lustre/lmv/lmv_obd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LMV
  41 #ifdef __KERNEL__
  42 #include <linux/slab.h>
  43 #include <linux/module.h>
  44 #include <linux/init.h>
  45 #include <linux/slab.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/mm.h>
  48 #include <asm/div64.h>
  49 #include <linux/seq_file.h>
  50 #include <linux/namei.h>
  51 #else
  52 #include <liblustre.h>
  53 #endif
  54
  55 #include <lustre/lustre_idl.h>
  56 #include <lustre_log.h>
  57 #include <obd_support.h>
  58 #include <lustre_lib.h>
  59 #include <lustre_net.h>
  60 #include <obd_class.h>
  61 #include <lprocfs_status.h>
  62 #include <lustre_lite.h>
  63 #include <lustre_fid.h>
  64 #include "lmv_internal.h"
  65
  66 /* object cache. */
  67 cfs_mem_cache_t *lmv_object_cache;
  68 atomic_t lmv_object_count = ATOMIC_INIT(0);
  69
  70 static void lmv_activate_target(struct lmv_obd *lmv,
  71                                 struct lmv_tgt_desc *tgt,
  72                                 int activate)
  73 {
  74         if (tgt->ltd_active == activate)
  75                 return;
  76
  77         tgt->ltd_active = activate;
  78         lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
  79 }
  80
  81 /**
  82  * Error codes:
  83  *
  84  *  -EINVAL  : UUID can't be found in the LMV's target list
  85  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  86  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  87  */
  88 static int lmv_set_mdc_active(struct lmv_obd *lmv, struct obd_uuid *uuid,
  89                               int activate)
  90 {
  91         struct lmv_tgt_desc    *tgt;
  92         struct obd_device      *obd;
  93         int                     i;
  94         int                     rc = 0;
  95         ENTRY;
  96
  97         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  98                lmv, uuid->uuid, activate);
  99
 100         spin_lock(&lmv->lmv_lock);
 101         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 102                 if (tgt->ltd_exp == NULL)
 103                         continue;
 104
 105                 CDEBUG(D_INFO, "Target idx %d is %s conn "LPX64"\n",
 106                        i, tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
 107
 108                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 109                         break;
 110         }
 111
 112         if (i == lmv->desc.ld_tgt_count)
 113                 GOTO(out_lmv_lock, rc = -EINVAL);
 114
 115         obd = class_exp2obd(tgt->ltd_exp);
 116         if (obd == NULL)
 117                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 118
 119         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 120                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 121                obd->obd_type->typ_name, i);
 122         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 123
 124         if (tgt->ltd_active == activate) {
 125                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 126                        activate ? "" : "in");
 127                 GOTO(out_lmv_lock, rc);
 128         }
 129
 130         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 131                activate ? "" : "in");
 132         lmv_activate_target(lmv, tgt, activate);
 133         EXIT;
 134
 135  out_lmv_lock:
 136         spin_unlock(&lmv->lmv_lock);
 137         return rc;
 138 }
 139
 140 static int lmv_set_mdc_data(struct lmv_obd *lmv, struct obd_uuid *uuid,
 141                             struct obd_connect_data *data)
 142 {
 143         struct lmv_tgt_desc    *tgt;
 144         int                     i;
 145         ENTRY;
 146
 147         LASSERT(data != NULL);
 148
 149         spin_lock(&lmv->lmv_lock);
 150         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 151                 if (tgt->ltd_exp == NULL)
 152                         continue;
 153
 154                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
 155                         lmv->datas[tgt->ltd_idx] = *data;
 156                         break;
 157                 }
 158         }
 159         spin_unlock(&lmv->lmv_lock);
 160         RETURN(0);
 161 }
 162
 163 struct obd_uuid *lmv_get_uuid(struct obd_export *exp) {
 164         struct obd_device *obd = exp->exp_obd;
 165         struct lmv_obd *lmv = &obd->u.lmv;
 166         return obd_get_uuid(lmv->tgts[0].ltd_exp);
 167 }
 168
 169 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 170                       enum obd_notify_event ev, void *data)
 171 {
 172         struct obd_connect_data *conn_data;
 173         struct lmv_obd          *lmv = &obd->u.lmv;
 174         struct obd_uuid         *uuid;
 175         int                      rc = 0;
 176         ENTRY;
 177
 178         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 179                 CERROR("unexpected notification of %s %s!\n",
 180                        watched->obd_type->typ_name,
 181                        watched->obd_name);
 182                 RETURN(-EINVAL);
 183         }
 184
 185         uuid = &watched->u.cli.cl_target_uuid;
 186         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 187                 /*
 188                  * Set MDC as active before notifying the observer, so the
 189                  * observer can use the MDC normally.
 190                  */
 191                 rc = lmv_set_mdc_active(lmv, uuid,
 192                                         ev == OBD_NOTIFY_ACTIVE);
 193                 if (rc) {
 194                         CERROR("%sactivation of %s failed: %d\n",
 195                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 196                                uuid->uuid, rc);
 197                         RETURN(rc);
 198                 }
 199         } else if (ev == OBD_NOTIFY_OCD) {
 200                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 201
 202                 /*
 203                  * Set connect data to desired target, update exp_connect_flags.
 204                  */
 205                 rc = lmv_set_mdc_data(lmv, uuid, conn_data);
 206                 if (rc) {
 207                         CERROR("can't set connect data to target %s, rc %d\n",
 208                                uuid->uuid, rc);
 209                         RETURN(rc);
 210                 }
 211
 212                 /*
 213                  * XXX: Make sure that ocd_connect_flags from all targets are
 214                  * the same. Otherwise one of MDTs runs wrong version or
 215                  * something like this.  --umka
 216                  */
 217                 obd->obd_self_export->exp_connect_flags =
 218                         conn_data->ocd_connect_flags;
 219         }
 220 #if 0
 221         else if (ev == OBD_NOTIFY_DISCON) {
 222                 /*
 223                  * For disconnect event, flush fld cache for failout MDS case.
 224                  */
 225                 fld_client_flush(&lmv->lmv_fld);
 226         }
 227 #endif
 228         /*
 229          * Pass the notification up the chain.
 230          */
 231         if (obd->obd_observer)
 232                 rc = obd_notify(obd->obd_observer, watched, ev, data);
 233
 234         RETURN(rc);
 235 }
 236
 237 /**
 238  * This is fake connect function. Its purpose is to initialize lmv and say
 239  * caller that everything is okay. Real connection will be performed later.
 240  */
 241 static int lmv_connect(const struct lu_env *env,
 242                        struct obd_export **exp, struct obd_device *obd,
 243                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 244                        void *localdata)
 245 {
 246 #ifdef __KERNEL__
 247         struct proc_dir_entry *lmv_proc_dir;
 248 #endif
 249         struct lmv_obd        *lmv = &obd->u.lmv;
 250         struct lustre_handle  conn = { 0 };
 251         int                    rc = 0;
 252         ENTRY;
 253
 254         /*
 255          * We don't want to actually do the underlying connections more than
 256          * once, so keep track.
 257          */
 258         lmv->refcount++;
 259         if (lmv->refcount > 1) {
 260                 *exp = NULL;
 261                 RETURN(0);
 262         }
 263
 264         rc = class_connect(&conn, obd, cluuid);
 265         if (rc) {
 266                 CERROR("class_connection() returned %d\n", rc);
 267                 RETURN(rc);
 268         }
 269
 270         *exp = class_conn2export(&conn);
 271         class_export_get(*exp);
 272
 273         lmv->exp = *exp;
 274         lmv->connected = 0;
 275         lmv->cluuid = *cluuid;
 276
 277         if (data)
 278                 lmv->conn_data = *data;
 279
 280 #ifdef __KERNEL__
 281         lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
 282                                         NULL, NULL);
 283         if (IS_ERR(lmv_proc_dir)) {
 284                 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
 285                        obd->obd_type->typ_name, obd->obd_name);
 286                 lmv_proc_dir = NULL;
 287         }
 288 #endif
 289
 290         /*
 291          * All real clients should perform actual connection right away, because
 292          * it is possible, that LMV will not have opportunity to connect targets
 293          * and MDC stuff will be called directly, for instance while reading
 294          * ../mdc/../kbytesfree procfs file, etc.
 295          */
 296         if (data->ocd_connect_flags & OBD_CONNECT_REAL)
 297                 rc = lmv_check_connect(obd);
 298
 299 #ifdef __KERNEL__
 300         if (rc) {
 301                 if (lmv_proc_dir)
 302                         lprocfs_remove(&lmv_proc_dir);
 303         }
 304 #endif
 305
 306         RETURN(rc);
 307 }
 308
 309 static void lmv_set_timeouts(struct obd_device *obd)
 310 {
 311         struct lmv_tgt_desc   *tgts;
 312         struct lmv_obd        *lmv;
 313         int                    i;
 314
 315         lmv = &obd->u.lmv;
 316         if (lmv->server_timeout == 0)
 317                 return;
 318
 319         if (lmv->connected == 0)
 320                 return;
 321
 322         for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
 323                 if (tgts->ltd_exp == NULL)
 324                         continue;
 325
 326                 obd_set_info_async(tgts->ltd_exp, sizeof(KEY_INTERMDS),
 327                                    KEY_INTERMDS, 0, NULL, NULL);
 328         }
 329 }
 330
 331 static int lmv_init_ea_size(struct obd_export *exp, int easize,
 332                             int def_easize, int cookiesize)
 333 {
 334         struct obd_device   *obd = exp->exp_obd;
 335         struct lmv_obd      *lmv = &obd->u.lmv;
 336         int                  i;
 337         int                  rc = 0;
 338         int                  change = 0;
 339         ENTRY;
 340
 341         if (lmv->max_easize < easize) {
 342                 lmv->max_easize = easize;
 343                 change = 1;
 344         }
 345         if (lmv->max_def_easize < def_easize) {
 346                 lmv->max_def_easize = def_easize;
 347                 change = 1;
 348         }
 349         if (lmv->max_cookiesize < cookiesize) {
 350                 lmv->max_cookiesize = cookiesize;
 351                 change = 1;
 352         }
 353         if (change == 0)
 354                 RETURN(0);
 355
 356         if (lmv->connected == 0)
 357                 RETURN(0);
 358
 359         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 360                 if (lmv->tgts[i].ltd_exp == NULL) {
 361                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
 362                         continue;
 363                 }
 364
 365                 rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize,
 366                                      cookiesize);
 367                 if (rc) {
 368                         CERROR("obd_init_ea_size() failed on MDT target %d, "
 369                                "error %d.\n", i, rc);
 370                         break;
 371                 }
 372         }
 373         RETURN(rc);
 374 }
 375
 376 #define MAX_STRING_SIZE 128
 377
 378 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 379 {
 380 #ifdef __KERNEL__
 381         struct proc_dir_entry   *lmv_proc_dir;
 382 #endif
 383         struct lmv_obd          *lmv = &obd->u.lmv;
 384         struct obd_uuid         *cluuid = &lmv->cluuid;
 385         struct obd_connect_data *mdc_data = NULL;
 386         struct obd_uuid          lmv_mdc_uuid = { "LMV_MDC_UUID" };
 387         struct obd_device       *mdc_obd;
 388         struct obd_export       *mdc_exp;
 389         struct lu_fld_target     target;
 390         int                      rc;
 391         ENTRY;
 392
 393         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 394                                         &obd->obd_uuid);
 395         if (!mdc_obd) {
 396                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 397                 RETURN(-EINVAL);
 398         }
 399
 400         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
 401                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 402                 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
 403                 cluuid->uuid);
 404
 405         if (!mdc_obd->obd_set_up) {
 406                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 407                 RETURN(-EINVAL);
 408         }
 409
 410         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &lmv_mdc_uuid,
 411                          &lmv->conn_data, NULL);
 412         if (rc) {
 413                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 414                 RETURN(rc);
 415         }
 416
 417         /*
 418          * Init fid sequence client for this mdc and add new fld target.
 419          */
 420         rc = obd_fid_init(mdc_exp);
 421         if (rc)
 422                 RETURN(rc);
 423
 424         target.ft_srv = NULL;
 425         target.ft_exp = mdc_exp;
 426         target.ft_idx = tgt->ltd_idx;
 427
 428         fld_client_add_target(&lmv->lmv_fld, &target);
 429
 430         mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data;
 431
 432         rc = obd_register_observer(mdc_obd, obd);
 433         if (rc) {
 434                 obd_disconnect(mdc_exp);
 435                 CERROR("target %s register_observer error %d\n",
 436                        tgt->ltd_uuid.uuid, rc);
 437                 RETURN(rc);
 438         }
 439
 440         if (obd->obd_observer) {
 441                 /*
 442                  * Tell the observer about the new target.
 443                  */
 444                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 445                                 OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts));
 446                 if (rc) {
 447                         obd_disconnect(mdc_exp);
 448                         RETURN(rc);
 449                 }
 450         }
 451
 452         tgt->ltd_active = 1;
 453         tgt->ltd_exp = mdc_exp;
 454         lmv->desc.ld_active_tgt_count++;
 455
 456         /*
 457          * Copy connect data, it may be used later.
 458          */
 459         lmv->datas[tgt->ltd_idx] = *mdc_data;
 460
 461         md_init_ea_size(tgt->ltd_exp, lmv->max_easize,
 462                         lmv->max_def_easize, lmv->max_cookiesize);
 463
 464         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 465                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 466                 atomic_read(&obd->obd_refcount));
 467
 468 #ifdef __KERNEL__
 469         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 470         if (lmv_proc_dir) {
 471                 struct proc_dir_entry *mdc_symlink;
 472
 473                 LASSERT(mdc_obd->obd_type != NULL);
 474                 LASSERT(mdc_obd->obd_type->typ_name != NULL);
 475                 mdc_symlink = lprocfs_add_symlink(mdc_obd->obd_name,
 476                                                   lmv_proc_dir,
 477                                                   "../../../%s/%s",
 478                                                   mdc_obd->obd_type->typ_name,
 479                                                   mdc_obd->obd_name);
 480                 if (mdc_symlink == NULL) {
 481                         CERROR("Could not register LMV target "
 482                                "/proc/fs/lustre/%s/%s/target_obds/%s.",
 483                                obd->obd_type->typ_name, obd->obd_name,
 484                                mdc_obd->obd_name);
 485                         lprocfs_remove(&lmv_proc_dir);
 486                         lmv_proc_dir = NULL;
 487                 }
 488         }
 489 #endif
 490         RETURN(0);
 491 }
 492
 493 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
 494 {
 495         struct lmv_obd      *lmv = &obd->u.lmv;
 496         struct lmv_tgt_desc *tgt;
 497         int                  rc = 0;
 498         ENTRY;
 499
 500         CDEBUG(D_CONFIG, "Target uuid: %s.\n", tgt_uuid->uuid);
 501
 502         lmv_init_lock(lmv);
 503
 504         if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
 505                 lmv_init_unlock(lmv);
 506                 CERROR("Can't add %s, LMV module compiled for %d MDCs. "
 507                        "That many MDCs already configured.\n",
 508                        tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
 509                 RETURN(-EINVAL);
 510         }
 511         if (lmv->desc.ld_tgt_count == 0) {
 512                 struct obd_device *mdc_obd;
 513
 514                 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
 515                                                 &obd->obd_uuid);
 516                 if (!mdc_obd) {
 517                         lmv_init_unlock(lmv);
 518                         CERROR("Target %s not attached\n", tgt_uuid->uuid);
 519                         RETURN(-EINVAL);
 520                 }
 521
 522                 rc = obd_llog_init(obd, &obd->obd_olg, mdc_obd, NULL);
 523                 if (rc) {
 524                         lmv_init_unlock(lmv);
 525                         CERROR("lmv failed to setup llogging subsystems\n");
 526                 }
 527         }
 528         spin_lock(&lmv->lmv_lock);
 529         tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
 530         tgt->ltd_uuid = *tgt_uuid;
 531         spin_unlock(&lmv->lmv_lock);
 532
 533         if (lmv->connected) {
 534                 rc = lmv_connect_mdc(obd, tgt);
 535                 if (rc) {
 536                         spin_lock(&lmv->lmv_lock);
 537                         lmv->desc.ld_tgt_count--;
 538                         memset(tgt, 0, sizeof(*tgt));
 539                         spin_unlock(&lmv->lmv_lock);
 540                 } else {
 541                         int easize = sizeof(struct lmv_stripe_md) +
 542                                      lmv->desc.ld_tgt_count *
 543                                      sizeof(struct lu_fid);
 544                         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 545                 }
 546         }
 547
 548         lmv_init_unlock(lmv);
 549         RETURN(rc);
 550 }
 551
 552 int lmv_check_connect(struct obd_device *obd)
 553 {
 554         struct lmv_obd       *lmv = &obd->u.lmv;
 555         struct lmv_tgt_desc  *tgt;
 556         int                   i;
 557         int                   rc;
 558         int                   easize;
 559         ENTRY;
 560
 561         if (lmv->connected)
 562                 RETURN(0);
 563
 564         lmv_init_lock(lmv);
 565         if (lmv->connected) {
 566                 lmv_init_unlock(lmv);
 567                 RETURN(0);
 568         }
 569
 570         if (lmv->desc.ld_tgt_count == 0) {
 571                 CERROR("%s: no targets configured.\n", obd->obd_name);
 572                 RETURN(-EINVAL);
 573         }
 574
 575         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 576                lmv->cluuid.uuid, obd->obd_name);
 577
 578         LASSERT(lmv->tgts != NULL);
 579
 580         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 581                 rc = lmv_connect_mdc(obd, tgt);
 582                 if (rc)
 583                         GOTO(out_disc, rc);
 584         }
 585
 586         lmv_set_timeouts(obd);
 587         class_export_put(lmv->exp);
 588         lmv->connected = 1;
 589         easize = lmv_get_easize(lmv);
 590         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 591         lmv_init_unlock(lmv);
 592         RETURN(0);
 593
 594  out_disc:
 595         while (i-- > 0) {
 596                 int rc2;
 597                 --tgt;
 598                 tgt->ltd_active = 0;
 599                 if (tgt->ltd_exp) {
 600                         --lmv->desc.ld_active_tgt_count;
 601                         rc2 = obd_disconnect(tgt->ltd_exp);
 602                         if (rc2) {
 603                                 CERROR("LMV target %s disconnect on "
 604                                        "MDC idx %d: error %d\n",
 605                                        tgt->ltd_uuid.uuid, i, rc2);
 606                         }
 607                 }
 608         }
 609         class_disconnect(lmv->exp);
 610         lmv_init_unlock(lmv);
 611         RETURN(rc);
 612 }
 613
 614 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 615 {
 616 #ifdef __KERNEL__
 617         struct proc_dir_entry  *lmv_proc_dir;
 618 #endif
 619         struct lmv_obd         *lmv = &obd->u.lmv;
 620         struct obd_device      *mdc_obd;
 621         int                     rc;
 622         ENTRY;
 623
 624         LASSERT(tgt != NULL);
 625         LASSERT(obd != NULL);
 626
 627         mdc_obd = class_exp2obd(tgt->ltd_exp);
 628
 629         if (mdc_obd)
 630                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 631
 632 #ifdef __KERNEL__
 633         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 634         if (lmv_proc_dir) {
 635                 struct proc_dir_entry *mdc_symlink;
 636
 637                 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
 638                 if (mdc_symlink) {
 639                         lprocfs_remove(&mdc_symlink);
 640                 } else {
 641                         CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
 642                                obd->obd_type->typ_name, obd->obd_name,
 643                                mdc_obd->obd_name);
 644                 }
 645         }
 646 #endif
 647         rc = obd_fid_fini(tgt->ltd_exp);
 648         if (rc)
 649                 CERROR("Can't finanize fids factory\n");
 650
 651         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 652                tgt->ltd_exp->exp_obd->obd_name,
 653                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 654
 655         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 656         rc = obd_disconnect(tgt->ltd_exp);
 657         if (rc) {
 658                 if (tgt->ltd_active) {
 659                         CERROR("Target %s disconnect error %d\n",
 660                                tgt->ltd_uuid.uuid, rc);
 661                 }
 662         }
 663
 664         lmv_activate_target(lmv, tgt, 0);
 665         tgt->ltd_exp = NULL;
 666         RETURN(0);
 667 }
 668
 669 static int lmv_disconnect(struct obd_export *exp)
 670 {
 671         struct obd_device     *obd = class_exp2obd(exp);
 672 #ifdef __KERNEL__
 673         struct proc_dir_entry *lmv_proc_dir;
 674 #endif
 675         struct lmv_obd        *lmv = &obd->u.lmv;
 676         int                    rc;
 677         int                    i;
 678         ENTRY;
 679
 680         if (!lmv->tgts)
 681                 goto out_local;
 682
 683         /*
 684          * Only disconnect the underlying layers on the final disconnect.
 685          */
 686         lmv->refcount--;
 687         if (lmv->refcount != 0)
 688                 goto out_local;
 689
 690         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 691                 if (lmv->tgts[i].ltd_exp == NULL)
 692                         continue;
 693                 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
 694         }
 695
 696 #ifdef __KERNEL__
 697         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 698         if (lmv_proc_dir) {
 699                 lprocfs_remove(&lmv_proc_dir);
 700         } else {
 701                 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
 702                        obd->obd_type->typ_name, obd->obd_name);
 703         }
 704 #endif
 705
 706 out_local:
 707         /*
 708          * This is the case when no real connection is established by
 709          * lmv_check_connect().
 710          */
 711         if (!lmv->connected)
 712                 class_export_put(exp);
 713         rc = class_disconnect(exp);
 714         if (lmv->refcount == 0)
 715                 lmv->connected = 0;
 716         RETURN(rc);
 717 }
 718
 719 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 720                          int len, void *karg, void *uarg)
 721 {
 722         struct obd_device    *obddev = class_exp2obd(exp);
 723         struct lmv_obd       *lmv = &obddev->u.lmv;
 724         int                   i = 0;
 725         int                   rc = 0;
 726         int                   set = 0;
 727         int                   count = lmv->desc.ld_tgt_count;
 728         ENTRY;
 729
 730         if (count == 0)
 731                 RETURN(-ENOTTY);
 732
 733         switch (cmd) {
 734         case IOC_OBD_STATFS: {
 735                 struct obd_ioctl_data *data = karg;
 736                 struct obd_device *mdc_obd;
 737                 struct obd_statfs stat_buf = {0};
 738                 __u32 index;
 739
 740                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 741                 LASSERT(data->ioc_plen1 == sizeof(struct obd_statfs));
 742
 743                 if ((index >= count))
 744                         RETURN(-ENODEV);
 745
 746                 if (!lmv->tgts[index].ltd_active)
 747                         RETURN(-ENODATA);
 748
 749                 mdc_obd = class_exp2obd(lmv->tgts[index].ltd_exp);
 750                 if (!mdc_obd)
 751                         RETURN(-EINVAL);
 752
 753                 rc = obd_statfs(mdc_obd, &stat_buf,
 754                                 cfs_time_current_64() - HZ, 0);
 755                 if (rc)
 756                         RETURN(rc);
 757                 if (copy_to_user(data->ioc_pbuf1, &stat_buf, data->ioc_plen1))
 758                         RETURN(-EFAULT);
 759                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 760                                   data->ioc_plen2))
 761                         RETURN(-EFAULT);
 762                 break;
 763         }
 764         case OBD_IOC_QUOTACTL: {
 765                 struct if_quotactl *qctl = karg;
 766                 struct lmv_tgt_desc *tgt = NULL;
 767                 struct obd_quotactl *oqctl;
 768
 769                 if (qctl->qc_valid == QC_MDTIDX) {
 770                         if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
 771                                 RETURN(-EINVAL);
 772
 773                         tgt = &lmv->tgts[qctl->qc_idx];
 774                         if (!tgt->ltd_exp)
 775                                 RETURN(-EINVAL);
 776                 } else if (qctl->qc_valid == QC_UUID) {
 777                         for (i = 0; i < count; i++) {
 778                                 tgt = &lmv->tgts[i];
 779                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 780                                                      &qctl->obd_uuid))
 781                                         continue;
 782
 783                                 if (tgt->ltd_exp == NULL)
 784                                         RETURN(-EINVAL);
 785
 786                                 break;
 787                         }
 788                 } else {
 789                         RETURN(-EINVAL);
 790                 }
 791
 792                 if (i >= count)
 793                         RETURN(-EAGAIN);
 794
 795                 LASSERT(tgt && tgt->ltd_exp);
 796                 OBD_ALLOC_PTR(oqctl);
 797                 if (!oqctl)
 798                         RETURN(-ENOMEM);
 799
 800                 QCTL_COPY(oqctl, qctl);
 801                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 802                 if (rc == 0) {
 803                         QCTL_COPY(qctl, oqctl);
 804                         qctl->qc_valid = QC_MDTIDX;
 805                         qctl->obd_uuid = tgt->ltd_uuid;
 806                 }
 807                 OBD_FREE_PTR(oqctl);
 808                 break;
 809         }
 810         case OBD_IOC_CHANGELOG_CLEAR: {
 811                 struct ioc_changelog_clear *icc = karg;
 812
 813                 if (icc->icc_mdtindex >= count)
 814                         RETURN(-ENODEV);
 815
 816                 rc = obd_iocontrol(cmd, lmv->tgts[icc->icc_mdtindex].ltd_exp,
 817                                    sizeof(*icc), icc, NULL);
 818                 break;
 819         }
 820
 821         default : {
 822                 for (i = 0; i < count; i++) {
 823                         int err;
 824
 825                         if (lmv->tgts[i].ltd_exp == NULL)
 826                                 continue;
 827
 828                         err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
 829                                             karg, uarg);
 830                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
 831                                 RETURN(err);
 832                         } else if (err) {
 833                                 if (lmv->tgts[i].ltd_active) {
 834                                         CERROR("error: iocontrol MDC %s on MDT"
 835                                                "idx %d cmd %x: err = %d\n",
 836                                                 lmv->tgts[i].ltd_uuid.uuid,
 837                                                 i, cmd, err);
 838                                         if (!rc)
 839                                                 rc = err;
 840                                 }
 841                         } else
 842                                 set = 1;
 843                 }
 844                 if (!set && !rc)
 845                         rc = -EIO;
 846         }
 847         }
 848         RETURN(rc);
 849 }
 850
 851 static int lmv_all_chars_policy(int count, const char *name,
 852                                 int len)
 853 {
 854         unsigned int c = 0;
 855
 856         while (len > 0)
 857                 c += name[--len];
 858         c = c % count;
 859         return c;
 860 }
 861
 862 static int lmv_nid_policy(struct lmv_obd *lmv)
 863 {
 864         struct obd_import *imp;
 865         __u32              id;
 866
 867         /*
 868          * XXX: To get nid we assume that underlying obd device is mdc.
 869          */
 870         imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
 871         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
 872         return id % lmv->desc.ld_tgt_count;
 873 }
 874
 875 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
 876                           placement_policy_t placement)
 877 {
 878         switch (placement) {
 879         case PLACEMENT_CHAR_POLICY:
 880                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
 881                                             op_data->op_name,
 882                                             op_data->op_namelen);
 883         case PLACEMENT_NID_POLICY:
 884                 return lmv_nid_policy(lmv);
 885
 886         default:
 887                 break;
 888         }
 889
 890         CERROR("Unsupported placement policy %x\n", placement);
 891         return -EINVAL;
 892 }
 893
 894 /**
 895  * This is _inode_ placement policy function (not name).
 896  */
 897 static int lmv_placement_policy(struct obd_device *obd,
 898                                 struct md_op_data *op_data,
 899                                 mdsno_t *mds)
 900 {
 901         struct lmv_obd          *lmv = &obd->u.lmv;
 902         struct lmv_object       *obj;
 903         int                      rc;
 904         ENTRY;
 905
 906         LASSERT(mds != NULL);
 907
 908         if (lmv->desc.ld_tgt_count == 1) {
 909                 *mds = 0;
 910                 RETURN(0);
 911         }
 912
 913         /*
 914          * Allocate new fid on target according to operation type and parent
 915          * home mds.
 916          */
 917         obj = lmv_object_find(obd, &op_data->op_fid1);
 918         if (obj != NULL || op_data->op_name == NULL ||
 919             op_data->op_opc != LUSTRE_OPC_MKDIR) {
 920                 /*
 921                  * Allocate fid for non-dir or for null name or for case parent
 922                  * dir is split.
 923                  */
 924                 if (obj) {
 925                         lmv_object_put(obj);
 926
 927                         /*
 928                          * If we have this flag turned on, and we see that
 929                          * parent dir is split, this means, that caller did not
 930                          * notice split yet. This is race and we would like to
 931                          * let caller know that.
 932                          */
 933                         if (op_data->op_bias & MDS_CHECK_SPLIT)
 934                                 RETURN(-ERESTART);
 935                 }
 936
 937                 /*
 938                  * Allocate new fid on same mds where parent fid is located and
 939                  * where operation will be sent. In case of split dir, ->op_fid1
 940                  * and ->op_mds here will contain fid and mds of slave directory
 941                  * object (assigned by caller).
 942                  */
 943                 *mds = op_data->op_mds;
 944                 rc = 0;
 945         } else {
 946                 /*
 947                  * Parent directory is not split and we want to create a
 948                  * directory in it. Let's calculate where to place it according
 949                  * to operation data @op_data.
 950                  */
 951                 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
 952                 rc = 0;
 953         }
 954
 955         if (rc) {
 956                 CERROR("Can't choose MDS, err = %d\n", rc);
 957         } else {
 958                 LASSERT(*mds < lmv->desc.ld_tgt_count);
 959         }
 960
 961         RETURN(rc);
 962 }
 963
 964 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
 965                     mdsno_t mds)
 966 {
 967         struct lmv_tgt_desc *tgt;
 968         int                  rc;
 969         ENTRY;
 970
 971         tgt = lmv_get_target(lmv, mds);
 972
 973         /*
 974          * New seq alloc and FLD setup should be atomic. Otherwise we may find
 975          * on server that seq in new allocated fid is not yet known.
 976          */
 977         down(&tgt->ltd_fid_sem);
 978
 979         if (!tgt->ltd_active)
 980                 GOTO(out, rc = -ENODEV);
 981
 982         /*
 983          * Asking underlaying tgt layer to allocate new fid.
 984          */
 985         rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
 986         if (rc > 0) {
 987                 LASSERT(fid_is_sane(fid));
 988                 rc = 0;
 989         }
 990
 991         EXIT;
 992 out:
 993         up(&tgt->ltd_fid_sem);
 994         return rc;
 995 }
 996
 997 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
 998                   struct md_op_data *op_data)
 999 {
1000         struct obd_device     *obd = class_exp2obd(exp);
1001         struct lmv_obd        *lmv = &obd->u.lmv;
1002         mdsno_t                mds;
1003         int                    rc;
1004         ENTRY;
1005
1006         LASSERT(op_data != NULL);
1007         LASSERT(fid != NULL);
1008
1009         rc = lmv_placement_policy(obd, op_data, &mds);
1010         if (rc) {
1011                 CERROR("Can't get target for allocating fid, "
1012                        "rc %d\n", rc);
1013                 RETURN(rc);
1014         }
1015
1016         rc = __lmv_fid_alloc(lmv, fid, mds);
1017         if (rc) {
1018                 CERROR("Can't alloc new fid, rc %d\n", rc);
1019                 RETURN(rc);
1020         }
1021
1022         RETURN(rc);
1023 }
1024
1025 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1026 {
1027         ENTRY;
1028         LASSERT(exp != NULL && fid != NULL);
1029         if (lmv_object_delete(exp, fid)) {
1030                 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1031                        PFID(fid));
1032         }
1033         RETURN(0);
1034 }
1035
1036 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1037 {
1038         struct lmv_obd             *lmv = &obd->u.lmv;
1039         struct lprocfs_static_vars  lvars;
1040         struct lmv_desc            *desc;
1041         int                         rc;
1042         int                         i = 0;
1043         ENTRY;
1044
1045         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1046                 CERROR("LMV setup requires a descriptor\n");
1047                 RETURN(-EINVAL);
1048         }
1049
1050         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1051         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1052                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1053                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1054                 RETURN(-EINVAL);
1055         }
1056
1057         lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1058
1059         OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1060         if (lmv->tgts == NULL)
1061                 RETURN(-ENOMEM);
1062
1063         for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1064                 sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
1065                 lmv->tgts[i].ltd_idx = i;
1066         }
1067
1068         lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1069
1070         OBD_ALLOC(lmv->datas, lmv->datas_size);
1071         if (lmv->datas == NULL)
1072                 GOTO(out_free_tgts, rc = -ENOMEM);
1073
1074         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1075         lmv->desc.ld_tgt_count = 0;
1076         lmv->desc.ld_active_tgt_count = 0;
1077         lmv->max_cookiesize = 0;
1078         lmv->max_def_easize = 0;
1079         lmv->max_easize = 0;
1080         lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1081
1082         spin_lock_init(&lmv->lmv_lock);
1083         sema_init(&lmv->init_sem, 1);
1084
1085         rc = lmv_object_setup(obd);
1086         if (rc) {
1087                 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1088                 GOTO(out_free_datas, rc);
1089         }
1090
1091         lprocfs_lmv_init_vars(&lvars);
1092         lprocfs_obd_setup(obd, lvars.obd_vars);
1093 #ifdef LPROCFS
1094         {
1095                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1096                                         0444, &lmv_proc_target_fops, obd);
1097                 if (rc)
1098                         CWARN("Error adding target_obd_stats file (%d)\n", rc);
1099        }
1100 #endif
1101         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1102                              LUSTRE_CLI_FLD_HASH_DHT);
1103         if (rc) {
1104                 CERROR("Can't init FLD, err %d\n", rc);
1105                 GOTO(out_free_datas, rc);
1106         }
1107
1108         RETURN(0);
1109
1110 out_free_datas:
1111         OBD_FREE(lmv->datas, lmv->datas_size);
1112         lmv->datas = NULL;
1113 out_free_tgts:
1114         OBD_FREE(lmv->tgts, lmv->tgts_size);
1115         lmv->tgts = NULL;
1116         return rc;
1117 }
1118
1119 static int lmv_cleanup(struct obd_device *obd)
1120 {
1121         struct lmv_obd   *lmv = &obd->u.lmv;
1122         ENTRY;
1123
1124         fld_client_fini(&lmv->lmv_fld);
1125         lprocfs_obd_cleanup(obd);
1126         lmv_object_cleanup(obd);
1127         OBD_FREE(lmv->datas, lmv->datas_size);
1128         OBD_FREE(lmv->tgts, lmv->tgts_size);
1129
1130         RETURN(0);
1131 }
1132
1133 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1134 {
1135         struct lustre_cfg     *lcfg = buf;
1136         struct obd_uuid        tgt_uuid;
1137         int                    rc;
1138         ENTRY;
1139
1140         switch(lcfg->lcfg_command) {
1141         case LCFG_ADD_MDC:
1142                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1143                         GOTO(out, rc = -EINVAL);
1144
1145                 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1146                 rc = lmv_add_target(obd, &tgt_uuid);
1147                 GOTO(out, rc);
1148         default: {
1149                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1150                 GOTO(out, rc = -EINVAL);
1151         }
1152         }
1153 out:
1154         RETURN(rc);
1155 }
1156
1157 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1158                       __u64 max_age, __u32 flags)
1159 {
1160         struct lmv_obd        *lmv = &obd->u.lmv;
1161         struct obd_statfs     *temp;
1162         int                    rc = 0;
1163         int                    i;
1164         ENTRY;
1165
1166         rc = lmv_check_connect(obd);
1167         if (rc)
1168                 RETURN(rc);
1169
1170         OBD_ALLOC(temp, sizeof(*temp));
1171         if (temp == NULL)
1172                 RETURN(-ENOMEM);
1173
1174         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1175                 if (lmv->tgts[i].ltd_exp == NULL)
1176                         continue;
1177
1178                 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1179                                 max_age, flags);
1180                 if (rc) {
1181                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1182                                lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1183                                rc);
1184                         GOTO(out_free_temp, rc);
1185                 }
1186                 if (i == 0) {
1187                         *osfs = *temp;
1188                 } else {
1189                         osfs->os_bavail += temp->os_bavail;
1190                         osfs->os_blocks += temp->os_blocks;
1191                         osfs->os_ffree += temp->os_ffree;
1192                         osfs->os_files += temp->os_files;
1193                 }
1194         }
1195
1196         EXIT;
1197 out_free_temp:
1198         OBD_FREE(temp, sizeof(*temp));
1199         return rc;
1200 }
1201
1202 static int lmv_getstatus(struct obd_export *exp,
1203                          struct lu_fid *fid,
1204                          struct obd_capa **pc)
1205 {
1206         struct obd_device    *obd = exp->exp_obd;
1207         struct lmv_obd       *lmv = &obd->u.lmv;
1208         int                   rc;
1209         ENTRY;
1210
1211         rc = lmv_check_connect(obd);
1212         if (rc)
1213                 RETURN(rc);
1214
1215         rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1216         RETURN(rc);
1217 }
1218
1219 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1220                         struct obd_capa *oc, obd_valid valid, const char *name,
1221                         const char *input, int input_size, int output_size,
1222                         int flags, struct ptlrpc_request **request)
1223 {
1224         struct obd_device      *obd = exp->exp_obd;
1225         struct lmv_obd         *lmv = &obd->u.lmv;
1226         struct lmv_tgt_desc    *tgt;
1227         int                     rc;
1228         ENTRY;
1229
1230         rc = lmv_check_connect(obd);
1231         if (rc)
1232                 RETURN(rc);
1233
1234         tgt = lmv_find_target(lmv, fid);
1235         if (IS_ERR(tgt))
1236                 RETURN(PTR_ERR(tgt));
1237
1238         rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1239                          input_size, output_size, flags, request);
1240
1241         RETURN(rc);
1242 }
1243
1244 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1245                         struct obd_capa *oc, obd_valid valid, const char *name,
1246                         const char *input, int input_size, int output_size,
1247                         int flags, __u32 suppgid,
1248                         struct ptlrpc_request **request)
1249 {
1250         struct obd_device      *obd = exp->exp_obd;
1251         struct lmv_obd         *lmv = &obd->u.lmv;
1252         struct lmv_tgt_desc    *tgt;
1253         int                     rc;
1254         ENTRY;
1255
1256         rc = lmv_check_connect(obd);
1257         if (rc)
1258                 RETURN(rc);
1259
1260         tgt = lmv_find_target(lmv, fid);
1261         if (IS_ERR(tgt))
1262                 RETURN(PTR_ERR(tgt));
1263
1264         rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1265                          input_size, output_size, flags, suppgid,
1266                          request);
1267
1268         RETURN(rc);
1269 }
1270
1271 static int lmv_getattr(struct obd_export *exp, const struct lu_fid *fid,
1272                        struct obd_capa *oc, obd_valid valid, int ea_size,
1273                        struct ptlrpc_request **request)
1274 {
1275         struct obd_device       *obd = exp->exp_obd;
1276         struct lmv_obd          *lmv = &obd->u.lmv;
1277         struct lmv_tgt_desc     *tgt;
1278         struct lmv_object       *obj;
1279         int                      rc;
1280         int                      i;
1281         ENTRY;
1282
1283         rc = lmv_check_connect(obd);
1284         if (rc)
1285                 RETURN(rc);
1286
1287         tgt = lmv_find_target(lmv, fid);
1288         if (IS_ERR(tgt))
1289                 RETURN(PTR_ERR(tgt));
1290
1291         rc = md_getattr(tgt->ltd_exp, fid, oc, valid, ea_size, request);
1292         if (rc)
1293                 RETURN(rc);
1294
1295         obj = lmv_object_find_lock(obd, fid);
1296
1297         CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(fid),
1298                obj ? "(split)" : "");
1299
1300         /*
1301          * If object is split, then we loop over all the slaves and gather size
1302          * attribute. In ideal world we would have to gather also mds field from
1303          * all slaves, as object is spread over the cluster and this is
1304          * definitely interesting information and it is not good to loss it,
1305          * but...
1306          */
1307         if (obj) {
1308                 struct mdt_body *body;
1309
1310                 if (*request == NULL) {
1311                         lmv_object_put(obj);
1312                         RETURN(rc);
1313                 }
1314
1315                 body = req_capsule_server_get(&(*request)->rq_pill,
1316                                               &RMF_MDT_BODY);
1317                 LASSERT(body != NULL);
1318
1319                 for (i = 0; i < obj->lo_objcount; i++) {
1320                         if (lmv->tgts[i].ltd_exp == NULL) {
1321                                 CWARN("%s: NULL export for %d\n",
1322                                       obd->obd_name, i);
1323                                 continue;
1324                         }
1325
1326                         /*
1327                          * Skip master object.
1328                          */
1329                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1330                                 continue;
1331
1332                         body->size += obj->lo_stripes[i].ls_size;
1333                 }
1334
1335                 lmv_object_put_unlock(obj);
1336         }
1337
1338         RETURN(rc);
1339 }
1340
1341 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1342                              ldlm_iterator_t it, void *data)
1343 {
1344         struct obd_device   *obd = exp->exp_obd;
1345         struct lmv_obd      *lmv = &obd->u.lmv;
1346         int                  i;
1347         int                  rc;
1348         ENTRY;
1349
1350         rc = lmv_check_connect(obd);
1351         if (rc)
1352                 RETURN(rc);
1353
1354         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1355
1356         /*
1357          * With CMD every object can have two locks in different namespaces:
1358          * lookup lock in space of mds storing direntry and update/open lock in
1359          * space of mds storing inode.
1360          */
1361         for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1362                 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1363
1364         RETURN(0);
1365 }
1366
1367 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1368                      struct md_open_data *mod, struct ptlrpc_request **request)
1369 {
1370         struct obd_device     *obd = exp->exp_obd;
1371         struct lmv_obd        *lmv = &obd->u.lmv;
1372         struct lmv_tgt_desc   *tgt;
1373         int                    rc;
1374         ENTRY;
1375
1376         rc = lmv_check_connect(obd);
1377         if (rc)
1378                 RETURN(rc);
1379
1380         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1381         if (IS_ERR(tgt))
1382                 RETURN(PTR_ERR(tgt));
1383
1384         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1385         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1386         RETURN(rc);
1387 }
1388
1389 /**
1390  * Called in the case MDS returns -ERESTART on create on open, what means that
1391  * directory is split and its LMV presentation object has to be updated.
1392  */
1393 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1394 {
1395         struct obd_device       *obd = exp->exp_obd;
1396         struct lmv_obd          *lmv = &obd->u.lmv;
1397         struct ptlrpc_request   *req = NULL;
1398         struct lmv_tgt_desc     *tgt;
1399         struct lmv_object       *obj;
1400         struct lustre_md         md;
1401         int                      mealen;
1402         int                      rc;
1403         __u64                    valid;
1404         ENTRY;
1405
1406         md.mea = NULL;
1407         mealen = lmv_get_easize(lmv);
1408
1409         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1410
1411         tgt = lmv_find_target(lmv, fid);
1412         if (IS_ERR(tgt))
1413                 RETURN(PTR_ERR(tgt));
1414
1415         /*
1416          * Time to update mea of parent fid.
1417          */
1418         rc = md_getattr(tgt->ltd_exp, fid, NULL, valid, mealen, &req);
1419         if (rc) {
1420                 CERROR("md_getattr() failed, error %d\n", rc);
1421                 GOTO(cleanup, rc);
1422         }
1423
1424         rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1425         if (rc) {
1426                 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1427                 GOTO(cleanup, rc);
1428         }
1429
1430         if (md.mea == NULL)
1431                 GOTO(cleanup, rc = -ENODATA);
1432
1433         obj = lmv_object_create(exp, fid, md.mea);
1434         if (IS_ERR(obj))
1435                 rc = PTR_ERR(obj);
1436         else
1437                 lmv_object_put(obj);
1438
1439         obd_free_memmd(exp, (void *)&md.mea);
1440         EXIT;
1441 cleanup:
1442         if (req)
1443                 ptlrpc_req_finished(req);
1444         return rc;
1445 }
1446
1447 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1448                const void *data, int datalen, int mode, __u32 uid,
1449                __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1450                struct ptlrpc_request **request)
1451 {
1452         struct obd_device       *obd = exp->exp_obd;
1453         struct lmv_obd          *lmv = &obd->u.lmv;
1454         struct lmv_tgt_desc     *tgt;
1455         struct lmv_object       *obj;
1456         int                      rc;
1457         int                      loop = 0;
1458         int                      sidx;
1459         ENTRY;
1460
1461         rc = lmv_check_connect(obd);
1462         if (rc)
1463                 RETURN(rc);
1464
1465         if (!lmv->desc.ld_active_tgt_count)
1466                 RETURN(-EIO);
1467 repeat:
1468         ++loop;
1469         LASSERT(loop <= 2);
1470
1471         obj = lmv_object_find(obd, &op_data->op_fid1);
1472         if (obj) {
1473                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1474                                        op_data->op_name, op_data->op_namelen);
1475                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1476                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1477                 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1478                 tgt = lmv_get_target(lmv, op_data->op_mds);
1479                 lmv_object_put(obj);
1480         } else {
1481                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1482                 op_data->op_bias |= MDS_CHECK_SPLIT;
1483                 op_data->op_mds = tgt->ltd_idx;
1484         }
1485
1486         if (IS_ERR(tgt))
1487                 RETURN(PTR_ERR(tgt));
1488
1489         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1490         if (rc == -ERESTART)
1491                 goto repeat;
1492         else if (rc)
1493                 RETURN(rc);
1494
1495         CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #%x\n",
1496                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1497                op_data->op_mds);
1498
1499         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1500         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1501                        cap_effective, rdev, request);
1502         if (rc == 0) {
1503                 if (*request == NULL)
1504                         RETURN(rc);
1505                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1506         } else if (rc == -ERESTART) {
1507                 LASSERT(*request != NULL);
1508                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1509                           "Got -ERESTART during create!\n");
1510                 ptlrpc_req_finished(*request);
1511                 *request = NULL;
1512
1513                 /*
1514                  * Directory got split. Time to update local object and repeat
1515                  * the request with proper MDS.
1516                  */
1517                 rc = lmv_handle_split(exp, &op_data->op_fid1);
1518                 if (rc == 0) {
1519                         rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1520                                                  op_data, &op_data->op_fid2);
1521                         if (rc)
1522                                 RETURN(rc);
1523                         goto repeat;
1524                 }
1525         }
1526         RETURN(rc);
1527 }
1528
1529 static int lmv_done_writing(struct obd_export *exp,
1530                             struct md_op_data *op_data,
1531                             struct md_open_data *mod)
1532 {
1533         struct obd_device     *obd = exp->exp_obd;
1534         struct lmv_obd        *lmv = &obd->u.lmv;
1535         struct lmv_tgt_desc   *tgt;
1536         int                    rc;
1537         ENTRY;
1538
1539         rc = lmv_check_connect(obd);
1540         if (rc)
1541                 RETURN(rc);
1542
1543         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1544         if (IS_ERR(tgt))
1545                 RETURN(PTR_ERR(tgt));
1546
1547         rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1548         RETURN(rc);
1549 }
1550
1551 static int
1552 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1553                    struct lookup_intent *it, struct md_op_data *op_data,
1554                    struct lustre_handle *lockh, void *lmm, int lmmsize)
1555 {
1556         struct obd_device     *obd = exp->exp_obd;
1557         struct lmv_obd        *lmv = &obd->u.lmv;
1558         struct lmv_stripe_md  *mea = op_data->op_mea1;
1559         struct md_op_data     *op_data2;
1560         struct lmv_tgt_desc   *tgt;
1561         int                    i;
1562         int                    rc = 0;
1563         ENTRY;
1564
1565         OBD_ALLOC_PTR(op_data2);
1566         if (op_data2 == NULL)
1567                 RETURN(-ENOMEM);
1568
1569         LASSERT(mea != NULL);
1570         for (i = 0; i < mea->mea_count; i++) {
1571                 memset(op_data2, 0, sizeof(*op_data2));
1572                 op_data2->op_fid1 = mea->mea_ids[i];
1573                 op_data2->op_bias = 0;
1574
1575                 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1576                 if (IS_ERR(tgt))
1577                         GOTO(cleanup, rc = PTR_ERR(tgt));
1578
1579                 if (tgt->ltd_exp == NULL)
1580                         continue;
1581
1582                 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1583                                 lockh + i, lmm, lmmsize, NULL, 0);
1584
1585                 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1586                        PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1587
1588                 if (rc)
1589                         GOTO(cleanup, rc);
1590
1591                 if (it->d.lustre.it_data) {
1592                         struct ptlrpc_request *req;
1593                         req = (struct ptlrpc_request *)it->d.lustre.it_data;
1594                         ptlrpc_req_finished(req);
1595                 }
1596
1597                 if (it->d.lustre.it_status)
1598                         GOTO(cleanup, rc = it->d.lustre.it_status);
1599         }
1600
1601         EXIT;
1602 cleanup:
1603         OBD_FREE_PTR(op_data2);
1604
1605         if (rc != 0) {
1606                 /*
1607                  * Drop all taken locks.
1608                  */
1609                 while (--i >= 0) {
1610                         if (lockh[i].cookie)
1611                                 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1612                         lockh[i].cookie = 0;
1613                 }
1614         }
1615         return rc;
1616 }
1617
1618 static int
1619 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1620                    struct lookup_intent *it, struct md_op_data *op_data,
1621                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1622                    int extra_lock_flags)
1623 {
1624         struct ptlrpc_request      *req = it->d.lustre.it_data;
1625         struct obd_device          *obd = exp->exp_obd;
1626         struct lmv_obd             *lmv = &obd->u.lmv;
1627         struct lustre_handle        plock;
1628         struct lmv_tgt_desc        *tgt;
1629         struct md_op_data          *rdata;
1630         struct lu_fid               fid1;
1631         struct mdt_body            *body;
1632         int                         rc = 0;
1633         int                         pmode;
1634         ENTRY;
1635
1636         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1637         LASSERT(body != NULL);
1638
1639         if (!(body->valid & OBD_MD_MDS))
1640                 RETURN(0);
1641
1642         CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1643                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1644
1645         /*
1646          * We got LOOKUP lock, but we really need attrs.
1647          */
1648         pmode = it->d.lustre.it_lock_mode;
1649         LASSERT(pmode != 0);
1650         memcpy(&plock, lockh, sizeof(plock));
1651         it->d.lustre.it_lock_mode = 0;
1652         it->d.lustre.it_data = NULL;
1653         fid1 = body->fid1;
1654
1655         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1656         ptlrpc_req_finished(req);
1657
1658         tgt = lmv_find_target(lmv, &fid1);
1659         if (IS_ERR(tgt))
1660                 GOTO(out, rc = PTR_ERR(tgt));
1661
1662         OBD_ALLOC_PTR(rdata);
1663         if (rdata == NULL)
1664                 GOTO(out, rc = -ENOMEM);
1665
1666         rdata->op_fid1 = fid1;
1667         rdata->op_bias = MDS_CROSS_REF;
1668
1669         rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1670                         lmm, lmmsize, NULL, extra_lock_flags);
1671         OBD_FREE_PTR(rdata);
1672         EXIT;
1673 out:
1674         ldlm_lock_decref(&plock, pmode);
1675         return rc;
1676 }
1677
1678 static int
1679 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1680             struct lookup_intent *it, struct md_op_data *op_data,
1681             struct lustre_handle *lockh, void *lmm, int lmmsize,
1682             struct ptlrpc_request **req, int extra_lock_flags)
1683 {
1684         struct obd_device        *obd = exp->exp_obd;
1685         struct lmv_obd           *lmv = &obd->u.lmv;
1686         struct lmv_tgt_desc      *tgt;
1687         struct lmv_object        *obj;
1688         int                       sidx;
1689         int                       rc;
1690         ENTRY;
1691
1692         rc = lmv_check_connect(obd);
1693         if (rc)
1694                 RETURN(rc);
1695
1696         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1697                LL_IT2STR(it), PFID(&op_data->op_fid1));
1698
1699         if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1700                 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1701                                         lockh, lmm, lmmsize);
1702                 RETURN(rc);
1703         }
1704
1705         obj = lmv_object_find(obd, &op_data->op_fid1);
1706         if (obj && op_data->op_namelen) {
1707                 sidx = raw_name2idx(obj->lo_hashtype,
1708                                        obj->lo_objcount,
1709                                        (char *)op_data->op_name,
1710                                        op_data->op_namelen);
1711                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1712                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1713         } else {
1714                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1715         }
1716         if (obj)
1717                 lmv_object_put(obj);
1718
1719         if (IS_ERR(tgt))
1720                 RETURN(PTR_ERR(tgt));
1721
1722         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1723                LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1724
1725         rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1726                         lmm, lmmsize, req, extra_lock_flags);
1727
1728         if (rc == 0 && it && it->it_op == IT_OPEN) {
1729                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1730                                         lmm, lmmsize, extra_lock_flags);
1731         }
1732         RETURN(rc);
1733 }
1734
1735 static int
1736 lmv_getattr_name(struct obd_export *exp, const struct lu_fid *fid,
1737                  struct obd_capa *oc, const char *name, int namelen,
1738                  obd_valid valid, int ea_size, __u32 suppgid,
1739                  struct ptlrpc_request **request)
1740 {
1741         struct ptlrpc_request   *req = NULL;
1742         struct obd_device       *obd = exp->exp_obd;
1743         struct lmv_obd          *lmv = &obd->u.lmv;
1744         struct lu_fid            rid = *fid;
1745         struct lmv_tgt_desc     *tgt;
1746         struct mdt_body         *body;
1747         struct lmv_object       *obj;
1748         int                      rc;
1749         int                      loop = 0;
1750         int                      sidx;
1751         ENTRY;
1752
1753         rc = lmv_check_connect(obd);
1754         if (rc)
1755                 RETURN(rc);
1756
1757 repeat:
1758         ++loop;
1759         LASSERT(loop <= 2);
1760         obj = lmv_object_find(obd, &rid);
1761         if (obj) {
1762                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1763                                        name, namelen - 1);
1764                 rid = obj->lo_stripes[sidx].ls_fid;
1765                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1766                 valid &= ~OBD_MD_FLCKSPLIT;
1767                 lmv_object_put(obj);
1768         } else {
1769                 tgt = lmv_find_target(lmv, &rid);
1770                 valid |= OBD_MD_FLCKSPLIT;
1771         }
1772         if (IS_ERR(tgt))
1773                 RETURN(PTR_ERR(tgt));
1774
1775         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1776                namelen, name, PFID(fid), PFID(&rid), tgt->ltd_idx);
1777
1778         rc = md_getattr_name(tgt->ltd_exp, &rid, oc, name, namelen, valid,
1779                              ea_size, suppgid, request);
1780         if (rc == 0) {
1781                 body = req_capsule_server_get(&(*request)->rq_pill,
1782                                               &RMF_MDT_BODY);
1783                 LASSERT(body != NULL);
1784
1785                 if (body->valid & OBD_MD_MDS) {
1786                         rid = body->fid1;
1787                         CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1788                                PFID(&rid));
1789
1790                         tgt = lmv_find_target(lmv, &rid);
1791                         if (IS_ERR(tgt)) {
1792                                 ptlrpc_req_finished(*request);
1793                                 RETURN(PTR_ERR(tgt));
1794                         }
1795
1796                         rc = md_getattr_name(tgt->ltd_exp, &rid, NULL, NULL,
1797                                              1, valid | OBD_MD_FLCROSSREF,
1798                                              ea_size, suppgid, &req);
1799                         ptlrpc_req_finished(*request);
1800                         *request = req;
1801                 }
1802         } else if (rc == -ERESTART) {
1803                 LASSERT(*request != NULL);
1804                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1805                           "Got -ERESTART during getattr!\n");
1806                 ptlrpc_req_finished(*request);
1807                 *request = NULL;
1808
1809                 /*
1810                  * Directory got split. Time to update local object and repeat
1811                  * the request with proper MDS.
1812                  */
1813                 rc = lmv_handle_split(exp, &rid);
1814                 if (rc == 0)
1815                         goto repeat;
1816         }
1817         RETURN(rc);
1818 }
1819
1820 #define md_op_data_fid(op_data, fl)                     \
1821         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1822          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1823          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1824          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1825          NULL)
1826
1827 static int lmv_early_cancel_slaves(struct obd_export *exp,
1828                                    struct md_op_data *op_data, int op_tgt,
1829                                    ldlm_mode_t mode, int bits, int flag)
1830 {
1831         struct obd_device      *obd = exp->exp_obd;
1832         struct lmv_obd         *lmv = &obd->u.lmv;
1833         ldlm_policy_data_t      policy = {{0}};
1834         struct lu_fid          *op_fid;
1835         struct lu_fid          *st_fid;
1836         struct lmv_tgt_desc    *tgt;
1837         struct lmv_object      *obj;
1838         int                     rc = 0;
1839         int                     i;
1840         ENTRY;
1841
1842         op_fid = md_op_data_fid(op_data, flag);
1843         if (!fid_is_sane(op_fid))
1844                 RETURN(0);
1845
1846         obj = lmv_object_find(obd, op_fid);
1847         if (obj == NULL)
1848                 RETURN(-EALREADY);
1849
1850         policy.l_inodebits.bits = bits;
1851         for (i = 0; i < obj->lo_objcount; i++) {
1852                 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1853                 st_fid = &obj->lo_stripes[i].ls_fid;
1854                 if (op_tgt != tgt->ltd_idx) {
1855                         CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1856                                PFID(st_fid), tgt->ltd_idx);
1857                         rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1858                                               mode, LDLM_FL_ASYNC, NULL);
1859                         if (rc)
1860                                 GOTO(out_put_obj, rc);
1861                 } else {
1862                         CDEBUG(D_INODE,
1863                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1864                                op_tgt, PFID(st_fid));
1865                         /*
1866                          * Do not cancel locks for operation target, they will
1867                          * be handled later in underlaying layer when calling
1868                          * function we run on behalf of.
1869                          */
1870                         *op_fid = *st_fid;
1871                         op_data->op_flags |= flag;
1872                 }
1873         }
1874         EXIT;
1875 out_put_obj:
1876         lmv_object_put(obj);
1877         return rc;
1878 }
1879
1880 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1881                             int op_tgt, ldlm_mode_t mode, int bits, int flag)
1882 {
1883         struct lu_fid          *fid = md_op_data_fid(op_data, flag);
1884         struct obd_device      *obd = exp->exp_obd;
1885         struct lmv_obd         *lmv = &obd->u.lmv;
1886         struct lmv_tgt_desc    *tgt;
1887         ldlm_policy_data_t      policy = {{0}};
1888         struct lmv_object      *obj;
1889         int                     rc = 0;
1890         ENTRY;
1891
1892         if (!fid_is_sane(fid))
1893                 RETURN(0);
1894
1895         obj = lmv_object_find(obd, fid);
1896         if (obj) {
1897                 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1898                                              bits, flag);
1899                 lmv_object_put(obj);
1900         } else {
1901                 tgt = lmv_find_target(lmv, fid);
1902                 if (IS_ERR(tgt))
1903                         RETURN(PTR_ERR(tgt));
1904
1905                 if (tgt->ltd_idx != op_tgt) {
1906                         CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1907                         policy.l_inodebits.bits = bits;
1908                         rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1909                                               mode, LDLM_FL_ASYNC, NULL);
1910                 } else {
1911                         CDEBUG(D_INODE,
1912                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1913                                op_tgt, PFID(fid));
1914                         op_data->op_flags |= flag;
1915                         rc = 0;
1916                 }
1917
1918         }
1919         RETURN(rc);
1920 }
1921
1922 /*
1923  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1924  * op_data->op_fid2
1925  */
1926 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1927                     struct ptlrpc_request **request)
1928 {
1929         struct obd_device       *obd = exp->exp_obd;
1930         struct lmv_obd          *lmv = &obd->u.lmv;
1931         struct lmv_tgt_desc     *tgt;
1932         struct lmv_object       *obj;
1933         int                      rc;
1934         int                      loop = 0;
1935         mdsno_t                  mds;
1936         int                      sidx;
1937         ENTRY;
1938
1939         rc = lmv_check_connect(obd);
1940         if (rc)
1941                 RETURN(rc);
1942
1943 repeat:
1944         ++loop;
1945         LASSERT(loop <= 2);
1946         LASSERT(op_data->op_namelen != 0);
1947
1948         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
1949                PFID(&op_data->op_fid2), op_data->op_namelen,
1950                op_data->op_name, PFID(&op_data->op_fid1));
1951
1952         obj = lmv_object_find(obd, &op_data->op_fid2);
1953         if (obj) {
1954                 sidx = raw_name2idx(obj->lo_hashtype,
1955                                     obj->lo_objcount,
1956                                     op_data->op_name,
1957                                     op_data->op_namelen);
1958                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
1959                 mds = obj->lo_stripes[sidx].ls_mds;
1960                 lmv_object_put(obj);
1961         } else {
1962                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1963                 if (rc)
1964                         RETURN(rc);
1965         }
1966
1967         CDEBUG(D_INODE, "Forward to mds #%x ("DFID")\n",
1968                mds, PFID(&op_data->op_fid1));
1969
1970         op_data->op_fsuid = cfs_curproc_fsuid();
1971         op_data->op_fsgid = cfs_curproc_fsgid();
1972         op_data->op_cap = cfs_curproc_cap_pack();
1973         tgt = lmv_get_target(lmv, mds);
1974
1975         /*
1976          * Cancel UPDATE lock on child (fid1).
1977          */
1978         op_data->op_flags |= MF_MDC_CANCEL_FID2;
1979         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
1980                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1981         if (rc == 0)
1982                 rc = md_link(tgt->ltd_exp, op_data, request);
1983         if (rc == -ERESTART) {
1984                 LASSERT(*request != NULL);
1985                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1986                           "Got -ERESTART during link!\n");
1987                 ptlrpc_req_finished(*request);
1988                 *request = NULL;
1989
1990                 /*
1991                  * Directory got split. Time to update local object and repeat
1992                  * the request with proper MDS.
1993                  */
1994                 rc = lmv_handle_split(exp, &op_data->op_fid2);
1995                 if (rc == 0)
1996                         goto repeat;
1997         }
1998
1999         RETURN(rc);
2000 }
2001
2002 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2003                       const char *old, int oldlen, const char *new, int newlen,
2004                       struct ptlrpc_request **request)
2005 {
2006         struct obd_device       *obd = exp->exp_obd;
2007         struct lmv_obd          *lmv = &obd->u.lmv;
2008         struct lmv_tgt_desc     *src_tgt;
2009         struct lmv_tgt_desc     *tgt_tgt;
2010         int                      rc;
2011         int                      sidx;
2012         int                      loop = 0;
2013         struct lmv_object       *obj;
2014         mdsno_t                  mds1;
2015         mdsno_t                  mds2;
2016         ENTRY;
2017
2018         LASSERT(oldlen != 0);
2019
2020         CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2021                oldlen, old, PFID(&op_data->op_fid1),
2022                newlen, new, PFID(&op_data->op_fid2));
2023
2024         rc = lmv_check_connect(obd);
2025         if (rc)
2026                 RETURN(rc);
2027
2028 repeat:
2029         ++loop;
2030         LASSERT(loop <= 2);
2031         obj = lmv_object_find(obd, &op_data->op_fid1);
2032         if (obj) {
2033                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2034                                     (char *)old, oldlen);
2035                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2036                 mds1 = obj->lo_stripes[sidx].ls_mds;
2037                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2038                 lmv_object_put(obj);
2039         } else {
2040                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2041                 if (rc)
2042                         RETURN(rc);
2043         }
2044
2045         obj = lmv_object_find(obd, &op_data->op_fid2);
2046         if (obj) {
2047                 /*
2048                  * Directory is already split, so we have to forward request to
2049                  * the right MDS.
2050                  */
2051                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2052                                        (char *)new, newlen);
2053
2054                 mds2 = obj->lo_stripes[sidx].ls_mds;
2055                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2056                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2057                 lmv_object_put(obj);
2058         } else {
2059                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2060                 if (rc)
2061                         RETURN(rc);
2062         }
2063
2064         op_data->op_fsuid = cfs_curproc_fsuid();
2065         op_data->op_fsgid = cfs_curproc_fsgid();
2066         op_data->op_cap = cfs_curproc_cap_pack();
2067
2068         src_tgt = lmv_get_target(lmv, mds1);
2069         tgt_tgt = lmv_get_target(lmv, mds2);
2070
2071         /*
2072          * LOOKUP lock on src child (fid3) should also be cancelled for
2073          * src_tgt in mdc_rename.
2074          */
2075         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2076
2077         /*
2078          * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2079          * own target.
2080          */
2081         rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2082                               LCK_EX, MDS_INODELOCK_UPDATE,
2083                               MF_MDC_CANCEL_FID2);
2084
2085         /*
2086          * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2087          */
2088         if (rc == 0) {
2089                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2090                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2091                                       MF_MDC_CANCEL_FID4);
2092         }
2093
2094         /*
2095          * Cancel all the locks on tgt child (fid4).
2096          */
2097         if (rc == 0)
2098                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2099                                       LCK_EX, MDS_INODELOCK_FULL,
2100                                       MF_MDC_CANCEL_FID4);
2101
2102         if (rc == 0)
2103                 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2104                                new, newlen, request);
2105
2106         if (rc == -ERESTART) {
2107                 LASSERT(*request != NULL);
2108                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2109                           "Got -ERESTART during rename!\n");
2110                 ptlrpc_req_finished(*request);
2111                 *request = NULL;
2112
2113                 /*
2114                  * Directory got split. Time to update local object and repeat
2115                  * the request with proper MDS.
2116                  */
2117                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2118                 if (rc == 0)
2119                         goto repeat;
2120         }
2121         RETURN(rc);
2122 }
2123
2124 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2125                        void *ea, int ealen, void *ea2, int ea2len,
2126                        struct ptlrpc_request **request,
2127                        struct md_open_data **mod)
2128 {
2129         struct obd_device       *obd = exp->exp_obd;
2130         struct lmv_obd          *lmv = &obd->u.lmv;
2131         struct ptlrpc_request   *req;
2132         struct lmv_tgt_desc     *tgt;
2133         struct lmv_object       *obj;
2134         int                      rc = 0;
2135         int                      i;
2136         ENTRY;
2137
2138         rc = lmv_check_connect(obd);
2139         if (rc)
2140                 RETURN(rc);
2141
2142         obj = lmv_object_find(obd, &op_data->op_fid1);
2143
2144         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2145                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2146                obj ? ", split" : "");
2147
2148         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2149         if (obj) {
2150                 for (i = 0; i < obj->lo_objcount; i++) {
2151                         op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2152
2153                         tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2154                         if (IS_ERR(tgt)) {
2155                                 rc = PTR_ERR(tgt);
2156                                 break;
2157                         }
2158
2159                         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2160                                         ea2, ea2len, &req, mod);
2161
2162                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2163                                 /*
2164                                  * This is master object and this request should
2165                                  * be returned back to llite.
2166                                  */
2167                                 *request = req;
2168                         } else {
2169                                 ptlrpc_req_finished(req);
2170                         }
2171
2172                         if (rc)
2173                                 break;
2174                 }
2175                 lmv_object_put(obj);
2176         } else {
2177                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2178                 if (IS_ERR(tgt))
2179                         RETURN(PTR_ERR(tgt));
2180
2181                 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2182                                 ea2len, request, mod);
2183         }
2184         RETURN(rc);
2185 }
2186
2187 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2188                     struct obd_capa *oc, struct ptlrpc_request **request)
2189 {
2190         struct obd_device         *obd = exp->exp_obd;
2191         struct lmv_obd            *lmv = &obd->u.lmv;
2192         struct lmv_tgt_desc       *tgt;
2193         int                        rc;
2194         ENTRY;
2195
2196         rc = lmv_check_connect(obd);
2197         if (rc)
2198                 RETURN(rc);
2199
2200         tgt = lmv_find_target(lmv, fid);
2201         if (IS_ERR(tgt))
2202                 RETURN(PTR_ERR(tgt));
2203
2204         rc = md_sync(tgt->ltd_exp, fid, oc, request);
2205         RETURN(rc);
2206 }
2207
2208 /**
2209  * Main purpose of LMV blocking ast is to remove split directory LMV
2210  * presentation object (struct lmv_object) attached to the lock being revoked.
2211  */
2212 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2213                      void *data, int flag)
2214 {
2215         struct lustre_handle    lockh;
2216         struct lmv_object      *obj;
2217         int                     rc;
2218         ENTRY;
2219
2220         switch (flag) {
2221         case LDLM_CB_BLOCKING:
2222                 ldlm_lock2handle(lock, &lockh);
2223                 rc = ldlm_cli_cancel(&lockh);
2224                 if (rc < 0) {
2225                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2226                         RETURN(rc);
2227                 }
2228                 break;
2229         case LDLM_CB_CANCELING:
2230                 /*
2231                  * Time to drop cached attrs for split directory object
2232                  */
2233                 obj = lock->l_ast_data;
2234                 if (obj) {
2235                         CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2236                                ", master "DFID"\n",
2237                                lock->l_resource->lr_name.name[3] == 1 ?
2238                                "LOOKUP" : "UPDATE",
2239                                lock->l_resource->lr_name.name[0],
2240                                lock->l_resource->lr_name.name[1],
2241                                PFID(&obj->lo_fid));
2242                         lmv_object_put(obj);
2243                 }
2244                 break;
2245         default:
2246                 LBUG();
2247         }
2248         RETURN(0);
2249 }
2250
2251 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2252 {
2253         __u64         val;
2254
2255         val = le64_to_cpu(*hash);
2256         if (val < hash_adj)
2257                 val += MAX_HASH_SIZE;
2258         if (val != DIR_END_OFF)
2259                 *hash = cpu_to_le64(val - hash_adj);
2260 }
2261
2262 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2263 {
2264         __u64              id;
2265         struct obd_import *imp;
2266
2267         /*
2268          * XXX: to get nid we assume that underlying obd device is mdc.
2269          */
2270         imp  = class_exp2cliimp(exp);
2271         id   = imp->imp_connection->c_self + fid_flatten(fid);
2272
2273         CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2274                imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2275
2276         return id ^ (id >> 32);
2277 }
2278
2279 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2280                         struct obd_capa *oc, __u64 offset64, struct page *page,
2281                         struct ptlrpc_request **request)
2282 {
2283         struct obd_device       *obd = exp->exp_obd;
2284         struct lmv_obd          *lmv = &obd->u.lmv;
2285         struct lu_fid            rid = *fid;
2286         struct lmv_object       *obj;
2287         __u64                    offset;
2288         __u64                    hash_adj = 0;
2289         __u32                    rank = 0;
2290         __u64                    seg_size = 0;
2291         __u64                    tgt_tmp = 0;
2292         int                      tgt_idx = 0;
2293         int                      tgt0_idx = 0;
2294         int                      rc;
2295         int                      nr = 0;
2296         struct lmv_stripe       *los;
2297         struct lmv_tgt_desc     *tgt;
2298         struct lu_dirpage       *dp;
2299         struct lu_dirent        *ent;
2300         ENTRY;
2301
2302         offset = offset64;
2303
2304         rc = lmv_check_connect(obd);
2305         if (rc)
2306                 RETURN(rc);
2307
2308         CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2309
2310         /*
2311          * This case handle directory lookup in clustered metadata case (i.e.
2312          * split directory is located on multiple md servers.)
2313          * each server keeps directory entries for certain range of hashes.
2314          * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2315          * first server will keep records with hashes [ 0 ... MAX_HASH / N  - 1],
2316          * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2317          * so on....
2318          *      readdir can simply start reading entries from 0 - N server in
2319          * order but that will not scale well as all client will request dir in
2320          * to server in same order.
2321          * Following algorithm does optimization:
2322          * Instead of doing readdir in 1, 2, ...., N order, client with a
2323          * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2324          * (every client has rank R)
2325          *      But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2326          * since client ask dir from MDS{R} client has pages with offsets
2327          * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2328          * on hash  values that we get.
2329          */
2330         obj = lmv_object_find_lock(obd, fid);
2331         if (obj) {
2332                 nr       = obj->lo_objcount;
2333                 LASSERT(nr > 0);
2334                 seg_size = MAX_HASH_SIZE;
2335                 do_div(seg_size, nr);
2336                 los      = obj->lo_stripes;
2337                 tgt      = lmv_get_target(lmv, los[0].ls_mds);
2338                 rank     = lmv_node_rank(tgt->ltd_exp, fid) % nr;
2339                 tgt_tmp  = offset;
2340                 do_div(tgt_tmp, seg_size);
2341                 tgt0_idx = do_div(tgt_tmp,  nr);
2342                 tgt_idx  = (tgt0_idx + rank) % nr;
2343
2344                 if (tgt_idx < tgt0_idx)
2345                         /*
2346                          * Wrap around.
2347                          *
2348                          * Last segment has unusual length due to division
2349                          * rounding.
2350                          */
2351                         hash_adj = MAX_HASH_SIZE - seg_size * nr;
2352                 else
2353                         hash_adj = 0;
2354
2355                 hash_adj += rank * seg_size;
2356
2357                 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2358                        LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2359                        offset, tgt0_idx, offset + hash_adj, tgt_idx);
2360
2361                 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2362                 rid = obj->lo_stripes[tgt_idx].ls_fid;
2363                 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2364
2365                 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2366                        PFID(&rid), (unsigned long)offset, tgt_idx);
2367         } else
2368                 tgt = lmv_find_target(lmv, &rid);
2369
2370         if (IS_ERR(tgt))
2371                 GOTO(cleanup, rc = PTR_ERR(tgt));
2372
2373         rc = md_readpage(tgt->ltd_exp, &rid, oc, offset, page, request);
2374         if (rc)
2375                 GOTO(cleanup, rc);
2376         if (obj) {
2377                 dp = cfs_kmap(page);
2378
2379                 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2380                 lmv_hash_adjust(&dp->ldp_hash_end,   hash_adj);
2381                 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2382
2383                 for (ent = lu_dirent_start(dp); ent != NULL;
2384                      ent = lu_dirent_next(ent))
2385                         lmv_hash_adjust(&ent->lde_hash, hash_adj);
2386
2387                 if (tgt0_idx != nr - 1) {
2388                         __u64 end;
2389
2390                         end = le64_to_cpu(dp->ldp_hash_end);
2391                         if (end == DIR_END_OFF) {
2392                                 dp->ldp_hash_end = cpu_to_le32(seg_size *
2393                                                                (tgt0_idx + 1));
2394                                 CDEBUG(D_INODE,
2395                                        ""DFID" reset end "LPX64" tgt %d\n",
2396                                        PFID(&rid),
2397                                        le64_to_cpu(dp->ldp_hash_end), tgt_idx);
2398                         }
2399                 }
2400                 cfs_kunmap(page);
2401         }
2402         EXIT;
2403 cleanup:
2404         if (obj)
2405                 lmv_object_put_unlock(obj);
2406         return rc;
2407 }
2408
2409 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2410                       struct ptlrpc_request **request)
2411 {
2412         struct obd_device       *obd = exp->exp_obd;
2413         struct lmv_obd          *lmv = &obd->u.lmv;
2414         struct lmv_tgt_desc     *tgt = NULL;
2415         struct lmv_object       *obj;
2416         int                      rc;
2417         int                      sidx;
2418         int                      loop = 0;
2419         ENTRY;
2420
2421         rc = lmv_check_connect(obd);
2422         if (rc)
2423                 RETURN(rc);
2424
2425 repeat:
2426         ++loop;
2427         LASSERT(loop <= 2);
2428         LASSERT(op_data->op_namelen != 0);
2429
2430         obj = lmv_object_find(obd, &op_data->op_fid1);
2431         if (obj) {
2432                 sidx = raw_name2idx(obj->lo_hashtype,
2433                                     obj->lo_objcount,
2434                                     op_data->op_name,
2435                                     op_data->op_namelen);
2436                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2437                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2438                 tgt = lmv_get_target(lmv,
2439                                      obj->lo_stripes[sidx].ls_mds);
2440                 lmv_object_put(obj);
2441                 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2442                        op_data->op_namelen, op_data->op_name,
2443                        PFID(&op_data->op_fid1), sidx);
2444         }
2445
2446         if (tgt == NULL) {
2447                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2448                 if (IS_ERR(tgt))
2449                         RETURN(PTR_ERR(tgt));
2450                 op_data->op_bias |= MDS_CHECK_SPLIT;
2451         }
2452
2453         op_data->op_fsuid = cfs_curproc_fsuid();
2454         op_data->op_fsgid = cfs_curproc_fsgid();
2455         op_data->op_cap = cfs_curproc_cap_pack();
2456
2457         /*
2458          * If child's fid is given, cancel unused locks for it if it is from
2459          * another export than parent.
2460          *
2461          * LOOKUP lock for child (fid3) should also be cancelled on parent
2462          * tgt_tgt in mdc_unlink().
2463          */
2464         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2465
2466         /*
2467          * Cancel FULL locks on child (fid3).
2468          */
2469         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2470                               MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2471
2472         if (rc == 0)
2473                 rc = md_unlink(tgt->ltd_exp, op_data, request);
2474
2475         if (rc == -ERESTART) {
2476                 LASSERT(*request != NULL);
2477                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2478                           "Got -ERESTART during unlink!\n");
2479                 ptlrpc_req_finished(*request);
2480                 *request = NULL;
2481
2482                 /*
2483                  * Directory got split. Time to update local object and repeat
2484                  * the request with proper MDS.
2485                  */
2486                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2487                 if (rc == 0)
2488                         goto repeat;
2489         }
2490         RETURN(rc);
2491 }
2492
2493 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2494 {
2495         int        rc = 0;
2496
2497         switch (stage) {
2498         case OBD_CLEANUP_EARLY:
2499                 /* XXX: here should be calling obd_precleanup() down to
2500                  * stack. */
2501                 break;
2502         case OBD_CLEANUP_EXPORTS:
2503                 rc = obd_llog_finish(obd, 0);
2504                 if (rc != 0)
2505                         CERROR("failed to cleanup llogging subsystems\n");
2506                 break;
2507         default:
2508                 break;
2509         }
2510         RETURN(rc);
2511 }
2512
2513 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2514                         void *key, __u32 *vallen, void *val,
2515                         struct lov_stripe_md *lsm)
2516 {
2517         struct obd_device       *obd;
2518         struct lmv_obd          *lmv;
2519         int                      rc = 0;
2520         ENTRY;
2521
2522         obd = class_exp2obd(exp);
2523         if (obd == NULL) {
2524                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2525                        exp->exp_handle.h_cookie);
2526                 RETURN(-EINVAL);
2527         }
2528
2529         lmv = &obd->u.lmv;
2530         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2531                 struct lmv_tgt_desc *tgts;
2532                 int i;
2533
2534                 rc = lmv_check_connect(obd);
2535                 if (rc)
2536                         RETURN(rc);
2537
2538                 LASSERT(*vallen == sizeof(__u32));
2539                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2540                      i++, tgts++) {
2541
2542                         /*
2543                          * All tgts should be connected when this gets called.
2544                          */
2545                         if (!tgts || !tgts->ltd_exp) {
2546                                 CERROR("target not setup?\n");
2547                                 continue;
2548                         }
2549
2550                         if (!obd_get_info(tgts->ltd_exp, keylen, key,
2551                                           vallen, val, NULL))
2552                                 RETURN(0);
2553                 }
2554                 RETURN(-EINVAL);
2555         } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2556                 rc = lmv_check_connect(obd);
2557                 if (rc)
2558                         RETURN(rc);
2559
2560                 /*
2561                  * Forwarding this request to first MDS, it should know LOV
2562                  * desc.
2563                  */
2564                 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2565                                   vallen, val, NULL);
2566                 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2567                         exp->exp_connect_flags =
2568                         ((struct obd_connect_data *)val)->ocd_connect_flags;
2569                 }
2570                 RETURN(rc);
2571         }
2572
2573         CDEBUG(D_IOCTL, "Invalid key\n");
2574         RETURN(-EINVAL);
2575 }
2576
2577 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2578                        void *key, obd_count vallen, void *val,
2579                        struct ptlrpc_request_set *set)
2580 {
2581         struct lmv_tgt_desc    *tgt;
2582         struct obd_device      *obd;
2583         struct lmv_obd         *lmv;
2584         int rc = 0;
2585         ENTRY;
2586
2587         obd = class_exp2obd(exp);
2588         if (obd == NULL) {
2589                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2590                        exp->exp_handle.h_cookie);
2591                 RETURN(-EINVAL);
2592         }
2593         lmv = &obd->u.lmv;
2594
2595         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2596             KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2597                 int i, err = 0;
2598
2599                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2600                         tgt = &lmv->tgts[i];
2601
2602                         if (!tgt->ltd_exp)
2603                                 continue;
2604
2605                         err = obd_set_info_async(tgt->ltd_exp,
2606                                                  keylen, key, vallen, val, set);
2607                         if (err && rc == 0)
2608                                 rc = err;
2609                 }
2610
2611                 RETURN(rc);
2612         }
2613
2614         RETURN(-EINVAL);
2615 }
2616
2617 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2618                struct lov_stripe_md *lsm)
2619 {
2620         struct obd_device         *obd = class_exp2obd(exp);
2621         struct lmv_obd            *lmv = &obd->u.lmv;
2622         struct lmv_stripe_md      *meap;
2623         struct lmv_stripe_md      *lsmp;
2624         int                        mea_size;
2625         int                        i;
2626         ENTRY;
2627
2628         mea_size = lmv_get_easize(lmv);
2629         if (!lmmp)
2630                 RETURN(mea_size);
2631
2632         if (*lmmp && !lsm) {
2633                 OBD_FREE(*lmmp, mea_size);
2634                 *lmmp = NULL;
2635                 RETURN(0);
2636         }
2637
2638         if (*lmmp == NULL) {
2639                 OBD_ALLOC(*lmmp, mea_size);
2640                 if (*lmmp == NULL)
2641                         RETURN(-ENOMEM);
2642         }
2643
2644         if (!lsm)
2645                 RETURN(mea_size);
2646
2647         lsmp = (struct lmv_stripe_md *)lsm;
2648         meap = (struct lmv_stripe_md *)*lmmp;
2649
2650         if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2651             lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2652                 RETURN(-EINVAL);
2653
2654         meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2655         meap->mea_count = cpu_to_le32(lsmp->mea_count);
2656         meap->mea_master = cpu_to_le32(lsmp->mea_master);
2657
2658         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2659                 meap->mea_ids[i] = meap->mea_ids[i];
2660                 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2661         }
2662
2663         RETURN(mea_size);
2664 }
2665
2666 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2667                  struct lov_mds_md *lmm, int lmm_size)
2668 {
2669         struct obd_device          *obd = class_exp2obd(exp);
2670         struct lmv_stripe_md      **tmea = (struct lmv_stripe_md **)lsmp;
2671         struct lmv_stripe_md       *mea = (struct lmv_stripe_md *)lmm;
2672         struct lmv_obd             *lmv = &obd->u.lmv;
2673         int                         mea_size;
2674         int                         i;
2675         __u32                       magic;
2676         ENTRY;
2677
2678         mea_size = lmv_get_easize(lmv);
2679         if (lsmp == NULL)
2680                 return mea_size;
2681
2682         if (*lsmp != NULL && lmm == NULL) {
2683                 OBD_FREE(*tmea, mea_size);
2684                 *lsmp = NULL;
2685                 RETURN(0);
2686         }
2687
2688         LASSERT(mea_size == lmm_size);
2689
2690         OBD_ALLOC(*tmea, mea_size);
2691         if (*tmea == NULL)
2692                 RETURN(-ENOMEM);
2693
2694         if (!lmm)
2695                 RETURN(mea_size);
2696
2697         if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2698             mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2699             mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2700         {
2701                 magic = le32_to_cpu(mea->mea_magic);
2702         } else {
2703                 /*
2704                  * Old mea is not handled here.
2705                  */
2706                 CERROR("Old not supportable EA is found\n");
2707                 LBUG();
2708         }
2709
2710         (*tmea)->mea_magic = magic;
2711         (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2712         (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2713
2714         for (i = 0; i < (*tmea)->mea_count; i++) {
2715                 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2716                 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2717         }
2718         RETURN(mea_size);
2719 }
2720
2721 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2722                              ldlm_policy_data_t *policy, ldlm_mode_t mode,
2723                              int flags, void *opaque)
2724 {
2725         struct obd_device       *obd = exp->exp_obd;
2726         struct lmv_obd          *lmv = &obd->u.lmv;
2727         int                      rc = 0;
2728         int                      err;
2729         int                      i;
2730         ENTRY;
2731
2732         LASSERT(fid != NULL);
2733
2734         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2735                 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2736                         continue;
2737
2738                 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2739                                        policy, mode, flags, opaque);
2740                 if (!rc)
2741                         rc = err;
2742         }
2743         RETURN(rc);
2744 }
2745
2746 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
2747                       __u32 *bits)
2748 {
2749         struct obd_device       *obd = exp->exp_obd;
2750         struct lmv_obd          *lmv = &obd->u.lmv;
2751         int                      rc;
2752         ENTRY;
2753
2754         rc =  md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data, bits);
2755         RETURN(rc);
2756 }
2757
2758 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2759                            const struct lu_fid *fid, ldlm_type_t type,
2760                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
2761                            struct lustre_handle *lockh)
2762 {
2763         struct obd_device       *obd = exp->exp_obd;
2764         struct lmv_obd          *lmv = &obd->u.lmv;
2765         ldlm_mode_t              rc;
2766         int                      i;
2767         ENTRY;
2768
2769         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2770
2771         /*
2772          * With CMD every object can have two locks in different namespaces:
2773          * lookup lock in space of mds storing direntry and update/open lock in
2774          * space of mds storing inode. Thus we check all targets, not only that
2775          * one fid was created in.
2776          */
2777         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2778                 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2779                                    type, policy, mode, lockh);
2780                 if (rc)
2781                         RETURN(rc);
2782         }
2783
2784         RETURN(0);
2785 }
2786
2787 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2788                       struct obd_export *dt_exp, struct obd_export *md_exp,
2789                       struct lustre_md *md)
2790 {
2791         struct obd_device       *obd = exp->exp_obd;
2792         struct lmv_obd          *lmv = &obd->u.lmv;
2793         int                      rc;
2794         ENTRY;
2795         rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2796         RETURN(rc);
2797 }
2798
2799 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2800 {
2801         struct obd_device       *obd = exp->exp_obd;
2802         struct lmv_obd          *lmv = &obd->u.lmv;
2803         ENTRY;
2804
2805         if (md->mea)
2806                 obd_free_memmd(exp, (void *)&md->mea);
2807         RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2808 }
2809
2810 int lmv_set_open_replay_data(struct obd_export *exp,
2811                              struct obd_client_handle *och,
2812                              struct ptlrpc_request *open_req)
2813 {
2814         struct obd_device       *obd = exp->exp_obd;
2815         struct lmv_obd          *lmv = &obd->u.lmv;
2816         struct lmv_tgt_desc     *tgt;
2817         ENTRY;
2818
2819         tgt = lmv_find_target(lmv, &och->och_fid);
2820         if (IS_ERR(tgt))
2821                 RETURN(PTR_ERR(tgt));
2822
2823         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2824 }
2825
2826 int lmv_clear_open_replay_data(struct obd_export *exp,
2827                                struct obd_client_handle *och)
2828 {
2829         struct obd_device       *obd = exp->exp_obd;
2830         struct lmv_obd          *lmv = &obd->u.lmv;
2831         struct lmv_tgt_desc     *tgt;
2832         ENTRY;
2833
2834         tgt = lmv_find_target(lmv, &och->och_fid);
2835         if (IS_ERR(tgt))
2836                 RETURN(PTR_ERR(tgt));
2837
2838         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2839 }
2840
2841 static int lmv_get_remote_perm(struct obd_export *exp,
2842                                const struct lu_fid *fid,
2843                                struct obd_capa *oc, __u32 suppgid,
2844                                struct ptlrpc_request **request)
2845 {
2846         struct obd_device       *obd = exp->exp_obd;
2847         struct lmv_obd          *lmv = &obd->u.lmv;
2848         struct lmv_tgt_desc     *tgt;
2849         int                      rc;
2850         ENTRY;
2851
2852         rc = lmv_check_connect(obd);
2853         if (rc)
2854                 RETURN(rc);
2855
2856         tgt = lmv_find_target(lmv, fid);
2857         if (IS_ERR(tgt))
2858                 RETURN(PTR_ERR(tgt));
2859
2860         rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
2861         RETURN(rc);
2862 }
2863
2864 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2865                           renew_capa_cb_t cb)
2866 {
2867         struct obd_device       *obd = exp->exp_obd;
2868         struct lmv_obd          *lmv = &obd->u.lmv;
2869         struct lmv_tgt_desc     *tgt;
2870         int                      rc;
2871         ENTRY;
2872
2873         rc = lmv_check_connect(obd);
2874         if (rc)
2875                 RETURN(rc);
2876
2877         tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
2878         if (IS_ERR(tgt))
2879                 RETURN(PTR_ERR(tgt));
2880
2881         rc = md_renew_capa(tgt->ltd_exp, oc, cb);
2882         RETURN(rc);
2883 }
2884
2885 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
2886                     const struct req_msg_field *field, struct obd_capa **oc)
2887 {
2888         struct obd_device *obd = exp->exp_obd;
2889         struct lmv_obd *lmv = &obd->u.lmv;
2890         int rc;
2891
2892         ENTRY;
2893         rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
2894         RETURN(rc);
2895 }
2896
2897 int lmv_intent_getattr_async(struct obd_export *exp,
2898                              struct md_enqueue_info *minfo,
2899                              struct ldlm_enqueue_info *einfo)
2900 {
2901         struct md_op_data       *op_data = &minfo->mi_data;
2902         struct obd_device       *obd = exp->exp_obd;
2903         struct lmv_obd          *lmv = &obd->u.lmv;
2904         struct lmv_object       *obj;
2905         struct lmv_tgt_desc     *tgt;
2906         int                      rc;
2907         int                      sidx;
2908         ENTRY;
2909
2910         rc = lmv_check_connect(obd);
2911         if (rc)
2912                 RETURN(rc);
2913
2914         if (!fid_is_sane(&op_data->op_fid2)) {
2915                 obj = lmv_object_find(obd, &op_data->op_fid1);
2916                 if (obj && op_data->op_namelen) {
2917                         sidx = raw_name2idx(obj->lo_hashtype,
2918                                             obj->lo_objcount,
2919                                             (char *)op_data->op_name,
2920                                             op_data->op_namelen);
2921                         op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2922                         tgt = lmv_get_target(lmv,
2923                                              obj->lo_stripes[sidx].ls_mds);
2924                         CDEBUG(D_INODE,
2925                                "Choose slave dir ("DFID") -> mds #%d\n",
2926                                PFID(&op_data->op_fid1), tgt->ltd_idx);
2927                 } else {
2928                         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2929                 }
2930                 if (obj)
2931                         lmv_object_put(obj);
2932         } else {
2933                 op_data->op_fid1 = op_data->op_fid2;
2934                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2935                 op_data->op_bias = MDS_CROSS_REF;
2936                 /*
2937                  * Unfortunately, we have to lie to MDC/MDS to retrieve
2938                  * attributes llite needs.
2939                 */
2940                 if (minfo->mi_it.it_op & IT_LOOKUP)
2941                         minfo->mi_it.it_op = IT_GETATTR;
2942         }
2943
2944         if (IS_ERR(tgt))
2945                 RETURN(PTR_ERR(tgt));
2946
2947         rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
2948         RETURN(rc);
2949 }
2950
2951 int lmv_revalidate_lock(struct obd_export *exp,
2952                         struct lookup_intent *it,
2953                         struct lu_fid *fid)
2954 {
2955         struct obd_device       *obd = exp->exp_obd;
2956         struct lmv_obd          *lmv = &obd->u.lmv;
2957         struct lmv_tgt_desc     *tgt;
2958         int                      rc;
2959         ENTRY;
2960
2961         rc = lmv_check_connect(obd);
2962         if (rc)
2963                 RETURN(rc);
2964
2965         tgt = lmv_find_target(lmv, fid);
2966         if (IS_ERR(tgt))
2967                 RETURN(PTR_ERR(tgt));
2968
2969         rc = md_revalidate_lock(tgt->ltd_exp, it, fid);
2970         RETURN(rc);
2971 }
2972
2973
2974 struct obd_ops lmv_obd_ops = {
2975         .o_owner                = THIS_MODULE,
2976         .o_setup                = lmv_setup,
2977         .o_cleanup              = lmv_cleanup,
2978         .o_precleanup           = lmv_precleanup,
2979         .o_process_config       = lmv_process_config,
2980         .o_connect              = lmv_connect,
2981         .o_disconnect           = lmv_disconnect,
2982         .o_statfs               = lmv_statfs,
2983         .o_get_info             = lmv_get_info,
2984         .o_set_info_async       = lmv_set_info_async,
2985         .o_packmd               = lmv_packmd,
2986         .o_unpackmd             = lmv_unpackmd,
2987         .o_notify               = lmv_notify,
2988         .o_get_uuid             = lmv_get_uuid,
2989         .o_iocontrol            = lmv_iocontrol,
2990         .o_fid_delete           = lmv_fid_delete
2991 };
2992
2993 struct md_ops lmv_md_ops = {
2994         .m_getstatus            = lmv_getstatus,
2995         .m_change_cbdata        = lmv_change_cbdata,
2996         .m_close                = lmv_close,
2997         .m_create               = lmv_create,
2998         .m_done_writing         = lmv_done_writing,
2999         .m_enqueue              = lmv_enqueue,
3000         .m_getattr              = lmv_getattr,
3001         .m_getxattr             = lmv_getxattr,
3002         .m_getattr_name         = lmv_getattr_name,
3003         .m_intent_lock          = lmv_intent_lock,
3004         .m_link                 = lmv_link,
3005         .m_rename               = lmv_rename,
3006         .m_setattr              = lmv_setattr,
3007         .m_setxattr             = lmv_setxattr,
3008         .m_sync                 = lmv_sync,
3009         .m_readpage             = lmv_readpage,
3010         .m_unlink               = lmv_unlink,
3011         .m_init_ea_size         = lmv_init_ea_size,
3012         .m_cancel_unused        = lmv_cancel_unused,
3013         .m_set_lock_data        = lmv_set_lock_data,
3014         .m_lock_match           = lmv_lock_match,
3015         .m_get_lustre_md        = lmv_get_lustre_md,
3016         .m_free_lustre_md       = lmv_free_lustre_md,
3017         .m_set_open_replay_data = lmv_set_open_replay_data,
3018         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3019         .m_renew_capa           = lmv_renew_capa,
3020         .m_unpack_capa          = lmv_unpack_capa,
3021         .m_get_remote_perm      = lmv_get_remote_perm,
3022         .m_intent_getattr_async = lmv_intent_getattr_async,
3023         .m_revalidate_lock      = lmv_revalidate_lock
3024 };
3025
3026 static quota_interface_t *quota_interface;
3027 extern quota_interface_t lmv_quota_interface;
3028
3029 int __init lmv_init(void)
3030 {
3031         struct lprocfs_static_vars lvars;
3032         int                        rc;
3033
3034         lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3035                                                 sizeof(struct lmv_object),
3036                                                 0, 0);
3037         if (!lmv_object_cache) {
3038                 CERROR("Error allocating lmv objects cache\n");
3039                 return -ENOMEM;
3040         }
3041
3042         lprocfs_lmv_init_vars(&lvars);
3043
3044         request_module("lquota");
3045         quota_interface = PORTAL_SYMBOL_GET(lmv_quota_interface);
3046         init_obd_quota_ops(quota_interface, &lmv_obd_ops);
3047
3048         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3049                                  lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3050         if (rc) {
3051                 if (quota_interface)
3052                         PORTAL_SYMBOL_PUT(lmv_quota_interface);
3053                 cfs_mem_cache_destroy(lmv_object_cache);
3054         }
3055
3056         return rc;
3057 }
3058
3059 #ifdef __KERNEL__
3060 static void lmv_exit(void)
3061 {
3062         if (quota_interface)
3063                 PORTAL_SYMBOL_PUT(lmv_quota_interface);
3064
3065         class_unregister_type(LUSTRE_LMV_NAME);
3066
3067         LASSERTF(atomic_read(&lmv_object_count) == 0,
3068                  "Can't free lmv objects cache, %d object(s) busy\n",
3069                  atomic_read(&lmv_object_count));
3070         cfs_mem_cache_destroy(lmv_object_cache);
3071 }
3072
3073 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3074 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3075 MODULE_LICENSE("GPL");
3076
3077 module_init(lmv_init);
3078 module_exit(lmv_exit);
3079 #endif