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                 char name[MAX_STRING_SIZE + 1];
 473
 474                 LASSERT(mdc_obd->obd_type != NULL);
 475                 LASSERT(mdc_obd->obd_type->typ_name != NULL);
 476                 name[MAX_STRING_SIZE] = '\0';
 477                 snprintf(name, MAX_STRING_SIZE, "../../../%s/%s",
 478                          mdc_obd->obd_type->typ_name,
 479                          mdc_obd->obd_name);
 480                 mdc_symlink = proc_symlink(mdc_obd->obd_name,
 481                                            lmv_proc_dir, name);
 482                 if (mdc_symlink == NULL) {
 483                         CERROR("Could not register LMV target "
 484                                "/proc/fs/lustre/%s/%s/target_obds/%s.",
 485                                obd->obd_type->typ_name, obd->obd_name,
 486                                mdc_obd->obd_name);
 487                         lprocfs_remove(&lmv_proc_dir);
 488                         lmv_proc_dir = NULL;
 489                 }
 490         }
 491 #endif
 492         RETURN(0);
 493 }
 494
 495 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
 496 {
 497         struct lmv_obd      *lmv = &obd->u.lmv;
 498         struct lmv_tgt_desc *tgt;
 499         int                  rc = 0;
 500         ENTRY;
 501
 502         CDEBUG(D_CONFIG, "Target uuid: %s.\n", tgt_uuid->uuid);
 503
 504         lmv_init_lock(lmv);
 505
 506         if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
 507                 lmv_init_unlock(lmv);
 508                 CERROR("Can't add %s, LMV module compiled for %d MDCs. "
 509                        "That many MDCs already configured.\n",
 510                        tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
 511                 RETURN(-EINVAL);
 512         }
 513         if (lmv->desc.ld_tgt_count == 0) {
 514                 struct obd_device *mdc_obd;
 515
 516                 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
 517                                                 &obd->obd_uuid);
 518                 if (!mdc_obd) {
 519                         lmv_init_unlock(lmv);
 520                         CERROR("Target %s not attached\n", tgt_uuid->uuid);
 521                         RETURN(-EINVAL);
 522                 }
 523
 524                 rc = obd_llog_init(obd, &obd->obd_olg, mdc_obd, NULL);
 525                 if (rc) {
 526                         lmv_init_unlock(lmv);
 527                         CERROR("lmv failed to setup llogging subsystems\n");
 528                 }
 529         }
 530         spin_lock(&lmv->lmv_lock);
 531         tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
 532         tgt->ltd_uuid = *tgt_uuid;
 533         spin_unlock(&lmv->lmv_lock);
 534
 535         if (lmv->connected) {
 536                 rc = lmv_connect_mdc(obd, tgt);
 537                 if (rc) {
 538                         spin_lock(&lmv->lmv_lock);
 539                         lmv->desc.ld_tgt_count--;
 540                         memset(tgt, 0, sizeof(*tgt));
 541                         spin_unlock(&lmv->lmv_lock);
 542                 } else {
 543                         int easize = sizeof(struct lmv_stripe_md) +
 544                                      lmv->desc.ld_tgt_count *
 545                                      sizeof(struct lu_fid);
 546                         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 547                 }
 548         }
 549
 550         lmv_init_unlock(lmv);
 551         RETURN(rc);
 552 }
 553
 554 int lmv_check_connect(struct obd_device *obd)
 555 {
 556         struct lmv_obd       *lmv = &obd->u.lmv;
 557         struct lmv_tgt_desc  *tgt;
 558         int                   i;
 559         int                   rc;
 560         int                   easize;
 561         ENTRY;
 562
 563         if (lmv->connected)
 564                 RETURN(0);
 565
 566         lmv_init_lock(lmv);
 567         if (lmv->connected) {
 568                 lmv_init_unlock(lmv);
 569                 RETURN(0);
 570         }
 571
 572         if (lmv->desc.ld_tgt_count == 0) {
 573                 CERROR("%s: no targets configured.\n", obd->obd_name);
 574                 RETURN(-EINVAL);
 575         }
 576
 577         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 578                lmv->cluuid.uuid, obd->obd_name);
 579
 580         LASSERT(lmv->tgts != NULL);
 581
 582         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 583                 rc = lmv_connect_mdc(obd, tgt);
 584                 if (rc)
 585                         GOTO(out_disc, rc);
 586         }
 587
 588         lmv_set_timeouts(obd);
 589         class_export_put(lmv->exp);
 590         lmv->connected = 1;
 591         easize = lmv_get_easize(lmv);
 592         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 593         lmv_init_unlock(lmv);
 594         RETURN(0);
 595
 596  out_disc:
 597         while (i-- > 0) {
 598                 int rc2;
 599                 --tgt;
 600                 tgt->ltd_active = 0;
 601                 if (tgt->ltd_exp) {
 602                         --lmv->desc.ld_active_tgt_count;
 603                         rc2 = obd_disconnect(tgt->ltd_exp);
 604                         if (rc2) {
 605                                 CERROR("LMV target %s disconnect on "
 606                                        "MDC idx %d: error %d\n",
 607                                        tgt->ltd_uuid.uuid, i, rc2);
 608                         }
 609                 }
 610         }
 611         class_disconnect(lmv->exp);
 612         lmv_init_unlock(lmv);
 613         RETURN(rc);
 614 }
 615
 616 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 617 {
 618 #ifdef __KERNEL__
 619         struct proc_dir_entry  *lmv_proc_dir;
 620 #endif
 621         struct lmv_obd         *lmv = &obd->u.lmv;
 622         struct obd_device      *mdc_obd;
 623         int                     rc;
 624         ENTRY;
 625
 626         LASSERT(tgt != NULL);
 627         LASSERT(obd != NULL);
 628
 629         mdc_obd = class_exp2obd(tgt->ltd_exp);
 630
 631         if (mdc_obd)
 632                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 633
 634 #ifdef __KERNEL__
 635         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 636         if (lmv_proc_dir) {
 637                 struct proc_dir_entry *mdc_symlink;
 638
 639                 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
 640                 if (mdc_symlink) {
 641                         lprocfs_remove(&mdc_symlink);
 642                 } else {
 643                         CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
 644                                obd->obd_type->typ_name, obd->obd_name,
 645                                mdc_obd->obd_name);
 646                 }
 647         }
 648 #endif
 649         rc = obd_fid_fini(tgt->ltd_exp);
 650         if (rc)
 651                 CERROR("Can't finanize fids factory\n");
 652
 653         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 654                tgt->ltd_exp->exp_obd->obd_name,
 655                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 656
 657         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 658         rc = obd_disconnect(tgt->ltd_exp);
 659         if (rc) {
 660                 if (tgt->ltd_active) {
 661                         CERROR("Target %s disconnect error %d\n",
 662                                tgt->ltd_uuid.uuid, rc);
 663                 }
 664         }
 665
 666         lmv_activate_target(lmv, tgt, 0);
 667         tgt->ltd_exp = NULL;
 668         RETURN(0);
 669 }
 670
 671 static int lmv_disconnect(struct obd_export *exp)
 672 {
 673         struct obd_device     *obd = class_exp2obd(exp);
 674 #ifdef __KERNEL__
 675         struct proc_dir_entry *lmv_proc_dir;
 676 #endif
 677         struct lmv_obd        *lmv = &obd->u.lmv;
 678         int                    rc;
 679         int                    i;
 680         ENTRY;
 681
 682         if (!lmv->tgts)
 683                 goto out_local;
 684
 685         /*
 686          * Only disconnect the underlying layers on the final disconnect.
 687          */
 688         lmv->refcount--;
 689         if (lmv->refcount != 0)
 690                 goto out_local;
 691
 692         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 693                 if (lmv->tgts[i].ltd_exp == NULL)
 694                         continue;
 695                 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
 696         }
 697
 698 #ifdef __KERNEL__
 699         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 700         if (lmv_proc_dir) {
 701                 lprocfs_remove(&lmv_proc_dir);
 702         } else {
 703                 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
 704                        obd->obd_type->typ_name, obd->obd_name);
 705         }
 706 #endif
 707
 708 out_local:
 709         /*
 710          * This is the case when no real connection is established by
 711          * lmv_check_connect().
 712          */
 713         if (!lmv->connected)
 714                 class_export_put(exp);
 715         rc = class_disconnect(exp);
 716         if (lmv->refcount == 0)
 717                 lmv->connected = 0;
 718         RETURN(rc);
 719 }
 720
 721 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 722                          int len, void *karg, void *uarg)
 723 {
 724         struct obd_device    *obddev = class_exp2obd(exp);
 725         struct lmv_obd       *lmv = &obddev->u.lmv;
 726         int                   i = 0;
 727         int                   rc = 0;
 728         int                   set = 0;
 729         int                   count = lmv->desc.ld_tgt_count;
 730         ENTRY;
 731
 732         if (count == 0)
 733                 RETURN(-ENOTTY);
 734
 735         switch (cmd) {
 736         case IOC_OBD_STATFS: {
 737                 struct obd_ioctl_data *data = karg;
 738                 struct obd_device *mdc_obd;
 739                 struct obd_statfs stat_buf = {0};
 740                 __u32 index;
 741
 742                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 743                 LASSERT(data->ioc_plen1 == sizeof(struct obd_statfs));
 744
 745                 if ((index >= count))
 746                         RETURN(-ENODEV);
 747
 748                 if (!lmv->tgts[index].ltd_active)
 749                         RETURN(-ENODATA);
 750
 751                 mdc_obd = class_exp2obd(lmv->tgts[index].ltd_exp);
 752                 if (!mdc_obd)
 753                         RETURN(-EINVAL);
 754
 755                 rc = obd_statfs(mdc_obd, &stat_buf,
 756                                 cfs_time_current_64() - HZ, 0);
 757                 if (rc)
 758                         RETURN(rc);
 759                 if (copy_to_user(data->ioc_pbuf1, &stat_buf, data->ioc_plen1))
 760                         RETURN(-EFAULT);
 761                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 762                                   data->ioc_plen2))
 763                         RETURN(-EFAULT);
 764                 break;
 765         }
 766         case OBD_IOC_QUOTACTL: {
 767                 struct if_quotactl *qctl = karg;
 768                 struct lmv_tgt_desc *tgt = NULL;
 769                 struct obd_quotactl *oqctl;
 770
 771                 if (qctl->qc_valid == QC_MDTIDX) {
 772                         if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
 773                                 RETURN(-EINVAL);
 774
 775                         tgt = &lmv->tgts[qctl->qc_idx];
 776                         if (!tgt->ltd_exp)
 777                                 RETURN(-EINVAL);
 778                 } else if (qctl->qc_valid == QC_UUID) {
 779                         for (i = 0; i < count; i++) {
 780                                 tgt = &lmv->tgts[i];
 781                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 782                                                      &qctl->obd_uuid))
 783                                         continue;
 784
 785                                 if (tgt->ltd_exp == NULL)
 786                                         RETURN(-EINVAL);
 787
 788                                 break;
 789                         }
 790                 } else {
 791                         RETURN(-EINVAL);
 792                 }
 793
 794                 if (i >= count)
 795                         RETURN(-EAGAIN);
 796
 797                 LASSERT(tgt && tgt->ltd_exp);
 798                 OBD_ALLOC_PTR(oqctl);
 799                 if (!oqctl)
 800                         RETURN(-ENOMEM);
 801
 802                 QCTL_COPY(oqctl, qctl);
 803                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 804                 if (rc == 0) {
 805                         QCTL_COPY(qctl, oqctl);
 806                         qctl->qc_valid = QC_MDTIDX;
 807                         qctl->obd_uuid = tgt->ltd_uuid;
 808                 }
 809                 OBD_FREE_PTR(oqctl);
 810                 break;
 811         }
 812         case OBD_IOC_CHANGELOG_CLEAR: {
 813                 struct ioc_changelog_clear *icc = karg;
 814
 815                 if (icc->icc_mdtindex >= count)
 816                         RETURN(-ENODEV);
 817
 818                 rc = obd_iocontrol(cmd, lmv->tgts[icc->icc_mdtindex].ltd_exp,
 819                                    sizeof(*icc), icc, NULL);
 820                 break;
 821         }
 822
 823         default : {
 824                 for (i = 0; i < count; i++) {
 825                         int err;
 826
 827                         if (lmv->tgts[i].ltd_exp == NULL)
 828                                 continue;
 829
 830                         err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
 831                                             karg, uarg);
 832                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
 833                                 RETURN(err);
 834                         } else if (err) {
 835                                 if (lmv->tgts[i].ltd_active) {
 836                                         CERROR("error: iocontrol MDC %s on MDT"
 837                                                "idx %d cmd %x: err = %d\n",
 838                                                 lmv->tgts[i].ltd_uuid.uuid,
 839                                                 i, cmd, err);
 840                                         if (!rc)
 841                                                 rc = err;
 842                                 }
 843                         } else
 844                                 set = 1;
 845                 }
 846                 if (!set && !rc)
 847                         rc = -EIO;
 848         }
 849         }
 850         RETURN(rc);
 851 }
 852
 853 static int lmv_all_chars_policy(int count, const char *name,
 854                                 int len)
 855 {
 856         unsigned int c = 0;
 857
 858         while (len > 0)
 859                 c += name[--len];
 860         c = c % count;
 861         return c;
 862 }
 863
 864 static int lmv_nid_policy(struct lmv_obd *lmv)
 865 {
 866         struct obd_import *imp;
 867         __u32              id;
 868
 869         /*
 870          * XXX: To get nid we assume that underlying obd device is mdc.
 871          */
 872         imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
 873         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
 874         return id % lmv->desc.ld_tgt_count;
 875 }
 876
 877 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
 878                           placement_policy_t placement)
 879 {
 880         switch (placement) {
 881         case PLACEMENT_CHAR_POLICY:
 882                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
 883                                             op_data->op_name,
 884                                             op_data->op_namelen);
 885         case PLACEMENT_NID_POLICY:
 886                 return lmv_nid_policy(lmv);
 887
 888         default:
 889                 break;
 890         }
 891
 892         CERROR("Unsupported placement policy %x\n", placement);
 893         return -EINVAL;
 894 }
 895
 896 /**
 897  * This is _inode_ placement policy function (not name).
 898  */
 899 static int lmv_placement_policy(struct obd_device *obd,
 900                                 struct md_op_data *op_data,
 901                                 mdsno_t *mds)
 902 {
 903         struct lmv_obd          *lmv = &obd->u.lmv;
 904         struct lmv_object       *obj;
 905         int                      rc;
 906         ENTRY;
 907
 908         LASSERT(mds != NULL);
 909
 910         if (lmv->desc.ld_tgt_count == 1) {
 911                 *mds = 0;
 912                 RETURN(0);
 913         }
 914
 915         /*
 916          * Allocate new fid on target according to operation type and parent
 917          * home mds.
 918          */
 919         obj = lmv_object_find(obd, &op_data->op_fid1);
 920         if (obj != NULL || op_data->op_name == NULL ||
 921             op_data->op_opc != LUSTRE_OPC_MKDIR) {
 922                 /*
 923                  * Allocate fid for non-dir or for null name or for case parent
 924                  * dir is split.
 925                  */
 926                 if (obj) {
 927                         lmv_object_put(obj);
 928
 929                         /*
 930                          * If we have this flag turned on, and we see that
 931                          * parent dir is split, this means, that caller did not
 932                          * notice split yet. This is race and we would like to
 933                          * let caller know that.
 934                          */
 935                         if (op_data->op_bias & MDS_CHECK_SPLIT)
 936                                 RETURN(-ERESTART);
 937                 }
 938
 939                 /*
 940                  * Allocate new fid on same mds where parent fid is located and
 941                  * where operation will be sent. In case of split dir, ->op_fid1
 942                  * and ->op_mds here will contain fid and mds of slave directory
 943                  * object (assigned by caller).
 944                  */
 945                 *mds = op_data->op_mds;
 946                 rc = 0;
 947         } else {
 948                 /*
 949                  * Parent directory is not split and we want to create a
 950                  * directory in it. Let's calculate where to place it according
 951                  * to operation data @op_data.
 952                  */
 953                 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
 954                 rc = 0;
 955         }
 956
 957         if (rc) {
 958                 CERROR("Can't choose MDS, err = %d\n", rc);
 959         } else {
 960                 LASSERT(*mds < lmv->desc.ld_tgt_count);
 961         }
 962
 963         RETURN(rc);
 964 }
 965
 966 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
 967                     mdsno_t mds)
 968 {
 969         struct lmv_tgt_desc *tgt;
 970         int                  rc;
 971         ENTRY;
 972
 973         tgt = lmv_get_target(lmv, mds);
 974
 975         /*
 976          * New seq alloc and FLD setup should be atomic. Otherwise we may find
 977          * on server that seq in new allocated fid is not yet known.
 978          */
 979         down(&tgt->ltd_fid_sem);
 980
 981         if (!tgt->ltd_active)
 982                 GOTO(out, rc = -ENODEV);
 983
 984         /*
 985          * Asking underlaying tgt layer to allocate new fid.
 986          */
 987         rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
 988         if (rc > 0) {
 989                 LASSERT(fid_is_sane(fid));
 990                 rc = 0;
 991         }
 992
 993         EXIT;
 994 out:
 995         up(&tgt->ltd_fid_sem);
 996         return rc;
 997 }
 998
 999 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
1000                   struct md_op_data *op_data)
1001 {
1002         struct obd_device     *obd = class_exp2obd(exp);
1003         struct lmv_obd        *lmv = &obd->u.lmv;
1004         mdsno_t                mds;
1005         int                    rc;
1006         ENTRY;
1007
1008         LASSERT(op_data != NULL);
1009         LASSERT(fid != NULL);
1010
1011         rc = lmv_placement_policy(obd, op_data, &mds);
1012         if (rc) {
1013                 CERROR("Can't get target for allocating fid, "
1014                        "rc %d\n", rc);
1015                 RETURN(rc);
1016         }
1017
1018         rc = __lmv_fid_alloc(lmv, fid, mds);
1019         if (rc) {
1020                 CERROR("Can't alloc new fid, rc %d\n", rc);
1021                 RETURN(rc);
1022         }
1023
1024         RETURN(rc);
1025 }
1026
1027 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1028 {
1029         ENTRY;
1030         LASSERT(exp != NULL && fid != NULL);
1031         if (lmv_object_delete(exp, fid)) {
1032                 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1033                        PFID(fid));
1034         }
1035         RETURN(0);
1036 }
1037
1038 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1039 {
1040         struct lmv_obd             *lmv = &obd->u.lmv;
1041         struct lprocfs_static_vars  lvars;
1042         struct lmv_desc            *desc;
1043         int                         rc;
1044         int                         i = 0;
1045         ENTRY;
1046
1047         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1048                 CERROR("LMV setup requires a descriptor\n");
1049                 RETURN(-EINVAL);
1050         }
1051
1052         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1053         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1054                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1055                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1056                 RETURN(-EINVAL);
1057         }
1058
1059         lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1060
1061         OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1062         if (lmv->tgts == NULL)
1063                 RETURN(-ENOMEM);
1064
1065         for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1066                 sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
1067                 lmv->tgts[i].ltd_idx = i;
1068         }
1069
1070         lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1071
1072         OBD_ALLOC(lmv->datas, lmv->datas_size);
1073         if (lmv->datas == NULL)
1074                 GOTO(out_free_tgts, rc = -ENOMEM);
1075
1076         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1077         lmv->desc.ld_tgt_count = 0;
1078         lmv->desc.ld_active_tgt_count = 0;
1079         lmv->max_cookiesize = 0;
1080         lmv->max_def_easize = 0;
1081         lmv->max_easize = 0;
1082         lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1083
1084         spin_lock_init(&lmv->lmv_lock);
1085         sema_init(&lmv->init_sem, 1);
1086
1087         rc = lmv_object_setup(obd);
1088         if (rc) {
1089                 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1090                 GOTO(out_free_datas, rc);
1091         }
1092
1093         lprocfs_lmv_init_vars(&lvars);
1094         lprocfs_obd_setup(obd, lvars.obd_vars);
1095 #ifdef LPROCFS
1096         {
1097                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1098                                         0444, &lmv_proc_target_fops, obd);
1099                 if (rc)
1100                         CWARN("Error adding target_obd_stats file (%d)\n", rc);
1101        }
1102 #endif
1103         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1104                              LUSTRE_CLI_FLD_HASH_DHT);
1105         if (rc) {
1106                 CERROR("Can't init FLD, err %d\n", rc);
1107                 GOTO(out_free_datas, rc);
1108         }
1109
1110         RETURN(0);
1111
1112 out_free_datas:
1113         OBD_FREE(lmv->datas, lmv->datas_size);
1114         lmv->datas = NULL;
1115 out_free_tgts:
1116         OBD_FREE(lmv->tgts, lmv->tgts_size);
1117         lmv->tgts = NULL;
1118         return rc;
1119 }
1120
1121 static int lmv_cleanup(struct obd_device *obd)
1122 {
1123         struct lmv_obd   *lmv = &obd->u.lmv;
1124         ENTRY;
1125
1126         fld_client_fini(&lmv->lmv_fld);
1127         lprocfs_obd_cleanup(obd);
1128         lmv_object_cleanup(obd);
1129         OBD_FREE(lmv->datas, lmv->datas_size);
1130         OBD_FREE(lmv->tgts, lmv->tgts_size);
1131
1132         RETURN(0);
1133 }
1134
1135 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1136 {
1137         struct lustre_cfg     *lcfg = buf;
1138         struct obd_uuid        tgt_uuid;
1139         int                    rc;
1140         ENTRY;
1141
1142         switch(lcfg->lcfg_command) {
1143         case LCFG_ADD_MDC:
1144                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1145                         GOTO(out, rc = -EINVAL);
1146
1147                 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1148                 rc = lmv_add_target(obd, &tgt_uuid);
1149                 GOTO(out, rc);
1150         default: {
1151                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1152                 GOTO(out, rc = -EINVAL);
1153         }
1154         }
1155 out:
1156         RETURN(rc);
1157 }
1158
1159 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1160                       __u64 max_age, __u32 flags)
1161 {
1162         struct lmv_obd        *lmv = &obd->u.lmv;
1163         struct obd_statfs     *temp;
1164         int                    rc = 0;
1165         int                    i;
1166         ENTRY;
1167
1168         rc = lmv_check_connect(obd);
1169         if (rc)
1170                 RETURN(rc);
1171
1172         OBD_ALLOC(temp, sizeof(*temp));
1173         if (temp == NULL)
1174                 RETURN(-ENOMEM);
1175
1176         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1177                 if (lmv->tgts[i].ltd_exp == NULL)
1178                         continue;
1179
1180                 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1181                                 max_age, flags);
1182                 if (rc) {
1183                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1184                                lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1185                                rc);
1186                         GOTO(out_free_temp, rc);
1187                 }
1188                 if (i == 0) {
1189                         *osfs = *temp;
1190                 } else {
1191                         osfs->os_bavail += temp->os_bavail;
1192                         osfs->os_blocks += temp->os_blocks;
1193                         osfs->os_ffree += temp->os_ffree;
1194                         osfs->os_files += temp->os_files;
1195                 }
1196         }
1197
1198         EXIT;
1199 out_free_temp:
1200         OBD_FREE(temp, sizeof(*temp));
1201         return rc;
1202 }
1203
1204 static int lmv_getstatus(struct obd_export *exp,
1205                          struct lu_fid *fid,
1206                          struct obd_capa **pc)
1207 {
1208         struct obd_device    *obd = exp->exp_obd;
1209         struct lmv_obd       *lmv = &obd->u.lmv;
1210         int                   rc;
1211         ENTRY;
1212
1213         rc = lmv_check_connect(obd);
1214         if (rc)
1215                 RETURN(rc);
1216
1217         rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1218         RETURN(rc);
1219 }
1220
1221 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1222                         struct obd_capa *oc, obd_valid valid, const char *name,
1223                         const char *input, int input_size, int output_size,
1224                         int flags, struct ptlrpc_request **request)
1225 {
1226         struct obd_device      *obd = exp->exp_obd;
1227         struct lmv_obd         *lmv = &obd->u.lmv;
1228         struct lmv_tgt_desc    *tgt;
1229         int                     rc;
1230         ENTRY;
1231
1232         rc = lmv_check_connect(obd);
1233         if (rc)
1234                 RETURN(rc);
1235
1236         tgt = lmv_find_target(lmv, fid);
1237         if (IS_ERR(tgt))
1238                 RETURN(PTR_ERR(tgt));
1239
1240         rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1241                          input_size, output_size, flags, request);
1242
1243         RETURN(rc);
1244 }
1245
1246 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1247                         struct obd_capa *oc, obd_valid valid, const char *name,
1248                         const char *input, int input_size, int output_size,
1249                         int flags, __u32 suppgid,
1250                         struct ptlrpc_request **request)
1251 {
1252         struct obd_device      *obd = exp->exp_obd;
1253         struct lmv_obd         *lmv = &obd->u.lmv;
1254         struct lmv_tgt_desc    *tgt;
1255         int                     rc;
1256         ENTRY;
1257
1258         rc = lmv_check_connect(obd);
1259         if (rc)
1260                 RETURN(rc);
1261
1262         tgt = lmv_find_target(lmv, fid);
1263         if (IS_ERR(tgt))
1264                 RETURN(PTR_ERR(tgt));
1265
1266         rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1267                          input_size, output_size, flags, suppgid,
1268                          request);
1269
1270         RETURN(rc);
1271 }
1272
1273 static int lmv_getattr(struct obd_export *exp, const struct lu_fid *fid,
1274                        struct obd_capa *oc, obd_valid valid, int ea_size,
1275                        struct ptlrpc_request **request)
1276 {
1277         struct obd_device       *obd = exp->exp_obd;
1278         struct lmv_obd          *lmv = &obd->u.lmv;
1279         struct lmv_tgt_desc     *tgt;
1280         struct lmv_object       *obj;
1281         int                      rc;
1282         int                      i;
1283         ENTRY;
1284
1285         rc = lmv_check_connect(obd);
1286         if (rc)
1287                 RETURN(rc);
1288
1289         tgt = lmv_find_target(lmv, fid);
1290         if (IS_ERR(tgt))
1291                 RETURN(PTR_ERR(tgt));
1292
1293         rc = md_getattr(tgt->ltd_exp, fid, oc, valid, ea_size, request);
1294         if (rc)
1295                 RETURN(rc);
1296
1297         obj = lmv_object_find_lock(obd, fid);
1298
1299         CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(fid),
1300                obj ? "(split)" : "");
1301
1302         /*
1303          * If object is split, then we loop over all the slaves and gather size
1304          * attribute. In ideal world we would have to gather also mds field from
1305          * all slaves, as object is spread over the cluster and this is
1306          * definitely interesting information and it is not good to loss it,
1307          * but...
1308          */
1309         if (obj) {
1310                 struct mdt_body *body;
1311
1312                 if (*request == NULL) {
1313                         lmv_object_put(obj);
1314                         RETURN(rc);
1315                 }
1316
1317                 body = req_capsule_server_get(&(*request)->rq_pill,
1318                                               &RMF_MDT_BODY);
1319                 LASSERT(body != NULL);
1320
1321                 for (i = 0; i < obj->lo_objcount; i++) {
1322                         if (lmv->tgts[i].ltd_exp == NULL) {
1323                                 CWARN("%s: NULL export for %d\n",
1324                                       obd->obd_name, i);
1325                                 continue;
1326                         }
1327
1328                         /*
1329                          * Skip master object.
1330                          */
1331                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1332                                 continue;
1333
1334                         body->size += obj->lo_stripes[i].ls_size;
1335                 }
1336
1337                 lmv_object_put_unlock(obj);
1338         }
1339
1340         RETURN(rc);
1341 }
1342
1343 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1344                              ldlm_iterator_t it, void *data)
1345 {
1346         struct obd_device   *obd = exp->exp_obd;
1347         struct lmv_obd      *lmv = &obd->u.lmv;
1348         int                  i;
1349         int                  rc;
1350         ENTRY;
1351
1352         rc = lmv_check_connect(obd);
1353         if (rc)
1354                 RETURN(rc);
1355
1356         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1357
1358         /*
1359          * With CMD every object can have two locks in different namespaces:
1360          * lookup lock in space of mds storing direntry and update/open lock in
1361          * space of mds storing inode.
1362          */
1363         for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1364                 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1365
1366         RETURN(0);
1367 }
1368
1369 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1370                      struct md_open_data *mod, struct ptlrpc_request **request)
1371 {
1372         struct obd_device     *obd = exp->exp_obd;
1373         struct lmv_obd        *lmv = &obd->u.lmv;
1374         struct lmv_tgt_desc   *tgt;
1375         int                    rc;
1376         ENTRY;
1377
1378         rc = lmv_check_connect(obd);
1379         if (rc)
1380                 RETURN(rc);
1381
1382         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1383         if (IS_ERR(tgt))
1384                 RETURN(PTR_ERR(tgt));
1385
1386         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1387         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1388         RETURN(rc);
1389 }
1390
1391 /**
1392  * Called in the case MDS returns -ERESTART on create on open, what means that
1393  * directory is split and its LMV presentation object has to be updated.
1394  */
1395 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1396 {
1397         struct obd_device       *obd = exp->exp_obd;
1398         struct lmv_obd          *lmv = &obd->u.lmv;
1399         struct ptlrpc_request   *req = NULL;
1400         struct lmv_tgt_desc     *tgt;
1401         struct lmv_object       *obj;
1402         struct lustre_md         md;
1403         int                      mealen;
1404         int                      rc;
1405         __u64                    valid;
1406         ENTRY;
1407
1408         md.mea = NULL;
1409         mealen = lmv_get_easize(lmv);
1410
1411         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1412
1413         tgt = lmv_find_target(lmv, fid);
1414         if (IS_ERR(tgt))
1415                 RETURN(PTR_ERR(tgt));
1416
1417         /*
1418          * Time to update mea of parent fid.
1419          */
1420         rc = md_getattr(tgt->ltd_exp, fid, NULL, valid, mealen, &req);
1421         if (rc) {
1422                 CERROR("md_getattr() failed, error %d\n", rc);
1423                 GOTO(cleanup, rc);
1424         }
1425
1426         rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1427         if (rc) {
1428                 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1429                 GOTO(cleanup, rc);
1430         }
1431
1432         if (md.mea == NULL)
1433                 GOTO(cleanup, rc = -ENODATA);
1434
1435         obj = lmv_object_create(exp, fid, md.mea);
1436         if (IS_ERR(obj))
1437                 rc = PTR_ERR(obj);
1438         else
1439                 lmv_object_put(obj);
1440
1441         obd_free_memmd(exp, (void *)&md.mea);
1442         EXIT;
1443 cleanup:
1444         if (req)
1445                 ptlrpc_req_finished(req);
1446         return rc;
1447 }
1448
1449 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1450                const void *data, int datalen, int mode, __u32 uid,
1451                __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1452                struct ptlrpc_request **request)
1453 {
1454         struct obd_device       *obd = exp->exp_obd;
1455         struct lmv_obd          *lmv = &obd->u.lmv;
1456         struct lmv_tgt_desc     *tgt;
1457         struct lmv_object       *obj;
1458         int                      rc;
1459         int                      loop = 0;
1460         int                      sidx;
1461         ENTRY;
1462
1463         rc = lmv_check_connect(obd);
1464         if (rc)
1465                 RETURN(rc);
1466
1467         if (!lmv->desc.ld_active_tgt_count)
1468                 RETURN(-EIO);
1469 repeat:
1470         ++loop;
1471         LASSERT(loop <= 2);
1472
1473         obj = lmv_object_find(obd, &op_data->op_fid1);
1474         if (obj) {
1475                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1476                                        op_data->op_name, op_data->op_namelen);
1477                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1478                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1479                 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1480                 tgt = lmv_get_target(lmv, op_data->op_mds);
1481                 lmv_object_put(obj);
1482         } else {
1483                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1484                 op_data->op_bias |= MDS_CHECK_SPLIT;
1485                 op_data->op_mds = tgt->ltd_idx;
1486         }
1487
1488         if (IS_ERR(tgt))
1489                 RETURN(PTR_ERR(tgt));
1490
1491         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1492         if (rc == -ERESTART)
1493                 goto repeat;
1494         else if (rc)
1495                 RETURN(rc);
1496
1497         CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #%x\n",
1498                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1499                op_data->op_mds);
1500
1501         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1502         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1503                        cap_effective, rdev, request);
1504         if (rc == 0) {
1505                 if (*request == NULL)
1506                         RETURN(rc);
1507                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1508         } else if (rc == -ERESTART) {
1509                 LASSERT(*request != NULL);
1510                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1511                           "Got -ERESTART during create!\n");
1512                 ptlrpc_req_finished(*request);
1513                 *request = NULL;
1514
1515                 /*
1516                  * Directory got split. Time to update local object and repeat
1517                  * the request with proper MDS.
1518                  */
1519                 rc = lmv_handle_split(exp, &op_data->op_fid1);
1520                 if (rc == 0) {
1521                         rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1522                                                  op_data, &op_data->op_fid2);
1523                         if (rc)
1524                                 RETURN(rc);
1525                         goto repeat;
1526                 }
1527         }
1528         RETURN(rc);
1529 }
1530
1531 static int lmv_done_writing(struct obd_export *exp,
1532                             struct md_op_data *op_data,
1533                             struct md_open_data *mod)
1534 {
1535         struct obd_device     *obd = exp->exp_obd;
1536         struct lmv_obd        *lmv = &obd->u.lmv;
1537         struct lmv_tgt_desc   *tgt;
1538         int                    rc;
1539         ENTRY;
1540
1541         rc = lmv_check_connect(obd);
1542         if (rc)
1543                 RETURN(rc);
1544
1545         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1546         if (IS_ERR(tgt))
1547                 RETURN(PTR_ERR(tgt));
1548
1549         rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1550         RETURN(rc);
1551 }
1552
1553 static int
1554 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1555                    struct lookup_intent *it, struct md_op_data *op_data,
1556                    struct lustre_handle *lockh, void *lmm, int lmmsize)
1557 {
1558         struct obd_device     *obd = exp->exp_obd;
1559         struct lmv_obd        *lmv = &obd->u.lmv;
1560         struct lmv_stripe_md  *mea = op_data->op_mea1;
1561         struct md_op_data     *op_data2;
1562         struct lmv_tgt_desc   *tgt;
1563         int                    i;
1564         int                    rc = 0;
1565         ENTRY;
1566
1567         OBD_ALLOC_PTR(op_data2);
1568         if (op_data2 == NULL)
1569                 RETURN(-ENOMEM);
1570
1571         LASSERT(mea != NULL);
1572         for (i = 0; i < mea->mea_count; i++) {
1573                 memset(op_data2, 0, sizeof(*op_data2));
1574                 op_data2->op_fid1 = mea->mea_ids[i];
1575                 op_data2->op_bias = 0;
1576
1577                 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1578                 if (IS_ERR(tgt))
1579                         GOTO(cleanup, rc = PTR_ERR(tgt));
1580
1581                 if (tgt->ltd_exp == NULL)
1582                         continue;
1583
1584                 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1585                                 lockh + i, lmm, lmmsize, NULL, 0);
1586
1587                 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1588                        PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1589
1590                 if (rc)
1591                         GOTO(cleanup, rc);
1592
1593                 if (it->d.lustre.it_data) {
1594                         struct ptlrpc_request *req;
1595                         req = (struct ptlrpc_request *)it->d.lustre.it_data;
1596                         ptlrpc_req_finished(req);
1597                 }
1598
1599                 if (it->d.lustre.it_status)
1600                         GOTO(cleanup, rc = it->d.lustre.it_status);
1601         }
1602
1603         EXIT;
1604 cleanup:
1605         OBD_FREE_PTR(op_data2);
1606
1607         if (rc != 0) {
1608                 /*
1609                  * Drop all taken locks.
1610                  */
1611                 while (--i >= 0) {
1612                         if (lockh[i].cookie)
1613                                 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1614                         lockh[i].cookie = 0;
1615                 }
1616         }
1617         return rc;
1618 }
1619
1620 static int
1621 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1622                    struct lookup_intent *it, struct md_op_data *op_data,
1623                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1624                    int extra_lock_flags)
1625 {
1626         struct ptlrpc_request      *req = it->d.lustre.it_data;
1627         struct obd_device          *obd = exp->exp_obd;
1628         struct lmv_obd             *lmv = &obd->u.lmv;
1629         struct lustre_handle        plock;
1630         struct lmv_tgt_desc        *tgt;
1631         struct md_op_data          *rdata;
1632         struct lu_fid               fid1;
1633         struct mdt_body            *body;
1634         int                         rc = 0;
1635         int                         pmode;
1636         ENTRY;
1637
1638         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1639         LASSERT(body != NULL);
1640
1641         if (!(body->valid & OBD_MD_MDS))
1642                 RETURN(0);
1643
1644         CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1645                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1646
1647         /*
1648          * We got LOOKUP lock, but we really need attrs.
1649          */
1650         pmode = it->d.lustre.it_lock_mode;
1651         LASSERT(pmode != 0);
1652         memcpy(&plock, lockh, sizeof(plock));
1653         it->d.lustre.it_lock_mode = 0;
1654         it->d.lustre.it_data = NULL;
1655         fid1 = body->fid1;
1656
1657         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1658         ptlrpc_req_finished(req);
1659
1660         tgt = lmv_find_target(lmv, &fid1);
1661         if (IS_ERR(tgt))
1662                 GOTO(out, rc = PTR_ERR(tgt));
1663
1664         OBD_ALLOC_PTR(rdata);
1665         if (rdata == NULL)
1666                 GOTO(out, rc = -ENOMEM);
1667
1668         rdata->op_fid1 = fid1;
1669         rdata->op_bias = MDS_CROSS_REF;
1670
1671         rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1672                         lmm, lmmsize, NULL, extra_lock_flags);
1673         OBD_FREE_PTR(rdata);
1674         EXIT;
1675 out:
1676         ldlm_lock_decref(&plock, pmode);
1677         return rc;
1678 }
1679
1680 static int
1681 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1682             struct lookup_intent *it, struct md_op_data *op_data,
1683             struct lustre_handle *lockh, void *lmm, int lmmsize,
1684             struct ptlrpc_request **req, int extra_lock_flags)
1685 {
1686         struct obd_device        *obd = exp->exp_obd;
1687         struct lmv_obd           *lmv = &obd->u.lmv;
1688         struct lmv_tgt_desc      *tgt;
1689         struct lmv_object        *obj;
1690         int                       sidx;
1691         int                       rc;
1692         ENTRY;
1693
1694         rc = lmv_check_connect(obd);
1695         if (rc)
1696                 RETURN(rc);
1697
1698         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1699                LL_IT2STR(it), PFID(&op_data->op_fid1));
1700
1701         if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1702                 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1703                                         lockh, lmm, lmmsize);
1704                 RETURN(rc);
1705         }
1706
1707         obj = lmv_object_find(obd, &op_data->op_fid1);
1708         if (obj && op_data->op_namelen) {
1709                 sidx = raw_name2idx(obj->lo_hashtype,
1710                                        obj->lo_objcount,
1711                                        (char *)op_data->op_name,
1712                                        op_data->op_namelen);
1713                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1714                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1715         } else {
1716                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1717         }
1718         if (obj)
1719                 lmv_object_put(obj);
1720
1721         if (IS_ERR(tgt))
1722                 RETURN(PTR_ERR(tgt));
1723
1724         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1725                LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1726
1727         rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1728                         lmm, lmmsize, req, extra_lock_flags);
1729
1730         if (rc == 0 && it && it->it_op == IT_OPEN) {
1731                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1732                                         lmm, lmmsize, extra_lock_flags);
1733         }
1734         RETURN(rc);
1735 }
1736
1737 static int
1738 lmv_getattr_name(struct obd_export *exp, const struct lu_fid *fid,
1739                  struct obd_capa *oc, const char *name, int namelen,
1740                  obd_valid valid, int ea_size, __u32 suppgid,
1741                  struct ptlrpc_request **request)
1742 {
1743         struct ptlrpc_request   *req = NULL;
1744         struct obd_device       *obd = exp->exp_obd;
1745         struct lmv_obd          *lmv = &obd->u.lmv;
1746         struct lu_fid            rid = *fid;
1747         struct lmv_tgt_desc     *tgt;
1748         struct mdt_body         *body;
1749         struct lmv_object       *obj;
1750         int                      rc;
1751         int                      loop = 0;
1752         int                      sidx;
1753         ENTRY;
1754
1755         rc = lmv_check_connect(obd);
1756         if (rc)
1757                 RETURN(rc);
1758
1759 repeat:
1760         ++loop;
1761         LASSERT(loop <= 2);
1762         obj = lmv_object_find(obd, &rid);
1763         if (obj) {
1764                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1765                                        name, namelen - 1);
1766                 rid = obj->lo_stripes[sidx].ls_fid;
1767                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1768                 valid &= ~OBD_MD_FLCKSPLIT;
1769                 lmv_object_put(obj);
1770         } else {
1771                 tgt = lmv_find_target(lmv, &rid);
1772                 valid |= OBD_MD_FLCKSPLIT;
1773         }
1774         if (IS_ERR(tgt))
1775                 RETURN(PTR_ERR(tgt));
1776
1777         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1778                namelen, name, PFID(fid), PFID(&rid), tgt->ltd_idx);
1779
1780         rc = md_getattr_name(tgt->ltd_exp, &rid, oc, name, namelen, valid,
1781                              ea_size, suppgid, request);
1782         if (rc == 0) {
1783                 body = req_capsule_server_get(&(*request)->rq_pill,
1784                                               &RMF_MDT_BODY);
1785                 LASSERT(body != NULL);
1786
1787                 if (body->valid & OBD_MD_MDS) {
1788                         rid = body->fid1;
1789                         CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1790                                PFID(&rid));
1791
1792                         tgt = lmv_find_target(lmv, &rid);
1793                         if (IS_ERR(tgt)) {
1794                                 ptlrpc_req_finished(*request);
1795                                 RETURN(PTR_ERR(tgt));
1796                         }
1797
1798                         rc = md_getattr_name(tgt->ltd_exp, &rid, NULL, NULL,
1799                                              1, valid | OBD_MD_FLCROSSREF,
1800                                              ea_size, suppgid, &req);
1801                         ptlrpc_req_finished(*request);
1802                         *request = req;
1803                 }
1804         } else if (rc == -ERESTART) {
1805                 LASSERT(*request != NULL);
1806                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1807                           "Got -ERESTART during getattr!\n");
1808                 ptlrpc_req_finished(*request);
1809                 *request = NULL;
1810
1811                 /*
1812                  * Directory got split. Time to update local object and repeat
1813                  * the request with proper MDS.
1814                  */
1815                 rc = lmv_handle_split(exp, &rid);
1816                 if (rc == 0)
1817                         goto repeat;
1818         }
1819         RETURN(rc);
1820 }
1821
1822 #define md_op_data_fid(op_data, fl)                     \
1823         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1824          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1825          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1826          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1827          NULL)
1828
1829 static int lmv_early_cancel_slaves(struct obd_export *exp,
1830                                    struct md_op_data *op_data, int op_tgt,
1831                                    ldlm_mode_t mode, int bits, int flag)
1832 {
1833         struct obd_device      *obd = exp->exp_obd;
1834         struct lmv_obd         *lmv = &obd->u.lmv;
1835         ldlm_policy_data_t      policy = {{0}};
1836         struct lu_fid          *op_fid;
1837         struct lu_fid          *st_fid;
1838         struct lmv_tgt_desc    *tgt;
1839         struct lmv_object      *obj;
1840         int                     rc = 0;
1841         int                     i;
1842         ENTRY;
1843
1844         op_fid = md_op_data_fid(op_data, flag);
1845         if (!fid_is_sane(op_fid))
1846                 RETURN(0);
1847
1848         obj = lmv_object_find(obd, op_fid);
1849         if (obj == NULL)
1850                 RETURN(-EALREADY);
1851
1852         policy.l_inodebits.bits = bits;
1853         for (i = 0; i < obj->lo_objcount; i++) {
1854                 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1855                 st_fid = &obj->lo_stripes[i].ls_fid;
1856                 if (op_tgt != tgt->ltd_idx) {
1857                         CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1858                                PFID(st_fid), tgt->ltd_idx);
1859                         rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1860                                               mode, LDLM_FL_ASYNC, NULL);
1861                         if (rc)
1862                                 GOTO(out_put_obj, rc);
1863                 } else {
1864                         CDEBUG(D_INODE,
1865                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1866                                op_tgt, PFID(st_fid));
1867                         /*
1868                          * Do not cancel locks for operation target, they will
1869                          * be handled later in underlaying layer when calling
1870                          * function we run on behalf of.
1871                          */
1872                         *op_fid = *st_fid;
1873                         op_data->op_flags |= flag;
1874                 }
1875         }
1876         EXIT;
1877 out_put_obj:
1878         lmv_object_put(obj);
1879         return rc;
1880 }
1881
1882 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1883                             int op_tgt, ldlm_mode_t mode, int bits, int flag)
1884 {
1885         struct lu_fid          *fid = md_op_data_fid(op_data, flag);
1886         struct obd_device      *obd = exp->exp_obd;
1887         struct lmv_obd         *lmv = &obd->u.lmv;
1888         struct lmv_tgt_desc    *tgt;
1889         ldlm_policy_data_t      policy = {{0}};
1890         struct lmv_object      *obj;
1891         int                     rc = 0;
1892         ENTRY;
1893
1894         if (!fid_is_sane(fid))
1895                 RETURN(0);
1896
1897         obj = lmv_object_find(obd, fid);
1898         if (obj) {
1899                 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1900                                              bits, flag);
1901                 lmv_object_put(obj);
1902         } else {
1903                 tgt = lmv_find_target(lmv, fid);
1904                 if (IS_ERR(tgt))
1905                         RETURN(PTR_ERR(tgt));
1906
1907                 if (tgt->ltd_idx != op_tgt) {
1908                         CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1909                         policy.l_inodebits.bits = bits;
1910                         rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1911                                               mode, LDLM_FL_ASYNC, NULL);
1912                 } else {
1913                         CDEBUG(D_INODE,
1914                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1915                                op_tgt, PFID(fid));
1916                         op_data->op_flags |= flag;
1917                         rc = 0;
1918                 }
1919
1920         }
1921         RETURN(rc);
1922 }
1923
1924 /*
1925  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1926  * op_data->op_fid2
1927  */
1928 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1929                     struct ptlrpc_request **request)
1930 {
1931         struct obd_device       *obd = exp->exp_obd;
1932         struct lmv_obd          *lmv = &obd->u.lmv;
1933         struct lmv_tgt_desc     *tgt;
1934         struct lmv_object       *obj;
1935         int                      rc;
1936         int                      loop = 0;
1937         mdsno_t                  mds;
1938         int                      sidx;
1939         ENTRY;
1940
1941         rc = lmv_check_connect(obd);
1942         if (rc)
1943                 RETURN(rc);
1944
1945 repeat:
1946         ++loop;
1947         LASSERT(loop <= 2);
1948         LASSERT(op_data->op_namelen != 0);
1949
1950         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
1951                PFID(&op_data->op_fid2), op_data->op_namelen,
1952                op_data->op_name, PFID(&op_data->op_fid1));
1953
1954         obj = lmv_object_find(obd, &op_data->op_fid2);
1955         if (obj) {
1956                 sidx = raw_name2idx(obj->lo_hashtype,
1957                                     obj->lo_objcount,
1958                                     op_data->op_name,
1959                                     op_data->op_namelen);
1960                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
1961                 mds = obj->lo_stripes[sidx].ls_mds;
1962                 lmv_object_put(obj);
1963         } else {
1964                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1965                 if (rc)
1966                         RETURN(rc);
1967         }
1968
1969         CDEBUG(D_INODE, "Forward to mds #%x ("DFID")\n",
1970                mds, PFID(&op_data->op_fid1));
1971
1972         op_data->op_fsuid = current->fsuid;
1973         op_data->op_fsgid = current->fsgid;
1974         op_data->op_cap = cfs_curproc_cap_pack();
1975         tgt = lmv_get_target(lmv, mds);
1976
1977         /*
1978          * Cancel UPDATE lock on child (fid1).
1979          */
1980         op_data->op_flags |= MF_MDC_CANCEL_FID2;
1981         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
1982                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1983         if (rc == 0)
1984                 rc = md_link(tgt->ltd_exp, op_data, request);
1985         if (rc == -ERESTART) {
1986                 LASSERT(*request != NULL);
1987                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1988                           "Got -ERESTART during link!\n");
1989                 ptlrpc_req_finished(*request);
1990                 *request = NULL;
1991
1992                 /*
1993                  * Directory got split. Time to update local object and repeat
1994                  * the request with proper MDS.
1995                  */
1996                 rc = lmv_handle_split(exp, &op_data->op_fid2);
1997                 if (rc == 0)
1998                         goto repeat;
1999         }
2000
2001         RETURN(rc);
2002 }
2003
2004 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2005                       const char *old, int oldlen, const char *new, int newlen,
2006                       struct ptlrpc_request **request)
2007 {
2008         struct obd_device       *obd = exp->exp_obd;
2009         struct lmv_obd          *lmv = &obd->u.lmv;
2010         struct lmv_tgt_desc     *src_tgt;
2011         struct lmv_tgt_desc     *tgt_tgt;
2012         int                      rc;
2013         int                      sidx;
2014         int                      loop = 0;
2015         struct lmv_object       *obj;
2016         mdsno_t                  mds1;
2017         mdsno_t                  mds2;
2018         ENTRY;
2019
2020         LASSERT(oldlen != 0);
2021
2022         CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2023                oldlen, old, PFID(&op_data->op_fid1),
2024                newlen, new, PFID(&op_data->op_fid2));
2025
2026         rc = lmv_check_connect(obd);
2027         if (rc)
2028                 RETURN(rc);
2029
2030 repeat:
2031         ++loop;
2032         LASSERT(loop <= 2);
2033         obj = lmv_object_find(obd, &op_data->op_fid1);
2034         if (obj) {
2035                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2036                                     (char *)old, oldlen);
2037                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2038                 mds1 = obj->lo_stripes[sidx].ls_mds;
2039                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2040                 lmv_object_put(obj);
2041         } else {
2042                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2043                 if (rc)
2044                         RETURN(rc);
2045         }
2046
2047         obj = lmv_object_find(obd, &op_data->op_fid2);
2048         if (obj) {
2049                 /*
2050                  * Directory is already split, so we have to forward request to
2051                  * the right MDS.
2052                  */
2053                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2054                                        (char *)new, newlen);
2055
2056                 mds2 = obj->lo_stripes[sidx].ls_mds;
2057                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2058                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2059                 lmv_object_put(obj);
2060         } else {
2061                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2062                 if (rc)
2063                         RETURN(rc);
2064         }
2065
2066         op_data->op_fsuid = current->fsuid;
2067         op_data->op_fsgid = current->fsgid;
2068         op_data->op_cap = cfs_curproc_cap_pack();
2069
2070         src_tgt = lmv_get_target(lmv, mds1);
2071         tgt_tgt = lmv_get_target(lmv, mds2);
2072
2073         /*
2074          * LOOKUP lock on src child (fid3) should also be cancelled for
2075          * src_tgt in mdc_rename.
2076          */
2077         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2078
2079         /*
2080          * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2081          * own target.
2082          */
2083         rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2084                               LCK_EX, MDS_INODELOCK_UPDATE,
2085                               MF_MDC_CANCEL_FID2);
2086
2087         /*
2088          * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2089          */
2090         if (rc == 0) {
2091                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2092                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2093                                       MF_MDC_CANCEL_FID4);
2094         }
2095
2096         /*
2097          * Cancel all the locks on tgt child (fid4).
2098          */
2099         if (rc == 0)
2100                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2101                                       LCK_EX, MDS_INODELOCK_FULL,
2102                                       MF_MDC_CANCEL_FID4);
2103
2104         if (rc == 0)
2105                 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2106                                new, newlen, request);
2107
2108         if (rc == -ERESTART) {
2109                 LASSERT(*request != NULL);
2110                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2111                           "Got -ERESTART during rename!\n");
2112                 ptlrpc_req_finished(*request);
2113                 *request = NULL;
2114
2115                 /*
2116                  * Directory got split. Time to update local object and repeat
2117                  * the request with proper MDS.
2118                  */
2119                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2120                 if (rc == 0)
2121                         goto repeat;
2122         }
2123         RETURN(rc);
2124 }
2125
2126 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2127                        void *ea, int ealen, void *ea2, int ea2len,
2128                        struct ptlrpc_request **request,
2129                        struct md_open_data **mod)
2130 {
2131         struct obd_device       *obd = exp->exp_obd;
2132         struct lmv_obd          *lmv = &obd->u.lmv;
2133         struct ptlrpc_request   *req;
2134         struct lmv_tgt_desc     *tgt;
2135         struct lmv_object       *obj;
2136         int                      rc = 0;
2137         int                      i;
2138         ENTRY;
2139
2140         rc = lmv_check_connect(obd);
2141         if (rc)
2142                 RETURN(rc);
2143
2144         obj = lmv_object_find(obd, &op_data->op_fid1);
2145
2146         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2147                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2148                obj ? ", split" : "");
2149
2150         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2151         if (obj) {
2152                 for (i = 0; i < obj->lo_objcount; i++) {
2153                         op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2154
2155                         tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2156                         if (IS_ERR(tgt)) {
2157                                 rc = PTR_ERR(tgt);
2158                                 break;
2159                         }
2160
2161                         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2162                                         ea2, ea2len, &req, mod);
2163
2164                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2165                                 /*
2166                                  * This is master object and this request should
2167                                  * be returned back to llite.
2168                                  */
2169                                 *request = req;
2170                         } else {
2171                                 ptlrpc_req_finished(req);
2172                         }
2173
2174                         if (rc)
2175                                 break;
2176                 }
2177                 lmv_object_put(obj);
2178         } else {
2179                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2180                 if (IS_ERR(tgt))
2181                         RETURN(PTR_ERR(tgt));
2182
2183                 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2184                                 ea2len, request, mod);
2185         }
2186         RETURN(rc);
2187 }
2188
2189 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2190                     struct obd_capa *oc, struct ptlrpc_request **request)
2191 {
2192         struct obd_device         *obd = exp->exp_obd;
2193         struct lmv_obd            *lmv = &obd->u.lmv;
2194         struct lmv_tgt_desc       *tgt;
2195         int                        rc;
2196         ENTRY;
2197
2198         rc = lmv_check_connect(obd);
2199         if (rc)
2200                 RETURN(rc);
2201
2202         tgt = lmv_find_target(lmv, fid);
2203         if (IS_ERR(tgt))
2204                 RETURN(PTR_ERR(tgt));
2205
2206         rc = md_sync(tgt->ltd_exp, fid, oc, request);
2207         RETURN(rc);
2208 }
2209
2210 /**
2211  * Main purpose of LMV blocking ast is to remove split directory LMV
2212  * presentation object (struct lmv_object) attached to the lock being revoked.
2213  */
2214 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2215                      void *data, int flag)
2216 {
2217         struct lustre_handle    lockh;
2218         struct lmv_object      *obj;
2219         int                     rc;
2220         ENTRY;
2221
2222         switch (flag) {
2223         case LDLM_CB_BLOCKING:
2224                 ldlm_lock2handle(lock, &lockh);
2225                 rc = ldlm_cli_cancel(&lockh);
2226                 if (rc < 0) {
2227                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2228                         RETURN(rc);
2229                 }
2230                 break;
2231         case LDLM_CB_CANCELING:
2232                 /*
2233                  * Time to drop cached attrs for split directory object
2234                  */
2235                 obj = lock->l_ast_data;
2236                 if (obj) {
2237                         CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2238                                ", master "DFID"\n",
2239                                lock->l_resource->lr_name.name[3] == 1 ?
2240                                "LOOKUP" : "UPDATE",
2241                                lock->l_resource->lr_name.name[0],
2242                                lock->l_resource->lr_name.name[1],
2243                                PFID(&obj->lo_fid));
2244                         lmv_object_put(obj);
2245                 }
2246                 break;
2247         default:
2248                 LBUG();
2249         }
2250         RETURN(0);
2251 }
2252
2253 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2254 {
2255         __u64         val;
2256
2257         val = le64_to_cpu(*hash);
2258         if (val < hash_adj)
2259                 val += MAX_HASH_SIZE;
2260         if (val != DIR_END_OFF)
2261                 *hash = cpu_to_le64(val - hash_adj);
2262 }
2263
2264 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2265 {
2266         __u64              id;
2267         struct obd_import *imp;
2268
2269         /*
2270          * XXX: to get nid we assume that underlying obd device is mdc.
2271          */
2272         imp  = class_exp2cliimp(exp);
2273         id   = imp->imp_connection->c_self + fid_flatten(fid);
2274
2275         CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2276                imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2277
2278         return id ^ (id >> 32);
2279 }
2280
2281 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2282                         struct obd_capa *oc, __u64 offset64, struct page *page,
2283                         struct ptlrpc_request **request)
2284 {
2285         struct obd_device       *obd = exp->exp_obd;
2286         struct lmv_obd          *lmv = &obd->u.lmv;
2287         struct lu_fid            rid = *fid;
2288         struct lmv_object       *obj;
2289         __u64                    offset;
2290         __u64                    hash_adj = 0;
2291         __u32                    rank = 0;
2292         __u64                    seg_size = 0;
2293         __u64                    tgt_tmp = 0;
2294         int                      tgt_idx = 0;
2295         int                      tgt0_idx = 0;
2296         int                      rc;
2297         int                      nr = 0;
2298         struct lmv_stripe       *los;
2299         struct lmv_tgt_desc     *tgt;
2300         struct lu_dirpage       *dp;
2301         struct lu_dirent        *ent;
2302         ENTRY;
2303
2304         offset = offset64;
2305
2306         rc = lmv_check_connect(obd);
2307         if (rc)
2308                 RETURN(rc);
2309
2310         CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2311
2312         /*
2313          * This case handle directory lookup in clustered metadata case (i.e.
2314          * split directory is located on multiple md servers.)
2315          * each server keeps directory entries for certain range of hashes.
2316          * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2317          * first server will keep records with hashes [ 0 ... MAX_HASH / N  - 1],
2318          * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2319          * so on....
2320          *      readdir can simply start reading entries from 0 - N server in
2321          * order but that will not scale well as all client will request dir in
2322          * to server in same order.
2323          * Following algorithm does optimization:
2324          * Instead of doing readdir in 1, 2, ...., N order, client with a
2325          * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2326          * (every client has rank R)
2327          *      But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2328          * since client ask dir from MDS{R} client has pages with offsets
2329          * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2330          * on hash  values that we get.
2331          */
2332         obj = lmv_object_find_lock(obd, fid);
2333         if (obj) {
2334                 nr       = obj->lo_objcount;
2335                 LASSERT(nr > 0);
2336                 seg_size = MAX_HASH_SIZE;
2337                 do_div(seg_size, nr);
2338                 los      = obj->lo_stripes;
2339                 tgt      = lmv_get_target(lmv, los[0].ls_mds);
2340                 rank     = lmv_node_rank(tgt->ltd_exp, fid) % nr;
2341                 tgt_tmp  = offset;
2342                 do_div(tgt_tmp, seg_size);
2343                 tgt0_idx = do_div(tgt_tmp,  nr);
2344                 tgt_idx  = (tgt0_idx + rank) % nr;
2345
2346                 if (tgt_idx < tgt0_idx)
2347                         /*
2348                          * Wrap around.
2349                          *
2350                          * Last segment has unusual length due to division
2351                          * rounding.
2352                          */
2353                         hash_adj = MAX_HASH_SIZE - seg_size * nr;
2354                 else
2355                         hash_adj = 0;
2356
2357                 hash_adj += rank * seg_size;
2358
2359                 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2360                        LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2361                        offset, tgt0_idx, offset + hash_adj, tgt_idx);
2362
2363                 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2364                 rid = obj->lo_stripes[tgt_idx].ls_fid;
2365                 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2366
2367                 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2368                        PFID(&rid), (unsigned long)offset, tgt_idx);
2369         } else
2370                 tgt = lmv_find_target(lmv, &rid);
2371
2372         if (IS_ERR(tgt))
2373                 GOTO(cleanup, rc = PTR_ERR(tgt));
2374
2375         rc = md_readpage(tgt->ltd_exp, &rid, oc, offset, page, request);
2376         if (rc)
2377                 GOTO(cleanup, rc);
2378         if (obj) {
2379                 dp = cfs_kmap(page);
2380
2381                 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2382                 lmv_hash_adjust(&dp->ldp_hash_end,   hash_adj);
2383                 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2384
2385                 for (ent = lu_dirent_start(dp); ent != NULL;
2386                      ent = lu_dirent_next(ent))
2387                         lmv_hash_adjust(&ent->lde_hash, hash_adj);
2388
2389                 if (tgt0_idx != nr - 1) {
2390                         __u64 end;
2391
2392                         end = le64_to_cpu(dp->ldp_hash_end);
2393                         if (end == DIR_END_OFF) {
2394                                 dp->ldp_hash_end = cpu_to_le32(seg_size *
2395                                                                (tgt0_idx + 1));
2396                                 CDEBUG(D_INODE,
2397                                        ""DFID" reset end "LPX64" tgt %d\n",
2398                                        PFID(&rid),
2399                                        le64_to_cpu(dp->ldp_hash_end), tgt_idx);
2400                         }
2401                 }
2402                 cfs_kunmap(page);
2403         }
2404         EXIT;
2405 cleanup:
2406         if (obj)
2407                 lmv_object_put_unlock(obj);
2408         return rc;
2409 }
2410
2411 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2412                       struct ptlrpc_request **request)
2413 {
2414         struct obd_device       *obd = exp->exp_obd;
2415         struct lmv_obd          *lmv = &obd->u.lmv;
2416         struct lmv_tgt_desc     *tgt = NULL;
2417         struct lmv_object       *obj;
2418         int                      rc;
2419         int                      sidx;
2420         int                      loop = 0;
2421         ENTRY;
2422
2423         rc = lmv_check_connect(obd);
2424         if (rc)
2425                 RETURN(rc);
2426
2427 repeat:
2428         ++loop;
2429         LASSERT(loop <= 2);
2430         LASSERT(op_data->op_namelen != 0);
2431
2432         obj = lmv_object_find(obd, &op_data->op_fid1);
2433         if (obj) {
2434                 sidx = raw_name2idx(obj->lo_hashtype,
2435                                     obj->lo_objcount,
2436                                     op_data->op_name,
2437                                     op_data->op_namelen);
2438                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2439                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2440                 tgt = lmv_get_target(lmv,
2441                                      obj->lo_stripes[sidx].ls_mds);
2442                 lmv_object_put(obj);
2443                 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2444                        op_data->op_namelen, op_data->op_name,
2445                        PFID(&op_data->op_fid1), sidx);
2446         }
2447
2448         if (tgt == NULL) {
2449                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2450                 if (IS_ERR(tgt))
2451                         RETURN(PTR_ERR(tgt));
2452                 op_data->op_bias |= MDS_CHECK_SPLIT;
2453         }
2454
2455         op_data->op_fsuid = current->fsuid;
2456         op_data->op_fsgid = current->fsgid;
2457         op_data->op_cap = cfs_curproc_cap_pack();
2458
2459         /*
2460          * If child's fid is given, cancel unused locks for it if it is from
2461          * another export than parent.
2462          *
2463          * LOOKUP lock for child (fid3) should also be cancelled on parent
2464          * tgt_tgt in mdc_unlink().
2465          */
2466         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2467
2468         /*
2469          * Cancel FULL locks on child (fid3).
2470          */
2471         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2472                               MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2473
2474         if (rc == 0)
2475                 rc = md_unlink(tgt->ltd_exp, op_data, request);
2476
2477         if (rc == -ERESTART) {
2478                 LASSERT(*request != NULL);
2479                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2480                           "Got -ERESTART during unlink!\n");
2481                 ptlrpc_req_finished(*request);
2482                 *request = NULL;
2483
2484                 /*
2485                  * Directory got split. Time to update local object and repeat
2486                  * the request with proper MDS.
2487                  */
2488                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2489                 if (rc == 0)
2490                         goto repeat;
2491         }
2492         RETURN(rc);
2493 }
2494
2495 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2496 {
2497         int        rc = 0;
2498
2499         switch (stage) {
2500         case OBD_CLEANUP_EARLY:
2501                 /* XXX: here should be calling obd_precleanup() down to
2502                  * stack. */
2503                 break;
2504         case OBD_CLEANUP_EXPORTS:
2505                 rc = obd_llog_finish(obd, 0);
2506                 if (rc != 0)
2507                         CERROR("failed to cleanup llogging subsystems\n");
2508                 break;
2509         default:
2510                 break;
2511         }
2512         RETURN(rc);
2513 }
2514
2515 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2516                         void *key, __u32 *vallen, void *val,
2517                         struct lov_stripe_md *lsm)
2518 {
2519         struct obd_device       *obd;
2520         struct lmv_obd          *lmv;
2521         int                      rc = 0;
2522         ENTRY;
2523
2524         obd = class_exp2obd(exp);
2525         if (obd == NULL) {
2526                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2527                        exp->exp_handle.h_cookie);
2528                 RETURN(-EINVAL);
2529         }
2530
2531         lmv = &obd->u.lmv;
2532         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2533                 struct lmv_tgt_desc *tgts;
2534                 int i;
2535
2536                 rc = lmv_check_connect(obd);
2537                 if (rc)
2538                         RETURN(rc);
2539
2540                 LASSERT(*vallen == sizeof(__u32));
2541                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2542                      i++, tgts++) {
2543
2544                         /*
2545                          * All tgts should be connected when this gets called.
2546                          */
2547                         if (!tgts || !tgts->ltd_exp) {
2548                                 CERROR("target not setup?\n");
2549                                 continue;
2550                         }
2551
2552                         if (!obd_get_info(tgts->ltd_exp, keylen, key,
2553                                           vallen, val, NULL))
2554                                 RETURN(0);
2555                 }
2556                 RETURN(-EINVAL);
2557         } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2558                 rc = lmv_check_connect(obd);
2559                 if (rc)
2560                         RETURN(rc);
2561
2562                 /*
2563                  * Forwarding this request to first MDS, it should know LOV
2564                  * desc.
2565                  */
2566                 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2567                                   vallen, val, NULL);
2568                 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2569                         exp->exp_connect_flags =
2570                         ((struct obd_connect_data *)val)->ocd_connect_flags;
2571                 }
2572                 RETURN(rc);
2573         }
2574
2575         CDEBUG(D_IOCTL, "Invalid key\n");
2576         RETURN(-EINVAL);
2577 }
2578
2579 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2580                        void *key, obd_count vallen, void *val,
2581                        struct ptlrpc_request_set *set)
2582 {
2583         struct lmv_tgt_desc    *tgt;
2584         struct obd_device      *obd;
2585         struct lmv_obd         *lmv;
2586         int rc = 0;
2587         ENTRY;
2588
2589         obd = class_exp2obd(exp);
2590         if (obd == NULL) {
2591                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2592                        exp->exp_handle.h_cookie);
2593                 RETURN(-EINVAL);
2594         }
2595         lmv = &obd->u.lmv;
2596
2597         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2598             KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2599                 int i, err = 0;
2600
2601                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2602                         tgt = &lmv->tgts[i];
2603
2604                         if (!tgt->ltd_exp)
2605                                 continue;
2606
2607                         err = obd_set_info_async(tgt->ltd_exp,
2608                                                  keylen, key, vallen, val, set);
2609                         if (err && rc == 0)
2610                                 rc = err;
2611                 }
2612
2613                 RETURN(rc);
2614         }
2615
2616         RETURN(-EINVAL);
2617 }
2618
2619 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2620                struct lov_stripe_md *lsm)
2621 {
2622         struct obd_device         *obd = class_exp2obd(exp);
2623         struct lmv_obd            *lmv = &obd->u.lmv;
2624         struct lmv_stripe_md      *meap;
2625         struct lmv_stripe_md      *lsmp;
2626         int                        mea_size;
2627         int                        i;
2628         ENTRY;
2629
2630         mea_size = lmv_get_easize(lmv);
2631         if (!lmmp)
2632                 RETURN(mea_size);
2633
2634         if (*lmmp && !lsm) {
2635                 OBD_FREE(*lmmp, mea_size);
2636                 *lmmp = NULL;
2637                 RETURN(0);
2638         }
2639
2640         if (*lmmp == NULL) {
2641                 OBD_ALLOC(*lmmp, mea_size);
2642                 if (*lmmp == NULL)
2643                         RETURN(-ENOMEM);
2644         }
2645
2646         if (!lsm)
2647                 RETURN(mea_size);
2648
2649         lsmp = (struct lmv_stripe_md *)lsm;
2650         meap = (struct lmv_stripe_md *)*lmmp;
2651
2652         if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2653             lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2654                 RETURN(-EINVAL);
2655
2656         meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2657         meap->mea_count = cpu_to_le32(lsmp->mea_count);
2658         meap->mea_master = cpu_to_le32(lsmp->mea_master);
2659
2660         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2661                 meap->mea_ids[i] = meap->mea_ids[i];
2662                 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2663         }
2664
2665         RETURN(mea_size);
2666 }
2667
2668 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2669                  struct lov_mds_md *lmm, int lmm_size)
2670 {
2671         struct obd_device          *obd = class_exp2obd(exp);
2672         struct lmv_stripe_md      **tmea = (struct lmv_stripe_md **)lsmp;
2673         struct lmv_stripe_md       *mea = (struct lmv_stripe_md *)lmm;
2674         struct lmv_obd             *lmv = &obd->u.lmv;
2675         int                         mea_size;
2676         int                         i;
2677         __u32                       magic;
2678         ENTRY;
2679
2680         mea_size = lmv_get_easize(lmv);
2681         if (lsmp == NULL)
2682                 return mea_size;
2683
2684         if (*lsmp != NULL && lmm == NULL) {
2685                 OBD_FREE(*tmea, mea_size);
2686                 *lsmp = NULL;
2687                 RETURN(0);
2688         }
2689
2690         LASSERT(mea_size == lmm_size);
2691
2692         OBD_ALLOC(*tmea, mea_size);
2693         if (*tmea == NULL)
2694                 RETURN(-ENOMEM);
2695
2696         if (!lmm)
2697                 RETURN(mea_size);
2698
2699         if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2700             mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2701             mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2702         {
2703                 magic = le32_to_cpu(mea->mea_magic);
2704         } else {
2705                 /*
2706                  * Old mea is not handled here.
2707                  */
2708                 CERROR("Old not supportable EA is found\n");
2709                 LBUG();
2710         }
2711
2712         (*tmea)->mea_magic = magic;
2713         (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2714         (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2715
2716         for (i = 0; i < (*tmea)->mea_count; i++) {
2717                 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2718                 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2719         }
2720         RETURN(mea_size);
2721 }
2722
2723 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2724                              ldlm_policy_data_t *policy, ldlm_mode_t mode,
2725                              int flags, void *opaque)
2726 {
2727         struct obd_device       *obd = exp->exp_obd;
2728         struct lmv_obd          *lmv = &obd->u.lmv;
2729         int                      rc = 0;
2730         int                      err;
2731         int                      i;
2732         ENTRY;
2733
2734         LASSERT(fid != NULL);
2735
2736         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2737                 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2738                         continue;
2739
2740                 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2741                                        policy, mode, flags, opaque);
2742                 if (!rc)
2743                         rc = err;
2744         }
2745         RETURN(rc);
2746 }
2747
2748 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
2749                       __u32 *bits)
2750 {
2751         struct obd_device       *obd = exp->exp_obd;
2752         struct lmv_obd          *lmv = &obd->u.lmv;
2753         int                      rc;
2754         ENTRY;
2755
2756         rc =  md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data, bits);
2757         RETURN(rc);
2758 }
2759
2760 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2761                            const struct lu_fid *fid, ldlm_type_t type,
2762                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
2763                            struct lustre_handle *lockh)
2764 {
2765         struct obd_device       *obd = exp->exp_obd;
2766         struct lmv_obd          *lmv = &obd->u.lmv;
2767         ldlm_mode_t              rc;
2768         int                      i;
2769         ENTRY;
2770
2771         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2772
2773         /*
2774          * With CMD every object can have two locks in different namespaces:
2775          * lookup lock in space of mds storing direntry and update/open lock in
2776          * space of mds storing inode. Thus we check all targets, not only that
2777          * one fid was created in.
2778          */
2779         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2780                 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2781                                    type, policy, mode, lockh);
2782                 if (rc)
2783                         RETURN(rc);
2784         }
2785
2786         RETURN(0);
2787 }
2788
2789 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2790                       struct obd_export *dt_exp, struct obd_export *md_exp,
2791                       struct lustre_md *md)
2792 {
2793         struct obd_device       *obd = exp->exp_obd;
2794         struct lmv_obd          *lmv = &obd->u.lmv;
2795         int                      rc;
2796         ENTRY;
2797         rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2798         RETURN(rc);
2799 }
2800
2801 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2802 {
2803         struct obd_device       *obd = exp->exp_obd;
2804         struct lmv_obd          *lmv = &obd->u.lmv;
2805         ENTRY;
2806
2807         if (md->mea)
2808                 obd_free_memmd(exp, (void *)&md->mea);
2809         RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2810 }
2811
2812 int lmv_set_open_replay_data(struct obd_export *exp,
2813                              struct obd_client_handle *och,
2814                              struct ptlrpc_request *open_req)
2815 {
2816         struct obd_device       *obd = exp->exp_obd;
2817         struct lmv_obd          *lmv = &obd->u.lmv;
2818         struct lmv_tgt_desc     *tgt;
2819         ENTRY;
2820
2821         tgt = lmv_find_target(lmv, &och->och_fid);
2822         if (IS_ERR(tgt))
2823                 RETURN(PTR_ERR(tgt));
2824
2825         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2826 }
2827
2828 int lmv_clear_open_replay_data(struct obd_export *exp,
2829                                struct obd_client_handle *och)
2830 {
2831         struct obd_device       *obd = exp->exp_obd;
2832         struct lmv_obd          *lmv = &obd->u.lmv;
2833         struct lmv_tgt_desc     *tgt;
2834         ENTRY;
2835
2836         tgt = lmv_find_target(lmv, &och->och_fid);
2837         if (IS_ERR(tgt))
2838                 RETURN(PTR_ERR(tgt));
2839
2840         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2841 }
2842
2843 static int lmv_get_remote_perm(struct obd_export *exp,
2844                                const struct lu_fid *fid,
2845                                struct obd_capa *oc, __u32 suppgid,
2846                                struct ptlrpc_request **request)
2847 {
2848         struct obd_device       *obd = exp->exp_obd;
2849         struct lmv_obd          *lmv = &obd->u.lmv;
2850         struct lmv_tgt_desc     *tgt;
2851         int                      rc;
2852         ENTRY;
2853
2854         rc = lmv_check_connect(obd);
2855         if (rc)
2856                 RETURN(rc);
2857
2858         tgt = lmv_find_target(lmv, fid);
2859         if (IS_ERR(tgt))
2860                 RETURN(PTR_ERR(tgt));
2861
2862         rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
2863         RETURN(rc);
2864 }
2865
2866 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2867                           renew_capa_cb_t cb)
2868 {
2869         struct obd_device       *obd = exp->exp_obd;
2870         struct lmv_obd          *lmv = &obd->u.lmv;
2871         struct lmv_tgt_desc     *tgt;
2872         int                      rc;
2873         ENTRY;
2874
2875         rc = lmv_check_connect(obd);
2876         if (rc)
2877                 RETURN(rc);
2878
2879         tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
2880         if (IS_ERR(tgt))
2881                 RETURN(PTR_ERR(tgt));
2882
2883         rc = md_renew_capa(tgt->ltd_exp, oc, cb);
2884         RETURN(rc);
2885 }
2886
2887 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
2888                     const struct req_msg_field *field, struct obd_capa **oc)
2889 {
2890         struct obd_device *obd = exp->exp_obd;
2891         struct lmv_obd *lmv = &obd->u.lmv;
2892         int rc;
2893
2894         ENTRY;
2895         rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
2896         RETURN(rc);
2897 }
2898
2899 int lmv_intent_getattr_async(struct obd_export *exp,
2900                              struct md_enqueue_info *minfo,
2901                              struct ldlm_enqueue_info *einfo)
2902 {
2903         struct md_op_data       *op_data = &minfo->mi_data;
2904         struct obd_device       *obd = exp->exp_obd;
2905         struct lmv_obd          *lmv = &obd->u.lmv;
2906         struct lmv_object       *obj;
2907         struct lmv_tgt_desc     *tgt;
2908         int                      rc;
2909         int                      sidx;
2910         ENTRY;
2911
2912         rc = lmv_check_connect(obd);
2913         if (rc)
2914                 RETURN(rc);
2915
2916         if (!fid_is_sane(&op_data->op_fid2)) {
2917                 obj = lmv_object_find(obd, &op_data->op_fid1);
2918                 if (obj && op_data->op_namelen) {
2919                         sidx = raw_name2idx(obj->lo_hashtype,
2920                                             obj->lo_objcount,
2921                                             (char *)op_data->op_name,
2922                                             op_data->op_namelen);
2923                         op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2924                         tgt = lmv_get_target(lmv,
2925                                              obj->lo_stripes[sidx].ls_mds);
2926                         CDEBUG(D_INODE,
2927                                "Choose slave dir ("DFID") -> mds #%d\n",
2928                                PFID(&op_data->op_fid1), tgt->ltd_idx);
2929                 } else {
2930                         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2931                 }
2932                 if (obj)
2933                         lmv_object_put(obj);
2934         } else {
2935                 op_data->op_fid1 = op_data->op_fid2;
2936                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2937                 op_data->op_bias = MDS_CROSS_REF;
2938                 /*
2939                  * Unfortunately, we have to lie to MDC/MDS to retrieve
2940                  * attributes llite needs.
2941                 */
2942                 if (minfo->mi_it.it_op & IT_LOOKUP)
2943                         minfo->mi_it.it_op = IT_GETATTR;
2944         }
2945
2946         if (IS_ERR(tgt))
2947                 RETURN(PTR_ERR(tgt));
2948
2949         rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
2950         RETURN(rc);
2951 }
2952
2953 int lmv_revalidate_lock(struct obd_export *exp,
2954                         struct lookup_intent *it,
2955                         struct lu_fid *fid)
2956 {
2957         struct obd_device       *obd = exp->exp_obd;
2958         struct lmv_obd          *lmv = &obd->u.lmv;
2959         struct lmv_tgt_desc     *tgt;
2960         int                      rc;
2961         ENTRY;
2962
2963         rc = lmv_check_connect(obd);
2964         if (rc)
2965                 RETURN(rc);
2966
2967         tgt = lmv_find_target(lmv, fid);
2968         if (IS_ERR(tgt))
2969                 RETURN(PTR_ERR(tgt));
2970
2971         rc = md_revalidate_lock(tgt->ltd_exp, it, fid);
2972         RETURN(rc);
2973 }
2974
2975
2976 struct obd_ops lmv_obd_ops = {
2977         .o_owner                = THIS_MODULE,
2978         .o_setup                = lmv_setup,
2979         .o_cleanup              = lmv_cleanup,
2980         .o_precleanup           = lmv_precleanup,
2981         .o_process_config       = lmv_process_config,
2982         .o_connect              = lmv_connect,
2983         .o_disconnect           = lmv_disconnect,
2984         .o_statfs               = lmv_statfs,
2985         .o_get_info             = lmv_get_info,
2986         .o_set_info_async       = lmv_set_info_async,
2987         .o_packmd               = lmv_packmd,
2988         .o_unpackmd             = lmv_unpackmd,
2989         .o_notify               = lmv_notify,
2990         .o_get_uuid             = lmv_get_uuid,
2991         .o_iocontrol            = lmv_iocontrol,
2992         .o_fid_delete           = lmv_fid_delete
2993 };
2994
2995 struct md_ops lmv_md_ops = {
2996         .m_getstatus            = lmv_getstatus,
2997         .m_change_cbdata        = lmv_change_cbdata,
2998         .m_close                = lmv_close,
2999         .m_create               = lmv_create,
3000         .m_done_writing         = lmv_done_writing,
3001         .m_enqueue              = lmv_enqueue,
3002         .m_getattr              = lmv_getattr,
3003         .m_getxattr             = lmv_getxattr,
3004         .m_getattr_name         = lmv_getattr_name,
3005         .m_intent_lock          = lmv_intent_lock,
3006         .m_link                 = lmv_link,
3007         .m_rename               = lmv_rename,
3008         .m_setattr              = lmv_setattr,
3009         .m_setxattr             = lmv_setxattr,
3010         .m_sync                 = lmv_sync,
3011         .m_readpage             = lmv_readpage,
3012         .m_unlink               = lmv_unlink,
3013         .m_init_ea_size         = lmv_init_ea_size,
3014         .m_cancel_unused        = lmv_cancel_unused,
3015         .m_set_lock_data        = lmv_set_lock_data,
3016         .m_lock_match           = lmv_lock_match,
3017         .m_get_lustre_md        = lmv_get_lustre_md,
3018         .m_free_lustre_md       = lmv_free_lustre_md,
3019         .m_set_open_replay_data = lmv_set_open_replay_data,
3020         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3021         .m_renew_capa           = lmv_renew_capa,
3022         .m_unpack_capa          = lmv_unpack_capa,
3023         .m_get_remote_perm      = lmv_get_remote_perm,
3024         .m_intent_getattr_async = lmv_intent_getattr_async,
3025         .m_revalidate_lock      = lmv_revalidate_lock
3026 };
3027
3028 static quota_interface_t *quota_interface;
3029 extern quota_interface_t lmv_quota_interface;
3030
3031 int __init lmv_init(void)
3032 {
3033         struct lprocfs_static_vars lvars;
3034         int                        rc;
3035
3036         lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3037                                                 sizeof(struct lmv_object),
3038                                                 0, 0);
3039         if (!lmv_object_cache) {
3040                 CERROR("Error allocating lmv objects cache\n");
3041                 return -ENOMEM;
3042         }
3043
3044         lprocfs_lmv_init_vars(&lvars);
3045
3046         request_module("lquota");
3047         quota_interface = PORTAL_SYMBOL_GET(lmv_quota_interface);
3048         init_obd_quota_ops(quota_interface, &lmv_obd_ops);
3049
3050         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3051                                  lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3052         if (rc) {
3053                 if (quota_interface)
3054                         PORTAL_SYMBOL_PUT(lmv_quota_interface);
3055                 cfs_mem_cache_destroy(lmv_object_cache);
3056         }
3057
3058         return rc;
3059 }
3060
3061 #ifdef __KERNEL__
3062 static void lmv_exit(void)
3063 {
3064         if (quota_interface)
3065                 PORTAL_SYMBOL_PUT(lmv_quota_interface);
3066
3067         class_unregister_type(LUSTRE_LMV_NAME);
3068
3069         LASSERTF(atomic_read(&lmv_object_count) == 0,
3070                  "Can't free lmv objects cache, %d object(s) busy\n",
3071                  atomic_read(&lmv_object_count));
3072         cfs_mem_cache_destroy(lmv_object_cache);
3073 }
3074
3075 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3076 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3077 MODULE_LICENSE("GPL");
3078
3079 module_init(lmv_init);
3080 module_exit(lmv_exit);
3081 #endif