         struct lmv_obd      *lmv = &obd->u.lmv;
1383         int                  i;
1384         int                  rc;
1385         ENTRY;
1386
1387         rc = lmv_check_connect(obd);
1388         if (rc)
1389                 RETURN(rc);
1390
1391         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1392
1393         /*
1394          * With CMD every object can have two locks in different namespaces:
1395          * lookup lock in space of mds storing direntry and update/open lock in
1396          * space of mds storing inode.
1397          */
1398         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1399                 rc = md_find_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1400                 if (rc)
1401                         RETURN(rc);
1402         }
1403
1404         RETURN(rc);
1405 }
1406
1407
1408 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1409                      struct md_open_data *mod, struct ptlrpc_request **request)
1410 {
1411         struct obd_device     *obd = exp->exp_obd;
1412         struct lmv_obd        *lmv = &obd->u.lmv;
1413         struct lmv_tgt_desc   *tgt;
1414         int                    rc;
1415         ENTRY;
1416
1417         rc = lmv_check_connect(obd);
1418         if (rc)
1419                 RETURN(rc);
1420
1421         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1422         if (IS_ERR(tgt))
1423                 RETURN(PTR_ERR(tgt));
1424
1425         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1426         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1427         RETURN(rc);
1428 }
1429
1430 /**
1431  * Called in the case MDS returns -ERESTART on create on open, what means that
1432  * directory is split and its LMV presentation object has to be updated.
1433  */
1434 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1435 {
1436         struct obd_device       *obd = exp->exp_obd;
1437         struct lmv_obd          *lmv = &obd->u.lmv;
1438         struct ptlrpc_request   *req = NULL;
1439         struct lmv_tgt_desc     *tgt;
1440         struct lmv_object       *obj;
1441         struct lustre_md         md;
1442         struct md_op_data       *op_data;
1443         int                      mealen;
1444         int                      rc;
1445         __u64                    valid;
1446         ENTRY;
1447
1448         md.mea = NULL;
1449         mealen = lmv_get_easize(lmv);
1450
1451         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1452
1453         tgt = lmv_find_target(lmv, fid);
1454         if (IS_ERR(tgt))
1455                 RETURN(PTR_ERR(tgt));
1456
1457         /*
1458          * Time to update mea of parent fid.
1459          */
1460
1461         OBD_ALLOC_PTR(op_data);
1462         if (op_data == NULL)
1463                 RETURN(-ENOMEM);
1464
1465         op_data->op_fid1 = *fid;
1466         op_data->op_mode = mealen;
1467         op_data->op_valid = valid;
1468
1469         rc = md_getattr(tgt->ltd_exp, op_data, &req);
1470         OBD_FREE_PTR(op_data);
1471         if (rc) {
1472                 CERROR("md_getattr() failed, error %d\n", rc);
1473                 GOTO(cleanup, rc);
1474         }
1475
1476         rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1477         if (rc) {
1478                 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1479                 GOTO(cleanup, rc);
1480         }
1481
1482         if (md.mea == NULL)
1483                 GOTO(cleanup, rc = -ENODATA);
1484
1485         obj = lmv_object_create(exp, fid, md.mea);
1486         if (IS_ERR(obj))
1487                 rc = PTR_ERR(obj);
1488         else
1489                 lmv_object_put(obj);
1490
1491         obd_free_memmd(exp, (void *)&md.mea);
1492         EXIT;
1493 cleanup:
1494         if (req)
1495                 ptlrpc_req_finished(req);
1496         return rc;
1497 }
1498
1499 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1500                const void *data, int datalen, int mode, __u32 uid,
1501                __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1502                struct ptlrpc_request **request)
1503 {
1504         struct obd_device       *obd = exp->exp_obd;
1505         struct lmv_obd          *lmv = &obd->u.lmv;
1506         struct lmv_tgt_desc     *tgt;
1507         struct lmv_object       *obj;
1508         int                      rc;
1509         int                      loop = 0;
1510         int                      sidx;
1511         ENTRY;
1512
1513         rc = lmv_check_connect(obd);
1514         if (rc)
1515                 RETURN(rc);
1516
1517         if (!lmv->desc.ld_active_tgt_count)
1518                 RETURN(-EIO);
1519 repeat:
1520         ++loop;
1521         LASSERT(loop <= 2);
1522
1523         obj = lmv_object_find(obd, &op_data->op_fid1);
1524         if (obj) {
1525                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1526                                        op_data->op_name, op_data->op_namelen);
1527                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1528                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1529                 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1530                 tgt = lmv_get_target(lmv, op_data->op_mds);
1531                 lmv_object_put(obj);
1532         } else {
1533                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1534                 op_data->op_bias |= MDS_CHECK_SPLIT;
1535                 op_data->op_mds = tgt->ltd_idx;
1536         }
1537
1538         if (IS_ERR(tgt))
1539                 RETURN(PTR_ERR(tgt));
1540
1541         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1542         if (rc == -ERESTART)
1543                 goto repeat;
1544         else if (rc)
1545                 RETURN(rc);
1546
1547         CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #%x\n",
1548                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1549                op_data->op_mds);
1550
1551         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1552         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1553                        cap_effective, rdev, request);
1554         if (rc == 0) {
1555                 if (*request == NULL)
1556                         RETURN(rc);
1557                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1558         } else if (rc == -ERESTART) {
1559                 LASSERT(*request != NULL);
1560                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1561                           "Got -ERESTART during create!\n");
1562                 ptlrpc_req_finished(*request);
1563                 *request = NULL;
1564
1565                 /*
1566                  * Directory got split. Time to update local object and repeat
1567                  * the request with proper MDS.
1568                  */
1569                 rc = lmv_handle_split(exp, &op_data->op_fid1);
1570                 if (rc == 0) {
1571                         rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1572                                                  op_data, &op_data->op_fid2);
1573                         if (rc)
1574                                 RETURN(rc);
1575                         goto repeat;
1576                 }
1577         }
1578         RETURN(rc);
1579 }
1580
1581 static int lmv_done_writing(struct obd_export *exp,
1582                             struct md_op_data *op_data,
1583                             struct md_open_data *mod)
1584 {
1585         struct obd_device     *obd = exp->exp_obd;
1586         struct lmv_obd        *lmv = &obd->u.lmv;
1587         struct lmv_tgt_desc   *tgt;
1588         int                    rc;
1589         ENTRY;
1590
1591         rc = lmv_check_connect(obd);
1592         if (rc)
1593                 RETURN(rc);
1594
1595         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1596         if (IS_ERR(tgt))
1597                 RETURN(PTR_ERR(tgt));
1598
1599         rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1600         RETURN(rc);
1601 }
1602
1603 static int
1604 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1605                    struct lookup_intent *it, struct md_op_data *op_data,
1606                    struct lustre_handle *lockh, void *lmm, int lmmsize)
1607 {
1608         struct obd_device     *obd = exp->exp_obd;
1609         struct lmv_obd        *lmv = &obd->u.lmv;
1610         struct lmv_stripe_md  *mea = op_data->op_mea1;
1611         struct md_op_data     *op_data2;
1612         struct lmv_tgt_desc   *tgt;
1613         int                    i;
1614         int                    rc = 0;
1615         ENTRY;
1616
1617         OBD_ALLOC_PTR(op_data2);
1618         if (op_data2 == NULL)
1619                 RETURN(-ENOMEM);
1620
1621         LASSERT(mea != NULL);
1622         for (i = 0; i < mea->mea_count; i++) {
1623                 memset(op_data2, 0, sizeof(*op_data2));
1624                 op_data2->op_fid1 = mea->mea_ids[i];
1625                 op_data2->op_bias = 0;
1626
1627                 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1628                 if (IS_ERR(tgt))
1629                         GOTO(cleanup, rc = PTR_ERR(tgt));
1630
1631                 if (tgt->ltd_exp == NULL)
1632                         continue;
1633
1634                 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1635                                 lockh + i, lmm, lmmsize, NULL, 0);
1636
1637                 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1638                        PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1639
1640                 if (rc)
1641                         GOTO(cleanup, rc);
1642
1643                 if (it->d.lustre.it_data) {
1644                         struct ptlrpc_request *req;
1645                         req = (struct ptlrpc_request *)it->d.lustre.it_data;
1646                         ptlrpc_req_finished(req);
1647                 }
1648
1649                 if (it->d.lustre.it_status)
1650                         GOTO(cleanup, rc = it->d.lustre.it_status);
1651         }
1652
1653         EXIT;
1654 cleanup:
1655         OBD_FREE_PTR(op_data2);
1656
1657         if (rc != 0) {
1658                 /*
1659                  * Drop all taken locks.
1660                  */
1661                 while (--i >= 0) {
1662                         if (lockh[i].cookie)
1663                                 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1664                         lockh[i].cookie = 0;
1665                 }
1666         }
1667         return rc;
1668 }
1669
1670 static int
1671 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1672                    struct lookup_intent *it, struct md_op_data *op_data,
1673                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1674                    int extra_lock_flags)
1675 {
1676         struct ptlrpc_request      *req = it->d.lustre.it_data;
1677         struct obd_device          *obd = exp->exp_obd;
1678         struct lmv_obd             *lmv = &obd->u.lmv;
1679         struct lustre_handle        plock;
1680         struct lmv_tgt_desc        *tgt;
1681         struct md_op_data          *rdata;
1682         struct lu_fid               fid1;
1683         struct mdt_body            *body;
1684         int                         rc = 0;
1685         int                         pmode;
1686         ENTRY;
1687
1688         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1689         LASSERT(body != NULL);
1690
1691         if (!(body->valid & OBD_MD_MDS))
1692                 RETURN(0);
1693
1694         CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1695                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1696
1697         /*
1698          * We got LOOKUP lock, but we really need attrs.
1699          */
1700         pmode = it->d.lustre.it_lock_mode;
1701         LASSERT(pmode != 0);
1702         memcpy(&plock, lockh, sizeof(plock));
1703         it->d.lustre.it_lock_mode = 0;
1704         it->d.lustre.it_data = NULL;
1705         fid1 = body->fid1;
1706
1707         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1708         ptlrpc_req_finished(req);
1709
1710         tgt = lmv_find_target(lmv, &fid1);
1711         if (IS_ERR(tgt))
1712                 GOTO(out, rc = PTR_ERR(tgt));
1713
1714         OBD_ALLOC_PTR(rdata);
1715         if (rdata == NULL)
1716                 GOTO(out, rc = -ENOMEM);
1717
1718         rdata->op_fid1 = fid1;
1719         rdata->op_bias = MDS_CROSS_REF;
1720
1721         rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1722                         lmm, lmmsize, NULL, extra_lock_flags);
1723         OBD_FREE_PTR(rdata);
1724         EXIT;
1725 out:
1726         ldlm_lock_decref(&plock, pmode);
1727         return rc;
1728 }
1729
1730 static int
1731 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1732             struct lookup_intent *it, struct md_op_data *op_data,
1733             struct lustre_handle *lockh, void *lmm, int lmmsize,
1734             struct ptlrpc_request **req, int extra_lock_flags)
1735 {
1736         struct obd_device        *obd = exp->exp_obd;
1737         struct lmv_obd           *lmv = &obd->u.lmv;
1738         struct lmv_tgt_desc      *tgt;
1739         struct lmv_object        *obj;
1740         int                       sidx;
1741         int                       rc;
1742         ENTRY;
1743
1744         rc = lmv_check_connect(obd);
1745         if (rc)
1746                 RETURN(rc);
1747
1748         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1749                LL_IT2STR(it), PFID(&op_data->op_fid1));
1750
1751         if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1752                 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1753                                         lockh, lmm, lmmsize);
1754                 RETURN(rc);
1755         }
1756
1757         obj = lmv_object_find(obd, &op_data->op_fid1);
1758         if (obj && op_data->op_namelen) {
1759                 sidx = raw_name2idx(obj->lo_hashtype,
1760                                        obj->lo_objcount,
1761                                        (char *)op_data->op_name,
1762                                        op_data->op_namelen);
1763                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1764                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1765         } else {
1766                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1767         }
1768         if (obj)
1769                 lmv_object_put(obj);
1770
1771         if (IS_ERR(tgt))
1772                 RETURN(PTR_ERR(tgt));
1773
1774         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1775                LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1776
1777         rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1778                         lmm, lmmsize, req, extra_lock_flags);
1779
1780         if (rc == 0 && it && it->it_op == IT_OPEN) {
1781                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1782                                         lmm, lmmsize, extra_lock_flags);
1783         }
1784         RETURN(rc);
1785 }
1786
1787 static int
1788 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1789                  struct ptlrpc_request **request)
1790 {
1791         struct ptlrpc_request   *req = NULL;
1792         struct obd_device       *obd = exp->exp_obd;
1793         struct lmv_obd          *lmv = &obd->u.lmv;
1794         struct lu_fid            rid = op_data->op_fid1;
1795         struct lmv_tgt_desc     *tgt;
1796         struct mdt_body         *body;
1797         struct lmv_object       *obj;
1798         obd_valid                valid = op_data->op_valid;
1799         int                      rc;
1800         int                      loop = 0;
1801         int                      sidx;
1802         ENTRY;
1803
1804         rc = lmv_check_connect(obd);
1805         if (rc)
1806                 RETURN(rc);
1807
1808 repeat:
1809         ++loop;
1810         LASSERT(loop <= 2);
1811         obj = lmv_object_find(obd, &rid);
1812         if (obj) {
1813                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1814                                     op_data->op_name, op_data->op_namelen);
1815                 rid = obj->lo_stripes[sidx].ls_fid;
1816                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1817                 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1818                 valid &= ~OBD_MD_FLCKSPLIT;
1819                 lmv_object_put(obj);
1820         } else {
1821                 tgt = lmv_find_target(lmv, &rid);
1822                 valid |= OBD_MD_FLCKSPLIT;
1823                 op_data->op_mds = tgt->ltd_idx;
1824         }
1825         if (IS_ERR(tgt))
1826                 RETURN(PTR_ERR(tgt));
1827
1828         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1829                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1830                PFID(&rid), tgt->ltd_idx);
1831
1832         op_data->op_valid = valid;
1833         op_data->op_fid1 = rid;
1834         rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1835         if (rc == 0) {
1836                 body = req_capsule_server_get(&(*request)->rq_pill,
1837                                               &RMF_MDT_BODY);
1838                 LASSERT(body != NULL);
1839
1840                 if (body->valid & OBD_MD_MDS) {
1841                         rid = body->fid1;
1842                         CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1843                                PFID(&rid));
1844
1845                         tgt = lmv_find_target(lmv, &rid);
1846                         if (IS_ERR(tgt)) {
1847                                 ptlrpc_req_finished(*request);
1848                                 RETURN(PTR_ERR(tgt));
1849                         }
1850
1851                         op_data->op_fid1 = rid;
1852                         op_data->op_valid |= OBD_MD_FLCROSSREF;
1853                         op_data->op_namelen = 0;
1854                         op_data->op_name = NULL;
1855                         rc = md_getattr_name(tgt->ltd_exp, op_data, &req);
1856                         ptlrpc_req_finished(*request);
1857                         *request = req;
1858                 }
1859         } else if (rc == -ERESTART) {
1860                 LASSERT(*request != NULL);
1861                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1862                           "Got -ERESTART during getattr!\n");
1863                 ptlrpc_req_finished(*request);
1864                 *request = NULL;
1865
1866                 /*
1867                  * Directory got split. Time to update local object and repeat
1868                  * the request with proper MDS.
1869                  */
1870                 rc = lmv_handle_split(exp, &rid);
1871                 if (rc == 0)
1872                         goto repeat;
1873         }
1874         RETURN(rc);
1875 }
1876
1877 #define md_op_data_fid(op_data, fl)                     \
1878         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1879          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1880          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1881          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1882          NULL)
1883
1884 static int lmv_early_cancel_slaves(struct obd_export *exp,
1885                                    struct md_op_data *op_data, int op_tgt,
1886                                    ldlm_mode_t mode, int bits, int flag)
1887 {
1888         struct obd_device      *obd = exp->exp_obd;
1889         struct lmv_obd         *lmv = &obd->u.lmv;
1890         ldlm_policy_data_t      policy = {{0}};
1891         struct lu_fid          *op_fid;
1892         struct lu_fid          *st_fid;
1893         struct lmv_tgt_desc    *tgt;
1894         struct lmv_object      *obj;
1895         int                     rc = 0;
1896         int                     i;
1897         ENTRY;
1898
1899         op_fid = md_op_data_fid(op_data, flag);
1900         if (!fid_is_sane(op_fid))
1901                 RETURN(0);
1902
1903         obj = lmv_object_find(obd, op_fid);
1904         if (obj == NULL)
1905                 RETURN(-EALREADY);
1906
1907         policy.l_inodebits.bits = bits;
1908         for (i = 0; i < obj->lo_objcount; i++) {
1909                 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1910                 st_fid = &obj->lo_stripes[i].ls_fid;
1911                 if (op_tgt != tgt->ltd_idx) {
1912                         CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1913                                PFID(st_fid), tgt->ltd_idx);
1914                         rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1915                                               mode, LCF_ASYNC, NULL);
1916                         if (rc)
1917                                 GOTO(out_put_obj, rc);
1918                 } else {
1919                         CDEBUG(D_INODE,
1920                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1921                                op_tgt, PFID(st_fid));
1922                         /*
1923                          * Do not cancel locks for operation target, they will
1924                          * be handled later in underlaying layer when calling
1925                          * function we run on behalf of.
1926                          */
1927                         *op_fid = *st_fid;
1928                         op_data->op_flags |= flag;
1929                 }
1930         }
1931         EXIT;
1932 out_put_obj:
1933         lmv_object_put(obj);
1934         return rc;
1935 }
1936
1937 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1938                             int op_tgt, ldlm_mode_t mode, int bits, int flag)
1939 {
1940         struct lu_fid          *fid = md_op_data_fid(op_data, flag);
1941         struct obd_device      *obd = exp->exp_obd;
1942         struct lmv_obd         *lmv = &obd->u.lmv;
1943         struct lmv_tgt_desc    *tgt;
1944         ldlm_policy_data_t      policy = {{0}};
1945         struct lmv_object      *obj;
1946         int                     rc = 0;
1947         ENTRY;
1948
1949         if (!fid_is_sane(fid))
1950                 RETURN(0);
1951
1952         obj = lmv_object_find(obd, fid);
1953         if (obj) {
1954                 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1955                                              bits, flag);
1956                 lmv_object_put(obj);
1957         } else {
1958                 tgt = lmv_find_target(lmv, fid);
1959                 if (IS_ERR(tgt))
1960                         RETURN(PTR_ERR(tgt));
1961
1962                 if (tgt->ltd_idx != op_tgt) {
1963                         CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1964                         policy.l_inodebits.bits = bits;
1965                         rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1966                                               mode, LCF_ASYNC, NULL);
1967                 } else {
1968                         CDEBUG(D_INODE,
1969                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1970                                op_tgt, PFID(fid));
1971                         op_data->op_flags |= flag;
1972                         rc = 0;
1973                 }
1974
1975         }
1976         RETURN(rc);
1977 }
1978
1979 /*
1980  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1981  * op_data->op_fid2
1982  */
1983 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1984                     struct ptlrpc_request **request)
1985 {
1986         struct obd_device       *obd = exp->exp_obd;
1987         struct lmv_obd          *lmv = &obd->u.lmv;
1988         struct lmv_tgt_desc     *tgt;
1989         struct lmv_object       *obj;
1990         int                      rc;
1991         int                      loop = 0;
1992         mdsno_t                  mds;
1993         int                      sidx;
1994         ENTRY;
1995
1996         rc = lmv_check_connect(obd);
1997         if (rc)
1998                 RETURN(rc);
1999
2000 repeat:
2001         ++loop;
2002         LASSERT(loop <= 2);
2003         LASSERT(op_data->op_namelen != 0);
2004
2005         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2006                PFID(&op_data->op_fid2), op_data->op_namelen,
2007                op_data->op_name, PFID(&op_data->op_fid1));
2008
2009         obj = lmv_object_find(obd, &op_data->op_fid2);
2010         if (obj) {
2011                 sidx = raw_name2idx(obj->lo_hashtype,
2012                                     obj->lo_objcount,
2013                                     op_data->op_name,
2014                                     op_data->op_namelen);
2015                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2016                 mds = obj->lo_stripes[sidx].ls_mds;
2017                 lmv_object_put(obj);
2018         } else {
2019                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
2020                 if (rc)
2021                         RETURN(rc);
2022         }
2023
2024         CDEBUG(D_INODE, "Forward to mds #%x ("DFID")\n",
2025                mds, PFID(&op_data->op_fid1));
2026
2027         op_data->op_fsuid = cfs_curproc_fsuid();
2028         op_data->op_fsgid = cfs_curproc_fsgid();
2029         op_data->op_cap = cfs_curproc_cap_pack();
2030         tgt = lmv_get_target(lmv, mds);
2031
2032         /*
2033          * Cancel UPDATE lock on child (fid1).
2034          */
2035         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2036         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2037                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2038         if (rc == 0)
2039                 rc = md_link(tgt->ltd_exp, op_data, request);
2040         if (rc == -ERESTART) {
2041                 LASSERT(*request != NULL);
2042                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2043                           "Got -ERESTART during link!\n");
2044                 ptlrpc_req_finished(*request);
2045                 *request = NULL;
2046
2047                 /*
2048                  * Directory got split. Time to update local object and repeat
2049                  * the request with proper MDS.
2050                  */
2051                 rc = lmv_handle_split(exp, &op_data->op_fid2);
2052                 if (rc == 0)
2053                         goto repeat;
2054         }
2055
2056         RETURN(rc);
2057 }
2058
2059 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2060                       const char *old, int oldlen, const char *new, int newlen,
2061                       struct ptlrpc_request **request)
2062 {
2063         struct obd_device       *obd = exp->exp_obd;
2064         struct lmv_obd          *lmv = &obd->u.lmv;
2065         struct lmv_tgt_desc     *src_tgt;
2066         int                      rc;
2067         int                      sidx;
2068         int                      loop = 0;
2069         struct lmv_object       *obj;
2070         mdsno_t                  mds1;
2071         mdsno_t                  mds2;
2072         ENTRY;
2073
2074         LASSERT(oldlen != 0);
2075
2076         CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2077                oldlen, old, PFID(&op_data->op_fid1),
2078                newlen, new, PFID(&op_data->op_fid2));
2079
2080         rc = lmv_check_connect(obd);
2081         if (rc)
2082                 RETURN(rc);
2083
2084 repeat:
2085         ++loop;
2086         LASSERT(loop <= 2);
2087         obj = lmv_object_find(obd, &op_data->op_fid1);
2088         if (obj) {
2089                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2090                                     (char *)old, oldlen);
2091                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2092                 mds1 = obj->lo_stripes[sidx].ls_mds;
2093                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2094                 lmv_object_put(obj);
2095         } else {
2096                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2097                 if (rc)
2098                         RETURN(rc);
2099         }
2100
2101         obj = lmv_object_find(obd, &op_data->op_fid2);
2102         if (obj) {
2103                 /*
2104                  * Directory is already split, so we have to forward request to
2105                  * the right MDS.
2106                  */
2107                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2108                                        (char *)new, newlen);
2109
2110                 mds2 = obj->lo_stripes[sidx].ls_mds;
2111                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2112                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2113                 lmv_object_put(obj);
2114         } else {
2115                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2116                 if (rc)
2117                         RETURN(rc);
2118         }
2119
2120         op_data->op_fsuid = cfs_curproc_fsuid();
2121         op_data->op_fsgid = cfs_curproc_fsgid();
2122         op_data->op_cap = cfs_curproc_cap_pack();
2123
2124         src_tgt = lmv_get_target(lmv, mds1);
2125
2126         /*
2127          * LOOKUP lock on src child (fid3) should also be cancelled for
2128          * src_tgt in mdc_rename.
2129          */
2130         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2131
2132         /*
2133          * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2134          * own target.
2135          */
2136         rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2137                               LCK_EX, MDS_INODELOCK_UPDATE,
2138                               MF_MDC_CANCEL_FID2);
2139
2140         /*
2141          * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2142          */
2143         if (rc == 0) {
2144                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2145                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2146                                       MF_MDC_CANCEL_FID4);
2147         }
2148
2149         /*
2150          * Cancel all the locks on tgt child (fid4).
2151          */
2152         if (rc == 0)
2153                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2154                                       LCK_EX, MDS_INODELOCK_FULL,
2155                                       MF_MDC_CANCEL_FID4);
2156
2157         if (rc == 0)
2158                 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2159                                new, newlen, request);
2160
2161         if (rc == -ERESTART) {
2162                 LASSERT(*request != NULL);
2163                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2164                           "Got -ERESTART during rename!\n");
2165                 ptlrpc_req_finished(*request);
2166                 *request = NULL;
2167
2168                 /*
2169                  * Directory got split. Time to update local object and repeat
2170                  * the request with proper MDS.
2171                  */
2172                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2173                 if (rc == 0)
2174                         goto repeat;
2175         }
2176         RETURN(rc);
2177 }
2178
2179 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2180                        void *ea, int ealen, void *ea2, int ea2len,
2181                        struct ptlrpc_request **request,
2182                        struct md_open_data **mod)
2183 {
2184         struct obd_device       *obd = exp->exp_obd;
2185         struct lmv_obd          *lmv = &obd->u.lmv;
2186         struct ptlrpc_request   *req;
2187         struct lmv_tgt_desc     *tgt;
2188         struct lmv_object       *obj;
2189         int                      rc = 0;
2190         int                      i;
2191         ENTRY;
2192
2193         rc = lmv_check_connect(obd);
2194         if (rc)
2195                 RETURN(rc);
2196
2197         obj = lmv_object_find(obd, &op_data->op_fid1);
2198
2199         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2200                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2201                obj ? ", split" : "");
2202
2203         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2204         if (obj) {
2205                 for (i = 0; i < obj->lo_objcount; i++) {
2206                         op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2207
2208                         tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2209                         if (IS_ERR(tgt)) {
2210                                 rc = PTR_ERR(tgt);
2211                                 break;
2212                         }
2213
2214                         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2215                                         ea2, ea2len, &req, mod);
2216
2217                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2218                                 /*
2219                                  * This is master object and this request should
2220                                  * be returned back to llite.
2221                                  */
2222                                 *request = req;
2223                         } else {
2224                                 ptlrpc_req_finished(req);
2225                         }
2226
2227                         if (rc)
2228                                 break;
2229                 }
2230                 lmv_object_put(obj);
2231         } else {
2232                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2233                 if (IS_ERR(tgt))
2234                         RETURN(PTR_ERR(tgt));
2235
2236                 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2237                                 ea2len, request, mod);
2238         }
2239         RETURN(rc);
2240 }
2241
2242 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2243                     struct obd_capa *oc, struct ptlrpc_request **request)
2244 {
2245         struct obd_device         *obd = exp->exp_obd;
2246         struct lmv_obd            *lmv = &obd->u.lmv;
2247         struct lmv_tgt_desc       *tgt;
2248         int                        rc;
2249         ENTRY;
2250
2251         rc = lmv_check_connect(obd);
2252         if (rc)
2253                 RETURN(rc);
2254
2255         tgt = lmv_find_target(lmv, fid);
2256         if (IS_ERR(tgt))
2257                 RETURN(PTR_ERR(tgt));
2258
2259         rc = md_sync(tgt->ltd_exp, fid, oc, request);
2260         RETURN(rc);
2261 }
2262
2263 /**
2264  * Main purpose of LMV blocking ast is to remove split directory LMV
2265  * presentation object (struct lmv_object) attached to the lock being revoked.
2266  */
2267 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2268                      void *data, int flag)
2269 {
2270         struct lustre_handle    lockh;
2271         struct lmv_object      *obj;
2272         int                     rc;
2273         ENTRY;
2274
2275         switch (flag) {
2276         case LDLM_CB_BLOCKING:
2277                 ldlm_lock2handle(lock, &lockh);
2278                 rc = ldlm_cli_cancel(&lockh);
2279                 if (rc < 0) {
2280                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2281                         RETURN(rc);
2282                 }
2283                 break;
2284         case LDLM_CB_CANCELING:
2285                 /*
2286                  * Time to drop cached attrs for split directory object
2287                  */
2288                 obj = lock->l_ast_data;
2289                 if (obj) {
2290                         CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2291                                ", master "DFID"\n",
2292                                lock->l_resource->lr_name.name[3] == 1 ?
2293                                "LOOKUP" : "UPDATE",
2294                                lock->l_resource->lr_name.name[0],
2295                                lock->l_resource->lr_name.name[1],
2296                                PFID(&obj->lo_fid));
2297                         lmv_object_put(obj);
2298                 }
2299                 break;
2300         default:
2301                 LBUG();
2302         }
2303         RETURN(0);
2304 }
2305
2306 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2307 {
2308         __u64         val;
2309
2310         val = le64_to_cpu(*hash);
2311         if (val < hash_adj)
2312                 val += MAX_HASH_SIZE;
2313         if (val != MDS_DIR_END_OFF)
2314                 *hash = cpu_to_le64(val - hash_adj);
2315 }
2316
2317 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2318 {
2319         __u64              id;
2320         struct obd_import *imp;
2321
2322         /*
2323          * XXX: to get nid we assume that underlying obd device is mdc.
2324          */
2325         imp  = class_exp2cliimp(exp);
2326         id   = imp->imp_connection->c_self + fid_flatten(fid);
2327
2328         CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2329                imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2330
2331         return id ^ (id >> 32);
2332 }
2333
2334 static int lmv_readpage(struct obd_export *exp, struct md_op_data *op_data,
2335                         struct page **pages, struct ptlrpc_request **request)
2336 {
2337         struct obd_device       *obd = exp->exp_obd;
2338         struct lmv_obd          *lmv = &obd->u.lmv;
2339         struct lmv_object       *obj;
2340         struct lu_fid            rid = op_data->op_fid1;
2341         __u64                    offset = op_data->op_offset;
2342         __u64                    hash_adj = 0;
2343         __u32                    rank = 0;
2344         __u64                    seg_size = 0;
2345         __u64                    tgt_tmp = 0;
2346         int                      tgt_idx = 0;
2347         int                      tgt0_idx = 0;
2348         int                      rc;
2349         int                      nr = 0;
2350         int                      i;
2351         /* number of pages read, in CFS_PAGE_SIZE */
2352         int                      nrdpgs;
2353         /* number of pages transferred in LU_PAGE_SIZE */
2354         int                      nlupgs;
2355         struct lmv_stripe       *los;
2356         struct lmv_tgt_desc     *tgt;
2357         struct lu_dirpage       *dp;
2358         struct lu_dirent        *ent;
2359         ENTRY;
2360
2361         rc = lmv_check_connect(obd);
2362         if (rc)
2363                 RETURN(rc);
2364
2365         CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2366
2367         /*
2368          * This case handle directory lookup in clustered metadata case (i.e.
2369          * split directory is located on multiple md servers.)
2370          * each server keeps directory entries for certain range of hashes.
2371          * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2372          * first server will keep records with hashes [ 0 ... MAX_HASH / N  - 1],
2373          * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2374          * so on....
2375          *      readdir can simply start reading entries from 0 - N server in
2376          * order but that will not scale well as all client will request dir in
2377          * to server in same order.
2378          * Following algorithm does optimization:
2379          * Instead of doing readdir in 1, 2, ...., N order, client with a
2380          * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2381          * (every client has rank R)
2382          *      But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2383          * since client ask dir from MDS{R} client has pages with offsets
2384          * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2385          * on hash  values that we get.
2386          */
2387         obj = lmv_object_find_lock(obd, &rid);
2388         if (obj) {
2389                 nr       = obj->lo_objcount;
2390                 LASSERT(nr > 0);
2391                 seg_size = MAX_HASH_SIZE;
2392                 do_div(seg_size, nr);
2393                 los      = obj->lo_stripes;
2394                 tgt      = lmv_get_target(lmv, los[0].ls_mds);
2395                 rank     = lmv_node_rank(tgt->ltd_exp, &rid) % nr;
2396                 tgt_tmp  = offset;
2397                 do_div(tgt_tmp, seg_size);
2398                 tgt0_idx = do_div(tgt_tmp,  nr);
2399                 tgt_idx  = (tgt0_idx + rank) % nr;
2400
2401                 if (tgt_idx < tgt0_idx)
2402                         /*
2403                          * Wrap around.
2404                          *
2405                          * Last segment has unusual length due to division
2406                          * rounding.
2407                          */
2408                         hash_adj = MAX_HASH_SIZE - seg_size * nr;
2409                 else
2410                         hash_adj = 0;
2411
2412                 hash_adj += rank * seg_size;
2413
2414                 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2415                        LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2416                        offset, tgt0_idx, offset + hash_adj, tgt_idx);
2417
2418                 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2419                 rid = obj->lo_stripes[tgt_idx].ls_fid;
2420                 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2421
2422                 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2423                        PFID(&rid), (unsigned long)offset, tgt_idx);
2424         } else
2425                 tgt = lmv_find_target(lmv, &rid);
2426
2427         if (IS_ERR(tgt))
2428                 GOTO(cleanup, rc = PTR_ERR(tgt));
2429
2430         op_data->op_fid1 = rid;
2431         rc = md_readpage(tgt->ltd_exp, op_data, pages, request);
2432         if (rc)
2433                 GOTO(cleanup, rc);
2434
2435         nrdpgs = ((*request)->rq_bulk->bd_nob_transferred + CFS_PAGE_SIZE - 1)
2436                  >> CFS_PAGE_SHIFT;
2437         nlupgs = (*request)->rq_bulk->bd_nob_transferred >> LU_PAGE_SHIFT;
2438         LASSERT(!((*request)->rq_bulk->bd_nob_transferred & ~LU_PAGE_MASK));
2439         LASSERT(nrdpgs > 0 && nrdpgs <= op_data->op_npages);
2440
2441         CDEBUG(D_INODE, "read %d(%d)/%d pages\n", nrdpgs, nlupgs,
2442                op_data->op_npages);
2443
2444         for (i = 0; i < nrdpgs; i++) {
2445 #if CFS_PAGE_SIZE > LU_PAGE_SIZE
2446                 struct lu_dirpage *first;
2447                 __u64 hash_end = 0;
2448                 __u32 flags = 0;
2449 #endif
2450                 struct lu_dirent *tmp = NULL;
2451
2452                 dp = cfs_kmap(pages[i]);
2453                 if (obj) {
2454                         lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2455                         lmv_hash_adjust(&dp->ldp_hash_end,   hash_adj);
2456                         LASSERT(le64_to_cpu(dp->ldp_hash_start) <=
2457                                 op_data->op_offset);
2458
2459                         if ((tgt0_idx != nr - 1) &&
2460                             (le64_to_cpu(dp->ldp_hash_end) == MDS_DIR_END_OFF))
2461                         {
2462                                 dp->ldp_hash_end = cpu_to_le32(seg_size *
2463                                                                (tgt0_idx + 1));
2464                                 CDEBUG(D_INODE,
2465                                        ""DFID" reset end "LPX64" tgt %d\n",
2466                                        PFID(&rid),
2467                                        (__u64)le64_to_cpu(dp->ldp_hash_end),
2468                                        tgt_idx);
2469                         }
2470                 }
2471
2472                 ent = lu_dirent_start(dp);
2473 #if CFS_PAGE_SIZE > LU_PAGE_SIZE
2474                 first = dp;
2475                 hash_end = dp->ldp_hash_end;
2476 repeat:
2477 #endif
2478                 nlupgs--;
2479                 for (tmp = ent; ent != NULL;
2480                      tmp = ent, ent = lu_dirent_next(ent)) {
2481                         if (obj)
2482                                 lmv_hash_adjust(&ent->lde_hash, hash_adj);
2483                 }
2484
2485 #if CFS_PAGE_SIZE > LU_PAGE_SIZE
2486                 dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
2487                 if (((unsigned long)dp & ~CFS_PAGE_MASK) && nlupgs > 0) {
2488                         ent = lu_dirent_start(dp);
2489
2490                         if (obj) {
2491                                 lmv_hash_adjust(&dp->ldp_hash_end, hash_adj);
2492                                 if ((tgt0_idx != nr - 1) &&
2493                                     (le64_to_cpu(dp->ldp_hash_end) ==
2494                                      MDS_DIR_END_OFF)) {
2495                                         hash_end = cpu_to_le32(seg_size *
2496                                                                (tgt0_idx + 1));
2497                                         CDEBUG(D_INODE,
2498                                             ""DFID" reset end "LPX64" tgt %d\n",
2499                                             PFID(&rid),
2500                                             (__u64)le64_to_cpu(hash_end),
2501                                             tgt_idx);
2502                                 }
2503                         }
2504                         hash_end = dp->ldp_hash_end;
2505                         flags = dp->ldp_flags;
2506
2507                         if (tmp) {
2508                                 /* enlarge the end entry lde_reclen from 0 to
2509                                  * first entry of next lu_dirpage, in this way
2510                                  * several lu_dirpages can be stored into one
2511                                  * client page on client. */
2512                                 tmp = ((void *)tmp) +
2513                                       le16_to_cpu(tmp->lde_reclen);
2514                                 tmp->lde_reclen =
2515                                         cpu_to_le16((char *)(dp->ldp_entries) -
2516                                                     (char *)tmp);
2517                                 goto repeat;
2518                         }
2519                 }
2520                 first->ldp_hash_end = hash_end;
2521                 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
2522                 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
2523 #else
2524                 SET_BUT_UNUSED(tmp);
2525 #endif
2526                 cfs_kunmap(pages[i]);
2527         }
2528         EXIT;
2529 cleanup:
2530         if (obj)
2531                 lmv_object_put_unlock(obj);
2532         return rc;
2533 }
2534
2535 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2536                       struct ptlrpc_request **request)
2537 {
2538         struct obd_device       *obd = exp->exp_obd;
2539         struct lmv_obd          *lmv = &obd->u.lmv;
2540         struct lmv_tgt_desc     *tgt = NULL;
2541         struct lmv_object       *obj;
2542         int                      rc;
2543         int                      sidx;
2544         int                      loop = 0;
2545         ENTRY;
2546
2547         rc = lmv_check_connect(obd);
2548         if (rc)
2549                 RETURN(rc);
2550
2551 repeat:
2552         ++loop;
2553         LASSERT(loop <= 2);
2554         LASSERT(op_data->op_namelen != 0);
2555
2556         obj = lmv_object_find(obd, &op_data->op_fid1);
2557         if (obj) {
2558                 sidx = raw_name2idx(obj->lo_hashtype,
2559                                     obj->lo_objcount,
2560                                     op_data->op_name,
2561                                     op_data->op_namelen);
2562                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2563                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2564                 tgt = lmv_get_target(lmv,
2565                                      obj->lo_stripes[sidx].ls_mds);
2566                 lmv_object_put(obj);
2567                 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2568                        op_data->op_namelen, op_data->op_name,
2569                        PFID(&op_data->op_fid1), sidx);
2570         }
2571
2572         if (tgt == NULL) {
2573                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2574                 if (IS_ERR(tgt))
2575                         RETURN(PTR_ERR(tgt));
2576                 op_data->op_bias |= MDS_CHECK_SPLIT;
2577         }
2578
2579         op_data->op_fsuid = cfs_curproc_fsuid();
2580         op_data->op_fsgid = cfs_curproc_fsgid();
2581         op_data->op_cap = cfs_curproc_cap_pack();
2582
2583         /*
2584          * If child's fid is given, cancel unused locks for it if it is from
2585          * another export than parent.
2586          *
2587          * LOOKUP lock for child (fid3) should also be cancelled on parent
2588          * tgt_tgt in mdc_unlink().
2589          */
2590         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2591
2592         /*
2593          * Cancel FULL locks on child (fid3).
2594          */
2595         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2596                               MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2597
2598         if (rc == 0)
2599                 rc = md_unlink(tgt->ltd_exp, op_data, request);
2600
2601         if (rc == -ERESTART) {
2602                 LASSERT(*request != NULL);
2603                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2604                           "Got -ERESTART during unlink!\n");
2605                 ptlrpc_req_finished(*request);
2606                 *request = NULL;
2607
2608                 /*
2609                  * Directory got split. Time to update local object and repeat
2610                  * the request with proper MDS.
2611                  */
2612                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2613                 if (rc == 0)
2614                         goto repeat;
2615         }
2616         RETURN(rc);
2617 }
2618
2619 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2620 {
2621         struct lmv_obd *lmv = &obd->u.lmv;
2622         int rc = 0;
2623
2624         switch (stage) {
2625         case OBD_CLEANUP_EARLY:
2626                 /* XXX: here should be calling obd_precleanup() down to
2627                  * stack. */
2628                 break;
2629         case OBD_CLEANUP_EXPORTS:
2630                 fld_client_proc_fini(&lmv->lmv_fld);
2631                 lprocfs_obd_cleanup(obd);
2632                 break;
2633         default:
2634                 break;
2635         }
2636         RETURN(rc);
2637 }
2638
2639 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2640                         __u32 keylen, void *key, __u32 *vallen, void *val,
2641                         struct lov_stripe_md *lsm)
2642 {
2643         struct obd_device       *obd;
2644         struct lmv_obd          *lmv;
2645         int                      rc = 0;
2646         ENTRY;
2647
2648         obd = class_exp2obd(exp);
2649         if (obd == NULL) {
2650                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2651                        exp->exp_handle.h_cookie);
2652                 RETURN(-EINVAL);
2653         }
2654
2655         lmv = &obd->u.lmv;
2656         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2657                 struct lmv_tgt_desc *tgts;
2658                 int i;
2659
2660                 rc = lmv_check_connect(obd);
2661                 if (rc)
2662                         RETURN(rc);
2663
2664                 LASSERT(*vallen == sizeof(__u32));
2665                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2666                      i++, tgts++) {
2667
2668                         /*
2669                          * All tgts should be connected when this gets called.
2670                          */
2671                         if (!tgts || !tgts->ltd_exp) {
2672                                 CERROR("target not setup?\n");
2673                                 continue;
2674                         }
2675
2676                         if (!obd_get_info(env, tgts->ltd_exp, keylen, key,
2677                                           vallen, val, NULL))
2678                                 RETURN(0);
2679                 }
2680                 RETURN(-EINVAL);
2681         } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2682                 rc = lmv_check_connect(obd);
2683                 if (rc)
2684                         RETURN(rc);
2685
2686                 /*
2687                  * Forwarding this request to first MDS, it should know LOV
2688                  * desc.
2689                  */
2690                 rc = obd_get_info(env, lmv->tgts[0].ltd_exp, keylen, key,
2691                                   vallen, val, NULL);
2692                 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2693                         exp->exp_connect_flags =
2694                         ((struct obd_connect_data *)val)->ocd_connect_flags;
2695                 }
2696                 RETURN(rc);
2697         } else if (KEY_IS(KEY_TGT_COUNT)) {
2698                 *((int *)val) = lmv->desc.ld_tgt_count;
2699                 RETURN(0);
2700         }
2701
2702         CDEBUG(D_IOCTL, "Invalid key\n");
2703         RETURN(-EINVAL);
2704 }
2705
2706 int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
2707                        obd_count keylen, void *key, obd_count vallen,
2708                        void *val, struct ptlrpc_request_set *set)
2709 {
2710         struct lmv_tgt_desc    *tgt;
2711         struct obd_device      *obd;
2712         struct lmv_obd         *lmv;
2713         int rc = 0;
2714         ENTRY;
2715
2716         obd = class_exp2obd(exp);
2717         if (obd == NULL) {
2718                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2719                        exp->exp_handle.h_cookie);
2720                 RETURN(-EINVAL);
2721         }
2722         lmv = &obd->u.lmv;
2723
2724         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX)) {
2725                 int i, err = 0;
2726
2727                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2728                         tgt = &lmv->tgts[i];
2729
2730                         if (!tgt->ltd_exp)
2731                                 continue;
2732
2733                         err = obd_set_info_async(env, tgt->ltd_exp,
2734                                                  keylen, key, vallen, val, set);
2735                         if (err && rc == 0)
2736                                 rc = err;
2737                 }
2738
2739                 RETURN(rc);
2740         }
2741
2742         RETURN(-EINVAL);
2743 }
2744
2745 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2746                struct lov_stripe_md *lsm)
2747 {
2748         struct obd_device         *obd = class_exp2obd(exp);
2749         struct lmv_obd            *lmv = &obd->u.lmv;
2750         struct lmv_stripe_md      *meap;
2751         struct lmv_stripe_md      *lsmp;
2752         int                        mea_size;
2753         int                        i;
2754         ENTRY;
2755
2756         mea_size = lmv_get_easize(lmv);
2757         if (!lmmp)
2758                 RETURN(mea_size);
2759
2760         if (*lmmp && !lsm) {
2761                 OBD_FREE_LARGE(*lmmp, mea_size);
2762                 *lmmp = NULL;
2763                 RETURN(0);
2764         }
2765
2766         if (*lmmp == NULL) {
2767                 OBD_ALLOC_LARGE(*lmmp, mea_size);
2768                 if (*lmmp == NULL)
2769                         RETURN(-ENOMEM);
2770         }
2771
2772         if (!lsm)
2773                 RETURN(mea_size);
2774
2775         lsmp = (struct lmv_stripe_md *)lsm;
2776         meap = (struct lmv_stripe_md *)*lmmp;
2777
2778         if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2779             lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2780                 RETURN(-EINVAL);
2781
2782         meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2783         meap->mea_count = cpu_to_le32(lsmp->mea_count);
2784         meap->mea_master = cpu_to_le32(lsmp->mea_master);
2785
2786         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2787                 meap->mea_ids[i] = meap->mea_ids[i];
2788                 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2789         }
2790
2791         RETURN(mea_size);
2792 }
2793
2794 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2795                  struct lov_mds_md *lmm, int lmm_size)
2796 {
2797         struct obd_device          *obd = class_exp2obd(exp);
2798         struct lmv_stripe_md      **tmea = (struct lmv_stripe_md **)lsmp;
2799         struct lmv_stripe_md       *mea = (struct lmv_stripe_md *)lmm;
2800         struct lmv_obd             *lmv = &obd->u.lmv;
2801         int                         mea_size;
2802         int                         i;
2803         __u32                       magic;
2804         ENTRY;
2805
2806         mea_size = lmv_get_easize(lmv);
2807         if (lsmp == NULL)
2808                 return mea_size;
2809
2810         if (*lsmp != NULL && lmm == NULL) {
2811                 OBD_FREE_LARGE(*tmea, mea_size);
2812                 *lsmp = NULL;
2813                 RETURN(0);
2814         }
2815
2816         LASSERT(mea_size == lmm_size);
2817
2818         OBD_ALLOC_LARGE(*tmea, mea_size);
2819         if (*tmea == NULL)
2820                 RETURN(-ENOMEM);
2821
2822         if (!lmm)
2823                 RETURN(mea_size);
2824
2825         if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2826             mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2827             mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2828         {
2829                 magic = le32_to_cpu(mea->mea_magic);
2830         } else {
2831                 /*
2832                  * Old mea is not handled here.
2833                  */
2834                 CERROR("Old not supportable EA is found\n");
2835                 LBUG();
2836         }
2837
2838         (*tmea)->mea_magic = magic;
2839         (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2840         (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2841
2842         for (i = 0; i < (*tmea)->mea_count; i++) {
2843                 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2844                 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2845         }
2846         RETURN(mea_size);
2847 }
2848
2849 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2850                              ldlm_policy_data_t *policy, ldlm_mode_t mode,
2851                              ldlm_cancel_flags_t flags, void *opaque)
2852 {
2853         struct obd_device       *obd = exp->exp_obd;
2854         struct lmv_obd          *lmv = &obd->u.lmv;
2855         int                      rc = 0;
2856         int                      err;
2857         int                      i;
2858         ENTRY;
2859
2860         LASSERT(fid != NULL);
2861
2862         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2863                 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2864                         continue;
2865
2866                 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2867                                        policy, mode, flags, opaque);
2868                 if (!rc)
2869                         rc = err;
2870         }
2871         RETURN(rc);
2872 }
2873
2874 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
2875                       __u64 *bits)
2876 {
2877         struct obd_device       *obd = exp->exp_obd;
2878         struct lmv_obd          *lmv = &obd->u.lmv;
2879         int                      rc;
2880         ENTRY;
2881
2882         rc =  md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data, bits);
2883         RETURN(rc);
2884 }
2885
2886 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2887                            const struct lu_fid *fid, ldlm_type_t type,
2888                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
2889                            struct lustre_handle *lockh)
2890 {
2891         struct obd_device       *obd = exp->exp_obd;
2892         struct lmv_obd          *lmv = &obd->u.lmv;
2893         ldlm_mode_t              rc;
2894         int                      i;
2895         ENTRY;
2896
2897         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2898
2899         /*
2900          * With CMD every object can have two locks in different namespaces:
2901          * lookup lock in space of mds storing direntry and update/open lock in
2902          * space of mds storing inode. Thus we check all targets, not only that
2903          * one fid was created in.
2904          */
2905         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2906                 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2907                                    type, policy, mode, lockh);
2908                 if (rc)
2909                         RETURN(rc);
2910         }
2911
2912         RETURN(0);
2913 }
2914
2915 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2916                       struct obd_export *dt_exp, struct obd_export *md_exp,
2917                       struct lustre_md *md)
2918 {
2919         struct obd_device       *obd = exp->exp_obd;
2920         struct lmv_obd          *lmv = &obd->u.lmv;
2921         int                      rc;
2922         ENTRY;
2923         rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2924         RETURN(rc);
2925 }
2926
2927 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2928 {
2929         struct obd_device       *obd = exp->exp_obd;
2930         struct lmv_obd          *lmv = &obd->u.lmv;
2931         ENTRY;
2932
2933         if (md->mea)
2934                 obd_free_memmd(exp, (void *)&md->mea);
2935         RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2936 }
2937
2938 int lmv_set_open_replay_data(struct obd_export *exp,
2939                              struct obd_client_handle *och,
2940                              struct ptlrpc_request *open_req)
2941 {
2942         struct obd_device       *obd = exp->exp_obd;
2943         struct lmv_obd          *lmv = &obd->u.lmv;
2944         struct lmv_tgt_desc     *tgt;
2945         ENTRY;
2946
2947         tgt = lmv_find_target(lmv, &och->och_fid);
2948         if (IS_ERR(tgt))
2949                 RETURN(PTR_ERR(tgt));
2950
2951         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2952 }
2953
2954 int lmv_clear_open_replay_data(struct obd_export *exp,
2955                                struct obd_client_handle *och)
2956 {
2957         struct obd_device       *obd = exp->exp_obd;
2958         struct lmv_obd          *lmv = &obd->u.lmv;
2959         struct lmv_tgt_desc     *tgt;
2960         ENTRY;
2961
2962         tgt = lmv_find_target(lmv, &och->och_fid);
2963         if (IS_ERR(tgt))
2964                 RETURN(PTR_ERR(tgt));
2965
2966         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2967 }
2968
2969 static int lmv_get_remote_perm(struct obd_export *exp,
2970                                const struct lu_fid *fid,
2971                                struct obd_capa *oc, __u32 suppgid,
2972                                struct ptlrpc_request **request)
2973 {
2974         struct obd_device       *obd = exp->exp_obd;
2975         struct lmv_obd          *lmv = &obd->u.lmv;
2976         struct lmv_tgt_desc     *tgt;
2977         int                      rc;
2978         ENTRY;
2979
2980         rc = lmv_check_connect(obd);
2981         if (rc)
2982                 RETURN(rc);
2983
2984         tgt = lmv_find_target(lmv, fid);
2985         if (IS_ERR(tgt))
2986                 RETURN(PTR_ERR(tgt));
2987
2988         rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
2989         RETURN(rc);
2990 }
2991
2992 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2993                           renew_capa_cb_t cb)
2994 {
2995         struct obd_device       *obd = exp->exp_obd;
2996         struct lmv_obd          *lmv = &obd->u.lmv;
2997         struct lmv_tgt_desc     *tgt;
2998         int                      rc;
2999         ENTRY;
3000
3001         rc = lmv_check_connect(obd);
3002         if (rc)
3003                 RETURN(rc);
3004
3005         tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
3006         if (IS_ERR(tgt))
3007                 RETURN(PTR_ERR(tgt));
3008
3009         rc = md_renew_capa(tgt->ltd_exp, oc, cb);
3010         RETURN(rc);
3011 }
3012
3013 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
3014                     const struct req_msg_field *field, struct obd_capa **oc)
3015 {
3016         struct obd_device *obd = exp->exp_obd;
3017         struct lmv_obd *lmv = &obd->u.lmv;
3018         int rc;
3019
3020         ENTRY;
3021         rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
3022         RETURN(rc);
3023 }
3024
3025 int lmv_intent_getattr_async(struct obd_export *exp,
3026                              struct md_enqueue_info *minfo,
3027                              struct ldlm_enqueue_info *einfo)
3028 {
3029         struct md_op_data       *op_data = &minfo->mi_data;
3030         struct obd_device       *obd = exp->exp_obd;
3031         struct lmv_obd          *lmv = &obd->u.lmv;
3032         struct lmv_object       *obj;
3033         struct lmv_tgt_desc     *tgt = NULL;
3034         int                      rc;
3035         int                      sidx;
3036         ENTRY;
3037
3038         rc = lmv_check_connect(obd);
3039         if (rc)
3040                 RETURN(rc);
3041
3042         if (op_data->op_namelen) {
3043                 obj = lmv_object_find(obd, &op_data->op_fid1);
3044                 if (obj) {
3045                         sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
3046                                             (char *)op_data->op_name,
3047                                             op_data->op_namelen);
3048                         op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
3049                         tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
3050                         lmv_object_put(obj);
3051                 }
3052         }
3053
3054         if (tgt == NULL)
3055                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
3056
3057         if (IS_ERR(tgt))
3058                 RETURN(PTR_ERR(tgt));
3059
3060         rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
3061         RETURN(rc);
3062 }
3063
3064 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3065                         struct lu_fid *fid, __u64 *bits)
3066 {
3067         struct obd_device       *obd = exp->exp_obd;
3068         struct lmv_obd          *lmv = &obd->u.lmv;
3069         struct lmv_tgt_desc     *tgt;
3070         int                      rc;
3071         ENTRY;
3072
3073         rc = lmv_check_connect(obd);
3074         if (rc)
3075                 RETURN(rc);
3076
3077         tgt = lmv_find_target(lmv, fid);
3078         if (IS_ERR(tgt))
3079                 RETURN(PTR_ERR(tgt));
3080
3081         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3082         RETURN(rc);
3083 }
3084
3085 /**
3086  * For lmv, only need to send request to master MDT, and the master MDT will
3087  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3088  * we directly fetch data from the slave MDTs.
3089  */
3090 int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3091                  struct obd_quotactl *oqctl)
3092 {
3093         struct obd_device   *obd = class_exp2obd(exp);
3094         struct lmv_obd      *lmv = &obd->u.lmv;
3095         struct lmv_tgt_desc *tgt = &lmv->tgts[0];
3096         int                  rc = 0, i;
3097         __u64                curspace, curinodes;
3098         ENTRY;
3099
3100         if (!lmv->desc.ld_tgt_count || !tgt->ltd_active) {
3101                 CERROR("master lmv inactive\n");
3102                 RETURN(-EIO);
3103         }
3104
3105         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3106                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3107                 RETURN(rc);
3108         }
3109
3110         curspace = curinodes = 0;
3111         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3112                 int err;
3113                 tgt = &lmv->tgts[i];
3114
3115                 if (tgt->ltd_exp == NULL)
3116                         continue;
3117                 if (!tgt->ltd_active) {
3118                         CDEBUG(D_HA, "mdt %d is inactive.\n", i);
3119                         continue;
3120                 }
3121
3122                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3123                 if (err) {
3124                         CERROR("getquota on mdt %d failed. %d\n", i, err);
3125                         if (!rc)
3126                                 rc = err;
3127                 } else {
3128                         curspace += oqctl->qc_dqblk.dqb_curspace;
3129                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3130                 }
3131         }
3132         oqctl->qc_dqblk.dqb_curspace = curspace;
3133         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3134
3135         RETURN(rc);
3136 }
3137
3138 int lmv_quotacheck(struct obd_device *unused, struct obd_export *exp,
3139                    struct obd_quotactl *oqctl)
3140 {
3141         struct obd_device   *obd = class_exp2obd(exp);
3142         struct lmv_obd      *lmv = &obd->u.lmv;
3143         struct lmv_tgt_desc *tgt;
3144         int                  i, rc = 0;
3145         ENTRY;
3146
3147         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
3148                 int err;
3149
3150                 if (!tgt->ltd_active) {
3151                         CERROR("lmv idx %d inactive\n", i);
3152                         RETURN(-EIO);
3153                 }
3154
3155                 err = obd_quotacheck(tgt->ltd_exp, oqctl);
3156                 if (err && !rc)
3157                         rc = err;
3158         }
3159
3160         RETURN(rc);
3161 }
3162
3163 struct obd_ops lmv_obd_ops = {
3164         .o_owner                = THIS_MODULE,
3165         .o_setup                = lmv_setup,
3166         .o_cleanup              = lmv_cleanup,
3167         .o_precleanup           = lmv_precleanup,
3168         .o_process_config       = lmv_process_config,
3169         .o_connect              = lmv_connect,
3170         .o_disconnect           = lmv_disconnect,
3171         .o_statfs               = lmv_statfs,
3172         .o_get_info             = lmv_get_info,
3173         .o_set_info_async       = lmv_set_info_async,
3174         .o_packmd               = lmv_packmd,
3175         .o_unpackmd             = lmv_unpackmd,
3176         .o_notify               = lmv_notify,
3177         .o_get_uuid             = lmv_get_uuid,
3178         .o_iocontrol            = lmv_iocontrol,
3179         .o_fid_delete           = lmv_fid_delete,
3180         .o_quotacheck           = lmv_quotacheck,
3181         .o_quotactl             = lmv_quotactl
3182 };
3183
3184 struct md_ops lmv_md_ops = {
3185         .m_getstatus            = lmv_getstatus,
3186         .m_change_cbdata        = lmv_change_cbdata,
3187         .m_find_cbdata          = lmv_find_cbdata,
3188         .m_close                = lmv_close,
3189         .m_create               = lmv_create,
3190         .m_done_writing         = lmv_done_writing,
3191         .m_enqueue              = lmv_enqueue,
3192         .m_getattr              = lmv_getattr,
3193         .m_getxattr             = lmv_getxattr,
3194         .m_getattr_name         = lmv_getattr_name,
3195         .m_intent_lock          = lmv_intent_lock,
3196         .m_link                 = lmv_link,
3197         .m_rename               = lmv_rename,
3198         .m_setattr              = lmv_setattr,
3199         .m_setxattr             = lmv_setxattr,
3200         .m_sync                 = lmv_sync,
3201         .m_readpage             = lmv_readpage,
3202         .m_unlink               = lmv_unlink,
3203         .m_init_ea_size         = lmv_init_ea_size,
3204         .m_cancel_unused        = lmv_cancel_unused,
3205         .m_set_lock_data        = lmv_set_lock_data,
3206         .m_lock_match           = lmv_lock_match,
3207         .m_get_lustre_md        = lmv_get_lustre_md,
3208         .m_free_lustre_md       = lmv_free_lustre_md,
3209         .m_set_open_replay_data = lmv_set_open_replay_data,
3210         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3211         .m_renew_capa           = lmv_renew_capa,
3212         .m_unpack_capa          = lmv_unpack_capa,
3213         .m_get_remote_perm      = lmv_get_remote_perm,
3214         .m_intent_getattr_async = lmv_intent_getattr_async,
3215         .m_revalidate_lock      = lmv_revalidate_lock
3216 };
3217
3218 int __init lmv_init(void)
3219 {
3220         struct lprocfs_static_vars lvars;
3221         int                        rc;
3222
3223         lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3224                                                 sizeof(struct lmv_object),
3225                                                 0, 0);
3226         if (!lmv_object_cache) {
3227                 CERROR("Error allocating lmv objects cache\n");
3228                 return -ENOMEM;
3229         }
3230
3231         lprocfs_lmv_init_vars(&lvars);
3232
3233         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3234                                  lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3235         if (rc)
3236                 cfs_mem_cache_destroy(lmv_object_cache);
3237
3238         return rc;
3239 }
3240
3241 #ifdef __KERNEL__
3242 static void lmv_exit(void)
3243 {
3244         class_unregister_type(LUSTRE_LMV_NAME);
3245
3246         LASSERTF(cfs_atomic_read(&lmv_object_count) == 0,
3247                  "Can't free lmv objects cache, %d object(s) busy\n",
3248                  cfs_atomic_read(&lmv_object_count));
3249         cfs_mem_cache_destroy(lmv_object_cache);
3250 }
3251
3252 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3253 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3254 MODULE_LICENSE("GPL");
3255
3256 module_init(lmv_init);
3257 module_exit(lmv_exit);
3258 #endif