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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_LMV
  34
  35 #include <linux/file.h>
  36 #include <linux/module.h>
  37 #include <linux/init.h>
  38 #include <linux/user_namespace.h>
  39 #ifdef HAVE_UIDGID_HEADER
  40 # include <linux/uidgid.h>
  41 #endif
  42 #include <linux/slab.h>
  43 #include <linux/pagemap.h>
  44 #include <linux/mm.h>
  45 #include <linux/math64.h>
  46 #include <linux/seq_file.h>
  47 #include <linux/namei.h>
  48
  49 #include <obd_support.h>
  50 #include <lustre_lib.h>
  51 #include <lustre_net.h>
  52 #include <obd_class.h>
  53 #include <lustre_lmv.h>
  54 #include <lprocfs_status.h>
  55 #include <cl_object.h>
  56 #include <lustre_fid.h>
  57 #include <uapi/linux/lustre/lustre_ioctl.h>
  58 #include <lustre_kernelcomm.h>
  59 #include "lmv_internal.h"
  60
  61 static int lmv_check_connect(struct obd_device *obd);
  62
  63 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
  64                          int activate)
  65 {
  66         if (tgt->ltd_active == activate)
  67                 return;
  68
  69         tgt->ltd_active = activate;
  70         lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
  71
  72         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
  73 }
  74
  75 /**
  76  * Error codes:
  77  *
  78  *  -EINVAL  : UUID can't be found in the LMV's target list
  79  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  80  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  81  */
  82 static int lmv_set_mdc_active(struct lmv_obd *lmv,
  83                               const struct obd_uuid *uuid,
  84                               int activate)
  85 {
  86         struct lmv_tgt_desc     *tgt = NULL;
  87         struct obd_device       *obd;
  88         __u32                    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         spin_lock(&lmv->lmv_lock);
  96         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
  97                 tgt = lmv->tgts[i];
  98                 if (tgt == NULL || tgt->ltd_exp == NULL)
  99                         continue;
 100
 101                 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n", i,
 102                        tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
 103
 104                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 105                         break;
 106         }
 107
 108         if (i == lmv->desc.ld_tgt_count)
 109                 GOTO(out_lmv_lock, rc = -EINVAL);
 110
 111         obd = class_exp2obd(tgt->ltd_exp);
 112         if (obd == NULL)
 113                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 114
 115         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 116                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 117                obd->obd_type->typ_name, i);
 118         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 119
 120         if (tgt->ltd_active == activate) {
 121                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 122                        activate ? "" : "in");
 123                 GOTO(out_lmv_lock, rc);
 124         }
 125
 126         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 127                activate ? "" : "in");
 128         lmv_activate_target(lmv, tgt, activate);
 129         EXIT;
 130
 131  out_lmv_lock:
 132         spin_unlock(&lmv->lmv_lock);
 133         return rc;
 134 }
 135
 136 struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
 137 {
 138         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
 139         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
 140
 141         return (tgt == NULL) ? NULL : obd_get_uuid(tgt->ltd_exp);
 142 }
 143
 144 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 145                       enum obd_notify_event ev)
 146 {
 147         struct obd_connect_data *conn_data;
 148         struct lmv_obd          *lmv = &obd->u.lmv;
 149         struct obd_uuid         *uuid;
 150         int                      rc = 0;
 151         ENTRY;
 152
 153         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 154                 CERROR("unexpected notification of %s %s!\n",
 155                        watched->obd_type->typ_name,
 156                        watched->obd_name);
 157                 RETURN(-EINVAL);
 158         }
 159
 160         uuid = &watched->u.cli.cl_target_uuid;
 161         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 162                 /*
 163                  * Set MDC as active before notifying the observer, so the
 164                  * observer can use the MDC normally.
 165                  */
 166                 rc = lmv_set_mdc_active(lmv, uuid,
 167                                         ev == OBD_NOTIFY_ACTIVE);
 168                 if (rc) {
 169                         CERROR("%sactivation of %s failed: %d\n",
 170                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 171                                uuid->uuid, rc);
 172                         RETURN(rc);
 173                 }
 174         } else if (ev == OBD_NOTIFY_OCD) {
 175                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 176                 /*
 177                  * XXX: Make sure that ocd_connect_flags from all targets are
 178                  * the same. Otherwise one of MDTs runs wrong version or
 179                  * something like this.  --umka
 180                  */
 181                 obd->obd_self_export->exp_connect_data = *conn_data;
 182         }
 183
 184         /*
 185          * Pass the notification up the chain.
 186          */
 187         if (obd->obd_observer)
 188                 rc = obd_notify(obd->obd_observer, watched, ev);
 189
 190         RETURN(rc);
 191 }
 192
 193 static int lmv_connect(const struct lu_env *env,
 194                        struct obd_export **pexp, struct obd_device *obd,
 195                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 196                        void *localdata)
 197 {
 198         struct lmv_obd *lmv = &obd->u.lmv;
 199         struct lustre_handle conn = { 0 };
 200         struct obd_export *exp;
 201         int rc;
 202         ENTRY;
 203
 204         rc = class_connect(&conn, obd, cluuid);
 205         if (rc) {
 206                 CERROR("class_connection() returned %d\n", rc);
 207                 RETURN(rc);
 208         }
 209
 210         exp = class_conn2export(&conn);
 211
 212         lmv->connected = 0;
 213         lmv->conn_data = *data;
 214         lmv->lmv_cache = localdata;
 215
 216         lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
 217                                                     &obd->obd_kset.kobj);
 218         if (!lmv->lmv_tgts_kobj) {
 219                 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
 220                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
 221         }
 222
 223         rc = lmv_check_connect(obd);
 224         if (rc != 0)
 225                 GOTO(out_sysfs, rc);
 226
 227         *pexp = exp;
 228
 229         RETURN(rc);
 230
 231 out_sysfs:
 232         if (lmv->lmv_tgts_kobj)
 233                 kobject_put(lmv->lmv_tgts_kobj);
 234
 235         class_disconnect(exp);
 236
 237         return rc;
 238 }
 239
 240 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
 241                             __u32 def_easize)
 242 {
 243         struct obd_device       *obd = exp->exp_obd;
 244         struct lmv_obd          *lmv = &obd->u.lmv;
 245         __u32                    i;
 246         int                      rc = 0;
 247         int                      change = 0;
 248         ENTRY;
 249
 250         if (lmv->max_easize < easize) {
 251                 lmv->max_easize = easize;
 252                 change = 1;
 253         }
 254         if (lmv->max_def_easize < def_easize) {
 255                 lmv->max_def_easize = def_easize;
 256                 change = 1;
 257         }
 258
 259         if (change == 0)
 260                 RETURN(0);
 261
 262         if (lmv->connected == 0)
 263                 RETURN(0);
 264
 265         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 266                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 267
 268                 if (tgt == NULL || tgt->ltd_exp == NULL) {
 269                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
 270                         continue;
 271                 }
 272                 if (!tgt->ltd_active)
 273                         continue;
 274
 275                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
 276                 if (rc) {
 277                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
 278                                " rc = %d\n", obd->obd_name, i, rc);
 279                         break;
 280                 }
 281         }
 282         RETURN(rc);
 283 }
 284
 285 #define MAX_STRING_SIZE 128
 286
 287 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 288 {
 289         struct lmv_obd          *lmv = &obd->u.lmv;
 290         struct obd_device       *mdc_obd;
 291         struct obd_export       *mdc_exp;
 292         struct lu_fld_target     target;
 293         int                      rc;
 294         ENTRY;
 295
 296         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 297                                         &obd->obd_uuid);
 298         if (!mdc_obd) {
 299                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 300                 RETURN(-EINVAL);
 301         }
 302
 303         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
 304                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 305                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
 306
 307         if (!mdc_obd->obd_set_up) {
 308                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 309                 RETURN(-EINVAL);
 310         }
 311
 312         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
 313                          &lmv->conn_data, lmv->lmv_cache);
 314         if (rc) {
 315                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 316                 RETURN(rc);
 317         }
 318
 319         /*
 320          * Init fid sequence client for this mdc and add new fld target.
 321          */
 322         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
 323         if (rc)
 324                 RETURN(rc);
 325
 326         target.ft_srv = NULL;
 327         target.ft_exp = mdc_exp;
 328         target.ft_idx = tgt->ltd_index;
 329
 330         fld_client_add_target(&lmv->lmv_fld, &target);
 331
 332         rc = obd_register_observer(mdc_obd, obd);
 333         if (rc) {
 334                 obd_disconnect(mdc_exp);
 335                 CERROR("target %s register_observer error %d\n",
 336                        tgt->ltd_uuid.uuid, rc);
 337                 RETURN(rc);
 338         }
 339
 340         if (obd->obd_observer) {
 341                 /*
 342                  * Tell the observer about the new target.
 343                  */
 344                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 345                                 OBD_NOTIFY_ACTIVE);
 346                 if (rc) {
 347                         obd_disconnect(mdc_exp);
 348                         RETURN(rc);
 349                 }
 350         }
 351
 352         tgt->ltd_active = 1;
 353         tgt->ltd_exp = mdc_exp;
 354         lmv->desc.ld_active_tgt_count++;
 355
 356         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
 357
 358         rc = lqos_add_tgt(&lmv->lmv_qos, tgt);
 359         if (rc) {
 360                 obd_disconnect(mdc_exp);
 361                 RETURN(rc);
 362         }
 363
 364         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 365                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 366                 atomic_read(&obd->obd_refcount));
 367
 368         lmv_statfs_check_update(obd, tgt);
 369
 370         if (lmv->lmv_tgts_kobj)
 371                 /* Even if we failed to create the link, that's fine */
 372                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
 373                                        &mdc_obd->obd_kset.kobj,
 374                                        mdc_obd->obd_name);
 375         RETURN(0);
 376 }
 377
 378 static void lmv_del_target(struct lmv_obd *lmv, int index)
 379 {
 380         if (lmv->tgts[index] == NULL)
 381                 return;
 382
 383         lqos_del_tgt(&lmv->lmv_qos, lmv->tgts[index]);
 384
 385         OBD_FREE_PTR(lmv->tgts[index]);
 386         lmv->tgts[index] = NULL;
 387         return;
 388 }
 389
 390 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 391                            __u32 index, int gen)
 392 {
 393         struct obd_device *mdc_obd;
 394         struct lmv_obd *lmv = &obd->u.lmv;
 395         struct lmv_tgt_desc *tgt;
 396         int orig_tgt_count = 0;
 397         int rc = 0;
 398
 399         ENTRY;
 400
 401         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
 402         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
 403                                         &obd->obd_uuid);
 404         if (!mdc_obd) {
 405                 CERROR("%s: Target %s not attached: rc = %d\n",
 406                        obd->obd_name, uuidp->uuid, -EINVAL);
 407                 RETURN(-EINVAL);
 408         }
 409
 410         mutex_lock(&lmv->lmv_init_mutex);
 411         if ((index < lmv->tgts_size) && (lmv->tgts[index] != NULL)) {
 412                 tgt = lmv->tgts[index];
 413                 CERROR("%s: UUID %s already assigned at LMV target index %d:"
 414                        " rc = %d\n", obd->obd_name,
 415                        obd_uuid2str(&tgt->ltd_uuid), index, -EEXIST);
 416                 mutex_unlock(&lmv->lmv_init_mutex);
 417                 RETURN(-EEXIST);
 418         }
 419
 420         if (index >= lmv->tgts_size) {
 421                 /* We need to reallocate the lmv target array. */
 422                 struct lmv_tgt_desc **newtgts, **old = NULL;
 423                 __u32 newsize = 1;
 424                 __u32 oldsize = 0;
 425
 426                 while (newsize < index + 1)
 427                         newsize = newsize << 1;
 428                 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
 429                 if (newtgts == NULL) {
 430                         mutex_unlock(&lmv->lmv_init_mutex);
 431                         RETURN(-ENOMEM);
 432                 }
 433
 434                 if (lmv->tgts_size) {
 435                         memcpy(newtgts, lmv->tgts,
 436                                sizeof(*newtgts) * lmv->tgts_size);
 437                         old = lmv->tgts;
 438                         oldsize = lmv->tgts_size;
 439                 }
 440
 441                 lmv->tgts = newtgts;
 442                 lmv->tgts_size = newsize;
 443                 smp_rmb();
 444                 if (old)
 445                         OBD_FREE(old, sizeof(*old) * oldsize);
 446
 447                 CDEBUG(D_CONFIG, "tgts: %p size: %d\n", lmv->tgts,
 448                        lmv->tgts_size);
 449         }
 450
 451         OBD_ALLOC_PTR(tgt);
 452         if (!tgt) {
 453                 mutex_unlock(&lmv->lmv_init_mutex);
 454                 RETURN(-ENOMEM);
 455         }
 456
 457         mutex_init(&tgt->ltd_fid_mutex);
 458         tgt->ltd_index = index;
 459         tgt->ltd_uuid = *uuidp;
 460         tgt->ltd_active = 0;
 461         lmv->tgts[index] = tgt;
 462         if (index >= lmv->desc.ld_tgt_count) {
 463                 orig_tgt_count = lmv->desc.ld_tgt_count;
 464                 lmv->desc.ld_tgt_count = index + 1;
 465         }
 466
 467         if (lmv->connected == 0) {
 468                 /* lmv_check_connect() will connect this target. */
 469                 mutex_unlock(&lmv->lmv_init_mutex);
 470                 RETURN(0);
 471         }
 472
 473         /* Otherwise let's connect it ourselves */
 474         mutex_unlock(&lmv->lmv_init_mutex);
 475         rc = lmv_connect_mdc(obd, tgt);
 476         if (rc != 0) {
 477                 spin_lock(&lmv->lmv_lock);
 478                 if (lmv->desc.ld_tgt_count == index + 1)
 479                         lmv->desc.ld_tgt_count = orig_tgt_count;
 480                 memset(tgt, 0, sizeof(*tgt));
 481                 spin_unlock(&lmv->lmv_lock);
 482         } else {
 483                 int easize = sizeof(struct lmv_stripe_md) +
 484                         lmv->desc.ld_tgt_count * sizeof(struct lu_fid);
 485                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
 486         }
 487
 488         RETURN(rc);
 489 }
 490
 491 static int lmv_check_connect(struct obd_device *obd)
 492 {
 493         struct lmv_obd          *lmv = &obd->u.lmv;
 494         struct lmv_tgt_desc     *tgt;
 495         __u32                    i;
 496         int                      rc;
 497         int                      easize;
 498         ENTRY;
 499
 500         if (lmv->connected)
 501                 RETURN(0);
 502
 503         mutex_lock(&lmv->lmv_init_mutex);
 504         if (lmv->connected) {
 505                 mutex_unlock(&lmv->lmv_init_mutex);
 506                 RETURN(0);
 507         }
 508
 509         if (lmv->desc.ld_tgt_count == 0) {
 510                 mutex_unlock(&lmv->lmv_init_mutex);
 511                 CERROR("%s: no targets configured.\n", obd->obd_name);
 512                 RETURN(-EINVAL);
 513         }
 514
 515         LASSERT(lmv->tgts != NULL);
 516
 517         if (lmv->tgts[0] == NULL) {
 518                 mutex_unlock(&lmv->lmv_init_mutex);
 519                 CERROR("%s: no target configured for index 0.\n",
 520                        obd->obd_name);
 521                 RETURN(-EINVAL);
 522         }
 523
 524         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 525                obd->obd_uuid.uuid, obd->obd_name);
 526
 527         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 528                 tgt = lmv->tgts[i];
 529                 if (tgt == NULL)
 530                         continue;
 531                 rc = lmv_connect_mdc(obd, tgt);
 532                 if (rc)
 533                         GOTO(out_disc, rc);
 534         }
 535
 536         lmv->connected = 1;
 537         easize = lmv_mds_md_size(lmv->desc.ld_tgt_count, LMV_MAGIC);
 538         lmv_init_ea_size(obd->obd_self_export, easize, 0);
 539         mutex_unlock(&lmv->lmv_init_mutex);
 540         RETURN(0);
 541
 542  out_disc:
 543         while (i-- > 0) {
 544                 int rc2;
 545                 tgt = lmv->tgts[i];
 546                 if (tgt == NULL)
 547                         continue;
 548                 tgt->ltd_active = 0;
 549                 if (tgt->ltd_exp) {
 550                         --lmv->desc.ld_active_tgt_count;
 551                         rc2 = obd_disconnect(tgt->ltd_exp);
 552                         if (rc2) {
 553                                 CERROR("LMV target %s disconnect on "
 554                                        "MDC idx %d: error %d\n",
 555                                        tgt->ltd_uuid.uuid, i, rc2);
 556                         }
 557                 }
 558         }
 559
 560         mutex_unlock(&lmv->lmv_init_mutex);
 561
 562         RETURN(rc);
 563 }
 564
 565 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 566 {
 567         struct lmv_obd         *lmv = &obd->u.lmv;
 568         struct obd_device      *mdc_obd;
 569         int                     rc;
 570         ENTRY;
 571
 572         LASSERT(tgt != NULL);
 573         LASSERT(obd != NULL);
 574
 575         mdc_obd = class_exp2obd(tgt->ltd_exp);
 576
 577         if (mdc_obd) {
 578                 mdc_obd->obd_force = obd->obd_force;
 579                 mdc_obd->obd_fail = obd->obd_fail;
 580                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 581
 582                 if (lmv->lmv_tgts_kobj)
 583                         sysfs_remove_link(lmv->lmv_tgts_kobj,
 584                                           mdc_obd->obd_name);
 585         }
 586
 587         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
 588         if (rc)
 589                 CERROR("Can't finanize fids factory\n");
 590
 591         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 592                tgt->ltd_exp->exp_obd->obd_name,
 593                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 594
 595         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 596         rc = obd_disconnect(tgt->ltd_exp);
 597         if (rc) {
 598                 if (tgt->ltd_active) {
 599                         CERROR("Target %s disconnect error %d\n",
 600                                tgt->ltd_uuid.uuid, rc);
 601                 }
 602         }
 603
 604         lmv_activate_target(lmv, tgt, 0);
 605         tgt->ltd_exp = NULL;
 606         RETURN(0);
 607 }
 608
 609 static int lmv_disconnect(struct obd_export *exp)
 610 {
 611         struct obd_device       *obd = class_exp2obd(exp);
 612         struct lmv_obd          *lmv = &obd->u.lmv;
 613         int                      rc;
 614         __u32                    i;
 615         ENTRY;
 616
 617         if (!lmv->tgts)
 618                 goto out_local;
 619
 620         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 621                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
 622                         continue;
 623
 624                 lmv_disconnect_mdc(obd, lmv->tgts[i]);
 625         }
 626
 627         if (lmv->lmv_tgts_kobj)
 628                 kobject_put(lmv->lmv_tgts_kobj);
 629
 630 out_local:
 631         /*
 632          * This is the case when no real connection is established by
 633          * lmv_check_connect().
 634          */
 635         if (!lmv->connected)
 636                 class_export_put(exp);
 637         rc = class_disconnect(exp);
 638         lmv->connected = 0;
 639
 640         RETURN(rc);
 641 }
 642
 643 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
 644                         void __user *uarg)
 645 {
 646         struct obd_device       *obddev = class_exp2obd(exp);
 647         struct lmv_obd          *lmv = &obddev->u.lmv;
 648         struct getinfo_fid2path *gf;
 649         struct lmv_tgt_desc     *tgt;
 650         struct getinfo_fid2path *remote_gf = NULL;
 651         struct lu_fid           root_fid;
 652         int                     remote_gf_size = 0;
 653         int                     rc;
 654
 655         gf = karg;
 656         tgt = lmv_find_target(lmv, &gf->gf_fid);
 657         if (IS_ERR(tgt))
 658                 RETURN(PTR_ERR(tgt));
 659
 660         root_fid = *gf->gf_u.gf_root_fid;
 661         LASSERT(fid_is_sane(&root_fid));
 662
 663 repeat_fid2path:
 664         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
 665         if (rc != 0 && rc != -EREMOTE)
 666                 GOTO(out_fid2path, rc);
 667
 668         /* If remote_gf != NULL, it means just building the
 669          * path on the remote MDT, copy this path segement to gf */
 670         if (remote_gf != NULL) {
 671                 struct getinfo_fid2path *ori_gf;
 672                 char *ptr;
 673                 int len;
 674
 675                 ori_gf = (struct getinfo_fid2path *)karg;
 676                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
 677                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
 678                         GOTO(out_fid2path, rc = -EOVERFLOW);
 679
 680                 ptr = ori_gf->gf_u.gf_path;
 681
 682                 len = strlen(gf->gf_u.gf_path);
 683                 /* move the current path to the right to release space
 684                  * for closer-to-root part */
 685                 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
 686                 memcpy(ptr, gf->gf_u.gf_path, len);
 687                 ptr[len] = '/';
 688         }
 689
 690         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
 691                tgt->ltd_exp->exp_obd->obd_name,
 692                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
 693                gf->gf_linkno);
 694
 695         if (rc == 0)
 696                 GOTO(out_fid2path, rc);
 697
 698         /* sigh, has to go to another MDT to do path building further */
 699         if (remote_gf == NULL) {
 700                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
 701                 OBD_ALLOC(remote_gf, remote_gf_size);
 702                 if (remote_gf == NULL)
 703                         GOTO(out_fid2path, rc = -ENOMEM);
 704                 remote_gf->gf_pathlen = PATH_MAX;
 705         }
 706
 707         if (!fid_is_sane(&gf->gf_fid)) {
 708                 CERROR("%s: invalid FID "DFID": rc = %d\n",
 709                        tgt->ltd_exp->exp_obd->obd_name,
 710                        PFID(&gf->gf_fid), -EINVAL);
 711                 GOTO(out_fid2path, rc = -EINVAL);
 712         }
 713
 714         tgt = lmv_find_target(lmv, &gf->gf_fid);
 715         if (IS_ERR(tgt))
 716                 GOTO(out_fid2path, rc = -EINVAL);
 717
 718         remote_gf->gf_fid = gf->gf_fid;
 719         remote_gf->gf_recno = -1;
 720         remote_gf->gf_linkno = -1;
 721         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
 722         *remote_gf->gf_u.gf_root_fid = root_fid;
 723         gf = remote_gf;
 724         goto repeat_fid2path;
 725
 726 out_fid2path:
 727         if (remote_gf != NULL)
 728                 OBD_FREE(remote_gf, remote_gf_size);
 729         RETURN(rc);
 730 }
 731
 732 static int lmv_hsm_req_count(struct lmv_obd *lmv,
 733                              const struct hsm_user_request *hur,
 734                              const struct lmv_tgt_desc *tgt_mds)
 735 {
 736         __u32                    i;
 737         int                      nr = 0;
 738         struct lmv_tgt_desc     *curr_tgt;
 739
 740         /* count how many requests must be sent to the given target */
 741         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
 742                 curr_tgt = lmv_find_target(lmv, &hur->hur_user_item[i].hui_fid);
 743                 if (IS_ERR(curr_tgt))
 744                         RETURN(PTR_ERR(curr_tgt));
 745                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
 746                         nr++;
 747         }
 748         return nr;
 749 }
 750
 751 static int lmv_hsm_req_build(struct lmv_obd *lmv,
 752                               struct hsm_user_request *hur_in,
 753                               const struct lmv_tgt_desc *tgt_mds,
 754                               struct hsm_user_request *hur_out)
 755 {
 756         __u32                    i, nr_out;
 757         struct lmv_tgt_desc     *curr_tgt;
 758
 759         /* build the hsm_user_request for the given target */
 760         hur_out->hur_request = hur_in->hur_request;
 761         nr_out = 0;
 762         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
 763                 curr_tgt = lmv_find_target(lmv,
 764                                            &hur_in->hur_user_item[i].hui_fid);
 765                 if (IS_ERR(curr_tgt))
 766                         RETURN(PTR_ERR(curr_tgt));
 767                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
 768                         hur_out->hur_user_item[nr_out] =
 769                                                 hur_in->hur_user_item[i];
 770                         nr_out++;
 771                 }
 772         }
 773         hur_out->hur_request.hr_itemcount = nr_out;
 774         memcpy(hur_data(hur_out), hur_data(hur_in),
 775                hur_in->hur_request.hr_data_len);
 776
 777         RETURN(0);
 778 }
 779
 780 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
 781                                  int len, struct lustre_kernelcomm *lk,
 782                                  void __user *uarg)
 783 {
 784         struct lmv_obd *lmv = &obd->u.lmv;
 785         __u32   i;
 786         int     rc;
 787         ENTRY;
 788
 789         /* unregister request (call from llapi_hsm_copytool_fini) */
 790         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 791                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 792
 793                 if (tgt == NULL || tgt->ltd_exp == NULL)
 794                         continue;
 795                 /* best effort: try to clean as much as possible
 796                  * (continue on error) */
 797                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 798         }
 799
 800         /* Whatever the result, remove copytool from kuc groups.
 801          * Unreached coordinators will get EPIPE on next requests
 802          * and will unregister automatically.
 803          */
 804         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 805
 806         RETURN(rc);
 807 }
 808
 809 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
 810                                int len, struct lustre_kernelcomm *lk,
 811                                void __user *uarg)
 812 {
 813         struct lmv_obd *lmv = &obd->u.lmv;
 814         struct file *filp;
 815         __u32 i, j;
 816         int err;
 817         bool any_set = false;
 818         struct kkuc_ct_data *kcd;
 819         size_t kcd_size;
 820         int rc = 0;
 821         ENTRY;
 822
 823         filp = fget(lk->lk_wfd);
 824         if (!filp)
 825                 RETURN(-EBADF);
 826
 827         if (lk->lk_flags & LK_FLG_DATANR)
 828                 kcd_size = offsetof(struct kkuc_ct_data,
 829                                     kcd_archives[lk->lk_data_count]);
 830         else
 831                 kcd_size = sizeof(*kcd);
 832
 833         OBD_ALLOC(kcd, kcd_size);
 834         if (kcd == NULL)
 835                 GOTO(err_fput, rc = -ENOMEM);
 836
 837         kcd->kcd_nr_archives = lk->lk_data_count;
 838         if (lk->lk_flags & LK_FLG_DATANR) {
 839                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
 840                 if (lk->lk_data_count > 0)
 841                         memcpy(kcd->kcd_archives, lk->lk_data,
 842                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
 843         } else {
 844                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
 845         }
 846
 847         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
 848                                    lk->lk_group, kcd, kcd_size);
 849         OBD_FREE(kcd, kcd_size);
 850         if (rc)
 851                 GOTO(err_fput, rc);
 852
 853         /* All or nothing: try to register to all MDS.
 854          * In case of failure, unregister from previous MDS,
 855          * except if it because of inactive target. */
 856         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 857                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 858
 859                 if (tgt == NULL || tgt->ltd_exp == NULL)
 860                         continue;
 861
 862                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 863                 if (err) {
 864                         if (tgt->ltd_active) {
 865                                 /* permanent error */
 866                                 CERROR("%s: iocontrol MDC %s on MDT"
 867                                        " idx %d cmd %x: err = %d\n",
 868                                        lmv2obd_dev(lmv)->obd_name,
 869                                        tgt->ltd_uuid.uuid, i, cmd, err);
 870                                 rc = err;
 871                                 lk->lk_flags |= LK_FLG_STOP;
 872                                 /* unregister from previous MDS */
 873                                 for (j = 0; j < i; j++) {
 874                                         tgt = lmv->tgts[j];
 875                                         if (tgt == NULL || tgt->ltd_exp == NULL)
 876                                                 continue;
 877                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
 878                                                       lk, uarg);
 879                                 }
 880                                 GOTO(err_kkuc_rem, rc);
 881                         }
 882                         /* else: transient error.
 883                          * kuc will register to the missing MDT
 884                          * when it is back */
 885                 } else {
 886                         any_set = true;
 887                 }
 888         }
 889
 890         if (!any_set)
 891                 /* no registration done: return error */
 892                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
 893
 894         RETURN(0);
 895
 896 err_kkuc_rem:
 897         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 898
 899 err_fput:
 900         fput(filp);
 901         return rc;
 902 }
 903
 904
 905
 906
 907 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 908                          int len, void *karg, void __user *uarg)
 909 {
 910         struct obd_device       *obddev = class_exp2obd(exp);
 911         struct lmv_obd          *lmv = &obddev->u.lmv;
 912         struct lmv_tgt_desc     *tgt = NULL;
 913         __u32                    i = 0;
 914         int                      rc = 0;
 915         int                      set = 0;
 916         __u32                    count = lmv->desc.ld_tgt_count;
 917         ENTRY;
 918
 919         if (count == 0)
 920                 RETURN(-ENOTTY);
 921
 922         switch (cmd) {
 923         case IOC_OBD_STATFS: {
 924                 struct obd_ioctl_data *data = karg;
 925                 struct obd_device *mdc_obd;
 926                 struct obd_statfs stat_buf = {0};
 927                 __u32 index;
 928
 929                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 930                 if ((index >= count))
 931                         RETURN(-ENODEV);
 932
 933                 tgt = lmv->tgts[index];
 934                 if (tgt == NULL || !tgt->ltd_active)
 935                         RETURN(-ENODATA);
 936
 937                 mdc_obd = class_exp2obd(tgt->ltd_exp);
 938                 if (!mdc_obd)
 939                         RETURN(-EINVAL);
 940
 941                 /* copy UUID */
 942                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 943                                  min((int) data->ioc_plen2,
 944                                      (int) sizeof(struct obd_uuid))))
 945                         RETURN(-EFAULT);
 946
 947                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
 948                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
 949                                 0);
 950                 if (rc)
 951                         RETURN(rc);
 952                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
 953                                  min((int) data->ioc_plen1,
 954                                      (int) sizeof(stat_buf))))
 955                         RETURN(-EFAULT);
 956                 break;
 957         }
 958         case OBD_IOC_QUOTACTL: {
 959                 struct if_quotactl *qctl = karg;
 960                 struct obd_quotactl *oqctl;
 961
 962                 if (qctl->qc_valid == QC_MDTIDX) {
 963                         if (count <= qctl->qc_idx)
 964                                 RETURN(-EINVAL);
 965
 966                         tgt = lmv->tgts[qctl->qc_idx];
 967                         if (tgt == NULL || tgt->ltd_exp == NULL)
 968                                 RETURN(-EINVAL);
 969                 } else if (qctl->qc_valid == QC_UUID) {
 970                         for (i = 0; i < count; i++) {
 971                                 tgt = lmv->tgts[i];
 972                                 if (tgt == NULL)
 973                                         continue;
 974                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 975                                                      &qctl->obd_uuid))
 976                                         continue;
 977
 978                                 if (tgt->ltd_exp == NULL)
 979                                         RETURN(-EINVAL);
 980
 981                                 break;
 982                         }
 983                 } else {
 984                         RETURN(-EINVAL);
 985                 }
 986
 987                 if (i >= count)
 988                         RETURN(-EAGAIN);
 989
 990                 LASSERT(tgt != NULL && tgt->ltd_exp != NULL);
 991                 OBD_ALLOC_PTR(oqctl);
 992                 if (!oqctl)
 993                         RETURN(-ENOMEM);
 994
 995                 QCTL_COPY(oqctl, qctl);
 996                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 997                 if (rc == 0) {
 998                         QCTL_COPY(qctl, oqctl);
 999                         qctl->qc_valid = QC_MDTIDX;
1000                         qctl->obd_uuid = tgt->ltd_uuid;
1001                 }
1002                 OBD_FREE_PTR(oqctl);
1003                 break;
1004         }
1005         case LL_IOC_GET_CONNECT_FLAGS: {
1006                 tgt = lmv->tgts[0];
1007                 if (tgt == NULL || tgt->ltd_exp == NULL)
1008                         RETURN(-ENODATA);
1009                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1010                 break;
1011         }
1012         case LL_IOC_FID2MDTIDX: {
1013                 struct lu_fid *fid = karg;
1014                 int             mdt_index;
1015
1016                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
1017                 if (rc != 0)
1018                         RETURN(rc);
1019
1020                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
1021                  * point to user space memory for FID2MDTIDX. */
1022                 *(__u32 *)uarg = mdt_index;
1023                 break;
1024         }
1025         case OBD_IOC_FID2PATH: {
1026                 rc = lmv_fid2path(exp, len, karg, uarg);
1027                 break;
1028         }
1029         case LL_IOC_HSM_STATE_GET:
1030         case LL_IOC_HSM_STATE_SET:
1031         case LL_IOC_HSM_ACTION: {
1032                 struct md_op_data       *op_data = karg;
1033
1034                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1035                 if (IS_ERR(tgt))
1036                         RETURN(PTR_ERR(tgt));
1037
1038                 if (tgt->ltd_exp == NULL)
1039                         RETURN(-EINVAL);
1040
1041                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1042                 break;
1043         }
1044         case LL_IOC_HSM_PROGRESS: {
1045                 const struct hsm_progress_kernel *hpk = karg;
1046
1047                 tgt = lmv_find_target(lmv, &hpk->hpk_fid);
1048                 if (IS_ERR(tgt))
1049                         RETURN(PTR_ERR(tgt));
1050                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1051                 break;
1052         }
1053         case LL_IOC_HSM_REQUEST: {
1054                 struct hsm_user_request *hur = karg;
1055                 unsigned int reqcount = hur->hur_request.hr_itemcount;
1056
1057                 if (reqcount == 0)
1058                         RETURN(0);
1059
1060                 /* if the request is about a single fid
1061                  * or if there is a single MDS, no need to split
1062                  * the request. */
1063                 if (reqcount == 1 || count == 1) {
1064                         tgt = lmv_find_target(lmv,
1065                                               &hur->hur_user_item[0].hui_fid);
1066                         if (IS_ERR(tgt))
1067                                 RETURN(PTR_ERR(tgt));
1068                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1069                 } else {
1070                         /* split fid list to their respective MDS */
1071                         for (i = 0; i < count; i++) {
1072                                 int nr, rc1;
1073                                 size_t reqlen;
1074                                 struct hsm_user_request *req;
1075
1076                                 tgt = lmv->tgts[i];
1077                                 if (tgt == NULL || tgt->ltd_exp == NULL)
1078                                         continue;
1079
1080                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
1081                                 if (nr < 0)
1082                                         RETURN(nr);
1083                                 if (nr == 0) /* nothing for this MDS */
1084                                         continue;
1085
1086                                 /* build a request with fids for this MDS */
1087                                 reqlen = offsetof(typeof(*hur),
1088                                                   hur_user_item[nr])
1089                                                 + hur->hur_request.hr_data_len;
1090                                 OBD_ALLOC_LARGE(req, reqlen);
1091                                 if (req == NULL)
1092                                         RETURN(-ENOMEM);
1093                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
1094                                 if (rc1 < 0)
1095                                         GOTO(hsm_req_err, rc1);
1096                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
1097                                                     req, uarg);
1098 hsm_req_err:
1099                                 if (rc1 != 0 && rc == 0)
1100                                         rc = rc1;
1101                                 OBD_FREE_LARGE(req, reqlen);
1102                         }
1103                 }
1104                 break;
1105         }
1106         case LL_IOC_LOV_SWAP_LAYOUTS: {
1107                 struct md_op_data       *op_data = karg;
1108                 struct lmv_tgt_desc     *tgt1, *tgt2;
1109
1110                 tgt1 = lmv_find_target(lmv, &op_data->op_fid1);
1111                 if (IS_ERR(tgt1))
1112                         RETURN(PTR_ERR(tgt1));
1113
1114                 tgt2 = lmv_find_target(lmv, &op_data->op_fid2);
1115                 if (IS_ERR(tgt2))
1116                         RETURN(PTR_ERR(tgt2));
1117
1118                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1119                         RETURN(-EINVAL);
1120
1121                 /* only files on same MDT can have their layouts swapped */
1122                 if (tgt1->ltd_index != tgt2->ltd_index)
1123                         RETURN(-EPERM);
1124
1125                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1126                 break;
1127         }
1128         case LL_IOC_HSM_CT_START: {
1129                 struct lustre_kernelcomm *lk = karg;
1130                 if (lk->lk_flags & LK_FLG_STOP)
1131                         rc = lmv_hsm_ct_unregister(obddev, cmd, len, lk, uarg);
1132                 else
1133                         rc = lmv_hsm_ct_register(obddev, cmd, len, lk, uarg);
1134                 break;
1135         }
1136         default:
1137                 for (i = 0; i < count; i++) {
1138                         struct obd_device *mdc_obd;
1139                         int err;
1140
1141                         tgt = lmv->tgts[i];
1142                         if (tgt == NULL || tgt->ltd_exp == NULL)
1143                                 continue;
1144                         /* ll_umount_begin() sets force flag but for lmv, not
1145                          * mdc. Let's pass it through */
1146                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1147                         mdc_obd->obd_force = obddev->obd_force;
1148                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1149                         if (err) {
1150                                 if (tgt->ltd_active) {
1151                                         CERROR("error: iocontrol MDC %s on MDT"
1152                                                " idx %d cmd %x: err = %d\n",
1153                                                tgt->ltd_uuid.uuid, i, cmd, err);
1154                                         if (!rc)
1155                                                 rc = err;
1156                                 }
1157                         } else
1158                                 set = 1;
1159                 }
1160                 if (!set && !rc)
1161                         rc = -EIO;
1162         }
1163         RETURN(rc);
1164 }
1165
1166 /**
1167  * This is _inode_ placement policy function (not name).
1168  */
1169 static u32 lmv_placement_policy(struct obd_device *obd,
1170                                 struct md_op_data *op_data)
1171 {
1172         struct lmv_obd *lmv = &obd->u.lmv;
1173         struct lmv_user_md *lum;
1174         u32 mdt;
1175
1176         ENTRY;
1177
1178         if (lmv->desc.ld_tgt_count == 1)
1179                 RETURN(0);
1180
1181         lum = op_data->op_data;
1182         /*
1183          * Choose MDT by
1184          * 1. See if the stripe offset is specified by lum.
1185          * 2. If parent has default LMV, and its hash type is "space", choose
1186          *    MDT with QoS. (see lmv_locate_tgt_qos()).
1187          * 3. Then check if default LMV stripe offset is not -1.
1188          * 4. Finally choose MDS by name hash if the parent
1189          *    is striped directory. (see lmv_locate_tgt()).
1190          *
1191          * presently explicit MDT location is not supported
1192          * for foreign dirs (as it can't be embedded into free
1193          * format LMV, like with lum_stripe_offset), so we only
1194          * rely on default stripe offset or then name hashing.
1195          */
1196         if (op_data->op_cli_flags & CLI_SET_MEA && lum != NULL &&
1197             le32_to_cpu(lum->lum_magic != LMV_MAGIC_FOREIGN) &&
1198             le32_to_cpu(lum->lum_stripe_offset) != (__u32)-1) {
1199                 mdt = le32_to_cpu(lum->lum_stripe_offset);
1200         } else if (op_data->op_code == LUSTRE_OPC_MKDIR &&
1201                    !lmv_dir_striped(op_data->op_mea1) &&
1202                    lmv_dir_qos_mkdir(op_data->op_default_mea1)) {
1203                 mdt = op_data->op_mds;
1204         } else if (op_data->op_code == LUSTRE_OPC_MKDIR &&
1205                    op_data->op_default_mea1 &&
1206                    op_data->op_default_mea1->lsm_md_master_mdt_index !=
1207                         (__u32)-1) {
1208                 mdt = op_data->op_default_mea1->lsm_md_master_mdt_index;
1209                 op_data->op_mds = mdt;
1210         } else {
1211                 mdt = op_data->op_mds;
1212         }
1213
1214         RETURN(mdt);
1215 }
1216
1217 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid, u32 mds)
1218 {
1219         struct lmv_tgt_desc     *tgt;
1220         int                      rc;
1221         ENTRY;
1222
1223         tgt = lmv_get_target(lmv, mds, NULL);
1224         if (IS_ERR(tgt))
1225                 RETURN(PTR_ERR(tgt));
1226
1227         /*
1228          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1229          * on server that seq in new allocated fid is not yet known.
1230          */
1231         mutex_lock(&tgt->ltd_fid_mutex);
1232
1233         if (tgt->ltd_active == 0 || tgt->ltd_exp == NULL)
1234                 GOTO(out, rc = -ENODEV);
1235
1236         /*
1237          * Asking underlying tgt layer to allocate new fid.
1238          */
1239         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1240         if (rc > 0) {
1241                 LASSERT(fid_is_sane(fid));
1242                 rc = 0;
1243         }
1244
1245         EXIT;
1246 out:
1247         mutex_unlock(&tgt->ltd_fid_mutex);
1248         return rc;
1249 }
1250
1251 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1252                   struct lu_fid *fid, struct md_op_data *op_data)
1253 {
1254         struct obd_device *obd = class_exp2obd(exp);
1255         struct lmv_obd *lmv = &obd->u.lmv;
1256         u32 mds;
1257         int rc;
1258
1259         ENTRY;
1260
1261         LASSERT(op_data != NULL);
1262         LASSERT(fid != NULL);
1263
1264         mds = lmv_placement_policy(obd, op_data);
1265
1266         rc = __lmv_fid_alloc(lmv, fid, mds);
1267         if (rc)
1268                 CERROR("Can't alloc new fid, rc %d\n", rc);
1269
1270         RETURN(rc);
1271 }
1272
1273 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1274 {
1275         struct lmv_obd *lmv = &obd->u.lmv;
1276         struct lmv_desc *desc;
1277         struct lnet_process_id lnet_id;
1278         int i = 0;
1279         int rc;
1280
1281         ENTRY;
1282
1283         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1284                 CERROR("LMV setup requires a descriptor\n");
1285                 RETURN(-EINVAL);
1286         }
1287
1288         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1289         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1290                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1291                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1292                 RETURN(-EINVAL);
1293         }
1294
1295         lmv->tgts_size = 32U;
1296         OBD_ALLOC(lmv->tgts, sizeof(*lmv->tgts) * lmv->tgts_size);
1297         if (lmv->tgts == NULL)
1298                 RETURN(-ENOMEM);
1299
1300         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1301         lmv->desc.ld_tgt_count = 0;
1302         lmv->desc.ld_active_tgt_count = 0;
1303         lmv->desc.ld_qos_maxage = LMV_DESC_QOS_MAXAGE_DEFAULT;
1304         lmv->max_def_easize = 0;
1305         lmv->max_easize = 0;
1306
1307         spin_lock_init(&lmv->lmv_lock);
1308         mutex_init(&lmv->lmv_init_mutex);
1309
1310         /* Set up allocation policy (QoS and RR) */
1311         INIT_LIST_HEAD(&lmv->lmv_qos.lq_svr_list);
1312         init_rwsem(&lmv->lmv_qos.lq_rw_sem);
1313         lmv->lmv_qos.lq_dirty = 1;
1314         lmv->lmv_qos.lq_rr.lqr_dirty = 1;
1315         lmv->lmv_qos.lq_reset = 1;
1316         /* Default priority is toward free space balance */
1317         lmv->lmv_qos.lq_prio_free = 232;
1318         /* Default threshold for rr (roughly 17%) */
1319         lmv->lmv_qos.lq_threshold_rr = 43;
1320
1321         /*
1322          * initialize rr_index to lower 32bit of netid, so that client
1323          * can distribute subdirs evenly from the beginning.
1324          */
1325         while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1326                 if (LNET_NETTYP(LNET_NIDNET(lnet_id.nid)) != LOLND) {
1327                         lmv->lmv_qos_rr_index = (u32)lnet_id.nid;
1328                         break;
1329                 }
1330         }
1331
1332         rc = lmv_tunables_init(obd);
1333         if (rc)
1334                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1335                       obd->obd_name, rc);
1336
1337         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1338                              LUSTRE_CLI_FLD_HASH_DHT);
1339         if (rc) {
1340                 CERROR("Can't init FLD, err %d\n", rc);
1341                 GOTO(out, rc);
1342         }
1343
1344         RETURN(0);
1345
1346 out:
1347         return rc;
1348 }
1349
1350 static int lmv_cleanup(struct obd_device *obd)
1351 {
1352         struct lmv_obd   *lmv = &obd->u.lmv;
1353         ENTRY;
1354
1355         fld_client_fini(&lmv->lmv_fld);
1356         if (lmv->tgts != NULL) {
1357                 int i;
1358                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1359                         if (lmv->tgts[i] == NULL)
1360                                 continue;
1361                         lmv_del_target(lmv, i);
1362                 }
1363                 OBD_FREE(lmv->tgts, sizeof(*lmv->tgts) * lmv->tgts_size);
1364                 lmv->tgts_size = 0;
1365         }
1366         RETURN(0);
1367 }
1368
1369 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1370 {
1371         struct lustre_cfg       *lcfg = buf;
1372         struct obd_uuid         obd_uuid;
1373         int                     gen;
1374         __u32                   index;
1375         int                     rc;
1376         ENTRY;
1377
1378         switch (lcfg->lcfg_command) {
1379         case LCFG_ADD_MDC:
1380                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1381                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1382                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1383                         GOTO(out, rc = -EINVAL);
1384
1385                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1386
1387                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1388                         GOTO(out, rc = -EINVAL);
1389                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1390                         GOTO(out, rc = -EINVAL);
1391                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1392                 GOTO(out, rc);
1393         default:
1394                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1395                 GOTO(out, rc = -EINVAL);
1396         }
1397 out:
1398         RETURN(rc);
1399 }
1400
1401 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1402 {
1403         int i;
1404
1405         if (flags & OBD_STATFS_FOR_MDT0)
1406                 return 0;
1407
1408         if (lmv->lmv_statfs_start || lmv->desc.ld_tgt_count == 1)
1409                 return lmv->lmv_statfs_start;
1410
1411         /* choose initial MDT for this client */
1412         for (i = 0;; i++) {
1413                 struct lnet_process_id lnet_id;
1414                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1415                         break;
1416
1417                 if (LNET_NETTYP(LNET_NIDNET(lnet_id.nid)) != LOLND) {
1418                         /* We dont need a full 64-bit modulus, just enough
1419                          * to distribute the requests across MDTs evenly.
1420                          */
1421                         lmv->lmv_statfs_start =
1422                                 (u32)lnet_id.nid % lmv->desc.ld_tgt_count;
1423                         break;
1424                 }
1425         }
1426
1427         return lmv->lmv_statfs_start;
1428 }
1429
1430 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1431                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1432 {
1433         struct obd_device       *obd = class_exp2obd(exp);
1434         struct lmv_obd          *lmv = &obd->u.lmv;
1435         struct obd_statfs       *temp;
1436         int                      rc = 0;
1437         __u32                    i, idx;
1438         ENTRY;
1439
1440         OBD_ALLOC(temp, sizeof(*temp));
1441         if (temp == NULL)
1442                 RETURN(-ENOMEM);
1443
1444         /* distribute statfs among MDTs */
1445         idx = lmv_select_statfs_mdt(lmv, flags);
1446
1447         for (i = 0; i < lmv->desc.ld_tgt_count; i++, idx++) {
1448                 idx = idx % lmv->desc.ld_tgt_count;
1449                 if (lmv->tgts[idx] == NULL || lmv->tgts[idx]->ltd_exp == NULL)
1450                         continue;
1451
1452                 rc = obd_statfs(env, lmv->tgts[idx]->ltd_exp, temp,
1453                                 max_age, flags);
1454                 if (rc) {
1455                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1456                                lmv->tgts[idx]->ltd_exp->exp_obd->obd_name, i,
1457                                rc);
1458                         GOTO(out_free_temp, rc);
1459                 }
1460
1461                 if (temp->os_state & OS_STATE_SUM ||
1462                     flags == OBD_STATFS_FOR_MDT0) {
1463                         /* reset to the last aggregated values
1464                          * and don't sum with non-aggrated data */
1465                         /* If the statfs is from mount, it needs to retrieve
1466                          * necessary information from MDT0. i.e. mount does
1467                          * not need the merged osfs from all of MDT. Also
1468                          * clients can be mounted as long as MDT0 is in
1469                          * service */
1470                         *osfs = *temp;
1471                         break;
1472                 }
1473
1474                 if (i == 0) {
1475                         *osfs = *temp;
1476                 } else {
1477                         osfs->os_bavail += temp->os_bavail;
1478                         osfs->os_blocks += temp->os_blocks;
1479                         osfs->os_ffree += temp->os_ffree;
1480                         osfs->os_files += temp->os_files;
1481                         osfs->os_granted += temp->os_granted;
1482                 }
1483         }
1484
1485         EXIT;
1486 out_free_temp:
1487         OBD_FREE(temp, sizeof(*temp));
1488         return rc;
1489 }
1490
1491 static int lmv_statfs_update(void *cookie, int rc)
1492 {
1493         struct obd_info *oinfo = cookie;
1494         struct obd_device *obd = oinfo->oi_obd;
1495         struct lmv_obd *lmv = &obd->u.lmv;
1496         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1497         struct obd_statfs *osfs = oinfo->oi_osfs;
1498
1499         /*
1500          * NB: don't deactivate TGT upon error, because we may not trigger async
1501          * statfs any longer, then there is no chance to activate TGT.
1502          */
1503         if (!rc) {
1504                 spin_lock(&lmv->lmv_lock);
1505                 tgt->ltd_statfs = *osfs;
1506                 tgt->ltd_statfs_age = ktime_get_seconds();
1507                 spin_unlock(&lmv->lmv_lock);
1508                 lmv->lmv_qos.lq_dirty = 1;
1509         }
1510
1511         return rc;
1512 }
1513
1514 /* update tgt statfs async if it's ld_qos_maxage old */
1515 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1516 {
1517         struct obd_info oinfo = {
1518                 .oi_obd = obd,
1519                 .oi_tgt = tgt,
1520                 .oi_cb_up = lmv_statfs_update,
1521         };
1522         int rc;
1523
1524         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1525             obd->u.lmv.desc.ld_qos_maxage)
1526                 return 0;
1527
1528         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1529
1530         return rc;
1531 }
1532
1533 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1534                         struct lu_fid *fid)
1535 {
1536         struct obd_device    *obd = exp->exp_obd;
1537         struct lmv_obd       *lmv = &obd->u.lmv;
1538         int                   rc;
1539         ENTRY;
1540
1541         rc = md_get_root(lmv->tgts[0]->ltd_exp, fileset, fid);
1542         RETURN(rc);
1543 }
1544
1545 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1546                         u64 obd_md_valid, const char *name, size_t buf_size,
1547                         struct ptlrpc_request **req)
1548 {
1549         struct obd_device      *obd = exp->exp_obd;
1550         struct lmv_obd         *lmv = &obd->u.lmv;
1551         struct lmv_tgt_desc    *tgt;
1552         int                     rc;
1553         ENTRY;
1554
1555         tgt = lmv_find_target(lmv, fid);
1556         if (IS_ERR(tgt))
1557                 RETURN(PTR_ERR(tgt));
1558
1559         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1560
1561         RETURN(rc);
1562 }
1563
1564 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1565                         u64 obd_md_valid, const char *name,
1566                         const void *value, size_t value_size,
1567                         unsigned int xattr_flags, u32 suppgid,
1568                         struct ptlrpc_request **req)
1569 {
1570         struct obd_device      *obd = exp->exp_obd;
1571         struct lmv_obd         *lmv = &obd->u.lmv;
1572         struct lmv_tgt_desc    *tgt;
1573         int                     rc;
1574         ENTRY;
1575
1576         tgt = lmv_find_target(lmv, fid);
1577         if (IS_ERR(tgt))
1578                 RETURN(PTR_ERR(tgt));
1579
1580         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1581                          value, value_size, xattr_flags, suppgid, req);
1582
1583         RETURN(rc);
1584 }
1585
1586 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1587                        struct ptlrpc_request **request)
1588 {
1589         struct obd_device       *obd = exp->exp_obd;
1590         struct lmv_obd          *lmv = &obd->u.lmv;
1591         struct lmv_tgt_desc     *tgt;
1592         int                      rc;
1593         ENTRY;
1594
1595         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1596         if (IS_ERR(tgt))
1597                 RETURN(PTR_ERR(tgt));
1598
1599         if (op_data->op_flags & MF_GET_MDT_IDX) {
1600                 op_data->op_mds = tgt->ltd_index;
1601                 RETURN(0);
1602         }
1603
1604         rc = md_getattr(tgt->ltd_exp, op_data, request);
1605
1606         RETURN(rc);
1607 }
1608
1609 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1610 {
1611         struct obd_device   *obd = exp->exp_obd;
1612         struct lmv_obd      *lmv = &obd->u.lmv;
1613         __u32                i;
1614         ENTRY;
1615
1616         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1617
1618         /*
1619          * With DNE every object can have two locks in different namespaces:
1620          * lookup lock in space of MDT storing direntry and update/open lock in
1621          * space of MDT storing inode.
1622          */
1623         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1624                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
1625                         continue;
1626                 md_null_inode(lmv->tgts[i]->ltd_exp, fid);
1627         }
1628
1629         RETURN(0);
1630 }
1631
1632 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1633                      struct md_open_data *mod, struct ptlrpc_request **request)
1634 {
1635         struct obd_device     *obd = exp->exp_obd;
1636         struct lmv_obd        *lmv = &obd->u.lmv;
1637         struct lmv_tgt_desc   *tgt;
1638         int                    rc;
1639         ENTRY;
1640
1641         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1642         if (IS_ERR(tgt))
1643                 RETURN(PTR_ERR(tgt));
1644
1645         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1646         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1647         RETURN(rc);
1648 }
1649
1650 static struct lmv_tgt_desc *
1651 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1652                        const char *name, int namelen, struct lu_fid *fid,
1653                        __u32 *mds, bool post_migrate)
1654 {
1655         struct lmv_tgt_desc *tgt;
1656         const struct lmv_oinfo *oinfo;
1657
1658         if (!lmv_dir_striped(lsm) || !namelen) {
1659                 tgt = lmv_find_target(lmv, fid);
1660                 if (IS_ERR(tgt))
1661                         return tgt;
1662
1663                 *mds = tgt->ltd_index;
1664                 return tgt;
1665         }
1666
1667         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1668                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1669                         return ERR_PTR(-EBADF);
1670                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1671         } else {
1672                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
1673                                                 post_migrate);
1674                 if (IS_ERR(oinfo))
1675                         return ERR_CAST(oinfo);
1676         }
1677
1678         *fid = oinfo->lmo_fid;
1679         *mds = oinfo->lmo_mds;
1680         tgt = lmv_get_target(lmv, oinfo->lmo_mds, NULL);
1681
1682         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1683
1684         return tgt;
1685 }
1686
1687 /**
1688  * Locate MDT of op_data->op_fid1
1689  *
1690  * For striped directory, it will locate the stripe by name hash, if hash_type
1691  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1692  * which is set outside, and if dir is migrating, 'op_data->op_post_migrate'
1693  * indicates whether old or new layout is used to locate.
1694  *
1695  * For plain direcotry, normally it will locate MDT by FID, but if this
1696  * directory has default LMV, and its hash type is "space", locate MDT with QoS.
1697  *
1698  * \param[in] lmv       LMV device
1699  * \param[in] op_data   client MD stack parameters, name, namelen
1700  *                      mds_num etc.
1701  *
1702  * retval               pointer to the lmv_tgt_desc if succeed.
1703  *                      ERR_PTR(errno) if failed.
1704  */
1705 struct lmv_tgt_desc *
1706 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1707 {
1708         struct lmv_stripe_md *lsm = op_data->op_mea1;
1709         struct lmv_oinfo *oinfo;
1710         struct lmv_tgt_desc *tgt;
1711
1712         if (lmv_dir_foreign(lsm))
1713                 return ERR_PTR(-ENODATA);
1714
1715         /* During creating VOLATILE file, it should honor the mdt
1716          * index if the file under striped dir is being restored, see
1717          * ct_restore(). */
1718         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1719             (int)op_data->op_mds != -1) {
1720                 tgt = lmv_get_target(lmv, op_data->op_mds, NULL);
1721                 if (IS_ERR(tgt))
1722                         return tgt;
1723
1724                 if (lmv_dir_striped(lsm)) {
1725                         int i;
1726
1727                         /* refill the right parent fid */
1728                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1729                                 oinfo = &lsm->lsm_md_oinfo[i];
1730                                 if (oinfo->lmo_mds == op_data->op_mds) {
1731                                         op_data->op_fid1 = oinfo->lmo_fid;
1732                                         break;
1733                                 }
1734                         }
1735
1736                         if (i == lsm->lsm_md_stripe_count)
1737                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1738                 }
1739         } else if (lmv_dir_bad_hash(lsm)) {
1740                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1741                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1742
1743                 op_data->op_fid1 = oinfo->lmo_fid;
1744                 op_data->op_mds = oinfo->lmo_mds;
1745                 tgt = lmv_get_target(lmv, oinfo->lmo_mds, NULL);
1746         } else if (op_data->op_code == LUSTRE_OPC_MKDIR &&
1747                    lmv_dir_qos_mkdir(op_data->op_default_mea1) &&
1748                    !lmv_dir_striped(lsm)) {
1749                 tgt = lmv_locate_tgt_qos(lmv, &op_data->op_mds);
1750                 if (tgt == ERR_PTR(-EAGAIN))
1751                         tgt = lmv_locate_tgt_rr(lmv, &op_data->op_mds);
1752                 /*
1753                  * only update statfs when mkdir under dir with "space" hash,
1754                  * this means the cached statfs may be stale, and current mkdir
1755                  * may not follow QoS accurately, but it's not serious, and it
1756                  * avoids periodic statfs when client doesn't mkdir under
1757                  * "space" hashed directories.
1758                  *
1759                  * TODO: after MDT support QoS object allocation, also update
1760                  * statfs for 'lfs mkdir -i -1 ...", currently it's done in user
1761                  * space.
1762                  */
1763                 if (!IS_ERR(tgt)) {
1764                         struct obd_device *obd;
1765
1766                         obd = container_of(lmv, struct obd_device, u.lmv);
1767                         lmv_statfs_check_update(obd, tgt);
1768                 }
1769         } else {
1770                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1771                                 op_data->op_name, op_data->op_namelen,
1772                                 &op_data->op_fid1, &op_data->op_mds,
1773                                 op_data->op_post_migrate);
1774         }
1775
1776         return tgt;
1777 }
1778
1779 /* Locate MDT of op_data->op_fid2 for link/rename */
1780 static struct lmv_tgt_desc *
1781 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1782 {
1783         struct lmv_tgt_desc *tgt;
1784         int rc;
1785
1786         LASSERT(op_data->op_name);
1787         if (lmv_dir_migrating(op_data->op_mea2)) {
1788                 struct lu_fid fid1 = op_data->op_fid1;
1789                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1790                 struct ptlrpc_request *request = NULL;
1791
1792                 /*
1793                  * avoid creating new file under old layout of migrating
1794                  * directory, check it here.
1795                  */
1796                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1797                                 op_data->op_name, op_data->op_namelen,
1798                                 &op_data->op_fid2, &op_data->op_mds, false);
1799                 if (IS_ERR(tgt))
1800                         RETURN(tgt);
1801
1802                 op_data->op_fid1 = op_data->op_fid2;
1803                 op_data->op_mea1 = op_data->op_mea2;
1804                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1805                 op_data->op_fid1 = fid1;
1806                 op_data->op_mea1 = lsm1;
1807                 if (!rc) {
1808                         ptlrpc_req_finished(request);
1809                         RETURN(ERR_PTR(-EEXIST));
1810                 }
1811
1812                 if (rc != -ENOENT)
1813                         RETURN(ERR_PTR(rc));
1814         }
1815
1816         return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1817                                 op_data->op_namelen, &op_data->op_fid2,
1818                                 &op_data->op_mds, true);
1819 }
1820
1821 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1822                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1823                 gid_t gid, cfs_cap_t cap_effective, __u64 rdev,
1824                 struct ptlrpc_request **request)
1825 {
1826         struct obd_device *obd = exp->exp_obd;
1827         struct lmv_obd *lmv = &obd->u.lmv;
1828         struct lmv_tgt_desc *tgt;
1829         int rc;
1830
1831         ENTRY;
1832
1833         if (!lmv->desc.ld_active_tgt_count)
1834                 RETURN(-EIO);
1835
1836         if (lmv_dir_bad_hash(op_data->op_mea1))
1837                 RETURN(-EBADF);
1838
1839         if (lmv_dir_migrating(op_data->op_mea1)) {
1840                 /*
1841                  * if parent is migrating, create() needs to lookup existing
1842                  * name, to avoid creating new file under old layout of
1843                  * migrating directory, check old layout here.
1844                  */
1845                 tgt = lmv_locate_tgt(lmv, op_data);
1846                 if (IS_ERR(tgt))
1847                         RETURN(PTR_ERR(tgt));
1848
1849                 rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1850                 if (!rc) {
1851                         ptlrpc_req_finished(*request);
1852                         *request = NULL;
1853                         RETURN(-EEXIST);
1854                 }
1855
1856                 if (rc != -ENOENT)
1857                         RETURN(rc);
1858
1859                 op_data->op_post_migrate = true;
1860         }
1861
1862         tgt = lmv_locate_tgt(lmv, op_data);
1863         if (IS_ERR(tgt))
1864                 RETURN(PTR_ERR(tgt));
1865
1866         CDEBUG(D_INODE, "CREATE name '%.*s' on "DFID" -> mds #%x\n",
1867                 (int)op_data->op_namelen, op_data->op_name,
1868                 PFID(&op_data->op_fid1), op_data->op_mds);
1869
1870         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1871         if (rc)
1872                 RETURN(rc);
1873
1874         if (exp_connect_flags(exp) & OBD_CONNECT_DIR_STRIPE) {
1875                 /* Send the create request to the MDT where the object
1876                  * will be located */
1877                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
1878                 if (IS_ERR(tgt))
1879                         RETURN(PTR_ERR(tgt));
1880
1881                 op_data->op_mds = tgt->ltd_index;
1882         }
1883
1884         CDEBUG(D_INODE, "CREATE obj "DFID" -> mds #%x\n",
1885                PFID(&op_data->op_fid2), op_data->op_mds);
1886
1887         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1888         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1889                        cap_effective, rdev, request);
1890         if (rc == 0) {
1891                 if (*request == NULL)
1892                         RETURN(rc);
1893                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1894         }
1895         RETURN(rc);
1896 }
1897
1898 static int
1899 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1900             const union ldlm_policy_data *policy, struct md_op_data *op_data,
1901             struct lustre_handle *lockh, __u64 extra_lock_flags)
1902 {
1903         struct obd_device        *obd = exp->exp_obd;
1904         struct lmv_obd           *lmv = &obd->u.lmv;
1905         struct lmv_tgt_desc      *tgt;
1906         int                       rc;
1907         ENTRY;
1908
1909         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
1910
1911         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1912         if (IS_ERR(tgt))
1913                 RETURN(PTR_ERR(tgt));
1914
1915         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
1916                PFID(&op_data->op_fid1), tgt->ltd_index);
1917
1918         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
1919                         extra_lock_flags);
1920
1921         RETURN(rc);
1922 }
1923
1924 int
1925 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1926                  struct ptlrpc_request **preq)
1927 {
1928         struct obd_device *obd = exp->exp_obd;
1929         struct lmv_obd *lmv = &obd->u.lmv;
1930         struct lmv_tgt_desc *tgt;
1931         struct mdt_body *body;
1932         int rc;
1933
1934         ENTRY;
1935
1936 retry:
1937         tgt = lmv_locate_tgt(lmv, op_data);
1938         if (IS_ERR(tgt))
1939                 RETURN(PTR_ERR(tgt));
1940
1941         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1942                 (int)op_data->op_namelen, op_data->op_name,
1943                 PFID(&op_data->op_fid1), tgt->ltd_index);
1944
1945         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
1946         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
1947                 ptlrpc_req_finished(*preq);
1948                 *preq = NULL;
1949                 goto retry;
1950         }
1951
1952         if (rc)
1953                 RETURN(rc);
1954
1955         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
1956         LASSERT(body != NULL);
1957
1958         if (body->mbo_valid & OBD_MD_MDS) {
1959                 op_data->op_fid1 = body->mbo_fid1;
1960                 op_data->op_valid |= OBD_MD_FLCROSSREF;
1961                 op_data->op_namelen = 0;
1962                 op_data->op_name = NULL;
1963
1964                 ptlrpc_req_finished(*preq);
1965                 *preq = NULL;
1966
1967                 goto retry;
1968         }
1969
1970         RETURN(rc);
1971 }
1972
1973 #define md_op_data_fid(op_data, fl)                     \
1974         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1975          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1976          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1977          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1978          NULL)
1979
1980 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
1981                             struct md_op_data *op_data, __u32 op_tgt,
1982                             enum ldlm_mode mode, int bits, int flag)
1983 {
1984         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1985         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
1986         union ldlm_policy_data policy = { { 0 } };
1987         int rc = 0;
1988         ENTRY;
1989
1990         if (!fid_is_sane(fid))
1991                 RETURN(0);
1992
1993         if (tgt == NULL) {
1994                 tgt = lmv_find_target(lmv, fid);
1995                 if (IS_ERR(tgt))
1996                         RETURN(PTR_ERR(tgt));
1997         }
1998
1999         if (tgt->ltd_index != op_tgt) {
2000                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2001                 policy.l_inodebits.bits = bits;
2002                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2003                                       mode, LCF_ASYNC, NULL);
2004         } else {
2005                 CDEBUG(D_INODE,
2006                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2007                        op_tgt, PFID(fid));
2008                 op_data->op_flags |= flag;
2009                 rc = 0;
2010         }
2011
2012         RETURN(rc);
2013 }
2014
2015 /*
2016  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2017  * op_data->op_fid2
2018  */
2019 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2020                     struct ptlrpc_request **request)
2021 {
2022         struct obd_device       *obd = exp->exp_obd;
2023         struct lmv_obd          *lmv = &obd->u.lmv;
2024         struct lmv_tgt_desc     *tgt;
2025         int                      rc;
2026         ENTRY;
2027
2028         LASSERT(op_data->op_namelen != 0);
2029
2030         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2031                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2032                op_data->op_name, PFID(&op_data->op_fid1));
2033
2034         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2035         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2036         op_data->op_cap = cfs_curproc_cap_pack();
2037
2038         tgt = lmv_locate_tgt2(lmv, op_data);
2039         if (IS_ERR(tgt))
2040                 RETURN(PTR_ERR(tgt));
2041
2042         /*
2043          * Cancel UPDATE lock on child (fid1).
2044          */
2045         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2046         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2047                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2048         if (rc != 0)
2049                 RETURN(rc);
2050
2051         rc = md_link(tgt->ltd_exp, op_data, request);
2052
2053         RETURN(rc);
2054 }
2055
2056 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2057                         const char *name, size_t namelen,
2058                         struct ptlrpc_request **request)
2059 {
2060         struct obd_device *obd = exp->exp_obd;
2061         struct lmv_obd *lmv = &obd->u.lmv;
2062         struct lmv_stripe_md *lsm = op_data->op_mea1;
2063         struct lmv_tgt_desc *parent_tgt;
2064         struct lmv_tgt_desc *sp_tgt;
2065         struct lmv_tgt_desc *tp_tgt = NULL;
2066         struct lmv_tgt_desc *child_tgt;
2067         struct lmv_tgt_desc *tgt;
2068         struct lu_fid target_fid;
2069         int rc;
2070
2071         ENTRY;
2072
2073         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2074
2075         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2076                PFID(&op_data->op_fid1), (int)namelen, name);
2077
2078         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2079         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2080         op_data->op_cap = cfs_curproc_cap_pack();
2081
2082         parent_tgt = lmv_find_target(lmv, &op_data->op_fid1);
2083         if (IS_ERR(parent_tgt))
2084                 RETURN(PTR_ERR(parent_tgt));
2085
2086         if (lmv_dir_striped(lsm)) {
2087                 __u32 hash_type = lsm->lsm_md_hash_type;
2088                 __u32 stripe_count = lsm->lsm_md_stripe_count;
2089
2090                 /*
2091                  * old stripes are appended after new stripes for migrating
2092                  * directory.
2093                  */
2094                 if (lmv_dir_migrating(lsm)) {
2095                         hash_type = lsm->lsm_md_migrate_hash;
2096                         stripe_count -= lsm->lsm_md_migrate_offset;
2097                 }
2098
2099                 rc = lmv_name_to_stripe_index(hash_type, stripe_count, name,
2100                                               namelen);
2101                 if (rc < 0)
2102                         RETURN(rc);
2103
2104                 if (lmv_dir_migrating(lsm))
2105                         rc += lsm->lsm_md_migrate_offset;
2106
2107                 /* save it in fid4 temporarily for early cancel */
2108                 op_data->op_fid4 = lsm->lsm_md_oinfo[rc].lmo_fid;
2109                 sp_tgt = lmv_get_target(lmv, lsm->lsm_md_oinfo[rc].lmo_mds,
2110                                         NULL);
2111                 if (IS_ERR(sp_tgt))
2112                         RETURN(PTR_ERR(sp_tgt));
2113
2114                 /*
2115                  * if parent is being migrated too, fill op_fid2 with target
2116                  * stripe fid, otherwise the target stripe is not created yet.
2117                  */
2118                 if (lmv_dir_migrating(lsm)) {
2119                         hash_type = lsm->lsm_md_hash_type &
2120                                     ~LMV_HASH_FLAG_MIGRATION;
2121                         stripe_count = lsm->lsm_md_migrate_offset;
2122
2123                         rc = lmv_name_to_stripe_index(hash_type, stripe_count,
2124                                                       name, namelen);
2125                         if (rc < 0)
2126                                 RETURN(rc);
2127
2128                         op_data->op_fid2 = lsm->lsm_md_oinfo[rc].lmo_fid;
2129                         tp_tgt = lmv_get_target(lmv,
2130                                                 lsm->lsm_md_oinfo[rc].lmo_mds,
2131                                                 NULL);
2132                         if (IS_ERR(tp_tgt))
2133                                 RETURN(PTR_ERR(tp_tgt));
2134                 }
2135         } else {
2136                 sp_tgt = parent_tgt;
2137         }
2138
2139         child_tgt = lmv_find_target(lmv, &op_data->op_fid3);
2140         if (IS_ERR(child_tgt))
2141                 RETURN(PTR_ERR(child_tgt));
2142
2143         if (!S_ISDIR(op_data->op_mode) && tp_tgt)
2144                 rc = __lmv_fid_alloc(lmv, &target_fid, tp_tgt->ltd_index);
2145         else
2146                 rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2147         if (rc)
2148                 RETURN(rc);
2149
2150         /*
2151          * for directory, send migrate request to the MDT where the object will
2152          * be migrated to, because we can't create a striped directory remotely.
2153          *
2154          * otherwise, send to the MDT where source is located because regular
2155          * file may open lease.
2156          *
2157          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2158          * backward compatibility.
2159          */
2160         if (S_ISDIR(op_data->op_mode) &&
2161             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2162                 tgt = lmv_find_target(lmv, &target_fid);
2163                 if (IS_ERR(tgt))
2164                         RETURN(PTR_ERR(tgt));
2165         } else {
2166                 tgt = child_tgt;
2167         }
2168
2169         /* cancel UPDATE lock of parent master object */
2170         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2171                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2172         if (rc)
2173                 RETURN(rc);
2174
2175         /* cancel UPDATE lock of source parent */
2176         if (sp_tgt != parent_tgt) {
2177                 /*
2178                  * migrate RPC packs master object FID, because we can only pack
2179                  * two FIDs in reint RPC, but MDS needs to know both source
2180                  * parent and target parent, and it will obtain them from master
2181                  * FID and LMV, the other FID in RPC is kept for target.
2182                  *
2183                  * since this FID is not passed to MDC, cancel it anyway.
2184                  */
2185                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2186                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2187                 if (rc)
2188                         RETURN(rc);
2189
2190                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2191         }
2192         op_data->op_fid4 = target_fid;
2193
2194         /* cancel UPDATE locks of target parent */
2195         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2196                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2197         if (rc)
2198                 RETURN(rc);
2199
2200         /* cancel LOOKUP lock of source if source is remote object */
2201         if (child_tgt != sp_tgt) {
2202                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2203                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2204                                       MF_MDC_CANCEL_FID3);
2205                 if (rc)
2206                         RETURN(rc);
2207         }
2208
2209         /* cancel ELC locks of source */
2210         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2211                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2212         if (rc)
2213                 RETURN(rc);
2214
2215         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2216
2217         RETURN(rc);
2218 }
2219
2220 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2221                       const char *old, size_t oldlen,
2222                       const char *new, size_t newlen,
2223                       struct ptlrpc_request **request)
2224 {
2225         struct obd_device *obd = exp->exp_obd;
2226         struct lmv_obd *lmv = &obd->u.lmv;
2227         struct lmv_tgt_desc *sp_tgt;
2228         struct lmv_tgt_desc *tp_tgt = NULL;
2229         struct lmv_tgt_desc *src_tgt = NULL;
2230         struct lmv_tgt_desc *tgt;
2231         struct mdt_body *body;
2232         int rc;
2233
2234         ENTRY;
2235
2236         LASSERT(oldlen != 0);
2237
2238         if (op_data->op_cli_flags & CLI_MIGRATE) {
2239                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2240                 RETURN(rc);
2241         }
2242
2243         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2244         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2245         op_data->op_cap = cfs_curproc_cap_pack();
2246
2247         op_data->op_name = new;
2248         op_data->op_namelen = newlen;
2249
2250         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2251         if (IS_ERR(tp_tgt))
2252                 RETURN(PTR_ERR(tp_tgt));
2253
2254         /* Since the target child might be destroyed, and it might become
2255          * orphan, and we can only check orphan on the local MDT right now, so
2256          * we send rename request to the MDT where target child is located. If
2257          * target child does not exist, then it will send the request to the
2258          * target parent */
2259         if (fid_is_sane(&op_data->op_fid4)) {
2260                 tgt = lmv_find_target(lmv, &op_data->op_fid4);
2261                 if (IS_ERR(tgt))
2262                         RETURN(PTR_ERR(tgt));
2263         } else {
2264                 tgt = tp_tgt;
2265         }
2266
2267         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2268
2269         /* cancel UPDATE locks of target parent */
2270         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2271                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2272         if (rc != 0)
2273                 RETURN(rc);
2274
2275         if (fid_is_sane(&op_data->op_fid4)) {
2276                 /* cancel LOOKUP lock of target on target parent */
2277                 if (tgt != tp_tgt) {
2278                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2279                                               tgt->ltd_index, LCK_EX,
2280                                               MDS_INODELOCK_LOOKUP,
2281                                               MF_MDC_CANCEL_FID4);
2282                         if (rc != 0)
2283                                 RETURN(rc);
2284                 }
2285         }
2286
2287         if (fid_is_sane(&op_data->op_fid3)) {
2288                 src_tgt = lmv_find_target(lmv, &op_data->op_fid3);
2289                 if (IS_ERR(src_tgt))
2290                         RETURN(PTR_ERR(src_tgt));
2291
2292                 /* cancel ELC locks of source */
2293                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2294                                       LCK_EX, MDS_INODELOCK_ELC,
2295                                       MF_MDC_CANCEL_FID3);
2296                 if (rc != 0)
2297                         RETURN(rc);
2298         }
2299
2300         op_data->op_name = old;
2301         op_data->op_namelen = oldlen;
2302 retry:
2303         sp_tgt = lmv_locate_tgt(lmv, op_data);
2304         if (IS_ERR(sp_tgt))
2305                 RETURN(PTR_ERR(sp_tgt));
2306
2307         /* cancel UPDATE locks of source parent */
2308         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2309                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2310         if (rc != 0)
2311                 RETURN(rc);
2312
2313         if (fid_is_sane(&op_data->op_fid3)) {
2314                 /* cancel LOOKUP lock of source on source parent */
2315                 if (src_tgt != sp_tgt) {
2316                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2317                                               tgt->ltd_index, LCK_EX,
2318                                               MDS_INODELOCK_LOOKUP,
2319                                               MF_MDC_CANCEL_FID3);
2320                         if (rc != 0)
2321                                 RETURN(rc);
2322                 }
2323         }
2324
2325 rename:
2326         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2327                 PFID(&op_data->op_fid1), (int)oldlen, old,
2328                 PFID(&op_data->op_fid2), (int)newlen, new);
2329
2330         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2331                         request);
2332         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2333                 ptlrpc_req_finished(*request);
2334                 *request = NULL;
2335                 goto retry;
2336         }
2337
2338         if (rc && rc != -EXDEV)
2339                 RETURN(rc);
2340
2341         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2342         if (body == NULL)
2343                 RETURN(-EPROTO);
2344
2345         /* Not cross-ref case, just get out of here. */
2346         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2347                 RETURN(rc);
2348
2349         op_data->op_fid4 = body->mbo_fid1;
2350
2351         ptlrpc_req_finished(*request);
2352         *request = NULL;
2353
2354         tgt = lmv_find_target(lmv, &op_data->op_fid4);
2355         if (IS_ERR(tgt))
2356                 RETURN(PTR_ERR(tgt));
2357
2358         if (fid_is_sane(&op_data->op_fid4)) {
2359                 /* cancel LOOKUP lock of target on target parent */
2360                 if (tgt != tp_tgt) {
2361                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2362                                               tgt->ltd_index, LCK_EX,
2363                                               MDS_INODELOCK_LOOKUP,
2364                                               MF_MDC_CANCEL_FID4);
2365                         if (rc != 0)
2366                                 RETURN(rc);
2367                 }
2368         }
2369
2370         goto rename;
2371 }
2372
2373 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2374                        void *ea, size_t ealen, struct ptlrpc_request **request)
2375 {
2376         struct obd_device       *obd = exp->exp_obd;
2377         struct lmv_obd          *lmv = &obd->u.lmv;
2378         struct lmv_tgt_desc     *tgt;
2379         int                      rc = 0;
2380         ENTRY;
2381
2382         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2383                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2384                op_data->op_xvalid);
2385
2386         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2387         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2388         if (IS_ERR(tgt))
2389                 RETURN(PTR_ERR(tgt));
2390
2391         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2392
2393         RETURN(rc);
2394 }
2395
2396 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2397                      struct ptlrpc_request **request)
2398 {
2399         struct obd_device       *obd = exp->exp_obd;
2400         struct lmv_obd          *lmv = &obd->u.lmv;
2401         struct lmv_tgt_desc     *tgt;
2402         int                      rc;
2403         ENTRY;
2404
2405         tgt = lmv_find_target(lmv, fid);
2406         if (IS_ERR(tgt))
2407                 RETURN(PTR_ERR(tgt));
2408
2409         rc = md_fsync(tgt->ltd_exp, fid, request);
2410         RETURN(rc);
2411 }
2412
2413 struct stripe_dirent {
2414         struct page             *sd_page;
2415         struct lu_dirpage       *sd_dp;
2416         struct lu_dirent        *sd_ent;
2417         bool                     sd_eof;
2418 };
2419
2420 struct lmv_dir_ctxt {
2421         struct lmv_obd          *ldc_lmv;
2422         struct md_op_data       *ldc_op_data;
2423         struct md_callback      *ldc_cb_op;
2424         __u64                    ldc_hash;
2425         int                      ldc_count;
2426         struct stripe_dirent     ldc_stripes[0];
2427 };
2428
2429 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2430 {
2431         if (stripe->sd_page) {
2432                 kunmap(stripe->sd_page);
2433                 put_page(stripe->sd_page);
2434                 stripe->sd_page = NULL;
2435                 stripe->sd_ent = NULL;
2436         }
2437 }
2438
2439 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2440 {
2441         int i;
2442
2443         for (i = 0; i < ctxt->ldc_count; i++)
2444                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2445 }
2446
2447 /* if @ent is dummy, or . .., get next */
2448 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2449                                            struct lu_dirent *ent,
2450                                            int stripe_index)
2451 {
2452         for (; ent; ent = lu_dirent_next(ent)) {
2453                 /* Skip dummy entry */
2454                 if (le16_to_cpu(ent->lde_namelen) == 0)
2455                         continue;
2456
2457                 /* skip . and .. for other stripes */
2458                 if (stripe_index &&
2459                     (strncmp(ent->lde_name, ".",
2460                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2461                      strncmp(ent->lde_name, "..",
2462                              le16_to_cpu(ent->lde_namelen)) == 0))
2463                         continue;
2464
2465                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2466                         break;
2467         }
2468
2469         return ent;
2470 }
2471
2472 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2473                                             struct stripe_dirent *stripe,
2474                                             int stripe_index)
2475 {
2476         struct md_op_data *op_data = ctxt->ldc_op_data;
2477         struct lmv_oinfo *oinfo;
2478         struct lu_fid fid = op_data->op_fid1;
2479         struct inode *inode = op_data->op_data;
2480         struct lmv_tgt_desc *tgt;
2481         struct lu_dirent *ent = stripe->sd_ent;
2482         __u64 hash = ctxt->ldc_hash;
2483         int rc = 0;
2484
2485         ENTRY;
2486
2487         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2488         LASSERT(!ent);
2489
2490         do {
2491                 if (stripe->sd_page) {
2492                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2493
2494                         /* @hash should be the last dirent hash */
2495                         LASSERTF(hash <= end,
2496                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2497                                  ctxt, stripe, hash, end);
2498                         /* unload last page */
2499                         stripe_dirent_unload(stripe);
2500                         /* eof */
2501                         if (end == MDS_DIR_END_OFF) {
2502                                 stripe->sd_eof = true;
2503                                 break;
2504                         }
2505                         hash = end;
2506                 }
2507
2508                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2509                 if (!oinfo->lmo_root) {
2510                         rc = -ENOENT;
2511                         break;
2512                 }
2513
2514                 tgt = lmv_get_target(ctxt->ldc_lmv, oinfo->lmo_mds, NULL);
2515                 if (IS_ERR(tgt)) {
2516                         rc = PTR_ERR(tgt);
2517                         break;
2518                 }
2519
2520                 /* op_data is shared by stripes, reset after use */
2521                 op_data->op_fid1 = oinfo->lmo_fid;
2522                 op_data->op_fid2 = oinfo->lmo_fid;
2523                 op_data->op_data = oinfo->lmo_root;
2524
2525                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2526                                   &stripe->sd_page);
2527
2528                 op_data->op_fid1 = fid;
2529                 op_data->op_fid2 = fid;
2530                 op_data->op_data = inode;
2531
2532                 if (rc)
2533                         break;
2534
2535                 stripe->sd_dp = page_address(stripe->sd_page);
2536                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2537                                         stripe_index);
2538                 /* in case a page filled with ., .. and dummy, read next */
2539         } while (!ent);
2540
2541         stripe->sd_ent = ent;
2542         if (rc) {
2543                 LASSERT(!ent);
2544                 /* treat error as eof, so dir can be partially accessed */
2545                 stripe->sd_eof = true;
2546                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2547                               "directory is partially accessed!\n",
2548                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2549                               rc);
2550         }
2551
2552         RETURN(ent);
2553 }
2554
2555 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2556 {
2557         struct obd_device       *obd = exp->exp_obd;
2558         struct lmv_obd          *lmv = &obd->u.lmv;
2559         struct lmv_tgt_desc     *tgt;
2560         int                      rc;
2561         ENTRY;
2562
2563         rc = lmv_check_connect(obd);
2564         if (rc != 0)
2565                 RETURN(rc);
2566
2567         tgt = lmv_find_target(lmv, &data->op_fid1);
2568         if (IS_ERR(tgt))
2569                 RETURN(PTR_ERR(tgt));
2570
2571         data->op_flags |= MF_MDC_CANCEL_FID1;
2572         rc = md_file_resync(tgt->ltd_exp, data);
2573         RETURN(rc);
2574 }
2575
2576 /**
2577  * Get dirent with the closest hash for striped directory
2578  *
2579  * This function will search the dir entry, whose hash value is the
2580  * closest(>=) to hash from all of sub-stripes, and it is only being called
2581  * for striped directory.
2582  *
2583  * \param[in] ctxt              dir read context
2584  *
2585  * \retval                      dirent get the entry successfully
2586  *                              NULL does not get the entry, normally it means
2587  *                              it reaches the end of the directory, while read
2588  *                              stripe dirent error is ignored to allow partial
2589  *                              access.
2590  */
2591 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2592 {
2593         struct stripe_dirent *stripe;
2594         struct lu_dirent *ent = NULL;
2595         int i;
2596         int min = -1;
2597
2598         /* TODO: optimize with k-way merge sort */
2599         for (i = 0; i < ctxt->ldc_count; i++) {
2600                 stripe = &ctxt->ldc_stripes[i];
2601                 if (stripe->sd_eof)
2602                         continue;
2603
2604                 if (!stripe->sd_ent) {
2605                         stripe_dirent_load(ctxt, stripe, i);
2606                         if (!stripe->sd_ent) {
2607                                 LASSERT(stripe->sd_eof);
2608                                 continue;
2609                         }
2610                 }
2611
2612                 if (min == -1 ||
2613                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2614                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2615                         min = i;
2616                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2617                             ctxt->ldc_hash)
2618                                 break;
2619                 }
2620         }
2621
2622         if (min != -1) {
2623                 stripe = &ctxt->ldc_stripes[min];
2624                 ent = stripe->sd_ent;
2625                 /* pop found dirent */
2626                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2627                                                    min);
2628         }
2629
2630         return ent;
2631 }
2632
2633 /**
2634  * Build dir entry page for striped directory
2635  *
2636  * This function gets one entry by @offset from a striped directory. It will
2637  * read entries from all of stripes, and choose one closest to the required
2638  * offset(&offset). A few notes
2639  * 1. skip . and .. for non-zero stripes, because there can only have one .
2640  * and .. in a directory.
2641  * 2. op_data will be shared by all of stripes, instead of allocating new
2642  * one, so need to restore before reusing.
2643  *
2644  * \param[in] exp       obd export refer to LMV
2645  * \param[in] op_data   hold those MD parameters of read_entry
2646  * \param[in] cb_op     ldlm callback being used in enqueue in mdc_read_entry
2647  * \param[in] offset    starting hash offset
2648  * \param[out] ppage    the page holding the entry. Note: because the entry
2649  *                      will be accessed in upper layer, so we need hold the
2650  *                      page until the usages of entry is finished, see
2651  *                      ll_dir_entry_next.
2652  *
2653  * retval               =0 if get entry successfully
2654  *                      <0 cannot get entry
2655  */
2656 static int lmv_striped_read_page(struct obd_export *exp,
2657                                  struct md_op_data *op_data,
2658                                  struct md_callback *cb_op,
2659                                  __u64 offset, struct page **ppage)
2660 {
2661         struct page *page = NULL;
2662         struct lu_dirpage *dp;
2663         void *start;
2664         struct lu_dirent *ent;
2665         struct lu_dirent *last_ent;
2666         int stripe_count;
2667         struct lmv_dir_ctxt *ctxt;
2668         struct lu_dirent *next = NULL;
2669         __u16 ent_size;
2670         size_t left_bytes;
2671         int rc = 0;
2672         ENTRY;
2673
2674         /* Allocate a page and read entries from all of stripes and fill
2675          * the page by hash order */
2676         page = alloc_page(GFP_KERNEL);
2677         if (!page)
2678                 RETURN(-ENOMEM);
2679
2680         /* Initialize the entry page */
2681         dp = kmap(page);
2682         memset(dp, 0, sizeof(*dp));
2683         dp->ldp_hash_start = cpu_to_le64(offset);
2684
2685         start = dp + 1;
2686         left_bytes = PAGE_SIZE - sizeof(*dp);
2687         ent = start;
2688         last_ent = ent;
2689
2690         /* initalize dir read context */
2691         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2692         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2693         if (!ctxt)
2694                 GOTO(free_page, rc = -ENOMEM);
2695         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2696         ctxt->ldc_op_data = op_data;
2697         ctxt->ldc_cb_op = cb_op;
2698         ctxt->ldc_hash = offset;
2699         ctxt->ldc_count = stripe_count;
2700
2701         while (1) {
2702                 next = lmv_dirent_next(ctxt);
2703
2704                 /* end of directory */
2705                 if (!next) {
2706                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2707                         break;
2708                 }
2709                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2710
2711                 ent_size = le16_to_cpu(next->lde_reclen);
2712
2713                 /* the last entry lde_reclen is 0, but it might not be the last
2714                  * one of this temporay dir page */
2715                 if (!ent_size)
2716                         ent_size = lu_dirent_calc_size(
2717                                         le16_to_cpu(next->lde_namelen),
2718                                         le32_to_cpu(next->lde_attrs));
2719                 /* page full */
2720                 if (ent_size > left_bytes)
2721                         break;
2722
2723                 memcpy(ent, next, ent_size);
2724
2725                 /* Replace . with master FID and Replace .. with the parent FID
2726                  * of master object */
2727                 if (strncmp(ent->lde_name, ".",
2728                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2729                     le16_to_cpu(ent->lde_namelen) == 1)
2730                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2731                 else if (strncmp(ent->lde_name, "..",
2732                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2733                            le16_to_cpu(ent->lde_namelen) == 2)
2734                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2735
2736                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2737                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2738                        le64_to_cpu(ent->lde_hash));
2739
2740                 left_bytes -= ent_size;
2741                 ent->lde_reclen = cpu_to_le16(ent_size);
2742                 last_ent = ent;
2743                 ent = (void *)ent + ent_size;
2744         };
2745
2746         last_ent->lde_reclen = 0;
2747
2748         if (ent == start)
2749                 dp->ldp_flags |= LDF_EMPTY;
2750         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2751                 dp->ldp_flags |= LDF_COLLIDE;
2752         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2753         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2754
2755         put_lmv_dir_ctxt(ctxt);
2756         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2757
2758         *ppage = page;
2759
2760         RETURN(0);
2761
2762 free_page:
2763         kunmap(page);
2764         __free_page(page);
2765
2766         return rc;
2767 }
2768
2769 int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2770                   struct md_callback *cb_op, __u64 offset,
2771                   struct page **ppage)
2772 {
2773         struct obd_device *obd = exp->exp_obd;
2774         struct lmv_obd *lmv = &obd->u.lmv;
2775         struct lmv_tgt_desc *tgt;
2776         int rc;
2777
2778         ENTRY;
2779
2780         if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2781                 RETURN(-ENODATA);
2782
2783         if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2784                 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2785                 RETURN(rc);
2786         }
2787
2788         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2789         if (IS_ERR(tgt))
2790                 RETURN(PTR_ERR(tgt));
2791
2792         rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2793
2794         RETURN(rc);
2795 }
2796
2797 /**
2798  * Unlink a file/directory
2799  *
2800  * Unlink a file or directory under the parent dir. The unlink request
2801  * usually will be sent to the MDT where the child is located, but if
2802  * the client does not have the child FID then request will be sent to the
2803  * MDT where the parent is located.
2804  *
2805  * If the parent is a striped directory then it also needs to locate which
2806  * stripe the name of the child is located, and replace the parent FID
2807  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2808  * it will walk through all of sub-stripes until the child is being
2809  * unlinked finally.
2810  *
2811  * \param[in] exp       export refer to LMV
2812  * \param[in] op_data   different parameters transferred beween client
2813  *                      MD stacks, name, namelen, FIDs etc.
2814  *                      op_fid1 is the parent FID, op_fid2 is the child
2815  *                      FID.
2816  * \param[out] request  point to the request of unlink.
2817  *
2818  * retval               0 if succeed
2819  *                      negative errno if failed.
2820  */
2821 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2822                       struct ptlrpc_request **request)
2823 {
2824         struct obd_device *obd = exp->exp_obd;
2825         struct lmv_obd *lmv = &obd->u.lmv;
2826         struct lmv_tgt_desc *tgt;
2827         struct lmv_tgt_desc *parent_tgt;
2828         struct mdt_body *body;
2829         int rc;
2830
2831         ENTRY;
2832
2833         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2834         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2835         op_data->op_cap = cfs_curproc_cap_pack();
2836
2837 retry:
2838         parent_tgt = lmv_locate_tgt(lmv, op_data);
2839         if (IS_ERR(parent_tgt))
2840                 RETURN(PTR_ERR(parent_tgt));
2841
2842         if (likely(!fid_is_zero(&op_data->op_fid2))) {
2843                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2844                 if (IS_ERR(tgt))
2845                         RETURN(PTR_ERR(tgt));
2846         } else {
2847                 tgt = parent_tgt;
2848         }
2849
2850         /*
2851          * If child's fid is given, cancel unused locks for it if it is from
2852          * another export than parent.
2853          *
2854          * LOOKUP lock for child (fid3) should also be cancelled on parent
2855          * tgt_tgt in mdc_unlink().
2856          */
2857         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2858
2859         if (parent_tgt != tgt)
2860                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
2861                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2862                                       MF_MDC_CANCEL_FID3);
2863
2864         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2865                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2866         if (rc)
2867                 RETURN(rc);
2868
2869         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
2870                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
2871                tgt->ltd_index);
2872
2873         rc = md_unlink(tgt->ltd_exp, op_data, request);
2874         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2875                 ptlrpc_req_finished(*request);
2876                 *request = NULL;
2877                 goto retry;
2878         }
2879
2880         if (rc != -EREMOTE)
2881                 RETURN(rc);
2882
2883         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2884         if (body == NULL)
2885                 RETURN(-EPROTO);
2886
2887         /* Not cross-ref case, just get out of here. */
2888         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2889                 RETURN(rc);
2890
2891         /* This is a remote object, try remote MDT. */
2892         op_data->op_fid2 = body->mbo_fid1;
2893         ptlrpc_req_finished(*request);
2894         *request = NULL;
2895
2896         tgt = lmv_find_target(lmv, &op_data->op_fid2);
2897         if (IS_ERR(tgt))
2898                 RETURN(PTR_ERR(tgt));
2899
2900         goto retry;
2901 }
2902
2903 static int lmv_precleanup(struct obd_device *obd)
2904 {
2905         ENTRY;
2906         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
2907         fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
2908         lprocfs_obd_cleanup(obd);
2909         lprocfs_free_md_stats(obd);
2910         RETURN(0);
2911 }
2912
2913 /**
2914  * Get by key a value associated with a LMV device.
2915  *
2916  * Dispatch request to lower-layer devices as needed.
2917  *
2918  * \param[in] env               execution environment for this thread
2919  * \param[in] exp               export for the LMV device
2920  * \param[in] keylen            length of key identifier
2921  * \param[in] key               identifier of key to get value for
2922  * \param[in] vallen            size of \a val
2923  * \param[out] val              pointer to storage location for value
2924  * \param[in] lsm               optional striping metadata of object
2925  *
2926  * \retval 0            on success
2927  * \retval negative     negated errno on failure
2928  */
2929 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2930                         __u32 keylen, void *key, __u32 *vallen, void *val)
2931 {
2932         struct obd_device       *obd;
2933         struct lmv_obd          *lmv;
2934         int                      rc = 0;
2935         ENTRY;
2936
2937         obd = class_exp2obd(exp);
2938         if (obd == NULL) {
2939                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2940                        exp->exp_handle.h_cookie);
2941                 RETURN(-EINVAL);
2942         }
2943
2944         lmv = &obd->u.lmv;
2945         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2946                 int i;
2947
2948                 LASSERT(*vallen == sizeof(__u32));
2949                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2950                         struct lmv_tgt_desc *tgt = lmv->tgts[i];
2951                         /*
2952                          * All tgts should be connected when this gets called.
2953                          */
2954                         if (tgt == NULL || tgt->ltd_exp == NULL)
2955                                 continue;
2956
2957                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2958                                           vallen, val))
2959                                 RETURN(0);
2960                 }
2961                 RETURN(-EINVAL);
2962         } else if (KEY_IS(KEY_MAX_EASIZE) ||
2963                    KEY_IS(KEY_DEFAULT_EASIZE) ||
2964                    KEY_IS(KEY_CONN_DATA)) {
2965                 /*
2966                  * Forwarding this request to first MDS, it should know LOV
2967                  * desc.
2968                  */
2969                 rc = obd_get_info(env, lmv->tgts[0]->ltd_exp, keylen, key,
2970                                   vallen, val);
2971                 if (!rc && KEY_IS(KEY_CONN_DATA))
2972                         exp->exp_connect_data = *(struct obd_connect_data *)val;
2973                 RETURN(rc);
2974         } else if (KEY_IS(KEY_TGT_COUNT)) {
2975                 *((int *)val) = lmv->desc.ld_tgt_count;
2976                 RETURN(0);
2977         }
2978
2979         CDEBUG(D_IOCTL, "Invalid key\n");
2980         RETURN(-EINVAL);
2981 }
2982
2983 /**
2984  * Asynchronously set by key a value associated with a LMV device.
2985  *
2986  * Dispatch request to lower-layer devices as needed.
2987  *
2988  * \param[in] env       execution environment for this thread
2989  * \param[in] exp       export for the LMV device
2990  * \param[in] keylen    length of key identifier
2991  * \param[in] key       identifier of key to store value for
2992  * \param[in] vallen    size of value to store
2993  * \param[in] val       pointer to data to be stored
2994  * \param[in] set       optional list of related ptlrpc requests
2995  *
2996  * \retval 0            on success
2997  * \retval negative     negated errno on failure
2998  */
2999 int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3000                         __u32 keylen, void *key, __u32 vallen, void *val,
3001                         struct ptlrpc_request_set *set)
3002 {
3003         struct lmv_tgt_desc     *tgt = NULL;
3004         struct obd_device       *obd;
3005         struct lmv_obd          *lmv;
3006         int rc = 0;
3007         ENTRY;
3008
3009         obd = class_exp2obd(exp);
3010         if (obd == NULL) {
3011                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3012                        exp->exp_handle.h_cookie);
3013                 RETURN(-EINVAL);
3014         }
3015         lmv = &obd->u.lmv;
3016
3017         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3018             KEY_IS(KEY_DEFAULT_EASIZE)) {
3019                 int i, err = 0;
3020
3021                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3022                         tgt = lmv->tgts[i];
3023
3024                         if (tgt == NULL || tgt->ltd_exp == NULL)
3025                                 continue;
3026
3027                         err = obd_set_info_async(env, tgt->ltd_exp,
3028                                                  keylen, key, vallen, val, set);
3029                         if (err && rc == 0)
3030                                 rc = err;
3031                 }
3032
3033                 RETURN(rc);
3034         }
3035
3036         RETURN(-EINVAL);
3037 }
3038
3039 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3040                             const struct lmv_mds_md_v1 *lmm1)
3041 {
3042         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3043         int             stripe_count;
3044         int             cplen;
3045         int             i;
3046         int             rc = 0;
3047         ENTRY;
3048
3049         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3050         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3051         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3052         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3053                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3054         else
3055                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3056         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3057         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3058         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3059         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3060                         sizeof(lsm->lsm_md_pool_name));
3061
3062         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3063                 RETURN(-E2BIG);
3064
3065         CDEBUG(D_INFO, "unpack lsm count %d, master %d hash_type %#x "
3066                "layout_version %d\n", lsm->lsm_md_stripe_count,
3067                lsm->lsm_md_master_mdt_index, lsm->lsm_md_hash_type,
3068                lsm->lsm_md_layout_version);
3069
3070         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3071         for (i = 0; i < stripe_count; i++) {
3072                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3073                               &lmm1->lmv_stripe_fids[i]);
3074                 /*
3075                  * set default value -1, so lmv_locate_tgt() knows this stripe
3076                  * target is not initialized.
3077                  */
3078                 lsm->lsm_md_oinfo[i].lmo_mds = (u32)-1;
3079                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3080                         continue;
3081
3082                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3083                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3084                 if (rc == -ENOENT)
3085                         continue;
3086
3087                 if (rc)
3088                         RETURN(rc);
3089
3090                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3091                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3092         }
3093
3094         RETURN(rc);
3095 }
3096
3097 static inline int lmv_unpack_user_md(struct obd_export *exp,
3098                                      struct lmv_stripe_md *lsm,
3099                                      const struct lmv_user_md *lmu)
3100 {
3101         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3102         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3103         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3104         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3105
3106         return 0;
3107 }
3108
3109 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3110                         const union lmv_mds_md *lmm, size_t lmm_size)
3111 {
3112         struct lmv_stripe_md     *lsm;
3113         int                      lsm_size;
3114         int                      rc;
3115         bool                     allocated = false;
3116         ENTRY;
3117
3118         LASSERT(lsmp != NULL);
3119
3120         lsm = *lsmp;
3121         /* Free memmd */
3122         if (lsm != NULL && lmm == NULL) {
3123                 int i;
3124                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3125
3126                 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3127                         size_t lfm_size;
3128
3129                         lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3130                                                               lfm_value[0]);
3131                         OBD_FREE_LARGE(lfm, lfm_size);
3132                         RETURN(0);
3133                 }
3134
3135                 if (lmv_dir_striped(lsm)) {
3136                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3137                                 if (lsm->lsm_md_oinfo[i].lmo_root)
3138                                         iput(lsm->lsm_md_oinfo[i].lmo_root);
3139                         }
3140                         lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3141                 } else {
3142                         lsm_size = lmv_stripe_md_size(0);
3143                 }
3144                 OBD_FREE(lsm, lsm_size);
3145                 *lsmp = NULL;
3146                 RETURN(0);
3147         }
3148
3149         /* foreign lmv case */
3150         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3151                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3152
3153                 if (lfm == NULL) {
3154                         OBD_ALLOC_LARGE(lfm, lmm_size);
3155                         if (lfm == NULL)
3156                                 RETURN(-ENOMEM);
3157                         *lsmp = (struct lmv_stripe_md *)lfm;
3158                 }
3159                 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3160                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3161                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3162                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3163                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3164                        lfm->lfm_length);
3165                 RETURN(lmm_size);
3166         }
3167
3168         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3169                 RETURN(-EPERM);
3170
3171         /* Unpack memmd */
3172         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3173             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3174                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3175                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3176                        -EIO);
3177                 RETURN(-EIO);
3178         }
3179
3180         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3181                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3182         else
3183                 /**
3184                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3185                  * stripecount should be 0 then.
3186                  */
3187                 lsm_size = lmv_stripe_md_size(0);
3188
3189         if (lsm == NULL) {
3190                 OBD_ALLOC(lsm, lsm_size);
3191                 if (lsm == NULL)
3192                         RETURN(-ENOMEM);
3193                 allocated = true;
3194                 *lsmp = lsm;
3195         }
3196
3197         switch (le32_to_cpu(lmm->lmv_magic)) {
3198         case LMV_MAGIC_V1:
3199                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3200                 break;
3201         case LMV_USER_MAGIC:
3202                 rc = lmv_unpack_user_md(exp, lsm, &lmm->lmv_user_md);
3203                 break;
3204         default:
3205                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3206                        le32_to_cpu(lmm->lmv_magic));
3207                 rc = -EINVAL;
3208                 break;
3209         }
3210
3211         if (rc != 0 && allocated) {
3212                 OBD_FREE(lsm, lsm_size);
3213                 *lsmp = NULL;
3214                 lsm_size = rc;
3215         }
3216         RETURN(lsm_size);
3217 }
3218
3219 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3220 {
3221         lmv_unpackmd(NULL, &lsm, NULL, 0);
3222 }
3223 EXPORT_SYMBOL(lmv_free_memmd);
3224
3225 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3226                              union ldlm_policy_data *policy,
3227                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3228                              void *opaque)
3229 {
3230         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3231         int rc = 0;
3232         __u32 i;
3233         ENTRY;
3234
3235         LASSERT(fid != NULL);
3236
3237         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3238                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
3239                 int err;
3240
3241                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
3242                         continue;
3243
3244                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3245                                        opaque);
3246                 if (!rc)
3247                         rc = err;
3248         }
3249         RETURN(rc);
3250 }
3251
3252 static int lmv_set_lock_data(struct obd_export *exp,
3253                              const struct lustre_handle *lockh,
3254                              void *data, __u64 *bits)
3255 {
3256         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
3257         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3258         int                      rc;
3259         ENTRY;
3260
3261         if (tgt == NULL || tgt->ltd_exp == NULL)
3262                 RETURN(-EINVAL);
3263         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3264         RETURN(rc);
3265 }
3266
3267 enum ldlm_mode lmv_lock_match(struct obd_export *exp, __u64 flags,
3268                               const struct lu_fid *fid, enum ldlm_type type,
3269                               union ldlm_policy_data *policy,
3270                               enum ldlm_mode mode, struct lustre_handle *lockh)
3271 {
3272         struct obd_device       *obd = exp->exp_obd;
3273         struct lmv_obd          *lmv = &obd->u.lmv;
3274         enum ldlm_mode          rc;
3275         int                     tgt;
3276         int                     i;
3277         ENTRY;
3278
3279         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
3280
3281         /*
3282          * With DNE every object can have two locks in different namespaces:
3283          * lookup lock in space of MDT storing direntry and update/open lock in
3284          * space of MDT storing inode.  Try the MDT that the FID maps to first,
3285          * since this can be easily found, and only try others if that fails.
3286          */
3287         for (i = 0, tgt = lmv_find_target_index(lmv, fid);
3288              i < lmv->desc.ld_tgt_count;
3289              i++, tgt = (tgt + 1) % lmv->desc.ld_tgt_count) {
3290                 if (tgt < 0) {
3291                         CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
3292                                obd->obd_name, PFID(fid), tgt);
3293                         tgt = 0;
3294                 }
3295
3296                 if (lmv->tgts[tgt] == NULL ||
3297                     lmv->tgts[tgt]->ltd_exp == NULL ||
3298                     lmv->tgts[tgt]->ltd_active == 0)
3299                         continue;
3300
3301                 rc = md_lock_match(lmv->tgts[tgt]->ltd_exp, flags, fid,
3302                                    type, policy, mode, lockh);
3303                 if (rc)
3304                         RETURN(rc);
3305         }
3306
3307         RETURN(0);
3308 }
3309
3310 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
3311                       struct obd_export *dt_exp, struct obd_export *md_exp,
3312                       struct lustre_md *md)
3313 {
3314         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
3315         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3316
3317         if (tgt == NULL || tgt->ltd_exp == NULL)
3318                 RETURN(-EINVAL);
3319
3320         return md_get_lustre_md(lmv->tgts[0]->ltd_exp, req, dt_exp, md_exp, md);
3321 }
3322
3323 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
3324 {
3325         struct obd_device       *obd = exp->exp_obd;
3326         struct lmv_obd          *lmv = &obd->u.lmv;
3327         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3328         ENTRY;
3329
3330         if (md->default_lmv) {
3331                 lmv_free_memmd(md->default_lmv);
3332                 md->default_lmv = NULL;
3333         }
3334         if (md->lmv != NULL) {
3335                 lmv_free_memmd(md->lmv);
3336                 md->lmv = NULL;
3337         }
3338         if (tgt == NULL || tgt->ltd_exp == NULL)
3339                 RETURN(-EINVAL);
3340         RETURN(md_free_lustre_md(lmv->tgts[0]->ltd_exp, md));
3341 }
3342
3343 int lmv_set_open_replay_data(struct obd_export *exp,
3344                              struct obd_client_handle *och,
3345                              struct lookup_intent *it)
3346 {
3347         struct obd_device       *obd = exp->exp_obd;
3348         struct lmv_obd          *lmv = &obd->u.lmv;
3349         struct lmv_tgt_desc     *tgt;
3350         ENTRY;
3351
3352         tgt = lmv_find_target(lmv, &och->och_fid);
3353         if (IS_ERR(tgt))
3354                 RETURN(PTR_ERR(tgt));
3355
3356         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3357 }
3358
3359 int lmv_clear_open_replay_data(struct obd_export *exp,
3360                                struct obd_client_handle *och)
3361 {
3362         struct obd_device       *obd = exp->exp_obd;
3363         struct lmv_obd          *lmv = &obd->u.lmv;
3364         struct lmv_tgt_desc     *tgt;
3365         ENTRY;
3366
3367         tgt = lmv_find_target(lmv, &och->och_fid);
3368         if (IS_ERR(tgt))
3369                 RETURN(PTR_ERR(tgt));
3370
3371         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3372 }
3373
3374 int lmv_intent_getattr_async(struct obd_export *exp,
3375                              struct md_enqueue_info *minfo)
3376 {
3377         struct md_op_data *op_data = &minfo->mi_data;
3378         struct obd_device *obd = exp->exp_obd;
3379         struct lmv_obd *lmv = &obd->u.lmv;
3380         struct lmv_tgt_desc *tgt = NULL;
3381         int rc;
3382         ENTRY;
3383
3384         if (!fid_is_sane(&op_data->op_fid2))
3385                 RETURN(-EINVAL);
3386
3387         tgt = lmv_find_target(lmv, &op_data->op_fid1);
3388         if (IS_ERR(tgt))
3389                 RETURN(PTR_ERR(tgt));
3390
3391         /*
3392          * no special handle for remote dir, which needs to fetch both LOOKUP
3393          * lock on parent, and then UPDATE lock on child MDT, which makes all
3394          * complicated because this is done async. So only LOOKUP lock is
3395          * fetched for remote dir, but considering remote dir is rare case,
3396          * and not supporting it in statahead won't cause any issue, just leave
3397          * it as is.
3398          */
3399
3400         rc = md_intent_getattr_async(tgt->ltd_exp, minfo);
3401         RETURN(rc);
3402 }
3403
3404 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3405                         struct lu_fid *fid, __u64 *bits)
3406 {
3407         struct obd_device       *obd = exp->exp_obd;
3408         struct lmv_obd          *lmv = &obd->u.lmv;
3409         struct lmv_tgt_desc     *tgt;
3410         int                      rc;
3411         ENTRY;
3412
3413         tgt = lmv_find_target(lmv, fid);
3414         if (IS_ERR(tgt))
3415                 RETURN(PTR_ERR(tgt));
3416
3417         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3418         RETURN(rc);
3419 }
3420
3421 int lmv_get_fid_from_lsm(struct obd_export *exp,
3422                          const struct lmv_stripe_md *lsm,
3423                          const char *name, int namelen, struct lu_fid *fid)
3424 {
3425         const struct lmv_oinfo *oinfo;
3426
3427         LASSERT(lmv_dir_striped(lsm));
3428
3429         oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
3430         if (IS_ERR(oinfo))
3431                 return PTR_ERR(oinfo);
3432
3433         *fid = oinfo->lmo_fid;
3434
3435         RETURN(0);
3436 }
3437
3438 /**
3439  * For lmv, only need to send request to master MDT, and the master MDT will
3440  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3441  * we directly fetch data from the slave MDTs.
3442  */
3443 int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3444                  struct obd_quotactl *oqctl)
3445 {
3446         struct obd_device   *obd = class_exp2obd(exp);
3447         struct lmv_obd      *lmv = &obd->u.lmv;
3448         struct lmv_tgt_desc *tgt = lmv->tgts[0];
3449         int                  rc = 0;
3450         __u32                i;
3451         __u64                curspace, curinodes;
3452         ENTRY;
3453
3454         if (tgt == NULL ||
3455             tgt->ltd_exp == NULL ||
3456             !tgt->ltd_active ||
3457             lmv->desc.ld_tgt_count == 0) {
3458                 CERROR("master lmv inactive\n");
3459                 RETURN(-EIO);
3460         }
3461
3462         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3463                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3464                 RETURN(rc);
3465         }
3466
3467         curspace = curinodes = 0;
3468         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3469                 int err;
3470                 tgt = lmv->tgts[i];
3471
3472                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
3473                         continue;
3474
3475                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3476                 if (err) {
3477                         CERROR("getquota on mdt %d failed. %d\n", i, err);
3478                         if (!rc)
3479                                 rc = err;
3480                 } else {
3481                         curspace += oqctl->qc_dqblk.dqb_curspace;
3482                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3483                 }
3484         }
3485         oqctl->qc_dqblk.dqb_curspace = curspace;
3486         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3487
3488         RETURN(rc);
3489 }
3490
3491 static int lmv_merge_attr(struct obd_export *exp,
3492                           const struct lmv_stripe_md *lsm,
3493                           struct cl_attr *attr,
3494                           ldlm_blocking_callback cb_blocking)
3495 {
3496         int rc;
3497         int i;
3498
3499         if (!lmv_dir_striped(lsm))
3500                 return 0;
3501
3502         rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3503         if (rc < 0)
3504                 return rc;
3505
3506         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3507                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3508
3509                 if (!inode)
3510                         continue;
3511
3512                 CDEBUG(D_INFO,
3513                        "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
3514                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3515                        i_size_read(inode), (unsigned long long)inode->i_blocks,
3516                        inode->i_nlink, (s64)inode->i_atime.tv_sec,
3517                        (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
3518
3519                 /* for slave stripe, it needs to subtract nlink for . and .. */
3520                 if (i != 0)
3521                         attr->cat_nlink += inode->i_nlink - 2;
3522                 else
3523                         attr->cat_nlink = inode->i_nlink;
3524
3525                 attr->cat_size += i_size_read(inode);
3526                 attr->cat_blocks += inode->i_blocks;
3527
3528                 if (attr->cat_atime < inode->i_atime.tv_sec)
3529                         attr->cat_atime = inode->i_atime.tv_sec;
3530
3531                 if (attr->cat_ctime < inode->i_ctime.tv_sec)
3532                         attr->cat_ctime = inode->i_ctime.tv_sec;
3533
3534                 if (attr->cat_mtime < inode->i_mtime.tv_sec)
3535                         attr->cat_mtime = inode->i_mtime.tv_sec;
3536         }
3537         return 0;
3538 }
3539
3540 struct obd_ops lmv_obd_ops = {
3541         .o_owner                = THIS_MODULE,
3542         .o_setup                = lmv_setup,
3543         .o_cleanup              = lmv_cleanup,
3544         .o_precleanup           = lmv_precleanup,
3545         .o_process_config       = lmv_process_config,
3546         .o_connect              = lmv_connect,
3547         .o_disconnect           = lmv_disconnect,
3548         .o_statfs               = lmv_statfs,
3549         .o_get_info             = lmv_get_info,
3550         .o_set_info_async       = lmv_set_info_async,
3551         .o_notify               = lmv_notify,
3552         .o_get_uuid             = lmv_get_uuid,
3553         .o_fid_alloc            = lmv_fid_alloc,
3554         .o_iocontrol            = lmv_iocontrol,
3555         .o_quotactl             = lmv_quotactl
3556 };
3557
3558 struct md_ops lmv_md_ops = {
3559         .m_get_root             = lmv_get_root,
3560         .m_null_inode           = lmv_null_inode,
3561         .m_close                = lmv_close,
3562         .m_create               = lmv_create,
3563         .m_enqueue              = lmv_enqueue,
3564         .m_getattr              = lmv_getattr,
3565         .m_getxattr             = lmv_getxattr,
3566         .m_getattr_name         = lmv_getattr_name,
3567         .m_intent_lock          = lmv_intent_lock,
3568         .m_link                 = lmv_link,
3569         .m_rename               = lmv_rename,
3570         .m_setattr              = lmv_setattr,
3571         .m_setxattr             = lmv_setxattr,
3572         .m_fsync                = lmv_fsync,
3573         .m_file_resync          = lmv_file_resync,
3574         .m_read_page            = lmv_read_page,
3575         .m_unlink               = lmv_unlink,
3576         .m_init_ea_size         = lmv_init_ea_size,
3577         .m_cancel_unused        = lmv_cancel_unused,
3578         .m_set_lock_data        = lmv_set_lock_data,
3579         .m_lock_match           = lmv_lock_match,
3580         .m_get_lustre_md        = lmv_get_lustre_md,
3581         .m_free_lustre_md       = lmv_free_lustre_md,
3582         .m_merge_attr           = lmv_merge_attr,
3583         .m_set_open_replay_data = lmv_set_open_replay_data,
3584         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3585         .m_intent_getattr_async = lmv_intent_getattr_async,
3586         .m_revalidate_lock      = lmv_revalidate_lock,
3587         .m_get_fid_from_lsm     = lmv_get_fid_from_lsm,
3588         .m_unpackmd             = lmv_unpackmd,
3589 };
3590
3591 static int __init lmv_init(void)
3592 {
3593         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true, NULL,
3594                                    LUSTRE_LMV_NAME, NULL);
3595 }
3596
3597 static void __exit lmv_exit(void)
3598 {
3599         class_unregister_type(LUSTRE_LMV_NAME);
3600 }
3601
3602 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3603 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
3604 MODULE_VERSION(LUSTRE_VERSION_STRING);
3605 MODULE_LICENSE("GPL");
3606
3607 module_init(lmv_init);
3608 module_exit(lmv_exit);