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