lustre/lmv/lmv_obd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LMV
  38 #ifdef __KERNEL__
  39 #include <linux/slab.h>
  40 #include <linux/module.h>
  41 #include <linux/init.h>
  42 #include <linux/slab.h>
  43 #include <linux/pagemap.h>
  44 #include <linux/mm.h>
  45 #include <linux/math64.h>
  46 #include <linux/seq_file.h>
  47 #include <linux/namei.h>
  48 #else
  49 #include <liblustre.h>
  50 #endif
  51
  52 #include <lustre/lustre_idl.h>
  53 #include <obd_support.h>
  54 #include <lustre_lib.h>
  55 #include <lustre_net.h>
  56 #include <obd_class.h>
  57 #include <lustre_lmv.h>
  58 #include <lprocfs_status.h>
  59 #include <cl_object.h>
  60 #include <lclient.h>
  61 #include <lustre_lite.h>
  62 #include <lustre_fid.h>
  63 #include "lmv_internal.h"
  64
  65 /* This hash is only for testing purpose */
  66 static inline unsigned int
  67 lmv_hash_all_chars(unsigned int count, const char *name, int namelen)
  68 {
  69         unsigned int c = 0;
  70         const unsigned char *p = (const unsigned char *)name;
  71
  72         while (--namelen >= 0)
  73                 c += p[namelen];
  74
  75         c = c % count;
  76
  77         return c;
  78 }
  79
  80 static inline unsigned int
  81 lmv_hash_fnv1a(unsigned int count, const char *name, int namelen)
  82 {
  83         __u64   hash;
  84
  85         hash = lustre_hash_fnv_1a_64(name, namelen);
  86
  87         hash = hash % count;
  88
  89         return hash;
  90 }
  91
  92 int lmv_name_to_stripe_index(enum lmv_hash_type hashtype,
  93                              unsigned int max_mdt_index,
  94                              const char *name, int namelen)
  95 {
  96         int     idx;
  97
  98         LASSERT(namelen > 0);
  99         if (max_mdt_index <= 1)
 100                 return 0;
 101
 102         switch (hashtype) {
 103         case LMV_HASH_TYPE_ALL_CHARS:
 104                 idx = lmv_hash_all_chars(max_mdt_index, name, namelen);
 105                 break;
 106         case LMV_HASH_TYPE_FNV_1A_64:
 107                 idx = lmv_hash_fnv1a(max_mdt_index, name, namelen);
 108                 break;
 109         /* LMV_HASH_TYPE_MIGRATION means the file is being migrated,
 110          * and the file should be accessed by client, except for
 111          * lookup(see lmv_intent_lookup), return -EACCES here */
 112         case LMV_HASH_TYPE_MIGRATION:
 113                 CERROR("%.*s is being migrated: rc = %d\n", namelen,
 114                        name, -EACCES);
 115                 return -EACCES;
 116         default:
 117                 CERROR("Unknown hash type 0x%x\n", hashtype);
 118                 return -EINVAL;
 119         }
 120
 121         CDEBUG(D_INFO, "name %.*s hash_type %d idx %d\n", namelen, name,
 122                hashtype, idx);
 123
 124         LASSERT(idx < max_mdt_index);
 125         return idx;
 126 }
 127
 128 static void lmv_activate_target(struct lmv_obd *lmv,
 129                                 struct lmv_tgt_desc *tgt,
 130                                 int activate)
 131 {
 132         if (tgt->ltd_active == activate)
 133                 return;
 134
 135         tgt->ltd_active = activate;
 136         lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
 137 }
 138
 139 /**
 140  * Error codes:
 141  *
 142  *  -EINVAL  : UUID can't be found in the LMV's target list
 143  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
 144  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
 145  */
 146 static int lmv_set_mdc_active(struct lmv_obd *lmv,
 147                               const struct obd_uuid *uuid,
 148                               int activate)
 149 {
 150         struct lmv_tgt_desc     *tgt = NULL;
 151         struct obd_device       *obd;
 152         __u32                    i;
 153         int                      rc = 0;
 154         ENTRY;
 155
 156         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
 157                         lmv, uuid->uuid, activate);
 158
 159         spin_lock(&lmv->lmv_lock);
 160         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 161                 tgt = lmv->tgts[i];
 162                 if (tgt == NULL || tgt->ltd_exp == NULL)
 163                         continue;
 164
 165                 CDEBUG(D_INFO, "Target idx %d is %s conn "LPX64"\n", i,
 166                        tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
 167
 168                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 169                         break;
 170         }
 171
 172         if (i == lmv->desc.ld_tgt_count)
 173                 GOTO(out_lmv_lock, rc = -EINVAL);
 174
 175         obd = class_exp2obd(tgt->ltd_exp);
 176         if (obd == NULL)
 177                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 178
 179         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 180                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 181                obd->obd_type->typ_name, i);
 182         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 183
 184         if (tgt->ltd_active == activate) {
 185                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 186                        activate ? "" : "in");
 187                 GOTO(out_lmv_lock, rc);
 188         }
 189
 190         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 191                activate ? "" : "in");
 192         lmv_activate_target(lmv, tgt, activate);
 193         EXIT;
 194
 195  out_lmv_lock:
 196         spin_unlock(&lmv->lmv_lock);
 197         return rc;
 198 }
 199
 200 struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
 201 {
 202         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
 203         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
 204
 205         return (tgt == NULL) ? NULL : obd_get_uuid(tgt->ltd_exp);
 206 }
 207
 208 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 209                       enum obd_notify_event ev, void *data)
 210 {
 211         struct obd_connect_data *conn_data;
 212         struct lmv_obd          *lmv = &obd->u.lmv;
 213         struct obd_uuid         *uuid;
 214         int                      rc = 0;
 215         ENTRY;
 216
 217         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 218                 CERROR("unexpected notification of %s %s!\n",
 219                        watched->obd_type->typ_name,
 220                        watched->obd_name);
 221                 RETURN(-EINVAL);
 222         }
 223
 224         uuid = &watched->u.cli.cl_target_uuid;
 225         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 226                 /*
 227                  * Set MDC as active before notifying the observer, so the
 228                  * observer can use the MDC normally.
 229                  */
 230                 rc = lmv_set_mdc_active(lmv, uuid,
 231                                         ev == OBD_NOTIFY_ACTIVE);
 232                 if (rc) {
 233                         CERROR("%sactivation of %s failed: %d\n",
 234                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 235                                uuid->uuid, rc);
 236                         RETURN(rc);
 237                 }
 238         } else if (ev == OBD_NOTIFY_OCD) {
 239                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 240                 /*
 241                  * XXX: Make sure that ocd_connect_flags from all targets are
 242                  * the same. Otherwise one of MDTs runs wrong version or
 243                  * something like this.  --umka
 244                  */
 245                 obd->obd_self_export->exp_connect_data = *conn_data;
 246         }
 247 #if 0
 248         else if (ev == OBD_NOTIFY_DISCON) {
 249                 /*
 250                  * For disconnect event, flush fld cache for failout MDS case.
 251                  */
 252                 fld_client_flush(&lmv->lmv_fld);
 253         }
 254 #endif
 255         /*
 256          * Pass the notification up the chain.
 257          */
 258         if (obd->obd_observer)
 259                 rc = obd_notify(obd->obd_observer, watched, ev, data);
 260
 261         RETURN(rc);
 262 }
 263
 264 /**
 265  * This is fake connect function. Its purpose is to initialize lmv and say
 266  * caller that everything is okay. Real connection will be performed later.
 267  */
 268 static int lmv_connect(const struct lu_env *env,
 269                        struct obd_export **exp, struct obd_device *obd,
 270                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 271                        void *localdata)
 272 {
 273         struct lmv_obd        *lmv = &obd->u.lmv;
 274         struct lustre_handle  conn = { 0 };
 275         int                    rc = 0;
 276         ENTRY;
 277
 278         /*
 279          * We don't want to actually do the underlying connections more than
 280          * once, so keep track.
 281          */
 282         lmv->refcount++;
 283         if (lmv->refcount > 1) {
 284                 *exp = NULL;
 285                 RETURN(0);
 286         }
 287
 288         rc = class_connect(&conn, obd, cluuid);
 289         if (rc) {
 290                 CERROR("class_connection() returned %d\n", rc);
 291                 RETURN(rc);
 292         }
 293
 294         *exp = class_conn2export(&conn);
 295         class_export_get(*exp);
 296
 297         lmv->exp = *exp;
 298         lmv->connected = 0;
 299         lmv->cluuid = *cluuid;
 300
 301         if (data)
 302                 lmv->conn_data = *data;
 303
 304         if (obd->obd_type->typ_procsym == NULL) {
 305                 obd->obd_type->typ_procsym = lprocfs_seq_register("target_obds",
 306                                                          obd->obd_proc_entry,
 307                                                          NULL, NULL);
 308                 if (IS_ERR(obd->obd_type->typ_procsym)) {
 309                         CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
 310                                obd->obd_type->typ_name, obd->obd_name);
 311                         obd->obd_type->typ_procsym = NULL;
 312                 }
 313         }
 314
 315         /*
 316          * All real clients should perform actual connection right away, because
 317          * it is possible, that LMV will not have opportunity to connect targets
 318          * and MDC stuff will be called directly, for instance while reading
 319          * ../mdc/../kbytesfree procfs file, etc.
 320          */
 321         if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_REAL))
 322                 rc = lmv_check_connect(obd);
 323
 324         if (rc && obd->obd_type->typ_procsym != NULL)
 325                 lprocfs_remove(&obd->obd_type->typ_procsym);
 326         RETURN(rc);
 327 }
 328
 329 static void lmv_set_timeouts(struct obd_device *obd)
 330 {
 331         struct lmv_obd          *lmv;
 332         __u32                    i;
 333
 334         lmv = &obd->u.lmv;
 335         if (lmv->server_timeout == 0)
 336                 return;
 337
 338         if (lmv->connected == 0)
 339                 return;
 340
 341         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 342                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 343
 344                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
 345                         continue;
 346
 347                 obd_set_info_async(NULL, tgt->ltd_exp, sizeof(KEY_INTERMDS),
 348                                    KEY_INTERMDS, 0, NULL, NULL);
 349         }
 350 }
 351
 352 static int lmv_init_ea_size(struct obd_export *exp, int easize,
 353                             int def_easize, int cookiesize, int def_cookiesize)
 354 {
 355         struct obd_device       *obd = exp->exp_obd;
 356         struct lmv_obd          *lmv = &obd->u.lmv;
 357         __u32                    i;
 358         int                      rc = 0;
 359         int                      change = 0;
 360         ENTRY;
 361
 362         if (lmv->max_easize < easize) {
 363                 lmv->max_easize = easize;
 364                 change = 1;
 365         }
 366         if (lmv->max_def_easize < def_easize) {
 367                 lmv->max_def_easize = def_easize;
 368                 change = 1;
 369         }
 370         if (lmv->max_cookiesize < cookiesize) {
 371                 lmv->max_cookiesize = cookiesize;
 372                 change = 1;
 373         }
 374         if (lmv->max_def_cookiesize < def_cookiesize) {
 375                 lmv->max_def_cookiesize = def_cookiesize;
 376                 change = 1;
 377         }
 378         if (change == 0)
 379                 RETURN(0);
 380
 381         if (lmv->connected == 0)
 382                 RETURN(0);
 383
 384         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 385                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 386
 387                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active) {
 388                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
 389                         continue;
 390                 }
 391
 392                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize,
 393                                      cookiesize, def_cookiesize);
 394                 if (rc) {
 395                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
 396                                " rc = %d.\n", obd->obd_name, i, rc);
 397                         break;
 398                 }
 399         }
 400         RETURN(rc);
 401 }
 402
 403 #define MAX_STRING_SIZE 128
 404
 405 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 406 {
 407         struct lmv_obd          *lmv = &obd->u.lmv;
 408         struct obd_uuid         *cluuid = &lmv->cluuid;
 409         struct obd_uuid          lmv_mdc_uuid = { "LMV_MDC_UUID" };
 410         struct obd_device       *mdc_obd;
 411         struct obd_export       *mdc_exp;
 412         struct lu_fld_target     target;
 413         int                      rc;
 414         ENTRY;
 415
 416         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 417                                         &obd->obd_uuid);
 418         if (!mdc_obd) {
 419                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 420                 RETURN(-EINVAL);
 421         }
 422
 423         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
 424                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 425                 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
 426                 cluuid->uuid);
 427
 428         if (!mdc_obd->obd_set_up) {
 429                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 430                 RETURN(-EINVAL);
 431         }
 432
 433         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &lmv_mdc_uuid,
 434                          &lmv->conn_data, NULL);
 435         if (rc) {
 436                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 437                 RETURN(rc);
 438         }
 439
 440         /*
 441          * Init fid sequence client for this mdc and add new fld target.
 442          */
 443         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
 444         if (rc)
 445                 RETURN(rc);
 446
 447         target.ft_srv = NULL;
 448         target.ft_exp = mdc_exp;
 449         target.ft_idx = tgt->ltd_idx;
 450
 451         fld_client_add_target(&lmv->lmv_fld, &target);
 452
 453         rc = obd_register_observer(mdc_obd, obd);
 454         if (rc) {
 455                 obd_disconnect(mdc_exp);
 456                 CERROR("target %s register_observer error %d\n",
 457                        tgt->ltd_uuid.uuid, rc);
 458                 RETURN(rc);
 459         }
 460
 461         if (obd->obd_observer) {
 462                 /*
 463                  * Tell the observer about the new target.
 464                  */
 465                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 466                                 OBD_NOTIFY_ACTIVE,
 467                                 (void *)(tgt - lmv->tgts[0]));
 468                 if (rc) {
 469                         obd_disconnect(mdc_exp);
 470                         RETURN(rc);
 471                 }
 472         }
 473
 474         tgt->ltd_active = 1;
 475         tgt->ltd_exp = mdc_exp;
 476         lmv->desc.ld_active_tgt_count++;
 477
 478         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize,
 479                         lmv->max_cookiesize, lmv->max_def_cookiesize);
 480
 481         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 482                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 483                 atomic_read(&obd->obd_refcount));
 484
 485         if (obd->obd_type->typ_procsym != NULL) {
 486                 struct proc_dir_entry *mdc_symlink;
 487
 488                 LASSERT(mdc_obd->obd_type != NULL);
 489                 LASSERT(mdc_obd->obd_type->typ_name != NULL);
 490                 mdc_symlink = lprocfs_add_symlink(mdc_obd->obd_name,
 491                                                   obd->obd_type->typ_procsym,
 492                                                   "../../../%s/%s",
 493                                                   mdc_obd->obd_type->typ_name,
 494                                                   mdc_obd->obd_name);
 495                 if (mdc_symlink == NULL) {
 496                         CERROR("Could not register LMV target "
 497                                "/proc/fs/lustre/%s/%s/target_obds/%s.",
 498                                obd->obd_type->typ_name, obd->obd_name,
 499                                mdc_obd->obd_name);
 500                         lprocfs_remove(&obd->obd_type->typ_procsym);
 501                 }
 502         }
 503         RETURN(0);
 504 }
 505
 506 static void lmv_del_target(struct lmv_obd *lmv, int index)
 507 {
 508         if (lmv->tgts[index] == NULL)
 509                 return;
 510
 511         OBD_FREE_PTR(lmv->tgts[index]);
 512         lmv->tgts[index] = NULL;
 513         return;
 514 }
 515
 516 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 517                            __u32 index, int gen)
 518 {
 519         struct lmv_obd      *lmv = &obd->u.lmv;
 520         struct lmv_tgt_desc *tgt;
 521         int                  rc = 0;
 522         ENTRY;
 523
 524         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
 525
 526         lmv_init_lock(lmv);
 527
 528         if (lmv->desc.ld_tgt_count == 0) {
 529                 struct obd_device *mdc_obd;
 530
 531                 mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
 532                                                 &obd->obd_uuid);
 533                 if (!mdc_obd) {
 534                         lmv_init_unlock(lmv);
 535                         CERROR("%s: Target %s not attached: rc = %d\n",
 536                                obd->obd_name, uuidp->uuid, -EINVAL);
 537                         RETURN(-EINVAL);
 538                 }
 539         }
 540
 541         if ((index < lmv->tgts_size) && (lmv->tgts[index] != NULL)) {
 542                 tgt = lmv->tgts[index];
 543                 CERROR("%s: UUID %s already assigned at LOV target index %d:"
 544                        " rc = %d\n", obd->obd_name,
 545                        obd_uuid2str(&tgt->ltd_uuid), index, -EEXIST);
 546                 lmv_init_unlock(lmv);
 547                 RETURN(-EEXIST);
 548         }
 549
 550         if (index >= lmv->tgts_size) {
 551                 /* We need to reallocate the lmv target array. */
 552                 struct lmv_tgt_desc **newtgts, **old = NULL;
 553                 __u32 newsize = 1;
 554                 __u32 oldsize = 0;
 555
 556                 while (newsize < index + 1)
 557                         newsize = newsize << 1;
 558                 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
 559                 if (newtgts == NULL) {
 560                         lmv_init_unlock(lmv);
 561                         RETURN(-ENOMEM);
 562                 }
 563
 564                 if (lmv->tgts_size) {
 565                         memcpy(newtgts, lmv->tgts,
 566                                sizeof(*newtgts) * lmv->tgts_size);
 567                         old = lmv->tgts;
 568                         oldsize = lmv->tgts_size;
 569                 }
 570
 571                 lmv->tgts = newtgts;
 572                 lmv->tgts_size = newsize;
 573                 smp_rmb();
 574                 if (old)
 575                         OBD_FREE(old, sizeof(*old) * oldsize);
 576
 577                 CDEBUG(D_CONFIG, "tgts: %p size: %d\n", lmv->tgts,
 578                        lmv->tgts_size);
 579         }
 580
 581         OBD_ALLOC_PTR(tgt);
 582         if (!tgt) {
 583                 lmv_init_unlock(lmv);
 584                 RETURN(-ENOMEM);
 585         }
 586
 587         mutex_init(&tgt->ltd_fid_mutex);
 588         tgt->ltd_idx = index;
 589         tgt->ltd_uuid = *uuidp;
 590         tgt->ltd_active = 0;
 591         lmv->tgts[index] = tgt;
 592         if (index >= lmv->desc.ld_tgt_count)
 593                 lmv->desc.ld_tgt_count = index + 1;
 594
 595         if (lmv->connected) {
 596                 rc = lmv_connect_mdc(obd, tgt);
 597                 if (rc) {
 598                         spin_lock(&lmv->lmv_lock);
 599                         lmv->desc.ld_tgt_count--;
 600                         memset(tgt, 0, sizeof(*tgt));
 601                         spin_unlock(&lmv->lmv_lock);
 602                 } else {
 603                         int easize = sizeof(struct lmv_stripe_md) +
 604                                 lmv->desc.ld_tgt_count * sizeof(struct lu_fid);
 605                         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0, 0);
 606                 }
 607         }
 608
 609         lmv_init_unlock(lmv);
 610         RETURN(rc);
 611 }
 612
 613 int lmv_check_connect(struct obd_device *obd)
 614 {
 615         struct lmv_obd          *lmv = &obd->u.lmv;
 616         struct lmv_tgt_desc     *tgt;
 617         __u32                    i;
 618         int                      rc;
 619         int                      easize;
 620         ENTRY;
 621
 622         if (lmv->connected)
 623                 RETURN(0);
 624
 625         lmv_init_lock(lmv);
 626         if (lmv->connected) {
 627                 lmv_init_unlock(lmv);
 628                 RETURN(0);
 629         }
 630
 631         if (lmv->desc.ld_tgt_count == 0) {
 632                 lmv_init_unlock(lmv);
 633                 CERROR("%s: no targets configured.\n", obd->obd_name);
 634                 RETURN(-EINVAL);
 635         }
 636
 637         LASSERT(lmv->tgts != NULL);
 638
 639         if (lmv->tgts[0] == NULL) {
 640                 lmv_init_unlock(lmv);
 641                 CERROR("%s: no target configured for index 0.\n",
 642                        obd->obd_name);
 643                 RETURN(-EINVAL);
 644         }
 645
 646         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 647                lmv->cluuid.uuid, obd->obd_name);
 648
 649         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 650                 tgt = lmv->tgts[i];
 651                 if (tgt == NULL)
 652                         continue;
 653                 rc = lmv_connect_mdc(obd, tgt);
 654                 if (rc)
 655                         GOTO(out_disc, rc);
 656         }
 657
 658         lmv_set_timeouts(obd);
 659         class_export_put(lmv->exp);
 660         lmv->connected = 1;
 661         easize = lmv_mds_md_size(lmv->desc.ld_tgt_count, LMV_MAGIC);
 662         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0, 0);
 663         lmv_init_unlock(lmv);
 664         RETURN(0);
 665
 666  out_disc:
 667         while (i-- > 0) {
 668                 int rc2;
 669                 tgt = lmv->tgts[i];
 670                 if (tgt == NULL)
 671                         continue;
 672                 tgt->ltd_active = 0;
 673                 if (tgt->ltd_exp) {
 674                         --lmv->desc.ld_active_tgt_count;
 675                         rc2 = obd_disconnect(tgt->ltd_exp);
 676                         if (rc2) {
 677                                 CERROR("LMV target %s disconnect on "
 678                                        "MDC idx %d: error %d\n",
 679                                        tgt->ltd_uuid.uuid, i, rc2);
 680                         }
 681                 }
 682         }
 683         class_disconnect(lmv->exp);
 684         lmv_init_unlock(lmv);
 685         RETURN(rc);
 686 }
 687
 688 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 689 {
 690         struct lmv_obd         *lmv = &obd->u.lmv;
 691         struct obd_device      *mdc_obd;
 692         int                     rc;
 693         ENTRY;
 694
 695         LASSERT(tgt != NULL);
 696         LASSERT(obd != NULL);
 697
 698         mdc_obd = class_exp2obd(tgt->ltd_exp);
 699
 700         if (mdc_obd) {
 701                 mdc_obd->obd_force = obd->obd_force;
 702                 mdc_obd->obd_fail = obd->obd_fail;
 703                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 704         }
 705
 706         if (obd->obd_type->typ_procsym != NULL)
 707                 lprocfs_remove_proc_entry(mdc_obd->obd_name,
 708                                           obd->obd_type->typ_procsym);
 709
 710         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
 711         if (rc)
 712                 CERROR("Can't finanize fids factory\n");
 713
 714         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 715                tgt->ltd_exp->exp_obd->obd_name,
 716                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 717
 718         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 719         rc = obd_disconnect(tgt->ltd_exp);
 720         if (rc) {
 721                 if (tgt->ltd_active) {
 722                         CERROR("Target %s disconnect error %d\n",
 723                                tgt->ltd_uuid.uuid, rc);
 724                 }
 725         }
 726
 727         lmv_activate_target(lmv, tgt, 0);
 728         tgt->ltd_exp = NULL;
 729         RETURN(0);
 730 }
 731
 732 static int lmv_disconnect(struct obd_export *exp)
 733 {
 734         struct obd_device       *obd = class_exp2obd(exp);
 735         struct lmv_obd          *lmv = &obd->u.lmv;
 736         int                      rc;
 737         __u32                    i;
 738         ENTRY;
 739
 740         if (!lmv->tgts)
 741                 goto out_local;
 742
 743         /*
 744          * Only disconnect the underlying layers on the final disconnect.
 745          */
 746         lmv->refcount--;
 747         if (lmv->refcount != 0)
 748                 goto out_local;
 749
 750         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 751                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
 752                         continue;
 753
 754                 lmv_disconnect_mdc(obd, lmv->tgts[i]);
 755         }
 756
 757         if (obd->obd_type->typ_procsym != NULL)
 758                 lprocfs_remove(&obd->obd_type->typ_procsym);
 759         else
 760                 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
 761                        obd->obd_type->typ_name, obd->obd_name);
 762
 763 out_local:
 764         /*
 765          * This is the case when no real connection is established by
 766          * lmv_check_connect().
 767          */
 768         if (!lmv->connected)
 769                 class_export_put(exp);
 770         rc = class_disconnect(exp);
 771         if (lmv->refcount == 0)
 772                 lmv->connected = 0;
 773         RETURN(rc);
 774 }
 775
 776 static int lmv_fid2path(struct obd_export *exp, int len, void *karg, void *uarg)
 777 {
 778         struct obd_device       *obddev = class_exp2obd(exp);
 779         struct lmv_obd          *lmv = &obddev->u.lmv;
 780         struct getinfo_fid2path *gf;
 781         struct lmv_tgt_desc     *tgt;
 782         struct getinfo_fid2path *remote_gf = NULL;
 783         int                     remote_gf_size = 0;
 784         int                     rc;
 785
 786         gf = (struct getinfo_fid2path *)karg;
 787         tgt = lmv_find_target(lmv, &gf->gf_fid);
 788         if (IS_ERR(tgt))
 789                 RETURN(PTR_ERR(tgt));
 790
 791 repeat_fid2path:
 792         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
 793         if (rc != 0 && rc != -EREMOTE)
 794                 GOTO(out_fid2path, rc);
 795
 796         /* If remote_gf != NULL, it means just building the
 797          * path on the remote MDT, copy this path segement to gf */
 798         if (remote_gf != NULL) {
 799                 struct getinfo_fid2path *ori_gf;
 800                 char *ptr;
 801
 802                 ori_gf = (struct getinfo_fid2path *)karg;
 803                 if (strlen(ori_gf->gf_path) +
 804                     strlen(gf->gf_path) > ori_gf->gf_pathlen)
 805                         GOTO(out_fid2path, rc = -EOVERFLOW);
 806
 807                 ptr = ori_gf->gf_path;
 808
 809                 memmove(ptr + strlen(gf->gf_path) + 1, ptr,
 810                         strlen(ori_gf->gf_path));
 811
 812                 strncpy(ptr, gf->gf_path, strlen(gf->gf_path));
 813                 ptr += strlen(gf->gf_path);
 814                 *ptr = '/';
 815         }
 816
 817         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: "LPU64" ln: %u\n",
 818                tgt->ltd_exp->exp_obd->obd_name,
 819                gf->gf_path, PFID(&gf->gf_fid), gf->gf_recno,
 820                gf->gf_linkno);
 821
 822         if (rc == 0)
 823                 GOTO(out_fid2path, rc);
 824
 825         /* sigh, has to go to another MDT to do path building further */
 826         if (remote_gf == NULL) {
 827                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
 828                 OBD_ALLOC(remote_gf, remote_gf_size);
 829                 if (remote_gf == NULL)
 830                         GOTO(out_fid2path, rc = -ENOMEM);
 831                 remote_gf->gf_pathlen = PATH_MAX;
 832         }
 833
 834         if (!fid_is_sane(&gf->gf_fid)) {
 835                 CERROR("%s: invalid FID "DFID": rc = %d\n",
 836                        tgt->ltd_exp->exp_obd->obd_name,
 837                        PFID(&gf->gf_fid), -EINVAL);
 838                 GOTO(out_fid2path, rc = -EINVAL);
 839         }
 840
 841         tgt = lmv_find_target(lmv, &gf->gf_fid);
 842         if (IS_ERR(tgt))
 843                 GOTO(out_fid2path, rc = -EINVAL);
 844
 845         remote_gf->gf_fid = gf->gf_fid;
 846         remote_gf->gf_recno = -1;
 847         remote_gf->gf_linkno = -1;
 848         memset(remote_gf->gf_path, 0, remote_gf->gf_pathlen);
 849         gf = remote_gf;
 850         goto repeat_fid2path;
 851
 852 out_fid2path:
 853         if (remote_gf != NULL)
 854                 OBD_FREE(remote_gf, remote_gf_size);
 855         RETURN(rc);
 856 }
 857
 858 static int lmv_hsm_req_count(struct lmv_obd *lmv,
 859                              const struct hsm_user_request *hur,
 860                              const struct lmv_tgt_desc *tgt_mds)
 861 {
 862         __u32                    i;
 863         int                      nr = 0;
 864         struct lmv_tgt_desc     *curr_tgt;
 865
 866         /* count how many requests must be sent to the given target */
 867         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
 868                 curr_tgt = lmv_find_target(lmv, &hur->hur_user_item[i].hui_fid);
 869                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
 870                         nr++;
 871         }
 872         return nr;
 873 }
 874
 875 static void lmv_hsm_req_build(struct lmv_obd *lmv,
 876                               struct hsm_user_request *hur_in,
 877                               const struct lmv_tgt_desc *tgt_mds,
 878                               struct hsm_user_request *hur_out)
 879 {
 880         __u32                    i, nr_out;
 881         struct lmv_tgt_desc     *curr_tgt;
 882
 883         /* build the hsm_user_request for the given target */
 884         hur_out->hur_request = hur_in->hur_request;
 885         nr_out = 0;
 886         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
 887                 curr_tgt = lmv_find_target(lmv,
 888                                            &hur_in->hur_user_item[i].hui_fid);
 889                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
 890                         hur_out->hur_user_item[nr_out] =
 891                                                 hur_in->hur_user_item[i];
 892                         nr_out++;
 893                 }
 894         }
 895         hur_out->hur_request.hr_itemcount = nr_out;
 896         memcpy(hur_data(hur_out), hur_data(hur_in),
 897                hur_in->hur_request.hr_data_len);
 898 }
 899
 900 static int lmv_hsm_ct_unregister(struct lmv_obd *lmv, unsigned int cmd, int len,
 901                                  struct lustre_kernelcomm *lk, void *uarg)
 902 {
 903         __u32                    i;
 904         int                      rc;
 905         struct kkuc_ct_data     *kcd = NULL;
 906         ENTRY;
 907
 908         /* unregister request (call from llapi_hsm_copytool_fini) */
 909         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 910                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 911
 912                 if (tgt == NULL || tgt->ltd_exp == NULL)
 913                         continue;
 914                 /* best effort: try to clean as much as possible
 915                  * (continue on error) */
 916                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 917         }
 918
 919         /* Whatever the result, remove copytool from kuc groups.
 920          * Unreached coordinators will get EPIPE on next requests
 921          * and will unregister automatically.
 922          */
 923         rc = libcfs_kkuc_group_rem(lk->lk_uid, lk->lk_group, (void **)&kcd);
 924         if (kcd != NULL)
 925                 OBD_FREE_PTR(kcd);
 926
 927         RETURN(rc);
 928 }
 929
 930 static int lmv_hsm_ct_register(struct lmv_obd *lmv, unsigned int cmd, int len,
 931                                struct lustre_kernelcomm *lk, void *uarg)
 932 {
 933         struct file             *filp;
 934         __u32                    i, j;
 935         int                      err, rc;
 936         bool                     any_set = false;
 937         struct kkuc_ct_data     *kcd;
 938         ENTRY;
 939
 940         /* All or nothing: try to register to all MDS.
 941          * In case of failure, unregister from previous MDS,
 942          * except if it because of inactive target. */
 943         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 944                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 945
 946                 if (tgt == NULL || tgt->ltd_exp == NULL)
 947                         continue;
 948                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 949                 if (err) {
 950                         if (tgt->ltd_active) {
 951                                 /* permanent error */
 952                                 CERROR("%s: iocontrol MDC %s on MDT"
 953                                        " idx %d cmd %x: err = %d\n",
 954                                        class_exp2obd(lmv->exp)->obd_name,
 955                                        tgt->ltd_uuid.uuid, i, cmd, err);
 956                                 rc = err;
 957                                 lk->lk_flags |= LK_FLG_STOP;
 958                                 /* unregister from previous MDS */
 959                                 for (j = 0; j < i; j++) {
 960                                         tgt = lmv->tgts[j];
 961                                         if (tgt == NULL || tgt->ltd_exp == NULL)
 962                                                 continue;
 963                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
 964                                                       lk, uarg);
 965                                 }
 966                                 RETURN(rc);
 967                         }
 968                         /* else: transient error.
 969                          * kuc will register to the missing MDT
 970                          * when it is back */
 971                 } else {
 972                         any_set = true;
 973                 }
 974         }
 975
 976         if (!any_set)
 977                 /* no registration done: return error */
 978                 RETURN(-ENOTCONN);
 979
 980         /* at least one registration done, with no failure */
 981         filp = fget(lk->lk_wfd);
 982         if (filp == NULL)
 983                 RETURN(-EBADF);
 984
 985         OBD_ALLOC_PTR(kcd);
 986         if (kcd == NULL) {
 987                 fput(filp);
 988                 RETURN(-ENOMEM);
 989         }
 990         kcd->kcd_magic = KKUC_CT_DATA_MAGIC;
 991         kcd->kcd_uuid = lmv->cluuid;
 992         kcd->kcd_archive = lk->lk_data;
 993
 994         rc = libcfs_kkuc_group_add(filp, lk->lk_uid, lk->lk_group, kcd);
 995         if (rc != 0) {
 996                 if (filp != NULL)
 997                         fput(filp);
 998                 OBD_FREE_PTR(kcd);
 999         }
1000
1001         RETURN(rc);
1002 }
1003
1004
1005
1006
1007 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
1008                          int len, void *karg, void *uarg)
1009 {
1010         struct obd_device       *obddev = class_exp2obd(exp);
1011         struct lmv_obd          *lmv = &obddev->u.lmv;
1012         struct lmv_tgt_desc     *tgt = NULL;
1013         __u32                    i = 0;
1014         int                      rc = 0;
1015         int                      set = 0;
1016         __u32                    count = lmv->desc.ld_tgt_count;
1017         ENTRY;
1018
1019         if (count == 0)
1020                 RETURN(-ENOTTY);
1021
1022         switch (cmd) {
1023         case IOC_OBD_STATFS: {
1024                 struct obd_ioctl_data *data = karg;
1025                 struct obd_device *mdc_obd;
1026                 struct obd_statfs stat_buf = {0};
1027                 __u32 index;
1028
1029                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
1030                 if ((index >= count))
1031                         RETURN(-ENODEV);
1032
1033                 tgt = lmv->tgts[index];
1034                 if (tgt == NULL || !tgt->ltd_active)
1035                         RETURN(-ENODATA);
1036
1037                 mdc_obd = class_exp2obd(tgt->ltd_exp);
1038                 if (!mdc_obd)
1039                         RETURN(-EINVAL);
1040
1041                 /* copy UUID */
1042                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
1043                                  min((int) data->ioc_plen2,
1044                                      (int) sizeof(struct obd_uuid))))
1045                         RETURN(-EFAULT);
1046
1047                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
1048                                 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
1049                                 0);
1050                 if (rc)
1051                         RETURN(rc);
1052                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
1053                                  min((int) data->ioc_plen1,
1054                                      (int) sizeof(stat_buf))))
1055                         RETURN(-EFAULT);
1056                 break;
1057         }
1058         case OBD_IOC_QUOTACTL: {
1059                 struct if_quotactl *qctl = karg;
1060                 struct obd_quotactl *oqctl;
1061
1062                 if (qctl->qc_valid == QC_MDTIDX) {
1063                         if (count <= qctl->qc_idx)
1064                                 RETURN(-EINVAL);
1065
1066                         tgt = lmv->tgts[qctl->qc_idx];
1067                         if (tgt == NULL || tgt->ltd_exp == NULL)
1068                                 RETURN(-EINVAL);
1069                 } else if (qctl->qc_valid == QC_UUID) {
1070                         for (i = 0; i < count; i++) {
1071                                 tgt = lmv->tgts[i];
1072                                 if (tgt == NULL)
1073                                         continue;
1074                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
1075                                                      &qctl->obd_uuid))
1076                                         continue;
1077
1078                                 if (tgt->ltd_exp == NULL)
1079                                         RETURN(-EINVAL);
1080
1081                                 break;
1082                         }
1083                 } else {
1084                         RETURN(-EINVAL);
1085                 }
1086
1087                 if (i >= count)
1088                         RETURN(-EAGAIN);
1089
1090                 LASSERT(tgt != NULL && tgt->ltd_exp != NULL);
1091                 OBD_ALLOC_PTR(oqctl);
1092                 if (!oqctl)
1093                         RETURN(-ENOMEM);
1094
1095                 QCTL_COPY(oqctl, qctl);
1096                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
1097                 if (rc == 0) {
1098                         QCTL_COPY(qctl, oqctl);
1099                         qctl->qc_valid = QC_MDTIDX;
1100                         qctl->obd_uuid = tgt->ltd_uuid;
1101                 }
1102                 OBD_FREE_PTR(oqctl);
1103                 break;
1104         }
1105         case OBD_IOC_CHANGELOG_SEND:
1106         case OBD_IOC_CHANGELOG_CLEAR: {
1107                 struct ioc_changelog *icc = karg;
1108
1109                 if (icc->icc_mdtindex >= count)
1110                         RETURN(-ENODEV);
1111
1112                 tgt = lmv->tgts[icc->icc_mdtindex];
1113                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
1114                         RETURN(-ENODEV);
1115                 rc = obd_iocontrol(cmd, tgt->ltd_exp, sizeof(*icc), icc, NULL);
1116                 break;
1117         }
1118         case LL_IOC_GET_CONNECT_FLAGS: {
1119                 tgt = lmv->tgts[0];
1120                 if (tgt == NULL || tgt->ltd_exp == NULL)
1121                         RETURN(-ENODATA);
1122                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1123                 break;
1124         }
1125         case OBD_IOC_FID2PATH: {
1126                 rc = lmv_fid2path(exp, len, karg, uarg);
1127                 break;
1128         }
1129         case LL_IOC_HSM_STATE_GET:
1130         case LL_IOC_HSM_STATE_SET:
1131         case LL_IOC_HSM_ACTION: {
1132                 struct md_op_data       *op_data = karg;
1133
1134                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1135                 if (IS_ERR(tgt))
1136                         RETURN(PTR_ERR(tgt));
1137
1138                 if (tgt->ltd_exp == NULL)
1139                         RETURN(-EINVAL);
1140
1141                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1142                 break;
1143         }
1144         case LL_IOC_HSM_PROGRESS: {
1145                 const struct hsm_progress_kernel *hpk = karg;
1146
1147                 tgt = lmv_find_target(lmv, &hpk->hpk_fid);
1148                 if (IS_ERR(tgt))
1149                         RETURN(PTR_ERR(tgt));
1150                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1151                 break;
1152         }
1153         case LL_IOC_HSM_REQUEST: {
1154                 struct hsm_user_request *hur = karg;
1155                 unsigned int reqcount = hur->hur_request.hr_itemcount;
1156
1157                 if (reqcount == 0)
1158                         RETURN(0);
1159
1160                 /* if the request is about a single fid
1161                  * or if there is a single MDS, no need to split
1162                  * the request. */
1163                 if (reqcount == 1 || count == 1) {
1164                         tgt = lmv_find_target(lmv,
1165                                               &hur->hur_user_item[0].hui_fid);
1166                         if (IS_ERR(tgt))
1167                                 RETURN(PTR_ERR(tgt));
1168                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1169                 } else {
1170                         /* split fid list to their respective MDS */
1171                         for (i = 0; i < count; i++) {
1172                                 unsigned int            nr, reqlen;
1173                                 int                     rc1;
1174                                 struct hsm_user_request *req;
1175
1176                                 tgt = lmv->tgts[i];
1177                                 if (tgt == NULL || tgt->ltd_exp == NULL)
1178                                         continue;
1179
1180                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
1181                                 if (nr == 0) /* nothing for this MDS */
1182                                         continue;
1183
1184                                 /* build a request with fids for this MDS */
1185                                 reqlen = offsetof(typeof(*hur),
1186                                                   hur_user_item[nr])
1187                                                 + hur->hur_request.hr_data_len;
1188                                 OBD_ALLOC_LARGE(req, reqlen);
1189                                 if (req == NULL)
1190                                         RETURN(-ENOMEM);
1191
1192                                 lmv_hsm_req_build(lmv, hur, tgt, req);
1193
1194                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
1195                                                     req, uarg);
1196                                 if (rc1 != 0 && rc == 0)
1197                                         rc = rc1;
1198                                 OBD_FREE_LARGE(req, reqlen);
1199                         }
1200                 }
1201                 break;
1202         }
1203         case LL_IOC_LOV_SWAP_LAYOUTS: {
1204                 struct md_op_data       *op_data = karg;
1205                 struct lmv_tgt_desc     *tgt1, *tgt2;
1206
1207                 tgt1 = lmv_find_target(lmv, &op_data->op_fid1);
1208                 if (IS_ERR(tgt1))
1209                         RETURN(PTR_ERR(tgt1));
1210
1211                 tgt2 = lmv_find_target(lmv, &op_data->op_fid2);
1212                 if (IS_ERR(tgt2))
1213                         RETURN(PTR_ERR(tgt2));
1214
1215                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1216                         RETURN(-EINVAL);
1217
1218                 /* only files on same MDT can have their layouts swapped */
1219                 if (tgt1->ltd_idx != tgt2->ltd_idx)
1220                         RETURN(-EPERM);
1221
1222                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1223                 break;
1224         }
1225         case LL_IOC_HSM_CT_START: {
1226                 struct lustre_kernelcomm *lk = karg;
1227                 if (lk->lk_flags & LK_FLG_STOP)
1228                         rc = lmv_hsm_ct_unregister(lmv, cmd, len, lk, uarg);
1229                 else
1230                         rc = lmv_hsm_ct_register(lmv, cmd, len, lk, uarg);
1231                 break;
1232         }
1233         default:
1234                 for (i = 0; i < count; i++) {
1235                         struct obd_device *mdc_obd;
1236                         int err;
1237
1238                         tgt = lmv->tgts[i];
1239                         if (tgt == NULL || tgt->ltd_exp == NULL)
1240                                 continue;
1241                         /* ll_umount_begin() sets force flag but for lmv, not
1242                          * mdc. Let's pass it through */
1243                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1244                         mdc_obd->obd_force = obddev->obd_force;
1245                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1246                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
1247                                 RETURN(err);
1248                         } else if (err) {
1249                                 if (tgt->ltd_active) {
1250                                         CERROR("error: iocontrol MDC %s on MDT"
1251                                                " idx %d cmd %x: err = %d\n",
1252                                                tgt->ltd_uuid.uuid, i, cmd, err);
1253                                         if (!rc)
1254                                                 rc = err;
1255                                 }
1256                         } else
1257                                 set = 1;
1258                 }
1259                 if (!set && !rc)
1260                         rc = -EIO;
1261         }
1262         RETURN(rc);
1263 }
1264
1265 #if 0
1266 static int lmv_all_chars_policy(int count, const char *name,
1267                                 int len)
1268 {
1269         unsigned int c = 0;
1270
1271         while (len > 0)
1272                 c += name[--len];
1273         c = c % count;
1274         return c;
1275 }
1276
1277 static int lmv_nid_policy(struct lmv_obd *lmv)
1278 {
1279         struct obd_import *imp;
1280         __u32              id;
1281
1282         /*
1283          * XXX: To get nid we assume that underlying obd device is mdc.
1284          */
1285         imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
1286         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
1287         return id % lmv->desc.ld_tgt_count;
1288 }
1289
1290 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
1291                           placement_policy_t placement)
1292 {
1293         switch (placement) {
1294         case PLACEMENT_CHAR_POLICY:
1295                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
1296                                             op_data->op_name,
1297                                             op_data->op_namelen);
1298         case PLACEMENT_NID_POLICY:
1299                 return lmv_nid_policy(lmv);
1300
1301         default:
1302                 break;
1303         }
1304
1305         CERROR("Unsupported placement policy %x\n", placement);
1306         return -EINVAL;
1307 }
1308 #endif
1309
1310 /**
1311  * This is _inode_ placement policy function (not name).
1312  */
1313 static int lmv_placement_policy(struct obd_device *obd,
1314                                 struct md_op_data *op_data,
1315                                 mdsno_t *mds)
1316 {
1317         struct lmv_obd          *lmv = &obd->u.lmv;
1318         ENTRY;
1319
1320         LASSERT(mds != NULL);
1321
1322         if (lmv->desc.ld_tgt_count == 1) {
1323                 *mds = 0;
1324                 RETURN(0);
1325         }
1326
1327         /**
1328          * If stripe_offset is provided during setdirstripe
1329          * (setdirstripe -i xx), xx MDS will be choosen.
1330          */
1331         if (op_data->op_cli_flags & CLI_SET_MEA && op_data->op_data != NULL) {
1332                 struct lmv_user_md *lum;
1333
1334                 lum = op_data->op_data;
1335
1336                 if (le32_to_cpu(lum->lum_stripe_offset) != (__u32)-1) {
1337                         *mds = le32_to_cpu(lum->lum_stripe_offset);
1338                 } else {
1339                         /* -1 means default, which will be in the same MDT with
1340                          * the stripe */
1341                         *mds = op_data->op_mds;
1342                         lum->lum_stripe_offset = cpu_to_le32(op_data->op_mds);
1343                 }
1344         } else {
1345                 /* Allocate new fid on target according to operation type and
1346                  * parent home mds. */
1347                 *mds = op_data->op_mds;
1348         }
1349
1350         RETURN(0);
1351 }
1352
1353 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
1354                     mdsno_t mds)
1355 {
1356         struct lmv_tgt_desc     *tgt;
1357         int                      rc;
1358         ENTRY;
1359
1360         tgt = lmv_get_target(lmv, mds);
1361         if (IS_ERR(tgt))
1362                 RETURN(PTR_ERR(tgt));
1363
1364         /*
1365          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1366          * on server that seq in new allocated fid is not yet known.
1367          */
1368         mutex_lock(&tgt->ltd_fid_mutex);
1369
1370         if (tgt->ltd_active == 0 || tgt->ltd_exp == NULL)
1371                 GOTO(out, rc = -ENODEV);
1372
1373         /*
1374          * Asking underlaying tgt layer to allocate new fid.
1375          */
1376         rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
1377         if (rc > 0) {
1378                 LASSERT(fid_is_sane(fid));
1379                 rc = 0;
1380         }
1381
1382         EXIT;
1383 out:
1384         mutex_unlock(&tgt->ltd_fid_mutex);
1385         return rc;
1386 }
1387
1388 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
1389                   struct md_op_data *op_data)
1390 {
1391         struct obd_device     *obd = class_exp2obd(exp);
1392         struct lmv_obd        *lmv = &obd->u.lmv;
1393         mdsno_t                mds = 0;
1394         int                    rc;
1395         ENTRY;
1396
1397         LASSERT(op_data != NULL);
1398         LASSERT(fid != NULL);
1399
1400         rc = lmv_placement_policy(obd, op_data, &mds);
1401         if (rc) {
1402                 CERROR("Can't get target for allocating fid, "
1403                        "rc %d\n", rc);
1404                 RETURN(rc);
1405         }
1406
1407         rc = __lmv_fid_alloc(lmv, fid, mds);
1408         if (rc) {
1409                 CERROR("Can't alloc new fid, rc %d\n", rc);
1410                 RETURN(rc);
1411         }
1412
1413         RETURN(rc);
1414 }
1415
1416 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1417 {
1418         struct lmv_obd  *lmv = &obd->u.lmv;
1419         struct lmv_desc *desc;
1420         int             rc;
1421         ENTRY;
1422
1423         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1424                 CERROR("LMV setup requires a descriptor\n");
1425                 RETURN(-EINVAL);
1426         }
1427
1428         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1429         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1430                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1431                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1432                 RETURN(-EINVAL);
1433         }
1434
1435         OBD_ALLOC(lmv->tgts, sizeof(*lmv->tgts) * 32);
1436         if (lmv->tgts == NULL)
1437                 RETURN(-ENOMEM);
1438         lmv->tgts_size = 32;
1439
1440         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1441         lmv->desc.ld_tgt_count = 0;
1442         lmv->desc.ld_active_tgt_count = 0;
1443         lmv->max_cookiesize = 0;
1444         lmv->max_def_easize = 0;
1445         lmv->max_easize = 0;
1446         lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1447
1448         spin_lock_init(&lmv->lmv_lock);
1449         mutex_init(&lmv->init_mutex);
1450
1451 #ifdef LPROCFS
1452         obd->obd_vars = lprocfs_lmv_obd_vars;
1453         lprocfs_seq_obd_setup(obd);
1454         lprocfs_alloc_md_stats(obd, 0);
1455         rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd",
1456                                 0444, &lmv_proc_target_fops, obd);
1457         if (rc)
1458                 CWARN("%s: error adding LMV target_obd file: rc = %d\n",
1459                       obd->obd_name, rc);
1460 #endif
1461         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1462                              LUSTRE_CLI_FLD_HASH_DHT);
1463         if (rc) {
1464                 CERROR("Can't init FLD, err %d\n", rc);
1465                 GOTO(out, rc);
1466         }
1467
1468         RETURN(0);
1469
1470 out:
1471         return rc;
1472 }
1473
1474 static int lmv_cleanup(struct obd_device *obd)
1475 {
1476         struct lmv_obd   *lmv = &obd->u.lmv;
1477         ENTRY;
1478
1479         fld_client_fini(&lmv->lmv_fld);
1480         if (lmv->tgts != NULL) {
1481                 int i;
1482                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1483                         if (lmv->tgts[i] == NULL)
1484                                 continue;
1485                         lmv_del_target(lmv, i);
1486                 }
1487                 OBD_FREE(lmv->tgts, sizeof(*lmv->tgts) * lmv->tgts_size);
1488                 lmv->tgts_size = 0;
1489         }
1490         RETURN(0);
1491 }
1492
1493 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1494 {
1495         struct lustre_cfg       *lcfg = buf;
1496         struct obd_uuid         obd_uuid;
1497         int                     gen;
1498         __u32                   index;
1499         int                     rc;
1500         ENTRY;
1501
1502         switch (lcfg->lcfg_command) {
1503         case LCFG_ADD_MDC:
1504                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1505                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1506                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1507                         GOTO(out, rc = -EINVAL);
1508
1509                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1510
1511                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1512                         GOTO(out, rc = -EINVAL);
1513                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1514                         GOTO(out, rc = -EINVAL);
1515                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1516                 GOTO(out, rc);
1517         default:
1518                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1519                 GOTO(out, rc = -EINVAL);
1520         }
1521 out:
1522         RETURN(rc);
1523 }
1524
1525 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1526                       struct obd_statfs *osfs, __u64 max_age, __u32 flags)
1527 {
1528         struct obd_device       *obd = class_exp2obd(exp);
1529         struct lmv_obd          *lmv = &obd->u.lmv;
1530         struct obd_statfs       *temp;
1531         int                      rc = 0;
1532         __u32                    i;
1533         ENTRY;
1534
1535         rc = lmv_check_connect(obd);
1536         if (rc)
1537                 RETURN(rc);
1538
1539         OBD_ALLOC(temp, sizeof(*temp));
1540         if (temp == NULL)
1541                 RETURN(-ENOMEM);
1542
1543         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1544                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
1545                         continue;
1546
1547                 rc = obd_statfs(env, lmv->tgts[i]->ltd_exp, temp,
1548                                 max_age, flags);
1549                 if (rc) {
1550                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1551                                lmv->tgts[i]->ltd_exp->exp_obd->obd_name,
1552                                rc);
1553                         GOTO(out_free_temp, rc);
1554                 }
1555
1556                 if (i == 0) {
1557                         *osfs = *temp;
1558                         /* If the statfs is from mount, it will needs
1559                          * retrieve necessary information from MDT0.
1560                          * i.e. mount does not need the merged osfs
1561                          * from all of MDT.
1562                          * And also clients can be mounted as long as
1563                          * MDT0 is in service*/
1564                         if (flags & OBD_STATFS_FOR_MDT0)
1565                                 GOTO(out_free_temp, rc);
1566                 } else {
1567                         osfs->os_bavail += temp->os_bavail;
1568                         osfs->os_blocks += temp->os_blocks;
1569                         osfs->os_ffree += temp->os_ffree;
1570                         osfs->os_files += temp->os_files;
1571                 }
1572         }
1573
1574         EXIT;
1575 out_free_temp:
1576         OBD_FREE(temp, sizeof(*temp));
1577         return rc;
1578 }
1579
1580 static int lmv_getstatus(struct obd_export *exp,
1581                          struct lu_fid *fid,
1582                          struct obd_capa **pc)
1583 {
1584         struct obd_device    *obd = exp->exp_obd;
1585         struct lmv_obd       *lmv = &obd->u.lmv;
1586         int                   rc;
1587         ENTRY;
1588
1589         rc = lmv_check_connect(obd);
1590         if (rc)
1591                 RETURN(rc);
1592
1593         rc = md_getstatus(lmv->tgts[0]->ltd_exp, fid, pc);
1594         RETURN(rc);
1595 }
1596
1597 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1598                         struct obd_capa *oc, obd_valid valid, const char *name,
1599                         const char *input, int input_size, int output_size,
1600                         int flags, struct ptlrpc_request **request)
1601 {
1602         struct obd_device      *obd = exp->exp_obd;
1603         struct lmv_obd         *lmv = &obd->u.lmv;
1604         struct lmv_tgt_desc    *tgt;
1605         int                     rc;
1606         ENTRY;
1607
1608         rc = lmv_check_connect(obd);
1609         if (rc)
1610                 RETURN(rc);
1611
1612         tgt = lmv_find_target(lmv, fid);
1613         if (IS_ERR(tgt))
1614                 RETURN(PTR_ERR(tgt));
1615
1616         rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1617                          input_size, output_size, flags, request);
1618
1619         RETURN(rc);
1620 }
1621
1622 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1623                         struct obd_capa *oc, obd_valid valid, const char *name,
1624                         const char *input, int input_size, int output_size,
1625                         int flags, __u32 suppgid,
1626                         struct ptlrpc_request **request)
1627 {
1628         struct obd_device      *obd = exp->exp_obd;
1629         struct lmv_obd         *lmv = &obd->u.lmv;
1630         struct lmv_tgt_desc    *tgt;
1631         int                     rc;
1632         ENTRY;
1633
1634         rc = lmv_check_connect(obd);
1635         if (rc)
1636                 RETURN(rc);
1637
1638         tgt = lmv_find_target(lmv, fid);
1639         if (IS_ERR(tgt))
1640                 RETURN(PTR_ERR(tgt));
1641
1642         rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1643                          input_size, output_size, flags, suppgid,
1644                          request);
1645
1646         RETURN(rc);
1647 }
1648
1649 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1650                        struct ptlrpc_request **request)
1651 {
1652         struct obd_device       *obd = exp->exp_obd;
1653         struct lmv_obd          *lmv = &obd->u.lmv;
1654         struct lmv_tgt_desc     *tgt;
1655         int                      rc;
1656         ENTRY;
1657
1658         rc = lmv_check_connect(obd);
1659         if (rc)
1660                 RETURN(rc);
1661
1662         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1663         if (IS_ERR(tgt))
1664                 RETURN(PTR_ERR(tgt));
1665
1666         if (op_data->op_flags & MF_GET_MDT_IDX) {
1667                 op_data->op_mds = tgt->ltd_idx;
1668                 RETURN(0);
1669         }
1670
1671         rc = md_getattr(tgt->ltd_exp, op_data, request);
1672
1673         RETURN(rc);
1674 }
1675
1676 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1677 {
1678         struct obd_device   *obd = exp->exp_obd;
1679         struct lmv_obd      *lmv = &obd->u.lmv;
1680         __u32                i;
1681         int                  rc;
1682         ENTRY;
1683
1684         rc = lmv_check_connect(obd);
1685         if (rc)
1686                 RETURN(rc);
1687
1688         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1689
1690         /*
1691          * With DNE every object can have two locks in different namespaces:
1692          * lookup lock in space of MDT storing direntry and update/open lock in
1693          * space of MDT storing inode.
1694          */
1695         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1696                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
1697                         continue;
1698                 md_null_inode(lmv->tgts[i]->ltd_exp, fid);
1699         }
1700
1701         RETURN(0);
1702 }
1703
1704 static int lmv_find_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1705                            ldlm_iterator_t it, void *data)
1706 {
1707         struct obd_device   *obd = exp->exp_obd;
1708         struct lmv_obd      *lmv = &obd->u.lmv;
1709         __u32                i;
1710         int                  rc;
1711         ENTRY;
1712
1713         rc = lmv_check_connect(obd);
1714         if (rc)
1715                 RETURN(rc);
1716
1717         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1718
1719         /*
1720          * With DNE every object can have two locks in different namespaces:
1721          * lookup lock in space of MDT storing direntry and update/open lock in
1722          * space of MDT storing inode.
1723          */
1724         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1725                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
1726                         continue;
1727                 rc = md_find_cbdata(lmv->tgts[i]->ltd_exp, fid, it, data);
1728                 if (rc)
1729                         RETURN(rc);
1730         }
1731
1732         RETURN(rc);
1733 }
1734
1735
1736 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1737                      struct md_open_data *mod, struct ptlrpc_request **request)
1738 {
1739         struct obd_device     *obd = exp->exp_obd;
1740         struct lmv_obd        *lmv = &obd->u.lmv;
1741         struct lmv_tgt_desc   *tgt;
1742         int                    rc;
1743         ENTRY;
1744
1745         rc = lmv_check_connect(obd);
1746         if (rc)
1747                 RETURN(rc);
1748
1749         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1750         if (IS_ERR(tgt))
1751                 RETURN(PTR_ERR(tgt));
1752
1753         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1754         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1755         RETURN(rc);
1756 }
1757
1758 /**
1759  * Choosing the MDT by name or FID in @op_data.
1760  * For non-striped directory, it will locate MDT by fid.
1761  * For striped-directory, it will locate MDT by name. And also
1762  * it will reset op_fid1 with the FID of the choosen stripe.
1763  **/
1764 struct lmv_tgt_desc *
1765 lmv_locate_target_for_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1766                            const char *name, int namelen, struct lu_fid *fid,
1767                            mdsno_t *mds)
1768 {
1769         struct lmv_tgt_desc     *tgt;
1770         const struct lmv_oinfo  *oinfo;
1771
1772         oinfo = lsm_name_to_stripe_info(lsm, name, namelen);
1773         if (IS_ERR(oinfo))
1774                 RETURN((void *)oinfo);
1775         *fid = oinfo->lmo_fid;
1776         *mds = oinfo->lmo_mds;
1777         tgt = lmv_get_target(lmv, *mds);
1778
1779         CDEBUG(D_INFO, "locate on mds %u "DFID"\n", *mds, PFID(fid));
1780         return tgt;
1781 }
1782
1783 struct lmv_tgt_desc
1784 *lmv_locate_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
1785                 struct lu_fid *fid)
1786 {
1787         struct lmv_stripe_md    *lsm = op_data->op_mea1;
1788         struct lmv_tgt_desc     *tgt;
1789
1790         if (lsm == NULL || lsm->lsm_md_stripe_count <= 1 ||
1791             op_data->op_namelen == 0 ||
1792             lsm->lsm_md_magic == LMV_MAGIC_MIGRATE) {
1793                 tgt = lmv_find_target(lmv, fid);
1794                 if (IS_ERR(tgt))
1795                         return tgt;
1796
1797                 op_data->op_mds = tgt->ltd_idx;
1798                 return tgt;
1799         }
1800
1801         return lmv_locate_target_for_name(lmv, lsm, op_data->op_name,
1802                                           op_data->op_namelen, fid,
1803                                           &op_data->op_mds);
1804 }
1805
1806 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1807                const void *data, int datalen, int mode, __u32 uid,
1808                __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1809                struct ptlrpc_request **request)
1810 {
1811         struct obd_device       *obd = exp->exp_obd;
1812         struct lmv_obd          *lmv = &obd->u.lmv;
1813         struct lmv_tgt_desc     *tgt;
1814         int                      rc;
1815         ENTRY;
1816
1817         rc = lmv_check_connect(obd);
1818         if (rc)
1819                 RETURN(rc);
1820
1821         if (!lmv->desc.ld_active_tgt_count)
1822                 RETURN(-EIO);
1823
1824         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
1825         if (IS_ERR(tgt))
1826                 RETURN(PTR_ERR(tgt));
1827
1828         CDEBUG(D_INODE, "CREATE name '%.*s' on "DFID" -> mds #%x\n",
1829                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1830                op_data->op_mds);
1831
1832         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1833         if (rc)
1834                 RETURN(rc);
1835
1836         /* Send the create request to the MDT where the object
1837          * will be located */
1838         tgt = lmv_find_target(lmv, &op_data->op_fid2);
1839         if (IS_ERR(tgt))
1840                 RETURN(PTR_ERR(tgt));
1841
1842         op_data->op_mds = tgt->ltd_idx;
1843
1844         CDEBUG(D_INODE, "CREATE obj "DFID" -> mds #%x\n",
1845                PFID(&op_data->op_fid2), op_data->op_mds);
1846
1847         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1848         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1849                        cap_effective, rdev, request);
1850         if (rc == 0) {
1851                 if (*request == NULL)
1852                         RETURN(rc);
1853                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1854         }
1855         RETURN(rc);
1856 }
1857
1858 static int lmv_done_writing(struct obd_export *exp,
1859                             struct md_op_data *op_data,
1860                             struct md_open_data *mod)
1861 {
1862         struct obd_device     *obd = exp->exp_obd;
1863         struct lmv_obd        *lmv = &obd->u.lmv;
1864         struct lmv_tgt_desc   *tgt;
1865         int                    rc;
1866         ENTRY;
1867
1868         rc = lmv_check_connect(obd);
1869         if (rc)
1870                 RETURN(rc);
1871
1872         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1873         if (IS_ERR(tgt))
1874                 RETURN(PTR_ERR(tgt));
1875
1876         rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1877         RETURN(rc);
1878 }
1879
1880 static int
1881 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1882                    struct lookup_intent *it, struct md_op_data *op_data,
1883                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1884                    __u64 extra_lock_flags)
1885 {
1886         struct ptlrpc_request      *req = it->d.lustre.it_data;
1887         struct obd_device          *obd = exp->exp_obd;
1888         struct lmv_obd             *lmv = &obd->u.lmv;
1889         struct lustre_handle        plock;
1890         struct lmv_tgt_desc        *tgt;
1891         struct md_op_data          *rdata;
1892         struct lu_fid               fid1;
1893         struct mdt_body            *body;
1894         int                         rc = 0;
1895         int                         pmode;
1896         ENTRY;
1897
1898         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1899         LASSERT(body != NULL);
1900
1901         if (!(body->valid & OBD_MD_MDS))
1902                 RETURN(0);
1903
1904         CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1905                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1906
1907         /*
1908          * We got LOOKUP lock, but we really need attrs.
1909          */
1910         pmode = it->d.lustre.it_lock_mode;
1911         LASSERT(pmode != 0);
1912         memcpy(&plock, lockh, sizeof(plock));
1913         it->d.lustre.it_lock_mode = 0;
1914         it->d.lustre.it_data = NULL;
1915         fid1 = body->fid1;
1916
1917         ptlrpc_req_finished(req);
1918
1919         tgt = lmv_find_target(lmv, &fid1);
1920         if (IS_ERR(tgt))
1921                 GOTO(out, rc = PTR_ERR(tgt));
1922
1923         OBD_ALLOC_PTR(rdata);
1924         if (rdata == NULL)
1925                 GOTO(out, rc = -ENOMEM);
1926
1927         rdata->op_fid1 = fid1;
1928         rdata->op_bias = MDS_CROSS_REF;
1929
1930         rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1931                         lmm, lmmsize, NULL, extra_lock_flags);
1932         OBD_FREE_PTR(rdata);
1933         EXIT;
1934 out:
1935         ldlm_lock_decref(&plock, pmode);
1936         return rc;
1937 }
1938
1939 static int
1940 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1941             struct lookup_intent *it, struct md_op_data *op_data,
1942             struct lustre_handle *lockh, void *lmm, int lmmsize,
1943             struct ptlrpc_request **req, __u64 extra_lock_flags)
1944 {
1945         struct obd_device        *obd = exp->exp_obd;
1946         struct lmv_obd           *lmv = &obd->u.lmv;
1947         struct lmv_tgt_desc      *tgt;
1948         int                       rc;
1949         ENTRY;
1950
1951         rc = lmv_check_connect(obd);
1952         if (rc)
1953                 RETURN(rc);
1954
1955         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1956                LL_IT2STR(it), PFID(&op_data->op_fid1));
1957
1958         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
1959         if (IS_ERR(tgt))
1960                 RETURN(PTR_ERR(tgt));
1961
1962         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1963                LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1964
1965         rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1966                         lmm, lmmsize, req, extra_lock_flags);
1967
1968         if (rc == 0 && it && it->it_op == IT_OPEN) {
1969                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1970                                         lmm, lmmsize, extra_lock_flags);
1971         }
1972         RETURN(rc);
1973 }
1974
1975 static int
1976 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1977                  struct ptlrpc_request **preq)
1978 {
1979         struct ptlrpc_request   *req = NULL;
1980         struct obd_device       *obd = exp->exp_obd;
1981         struct lmv_obd          *lmv = &obd->u.lmv;
1982         struct lmv_tgt_desc     *tgt;
1983         struct mdt_body         *body;
1984         int                      rc;
1985         ENTRY;
1986
1987         rc = lmv_check_connect(obd);
1988         if (rc)
1989                 RETURN(rc);
1990
1991         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
1992         if (IS_ERR(tgt))
1993                 RETURN(PTR_ERR(tgt));
1994
1995         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1996                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1997                tgt->ltd_idx);
1998
1999         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
2000         if (rc != 0)
2001                 RETURN(rc);
2002
2003         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
2004         LASSERT(body != NULL);
2005
2006         if (body->valid & OBD_MD_MDS) {
2007                 struct lu_fid rid = body->fid1;
2008                 CDEBUG(D_INODE, "Request attrs for "DFID"\n",
2009                        PFID(&rid));
2010
2011                 tgt = lmv_find_target(lmv, &rid);
2012                 if (IS_ERR(tgt)) {
2013                         ptlrpc_req_finished(*preq);
2014                         preq = NULL;
2015                         RETURN(PTR_ERR(tgt));
2016                 }
2017
2018                 op_data->op_fid1 = rid;
2019                 op_data->op_valid |= OBD_MD_FLCROSSREF;
2020                 op_data->op_namelen = 0;
2021                 op_data->op_name = NULL;
2022                 rc = md_getattr_name(tgt->ltd_exp, op_data, &req);
2023                 ptlrpc_req_finished(*preq);
2024                 *preq = req;
2025         }
2026
2027         RETURN(rc);
2028 }
2029
2030 #define md_op_data_fid(op_data, fl)                     \
2031         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2032          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2033          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2034          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2035          NULL)
2036
2037 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2038                             struct md_op_data *op_data,
2039                             int op_tgt, ldlm_mode_t mode, int bits, int flag)
2040 {
2041         struct lu_fid          *fid = md_op_data_fid(op_data, flag);
2042         struct obd_device      *obd = exp->exp_obd;
2043         struct lmv_obd         *lmv = &obd->u.lmv;
2044         ldlm_policy_data_t      policy = {{ 0 }};
2045         int                     rc = 0;
2046         ENTRY;
2047
2048         if (!fid_is_sane(fid))
2049                 RETURN(0);
2050
2051         if (tgt == NULL) {
2052                 tgt = lmv_find_target(lmv, fid);
2053                 if (IS_ERR(tgt))
2054                         RETURN(PTR_ERR(tgt));
2055         }
2056
2057         if (tgt->ltd_idx != op_tgt) {
2058                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2059                 policy.l_inodebits.bits = bits;
2060                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2061                                       mode, LCF_ASYNC, NULL);
2062         } else {
2063                 CDEBUG(D_INODE,
2064                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2065                        op_tgt, PFID(fid));
2066                 op_data->op_flags |= flag;
2067                 rc = 0;
2068         }
2069
2070         RETURN(rc);
2071 }
2072
2073 /*
2074  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2075  * op_data->op_fid2
2076  */
2077 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2078                     struct ptlrpc_request **request)
2079 {
2080         struct obd_device       *obd = exp->exp_obd;
2081         struct lmv_obd          *lmv = &obd->u.lmv;
2082         struct lmv_tgt_desc     *tgt;
2083         int                      rc;
2084         ENTRY;
2085
2086         rc = lmv_check_connect(obd);
2087         if (rc)
2088                 RETURN(rc);
2089
2090         LASSERT(op_data->op_namelen != 0);
2091
2092         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2093                PFID(&op_data->op_fid2), op_data->op_namelen,
2094                op_data->op_name, PFID(&op_data->op_fid1));
2095
2096         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2097         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2098         op_data->op_cap = cfs_curproc_cap_pack();
2099         if (op_data->op_mea2 != NULL) {
2100                 struct lmv_stripe_md    *lsm = op_data->op_mea2;
2101                 const struct lmv_oinfo  *oinfo;
2102
2103                 oinfo = lsm_name_to_stripe_info(lsm, op_data->op_name,
2104                                                 op_data->op_namelen);
2105                 if (IS_ERR(oinfo))
2106                         RETURN(PTR_ERR(oinfo));
2107
2108                 op_data->op_fid2 = oinfo->lmo_fid;
2109         }
2110
2111         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid2);
2112         if (IS_ERR(tgt))
2113                 RETURN(PTR_ERR(tgt));
2114
2115         /*
2116          * Cancel UPDATE lock on child (fid1).
2117          */
2118         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2119         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_idx, LCK_EX,
2120                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2121         if (rc != 0)
2122                 RETURN(rc);
2123
2124         rc = md_link(tgt->ltd_exp, op_data, request);
2125
2126         RETURN(rc);
2127 }
2128
2129 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2130                       const char *old, int oldlen, const char *new, int newlen,
2131                       struct ptlrpc_request **request)
2132 {
2133         struct obd_device       *obd = exp->exp_obd;
2134         struct lmv_obd          *lmv = &obd->u.lmv;
2135         struct lmv_tgt_desc     *src_tgt;
2136         int                     rc;
2137         ENTRY;
2138
2139         LASSERT(oldlen != 0);
2140
2141         CDEBUG(D_INODE, "RENAME %.*s in "DFID":%d to %.*s in "DFID":%d\n",
2142                oldlen, old, PFID(&op_data->op_fid1),
2143                op_data->op_mea1 ? op_data->op_mea1->lsm_md_stripe_count : 0,
2144                newlen, new, PFID(&op_data->op_fid2),
2145                op_data->op_mea2 ? op_data->op_mea2->lsm_md_stripe_count : 0);
2146
2147         rc = lmv_check_connect(obd);
2148         if (rc)
2149                 RETURN(rc);
2150
2151         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2152         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2153         op_data->op_cap = cfs_curproc_cap_pack();
2154         if (op_data->op_cli_flags & CLI_MIGRATE) {
2155                 LASSERTF(fid_is_sane(&op_data->op_fid3), "invalid FID "DFID"\n",
2156                          PFID(&op_data->op_fid3));
2157                 rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
2158                 if (rc)
2159                         RETURN(rc);
2160                 src_tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid3);
2161         } else {
2162                 if (op_data->op_mea1 != NULL) {
2163                         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2164
2165                         src_tgt = lmv_locate_target_for_name(lmv, lsm, old,
2166                                                              oldlen,
2167                                                              &op_data->op_fid1,
2168                                                              &op_data->op_mds);
2169                         if (IS_ERR(src_tgt))
2170                                 RETURN(PTR_ERR(src_tgt));
2171                 } else {
2172                         src_tgt = lmv_find_target(lmv, &op_data->op_fid1);
2173                         if (IS_ERR(src_tgt))
2174                                 RETURN(PTR_ERR(src_tgt));
2175
2176                         op_data->op_mds = src_tgt->ltd_idx;
2177                 }
2178
2179                 if (op_data->op_mea2) {
2180                         struct lmv_stripe_md    *lsm = op_data->op_mea2;
2181                         const struct lmv_oinfo  *oinfo;
2182
2183                         oinfo = lsm_name_to_stripe_info(lsm, new, newlen);
2184                         if (IS_ERR(oinfo))
2185                                 RETURN(PTR_ERR(oinfo));
2186
2187                         op_data->op_fid2 = oinfo->lmo_fid;
2188                 }
2189         }
2190         if (IS_ERR(src_tgt))
2191                 RETURN(PTR_ERR(src_tgt));
2192
2193         /*
2194          * LOOKUP lock on src child (fid3) should also be cancelled for
2195          * src_tgt in mdc_rename.
2196          */
2197         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2198
2199         /*
2200          * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2201          * own target.
2202          */
2203         rc = lmv_early_cancel(exp, NULL, op_data, src_tgt->ltd_idx,
2204                               LCK_EX, MDS_INODELOCK_UPDATE,
2205                               MF_MDC_CANCEL_FID2);
2206
2207         if (rc != 0)
2208                 RETURN(rc);
2209         /*
2210          * Cancel LOOKUP locks on source child (fid3) for parent tgt_tgt.
2211          */
2212         if (fid_is_sane(&op_data->op_fid3)) {
2213                 struct lmv_tgt_desc *tgt;
2214
2215                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2216                 if (IS_ERR(tgt))
2217                         RETURN(PTR_ERR(tgt));
2218
2219                 /* Cancel LOOKUP lock on its parent */
2220                 rc = lmv_early_cancel(exp, tgt, op_data, src_tgt->ltd_idx,
2221                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2222                                       MF_MDC_CANCEL_FID3);
2223                 if (rc != 0)
2224                         RETURN(rc);
2225
2226                 rc = lmv_early_cancel(exp, NULL, op_data, src_tgt->ltd_idx,
2227                                       LCK_EX, MDS_INODELOCK_FULL,
2228                                       MF_MDC_CANCEL_FID3);
2229                 if (rc != 0)
2230                         RETURN(rc);
2231         }
2232
2233         /*
2234          * Cancel all the locks on tgt child (fid4).
2235          */
2236         if (fid_is_sane(&op_data->op_fid4))
2237                 rc = lmv_early_cancel(exp, NULL, op_data, src_tgt->ltd_idx,
2238                                       LCK_EX, MDS_INODELOCK_FULL,
2239                                       MF_MDC_CANCEL_FID4);
2240
2241         CDEBUG(D_INODE, DFID":m%d to "DFID"\n", PFID(&op_data->op_fid1),
2242                op_data->op_mds, PFID(&op_data->op_fid2));
2243
2244         rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2245                        request);
2246
2247         RETURN(rc);
2248 }
2249
2250 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2251                        void *ea, int ealen, void *ea2, int ea2len,
2252                        struct ptlrpc_request **request,
2253                        struct md_open_data **mod)
2254 {
2255         struct obd_device       *obd = exp->exp_obd;
2256         struct lmv_obd          *lmv = &obd->u.lmv;
2257         struct lmv_tgt_desc     *tgt;
2258         int                      rc = 0;
2259         ENTRY;
2260
2261         rc = lmv_check_connect(obd);
2262         if (rc)
2263                 RETURN(rc);
2264
2265         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x\n",
2266                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid);
2267
2268         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2269         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2270         if (IS_ERR(tgt))
2271                 RETURN(PTR_ERR(tgt));
2272
2273         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2274                         ea2len, request, mod);
2275
2276         RETURN(rc);
2277 }
2278
2279 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2280                      struct obd_capa *oc, struct ptlrpc_request **request)
2281 {
2282         struct obd_device       *obd = exp->exp_obd;
2283         struct lmv_obd          *lmv = &obd->u.lmv;
2284         struct lmv_tgt_desc     *tgt;
2285         int                      rc;
2286         ENTRY;
2287
2288         rc = lmv_check_connect(obd);
2289         if (rc != 0)
2290                 RETURN(rc);
2291
2292         tgt = lmv_find_target(lmv, fid);
2293         if (IS_ERR(tgt))
2294                 RETURN(PTR_ERR(tgt));
2295
2296         rc = md_fsync(tgt->ltd_exp, fid, oc, request);
2297         RETURN(rc);
2298 }
2299
2300 /*
2301  * Adjust a set of pages, each page containing an array of lu_dirpages,
2302  * so that each page can be used as a single logical lu_dirpage.
2303  *
2304  * A lu_dirpage is laid out as follows, where s = ldp_hash_start,
2305  * e = ldp_hash_end, f = ldp_flags, p = padding, and each "ent" is a
2306  * struct lu_dirent.  It has size up to LU_PAGE_SIZE. The ldp_hash_end
2307  * value is used as a cookie to request the next lu_dirpage in a
2308  * directory listing that spans multiple pages (two in this example):
2309  *   ________
2310  *  |        |
2311  * .|--------v-------   -----.
2312  * |s|e|f|p|ent|ent| ... |ent|
2313  * '--|--------------   -----'   Each CFS_PAGE contains a single
2314  *    '------.                   lu_dirpage.
2315  * .---------v-------   -----.
2316  * |s|e|f|p|ent| 0 | ... | 0 |
2317  * '-----------------   -----'
2318  *
2319  * However, on hosts where the native VM page size (PAGE_CACHE_SIZE) is
2320  * larger than LU_PAGE_SIZE, a single host page may contain multiple
2321  * lu_dirpages. After reading the lu_dirpages from the MDS, the
2322  * ldp_hash_end of the first lu_dirpage refers to the one immediately
2323  * after it in the same CFS_PAGE (arrows simplified for brevity, but
2324  * in general e0==s1, e1==s2, etc.):
2325  *
2326  * .--------------------   -----.
2327  * |s0|e0|f0|p|ent|ent| ... |ent|
2328  * |---v----------------   -----|
2329  * |s1|e1|f1|p|ent|ent| ... |ent|
2330  * |---v----------------   -----|  Here, each CFS_PAGE contains
2331  *             ...                 multiple lu_dirpages.
2332  * |---v----------------   -----|
2333  * |s'|e'|f'|p|ent|ent| ... |ent|
2334  * '---|----------------   -----'
2335  *     v
2336  * .----------------------------.
2337  * |        next CFS_PAGE       |
2338  *
2339  * This structure is transformed into a single logical lu_dirpage as follows:
2340  *
2341  * - Replace e0 with e' so the request for the next lu_dirpage gets the page
2342  *   labeled 'next CFS_PAGE'.
2343  *
2344  * - Copy the LDF_COLLIDE flag from f' to f0 to correctly reflect whether
2345  *   a hash collision with the next page exists.
2346  *
2347  * - Adjust the lde_reclen of the ending entry of each lu_dirpage to span
2348  *   to the first entry of the next lu_dirpage.
2349  */
2350 #if PAGE_CACHE_SIZE > LU_PAGE_SIZE
2351 static void lmv_adjust_dirpages(struct page **pages, int ncfspgs, int nlupgs)
2352 {
2353         int i;
2354
2355         for (i = 0; i < ncfspgs; i++) {
2356                 struct lu_dirpage       *dp = kmap(pages[i]);
2357                 struct lu_dirpage       *first = dp;
2358                 struct lu_dirent        *end_dirent = NULL;
2359                 struct lu_dirent        *ent;
2360                 __u64                   hash_end = dp->ldp_hash_end;
2361                 __u32                   flags = dp->ldp_flags;
2362
2363                 while (--nlupgs > 0) {
2364                         ent = lu_dirent_start(dp);
2365                         for (end_dirent = ent; ent != NULL;
2366                              end_dirent = ent, ent = lu_dirent_next(ent));
2367
2368                         /* Advance dp to next lu_dirpage. */
2369                         dp = (struct lu_dirpage *)((char *)dp + LU_PAGE_SIZE);
2370
2371                         /* Check if we've reached the end of the CFS_PAGE. */
2372                         if (!((unsigned long)dp & ~CFS_PAGE_MASK))
2373                                 break;
2374
2375                         /* Save the hash and flags of this lu_dirpage. */
2376                         hash_end = dp->ldp_hash_end;
2377                         flags = dp->ldp_flags;
2378
2379                         /* Check if lu_dirpage contains no entries. */
2380                         if (!end_dirent)
2381                                 break;
2382
2383                         /* Enlarge the end entry lde_reclen from 0 to
2384                          * first entry of next lu_dirpage. */
2385                         LASSERT(le16_to_cpu(end_dirent->lde_reclen) == 0);
2386                         end_dirent->lde_reclen =
2387                                 cpu_to_le16((char *)(dp->ldp_entries) -
2388                                             (char *)end_dirent);
2389                 }
2390
2391                 first->ldp_hash_end = hash_end;
2392                 first->ldp_flags &= ~cpu_to_le32(LDF_COLLIDE);
2393                 first->ldp_flags |= flags & cpu_to_le32(LDF_COLLIDE);
2394
2395                 kunmap(pages[i]);
2396         }
2397         LASSERTF(nlupgs == 0, "left = %d", nlupgs);
2398 }
2399 #else
2400 #define lmv_adjust_dirpages(pages, ncfspgs, nlupgs) do {} while (0)
2401 #endif  /* PAGE_CACHE_SIZE > LU_PAGE_SIZE */
2402
2403 #define NORMAL_MAX_STRIPES 4
2404 int lmv_read_entry(struct obd_export *exp, struct md_op_data *op_data,
2405                    struct md_callback *cb_op, struct lu_dirent **ldp,
2406                    struct page **ppage)
2407 {
2408         struct obd_device       *obd = exp->exp_obd;
2409         struct lmv_obd          *lmv = &obd->u.lmv;
2410         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2411         struct lu_dirent        *tmp_ents[NORMAL_MAX_STRIPES];
2412         struct lu_dirent        **ents = NULL;
2413         int                     stripe_count;
2414         __u64                   min_hash;
2415         int                     min_idx = 0;
2416         struct page             *min_page = NULL;
2417         int                     i;
2418         int                     rc;
2419         ENTRY;
2420
2421         rc = lmv_check_connect(obd);
2422         if (rc)
2423                 RETURN(rc);
2424
2425         if (lsm == NULL)
2426                 stripe_count = 1;
2427         else
2428                 stripe_count = lsm->lsm_md_stripe_count;
2429
2430         if (stripe_count > NORMAL_MAX_STRIPES) {
2431                 OBD_ALLOC(ents, sizeof(ents[0]) * stripe_count);
2432                 if (ents == NULL)
2433                         GOTO(out, rc = -ENOMEM);
2434         } else {
2435                 ents = tmp_ents;
2436                 memset(ents, 0, sizeof(ents[0]) * stripe_count);
2437         }
2438
2439         min_hash = MDS_DIR_END_OFF;
2440         for (i = 0; i < stripe_count; i++) {
2441                 struct lmv_tgt_desc *tgt;
2442                 struct page *page = NULL;
2443
2444                 if (likely(lsm == NULL)) {
2445                         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2446                         if (IS_ERR(tgt))
2447                                 GOTO(out, rc = PTR_ERR(tgt));
2448                         LASSERT(op_data->op_data != NULL);
2449                 } else {
2450                         tgt = lmv_get_target(lmv, lsm->lsm_md_oinfo[i].lmo_mds);
2451                         if (IS_ERR(tgt))
2452                                 GOTO(out, rc = PTR_ERR(tgt));
2453                         op_data->op_fid1 = lsm->lsm_md_oinfo[i].lmo_fid;
2454                         op_data->op_fid2 = lsm->lsm_md_oinfo[i].lmo_fid;
2455                         op_data->op_stripe_offset = i;
2456                 }
2457
2458                 rc = md_read_entry(tgt->ltd_exp, op_data, cb_op, &ents[i],
2459                                    &page);
2460                 if (rc != 0)
2461                         GOTO(out, rc);
2462
2463                 if (ents[i] != NULL &&
2464                     le64_to_cpu(ents[i]->lde_hash) <= min_hash) {
2465                         if (min_page != NULL)
2466                                 page_cache_release(min_page);
2467                         min_page = page;
2468                         min_hash = le64_to_cpu(ents[i]->lde_hash);
2469                         min_idx = i;
2470                 }
2471         }
2472
2473         if (min_hash != MDS_DIR_END_OFF)
2474                 *ldp = ents[min_idx];
2475         else
2476                 *ldp = NULL;
2477 out:
2478         if (stripe_count > NORMAL_MAX_STRIPES && ents != NULL)
2479                 OBD_FREE(ents, sizeof(ents[0]) * stripe_count);
2480
2481         if (rc != 0 && min_page != NULL) {
2482                 kunmap(min_page);
2483                 page_cache_release(min_page);
2484         } else {
2485                 *ppage = min_page;
2486         }
2487
2488         RETURN(rc);
2489 }
2490
2491 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2492                       struct ptlrpc_request **request)
2493 {
2494         struct obd_device       *obd = exp->exp_obd;
2495         struct lmv_obd          *lmv = &obd->u.lmv;
2496         struct lmv_tgt_desc     *tgt = NULL;
2497         struct lmv_tgt_desc     *parent_tgt = NULL;
2498         struct mdt_body         *body;
2499         int                     rc;
2500         ENTRY;
2501
2502         rc = lmv_check_connect(obd);
2503         if (rc)
2504                 RETURN(rc);
2505 retry:
2506         /* Send unlink requests to the MDT where the child is located */
2507         if (likely(!fid_is_zero(&op_data->op_fid2))) {
2508                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2509                 if (IS_ERR(tgt))
2510                         RETURN(PTR_ERR(tgt));
2511
2512                 /* For striped dir, we need to locate the parent as well */
2513                 if (op_data->op_mea1 != NULL &&
2514                     op_data->op_mea1->lsm_md_stripe_count > 1) {
2515                         struct lmv_tgt_desc *tmp;
2516
2517                         LASSERT(op_data->op_name != NULL &&
2518                                 op_data->op_namelen != 0);
2519                         tmp = lmv_locate_target_for_name(lmv,
2520                                                    op_data->op_mea1,
2521                                                    op_data->op_name,
2522                                                    op_data->op_namelen,
2523                                                    &op_data->op_fid1,
2524                                                    &op_data->op_mds);
2525                         if (IS_ERR(tmp))
2526                                 RETURN(PTR_ERR(tmp));
2527                 }
2528         } else {
2529                 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
2530                 if (IS_ERR(tgt))
2531                         RETURN(PTR_ERR(tgt));
2532         }
2533
2534         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2535         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2536         op_data->op_cap = cfs_curproc_cap_pack();
2537
2538         /*
2539          * If child's fid is given, cancel unused locks for it if it is from
2540          * another export than parent.
2541          *
2542          * LOOKUP lock for child (fid3) should also be cancelled on parent
2543          * tgt_tgt in mdc_unlink().
2544          */
2545         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2546
2547         /*
2548          * Cancel FULL locks on child (fid3).
2549          */
2550         parent_tgt = lmv_find_target(lmv, &op_data->op_fid1);
2551         if (IS_ERR(parent_tgt))
2552                 RETURN(PTR_ERR(parent_tgt));
2553
2554         if (parent_tgt != tgt) {
2555                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_idx,
2556                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2557                                       MF_MDC_CANCEL_FID3);
2558         }
2559
2560         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_idx, LCK_EX,
2561                               MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2562         if (rc != 0)
2563                 RETURN(rc);
2564
2565         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%d\n",
2566                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2), tgt->ltd_idx);
2567
2568         rc = md_unlink(tgt->ltd_exp, op_data, request);
2569         if (rc != 0 && rc != -EREMOTE)
2570                 RETURN(rc);
2571
2572         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2573         if (body == NULL)
2574                 RETURN(-EPROTO);
2575
2576         /* Not cross-ref case, just get out of here. */
2577         if (likely(!(body->valid & OBD_MD_MDS)))
2578                 RETURN(0);
2579
2580         CDEBUG(D_INODE, "%s: try unlink to another MDT for "DFID"\n",
2581                exp->exp_obd->obd_name, PFID(&body->fid1));
2582
2583         /* This is a remote object, try remote MDT, Note: it may
2584          * try more than 1 time here, Considering following case
2585          * /mnt/lustre is root on MDT0, remote1 is on MDT1
2586          * 1. Initially A does not know where remote1 is, it send
2587          *    unlink RPC to MDT0, MDT0 return -EREMOTE, it will
2588          *    resend unlink RPC to MDT1 (retry 1st time).
2589          *
2590          * 2. During the unlink RPC in flight,
2591          *    client B mv /mnt/lustre/remote1 /mnt/lustre/remote2
2592          *    and create new remote1, but on MDT0
2593          *
2594          * 3. MDT1 get unlink RPC(from A), then do remote lock on
2595          *    /mnt/lustre, then lookup get fid of remote1, and find
2596          *    it is remote dir again, and replay -EREMOTE again.
2597          *
2598          * 4. Then A will resend unlink RPC to MDT0. (retry 2nd times).
2599          *
2600          * In theory, it might try unlimited time here, but it should
2601          * be very rare case.  */
2602         op_data->op_fid2 = body->fid1;
2603         ptlrpc_req_finished(*request);
2604         *request = NULL;
2605
2606         goto retry;
2607 }
2608
2609 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2610 {
2611         struct lmv_obd *lmv = &obd->u.lmv;
2612         int rc = 0;
2613
2614         switch (stage) {
2615         case OBD_CLEANUP_EARLY:
2616                 /* XXX: here should be calling obd_precleanup() down to
2617                  * stack. */
2618                 break;
2619         case OBD_CLEANUP_EXPORTS:
2620                 fld_client_proc_fini(&lmv->lmv_fld);
2621                 lprocfs_obd_cleanup(obd);
2622                 lprocfs_free_md_stats(obd);
2623                 break;
2624         default:
2625                 break;
2626         }
2627         RETURN(rc);
2628 }
2629
2630 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2631                         __u32 keylen, void *key, __u32 *vallen, void *val,
2632                         struct lov_stripe_md *lsm)
2633 {
2634         struct obd_device       *obd;
2635         struct lmv_obd          *lmv;
2636         int                      rc = 0;
2637         ENTRY;
2638
2639         obd = class_exp2obd(exp);
2640         if (obd == NULL) {
2641                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2642                        exp->exp_handle.h_cookie);
2643                 RETURN(-EINVAL);
2644         }
2645
2646         lmv = &obd->u.lmv;
2647         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2648                 int i;
2649
2650                 rc = lmv_check_connect(obd);
2651                 if (rc)
2652                         RETURN(rc);
2653
2654                 LASSERT(*vallen == sizeof(__u32));
2655                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2656                         struct lmv_tgt_desc *tgt = lmv->tgts[i];
2657                         /*
2658                          * All tgts should be connected when this gets called.
2659                          */
2660                         if (tgt == NULL || tgt->ltd_exp == NULL)
2661                                 continue;
2662
2663                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2664                                           vallen, val, NULL))
2665                                 RETURN(0);
2666                 }
2667                 RETURN(-EINVAL);
2668         } else if (KEY_IS(KEY_MAX_EASIZE) ||
2669                    KEY_IS(KEY_DEFAULT_EASIZE) ||
2670                    KEY_IS(KEY_MAX_COOKIESIZE) ||
2671                    KEY_IS(KEY_DEFAULT_COOKIESIZE) ||
2672                    KEY_IS(KEY_CONN_DATA)) {
2673                 rc = lmv_check_connect(obd);
2674                 if (rc)
2675                         RETURN(rc);
2676
2677                 /*
2678                  * Forwarding this request to first MDS, it should know LOV
2679                  * desc.
2680                  */
2681                 rc = obd_get_info(env, lmv->tgts[0]->ltd_exp, keylen, key,
2682                                   vallen, val, NULL);
2683                 if (!rc && KEY_IS(KEY_CONN_DATA))
2684                         exp->exp_connect_data = *(struct obd_connect_data *)val;
2685                 RETURN(rc);
2686         } else if (KEY_IS(KEY_TGT_COUNT)) {
2687                 *((int *)val) = lmv->desc.ld_tgt_count;
2688                 RETURN(0);
2689         }
2690
2691         CDEBUG(D_IOCTL, "Invalid key\n");
2692         RETURN(-EINVAL);
2693 }
2694
2695 int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
2696                        obd_count keylen, void *key, obd_count vallen,
2697                        void *val, struct ptlrpc_request_set *set)
2698 {
2699         struct lmv_tgt_desc    *tgt = NULL;
2700         struct obd_device      *obd;
2701         struct lmv_obd         *lmv;
2702         int rc = 0;
2703         ENTRY;
2704
2705         obd = class_exp2obd(exp);
2706         if (obd == NULL) {
2707                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2708                        exp->exp_handle.h_cookie);
2709                 RETURN(-EINVAL);
2710         }
2711         lmv = &obd->u.lmv;
2712
2713         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX)) {
2714                 int i, err = 0;
2715
2716                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2717                         tgt = lmv->tgts[i];
2718
2719                         if (tgt == NULL || tgt->ltd_exp == NULL)
2720                                 continue;
2721
2722                         err = obd_set_info_async(env, tgt->ltd_exp,
2723                                                  keylen, key, vallen, val, set);
2724                         if (err && rc == 0)
2725                                 rc = err;
2726                 }
2727
2728                 RETURN(rc);
2729         }
2730
2731         RETURN(-EINVAL);
2732 }
2733
2734 static int lmv_pack_md_v1(const struct lmv_stripe_md *lsm,
2735                           struct lmv_mds_md_v1 *lmm1)
2736 {
2737         int     cplen;
2738         int     i;
2739
2740         lmm1->lmv_magic = cpu_to_le32(lsm->lsm_md_magic);
2741         lmm1->lmv_stripe_count = cpu_to_le32(lsm->lsm_md_stripe_count);
2742         lmm1->lmv_master_mdt_index = cpu_to_le32(lsm->lsm_md_master_mdt_index);
2743         lmm1->lmv_hash_type = cpu_to_le32(lsm->lsm_md_hash_type);
2744         cplen = strlcpy(lmm1->lmv_pool_name, lsm->lsm_md_pool_name,
2745                         sizeof(lmm1->lmv_pool_name));
2746         if (cplen >= sizeof(lmm1->lmv_pool_name))
2747                 return -E2BIG;
2748
2749         for (i = 0; i < lsm->lsm_md_stripe_count; i++)
2750                 fid_cpu_to_le(&lmm1->lmv_stripe_fids[i],
2751                               &lsm->lsm_md_oinfo[i].lmo_fid);
2752         return 0;
2753 }
2754
2755 int lmv_pack_md(union lmv_mds_md **lmmp, const struct lmv_stripe_md *lsm,
2756                 int stripe_count)
2757 {
2758         int     lmm_size = 0;
2759         bool    allocated = false;
2760         int     rc = 0;
2761         ENTRY;
2762
2763         LASSERT(lmmp != NULL);
2764         /* Free lmm */
2765         if (*lmmp != NULL && lsm == NULL) {
2766                 int stripe_count;
2767
2768                 stripe_count = lmv_mds_md_stripe_count_get(*lmmp);
2769                 lmm_size = lmv_mds_md_size(stripe_count,
2770                                            le32_to_cpu((*lmmp)->lmv_magic));
2771                 if (lmm_size == 0)
2772                         RETURN(-EINVAL);
2773                 OBD_FREE(*lmmp, lmm_size);
2774                 *lmmp = NULL;
2775                 RETURN(0);
2776         }
2777
2778         /* Alloc lmm */
2779         if (*lmmp == NULL && lsm == NULL) {
2780                 lmm_size = lmv_mds_md_size(stripe_count, LMV_MAGIC);
2781                 LASSERT(lmm_size > 0);
2782                 OBD_ALLOC(*lmmp, lmm_size);
2783                 if (*lmmp == NULL)
2784                         RETURN(-ENOMEM);
2785                 lmv_mds_md_stripe_count_set(*lmmp, stripe_count);
2786                 (*lmmp)->lmv_magic = cpu_to_le32(LMV_MAGIC);
2787                 RETURN(lmm_size);
2788         }
2789
2790         /* pack lmm */
2791         LASSERT(lsm != NULL);
2792         lmm_size = lmv_mds_md_size(lsm->lsm_md_stripe_count, lsm->lsm_md_magic);
2793         if (*lmmp == NULL) {
2794                 OBD_ALLOC(*lmmp, lmm_size);
2795                 if (*lmmp == NULL)
2796                         RETURN(-ENOMEM);
2797                 allocated = true;
2798         }
2799
2800         switch (lsm->lsm_md_magic) {
2801         case LMV_MAGIC_V1:
2802                 rc = lmv_pack_md_v1(lsm, &(*lmmp)->lmv_md_v1);
2803                 break;
2804         default:
2805                 rc = -EINVAL;
2806                 break;
2807         }
2808
2809         if (rc != 0 && allocated) {
2810                 OBD_FREE(*lmmp, lmm_size);
2811                 *lmmp = NULL;
2812         }
2813
2814         RETURN(lmm_size);
2815 }
2816 EXPORT_SYMBOL(lmv_pack_md);
2817
2818 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
2819                             const struct lmv_mds_md_v1 *lmm1)
2820 {
2821         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
2822         int             stripe_count;
2823         int             cplen;
2824         int             i;
2825         int             rc = 0;
2826         ENTRY;
2827
2828         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
2829         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
2830         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
2831         lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
2832         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
2833         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
2834                         sizeof(lsm->lsm_md_pool_name));
2835
2836         if (cplen >= sizeof(lsm->lsm_md_pool_name))
2837                 RETURN(-E2BIG);
2838
2839         CDEBUG(D_INFO, "unpack lsm count %d, master %d hash_type %d"
2840                "layout_version %d\n", lsm->lsm_md_stripe_count,
2841                lsm->lsm_md_master_mdt_index, lsm->lsm_md_hash_type,
2842                lsm->lsm_md_layout_version);
2843
2844         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
2845         for (i = 0; i < le32_to_cpu(stripe_count); i++) {
2846                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
2847                               &lmm1->lmv_stripe_fids[i]);
2848                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
2849                                     &lsm->lsm_md_oinfo[i].lmo_mds);
2850                 if (rc != 0)
2851                         RETURN(rc);
2852                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
2853                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
2854         }
2855
2856         RETURN(rc);
2857 }
2858
2859 int lmv_unpack_md(struct obd_export *exp, struct lmv_stripe_md **lsmp,
2860                   const union lmv_mds_md *lmm, int stripe_count)
2861 {
2862         struct lmv_stripe_md     *lsm;
2863         int                      lsm_size;
2864         int                      rc;
2865         bool                     allocated = false;
2866         ENTRY;
2867
2868         LASSERT(lsmp != NULL);
2869
2870         lsm = *lsmp;
2871         /* Free memmd */
2872         if (lsm != NULL && lmm == NULL) {
2873 #ifdef __KERNEL__
2874                 int i;
2875                 for (i = 1; i < lsm->lsm_md_stripe_count; i++) {
2876                         if (lsm->lsm_md_oinfo[i].lmo_root != NULL)
2877                                 iput(lsm->lsm_md_oinfo[i].lmo_root);
2878                 }
2879 #endif
2880                 lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
2881                 OBD_FREE(lsm, lsm_size);
2882                 *lsmp = NULL;
2883                 RETURN(0);
2884         }
2885
2886         /* Alloc memmd */
2887         if (lsm == NULL && lmm == NULL) {
2888                 lsm_size = lmv_stripe_md_size(stripe_count);
2889                 OBD_ALLOC(lsm, lsm_size);
2890                 if (lsm == NULL)
2891                         RETURN(-ENOMEM);
2892                 lsm->lsm_md_stripe_count = stripe_count;
2893                 *lsmp = lsm;
2894                 RETURN(0);
2895         }
2896
2897         /* Unpack memmd */
2898         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
2899             le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_MIGRATE &&
2900             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
2901                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
2902                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
2903                        -EIO);
2904                 RETURN(-EIO);
2905         }
2906
2907         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1 ||
2908             le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_MIGRATE)
2909                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
2910         else
2911                 /**
2912                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
2913                  * stripecount should be 0 then.
2914                  */
2915                 lsm_size = lmv_stripe_md_size(0);
2916
2917         lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
2918         if (lsm == NULL) {
2919                 OBD_ALLOC(lsm, lsm_size);
2920                 if (lsm == NULL)
2921                         RETURN(-ENOMEM);
2922                 allocated = true;
2923                 *lsmp = lsm;
2924         }
2925
2926         switch (le32_to_cpu(lmm->lmv_magic)) {
2927         case LMV_MAGIC_V1:
2928         case LMV_MAGIC_MIGRATE:
2929                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
2930                 break;
2931         default:
2932                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
2933                        le32_to_cpu(lmm->lmv_magic));
2934                 rc = -EINVAL;
2935                 break;
2936         }
2937
2938         if (rc != 0 && allocated) {
2939                 OBD_FREE(lsm, lsm_size);
2940                 *lsmp = NULL;
2941                 lsm_size = rc;
2942         }
2943         RETURN(lsm_size);
2944 }
2945
2946 int lmv_alloc_memmd(struct lmv_stripe_md **lsmp, int stripes)
2947 {
2948         return lmv_unpack_md(NULL, lsmp, NULL, stripes);
2949 }
2950 EXPORT_SYMBOL(lmv_alloc_memmd);
2951
2952 void lmv_free_memmd(struct lmv_stripe_md *lsm)
2953 {
2954         lmv_unpack_md(NULL, &lsm, NULL, 0);
2955 }
2956 EXPORT_SYMBOL(lmv_free_memmd);
2957
2958 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2959                  struct lov_mds_md *lmm, int disk_len)
2960 {
2961         return lmv_unpack_md(exp, (struct lmv_stripe_md **)lsmp,
2962                              (union lmv_mds_md *)lmm, disk_len);
2963 }
2964
2965 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2966                struct lov_stripe_md *lsm)
2967 {
2968         struct obd_device               *obd = exp->exp_obd;
2969         struct lmv_obd                  *lmv_obd = &obd->u.lmv;
2970         const struct lmv_stripe_md      *lmv = (struct lmv_stripe_md *)lsm;
2971         int                             stripe_count;
2972
2973         if (lmmp == NULL) {
2974                 if (lsm != NULL)
2975                         stripe_count = lmv->lsm_md_stripe_count;
2976                 else
2977                         stripe_count = lmv_obd->desc.ld_tgt_count;
2978
2979                 return lmv_mds_md_size(stripe_count, LMV_MAGIC_V1);
2980         }
2981
2982         return lmv_pack_md((union lmv_mds_md **)lmmp, lmv, 0);
2983 }
2984
2985 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2986                              ldlm_policy_data_t *policy, ldlm_mode_t mode,
2987                              ldlm_cancel_flags_t flags, void *opaque)
2988 {
2989         struct obd_device       *obd = exp->exp_obd;
2990         struct lmv_obd          *lmv = &obd->u.lmv;
2991         int                      rc = 0;
2992         int                      err;
2993         __u32                    i;
2994         ENTRY;
2995
2996         LASSERT(fid != NULL);
2997
2998         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2999                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
3000
3001                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
3002                         continue;
3003
3004                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3005                                        opaque);
3006                 if (!rc)
3007                         rc = err;
3008         }
3009         RETURN(rc);
3010 }
3011
3012 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
3013                       __u64 *bits)
3014 {
3015         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
3016         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3017         int                      rc;
3018         ENTRY;
3019
3020         if (tgt == NULL || tgt->ltd_exp == NULL)
3021                 RETURN(-EINVAL);
3022         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3023         RETURN(rc);
3024 }
3025
3026 ldlm_mode_t lmv_lock_match(struct obd_export *exp, __u64 flags,
3027                            const struct lu_fid *fid, ldlm_type_t type,
3028                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
3029                            struct lustre_handle *lockh)
3030 {
3031         struct obd_device       *obd = exp->exp_obd;
3032         struct lmv_obd          *lmv = &obd->u.lmv;
3033         ldlm_mode_t              rc;
3034         __u32                    i;
3035         ENTRY;
3036
3037         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
3038
3039         /*
3040          * With CMD every object can have two locks in different namespaces:
3041          * lookup lock in space of mds storing direntry and update/open lock in
3042          * space of mds storing inode. Thus we check all targets, not only that
3043          * one fid was created in.
3044          */
3045         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3046                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
3047
3048                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
3049                         continue;
3050
3051                 rc = md_lock_match(tgt->ltd_exp, flags, fid, type, policy, mode,
3052                                    lockh);
3053                 if (rc)
3054                         RETURN(rc);
3055         }
3056
3057         RETURN(0);
3058 }
3059
3060 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
3061                       struct obd_export *dt_exp, struct obd_export *md_exp,
3062                       struct lustre_md *md)
3063 {
3064         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
3065         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3066
3067         if (tgt == NULL || tgt->ltd_exp == NULL)
3068                 RETURN(-EINVAL);
3069
3070         return md_get_lustre_md(lmv->tgts[0]->ltd_exp, req, dt_exp, md_exp, md);
3071 }
3072
3073 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
3074 {
3075         struct obd_device       *obd = exp->exp_obd;
3076         struct lmv_obd          *lmv = &obd->u.lmv;
3077         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3078         ENTRY;
3079
3080         if (md->lmv != NULL) {
3081                 lmv_free_memmd(md->lmv);
3082                 md->lmv = NULL;
3083         }
3084         if (tgt == NULL || tgt->ltd_exp == NULL)
3085                 RETURN(-EINVAL);
3086         RETURN(md_free_lustre_md(lmv->tgts[0]->ltd_exp, md));
3087 }
3088
3089 int lmv_set_open_replay_data(struct obd_export *exp,
3090                              struct obd_client_handle *och,
3091                              struct lookup_intent *it)
3092 {
3093         struct obd_device       *obd = exp->exp_obd;
3094         struct lmv_obd          *lmv = &obd->u.lmv;
3095         struct lmv_tgt_desc     *tgt;
3096         ENTRY;
3097
3098         tgt = lmv_find_target(lmv, &och->och_fid);
3099         if (IS_ERR(tgt))
3100                 RETURN(PTR_ERR(tgt));
3101
3102         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3103 }
3104
3105 int lmv_clear_open_replay_data(struct obd_export *exp,
3106                                struct obd_client_handle *och)
3107 {
3108         struct obd_device       *obd = exp->exp_obd;
3109         struct lmv_obd          *lmv = &obd->u.lmv;
3110         struct lmv_tgt_desc     *tgt;
3111         ENTRY;
3112
3113         tgt = lmv_find_target(lmv, &och->och_fid);
3114         if (IS_ERR(tgt))
3115                 RETURN(PTR_ERR(tgt));
3116
3117         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3118 }
3119
3120 static int lmv_get_remote_perm(struct obd_export *exp,
3121                                const struct lu_fid *fid,
3122                                struct obd_capa *oc, __u32 suppgid,
3123                                struct ptlrpc_request **request)
3124 {
3125         struct obd_device       *obd = exp->exp_obd;
3126         struct lmv_obd          *lmv = &obd->u.lmv;
3127         struct lmv_tgt_desc     *tgt;
3128         int                      rc;
3129         ENTRY;
3130
3131         rc = lmv_check_connect(obd);
3132         if (rc)
3133                 RETURN(rc);
3134
3135         tgt = lmv_find_target(lmv, fid);
3136         if (IS_ERR(tgt))
3137                 RETURN(PTR_ERR(tgt));
3138
3139         rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
3140         RETURN(rc);
3141 }
3142
3143 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
3144                           renew_capa_cb_t cb)
3145 {
3146         struct obd_device       *obd = exp->exp_obd;
3147         struct lmv_obd          *lmv = &obd->u.lmv;
3148         struct lmv_tgt_desc     *tgt;
3149         int                      rc;
3150         ENTRY;
3151
3152         rc = lmv_check_connect(obd);
3153         if (rc)
3154                 RETURN(rc);
3155
3156         tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
3157         if (IS_ERR(tgt))
3158                 RETURN(PTR_ERR(tgt));
3159
3160         rc = md_renew_capa(tgt->ltd_exp, oc, cb);
3161         RETURN(rc);
3162 }
3163
3164 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
3165                     const struct req_msg_field *field, struct obd_capa **oc)
3166 {
3167         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
3168         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3169
3170         if (tgt == NULL || tgt->ltd_exp == NULL)
3171                 RETURN(-EINVAL);
3172         return md_unpack_capa(tgt->ltd_exp, req, field, oc);
3173 }
3174
3175 int lmv_intent_getattr_async(struct obd_export *exp,
3176                              struct md_enqueue_info *minfo,
3177                              struct ldlm_enqueue_info *einfo)
3178 {
3179         struct md_op_data       *op_data = &minfo->mi_data;
3180         struct obd_device       *obd = exp->exp_obd;
3181         struct lmv_obd          *lmv = &obd->u.lmv;
3182         struct lmv_tgt_desc     *tgt = NULL;
3183         int                      rc;
3184         ENTRY;
3185
3186         rc = lmv_check_connect(obd);
3187         if (rc)
3188                 RETURN(rc);
3189
3190         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
3191         if (IS_ERR(tgt))
3192                 RETURN(PTR_ERR(tgt));
3193
3194         rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
3195         RETURN(rc);
3196 }
3197
3198 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3199                         struct lu_fid *fid, __u64 *bits)
3200 {
3201         struct obd_device       *obd = exp->exp_obd;
3202         struct lmv_obd          *lmv = &obd->u.lmv;
3203         struct lmv_tgt_desc     *tgt;
3204         int                      rc;
3205         ENTRY;
3206
3207         rc = lmv_check_connect(obd);
3208         if (rc)
3209                 RETURN(rc);
3210
3211         tgt = lmv_find_target(lmv, fid);
3212         if (IS_ERR(tgt))
3213                 RETURN(PTR_ERR(tgt));
3214
3215         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3216         RETURN(rc);
3217 }
3218
3219 /**
3220  * For lmv, only need to send request to master MDT, and the master MDT will
3221  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3222  * we directly fetch data from the slave MDTs.
3223  */
3224 int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3225                  struct obd_quotactl *oqctl)
3226 {
3227         struct obd_device   *obd = class_exp2obd(exp);
3228         struct lmv_obd      *lmv = &obd->u.lmv;
3229         struct lmv_tgt_desc *tgt = lmv->tgts[0];
3230         int                  rc = 0;
3231         __u32                i;
3232         __u64                curspace, curinodes;
3233         ENTRY;
3234
3235         if (tgt == NULL ||
3236             tgt->ltd_exp == NULL ||
3237             !tgt->ltd_active ||
3238             lmv->desc.ld_tgt_count == 0) {
3239                 CERROR("master lmv inactive\n");
3240                 RETURN(-EIO);
3241         }
3242
3243         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3244                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3245                 RETURN(rc);
3246         }
3247
3248         curspace = curinodes = 0;
3249         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3250                 int err;
3251                 tgt = lmv->tgts[i];
3252
3253                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
3254                         continue;
3255
3256                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3257                 if (err) {
3258                         CERROR("getquota on mdt %d failed. %d\n", i, err);
3259                         if (!rc)
3260                                 rc = err;
3261                 } else {
3262                         curspace += oqctl->qc_dqblk.dqb_curspace;
3263                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3264                 }
3265         }
3266         oqctl->qc_dqblk.dqb_curspace = curspace;
3267         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3268
3269         RETURN(rc);
3270 }
3271
3272 int lmv_quotacheck(struct obd_device *unused, struct obd_export *exp,
3273                    struct obd_quotactl *oqctl)
3274 {
3275         struct obd_device       *obd = class_exp2obd(exp);
3276         struct lmv_obd          *lmv = &obd->u.lmv;
3277         struct lmv_tgt_desc     *tgt;
3278         __u32                    i;
3279         int                      rc = 0;
3280         ENTRY;
3281
3282         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3283                 int err;
3284                 tgt = lmv->tgts[i];
3285                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active) {
3286                         CERROR("lmv idx %d inactive\n", i);
3287                         RETURN(-EIO);
3288                 }
3289
3290                 err = obd_quotacheck(tgt->ltd_exp, oqctl);
3291                 if (err && !rc)
3292                         rc = err;
3293         }
3294
3295         RETURN(rc);
3296 }
3297
3298 int lmv_update_lsm_md(struct obd_export *exp, struct lmv_stripe_md *lsm,
3299                       struct mdt_body *body, ldlm_blocking_callback cb_blocking)
3300 {
3301         if (lsm->lsm_md_stripe_count <= 1)
3302                 return 0;
3303
3304         return lmv_revalidate_slaves(exp, body, lsm, cb_blocking, 0);
3305 }
3306
3307 int lmv_merge_attr(struct obd_export *exp, const struct lmv_stripe_md *lsm,
3308                    struct cl_attr *attr)
3309 {
3310 #ifdef __KERNEL__
3311         int i;
3312
3313         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3314                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3315
3316                 CDEBUG(D_INFO, ""DFID" size %llu, nlink %u, atime %lu ctime"
3317                        "%lu, mtime %lu.\n", PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3318                        i_size_read(inode), inode->i_nlink,
3319                        LTIME_S(inode->i_atime), LTIME_S(inode->i_ctime),
3320                        LTIME_S(inode->i_mtime));
3321
3322                 /* for slave stripe, it needs to subtract nlink for . and .. */
3323                 if (i != 0)
3324                         attr->cat_nlink += inode->i_nlink - 2;
3325                 else
3326                         attr->cat_nlink = inode->i_nlink;
3327
3328                 attr->cat_size += i_size_read(inode);
3329
3330                 if (attr->cat_atime < LTIME_S(inode->i_atime))
3331                         attr->cat_atime = LTIME_S(inode->i_atime);
3332
3333                 if (attr->cat_ctime < LTIME_S(inode->i_ctime))
3334                         attr->cat_ctime = LTIME_S(inode->i_ctime);
3335
3336                 if (attr->cat_mtime < LTIME_S(inode->i_mtime))
3337                         attr->cat_mtime = LTIME_S(inode->i_mtime);
3338         }
3339 #endif
3340         return 0;
3341 }
3342
3343 struct obd_ops lmv_obd_ops = {
3344         .o_owner                = THIS_MODULE,
3345         .o_setup                = lmv_setup,
3346         .o_cleanup              = lmv_cleanup,
3347         .o_precleanup           = lmv_precleanup,
3348         .o_process_config       = lmv_process_config,
3349         .o_connect              = lmv_connect,
3350         .o_disconnect           = lmv_disconnect,
3351         .o_statfs               = lmv_statfs,
3352         .o_get_info             = lmv_get_info,
3353         .o_set_info_async       = lmv_set_info_async,
3354         .o_packmd               = lmv_packmd,
3355         .o_unpackmd             = lmv_unpackmd,
3356         .o_notify               = lmv_notify,
3357         .o_get_uuid             = lmv_get_uuid,
3358         .o_iocontrol            = lmv_iocontrol,
3359         .o_quotacheck           = lmv_quotacheck,
3360         .o_quotactl             = lmv_quotactl
3361 };
3362
3363 struct md_ops lmv_md_ops = {
3364         .m_getstatus            = lmv_getstatus,
3365         .m_null_inode           = lmv_null_inode,
3366         .m_find_cbdata          = lmv_find_cbdata,
3367         .m_close                = lmv_close,
3368         .m_create               = lmv_create,
3369         .m_done_writing         = lmv_done_writing,
3370         .m_enqueue              = lmv_enqueue,
3371         .m_getattr              = lmv_getattr,
3372         .m_getxattr             = lmv_getxattr,
3373         .m_getattr_name         = lmv_getattr_name,
3374         .m_intent_lock          = lmv_intent_lock,
3375         .m_link                 = lmv_link,
3376         .m_rename               = lmv_rename,
3377         .m_setattr              = lmv_setattr,
3378         .m_setxattr             = lmv_setxattr,
3379         .m_fsync                = lmv_fsync,
3380         .m_read_entry           = lmv_read_entry,
3381         .m_unlink               = lmv_unlink,
3382         .m_init_ea_size         = lmv_init_ea_size,
3383         .m_cancel_unused        = lmv_cancel_unused,
3384         .m_set_lock_data        = lmv_set_lock_data,
3385         .m_lock_match           = lmv_lock_match,
3386         .m_get_lustre_md        = lmv_get_lustre_md,
3387         .m_free_lustre_md       = lmv_free_lustre_md,
3388         .m_update_lsm_md        = lmv_update_lsm_md,
3389         .m_merge_attr           = lmv_merge_attr,
3390         .m_set_open_replay_data = lmv_set_open_replay_data,
3391         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3392         .m_renew_capa           = lmv_renew_capa,
3393         .m_unpack_capa          = lmv_unpack_capa,
3394         .m_get_remote_perm      = lmv_get_remote_perm,
3395         .m_intent_getattr_async = lmv_intent_getattr_async,
3396         .m_revalidate_lock      = lmv_revalidate_lock
3397 };
3398
3399 int __init lmv_init(void)
3400 {
3401         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true, NULL,
3402 #ifndef HAVE_ONLY_PROCFS_SEQ
3403                                    NULL,
3404 #endif
3405                                    LUSTRE_LMV_NAME, NULL);
3406 }
3407
3408 #ifdef __KERNEL__
3409 static void lmv_exit(void)
3410 {
3411         class_unregister_type(LUSTRE_LMV_NAME);
3412 }
3413
3414 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3415 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3416 MODULE_LICENSE("GPL");
3417
3418 module_init(lmv_init);
3419 module_exit(lmv_exit);
3420 #endif