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 lustre_handle *conn, 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 obd_export     *exp;
 251         int                    rc = 0;
 252         ENTRY;
 253
 254         rc = class_connect(conn, obd, cluuid);
 255         if (rc) {
 256                 CERROR("class_connection() returned %d\n", rc);
 257                 RETURN(rc);
 258         }
 259
 260         exp = class_conn2export(conn);
 261
 262         /*
 263          * We don't want to actually do the underlying connections more than
 264          * once, so keep track.
 265          */
 266         lmv->refcount++;
 267         if (lmv->refcount > 1) {
 268                 class_export_put(exp);
 269                 RETURN(0);
 270         }
 271
 272         lmv->exp = exp;
 273         lmv->connected = 0;
 274         lmv->cluuid = *cluuid;
 275
 276         if (data)
 277                 lmv->conn_data = *data;
 278
 279 #ifdef __KERNEL__
 280         lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
 281                                         NULL, NULL);
 282         if (IS_ERR(lmv_proc_dir)) {
 283                 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
 284                        obd->obd_type->typ_name, obd->obd_name);
 285                 lmv_proc_dir = NULL;
 286         }
 287 #endif
 288
 289         /*
 290          * All real clients should perform actual connection right away, because
 291          * it is possible, that LMV will not have opportunity to connect targets
 292          * and MDC stuff will be called directly, for instance while reading
 293          * ../mdc/../kbytesfree procfs file, etc.
 294          */
 295         if (data->ocd_connect_flags & OBD_CONNECT_REAL)
 296                 rc = lmv_check_connect(obd);
 297
 298 #ifdef __KERNEL__
 299         if (rc) {
 300                 if (lmv_proc_dir)
 301                         lprocfs_remove(&lmv_proc_dir);
 302         }
 303 #endif
 304
 305         RETURN(rc);
 306 }
 307
 308 static void lmv_set_timeouts(struct obd_device *obd)
 309 {
 310         struct lmv_tgt_desc   *tgts;
 311         struct lmv_obd        *lmv;
 312         int                    i;
 313
 314         lmv = &obd->u.lmv;
 315         if (lmv->server_timeout == 0)
 316                 return;
 317
 318         if (lmv->connected == 0)
 319                 return;
 320
 321         for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
 322                 if (tgts->ltd_exp == NULL)
 323                         continue;
 324
 325                 obd_set_info_async(tgts->ltd_exp, sizeof(KEY_INTERMDS),
 326                                    KEY_INTERMDS, 0, NULL, NULL);
 327         }
 328 }
 329
 330 static int lmv_init_ea_size(struct obd_export *exp, int easize,
 331                             int def_easize, int cookiesize)
 332 {
 333         struct obd_device   *obd = exp->exp_obd;
 334         struct lmv_obd      *lmv = &obd->u.lmv;
 335         int                  i;
 336         int                  rc = 0;
 337         int                  change = 0;
 338         ENTRY;
 339
 340         if (lmv->max_easize < easize) {
 341                 lmv->max_easize = easize;
 342                 change = 1;
 343         }
 344         if (lmv->max_def_easize < def_easize) {
 345                 lmv->max_def_easize = def_easize;
 346                 change = 1;
 347         }
 348         if (lmv->max_cookiesize < cookiesize) {
 349                 lmv->max_cookiesize = cookiesize;
 350                 change = 1;
 351         }
 352         if (change == 0)
 353                 RETURN(0);
 354
 355         if (lmv->connected == 0)
 356                 RETURN(0);
 357
 358         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 359                 if (lmv->tgts[i].ltd_exp == NULL) {
 360                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
 361                         continue;
 362                 }
 363
 364                 rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize,
 365                                      cookiesize);
 366                 if (rc) {
 367                         CERROR("obd_init_ea_size() failed on MDT target %d, "
 368                                "error %d.\n", i, rc);
 369                         break;
 370                 }
 371         }
 372         RETURN(rc);
 373 }
 374
 375 #define MAX_STRING_SIZE 128
 376
 377 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 378 {
 379 #ifdef __KERNEL__
 380         struct proc_dir_entry   *lmv_proc_dir;
 381 #endif
 382         struct lmv_obd          *lmv = &obd->u.lmv;
 383         struct obd_uuid         *cluuid = &lmv->cluuid;
 384         struct obd_connect_data *mdc_data = NULL;
 385         struct obd_uuid          lmv_mdc_uuid = { "LMV_MDC_UUID" };
 386         struct lustre_handle     conn = {0, };
 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, &conn, 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         mdc_exp = class_conn2export(&conn);
 418
 419         /*
 420          * Init fid sequence client for this mdc and add new fld target.
 421          */
 422         rc = obd_fid_init(mdc_exp);
 423         if (rc)
 424                 RETURN(rc);
 425
 426         target.ft_srv = NULL;
 427         target.ft_exp = mdc_exp;
 428         target.ft_idx = tgt->ltd_idx;
 429
 430         fld_client_add_target(&lmv->lmv_fld, &target);
 431
 432         mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data;
 433
 434         rc = obd_register_observer(mdc_obd, obd);
 435         if (rc) {
 436                 obd_disconnect(mdc_exp);
 437                 CERROR("target %s register_observer error %d\n",
 438                        tgt->ltd_uuid.uuid, rc);
 439                 RETURN(rc);
 440         }
 441
 442         if (obd->obd_observer) {
 443                 /*
 444                  * Tell the observer about the new target.
 445                  */
 446                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 447                                 OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts));
 448                 if (rc) {
 449                         obd_disconnect(mdc_exp);
 450                         RETURN(rc);
 451                 }
 452         }
 453
 454         tgt->ltd_active = 1;
 455         tgt->ltd_exp = mdc_exp;
 456         lmv->desc.ld_active_tgt_count++;
 457
 458         /*
 459          * Copy connect data, it may be used later.
 460          */
 461         lmv->datas[tgt->ltd_idx] = *mdc_data;
 462
 463         md_init_ea_size(tgt->ltd_exp, lmv->max_easize,
 464                         lmv->max_def_easize, lmv->max_cookiesize);
 465
 466         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 467                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 468                 atomic_read(&obd->obd_refcount));
 469
 470 #ifdef __KERNEL__
 471         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 472         if (lmv_proc_dir) {
 473                 struct proc_dir_entry *mdc_symlink;
 474                 char name[MAX_STRING_SIZE + 1];
 475
 476                 LASSERT(mdc_obd->obd_type != NULL);
 477                 LASSERT(mdc_obd->obd_type->typ_name != NULL);
 478                 name[MAX_STRING_SIZE] = '\0';
 479                 snprintf(name, MAX_STRING_SIZE, "../../../%s/%s",
 480                          mdc_obd->obd_type->typ_name,
 481                          mdc_obd->obd_name);
 482                 mdc_symlink = proc_symlink(mdc_obd->obd_name,
 483                                            lmv_proc_dir, name);
 484                 if (mdc_symlink == NULL) {
 485                         CERROR("Could not register LMV target "
 486                                "/proc/fs/lustre/%s/%s/target_obds/%s.",
 487                                obd->obd_type->typ_name, obd->obd_name,
 488                                mdc_obd->obd_name);
 489                         lprocfs_remove(&lmv_proc_dir);
 490                         lmv_proc_dir = NULL;
 491                 }
 492         }
 493 #endif
 494         RETURN(0);
 495 }
 496
 497 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
 498 {
 499         struct lmv_obd      *lmv = &obd->u.lmv;
 500         struct lmv_tgt_desc *tgt;
 501         int                  rc = 0;
 502         ENTRY;
 503
 504         CDEBUG(D_CONFIG, "Target uuid: %s.\n", tgt_uuid->uuid);
 505
 506         lmv_init_lock(lmv);
 507
 508         if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
 509                 lmv_init_unlock(lmv);
 510                 CERROR("Can't add %s, LMV module compiled for %d MDCs. "
 511                        "That many MDCs already configured.\n",
 512                        tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
 513                 RETURN(-EINVAL);
 514         }
 515         if (lmv->desc.ld_tgt_count == 0) {
 516                 struct obd_device *mdc_obd;
 517
 518                 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
 519                                                 &obd->obd_uuid);
 520                 if (!mdc_obd) {
 521                         lmv_init_unlock(lmv);
 522                         CERROR("Target %s not attached\n", tgt_uuid->uuid);
 523                         RETURN(-EINVAL);
 524                 }
 525
 526                 rc = obd_llog_init(obd, &obd->obd_olg, mdc_obd, 0, NULL, tgt_uuid);
 527                 if (rc) {
 528                         lmv_init_unlock(lmv);
 529                         CERROR("lmv failed to setup llogging subsystems\n");
 530                 }
 531         }
 532         spin_lock(&lmv->lmv_lock);
 533         tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
 534         tgt->ltd_uuid = *tgt_uuid;
 535         spin_unlock(&lmv->lmv_lock);
 536
 537         if (lmv->connected) {
 538                 rc = lmv_connect_mdc(obd, tgt);
 539                 if (rc) {
 540                         spin_lock(&lmv->lmv_lock);
 541                         lmv->desc.ld_tgt_count--;
 542                         memset(tgt, 0, sizeof(*tgt));
 543                         spin_unlock(&lmv->lmv_lock);
 544                 } else {
 545                         int easize = sizeof(struct lmv_stripe_md) +
 546                                      lmv->desc.ld_tgt_count *
 547                                      sizeof(struct lu_fid);
 548                         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 549                 }
 550         }
 551
 552         lmv_init_unlock(lmv);
 553         RETURN(rc);
 554 }
 555
 556 int lmv_check_connect(struct obd_device *obd)
 557 {
 558         struct lmv_obd       *lmv = &obd->u.lmv;
 559         struct lmv_tgt_desc  *tgt;
 560         int                   i;
 561         int                   rc;
 562         int                   easize;
 563         ENTRY;
 564
 565         if (lmv->connected)
 566                 RETURN(0);
 567
 568         lmv_init_lock(lmv);
 569         if (lmv->connected) {
 570                 lmv_init_unlock(lmv);
 571                 RETURN(0);
 572         }
 573
 574         if (lmv->desc.ld_tgt_count == 0) {
 575                 CERROR("%s: no targets configured.\n", obd->obd_name);
 576                 RETURN(-EINVAL);
 577         }
 578
 579         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 580                lmv->cluuid.uuid, obd->obd_name);
 581
 582         LASSERT(lmv->tgts != NULL);
 583
 584         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 585                 rc = lmv_connect_mdc(obd, tgt);
 586                 if (rc)
 587                         GOTO(out_disc, rc);
 588         }
 589
 590         lmv_set_timeouts(obd);
 591         class_export_put(lmv->exp);
 592         lmv->connected = 1;
 593         easize = lmv_get_easize(lmv);
 594         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 595         lmv_init_unlock(lmv);
 596         RETURN(0);
 597
 598  out_disc:
 599         while (i-- > 0) {
 600                 int rc2;
 601                 --tgt;
 602                 tgt->ltd_active = 0;
 603                 if (tgt->ltd_exp) {
 604                         --lmv->desc.ld_active_tgt_count;
 605                         rc2 = obd_disconnect(tgt->ltd_exp);
 606                         if (rc2) {
 607                                 CERROR("LMV target %s disconnect on "
 608                                        "MDC idx %d: error %d\n",
 609                                        tgt->ltd_uuid.uuid, i, rc2);
 610                         }
 611                 }
 612         }
 613         class_disconnect(lmv->exp);
 614         lmv_init_unlock(lmv);
 615         RETURN(rc);
 616 }
 617
 618 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 619 {
 620 #ifdef __KERNEL__
 621         struct proc_dir_entry  *lmv_proc_dir;
 622 #endif
 623         struct lmv_obd         *lmv = &obd->u.lmv;
 624         struct obd_device      *mdc_obd;
 625         int                     rc;
 626         ENTRY;
 627
 628         LASSERT(tgt != NULL);
 629         LASSERT(obd != NULL);
 630
 631         mdc_obd = class_exp2obd(tgt->ltd_exp);
 632
 633         if (mdc_obd)
 634                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 635
 636 #ifdef __KERNEL__
 637         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 638         if (lmv_proc_dir) {
 639                 struct proc_dir_entry *mdc_symlink;
 640
 641                 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
 642                 if (mdc_symlink) {
 643                         lprocfs_remove(&mdc_symlink);
 644                 } else {
 645                         CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
 646                                obd->obd_type->typ_name, obd->obd_name,
 647                                mdc_obd->obd_name);
 648                 }
 649         }
 650 #endif
 651         rc = obd_fid_fini(tgt->ltd_exp);
 652         if (rc)
 653                 CERROR("Can't finanize fids factory\n");
 654
 655         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 656                tgt->ltd_exp->exp_obd->obd_name,
 657                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 658
 659         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 660         rc = obd_disconnect(tgt->ltd_exp);
 661         if (rc) {
 662                 if (tgt->ltd_active) {
 663                         CERROR("Target %s disconnect error %d\n",
 664                                tgt->ltd_uuid.uuid, rc);
 665                 }
 666         }
 667
 668         lmv_activate_target(lmv, tgt, 0);
 669         tgt->ltd_exp = NULL;
 670         RETURN(0);
 671 }
 672
 673 static int lmv_disconnect(struct obd_export *exp)
 674 {
 675         struct obd_device     *obd = class_exp2obd(exp);
 676 #ifdef __KERNEL__
 677         struct proc_dir_entry *lmv_proc_dir;
 678 #endif
 679         struct lmv_obd        *lmv = &obd->u.lmv;
 680         int                    rc;
 681         int                    i;
 682         ENTRY;
 683
 684         if (!lmv->tgts)
 685                 goto out_local;
 686
 687         /*
 688          * Only disconnect the underlying layers on the final disconnect.
 689          */
 690         lmv->refcount--;
 691         if (lmv->refcount != 0)
 692                 goto out_local;
 693
 694         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 695                 if (lmv->tgts[i].ltd_exp == NULL)
 696                         continue;
 697                 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
 698         }
 699
 700 #ifdef __KERNEL__
 701         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 702         if (lmv_proc_dir) {
 703                 lprocfs_remove(&lmv_proc_dir);
 704         } else {
 705                 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
 706                        obd->obd_type->typ_name, obd->obd_name);
 707         }
 708 #endif
 709
 710 out_local:
 711         /*
 712          * This is the case when no real connection is established by
 713          * lmv_check_connect().
 714          */
 715         if (!lmv->connected)
 716                 class_export_put(exp);
 717         rc = class_disconnect(exp);
 718         if (lmv->refcount == 0)
 719                 lmv->connected = 0;
 720         RETURN(rc);
 721 }
 722
 723 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 724                          int len, void *karg, void *uarg)
 725 {
 726         struct obd_device    *obddev = class_exp2obd(exp);
 727         struct lmv_obd       *lmv = &obddev->u.lmv;
 728         int                   i = 0;
 729         int                   rc = 0;
 730         int                   set = 0;
 731         int                   count = lmv->desc.ld_tgt_count;
 732         ENTRY;
 733
 734         if (count == 0)
 735                 RETURN(-ENOTTY);
 736
 737         switch (cmd) {
 738         case IOC_OBD_STATFS: {
 739                 struct obd_ioctl_data *data = karg;
 740                 struct obd_device *mdc_obd;
 741                 struct obd_statfs stat_buf = {0};
 742                 __u32 index;
 743
 744                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 745                 LASSERT(data->ioc_plen1 == sizeof(struct obd_statfs));
 746
 747                 if ((index >= count))
 748                         RETURN(-ENODEV);
 749
 750                 if (!lmv->tgts[index].ltd_active)
 751                         RETURN(-ENODATA);
 752
 753                 mdc_obd = class_exp2obd(lmv->tgts[index].ltd_exp);
 754                 if (!mdc_obd)
 755                         RETURN(-EINVAL);
 756
 757                 rc = obd_statfs(mdc_obd, &stat_buf,
 758                                 cfs_time_current_64() - HZ, 0);
 759                 if (rc)
 760                         RETURN(rc);
 761                 if (copy_to_user(data->ioc_pbuf1, &stat_buf, data->ioc_plen1))
 762                         RETURN(-EFAULT);
 763                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 764                                   data->ioc_plen2))
 765                         RETURN(-EFAULT);
 766                 break;
 767         }
 768         case OBD_IOC_QUOTACTL: {
 769                 struct if_quotactl *qctl = karg;
 770                 struct lmv_tgt_desc *tgt = NULL;
 771                 struct obd_quotactl *oqctl;
 772
 773                 if (qctl->qc_valid == QC_MDTIDX) {
 774                         if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
 775                                 RETURN(-EINVAL);
 776
 777                         tgt = &lmv->tgts[qctl->qc_idx];
 778                         if (!tgt->ltd_exp)
 779                                 RETURN(-EINVAL);
 780                 } else if (qctl->qc_valid == QC_UUID) {
 781                         for (i = 0; i < count; i++) {
 782                                 tgt = &lmv->tgts[i];
 783                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 784                                                      &qctl->obd_uuid))
 785                                         continue;
 786
 787                                 if (tgt->ltd_exp == NULL)
 788                                         RETURN(-EINVAL);
 789
 790                                 break;
 791                         }
 792                 } else {
 793                         RETURN(-EINVAL);
 794                 }
 795
 796                 if (i >= count)
 797                         RETURN(-EAGAIN);
 798
 799                 LASSERT(tgt && tgt->ltd_exp);
 800                 OBD_ALLOC_PTR(oqctl);
 801                 if (!oqctl)
 802                         RETURN(-ENOMEM);
 803
 804                 QCTL_COPY(oqctl, qctl);
 805                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 806                 if (rc == 0) {
 807                         QCTL_COPY(qctl, oqctl);
 808                         qctl->qc_valid = QC_MDTIDX;
 809                         qctl->obd_uuid = tgt->ltd_uuid;
 810                 }
 811                 OBD_FREE_PTR(oqctl);
 812                 break;
 813         }
 814         default : {
 815                 for (i = 0; i < count; i++) {
 816                         int err;
 817
 818                         if (lmv->tgts[i].ltd_exp == NULL)
 819                                 continue;
 820
 821                         err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
 822                                             karg, uarg);
 823                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
 824                                 RETURN(err);
 825                         } else if (err) {
 826                                 if (lmv->tgts[i].ltd_active) {
 827                                         CERROR("error: iocontrol MDC %s on MDT"
 828                                                "idx %d cmd %x: err = %d\n",
 829                                                 lmv->tgts[i].ltd_uuid.uuid,
 830                                                 i, cmd, err);
 831                                         if (!rc)
 832                                                 rc = err;
 833                                 }
 834                         } else
 835                                 set = 1;
 836                 }
 837                 if (!set && !rc)
 838                         rc = -EIO;
 839         }
 840         }
 841         RETURN(rc);
 842 }
 843
 844 static int lmv_all_chars_policy(int count, const char *name,
 845                                 int len)
 846 {
 847         unsigned int c = 0;
 848
 849         while (len > 0)
 850                 c += name[--len];
 851         c = c % count;
 852         return c;
 853 }
 854
 855 static int lmv_nid_policy(struct lmv_obd *lmv)
 856 {
 857         struct obd_import *imp;
 858         __u32              id;
 859
 860         /*
 861          * XXX: To get nid we assume that underlying obd device is mdc.
 862          */
 863         imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
 864         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
 865         return id % lmv->desc.ld_tgt_count;
 866 }
 867
 868 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
 869                           placement_policy_t placement)
 870 {
 871         switch (placement) {
 872         case PLACEMENT_CHAR_POLICY:
 873                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
 874                                             op_data->op_name,
 875                                             op_data->op_namelen);
 876         case PLACEMENT_NID_POLICY:
 877                 return lmv_nid_policy(lmv);
 878
 879         default:
 880                 break;
 881         }
 882
 883         CERROR("Unsupported placement policy %x\n", placement);
 884         return -EINVAL;
 885 }
 886
 887 /**
 888  * This is _inode_ placement policy function (not name).
 889  */
 890 static int lmv_placement_policy(struct obd_device *obd,
 891                                 struct md_op_data *op_data,
 892                                 mdsno_t *mds)
 893 {
 894         struct lmv_obd          *lmv = &obd->u.lmv;
 895         struct lmv_object       *obj;
 896         int                      rc;
 897         ENTRY;
 898
 899         LASSERT(mds != NULL);
 900
 901         if (lmv->desc.ld_tgt_count == 1) {
 902                 *mds = 0;
 903                 RETURN(0);
 904         }
 905
 906         /*
 907          * Allocate new fid on target according to operation type and parent
 908          * home mds.
 909          */
 910         obj = lmv_object_find(obd, &op_data->op_fid1);
 911         if (obj != NULL || op_data->op_name == NULL ||
 912             op_data->op_opc != LUSTRE_OPC_MKDIR) {
 913                 /*
 914                  * Allocate fid for non-dir or for null name or for case parent
 915                  * dir is split.
 916                  */
 917                 if (obj) {
 918                         lmv_object_put(obj);
 919
 920                         /*
 921                          * If we have this flag turned on, and we see that
 922                          * parent dir is split, this means, that caller did not
 923                          * notice split yet. This is race and we would like to
 924                          * let caller know that.
 925                          */
 926                         if (op_data->op_bias & MDS_CHECK_SPLIT)
 927                                 RETURN(-ERESTART);
 928                 }
 929
 930                 /*
 931                  * Allocate new fid on same mds where parent fid is located and
 932                  * where operation will be sent. In case of split dir, ->op_fid1
 933                  * and ->op_mds here will contain fid and mds of slave directory
 934                  * object (assigned by caller).
 935                  */
 936                 *mds = op_data->op_mds;
 937                 rc = 0;
 938         } else {
 939                 /*
 940                  * Parent directory is not split and we want to create a
 941                  * directory in it. Let's calculate where to place it according
 942                  * to operation data @op_data.
 943                  */
 944                 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
 945                 rc = 0;
 946         }
 947
 948         if (rc) {
 949                 CERROR("Can't choose MDS, err = %d\n", rc);
 950         } else {
 951                 LASSERT(*mds < lmv->desc.ld_tgt_count);
 952         }
 953
 954         RETURN(rc);
 955 }
 956
 957 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
 958                     mdsno_t mds)
 959 {
 960         struct lmv_tgt_desc *tgt;
 961         int                  rc;
 962         ENTRY;
 963
 964         tgt = lmv_get_target(lmv, mds);
 965
 966         /*
 967          * New seq alloc and FLD setup should be atomic. Otherwise we may find
 968          * on server that seq in new allocated fid is not yet known.
 969          */
 970         down(&tgt->ltd_fid_sem);
 971
 972         if (!tgt->ltd_active)
 973                 GOTO(out, rc = -ENODEV);
 974
 975         /*
 976          * Asking underlaying tgt layer to allocate new fid.
 977          */
 978         rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
 979         if (rc > 0) {
 980                 LASSERT(fid_is_sane(fid));
 981
 982                 /*
 983                  * Client switches to new sequence, setup FLD.
 984                  */
 985                 rc = fld_client_create(&lmv->lmv_fld, fid_seq(fid),
 986                                        mds, NULL);
 987                 if (rc) {
 988                         /*
 989                          * Delete just allocated fid sequence in case
 990                          * of fail back.
 991                          */
 992                         CERROR("Can't create fld entry, rc %d\n", rc);
 993                         obd_fid_delete(tgt->ltd_exp, NULL);
 994                 }
 995         }
 996
 997         EXIT;
 998 out:
 999         up(&tgt->ltd_fid_sem);
1000         return rc;
1001 }
1002
1003 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
1004                   struct md_op_data *op_data)
1005 {
1006         struct obd_device     *obd = class_exp2obd(exp);
1007         struct lmv_obd        *lmv = &obd->u.lmv;
1008         mdsno_t                mds;
1009         int                    rc;
1010         ENTRY;
1011
1012         LASSERT(op_data != NULL);
1013         LASSERT(fid != NULL);
1014
1015         rc = lmv_placement_policy(obd, op_data, &mds);
1016         if (rc) {
1017                 CERROR("Can't get target for allocating fid, "
1018                        "rc %d\n", rc);
1019                 RETURN(rc);
1020         }
1021
1022         rc = __lmv_fid_alloc(lmv, fid, mds);
1023         if (rc) {
1024                 CERROR("Can't alloc new fid, rc %d\n", rc);
1025                 RETURN(rc);
1026         }
1027
1028         RETURN(rc);
1029 }
1030
1031 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1032 {
1033         ENTRY;
1034         LASSERT(exp != NULL && fid != NULL);
1035         if (lmv_object_delete(exp, fid)) {
1036                 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1037                        PFID(fid));
1038         }
1039         RETURN(0);
1040 }
1041
1042 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1043 {
1044         struct lmv_obd             *lmv = &obd->u.lmv;
1045         struct lprocfs_static_vars  lvars;
1046         struct lmv_desc            *desc;
1047         int                         rc;
1048         int                         i = 0;
1049         ENTRY;
1050
1051         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1052                 CERROR("LMV setup requires a descriptor\n");
1053                 RETURN(-EINVAL);
1054         }
1055
1056         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1057         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1058                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1059                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1060                 RETURN(-EINVAL);
1061         }
1062
1063         lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1064
1065         OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1066         if (lmv->tgts == NULL)
1067                 RETURN(-ENOMEM);
1068
1069         for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1070                 sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
1071                 lmv->tgts[i].ltd_idx = i;
1072         }
1073
1074         lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1075
1076         OBD_ALLOC(lmv->datas, lmv->datas_size);
1077         if (lmv->datas == NULL)
1078                 GOTO(out_free_tgts, rc = -ENOMEM);
1079
1080         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1081         lmv->desc.ld_tgt_count = 0;
1082         lmv->desc.ld_active_tgt_count = 0;
1083         lmv->max_cookiesize = 0;
1084         lmv->max_def_easize = 0;
1085         lmv->max_easize = 0;
1086         lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1087
1088         spin_lock_init(&lmv->lmv_lock);
1089         sema_init(&lmv->init_sem, 1);
1090
1091         rc = lmv_object_setup(obd);
1092         if (rc) {
1093                 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1094                 GOTO(out_free_datas, rc);
1095         }
1096
1097         lprocfs_lmv_init_vars(&lvars);
1098         lprocfs_obd_setup(obd, lvars.obd_vars);
1099 #ifdef LPROCFS
1100         {
1101                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1102                                         0444, &lmv_proc_target_fops, obd);
1103                 if (rc)
1104                         CWARN("Error adding target_obd_stats file (%d)\n", rc);
1105        }
1106 #endif
1107         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1108                              LUSTRE_CLI_FLD_HASH_DHT);
1109         if (rc) {
1110                 CERROR("Can't init FLD, err %d\n", rc);
1111                 GOTO(out_free_datas, rc);
1112         }
1113
1114         RETURN(0);
1115
1116 out_free_datas:
1117         OBD_FREE(lmv->datas, lmv->datas_size);
1118         lmv->datas = NULL;
1119 out_free_tgts:
1120         OBD_FREE(lmv->tgts, lmv->tgts_size);
1121         lmv->tgts = NULL;
1122         return rc;
1123 }
1124
1125 static int lmv_cleanup(struct obd_device *obd)
1126 {
1127         struct lmv_obd   *lmv = &obd->u.lmv;
1128         ENTRY;
1129
1130         fld_client_fini(&lmv->lmv_fld);
1131         lprocfs_obd_cleanup(obd);
1132         lmv_object_cleanup(obd);
1133         OBD_FREE(lmv->datas, lmv->datas_size);
1134         OBD_FREE(lmv->tgts, lmv->tgts_size);
1135
1136         RETURN(0);
1137 }
1138
1139 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1140 {
1141         struct lustre_cfg     *lcfg = buf;
1142         struct obd_uuid        tgt_uuid;
1143         int                    rc;
1144         ENTRY;
1145
1146         switch(lcfg->lcfg_command) {
1147         case LCFG_ADD_MDC:
1148                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1149                         GOTO(out, rc = -EINVAL);
1150
1151                 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1152                 rc = lmv_add_target(obd, &tgt_uuid);
1153                 GOTO(out, rc);
1154         default: {
1155                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1156                 GOTO(out, rc = -EINVAL);
1157         }
1158         }
1159 out:
1160         RETURN(rc);
1161 }
1162
1163 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1164                       __u64 max_age, __u32 flags)
1165 {
1166         struct lmv_obd        *lmv = &obd->u.lmv;
1167         struct obd_statfs     *temp;
1168         int                    rc = 0;
1169         int                    i;
1170         ENTRY;
1171
1172         rc = lmv_check_connect(obd);
1173         if (rc)
1174                 RETURN(rc);
1175
1176         OBD_ALLOC(temp, sizeof(*temp));
1177         if (temp == NULL)
1178                 RETURN(-ENOMEM);
1179
1180         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1181                 if (lmv->tgts[i].ltd_exp == NULL)
1182                         continue;
1183
1184                 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1185                                 max_age, flags);
1186                 if (rc) {
1187                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1188                                lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1189                                rc);
1190                         GOTO(out_free_temp, rc);
1191                 }
1192                 if (i == 0) {
1193                         *osfs = *temp;
1194                 } else {
1195                         osfs->os_bavail += temp->os_bavail;
1196                         osfs->os_blocks += temp->os_blocks;
1197                         osfs->os_ffree += temp->os_ffree;
1198                         osfs->os_files += temp->os_files;
1199                 }
1200         }
1201
1202         EXIT;
1203 out_free_temp:
1204         OBD_FREE(temp, sizeof(*temp));
1205         return rc;
1206 }
1207
1208 static int lmv_getstatus(struct obd_export *exp,
1209                          struct lu_fid *fid,
1210                          struct obd_capa **pc)
1211 {
1212         struct obd_device    *obd = exp->exp_obd;
1213         struct lmv_obd       *lmv = &obd->u.lmv;
1214         int                   rc;
1215         ENTRY;
1216
1217         rc = lmv_check_connect(obd);
1218         if (rc)
1219                 RETURN(rc);
1220
1221         rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1222         RETURN(rc);
1223 }
1224
1225 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1226                         struct obd_capa *oc, obd_valid valid, const char *name,
1227                         const char *input, int input_size, int output_size,
1228                         int flags, struct ptlrpc_request **request)
1229 {
1230         struct obd_device      *obd = exp->exp_obd;
1231         struct lmv_obd         *lmv = &obd->u.lmv;
1232         struct lmv_tgt_desc    *tgt;
1233         int                     rc;
1234         ENTRY;
1235
1236         rc = lmv_check_connect(obd);
1237         if (rc)
1238                 RETURN(rc);
1239
1240         tgt = lmv_find_target(lmv, fid);
1241         if (IS_ERR(tgt))
1242                 RETURN(PTR_ERR(tgt));
1243
1244         rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1245                          input_size, output_size, flags, request);
1246
1247         RETURN(rc);
1248 }
1249
1250 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1251                         struct obd_capa *oc, obd_valid valid, const char *name,
1252                         const char *input, int input_size, int output_size,
1253                         int flags, __u32 suppgid,
1254                         struct ptlrpc_request **request)
1255 {
1256         struct obd_device      *obd = exp->exp_obd;
1257         struct lmv_obd         *lmv = &obd->u.lmv;
1258         struct lmv_tgt_desc    *tgt;
1259         int                     rc;
1260         ENTRY;
1261
1262         rc = lmv_check_connect(obd);
1263         if (rc)
1264                 RETURN(rc);
1265
1266         tgt = lmv_find_target(lmv, fid);
1267         if (IS_ERR(tgt))
1268                 RETURN(PTR_ERR(tgt));
1269
1270         rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1271                          input_size, output_size, flags, suppgid,
1272                          request);
1273
1274         RETURN(rc);
1275 }
1276
1277 static int lmv_getattr(struct obd_export *exp, const struct lu_fid *fid,
1278                        struct obd_capa *oc, obd_valid valid, int ea_size,
1279                        struct ptlrpc_request **request)
1280 {
1281         struct obd_device       *obd = exp->exp_obd;
1282         struct lmv_obd          *lmv = &obd->u.lmv;
1283         struct lmv_tgt_desc     *tgt;
1284         struct lmv_object       *obj;
1285         int                      rc;
1286         int                      i;
1287         ENTRY;
1288
1289         rc = lmv_check_connect(obd);
1290         if (rc)
1291                 RETURN(rc);
1292
1293         tgt = lmv_find_target(lmv, fid);
1294         if (IS_ERR(tgt))
1295                 RETURN(PTR_ERR(tgt));
1296
1297         rc = md_getattr(tgt->ltd_exp, fid, oc, valid, ea_size, request);
1298         if (rc)
1299                 RETURN(rc);
1300
1301         obj = lmv_object_find_lock(obd, fid);
1302
1303         CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(fid),
1304                obj ? "(split)" : "");
1305
1306         /*
1307          * If object is split, then we loop over all the slaves and gather size
1308          * attribute. In ideal world we would have to gather also mds field from
1309          * all slaves, as object is spread over the cluster and this is
1310          * definitely interesting information and it is not good to loss it,
1311          * but...
1312          */
1313         if (obj) {
1314                 struct mdt_body *body;
1315
1316                 if (*request == NULL) {
1317                         lmv_object_put(obj);
1318                         RETURN(rc);
1319                 }
1320
1321                 body = req_capsule_server_get(&(*request)->rq_pill,
1322                                               &RMF_MDT_BODY);
1323                 LASSERT(body != NULL);
1324
1325                 for (i = 0; i < obj->lo_objcount; i++) {
1326                         if (lmv->tgts[i].ltd_exp == NULL) {
1327                                 CWARN("%s: NULL export for %d\n",
1328                                       obd->obd_name, i);
1329                                 continue;
1330                         }
1331
1332                         /*
1333                          * Skip master object.
1334                          */
1335                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1336                                 continue;
1337
1338                         body->size += obj->lo_stripes[i].ls_size;
1339                 }
1340
1341                 lmv_object_put_unlock(obj);
1342         }
1343
1344         RETURN(rc);
1345 }
1346
1347 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1348                              ldlm_iterator_t it, void *data)
1349 {
1350         struct obd_device   *obd = exp->exp_obd;
1351         struct lmv_obd      *lmv = &obd->u.lmv;
1352         int                  i;
1353         int                  rc;
1354         ENTRY;
1355
1356         rc = lmv_check_connect(obd);
1357         if (rc)
1358                 RETURN(rc);
1359
1360         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1361
1362         /*
1363          * With CMD every object can have two locks in different namespaces:
1364          * lookup lock in space of mds storing direntry and update/open lock in
1365          * space of mds storing inode.
1366          */
1367         for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1368                 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1369
1370         RETURN(0);
1371 }
1372
1373 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1374                      struct md_open_data *mod, struct ptlrpc_request **request)
1375 {
1376         struct obd_device     *obd = exp->exp_obd;
1377         struct lmv_obd        *lmv = &obd->u.lmv;
1378         struct lmv_tgt_desc   *tgt;
1379         int                    rc;
1380         ENTRY;
1381
1382         rc = lmv_check_connect(obd);
1383         if (rc)
1384                 RETURN(rc);
1385
1386         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1387         if (IS_ERR(tgt))
1388                 RETURN(PTR_ERR(tgt));
1389
1390         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1391         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1392         RETURN(rc);
1393 }
1394
1395 /**
1396  * Called in the case MDS returns -ERESTART on create on open, what means that
1397  * directory is split and its LMV presentation object has to be updated.
1398  */
1399 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1400 {
1401         struct obd_device       *obd = exp->exp_obd;
1402         struct lmv_obd          *lmv = &obd->u.lmv;
1403         struct ptlrpc_request   *req = NULL;
1404         struct lmv_tgt_desc     *tgt;
1405         struct lmv_object       *obj;
1406         struct lustre_md         md;
1407         int                      mealen;
1408         int                      rc;
1409         __u64                    valid;
1410         ENTRY;
1411
1412         md.mea = NULL;
1413         mealen = lmv_get_easize(lmv);
1414
1415         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1416
1417         tgt = lmv_find_target(lmv, fid);
1418         if (IS_ERR(tgt))
1419                 RETURN(PTR_ERR(tgt));
1420
1421         /*
1422          * Time to update mea of parent fid.
1423          */
1424         rc = md_getattr(tgt->ltd_exp, fid, NULL, valid, mealen, &req);
1425         if (rc) {
1426                 CERROR("md_getattr() failed, error %d\n", rc);
1427                 GOTO(cleanup, rc);
1428         }
1429
1430         rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1431         if (rc) {
1432                 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1433                 GOTO(cleanup, rc);
1434         }
1435
1436         if (md.mea == NULL)
1437                 GOTO(cleanup, rc = -ENODATA);
1438
1439         obj = lmv_object_create(exp, fid, md.mea);
1440         if (IS_ERR(obj))
1441                 rc = PTR_ERR(obj);
1442         else
1443                 lmv_object_put(obj);
1444
1445         obd_free_memmd(exp, (void *)&md.mea);
1446         EXIT;
1447 cleanup:
1448         if (req)
1449                 ptlrpc_req_finished(req);
1450         return rc;
1451 }
1452
1453 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1454                const void *data, int datalen, int mode, __u32 uid,
1455                __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1456                struct ptlrpc_request **request)
1457 {
1458         struct obd_device       *obd = exp->exp_obd;
1459         struct lmv_obd          *lmv = &obd->u.lmv;
1460         struct lmv_tgt_desc     *tgt;
1461         struct lmv_object       *obj;
1462         int                      rc;
1463         int                      loop = 0;
1464         int                      sidx;
1465         ENTRY;
1466
1467         rc = lmv_check_connect(obd);
1468         if (rc)
1469                 RETURN(rc);
1470
1471         if (!lmv->desc.ld_active_tgt_count)
1472                 RETURN(-EIO);
1473 repeat:
1474         ++loop;
1475         LASSERT(loop <= 2);
1476
1477         obj = lmv_object_find(obd, &op_data->op_fid1);
1478         if (obj) {
1479                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1480                                        op_data->op_name, op_data->op_namelen);
1481                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1482                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1483                 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1484                 tgt = lmv_get_target(lmv, op_data->op_mds);
1485                 lmv_object_put(obj);
1486         } else {
1487                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1488                 op_data->op_bias |= MDS_CHECK_SPLIT;
1489                 op_data->op_mds = tgt->ltd_idx;
1490         }
1491
1492         if (IS_ERR(tgt))
1493                 RETURN(PTR_ERR(tgt));
1494
1495         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1496         if (rc == -ERESTART)
1497                 goto repeat;
1498         else if (rc)
1499                 RETURN(rc);
1500
1501         CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #"LPU64"\n",
1502                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1503                op_data->op_mds);
1504
1505         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1506         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1507                        cap_effective, rdev, request);
1508         if (rc == 0) {
1509                 if (*request == NULL)
1510                         RETURN(rc);
1511                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1512         } else if (rc == -ERESTART) {
1513                 LASSERT(*request != NULL);
1514                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1515                           "Got -ERESTART during create!\n");
1516                 ptlrpc_req_finished(*request);
1517                 *request = NULL;
1518
1519                 /*
1520                  * Directory got split. Time to update local object and repeat
1521                  * the request with proper MDS.
1522                  */
1523                 rc = lmv_handle_split(exp, &op_data->op_fid1);
1524                 if (rc == 0) {
1525                         rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1526                                                  op_data, &op_data->op_fid2);
1527                         if (rc)
1528                                 RETURN(rc);
1529                         goto repeat;
1530                 }
1531         }
1532         RETURN(rc);
1533 }
1534
1535 static int lmv_done_writing(struct obd_export *exp,
1536                             struct md_op_data *op_data,
1537                             struct md_open_data *mod)
1538 {
1539         struct obd_device     *obd = exp->exp_obd;
1540         struct lmv_obd        *lmv = &obd->u.lmv;
1541         struct lmv_tgt_desc   *tgt;
1542         int                    rc;
1543         ENTRY;
1544
1545         rc = lmv_check_connect(obd);
1546         if (rc)
1547                 RETURN(rc);
1548
1549         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1550         if (IS_ERR(tgt))
1551                 RETURN(PTR_ERR(tgt));
1552
1553         rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1554         RETURN(rc);
1555 }
1556
1557 static int
1558 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1559                    struct lookup_intent *it, struct md_op_data *op_data,
1560                    struct lustre_handle *lockh, void *lmm, int lmmsize)
1561 {
1562         struct obd_device     *obd = exp->exp_obd;
1563         struct lmv_obd        *lmv = &obd->u.lmv;
1564         struct lmv_stripe_md  *mea = op_data->op_mea1;
1565         struct md_op_data     *op_data2;
1566         struct lmv_tgt_desc   *tgt;
1567         int                    i;
1568         int                    rc = 0;
1569         ENTRY;
1570
1571         OBD_ALLOC_PTR(op_data2);
1572         if (op_data2 == NULL)
1573                 RETURN(-ENOMEM);
1574
1575         LASSERT(mea != NULL);
1576         for (i = 0; i < mea->mea_count; i++) {
1577                 memset(op_data2, 0, sizeof(*op_data2));
1578                 op_data2->op_fid1 = mea->mea_ids[i];
1579                 op_data2->op_bias = 0;
1580
1581                 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1582                 if (IS_ERR(tgt))
1583                         GOTO(cleanup, rc = PTR_ERR(tgt));
1584
1585                 if (tgt->ltd_exp == NULL)
1586                         continue;
1587
1588                 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1589                                 lockh + i, lmm, lmmsize, NULL, 0);
1590
1591                 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1592                        PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1593
1594                 if (rc)
1595                         GOTO(cleanup, rc);
1596
1597                 if (it->d.lustre.it_data) {
1598                         struct ptlrpc_request *req;
1599                         req = (struct ptlrpc_request *)it->d.lustre.it_data;
1600                         ptlrpc_req_finished(req);
1601                 }
1602
1603                 if (it->d.lustre.it_status)
1604                         GOTO(cleanup, rc = it->d.lustre.it_status);
1605         }
1606
1607         EXIT;
1608 cleanup:
1609         OBD_FREE_PTR(op_data2);
1610
1611         if (rc != 0) {
1612                 /*
1613                  * Drop all taken locks.
1614                  */
1615                 while (--i >= 0) {
1616                         if (lockh[i].cookie)
1617                                 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1618                         lockh[i].cookie = 0;
1619                 }
1620         }
1621         return rc;
1622 }
1623
1624 static int
1625 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1626                    struct lookup_intent *it, struct md_op_data *op_data,
1627                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1628                    int extra_lock_flags)
1629 {
1630         struct ptlrpc_request      *req = it->d.lustre.it_data;
1631         struct obd_device          *obd = exp->exp_obd;
1632         struct lmv_obd             *lmv = &obd->u.lmv;
1633         struct lustre_handle        plock;
1634         struct lmv_tgt_desc        *tgt;
1635         struct md_op_data          *rdata;
1636         struct lu_fid               fid1;
1637         struct mdt_body            *body;
1638         int                         rc = 0;
1639         int                         pmode;
1640         ENTRY;
1641
1642         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1643         LASSERT(body != NULL);
1644
1645         if (!(body->valid & OBD_MD_MDS))
1646                 RETURN(0);
1647
1648         CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1649                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1650
1651         /*
1652          * We got LOOKUP lock, but we really need attrs.
1653          */
1654         pmode = it->d.lustre.it_lock_mode;
1655         LASSERT(pmode != 0);
1656         memcpy(&plock, lockh, sizeof(plock));
1657         it->d.lustre.it_lock_mode = 0;
1658         it->d.lustre.it_data = NULL;
1659         fid1 = body->fid1;
1660
1661         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1662         ptlrpc_req_finished(req);
1663
1664         tgt = lmv_find_target(lmv, &fid1);
1665         if (IS_ERR(tgt))
1666                 GOTO(out, rc = PTR_ERR(tgt));
1667
1668         OBD_ALLOC_PTR(rdata);
1669         if (rdata == NULL)
1670                 GOTO(out, rc = -ENOMEM);
1671
1672         rdata->op_fid1 = fid1;
1673         rdata->op_bias = MDS_CROSS_REF;
1674
1675         rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1676                         lmm, lmmsize, NULL, extra_lock_flags);
1677         OBD_FREE_PTR(rdata);
1678         EXIT;
1679 out:
1680         ldlm_lock_decref(&plock, pmode);
1681         return rc;
1682 }
1683
1684 static int
1685 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1686             struct lookup_intent *it, struct md_op_data *op_data,
1687             struct lustre_handle *lockh, void *lmm, int lmmsize,
1688             struct ptlrpc_request **req, int extra_lock_flags)
1689 {
1690         struct obd_device        *obd = exp->exp_obd;
1691         struct lmv_obd           *lmv = &obd->u.lmv;
1692         struct lmv_tgt_desc      *tgt;
1693         struct lmv_object        *obj;
1694         int                       sidx;
1695         int                       rc;
1696         ENTRY;
1697
1698         rc = lmv_check_connect(obd);
1699         if (rc)
1700                 RETURN(rc);
1701
1702         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1703                LL_IT2STR(it), PFID(&op_data->op_fid1));
1704
1705         if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1706                 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1707                                         lockh, lmm, lmmsize);
1708                 RETURN(rc);
1709         }
1710
1711         obj = lmv_object_find(obd, &op_data->op_fid1);
1712         if (obj && op_data->op_namelen) {
1713                 sidx = raw_name2idx(obj->lo_hashtype,
1714                                        obj->lo_objcount,
1715                                        (char *)op_data->op_name,
1716                                        op_data->op_namelen);
1717                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1718                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1719         } else {
1720                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1721         }
1722         if (obj)
1723                 lmv_object_put(obj);
1724
1725         if (IS_ERR(tgt))
1726                 RETURN(PTR_ERR(tgt));
1727
1728         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1729                LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1730
1731         rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1732                         lmm, lmmsize, req, extra_lock_flags);
1733
1734         if (rc == 0 && it && it->it_op == IT_OPEN) {
1735                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1736                                         lmm, lmmsize, extra_lock_flags);
1737         }
1738         RETURN(rc);
1739 }
1740
1741 static int
1742 lmv_getattr_name(struct obd_export *exp, const struct lu_fid *fid,
1743                  struct obd_capa *oc, const char *name, int namelen,
1744                  obd_valid valid, int ea_size, __u32 suppgid,
1745                  struct ptlrpc_request **request)
1746 {
1747         struct ptlrpc_request   *req = NULL;
1748         struct obd_device       *obd = exp->exp_obd;
1749         struct lmv_obd          *lmv = &obd->u.lmv;
1750         struct lu_fid            rid = *fid;
1751         struct lmv_tgt_desc     *tgt;
1752         struct mdt_body         *body;
1753         struct lmv_object       *obj;
1754         int                      rc;
1755         int                      loop = 0;
1756         int                      sidx;
1757         ENTRY;
1758
1759         rc = lmv_check_connect(obd);
1760         if (rc)
1761                 RETURN(rc);
1762
1763 repeat:
1764         ++loop;
1765         LASSERT(loop <= 2);
1766         obj = lmv_object_find(obd, &rid);
1767         if (obj) {
1768                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1769                                        name, namelen - 1);
1770                 rid = obj->lo_stripes[sidx].ls_fid;
1771                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1772                 valid &= ~OBD_MD_FLCKSPLIT;
1773                 lmv_object_put(obj);
1774         } else {
1775                 tgt = lmv_find_target(lmv, &rid);
1776                 valid |= OBD_MD_FLCKSPLIT;
1777         }
1778         if (IS_ERR(tgt))
1779                 RETURN(PTR_ERR(tgt));
1780
1781         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1782                namelen, name, PFID(fid), PFID(&rid), tgt->ltd_idx);
1783
1784         rc = md_getattr_name(tgt->ltd_exp, &rid, oc, name, namelen, valid,
1785                              ea_size, suppgid, request);
1786         if (rc == 0) {
1787                 body = req_capsule_server_get(&(*request)->rq_pill,
1788                                               &RMF_MDT_BODY);
1789                 LASSERT(body != NULL);
1790
1791                 if (body->valid & OBD_MD_MDS) {
1792                         rid = body->fid1;
1793                         CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1794                                PFID(&rid));
1795
1796                         tgt = lmv_find_target(lmv, &rid);
1797                         if (IS_ERR(tgt)) {
1798                                 ptlrpc_req_finished(*request);
1799                                 RETURN(PTR_ERR(tgt));
1800                         }
1801
1802                         rc = md_getattr_name(tgt->ltd_exp, &rid, NULL, NULL,
1803                                              1, valid | OBD_MD_FLCROSSREF,
1804                                              ea_size, suppgid, &req);
1805                         ptlrpc_req_finished(*request);
1806                         *request = req;
1807                 }
1808         } else if (rc == -ERESTART) {
1809                 LASSERT(*request != NULL);
1810                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1811                           "Got -ERESTART during getattr!\n");
1812                 ptlrpc_req_finished(*request);
1813                 *request = NULL;
1814
1815                 /*
1816                  * Directory got split. Time to update local object and repeat
1817                  * the request with proper MDS.
1818                  */
1819                 rc = lmv_handle_split(exp, &rid);
1820                 if (rc == 0)
1821                         goto repeat;
1822         }
1823         RETURN(rc);
1824 }
1825
1826 #define md_op_data_fid(op_data, fl)                     \
1827         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1828          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1829          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1830          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1831          NULL)
1832
1833 static int lmv_early_cancel_slaves(struct obd_export *exp,
1834                                    struct md_op_data *op_data, int op_tgt,
1835                                    ldlm_mode_t mode, int bits, int flag)
1836 {
1837         struct obd_device      *obd = exp->exp_obd;
1838         struct lmv_obd         *lmv = &obd->u.lmv;
1839         ldlm_policy_data_t      policy = {{0}};
1840         struct lu_fid          *op_fid;
1841         struct lu_fid          *st_fid;
1842         struct lmv_tgt_desc    *tgt;
1843         struct lmv_object      *obj;
1844         int                     rc = 0;
1845         int                     i;
1846         ENTRY;
1847
1848         op_fid = md_op_data_fid(op_data, flag);
1849         if (!fid_is_sane(op_fid))
1850                 RETURN(0);
1851
1852         obj = lmv_object_find(obd, op_fid);
1853         if (obj == NULL)
1854                 RETURN(-EALREADY);
1855
1856         policy.l_inodebits.bits = bits;
1857         for (i = 0; i < obj->lo_objcount; i++) {
1858                 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1859                 st_fid = &obj->lo_stripes[i].ls_fid;
1860                 if (op_tgt != tgt->ltd_idx) {
1861                         CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1862                                PFID(st_fid), tgt->ltd_idx);
1863                         rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1864                                               mode, LDLM_FL_ASYNC, NULL);
1865                         if (rc)
1866                                 GOTO(out_put_obj, rc);
1867                 } else {
1868                         CDEBUG(D_INODE,
1869                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1870                                op_tgt, PFID(st_fid));
1871                         /*
1872                          * Do not cancel locks for operation target, they will
1873                          * be handled later in underlaying layer when calling
1874                          * function we run on behalf of.
1875                          */
1876                         *op_fid = *st_fid;
1877                         op_data->op_flags |= flag;
1878                 }
1879         }
1880         EXIT;
1881 out_put_obj:
1882         lmv_object_put(obj);
1883         return rc;
1884 }
1885
1886 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1887                             int op_tgt, ldlm_mode_t mode, int bits, int flag)
1888 {
1889         struct lu_fid          *fid = md_op_data_fid(op_data, flag);
1890         struct obd_device      *obd = exp->exp_obd;
1891         struct lmv_obd         *lmv = &obd->u.lmv;
1892         struct lmv_tgt_desc    *tgt;
1893         ldlm_policy_data_t      policy = {{0}};
1894         struct lmv_object      *obj;
1895         int                     rc = 0;
1896         ENTRY;
1897
1898         if (!fid_is_sane(fid))
1899                 RETURN(0);
1900
1901         obj = lmv_object_find(obd, fid);
1902         if (obj) {
1903                 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1904                                              bits, flag);
1905                 lmv_object_put(obj);
1906         } else {
1907                 tgt = lmv_find_target(lmv, fid);
1908                 if (IS_ERR(tgt))
1909                         RETURN(PTR_ERR(tgt));
1910
1911                 if (tgt->ltd_idx != op_tgt) {
1912                         CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1913                         policy.l_inodebits.bits = bits;
1914                         rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1915                                               mode, LDLM_FL_ASYNC, NULL);
1916                 } else {
1917                         CDEBUG(D_INODE,
1918                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1919                                op_tgt, PFID(fid));
1920                         op_data->op_flags |= flag;
1921                         rc = 0;
1922                 }
1923
1924         }
1925         RETURN(rc);
1926 }
1927
1928 /*
1929  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1930  * op_data->op_fid2
1931  */
1932 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1933                     struct ptlrpc_request **request)
1934 {
1935         struct obd_device       *obd = exp->exp_obd;
1936         struct lmv_obd          *lmv = &obd->u.lmv;
1937         struct lmv_tgt_desc     *tgt;
1938         struct lmv_object       *obj;
1939         int                      rc;
1940         int                      loop = 0;
1941         mdsno_t                  mds;
1942         int                      sidx;
1943         ENTRY;
1944
1945         rc = lmv_check_connect(obd);
1946         if (rc)
1947                 RETURN(rc);
1948
1949 repeat:
1950         ++loop;
1951         LASSERT(loop <= 2);
1952         LASSERT(op_data->op_namelen != 0);
1953
1954         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
1955                PFID(&op_data->op_fid2), op_data->op_namelen,
1956                op_data->op_name, PFID(&op_data->op_fid1));
1957
1958         obj = lmv_object_find(obd, &op_data->op_fid2);
1959         if (obj) {
1960                 sidx = raw_name2idx(obj->lo_hashtype,
1961                                     obj->lo_objcount,
1962                                     op_data->op_name,
1963                                     op_data->op_namelen);
1964                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
1965                 mds = obj->lo_stripes[sidx].ls_mds;
1966                 lmv_object_put(obj);
1967         } else {
1968                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1969                 if (rc)
1970                         RETURN(rc);
1971         }
1972
1973         CDEBUG(D_INODE, "Forward to mds #"LPU64" ("DFID")\n",
1974                mds, PFID(&op_data->op_fid1));
1975
1976         op_data->op_fsuid = current->fsuid;
1977         op_data->op_fsgid = current->fsgid;
1978         op_data->op_cap = cfs_curproc_cap_pack();
1979         tgt = lmv_get_target(lmv, mds);
1980
1981         /*
1982          * Cancel UPDATE lock on child (fid1).
1983          */
1984         op_data->op_flags |= MF_MDC_CANCEL_FID2;
1985         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
1986                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1987         if (rc == 0)
1988                 rc = md_link(tgt->ltd_exp, op_data, request);
1989         if (rc == -ERESTART) {
1990                 LASSERT(*request != NULL);
1991                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1992                           "Got -ERESTART during link!\n");
1993                 ptlrpc_req_finished(*request);
1994                 *request = NULL;
1995
1996                 /*
1997                  * Directory got split. Time to update local object and repeat
1998                  * the request with proper MDS.
1999                  */
2000                 rc = lmv_handle_split(exp, &op_data->op_fid2);
2001                 if (rc == 0)
2002                         goto repeat;
2003         }
2004
2005         RETURN(rc);
2006 }
2007
2008 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2009                       const char *old, int oldlen, const char *new, int newlen,
2010                       struct ptlrpc_request **request)
2011 {
2012         struct obd_device       *obd = exp->exp_obd;
2013         struct lmv_obd          *lmv = &obd->u.lmv;
2014         struct lmv_tgt_desc     *src_tgt;
2015         struct lmv_tgt_desc     *tgt_tgt;
2016         int                      rc;
2017         int                      sidx;
2018         int                      loop = 0;
2019         struct lmv_object       *obj;
2020         mdsno_t                  mds1;
2021         mdsno_t                  mds2;
2022         ENTRY;
2023
2024         LASSERT(oldlen != 0);
2025
2026         CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2027                oldlen, old, PFID(&op_data->op_fid1),
2028                newlen, new, PFID(&op_data->op_fid2));
2029
2030         rc = lmv_check_connect(obd);
2031         if (rc)
2032                 RETURN(rc);
2033
2034 repeat:
2035         ++loop;
2036         LASSERT(loop <= 2);
2037         obj = lmv_object_find(obd, &op_data->op_fid1);
2038         if (obj) {
2039                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2040                                     (char *)old, oldlen);
2041                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2042                 mds1 = obj->lo_stripes[sidx].ls_mds;
2043                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2044                 lmv_object_put(obj);
2045         } else {
2046                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2047                 if (rc)
2048                         RETURN(rc);
2049         }
2050
2051         obj = lmv_object_find(obd, &op_data->op_fid2);
2052         if (obj) {
2053                 /*
2054                  * Directory is already split, so we have to forward request to
2055                  * the right MDS.
2056                  */
2057                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2058                                        (char *)new, newlen);
2059
2060                 mds2 = obj->lo_stripes[sidx].ls_mds;
2061                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2062                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2063                 lmv_object_put(obj);
2064         } else {
2065                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2066                 if (rc)
2067                         RETURN(rc);
2068         }
2069
2070         op_data->op_fsuid = current->fsuid;
2071         op_data->op_fsgid = current->fsgid;
2072         op_data->op_cap = cfs_curproc_cap_pack();
2073
2074         src_tgt = lmv_get_target(lmv, mds1);
2075         tgt_tgt = lmv_get_target(lmv, mds2);
2076
2077         /*
2078          * LOOKUP lock on src child (fid3) should also be cancelled for
2079          * src_tgt in mdc_rename.
2080          */
2081         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2082
2083         /*
2084          * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2085          * own target.
2086          */
2087         rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2088                               LCK_EX, MDS_INODELOCK_UPDATE,
2089                               MF_MDC_CANCEL_FID2);
2090
2091         /*
2092          * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2093          */
2094         if (rc == 0) {
2095                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2096                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2097                                       MF_MDC_CANCEL_FID4);
2098         }
2099
2100         /*
2101          * Cancel all the locks on tgt child (fid4).
2102          */
2103         if (rc == 0)
2104                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2105                                       LCK_EX, MDS_INODELOCK_FULL,
2106                                       MF_MDC_CANCEL_FID4);
2107
2108         if (rc == 0)
2109                 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2110                                new, newlen, request);
2111
2112         if (rc == -ERESTART) {
2113                 LASSERT(*request != NULL);
2114                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2115                           "Got -ERESTART during rename!\n");
2116                 ptlrpc_req_finished(*request);
2117                 *request = NULL;
2118
2119                 /*
2120                  * Directory got split. Time to update local object and repeat
2121                  * the request with proper MDS.
2122                  */
2123                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2124                 if (rc == 0)
2125                         goto repeat;
2126         }
2127         RETURN(rc);
2128 }
2129
2130 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2131                        void *ea, int ealen, void *ea2, int ea2len,
2132                        struct ptlrpc_request **request,
2133                        struct md_open_data **mod)
2134 {
2135         struct obd_device       *obd = exp->exp_obd;
2136         struct lmv_obd          *lmv = &obd->u.lmv;
2137         struct ptlrpc_request   *req;
2138         struct lmv_tgt_desc     *tgt;
2139         struct lmv_object       *obj;
2140         int                      rc = 0;
2141         int                      i;
2142         ENTRY;
2143
2144         rc = lmv_check_connect(obd);
2145         if (rc)
2146                 RETURN(rc);
2147
2148         obj = lmv_object_find(obd, &op_data->op_fid1);
2149
2150         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2151                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2152                obj ? ", split" : "");
2153
2154         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2155         if (obj) {
2156                 for (i = 0; i < obj->lo_objcount; i++) {
2157                         op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2158
2159                         tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2160                         if (IS_ERR(tgt)) {
2161                                 rc = PTR_ERR(tgt);
2162                                 break;
2163                         }
2164
2165                         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2166                                         ea2, ea2len, &req, mod);
2167
2168                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2169                                 /*
2170                                  * This is master object and this request should
2171                                  * be returned back to llite.
2172                                  */
2173                                 *request = req;
2174                         } else {
2175                                 ptlrpc_req_finished(req);
2176                         }
2177
2178                         if (rc)
2179                                 break;
2180                 }
2181                 lmv_object_put(obj);
2182         } else {
2183                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2184                 if (IS_ERR(tgt))
2185                         RETURN(PTR_ERR(tgt));
2186
2187                 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2188                                 ea2len, request, mod);
2189         }
2190         RETURN(rc);
2191 }
2192
2193 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2194                     struct obd_capa *oc, struct ptlrpc_request **request)
2195 {
2196         struct obd_device         *obd = exp->exp_obd;
2197         struct lmv_obd            *lmv = &obd->u.lmv;
2198         struct lmv_tgt_desc       *tgt;
2199         int                        rc;
2200         ENTRY;
2201
2202         rc = lmv_check_connect(obd);
2203         if (rc)
2204                 RETURN(rc);
2205
2206         tgt = lmv_find_target(lmv, fid);
2207         if (IS_ERR(tgt))
2208                 RETURN(PTR_ERR(tgt));
2209
2210         rc = md_sync(tgt->ltd_exp, fid, oc, request);
2211         RETURN(rc);
2212 }
2213
2214 /**
2215  * Main purpose of LMV blocking ast is to remove split directory LMV
2216  * presentation object (struct lmv_object) attached to the lock being revoked.
2217  */
2218 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2219                      void *data, int flag)
2220 {
2221         struct lustre_handle    lockh;
2222         struct lmv_object      *obj;
2223         int                     rc;
2224         ENTRY;
2225
2226         switch (flag) {
2227         case LDLM_CB_BLOCKING:
2228                 ldlm_lock2handle(lock, &lockh);
2229                 rc = ldlm_cli_cancel(&lockh);
2230                 if (rc < 0) {
2231                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2232                         RETURN(rc);
2233                 }
2234                 break;
2235         case LDLM_CB_CANCELING:
2236                 /*
2237                  * Time to drop cached attrs for split directory object
2238                  */
2239                 obj = lock->l_ast_data;
2240                 if (obj) {
2241                         CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2242                                ", master "DFID"\n",
2243                                lock->l_resource->lr_name.name[3] == 1 ?
2244                                "LOOKUP" : "UPDATE",
2245                                lock->l_resource->lr_name.name[0],
2246                                lock->l_resource->lr_name.name[1],
2247                                PFID(&obj->lo_fid));
2248                         lmv_object_put(obj);
2249                 }
2250                 break;
2251         default:
2252                 LBUG();
2253         }
2254         RETURN(0);
2255 }
2256
2257 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2258 {
2259         __u64         val;
2260
2261         val = le64_to_cpu(*hash);
2262         if (val < hash_adj)
2263                 val += MAX_HASH_SIZE;
2264         if (val != DIR_END_OFF)
2265                 *hash = cpu_to_le64(val - hash_adj);
2266 }
2267
2268 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2269 {
2270         __u64              id;
2271         struct obd_import *imp;
2272
2273         /*
2274          * XXX: to get nid we assume that underlying obd device is mdc.
2275          */
2276         imp  = class_exp2cliimp(exp);
2277         id   = imp->imp_connection->c_self + fid_flatten(fid);
2278
2279         CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2280                imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2281
2282         return id ^ (id >> 32);
2283 }
2284
2285 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2286                         struct obd_capa *oc, __u64 offset64, struct page *page,
2287                         struct ptlrpc_request **request)
2288 {
2289         struct obd_device       *obd = exp->exp_obd;
2290         struct lmv_obd          *lmv = &obd->u.lmv;
2291         struct lu_fid            rid = *fid;
2292         struct lmv_object       *obj;
2293         __u64                    offset;
2294         __u64                    hash_adj = 0;
2295         __u32                    rank = 0;
2296         __u64                    seg_size = 0;
2297         __u64                    tgt_tmp = 0;
2298         int                      tgt_idx = 0;
2299         int                      tgt0_idx = 0;
2300         int                      rc;
2301         int                      nr = 0;
2302         struct lmv_stripe       *los;
2303         struct lmv_tgt_desc     *tgt;
2304         struct lu_dirpage       *dp;
2305         struct lu_dirent        *ent;
2306         ENTRY;
2307
2308         offset = offset64;
2309
2310         rc = lmv_check_connect(obd);
2311         if (rc)
2312                 RETURN(rc);
2313
2314         CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2315
2316         /*
2317          * This case handle directory lookup in clustered metadata case (i.e.
2318          * split directory is located on multiple md servers.)
2319          * each server keeps directory entries for certain range of hashes.
2320          * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2321          * first server will keep records with hashes [ 0 ... MAX_HASH / N  - 1],
2322          * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2323          * so on....
2324          *      readdir can simply start reading entries from 0 - N server in
2325          * order but that will not scale well as all client will request dir in
2326          * to server in same order.
2327          * Following algorithm does optimization:
2328          * Instead of doing readdir in 1, 2, ...., N order, client with a
2329          * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2330          * (every client has rank R)
2331          *      But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2332          * since client ask dir from MDS{R} client has pages with offsets
2333          * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2334          * on hash  values that we get.
2335          */
2336         obj = lmv_object_find_lock(obd, fid);
2337         if (obj) {
2338                 nr       = obj->lo_objcount;
2339                 LASSERT(nr > 0);
2340                 seg_size = MAX_HASH_SIZE;
2341                 do_div(seg_size, nr);
2342                 los      = obj->lo_stripes;
2343                 tgt      = lmv_get_target(lmv, los[0].ls_mds);
2344                 rank     = lmv_node_rank(tgt->ltd_exp, fid) % nr;
2345                 tgt_tmp  = offset;
2346                 do_div(tgt_tmp, seg_size);
2347                 tgt0_idx = do_div(tgt_tmp,  nr);
2348                 tgt_idx  = (tgt0_idx + rank) % nr;
2349
2350                 if (tgt_idx < tgt0_idx)
2351                         /*
2352                          * Wrap around.
2353                          *
2354                          * Last segment has unusual length due to division
2355                          * rounding.
2356                          */
2357                         hash_adj = MAX_HASH_SIZE - seg_size * nr;
2358                 else
2359                         hash_adj = 0;
2360
2361                 hash_adj += rank * seg_size;
2362
2363                 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2364                        LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2365                        offset, tgt0_idx, offset + hash_adj, tgt_idx);
2366
2367                 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2368                 rid = obj->lo_stripes[tgt_idx].ls_fid;
2369                 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2370
2371                 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2372                        PFID(&rid), (unsigned long)offset, tgt_idx);
2373         } else
2374                 tgt = lmv_find_target(lmv, &rid);
2375
2376         if (IS_ERR(tgt))
2377                 GOTO(cleanup, rc = PTR_ERR(tgt));
2378
2379         rc = md_readpage(tgt->ltd_exp, &rid, oc, offset, page, request);
2380         if (rc)
2381                 GOTO(cleanup, rc);
2382         if (obj) {
2383                 dp = cfs_kmap(page);
2384
2385                 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2386                 lmv_hash_adjust(&dp->ldp_hash_end,   hash_adj);
2387                 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2388
2389                 for (ent = lu_dirent_start(dp); ent != NULL;
2390                      ent = lu_dirent_next(ent))
2391                         lmv_hash_adjust(&ent->lde_hash, hash_adj);
2392
2393                 if (tgt0_idx != nr - 1) {
2394                         __u64 end;
2395
2396                         end = le64_to_cpu(dp->ldp_hash_end);
2397                         if (end == DIR_END_OFF) {
2398                                 dp->ldp_hash_end = cpu_to_le32(seg_size *
2399                                                                (tgt0_idx + 1));
2400                                 CDEBUG(D_INODE,
2401                                        ""DFID" reset end "LPX64" tgt %d\n",
2402                                        PFID(&rid),
2403                                        le64_to_cpu(dp->ldp_hash_end), tgt_idx);
2404                         }
2405                 }
2406                 cfs_kunmap(page);
2407         }
2408         EXIT;
2409 cleanup:
2410         if (obj)
2411                 lmv_object_put_unlock(obj);
2412         return rc;
2413 }
2414
2415 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2416                       struct ptlrpc_request **request)
2417 {
2418         struct obd_device       *obd = exp->exp_obd;
2419         struct lmv_obd          *lmv = &obd->u.lmv;
2420         struct lmv_tgt_desc     *tgt = NULL;
2421         struct lmv_object       *obj;
2422         int                      rc;
2423         int                      sidx;
2424         int                      loop = 0;
2425         ENTRY;
2426
2427         rc = lmv_check_connect(obd);
2428         if (rc)
2429                 RETURN(rc);
2430
2431 repeat:
2432         ++loop;
2433         LASSERT(loop <= 2);
2434         LASSERT(op_data->op_namelen != 0);
2435
2436         obj = lmv_object_find(obd, &op_data->op_fid1);
2437         if (obj) {
2438                 sidx = raw_name2idx(obj->lo_hashtype,
2439                                     obj->lo_objcount,
2440                                     op_data->op_name,
2441                                     op_data->op_namelen);
2442                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2443                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2444                 tgt = lmv_get_target(lmv,
2445                                      obj->lo_stripes[sidx].ls_mds);
2446                 lmv_object_put(obj);
2447                 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2448                        op_data->op_namelen, op_data->op_name,
2449                        PFID(&op_data->op_fid1), sidx);
2450         }
2451
2452         if (tgt == NULL) {
2453                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2454                 if (IS_ERR(tgt))
2455                         RETURN(PTR_ERR(tgt));
2456                 op_data->op_bias |= MDS_CHECK_SPLIT;
2457         }
2458
2459         op_data->op_fsuid = current->fsuid;
2460         op_data->op_fsgid = current->fsgid;
2461         op_data->op_cap = cfs_curproc_cap_pack();
2462
2463         /*
2464          * If child's fid is given, cancel unused locks for it if it is from
2465          * another export than parent.
2466          *
2467          * LOOKUP lock for child (fid3) should also be cancelled on parent
2468          * tgt_tgt in mdc_unlink().
2469          */
2470         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2471
2472         /*
2473          * Cancel FULL locks on child (fid3).
2474          */
2475         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2476                               MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2477
2478         if (rc == 0)
2479                 rc = md_unlink(tgt->ltd_exp, op_data, request);
2480
2481         if (rc == -ERESTART) {
2482                 LASSERT(*request != NULL);
2483                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2484                           "Got -ERESTART during unlink!\n");
2485                 ptlrpc_req_finished(*request);
2486                 *request = NULL;
2487
2488                 /*
2489                  * Directory got split. Time to update local object and repeat
2490                  * the request with proper MDS.
2491                  */
2492                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2493                 if (rc == 0)
2494                         goto repeat;
2495         }
2496         RETURN(rc);
2497 }
2498
2499 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2500 {
2501         int        rc = 0;
2502
2503         switch (stage) {
2504         case OBD_CLEANUP_EARLY:
2505                 /* XXX: here should be calling obd_precleanup() down to
2506                  * stack. */
2507                 break;
2508         case OBD_CLEANUP_EXPORTS:
2509                 rc = obd_llog_finish(obd, 0);
2510                 if (rc != 0)
2511                         CERROR("failed to cleanup llogging subsystems\n");
2512                 break;
2513         default:
2514                 break;
2515         }
2516         RETURN(rc);
2517 }
2518
2519 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2520                         void *key, __u32 *vallen, void *val,
2521                         struct lov_stripe_md *lsm)
2522 {
2523         struct obd_device       *obd;
2524         struct lmv_obd          *lmv;
2525         int                      rc = 0;
2526         ENTRY;
2527
2528         obd = class_exp2obd(exp);
2529         if (obd == NULL) {
2530                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2531                        exp->exp_handle.h_cookie);
2532                 RETURN(-EINVAL);
2533         }
2534
2535         lmv = &obd->u.lmv;
2536         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2537                 struct lmv_tgt_desc *tgts;
2538                 int i;
2539
2540                 rc = lmv_check_connect(obd);
2541                 if (rc)
2542                         RETURN(rc);
2543
2544                 LASSERT(*vallen == sizeof(__u32));
2545                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2546                      i++, tgts++) {
2547
2548                         /*
2549                          * All tgts should be connected when this gets called.
2550                          */
2551                         if (!tgts || !tgts->ltd_exp) {
2552                                 CERROR("target not setup?\n");
2553                                 continue;
2554                         }
2555
2556                         if (!obd_get_info(tgts->ltd_exp, keylen, key,
2557                                           vallen, val, NULL))
2558                                 RETURN(0);
2559                 }
2560                 RETURN(-EINVAL);
2561         } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2562                 rc = lmv_check_connect(obd);
2563                 if (rc)
2564                         RETURN(rc);
2565
2566                 /*
2567                  * Forwarding this request to first MDS, it should know LOV
2568                  * desc.
2569                  */
2570                 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2571                                   vallen, val, NULL);
2572                 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2573                         exp->exp_connect_flags =
2574                         ((struct obd_connect_data *)val)->ocd_connect_flags;
2575                 }
2576                 RETURN(rc);
2577         }
2578
2579         CDEBUG(D_IOCTL, "Invalid key\n");
2580         RETURN(-EINVAL);
2581 }
2582
2583 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2584                        void *key, obd_count vallen, void *val,
2585                        struct ptlrpc_request_set *set)
2586 {
2587         struct lmv_tgt_desc    *tgt;
2588         struct obd_device      *obd;
2589         struct lmv_obd         *lmv;
2590         int rc = 0;
2591         ENTRY;
2592
2593         obd = class_exp2obd(exp);
2594         if (obd == NULL) {
2595                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2596                        exp->exp_handle.h_cookie);
2597                 RETURN(-EINVAL);
2598         }
2599         lmv = &obd->u.lmv;
2600
2601         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2602             KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2603                 int i, err = 0;
2604
2605                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2606                         tgt = &lmv->tgts[i];
2607
2608                         if (!tgt->ltd_exp)
2609                                 continue;
2610
2611                         err = obd_set_info_async(tgt->ltd_exp,
2612                                                  keylen, key, vallen, val, set);
2613                         if (err && rc == 0)
2614                                 rc = err;
2615                 }
2616
2617                 RETURN(rc);
2618         }
2619
2620         RETURN(-EINVAL);
2621 }
2622
2623 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2624                struct lov_stripe_md *lsm)
2625 {
2626         struct obd_device         *obd = class_exp2obd(exp);
2627         struct lmv_obd            *lmv = &obd->u.lmv;
2628         struct lmv_stripe_md      *meap;
2629         struct lmv_stripe_md      *lsmp;
2630         int                        mea_size;
2631         int                        i;
2632         ENTRY;
2633
2634         mea_size = lmv_get_easize(lmv);
2635         if (!lmmp)
2636                 RETURN(mea_size);
2637
2638         if (*lmmp && !lsm) {
2639                 OBD_FREE(*lmmp, mea_size);
2640                 *lmmp = NULL;
2641                 RETURN(0);
2642         }
2643
2644         if (*lmmp == NULL) {
2645                 OBD_ALLOC(*lmmp, mea_size);
2646                 if (*lmmp == NULL)
2647                         RETURN(-ENOMEM);
2648         }
2649
2650         if (!lsm)
2651                 RETURN(mea_size);
2652
2653         lsmp = (struct lmv_stripe_md *)lsm;
2654         meap = (struct lmv_stripe_md *)*lmmp;
2655
2656         if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2657             lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2658                 RETURN(-EINVAL);
2659
2660         meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2661         meap->mea_count = cpu_to_le32(lsmp->mea_count);
2662         meap->mea_master = cpu_to_le32(lsmp->mea_master);
2663
2664         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2665                 meap->mea_ids[i] = meap->mea_ids[i];
2666                 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2667         }
2668
2669         RETURN(mea_size);
2670 }
2671
2672 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2673                  struct lov_mds_md *lmm, int lmm_size)
2674 {
2675         struct obd_device          *obd = class_exp2obd(exp);
2676         struct lmv_stripe_md      **tmea = (struct lmv_stripe_md **)lsmp;
2677         struct lmv_stripe_md       *mea = (struct lmv_stripe_md *)lmm;
2678         struct lmv_obd             *lmv = &obd->u.lmv;
2679         int                         mea_size;
2680         int                         i;
2681         __u32                       magic;
2682         ENTRY;
2683
2684         mea_size = lmv_get_easize(lmv);
2685         if (lsmp == NULL)
2686                 return mea_size;
2687
2688         if (*lsmp != NULL && lmm == NULL) {
2689                 OBD_FREE(*tmea, mea_size);
2690                 *lsmp = NULL;
2691                 RETURN(0);
2692         }
2693
2694         LASSERT(mea_size == lmm_size);
2695
2696         OBD_ALLOC(*tmea, mea_size);
2697         if (*tmea == NULL)
2698                 RETURN(-ENOMEM);
2699
2700         if (!lmm)
2701                 RETURN(mea_size);
2702
2703         if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2704             mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2705             mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2706         {
2707                 magic = le32_to_cpu(mea->mea_magic);
2708         } else {
2709                 /*
2710                  * Old mea is not handled here.
2711                  */
2712                 CERROR("Old not supportable EA is found\n");
2713                 LBUG();
2714         }
2715
2716         (*tmea)->mea_magic = magic;
2717         (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2718         (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2719
2720         for (i = 0; i < (*tmea)->mea_count; i++) {
2721                 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2722                 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2723         }
2724         RETURN(mea_size);
2725 }
2726
2727 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2728                              ldlm_policy_data_t *policy, ldlm_mode_t mode,
2729                              int flags, void *opaque)
2730 {
2731         struct obd_device       *obd = exp->exp_obd;
2732         struct lmv_obd          *lmv = &obd->u.lmv;
2733         int                      rc = 0;
2734         int                      err;
2735         int                      i;
2736         ENTRY;
2737
2738         LASSERT(fid != NULL);
2739
2740         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2741                 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2742                         continue;
2743
2744                 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2745                                        policy, mode, flags, opaque);
2746                 if (!rc)
2747                         rc = err;
2748         }
2749         RETURN(rc);
2750 }
2751
2752 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data)
2753 {
2754         struct obd_device       *obd = exp->exp_obd;
2755         struct lmv_obd          *lmv = &obd->u.lmv;
2756         int                      rc;
2757         ENTRY;
2758         rc =  md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data);
2759         RETURN(rc);
2760 }
2761
2762 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2763                            const struct lu_fid *fid, ldlm_type_t type,
2764                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
2765                            struct lustre_handle *lockh)
2766 {
2767         struct obd_device       *obd = exp->exp_obd;
2768         struct lmv_obd          *lmv = &obd->u.lmv;
2769         ldlm_mode_t              rc;
2770         int                      i;
2771         ENTRY;
2772
2773         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2774
2775         /*
2776          * With CMD every object can have two locks in different namespaces:
2777          * lookup lock in space of mds storing direntry and update/open lock in
2778          * space of mds storing inode. Thus we check all targets, not only that
2779          * one fid was created in.
2780          */
2781         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2782                 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2783                                    type, policy, mode, lockh);
2784                 if (rc)
2785                         RETURN(rc);
2786         }
2787
2788         RETURN(0);
2789 }
2790
2791 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2792                       struct obd_export *dt_exp, struct obd_export *md_exp,
2793                       struct lustre_md *md)
2794 {
2795         struct obd_device       *obd = exp->exp_obd;
2796         struct lmv_obd          *lmv = &obd->u.lmv;
2797         int                      rc;
2798         ENTRY;
2799         rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2800         RETURN(rc);
2801 }
2802
2803 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2804 {
2805         struct obd_device       *obd = exp->exp_obd;
2806         struct lmv_obd          *lmv = &obd->u.lmv;
2807         ENTRY;
2808
2809         if (md->mea)
2810                 obd_free_memmd(exp, (void *)&md->mea);
2811         RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2812 }
2813
2814 int lmv_set_open_replay_data(struct obd_export *exp,
2815                              struct obd_client_handle *och,
2816                              struct ptlrpc_request *open_req)
2817 {
2818         struct obd_device       *obd = exp->exp_obd;
2819         struct lmv_obd          *lmv = &obd->u.lmv;
2820         struct lmv_tgt_desc     *tgt;
2821         ENTRY;
2822
2823         tgt = lmv_find_target(lmv, &och->och_fid);
2824         if (IS_ERR(tgt))
2825                 RETURN(PTR_ERR(tgt));
2826
2827         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2828 }
2829
2830 int lmv_clear_open_replay_data(struct obd_export *exp,
2831                                struct obd_client_handle *och)
2832 {
2833         struct obd_device       *obd = exp->exp_obd;
2834         struct lmv_obd          *lmv = &obd->u.lmv;
2835         struct lmv_tgt_desc     *tgt;
2836         ENTRY;
2837
2838         tgt = lmv_find_target(lmv, &och->och_fid);
2839         if (IS_ERR(tgt))
2840                 RETURN(PTR_ERR(tgt));
2841
2842         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2843 }
2844
2845 static int lmv_get_remote_perm(struct obd_export *exp,
2846                                const struct lu_fid *fid,
2847                                struct obd_capa *oc, __u32 suppgid,
2848                                struct ptlrpc_request **request)
2849 {
2850         struct obd_device       *obd = exp->exp_obd;
2851         struct lmv_obd          *lmv = &obd->u.lmv;
2852         struct lmv_tgt_desc     *tgt;
2853         int                      rc;
2854         ENTRY;
2855
2856         rc = lmv_check_connect(obd);
2857         if (rc)
2858                 RETURN(rc);
2859
2860         tgt = lmv_find_target(lmv, fid);
2861         if (IS_ERR(tgt))
2862                 RETURN(PTR_ERR(tgt));
2863
2864         rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
2865         RETURN(rc);
2866 }
2867
2868 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2869                           renew_capa_cb_t cb)
2870 {
2871         struct obd_device       *obd = exp->exp_obd;
2872         struct lmv_obd          *lmv = &obd->u.lmv;
2873         struct lmv_tgt_desc     *tgt;
2874         int                      rc;
2875         ENTRY;
2876
2877         rc = lmv_check_connect(obd);
2878         if (rc)
2879                 RETURN(rc);
2880
2881         tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
2882         if (IS_ERR(tgt))
2883                 RETURN(PTR_ERR(tgt));
2884
2885         rc = md_renew_capa(tgt->ltd_exp, oc, cb);
2886         RETURN(rc);
2887 }
2888
2889 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
2890                     const struct req_msg_field *field, struct obd_capa **oc)
2891 {
2892         struct obd_device *obd = exp->exp_obd;
2893         struct lmv_obd *lmv = &obd->u.lmv;
2894         int rc;
2895
2896         ENTRY;
2897         rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
2898         RETURN(rc);
2899 }
2900
2901 int lmv_intent_getattr_async(struct obd_export *exp,
2902                              struct md_enqueue_info *minfo,
2903                              struct ldlm_enqueue_info *einfo)
2904 {
2905         struct md_op_data       *op_data = &minfo->mi_data;
2906         struct obd_device       *obd = exp->exp_obd;
2907         struct lmv_obd          *lmv = &obd->u.lmv;
2908         struct lmv_object       *obj;
2909         struct lmv_tgt_desc     *tgt;
2910         int                      rc;
2911         int                      sidx;
2912         ENTRY;
2913
2914         rc = lmv_check_connect(obd);
2915         if (rc)
2916                 RETURN(rc);
2917
2918         if (!fid_is_sane(&op_data->op_fid2)) {
2919                 obj = lmv_object_find(obd, &op_data->op_fid1);
2920                 if (obj && op_data->op_namelen) {
2921                         sidx = raw_name2idx(obj->lo_hashtype,
2922                                             obj->lo_objcount,
2923                                             (char *)op_data->op_name,
2924                                             op_data->op_namelen);
2925                         op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2926                         tgt = lmv_get_target(lmv,
2927                                              obj->lo_stripes[sidx].ls_mds);
2928                         CDEBUG(D_INODE,
2929                                "Choose slave dir ("DFID") -> mds #%d\n",
2930                                PFID(&op_data->op_fid1), tgt->ltd_idx);
2931                 } else {
2932                         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2933                 }
2934                 if (obj)
2935                         lmv_object_put(obj);
2936         } else {
2937                 op_data->op_fid1 = op_data->op_fid2;
2938                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2939                 op_data->op_bias = MDS_CROSS_REF;
2940                 /*
2941                  * Unfortunately, we have to lie to MDC/MDS to retrieve
2942                  * attributes llite needs.
2943                 */
2944                 if (minfo->mi_it.it_op & IT_LOOKUP)
2945                         minfo->mi_it.it_op = IT_GETATTR;
2946         }
2947
2948         if (IS_ERR(tgt))
2949                 RETURN(PTR_ERR(tgt));
2950
2951         rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
2952         RETURN(rc);
2953 }
2954
2955 int lmv_revalidate_lock(struct obd_export *exp,
2956                         struct lookup_intent *it,
2957                         struct lu_fid *fid)
2958 {
2959         struct obd_device       *obd = exp->exp_obd;
2960         struct lmv_obd          *lmv = &obd->u.lmv;
2961         struct lmv_tgt_desc     *tgt;
2962         int                      rc;
2963         ENTRY;
2964
2965         rc = lmv_check_connect(obd);
2966         if (rc)
2967                 RETURN(rc);
2968
2969         tgt = lmv_find_target(lmv, fid);
2970         if (IS_ERR(tgt))
2971                 RETURN(PTR_ERR(tgt));
2972
2973         rc = md_revalidate_lock(tgt->ltd_exp, it, fid);
2974         RETURN(rc);
2975 }
2976
2977
2978 struct obd_ops lmv_obd_ops = {
2979         .o_owner                = THIS_MODULE,
2980         .o_setup                = lmv_setup,
2981         .o_cleanup              = lmv_cleanup,
2982         .o_precleanup           = lmv_precleanup,
2983         .o_process_config       = lmv_process_config,
2984         .o_connect              = lmv_connect,
2985         .o_disconnect           = lmv_disconnect,
2986         .o_statfs               = lmv_statfs,
2987         .o_get_info             = lmv_get_info,
2988         .o_set_info_async       = lmv_set_info_async,
2989         .o_packmd               = lmv_packmd,
2990         .o_unpackmd             = lmv_unpackmd,
2991         .o_notify               = lmv_notify,
2992         .o_get_uuid             = lmv_get_uuid,
2993         .o_iocontrol            = lmv_iocontrol,
2994         .o_fid_delete           = lmv_fid_delete
2995 };
2996
2997 struct md_ops lmv_md_ops = {
2998         .m_getstatus            = lmv_getstatus,
2999         .m_change_cbdata        = lmv_change_cbdata,
3000         .m_close                = lmv_close,
3001         .m_create               = lmv_create,
3002         .m_done_writing         = lmv_done_writing,
3003         .m_enqueue              = lmv_enqueue,
3004         .m_getattr              = lmv_getattr,
3005         .m_getxattr             = lmv_getxattr,
3006         .m_getattr_name         = lmv_getattr_name,
3007         .m_intent_lock          = lmv_intent_lock,
3008         .m_link                 = lmv_link,
3009         .m_rename               = lmv_rename,
3010         .m_setattr              = lmv_setattr,
3011         .m_setxattr             = lmv_setxattr,
3012         .m_sync                 = lmv_sync,
3013         .m_readpage             = lmv_readpage,
3014         .m_unlink               = lmv_unlink,
3015         .m_init_ea_size         = lmv_init_ea_size,
3016         .m_cancel_unused        = lmv_cancel_unused,
3017         .m_set_lock_data        = lmv_set_lock_data,
3018         .m_lock_match           = lmv_lock_match,
3019         .m_get_lustre_md        = lmv_get_lustre_md,
3020         .m_free_lustre_md       = lmv_free_lustre_md,
3021         .m_set_open_replay_data = lmv_set_open_replay_data,
3022         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3023         .m_renew_capa           = lmv_renew_capa,
3024         .m_unpack_capa          = lmv_unpack_capa,
3025         .m_get_remote_perm      = lmv_get_remote_perm,
3026         .m_intent_getattr_async = lmv_intent_getattr_async,
3027         .m_revalidate_lock      = lmv_revalidate_lock
3028 };
3029
3030 static quota_interface_t *quota_interface;
3031 extern quota_interface_t lmv_quota_interface;
3032
3033 int __init lmv_init(void)
3034 {
3035         struct lprocfs_static_vars lvars;
3036         int                        rc;
3037
3038         lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3039                                                 sizeof(struct lmv_object),
3040                                                 0, 0);
3041         if (!lmv_object_cache) {
3042                 CERROR("Error allocating lmv objects cache\n");
3043                 return -ENOMEM;
3044         }
3045
3046         lprocfs_lmv_init_vars(&lvars);
3047
3048         request_module("lquota");
3049         quota_interface = PORTAL_SYMBOL_GET(lmv_quota_interface);
3050         init_obd_quota_ops(quota_interface, &lmv_obd_ops);
3051
3052         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3053                                  lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3054         if (rc) {
3055                 if (quota_interface)
3056                         PORTAL_SYMBOL_PUT(lmv_quota_interface);
3057                 cfs_mem_cache_destroy(lmv_object_cache);
3058         }
3059
3060         return rc;
3061 }
3062
3063 #ifdef __KERNEL__
3064 static void lmv_exit(void)
3065 {
3066         if (quota_interface)
3067                 PORTAL_SYMBOL_PUT(lmv_quota_interface);
3068
3069         class_unregister_type(LUSTRE_LMV_NAME);
3070
3071         LASSERTF(atomic_read(&lmv_object_count) == 0,
3072                  "Can't free lmv objects cache, %d object(s) busy\n",
3073                  atomic_read(&lmv_object_count));
3074         cfs_mem_cache_destroy(lmv_object_cache);
3075 }
3076
3077 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3078 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3079 MODULE_LICENSE("GPL");
3080
3081 module_init(lmv_init);
3082 module_exit(lmv_exit);
3083 #endif