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  * Copyright (C) 2002, 2003, 2004, 2005, 2006 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #ifndef EXPORT_SYMTAB
  23 # define EXPORT_SYMTAB
  24 #endif
  25 #define DEBUG_SUBSYSTEM S_LMV
  26 #ifdef __KERNEL__
  27 #include <linux/slab.h>
  28 #include <linux/module.h>
  29 #include <linux/init.h>
  30 #include <linux/slab.h>
  31 #include <linux/pagemap.h>
  32 #include <linux/mm.h>
  33 #include <asm/div64.h>
  34 #include <linux/seq_file.h>
  35 #include <linux/namei.h>
  36 #else
  37 #include <liblustre.h>
  38 #endif
  39
  40 #include <linux/ext2_fs.h>
  41
  42 #include <lustre/lustre_idl.h>
  43 #include <lustre_log.h>
  44 #include <obd_support.h>
  45 #include <lustre_lib.h>
  46 #include <lustre_net.h>
  47 #include <obd_class.h>
  48 #include <lprocfs_status.h>
  49 #include <lustre_lite.h>
  50 #include <lustre_fid.h>
  51 #include "lmv_internal.h"
  52
  53 /* not defined for liblustre building */
  54 #if !defined(ATOMIC_INIT)
  55 #define ATOMIC_INIT(val) { (val) }
  56 #endif
  57
  58 /* object cache. */
  59 cfs_mem_cache_t *obj_cache;
  60 atomic_t obj_cache_count = ATOMIC_INIT(0);
  61
  62 static void lmv_activate_target(struct lmv_obd *lmv,
  63                                 struct lmv_tgt_desc *tgt,
  64                                 int activate)
  65 {
  66         if (tgt->ltd_active == activate)
  67                 return;
  68
  69         tgt->ltd_active = activate;
  70         lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
  71 }
  72
  73 /* Error codes:
  74  *
  75  *  -EINVAL  : UUID can't be found in the LMV's target list
  76  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  77  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  78  */
  79 static int lmv_set_mdc_active(struct lmv_obd *lmv, struct obd_uuid *uuid,
  80                               int activate)
  81 {
  82         struct lmv_tgt_desc *tgt;
  83         struct obd_device *obd;
  84         int i, rc = 0;
  85         ENTRY;
  86
  87         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  88                lmv, uuid->uuid, activate);
  89
  90         spin_lock(&lmv->lmv_lock);
  91         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
  92                 if (tgt->ltd_exp == NULL)
  93                         continue;
  94
  95                 CDEBUG(D_INFO, "lmv idx %d is %s conn "LPX64"\n",
  96                        i, tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
  97
  98                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
  99                         break;
 100         }
 101
 102         if (i == lmv->desc.ld_tgt_count)
 103                 GOTO(out_lmv_lock, rc = -EINVAL);
 104
 105         obd = class_exp2obd(tgt->ltd_exp);
 106         if (obd == NULL)
 107                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 108
 109         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 110                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 111                obd->obd_type->typ_name, i);
 112         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 113
 114         if (tgt->ltd_active == activate) {
 115                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 116                        activate ? "" : "in");
 117                 GOTO(out_lmv_lock, rc);
 118         }
 119
 120         CDEBUG(D_INFO, "Marking OBD %p %sactive\n",
 121                obd, activate ? "" : "in");
 122
 123         lmv_activate_target(lmv, tgt, activate);
 124
 125         EXIT;
 126
 127  out_lmv_lock:
 128         spin_unlock(&lmv->lmv_lock);
 129         return rc;
 130 }
 131
 132 static int lmv_set_mdc_data(struct lmv_obd *lmv, struct obd_uuid *uuid,
 133                             struct obd_connect_data *data)
 134 {
 135         struct lmv_tgt_desc *tgt;
 136         int i;
 137         ENTRY;
 138
 139         LASSERT(data != NULL);
 140
 141         spin_lock(&lmv->lmv_lock);
 142         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 143                 if (tgt->ltd_exp == NULL)
 144                         continue;
 145
 146                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
 147                         lmv->datas[tgt->ltd_idx] = *data;
 148                         break;
 149                 }
 150         }
 151         spin_unlock(&lmv->lmv_lock);
 152         RETURN(0);
 153 }
 154
 155 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 156                       enum obd_notify_event ev, void *data)
 157 {
 158         struct lmv_obd *lmv = &obd->u.lmv;
 159         struct obd_uuid *uuid;
 160         int rc = 0;
 161         ENTRY;
 162
 163         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 164                 CERROR("unexpected notification of %s %s!\n",
 165                        watched->obd_type->typ_name,
 166                        watched->obd_name);
 167                 RETURN(-EINVAL);
 168         }
 169
 170         uuid = &watched->u.cli.cl_target_uuid;
 171         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 172                 /*
 173                  * Set MDC as active before notifying the observer, so the
 174                  * observer can use the MDC normally.
 175                  */
 176                 rc = lmv_set_mdc_active(lmv, uuid,
 177                                         ev == OBD_NOTIFY_ACTIVE);
 178                 if (rc) {
 179                         CERROR("%sactivation of %s failed: %d\n",
 180                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 181                                uuid->uuid, rc);
 182                         RETURN(rc);
 183                 }
 184         } else if (ev == OBD_NOTIFY_OCD) {
 185                 struct obd_connect_data *conn_data =
 186                         &watched->u.cli.cl_import->imp_connect_data;
 187
 188                 /* Set connect data to desired target, update
 189                  * exp_connect_flags. */
 190                 rc = lmv_set_mdc_data(lmv, uuid, conn_data);
 191                 if (rc) {
 192                         CERROR("can't set connect data to target %s, rc %d\n",
 193                                uuid->uuid, rc);
 194                         RETURN(rc);
 195                 }
 196
 197                 /*
 198                  * XXX: Make sure that ocd_connect_flags from all targets are
 199                  * the same. Otherwise one of MDTs runs wrong version or
 200                  * something like this.  --umka
 201                  */
 202                 obd->obd_self_export->exp_connect_flags =
 203                         conn_data->ocd_connect_flags;
 204         }
 205 #if 0
 206         else if (ev == OBD_NOTIFY_DISCON) {
 207                 /* For disconnect event, flush fld cache for failout MDS case. */
 208                 fld_client_flush(&lmv->lmv_fld);
 209         }
 210 #endif
 211         /* Pass the notification up the chain. */
 212         if (obd->obd_observer)
 213                 rc = obd_notify(obd->obd_observer, watched, ev, data);
 214
 215         RETURN(rc);
 216 }
 217
 218 /* this is fake connect function. Its purpose is to initialize lmv and say
 219  * caller that everything is okay. Real connection will be performed later. */
 220 static int lmv_connect(const struct lu_env *env,
 221                        struct lustre_handle *conn, struct obd_device *obd,
 222                        struct obd_uuid *cluuid, struct obd_connect_data *data)
 223 {
 224 #ifdef __KERNEL__
 225         struct proc_dir_entry *lmv_proc_dir;
 226 #endif
 227         struct lmv_obd *lmv = &obd->u.lmv;
 228         struct obd_export *exp;
 229         int rc = 0;
 230         ENTRY;
 231
 232         rc = class_connect(conn, obd, cluuid);
 233         if (rc) {
 234                 CERROR("class_connection() returned %d\n", rc);
 235                 RETURN(rc);
 236         }
 237
 238         exp = class_conn2export(conn);
 239
 240         /* we don't want to actually do the underlying connections more than
 241          * once, so keep track. */
 242         lmv->refcount++;
 243         if (lmv->refcount > 1) {
 244                 class_export_put(exp);
 245                 RETURN(0);
 246         }
 247
 248         lmv->exp = exp;
 249         lmv->connected = 0;
 250         lmv->cluuid = *cluuid;
 251
 252         if (data)
 253                 lmv->conn_data = *data;
 254
 255 #ifdef __KERNEL__
 256         lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
 257                                         NULL, NULL);
 258         if (IS_ERR(lmv_proc_dir)) {
 259                 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
 260                        obd->obd_type->typ_name, obd->obd_name);
 261                 lmv_proc_dir = NULL;
 262         }
 263 #endif
 264
 265         /* all real clients should perform actual connection right away, because
 266          * it is possible, that LMV will not have opportunity to connect targets
 267          * and MDC stuff will be called directly, for instance while reading
 268          * ../mdc/../kbytesfree procfs file, etc. */
 269         if (data->ocd_connect_flags & OBD_CONNECT_REAL)
 270                 rc = lmv_check_connect(obd);
 271
 272 #ifdef __KERNEL__
 273         if (rc) {
 274                 if (lmv_proc_dir)
 275                         lprocfs_remove(&lmv_proc_dir);
 276         }
 277 #endif
 278
 279         RETURN(rc);
 280 }
 281
 282 static void lmv_set_timeouts(struct obd_device *obd)
 283 {
 284         struct lmv_tgt_desc *tgts;
 285         struct lmv_obd *lmv;
 286         int i;
 287
 288         lmv = &obd->u.lmv;
 289         if (lmv->server_timeout == 0)
 290                 return;
 291
 292         if (lmv->connected == 0)
 293                 return;
 294
 295         for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
 296                 if (tgts->ltd_exp == NULL)
 297                         continue;
 298
 299                 obd_set_info_async(tgts->ltd_exp, strlen("inter_mds"),
 300                                    "inter_mds", 0, NULL, NULL);
 301         }
 302 }
 303
 304 static int lmv_init_ea_size(struct obd_export *exp, int easize,
 305                             int def_easize, int cookiesize)
 306 {
 307         struct obd_device *obd = exp->exp_obd;
 308         struct lmv_obd *lmv = &obd->u.lmv;
 309         int i, rc = 0, change = 0;
 310         ENTRY;
 311
 312         if (lmv->max_easize < easize) {
 313                 lmv->max_easize = easize;
 314                 change = 1;
 315         }
 316         if (lmv->max_def_easize < def_easize) {
 317                 lmv->max_def_easize = def_easize;
 318                 change = 1;
 319         }
 320         if (lmv->max_cookiesize < cookiesize) {
 321                 lmv->max_cookiesize = cookiesize;
 322                 change = 1;
 323         }
 324         if (change == 0)
 325                 RETURN(0);
 326
 327         if (lmv->connected == 0)
 328                 RETURN(0);
 329
 330         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 331                 if (lmv->tgts[i].ltd_exp == NULL) {
 332                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
 333                         continue;
 334                 }
 335
 336                 rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize,
 337                                      cookiesize);
 338                 if (rc) {
 339                         CERROR("obd_init_ea_size() failed on MDT target %d, "
 340                                "error %d.\n", i, rc);
 341                         break;
 342                 }
 343         }
 344         RETURN(rc);
 345 }
 346
 347 #define MAX_STRING_SIZE 128
 348
 349 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 350 {
 351         struct lmv_obd *lmv = &obd->u.lmv;
 352         struct obd_uuid *cluuid = &lmv->cluuid;
 353         struct obd_connect_data *mdc_data = NULL;
 354         struct obd_uuid lmv_mdc_uuid = { "LMV_MDC_UUID" };
 355         struct lustre_handle conn = {0, };
 356         struct obd_device *mdc_obd;
 357         struct obd_export *mdc_exp;
 358         struct lu_fld_target target;
 359         int rc;
 360 #ifdef __KERNEL__
 361         struct proc_dir_entry *lmv_proc_dir;
 362 #endif
 363         ENTRY;
 364
 365         /* for MDS: don't connect to yourself */
 366         if (obd_uuid_equals(&tgt->ltd_uuid, cluuid)) {
 367                 CDEBUG(D_CONFIG, "don't connect back to %s\n", cluuid->uuid);
 368                 /* XXX - the old code didn't increment active tgt count.
 369                  *       should we ? */
 370                 RETURN(0);
 371         }
 372
 373         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 374                                         &obd->obd_uuid);
 375         if (!mdc_obd) {
 376                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 377                 RETURN(-EINVAL);
 378         }
 379
 380         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
 381                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 382                 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
 383                 cluuid->uuid);
 384
 385         if (!mdc_obd->obd_set_up) {
 386                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 387                 RETURN(-EINVAL);
 388         }
 389
 390         rc = obd_connect(NULL, &conn, mdc_obd, &lmv_mdc_uuid,
 391                          &lmv->conn_data);
 392         if (rc) {
 393                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 394                 RETURN(rc);
 395         }
 396
 397         mdc_exp = class_conn2export(&conn);
 398
 399         /* Init fid sequence client for this mdc. */
 400         rc = obd_fid_init(mdc_exp);
 401         if (rc)
 402                 RETURN(rc);
 403
 404         /* Add new FLD target. */
 405         target.ft_srv = NULL;
 406         target.ft_exp = mdc_exp;
 407         target.ft_idx = tgt->ltd_idx;
 408
 409         fld_client_add_target(&lmv->lmv_fld, &target);
 410
 411         mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data;
 412
 413         rc = obd_register_observer(mdc_obd, obd);
 414         if (rc) {
 415                 obd_disconnect(mdc_exp);
 416                 CERROR("target %s register_observer error %d\n",
 417                        tgt->ltd_uuid.uuid, rc);
 418                 RETURN(rc);
 419         }
 420
 421         if (obd->obd_observer) {
 422                 /* tell the mds_lmv about the new target */
 423                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 424                                 OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts));
 425                 if (rc) {
 426                         obd_disconnect(mdc_exp);
 427                         RETURN(rc);
 428                 }
 429         }
 430
 431         tgt->ltd_active = 1;
 432         tgt->ltd_exp = mdc_exp;
 433         lmv->desc.ld_active_tgt_count++;
 434
 435         /* copy connect data, it may be used later */
 436         lmv->datas[tgt->ltd_idx] = *mdc_data;
 437
 438         md_init_ea_size(tgt->ltd_exp, lmv->max_easize,
 439                         lmv->max_def_easize, lmv->max_cookiesize);
 440
 441         CDEBUG(D_CONFIG, "connected to %s(%s) successfully (%d)\n",
 442                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 443                 atomic_read(&obd->obd_refcount));
 444
 445 #ifdef __KERNEL__
 446         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 447         if (lmv_proc_dir) {
 448                 struct proc_dir_entry *mdc_symlink;
 449                 char name[MAX_STRING_SIZE + 1];
 450
 451                 LASSERT(mdc_obd->obd_type != NULL);
 452                 LASSERT(mdc_obd->obd_type->typ_name != NULL);
 453                 name[MAX_STRING_SIZE] = '\0';
 454                 snprintf(name, MAX_STRING_SIZE, "../../../%s/%s",
 455                          mdc_obd->obd_type->typ_name,
 456                          mdc_obd->obd_name);
 457                 mdc_symlink = proc_symlink(mdc_obd->obd_name,
 458                                            lmv_proc_dir, name);
 459                 if (mdc_symlink == NULL) {
 460                         CERROR("could not register LMV target "
 461                                "/proc/fs/lustre/%s/%s/target_obds/%s.",
 462                                obd->obd_type->typ_name, obd->obd_name,
 463                                mdc_obd->obd_name);
 464                         lprocfs_remove(&lmv_proc_dir);
 465                         lmv_proc_dir = NULL;
 466                 }
 467         }
 468 #endif
 469         RETURN(0);
 470 }
 471
 472 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
 473 {
 474         struct lmv_obd *lmv = &obd->u.lmv;
 475         struct lmv_tgt_desc *tgt;
 476         int rc = 0;
 477         ENTRY;
 478
 479         CDEBUG(D_CONFIG, "tgt_uuid: %s.\n", tgt_uuid->uuid);
 480
 481         lmv_init_lock(lmv);
 482
 483         if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
 484                 lmv_init_unlock(lmv);
 485                 CERROR("can't add %s, LMV module compiled for %d MDCs. "
 486                        "That many MDCs already configured.\n",
 487                        tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
 488                 RETURN(-EINVAL);
 489         }
 490         if (lmv->desc.ld_tgt_count == 0) {
 491                 struct obd_device *mdc_obd;
 492
 493                 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
 494                                                 &obd->obd_uuid);
 495                 if (!mdc_obd) {
 496                         lmv_init_unlock(lmv);
 497                         CERROR("Target %s not attached\n", tgt_uuid->uuid);
 498                         RETURN(-EINVAL);
 499                 }
 500
 501                 rc = obd_llog_init(obd, NULL, mdc_obd, 0, NULL, tgt_uuid);
 502                 if (rc) {
 503                         lmv_init_unlock(lmv);
 504                         CERROR("lmv failed to setup llogging subsystems\n");
 505                 }
 506         }
 507         spin_lock(&lmv->lmv_lock);
 508         tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
 509         tgt->ltd_uuid = *tgt_uuid;
 510         spin_unlock(&lmv->lmv_lock);
 511
 512         if (lmv->connected) {
 513                 rc = lmv_connect_mdc(obd, tgt);
 514                 if (rc) {
 515                         spin_lock(&lmv->lmv_lock);
 516                         lmv->desc.ld_tgt_count--;
 517                         memset(tgt, 0, sizeof(*tgt));
 518                         spin_unlock(&lmv->lmv_lock);
 519                 } else {
 520                         int easize = sizeof(struct lmv_stripe_md) +
 521                                      lmv->desc.ld_tgt_count *
 522                                      sizeof(struct lu_fid);
 523                         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 524                 }
 525         }
 526
 527         lmv_init_unlock(lmv);
 528         RETURN(rc);
 529 }
 530
 531 /* performs a check if passed obd is connected. If no - connect it. */
 532 int lmv_check_connect(struct obd_device *obd)
 533 {
 534         struct lmv_obd *lmv = &obd->u.lmv;
 535         struct lmv_tgt_desc *tgt;
 536         int i, rc, easize;
 537         ENTRY;
 538
 539         if (lmv->connected)
 540                 RETURN(0);
 541
 542         lmv_init_lock(lmv);
 543         if (lmv->connected) {
 544                 lmv_init_unlock(lmv);
 545                 RETURN(0);
 546         }
 547
 548         if (lmv->desc.ld_tgt_count == 0) {
 549                 CERROR("%s: no targets configured.\n", obd->obd_name);
 550                 RETURN(-EINVAL);
 551         }
 552
 553         CDEBUG(D_CONFIG, "time to connect %s to %s\n",
 554                lmv->cluuid.uuid, obd->obd_name);
 555
 556         LASSERT(lmv->tgts != NULL);
 557
 558         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 559                 rc = lmv_connect_mdc(obd, tgt);
 560                 if (rc)
 561                         GOTO(out_disc, rc);
 562         }
 563
 564         lmv_set_timeouts(obd);
 565         class_export_put(lmv->exp);
 566         lmv->connected = 1;
 567         easize = lmv_get_easize(lmv);
 568         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 569         lmv_init_unlock(lmv);
 570         RETURN(0);
 571
 572  out_disc:
 573         while (i-- > 0) {
 574                 int rc2;
 575                 --tgt;
 576                 tgt->ltd_active = 0;
 577                 if (tgt->ltd_exp) {
 578                         --lmv->desc.ld_active_tgt_count;
 579                         rc2 = obd_disconnect(tgt->ltd_exp);
 580                         if (rc2) {
 581                                 CERROR("error: LMV target %s disconnect on "
 582                                        "MDC idx %d: error %d\n",
 583                                        tgt->ltd_uuid.uuid, i, rc2);
 584                         }
 585                 }
 586         }
 587         class_disconnect(lmv->exp);
 588         lmv_init_unlock(lmv);
 589         RETURN(rc);
 590 }
 591
 592 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 593 {
 594 #ifdef __KERNEL__
 595         struct proc_dir_entry *lmv_proc_dir;
 596 #endif
 597         struct lmv_obd *lmv = &obd->u.lmv;
 598         struct obd_device *mdc_obd;
 599         int rc;
 600         ENTRY;
 601
 602         LASSERT(tgt != NULL);
 603         LASSERT(obd != NULL);
 604
 605         mdc_obd = class_exp2obd(tgt->ltd_exp);
 606
 607         if (mdc_obd)
 608                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 609
 610 #ifdef __KERNEL__
 611         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 612         if (lmv_proc_dir) {
 613                 struct proc_dir_entry *mdc_symlink;
 614
 615                 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
 616                 if (mdc_symlink) {
 617                         lprocfs_remove(&mdc_symlink);
 618                 } else {
 619                         CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
 620                                obd->obd_type->typ_name, obd->obd_name,
 621                                mdc_obd->obd_name);
 622                 }
 623         }
 624 #endif
 625         rc = obd_fid_fini(tgt->ltd_exp);
 626         if (rc)
 627                 CERROR("Can't finanize fids factory\n");
 628
 629         CDEBUG(D_OTHER, "Disconnected from %s(%s) successfully\n",
 630                tgt->ltd_exp->exp_obd->obd_name,
 631                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 632
 633         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 634         rc = obd_disconnect(tgt->ltd_exp);
 635         if (rc) {
 636                 if (tgt->ltd_active) {
 637                         CERROR("Target %s disconnect error %d\n",
 638                                tgt->ltd_uuid.uuid, rc);
 639                 }
 640         }
 641
 642         lmv_activate_target(lmv, tgt, 0);
 643         tgt->ltd_exp = NULL;
 644         RETURN(0);
 645 }
 646
 647 static int lmv_disconnect(struct obd_export *exp)
 648 {
 649         struct obd_device *obd = class_exp2obd(exp);
 650 #ifdef __KERNEL__
 651         struct proc_dir_entry *lmv_proc_dir;
 652 #endif
 653         struct lmv_obd *lmv = &obd->u.lmv;
 654         int rc, i;
 655         ENTRY;
 656
 657         if (!lmv->tgts)
 658                 goto out_local;
 659
 660         /* Only disconnect the underlying layers on the final disconnect. */
 661         lmv->refcount--;
 662         if (lmv->refcount != 0)
 663                 goto out_local;
 664
 665         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 666                 if (lmv->tgts[i].ltd_exp == NULL)
 667                         continue;
 668                 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
 669         }
 670
 671 #ifdef __KERNEL__
 672         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 673         if (lmv_proc_dir) {
 674                 lprocfs_remove(&lmv_proc_dir);
 675         } else {
 676                 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
 677                        obd->obd_type->typ_name, obd->obd_name);
 678         }
 679 #endif
 680
 681 out_local:
 682         /*
 683          * This is the case when no real connection is established by
 684          * lmv_check_connect().
 685          */
 686         if (!lmv->connected)
 687                 class_export_put(exp);
 688         rc = class_disconnect(exp);
 689         if (lmv->refcount == 0)
 690                 lmv->connected = 0;
 691         RETURN(rc);
 692 }
 693
 694 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 695                          int len, void *karg, void *uarg)
 696 {
 697         struct obd_device *obddev = class_exp2obd(exp);
 698         struct lmv_obd *lmv = &obddev->u.lmv;
 699         int i, rc = 0, set = 0;
 700         ENTRY;
 701
 702         if (lmv->desc.ld_tgt_count == 0)
 703                 RETURN(-ENOTTY);
 704
 705         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 706                 int err;
 707
 708                 if (lmv->tgts[i].ltd_exp == NULL)
 709                         continue;
 710
 711                 err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len, karg, uarg);
 712                 if (err) {
 713                         if (lmv->tgts[i].ltd_active) {
 714                                 CERROR("error: iocontrol MDC %s on MDT"
 715                                        "idx %d: err = %d\n",
 716                                        lmv->tgts[i].ltd_uuid.uuid, i, err);
 717                                 if (!rc)
 718                                         rc = err;
 719                         }
 720                 } else
 721                         set = 1;
 722         }
 723         if (!set && !rc)
 724                 rc = -EIO;
 725
 726         RETURN(rc);
 727 }
 728
 729 enum MDS_POLICY {
 730      CHAR_TYPE,
 731      NID_TYPE
 732 };
 733
 734 static int lmv_all_chars_policy(int count, const char *name,
 735                                 int len)
 736 {
 737         unsigned int c = 0;
 738
 739         while (len > 0)
 740                 c += name[--len];
 741         c = c % count;
 742         return c;
 743 }
 744
 745 static int lmv_nid_policy(struct lmv_obd *lmv)
 746 {
 747         struct obd_import *imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
 748         __u32 id;
 749         /*
 750          * XXX Hack: to get nid we assume that underlying obd device is mdc.
 751          */
 752         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
 753         return id % lmv->desc.ld_tgt_count;
 754 }
 755
 756 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
 757                           int type)
 758 {
 759         switch (type) {
 760         case CHAR_TYPE:
 761                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
 762                                             op_data->op_name,
 763                                             op_data->op_namelen);
 764         case NID_TYPE:
 765                 return lmv_nid_policy(lmv);
 766
 767         default:
 768                 break;
 769         }
 770
 771         CERROR("unsupport type %d \n", type);
 772         return -EINVAL;
 773 }
 774
 775 /* This is _inode_ placement policy function (not name). */
 776 static int lmv_placement_policy(struct obd_device *obd,
 777                                 struct md_op_data *op_data,
 778                                 mdsno_t *mds)
 779 {
 780         struct lmv_obd *lmv = &obd->u.lmv;
 781         struct lmv_obj *obj;
 782         int rc;
 783         ENTRY;
 784
 785         LASSERT(mds != NULL);
 786
 787         /*
 788          * Allocate new fid on target according to operation type and parent
 789          * home mds.
 790          */
 791         obj = lmv_obj_grab(obd, &op_data->op_fid1);
 792         if (obj != NULL || op_data->op_name == NULL ||
 793             op_data->op_opc != LUSTRE_OPC_MKDIR) {
 794                 /*
 795                  * Allocate fid for non-dir or for null name or for case parent
 796                  * dir is split.
 797                  */
 798                 if (obj) {
 799                         lmv_obj_put(obj);
 800
 801                         /*
 802                          * If we have this flag turned on, and we see that
 803                          * parent dir is split, this means, that caller did not
 804                          * notice split yet. This is race and we would like to
 805                          * let caller know that.
 806                          */
 807                         if (op_data->op_bias & MDS_CHECK_SPLIT)
 808                                 RETURN(-ERESTART);
 809                 }
 810
 811                 /*
 812                  * Allocate new fid on same mds where parent fid is located and
 813                  * where operation will be sent. In case of split dir, ->op_fid1
 814                  * and ->op_mds here will contain fid and mds of slave directory
 815                  * object (assigned by caller).
 816                  */
 817                 *mds = op_data->op_mds;
 818                 rc = 0;
 819
 820 #if 0
 821                 /* XXX: This should be removed later wehn we sure it is not
 822                  * needed. */
 823                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, mds);
 824                 if (rc)
 825                         GOTO(out, rc);
 826 #endif
 827         } else {
 828                 /*
 829                  * Parent directory is not split and we want to create a
 830                  * directory in it. Let's calculate where to place it according
 831                  * to name.
 832                  */
 833                 *mds = lmv_choose_mds(lmv, op_data, NID_TYPE);
 834                 rc = 0;
 835         }
 836         EXIT;
 837 #if 0
 838 out:
 839 #endif
 840         if (rc) {
 841                 CERROR("Can't choose MDS, err = %d\n", rc);
 842         } else {
 843                 LASSERT(*mds < lmv->desc.ld_tgt_count);
 844         }
 845
 846         return rc;
 847 }
 848
 849 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
 850                     mdsno_t mds)
 851 {
 852         struct lmv_tgt_desc *tgt = &lmv->tgts[mds];
 853         int rc;
 854         ENTRY;
 855
 856         /* New seq alloc and FLD setup should be atomic. */
 857         down(&tgt->ltd_fid_sem);
 858
 859         /* Asking underlaying tgt layer to allocate new fid. */
 860         rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
 861         if (rc > 0) {
 862                 LASSERT(fid_is_sane(fid));
 863
 864                 /* Client switches to new sequence, setup FLD. */
 865                 rc = fld_client_create(&lmv->lmv_fld, fid_seq(fid),
 866                                        mds, NULL);
 867                 if (rc) {
 868                         CERROR("Can't create fld entry, rc %d\n", rc);
 869                         /* Delete just allocated fid sequence */
 870                         obd_fid_delete(tgt->ltd_exp, NULL);
 871                 }
 872         }
 873         up(&tgt->ltd_fid_sem);
 874         RETURN(rc);
 875 }
 876
 877 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
 878                   struct md_op_data *op_data)
 879 {
 880         struct obd_device *obd = class_exp2obd(exp);
 881         struct lmv_obd *lmv = &obd->u.lmv;
 882         mdsno_t mds;
 883         int rc;
 884         ENTRY;
 885
 886         LASSERT(op_data != NULL);
 887         LASSERT(fid != NULL);
 888
 889         rc = lmv_placement_policy(obd, op_data, &mds);
 890         if (rc) {
 891                 CERROR("Can't get target for allocating fid, "
 892                        "rc %d\n", rc);
 893                 RETURN(rc);
 894         }
 895
 896         rc = __lmv_fid_alloc(lmv, fid, mds);
 897         if (rc) {
 898                 CERROR("Can't alloc new fid, rc %d\n", rc);
 899                 RETURN(rc);
 900         }
 901
 902         RETURN(rc);
 903 }
 904
 905 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
 906 {
 907         ENTRY;
 908
 909         LASSERT(exp && fid);
 910         if (lmv_obj_delete(exp, fid)) {
 911                 CDEBUG(D_OTHER, "lmv object "DFID" is destroyed.\n",
 912                        PFID(fid));
 913         }
 914         RETURN(0);
 915 }
 916
 917 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 918 {
 919         struct lmv_obd *lmv = &obd->u.lmv;
 920         struct lprocfs_static_vars lvars;
 921         struct lmv_desc *desc;
 922         int rc, i = 0;
 923         ENTRY;
 924
 925         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 926                 CERROR("LMV setup requires a descriptor\n");
 927                 RETURN(-EINVAL);
 928         }
 929
 930         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
 931         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
 932                 CERROR("descriptor size wrong: %d > %d\n",
 933                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
 934                 RETURN(-EINVAL);
 935         }
 936
 937         lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
 938
 939         OBD_ALLOC(lmv->tgts, lmv->tgts_size);
 940         if (lmv->tgts == NULL)
 941                 RETURN(-ENOMEM);
 942
 943         for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
 944                 sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
 945                 lmv->tgts[i].ltd_idx = i;
 946         }
 947
 948         lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
 949
 950         OBD_ALLOC(lmv->datas, lmv->datas_size);
 951         if (lmv->datas == NULL)
 952                 GOTO(out_free_tgts, rc = -ENOMEM);
 953
 954         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
 955         lmv->desc.ld_tgt_count = 0;
 956         lmv->desc.ld_active_tgt_count = 0;
 957         lmv->max_cookiesize = 0;
 958         lmv->max_def_easize = 0;
 959         lmv->max_easize = 0;
 960
 961         spin_lock_init(&lmv->lmv_lock);
 962         sema_init(&lmv->init_sem, 1);
 963
 964         rc = lmv_obj_setup(obd);
 965         if (rc) {
 966                 CERROR("Can't setup LMV object manager, "
 967                        "error %d.\n", rc);
 968                 GOTO(out_free_datas, rc);
 969         }
 970
 971         lprocfs_lmv_init_vars(&lvars);
 972         lprocfs_obd_setup(obd, lvars.obd_vars);
 973 #ifdef LPROCFS
 974         {
 975                 struct proc_dir_entry *entry;
 976
 977                 entry = create_proc_entry("target_obd_status", 0444,
 978                                           obd->obd_proc_entry);
 979                 if (entry != NULL) {
 980                         entry->proc_fops = &lmv_proc_target_fops;
 981                         entry->data = obd;
 982                 }
 983        }
 984 #endif
 985         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
 986                              LUSTRE_CLI_FLD_HASH_DHT);
 987         if (rc) {
 988                 CERROR("can't init FLD, err %d\n",
 989                        rc);
 990                 GOTO(out_free_datas, rc);
 991         }
 992
 993         RETURN(0);
 994
 995 out_free_datas:
 996         OBD_FREE(lmv->datas, lmv->datas_size);
 997         lmv->datas = NULL;
 998 out_free_tgts:
 999         OBD_FREE(lmv->tgts, lmv->tgts_size);
1000         lmv->tgts = NULL;
1001         return rc;
1002 }
1003
1004 static int lmv_cleanup(struct obd_device *obd)
1005 {
1006         struct lmv_obd *lmv = &obd->u.lmv;
1007         ENTRY;
1008
1009         fld_client_fini(&lmv->lmv_fld);
1010         lprocfs_obd_cleanup(obd);
1011         lmv_obj_cleanup(obd);
1012         OBD_FREE(lmv->datas, lmv->datas_size);
1013         OBD_FREE(lmv->tgts, lmv->tgts_size);
1014
1015         RETURN(0);
1016 }
1017
1018 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1019 {
1020         struct lustre_cfg *lcfg = buf;
1021         struct obd_uuid tgt_uuid;
1022         int rc;
1023         ENTRY;
1024
1025         switch(lcfg->lcfg_command) {
1026         case LCFG_ADD_MDC:
1027                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1028                         GOTO(out, rc = -EINVAL);
1029
1030                 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1031                 rc = lmv_add_target(obd, &tgt_uuid);
1032                 GOTO(out, rc);
1033         default: {
1034                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1035                 GOTO(out, rc = -EINVAL);
1036         }
1037         }
1038 out:
1039         RETURN(rc);
1040 }
1041
1042 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1043                       __u64 max_age)
1044 {
1045         struct lmv_obd *lmv = &obd->u.lmv;
1046         struct obd_statfs *temp;
1047         int rc = 0, i;
1048         ENTRY;
1049
1050         rc = lmv_check_connect(obd);
1051         if (rc)
1052                 RETURN(rc);
1053
1054         OBD_ALLOC(temp, sizeof(*temp));
1055         if (temp == NULL)
1056                 RETURN(-ENOMEM);
1057
1058         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1059                 if (lmv->tgts[i].ltd_exp == NULL)
1060                         continue;
1061
1062                 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp, max_age);
1063                 if (rc) {
1064                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1065                                lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1066                                rc);
1067                         GOTO(out_free_temp, rc);
1068                 }
1069                 if (i == 0) {
1070                         *osfs = *temp;
1071                 } else {
1072                         osfs->os_bavail += temp->os_bavail;
1073                         osfs->os_blocks += temp->os_blocks;
1074                         osfs->os_ffree += temp->os_ffree;
1075                         osfs->os_files += temp->os_files;
1076                 }
1077         }
1078
1079         EXIT;
1080 out_free_temp:
1081         OBD_FREE(temp, sizeof(*temp));
1082         return rc;
1083 }
1084
1085 static int lmv_getstatus(struct obd_export *exp,
1086                          struct lu_fid *fid,
1087                          struct obd_capa **pc)
1088 {
1089         struct obd_device *obd = exp->exp_obd;
1090         struct lmv_obd *lmv = &obd->u.lmv;
1091         int rc;
1092         ENTRY;
1093
1094         rc = lmv_check_connect(obd);
1095         if (rc)
1096                 RETURN(rc);
1097
1098         rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1099
1100         RETURN(rc);
1101 }
1102
1103 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1104                         struct obd_capa *oc, obd_valid valid, const char *name,
1105                         const char *input, int input_size, int output_size,
1106                         int flags, struct ptlrpc_request **request)
1107 {
1108         struct obd_device *obd = exp->exp_obd;
1109         struct lmv_obd *lmv = &obd->u.lmv;
1110         struct obd_export *tgt_exp;
1111         int rc;
1112         ENTRY;
1113
1114         rc = lmv_check_connect(obd);
1115         if (rc)
1116                 RETURN(rc);
1117
1118         tgt_exp = lmv_find_export(lmv, fid);
1119         if (IS_ERR(tgt_exp))
1120                 RETURN(PTR_ERR(tgt_exp));
1121
1122         rc = md_getxattr(tgt_exp, fid, oc, valid, name, input, input_size,
1123                          output_size, flags, request);
1124
1125         RETURN(rc);
1126 }
1127
1128 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1129                         struct obd_capa *oc, obd_valid valid, const char *name,
1130                         const char *input, int input_size, int output_size,
1131                         int flags, __u32 suppgid,
1132                         struct ptlrpc_request **request)
1133 {
1134         struct obd_device *obd = exp->exp_obd;
1135         struct lmv_obd *lmv = &obd->u.lmv;
1136         struct obd_export *tgt_exp;
1137         int rc;
1138         ENTRY;
1139
1140         rc = lmv_check_connect(obd);
1141         if (rc)
1142                 RETURN(rc);
1143
1144         tgt_exp = lmv_find_export(lmv, fid);
1145         if (IS_ERR(tgt_exp))
1146                 RETURN(PTR_ERR(tgt_exp));
1147
1148         rc = md_setxattr(tgt_exp, fid, oc, valid, name,
1149                          input, input_size, output_size, flags, suppgid,
1150                          request);
1151
1152         RETURN(rc);
1153 }
1154
1155 static int lmv_getattr(struct obd_export *exp, const struct lu_fid *fid,
1156                        struct obd_capa *oc, obd_valid valid, int ea_size,
1157                        struct ptlrpc_request **request)
1158 {
1159         struct obd_device *obd = exp->exp_obd;
1160         struct lmv_obd *lmv = &obd->u.lmv;
1161         struct obd_export *tgt_exp;
1162         struct lmv_obj *obj;
1163         int rc, i;
1164         ENTRY;
1165
1166         rc = lmv_check_connect(obd);
1167         if (rc)
1168                 RETURN(rc);
1169
1170         tgt_exp = lmv_find_export(lmv, fid);
1171         if (IS_ERR(tgt_exp))
1172                 RETURN(PTR_ERR(tgt_exp));
1173
1174         rc = md_getattr(tgt_exp, fid, oc, valid, ea_size, request);
1175         if (rc)
1176                 RETURN(rc);
1177
1178         obj = lmv_obj_grab(obd, fid);
1179
1180         CDEBUG(D_OTHER, "GETATTR for "DFID" %s\n", PFID(fid),
1181                obj ? "(split)" : "");
1182
1183         /*
1184          * If object is split, then we loop over all the slaves and gather size
1185          * attribute. In ideal world we would have to gather also mds field from
1186          * all slaves, as object is spread over the cluster and this is
1187          * definitely interesting information and it is not good to loss it,
1188          * but...
1189          */
1190         if (obj) {
1191                 struct mdt_body *body;
1192
1193                 if (*request == NULL) {
1194                         lmv_obj_put(obj);
1195                         RETURN(rc);
1196                 }
1197
1198                 body = lustre_msg_buf((*request)->rq_repmsg, REPLY_REC_OFF,
1199                                       sizeof(*body));
1200                 LASSERT(body != NULL);
1201                 LASSERT(lustre_rep_swabbed(*request, REPLY_REC_OFF));
1202
1203                 lmv_obj_lock(obj);
1204
1205                 for (i = 0; i < obj->lo_objcount; i++) {
1206                         if (lmv->tgts[i].ltd_exp == NULL) {
1207                                 CWARN("%s: NULL export for %d\n",
1208                                       obd->obd_name, i);
1209                                 continue;
1210                         }
1211
1212                         /* skip master obj. */
1213                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_inodes[i].li_fid))
1214                                 continue;
1215
1216                         lmv_update_body(body, &obj->lo_inodes[i]);
1217                 }
1218
1219                 lmv_obj_unlock(obj);
1220                 lmv_obj_put(obj);
1221         }
1222
1223         RETURN(rc);
1224 }
1225
1226 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1227                              ldlm_iterator_t it, void *data)
1228 {
1229         struct obd_device *obd = exp->exp_obd;
1230         struct lmv_obd *lmv = &obd->u.lmv;
1231         int i, rc;
1232         ENTRY;
1233
1234         rc = lmv_check_connect(obd);
1235         if (rc)
1236                 RETURN(rc);
1237
1238         CDEBUG(D_OTHER, "CBDATA for "DFID"\n", PFID(fid));
1239
1240         /*
1241          * With CMD every object can have two locks in different namespaces:
1242          * lookup lock in space of mds storing direntry and update/open lock in
1243          * space of mds storing inode.
1244          */
1245         for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1246                 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1247
1248         RETURN(0);
1249 }
1250
1251 static int lmv_close(struct obd_export *exp,
1252                      struct md_op_data *op_data,
1253                      struct md_open_data *mod,
1254                      struct ptlrpc_request **request)
1255 {
1256         struct obd_device *obd = exp->exp_obd;
1257         struct lmv_obd *lmv = &obd->u.lmv;
1258         struct obd_export *tgt_exp;
1259         int rc;
1260         ENTRY;
1261
1262         rc = lmv_check_connect(obd);
1263         if (rc)
1264                 RETURN(rc);
1265
1266         tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1267         if (IS_ERR(tgt_exp))
1268                 RETURN(PTR_ERR(tgt_exp));
1269
1270         CDEBUG(D_OTHER, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1271         rc = md_close(tgt_exp, op_data, mod, request);
1272         RETURN(rc);
1273 }
1274
1275 /*
1276  * Called in the case MDS returns -ERESTART on create on open, what means that
1277  * directory is split and its LMV presentation object has to be updated.
1278  */
1279 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1280 {
1281         struct obd_device *obd = exp->exp_obd;
1282         struct lmv_obd *lmv = &obd->u.lmv;
1283         struct ptlrpc_request *req = NULL;
1284         struct obd_export *tgt_exp;
1285         struct lmv_obj *obj;
1286         struct lustre_md md;
1287         int mealen, rc;
1288         __u64 valid;
1289         ENTRY;
1290
1291         md.mea = NULL;
1292         mealen = lmv_get_easize(lmv);
1293
1294         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1295
1296         tgt_exp = lmv_find_export(lmv, fid);
1297         if (IS_ERR(tgt_exp))
1298                 RETURN(PTR_ERR(tgt_exp));
1299
1300         /* time to update mea of parent fid */
1301         rc = md_getattr(tgt_exp, fid, NULL, valid, mealen, &req);
1302         if (rc) {
1303                 CERROR("md_getattr() failed, error %d\n", rc);
1304                 GOTO(cleanup, rc);
1305         }
1306
1307         rc = md_get_lustre_md(tgt_exp, req, 1, NULL, exp, &md);
1308         if (rc) {
1309                 CERROR("mdc_get_lustre_md() failed, error %d\n", rc);
1310                 GOTO(cleanup, rc);
1311         }
1312
1313         if (md.mea == NULL)
1314                 GOTO(cleanup, rc = -ENODATA);
1315
1316         obj = lmv_obj_create(exp, fid, md.mea);
1317         if (IS_ERR(obj))
1318                 rc = PTR_ERR(obj);
1319         else
1320                 lmv_obj_put(obj);
1321
1322         obd_free_memmd(exp, (struct lov_stripe_md **)&md.mea);
1323
1324         EXIT;
1325 cleanup:
1326         if (req)
1327                 ptlrpc_req_finished(req);
1328         return rc;
1329 }
1330
1331 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1332                const void *data, int datalen, int mode, __u32 uid,
1333                __u32 gid, __u32 cap_effective,  __u64 rdev,
1334                struct ptlrpc_request **request)
1335 {
1336         struct obd_device *obd = exp->exp_obd;
1337         struct lmv_obd *lmv = &obd->u.lmv;
1338         struct obd_export *tgt_exp;
1339         struct lmv_obj *obj;
1340         int rc, loop = 0;
1341         ENTRY;
1342
1343         rc = lmv_check_connect(obd);
1344         if (rc)
1345                 RETURN(rc);
1346
1347         if (!lmv->desc.ld_active_tgt_count)
1348                 RETURN(-EIO);
1349 repeat:
1350         ++loop;
1351         LASSERT(loop <= 2);
1352         obj = lmv_obj_grab(obd, &op_data->op_fid1);
1353         if (obj) {
1354                 int mea_idx;
1355
1356                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1357                                        op_data->op_name, op_data->op_namelen);
1358                 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1359                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1360                 op_data->op_mds = obj->lo_inodes[mea_idx].li_mds;
1361                 tgt_exp = lmv_get_export(lmv, op_data->op_mds);
1362                 lmv_obj_put(obj);
1363         } else {
1364                 struct lmv_tgt_desc *tgt;
1365
1366                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1367                 op_data->op_bias |= MDS_CHECK_SPLIT;
1368                 op_data->op_mds = tgt->ltd_idx;
1369                 tgt_exp = tgt->ltd_exp;
1370         }
1371
1372         if (IS_ERR(tgt_exp))
1373                 RETURN(PTR_ERR(tgt_exp));
1374
1375         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1376         if (rc == -ERESTART)
1377                 goto repeat;
1378         else if (rc)
1379                 RETURN(rc);
1380
1381         CDEBUG(D_OTHER, "CREATE '%*s' on "DFID"\n", op_data->op_namelen,
1382                op_data->op_name, PFID(&op_data->op_fid1));
1383
1384         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1385         rc = md_create(tgt_exp, op_data, data, datalen, mode, uid, gid,
1386                        cap_effective, rdev, request);
1387         if (rc == 0) {
1388                 if (*request == NULL)
1389                         RETURN(rc);
1390                 CDEBUG(D_OTHER, "created - "DFID"\n", PFID(&op_data->op_fid1));
1391         } else if (rc == -ERESTART) {
1392                 LASSERT(*request != NULL);
1393                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1394                           "Got -ERESTART during create!\n");
1395                 ptlrpc_req_finished(*request);
1396                 *request = NULL;
1397
1398                 /*
1399                  * Directory got split. Time to update local object and repeat
1400                  * the request with proper MDS.
1401                  */
1402                 rc = lmv_handle_split(exp, &op_data->op_fid1);
1403                 if (rc == 0) {
1404                         rc = lmv_alloc_slave_fids(obd, &op_data->op_fid1,
1405                                                   op_data, &op_data->op_fid2);
1406                         if (rc)
1407                                 RETURN(rc);
1408                         goto repeat;
1409                 }
1410         }
1411         RETURN(rc);
1412 }
1413
1414 static int lmv_done_writing(struct obd_export *exp,
1415                             struct md_op_data *op_data,
1416                             struct md_open_data *mod)
1417 {
1418         struct obd_device *obd = exp->exp_obd;
1419         struct lmv_obd *lmv = &obd->u.lmv;
1420         struct obd_export *tgt_exp;
1421         int rc;
1422         ENTRY;
1423
1424         rc = lmv_check_connect(obd);
1425         if (rc)
1426                 RETURN(rc);
1427
1428         tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1429         if (IS_ERR(tgt_exp))
1430                 RETURN(PTR_ERR(tgt_exp));
1431
1432         rc = md_done_writing(tgt_exp, op_data, mod);
1433         RETURN(rc);
1434 }
1435
1436 static int
1437 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1438                    struct lookup_intent *it, struct md_op_data *op_data,
1439                    struct lustre_handle *lockh, void *lmm, int lmmsize)
1440 {
1441         struct obd_device *obd = exp->exp_obd;
1442         struct lmv_obd *lmv = &obd->u.lmv;
1443         struct lmv_stripe_md *mea = op_data->op_mea1;
1444         struct md_op_data *op_data2;
1445         struct obd_export *tgt_exp;
1446         int i, rc = 0;
1447         ENTRY;
1448
1449         OBD_ALLOC_PTR(op_data2);
1450         if (op_data2 == NULL)
1451                 RETURN(-ENOMEM);
1452
1453         LASSERT(mea != NULL);
1454         for (i = 0; i < mea->mea_count; i++) {
1455                 memset(op_data2, 0, sizeof(*op_data2));
1456                 op_data2->op_fid1 = mea->mea_ids[i];
1457                 op_data2->op_bias = 0;
1458
1459                 tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1);
1460                 if (IS_ERR(tgt_exp))
1461                         GOTO(cleanup, rc = PTR_ERR(tgt_exp));
1462
1463                 if (tgt_exp == NULL)
1464                         continue;
1465
1466                 rc = md_enqueue(tgt_exp, einfo, it, op_data2,
1467                                 lockh + i, lmm, lmmsize, 0);
1468
1469                 CDEBUG(D_OTHER, "take lock on slave "DFID" -> %d/%d\n",
1470                        PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1471
1472                 if (rc)
1473                         GOTO(cleanup, rc);
1474
1475                 if (it->d.lustre.it_data) {
1476                         struct ptlrpc_request *req;
1477                         req = (struct ptlrpc_request *)it->d.lustre.it_data;
1478                         ptlrpc_req_finished(req);
1479                 }
1480
1481                 if (it->d.lustre.it_status)
1482                         GOTO(cleanup, rc = it->d.lustre.it_status);
1483         }
1484
1485         EXIT;
1486 cleanup:
1487         OBD_FREE_PTR(op_data2);
1488
1489         if (rc != 0) {
1490                 /* drop all taken locks */
1491                 while (--i >= 0) {
1492                         if (lockh[i].cookie)
1493                                 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1494                         lockh[i].cookie = 0;
1495                 }
1496         }
1497         return rc;
1498 }
1499
1500 static int
1501 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1502                    struct lookup_intent *it, struct md_op_data *op_data,
1503                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1504                    int extra_lock_flags)
1505 {
1506         struct ptlrpc_request *req = it->d.lustre.it_data;
1507         struct obd_device *obd = exp->exp_obd;
1508         struct lmv_obd *lmv = &obd->u.lmv;
1509         struct lustre_handle plock;
1510         struct obd_export *tgt_exp;
1511         struct md_op_data *rdata;
1512         struct lu_fid fid_copy;
1513         struct mdt_body *body;
1514         int rc = 0, pmode;
1515         ENTRY;
1516
1517         body = lustre_msg_buf(req->rq_repmsg,
1518                               DLM_REPLY_REC_OFF, sizeof(*body));
1519         LASSERT(body != NULL);
1520         LASSERT(lustre_rep_swabbed(req, DLM_REPLY_REC_OFF));
1521
1522         if (!(body->valid & OBD_MD_MDS))
1523                 RETURN(0);
1524
1525         CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID" -> "DFID"\n",
1526                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1527
1528         /* We got LOOKUP lock, but we really need attrs */
1529         pmode = it->d.lustre.it_lock_mode;
1530         LASSERT(pmode != 0);
1531         memcpy(&plock, lockh, sizeof(plock));
1532         it->d.lustre.it_lock_mode = 0;
1533         it->d.lustre.it_data = NULL;
1534         fid_copy = body->fid1;
1535
1536         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1537         ptlrpc_req_finished(req);
1538
1539         tgt_exp = lmv_find_export(lmv, &fid_copy);
1540         if (IS_ERR(tgt_exp))
1541                 GOTO(out, rc = PTR_ERR(tgt_exp));
1542
1543         OBD_ALLOC_PTR(rdata);
1544         if (rdata == NULL)
1545                 GOTO(out, rc = -ENOMEM);
1546
1547         rdata->op_fid1 = fid_copy;
1548         rdata->op_bias = MDS_CROSS_REF;
1549
1550         rc = md_enqueue(tgt_exp, einfo, it, rdata, lockh,
1551                         lmm, lmmsize, extra_lock_flags);
1552         OBD_FREE_PTR(rdata);
1553         EXIT;
1554 out:
1555         ldlm_lock_decref(&plock, pmode);
1556         return rc;
1557 }
1558
1559 static int
1560 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1561             struct lookup_intent *it, struct md_op_data *op_data,
1562             struct lustre_handle *lockh, void *lmm, int lmmsize,
1563             int extra_lock_flags)
1564 {
1565         struct obd_device *obd = exp->exp_obd;
1566         struct lmv_obd *lmv = &obd->u.lmv;
1567         struct obd_export *tgt_exp = NULL;
1568         struct lmv_obj *obj;
1569         int rc;
1570         ENTRY;
1571
1572         rc = lmv_check_connect(obd);
1573         if (rc)
1574                 RETURN(rc);
1575
1576         if (op_data->op_mea1 && it->it_op == IT_UNLINK) {
1577                 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1578                                         lockh, lmm, lmmsize);
1579                 RETURN(rc);
1580         }
1581
1582         if (op_data->op_namelen) {
1583                 obj = lmv_obj_grab(obd, &op_data->op_fid1);
1584                 if (obj) {
1585                         int mea_idx;
1586
1587                         /* directory is split. look for right mds for this
1588                          * name */
1589                         mea_idx = raw_name2idx(obj->lo_hashtype,
1590                                                obj->lo_objcount,
1591                                                (char *)op_data->op_name,
1592                                                op_data->op_namelen);
1593                         op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1594                         tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds);
1595                         lmv_obj_put(obj);
1596                 }
1597         }
1598
1599         if (tgt_exp == NULL)
1600                 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1601         if (IS_ERR(tgt_exp))
1602                 RETURN(PTR_ERR(tgt_exp));
1603
1604         CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID"\n", LL_IT2STR(it),
1605                PFID(&op_data->op_fid1));
1606
1607         rc = md_enqueue(tgt_exp, einfo, it, op_data, lockh,
1608                         lmm, lmmsize, extra_lock_flags);
1609
1610         if (rc == 0 && it->it_op == IT_OPEN)
1611                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1612                                         lmm, lmmsize, extra_lock_flags);
1613         RETURN(rc);
1614 }
1615
1616 static int
1617 lmv_getattr_name(struct obd_export *exp, const struct lu_fid *fid,
1618                  struct obd_capa *oc, const char *filename, int namelen,
1619                  obd_valid valid, int ea_size, __u32 suppgid,
1620                  struct ptlrpc_request **request)
1621 {
1622         struct obd_device *obd = exp->exp_obd;
1623         struct lmv_obd *lmv = &obd->u.lmv;
1624         struct lu_fid rid = *fid;
1625         struct obd_export *tgt_exp;
1626         struct mdt_body *body;
1627         struct lmv_obj *obj;
1628         int rc, loop = 0;
1629         ENTRY;
1630
1631         rc = lmv_check_connect(obd);
1632         if (rc)
1633                 RETURN(rc);
1634
1635 repeat:
1636         ++loop;
1637         LASSERT(loop <= 2);
1638         obj = lmv_obj_grab(obd, &rid);
1639         if (obj) {
1640                 int mea_idx;
1641
1642                 /* Directory is split. Look for right mds for this name */
1643                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1644                                        filename, namelen - 1);
1645                 rid = obj->lo_inodes[mea_idx].li_fid;
1646                 tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds);
1647                 lmv_obj_put(obj);
1648                 valid &= ~OBD_MD_FLCKSPLIT;
1649         } else {
1650                 tgt_exp = lmv_find_export(lmv, &rid);
1651                 valid |= OBD_MD_FLCKSPLIT;
1652         }
1653         if (IS_ERR(tgt_exp))
1654                 RETURN(PTR_ERR(tgt_exp));
1655
1656         CDEBUG(D_OTHER, "getattr_name for %*s on "DFID" -> "DFID"\n",
1657                namelen, filename, PFID(fid), PFID(&rid));
1658
1659         rc = md_getattr_name(tgt_exp, &rid, oc, filename, namelen, valid,
1660                              ea_size, suppgid, request);
1661         if (rc == 0) {
1662                 body = lustre_msg_buf((*request)->rq_repmsg,
1663                                       REQ_REC_OFF, sizeof(*body));
1664                 LASSERT(body != NULL);
1665                 LASSERT(lustre_rep_swabbed(*request, REQ_REC_OFF));
1666
1667                 if (body->valid & OBD_MD_MDS) {
1668                         struct ptlrpc_request *req = NULL;
1669
1670                         rid = body->fid1;
1671                         CDEBUG(D_OTHER, "request attrs for "DFID"\n",
1672                                PFID(&rid));
1673
1674                         tgt_exp = lmv_find_export(lmv, &rid);
1675                         if (IS_ERR(tgt_exp)) {
1676                                 ptlrpc_req_finished(*request);
1677                                 RETURN(PTR_ERR(tgt_exp));
1678                         }
1679
1680                         rc = md_getattr_name(tgt_exp, &rid, NULL, NULL, 1,
1681                                              valid | OBD_MD_FLCROSSREF,
1682                                              ea_size, suppgid, &req);
1683                         ptlrpc_req_finished(*request);
1684                         *request = req;
1685                 }
1686         } else if (rc == -ERESTART) {
1687                 LASSERT(*request != NULL);
1688                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1689                           "Got -ERESTART during getattr!\n");
1690                 ptlrpc_req_finished(*request);
1691                 *request = NULL;
1692
1693                 /*
1694                  * Directory got split. Time to update local object and repeat
1695                  * the request with proper MDS.
1696                  */
1697                 rc = lmv_handle_split(exp, &rid);
1698                 if (rc == 0)
1699                         goto repeat;
1700         }
1701         RETURN(rc);
1702 }
1703
1704 #define md_op_data_fid(op_data, fl)                     \
1705         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1706          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1707          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1708          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1709          NULL)
1710
1711 /* @tgt_exp is the export the metadata request is sent.
1712  * @fid_exp is the export the cancel should be sent for the current fid.
1713  * if @fid_exp is NULL, the export is found for the current fid.
1714  * @op_data keeps the current fid, which is pointed through @flag.
1715  * @mode, @bits -- lock match parameters. */
1716 static int lmv_early_cancel(struct lmv_obd *lmv, struct obd_export *tgt_exp,
1717                             struct obd_export *fid_exp,
1718                             struct md_op_data *op_data,
1719                             ldlm_mode_t mode, int bits, int flag)
1720 {
1721         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1722         ldlm_policy_data_t policy = {{0}};
1723         int rc = 0;
1724         ENTRY;
1725
1726         if (!fid_is_sane(fid))
1727                 RETURN(0);
1728
1729         if (fid_exp == NULL)
1730                 fid_exp = lmv_find_export(lmv, fid);
1731
1732         if (tgt_exp == fid_exp) {
1733                 /* The export is the same as on the target server, cancel
1734                  * will be sent along with the main metadata operation. */
1735                 op_data->op_flags |= flag;
1736                 RETURN(0);
1737         }
1738
1739         policy.l_inodebits.bits = bits;
1740         rc = md_cancel_unused(fid_exp, fid, &policy, mode, LDLM_FL_ASYNC, NULL);
1741         RETURN(rc);
1742 }
1743
1744 #ifdef EARLY_CANCEL_FOR_STRIPED_DIR_IS_READY
1745 /* Check if the fid in @op_data pointed to by flag is of the same export(s)
1746  * as @tgt_exp. Early cancels will be sent later by mdc code, otherwise, call
1747  * md_cancel_unused for child export(s). */
1748 static int lmv_early_cancel_stripes(struct obd_export *exp,
1749                                     struct obd_export *tgt_exp,
1750                                     struct md_op_data *op_data,
1751                                     ldlm_mode_t mode, int bits, int flag)
1752 {
1753         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1754         struct obd_device *obd = exp->exp_obd;
1755         struct lmv_obd *lmv = &obd->u.lmv;
1756         struct obd_export *st_exp;
1757         struct lmv_obj *obj;
1758         int rc = 0;
1759         ENTRY;
1760
1761         if (!fid_is_sane(fid))
1762                 RETURN(0);
1763
1764         obj = lmv_obj_grab(obd, fid);
1765         if (obj) {
1766                 ldlm_policy_data_t policy = {{0}};
1767                 struct lu_fid *st_fid;
1768                 int i;
1769
1770                 policy.l_inodebits.bits = bits;
1771                 for (i = 0; i < obj->lo_objcount; i++) {
1772                         st_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds);
1773                         st_fid = &obj->lo_inodes[i].li_fid;
1774                         if (tgt_exp != st_exp) {
1775                                 rc = md_cancel_unused(st_exp, st_fid, &policy,
1776                                                       mode, LDLM_FL_ASYNC,
1777                                                       NULL);
1778                                 if (rc)
1779                                         break;
1780                         } else {
1781                                 /* Some export matches to @tgt_exp, do cancel
1782                                  * for its fid in mdc */
1783                                 *fid = *st_fid;
1784                                 op_data->op_flags |= flag;
1785                         }
1786                 }
1787                 lmv_obj_put(obj);
1788         } else {
1789                 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data,
1790                                       mode, bits, flag);
1791         }
1792         RETURN(rc);
1793 }
1794 #endif
1795
1796 /*
1797  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1798  * op_data->op_fid2
1799  */
1800 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1801                     struct ptlrpc_request **request)
1802 {
1803         struct obd_device *obd = exp->exp_obd;
1804         struct lmv_obd *lmv = &obd->u.lmv;
1805         struct obd_export *tgt_exp;
1806         struct lmv_obj *obj;
1807         int rc, loop = 0;
1808         mdsno_t mds;
1809         ENTRY;
1810
1811         rc = lmv_check_connect(obd);
1812         if (rc)
1813                 RETURN(rc);
1814
1815 repeat:
1816         ++loop;
1817         LASSERT(loop <= 2);
1818         if (op_data->op_namelen != 0) {
1819                 int mea_idx;
1820
1821                 /* Usual link request */
1822                 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1823                 if (obj) {
1824                         mea_idx = raw_name2idx(obj->lo_hashtype,
1825                                                obj->lo_objcount,
1826                                                op_data->op_name,
1827                                                op_data->op_namelen);
1828                         op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1829                         mds = obj->lo_inodes[mea_idx].li_mds;
1830                         lmv_obj_put(obj);
1831                 } else {
1832                         rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1833                         if (rc)
1834                                 RETURN(rc);
1835                 }
1836
1837                 CDEBUG(D_OTHER,"link "DFID":%*s to "DFID"\n",
1838                        PFID(&op_data->op_fid2), op_data->op_namelen,
1839                        op_data->op_name, PFID(&op_data->op_fid1));
1840         } else {
1841                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds);
1842                 if (rc)
1843                         RETURN(rc);
1844
1845                 /* request from MDS to acquire i_links for inode by fid1 */
1846                 CDEBUG(D_OTHER, "inc i_nlinks for "DFID"\n",
1847                        PFID(&op_data->op_fid1));
1848         }
1849
1850         CDEBUG(D_OTHER, "forward to MDS #"LPU64" ("DFID")\n",
1851                mds, PFID(&op_data->op_fid1));
1852
1853         op_data->op_fsuid = current->fsuid;
1854         op_data->op_fsgid = current->fsgid;
1855         op_data->op_cap   = current->cap_effective;
1856
1857         tgt_exp = lmv->tgts[mds].ltd_exp;
1858         if (op_data->op_namelen) {
1859                 op_data->op_flags |= MF_MDC_CANCEL_FID2;
1860                 /* Cancel UPDATE lock on child (fid1). */
1861                 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX,
1862                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1863         }
1864         if (rc == 0)
1865                 rc = md_link(tgt_exp, op_data, request);
1866         if (rc == -ERESTART) {
1867                 LASSERT(*request != NULL);
1868                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1869                           "Got -ERESTART during link!\n");
1870                 ptlrpc_req_finished(*request);
1871                 *request = NULL;
1872
1873                 /*
1874                  * Directory got split. Time to update local object and repeat
1875                  * the request with proper MDS.
1876                  */
1877                 rc = lmv_handle_split(exp, &op_data->op_fid2);
1878                 if (rc == 0)
1879                         goto repeat;
1880         }
1881
1882         RETURN(rc);
1883 }
1884
1885 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
1886                       const char *old, int oldlen, const char *new, int newlen,
1887                       struct ptlrpc_request **request)
1888 {
1889         struct obd_export *tgt_exp = NULL, *src_exp;
1890         struct obd_device *obd = exp->exp_obd;
1891         struct lmv_obd *lmv = &obd->u.lmv;
1892         int rc, mea_idx, loop = 0;
1893         struct lmv_obj *obj;
1894         mdsno_t mds1, mds2;
1895         ENTRY;
1896
1897         CDEBUG(D_OTHER, "rename %*s in "DFID" to %*s in "DFID"\n",
1898                oldlen, old, PFID(&op_data->op_fid1),
1899                newlen, new, PFID(&op_data->op_fid2));
1900
1901         rc = lmv_check_connect(obd);
1902         if (rc)
1903                 RETURN(rc);
1904
1905         if (oldlen == 0) {
1906                 /*
1907                  * MDS with old dir entry is asking another MDS to create name
1908                  * there.
1909                  */
1910                 CDEBUG(D_OTHER,
1911                        "create %*s(%d/%d) in "DFID" pointing "
1912                        "to "DFID"\n", newlen, new, oldlen, newlen,
1913                        PFID(&op_data->op_fid2), PFID(&op_data->op_fid1));
1914
1915                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds1);
1916                 if (rc)
1917                         RETURN(rc);
1918
1919                 /*
1920                  * Target directory can be split, sowe should forward request to
1921                  * the right MDS.
1922                  */
1923                 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1924                 if (obj) {
1925                         mea_idx = raw_name2idx(obj->lo_hashtype,
1926                                                obj->lo_objcount,
1927                                                (char *)new, newlen);
1928                         op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1929                         CDEBUG(D_OTHER, "Parent obj "DFID"\n",
1930                                PFID(&op_data->op_fid2));
1931                         lmv_obj_put(obj);
1932                 }
1933                 goto request;
1934         }
1935
1936 repeat:
1937         ++loop;
1938         LASSERT(loop <= 2);
1939         obj = lmv_obj_grab(obd, &op_data->op_fid1);
1940         if (obj) {
1941                 /*
1942                  * directory is already split, so we have to forward request to
1943                  * the right MDS.
1944                  */
1945                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1946                                        (char *)old, oldlen);
1947                 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1948                 mds1 = obj->lo_inodes[mea_idx].li_mds;
1949                 CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
1950                 lmv_obj_put(obj);
1951         } else {
1952                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
1953                 if (rc)
1954                         RETURN(rc);
1955         }
1956
1957         obj = lmv_obj_grab(obd, &op_data->op_fid2);
1958         if (obj) {
1959                 /*
1960                  * Directory is already split, so we have to forward request to
1961                  * the right MDS.
1962                  */
1963                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1964                                        (char *)new, newlen);
1965
1966                 mds2 = obj->lo_inodes[mea_idx].li_mds;
1967                 op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1968                 CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
1969                 lmv_obj_put(obj);
1970         } else {
1971                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
1972                 if (rc)
1973                         RETURN(rc);
1974         }
1975
1976 request:
1977         op_data->op_fsuid = current->fsuid;
1978         op_data->op_fsgid = current->fsgid;
1979         op_data->op_cap   = current->cap_effective;
1980
1981         src_exp = lmv_get_export(lmv, mds1);
1982         tgt_exp = lmv_get_export(lmv, mds2);
1983         if (oldlen) {
1984                 /* LOOKUP lock on src child (fid3) should also be cancelled for
1985                  * src_exp in mdc_rename. */
1986                 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
1987
1988                 /* Cancel UPDATE locks on tgt parent (fid2), tgt_exp is its
1989                  * own export. */
1990                 rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data, LCK_EX,
1991                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
1992
1993                 /* Cancel LOOKUP locks on tgt child (fid4) for parent tgt_exp.*/
1994                 if (rc == 0)
1995                         rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data,
1996                                               LCK_EX, MDS_INODELOCK_LOOKUP,
1997                                               MF_MDC_CANCEL_FID4);
1998
1999                 /* XXX: the case when child is a striped dir is not supported.
2000                  * Only the master stripe has all locks cancelled early. */
2001                 /* Cancel all the locks on tgt child (fid4). */
2002                 if (rc == 0)
2003                         rc = lmv_early_cancel(lmv, src_exp, NULL, op_data,
2004                                               LCK_EX, MDS_INODELOCK_FULL,
2005                                               MF_MDC_CANCEL_FID4);
2006         }
2007
2008         if (rc == 0)
2009                 rc = md_rename(src_exp, op_data, old, oldlen,
2010                                new, newlen, request);
2011         if (rc == -ERESTART) {
2012                 LASSERT(*request != NULL);
2013                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2014                           "Got -ERESTART during rename!\n");
2015                 ptlrpc_req_finished(*request);
2016                 *request = NULL;
2017
2018                 /*
2019                  * Directory got split. Time to update local object and repeat
2020                  * the request with proper MDS.
2021                  */
2022                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2023                 if (rc == 0)
2024                         goto repeat;
2025         }
2026         RETURN(rc);
2027 }
2028
2029 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2030                        void *ea, int ealen, void *ea2, int ea2len,
2031                        struct ptlrpc_request **request,
2032                        struct md_open_data **mod)
2033 {
2034         struct obd_device *obd = exp->exp_obd;
2035         struct lmv_obd *lmv = &obd->u.lmv;
2036         struct ptlrpc_request *req;
2037         struct obd_export *tgt_exp;
2038         struct lmv_obj *obj;
2039         int rc = 0, i;
2040         ENTRY;
2041
2042         rc = lmv_check_connect(obd);
2043         if (rc)
2044                 RETURN(rc);
2045
2046         obj = lmv_obj_grab(obd, &op_data->op_fid1);
2047
2048         CDEBUG(D_OTHER, "SETATTR for "DFID", valid 0x%x%s\n",
2049                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2050                obj ? ", split" : "");
2051
2052         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2053         if (obj) {
2054                 for (i = 0; i < obj->lo_objcount; i++) {
2055                         op_data->op_fid1 = obj->lo_inodes[i].li_fid;
2056
2057                         tgt_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds);
2058                         if (IS_ERR(tgt_exp)) {
2059                                 rc = PTR_ERR(tgt_exp);
2060                                 break;
2061                         }
2062
2063                         rc = md_setattr(tgt_exp, op_data, ea, ealen,
2064                                         ea2, ea2len, &req, mod);
2065
2066                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_inodes[i].li_fid)) {
2067                                 /*
2068                                  * this is master object and this request should
2069                                  * be returned back to llite.
2070                                  */
2071                                 *request = req;
2072                         } else {
2073                                 ptlrpc_req_finished(req);
2074                         }
2075
2076                         if (rc)
2077                                 break;
2078                 }
2079                 lmv_obj_put(obj);
2080         } else {
2081                 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
2082                 if (IS_ERR(tgt_exp))
2083                         RETURN(PTR_ERR(tgt_exp));
2084
2085                 rc = md_setattr(tgt_exp, op_data, ea, ealen, ea2,
2086                                 ea2len, request, mod);
2087         }
2088         RETURN(rc);
2089 }
2090
2091 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2092                     struct obd_capa *oc, struct ptlrpc_request **request)
2093 {
2094         struct obd_device *obd = exp->exp_obd;
2095         struct lmv_obd *lmv = &obd->u.lmv;
2096         struct obd_export *tgt_exp;
2097         int rc;
2098         ENTRY;
2099
2100         rc = lmv_check_connect(obd);
2101         if (rc)
2102                 RETURN(rc);
2103
2104         tgt_exp = lmv_find_export(lmv, fid);
2105         if (IS_ERR(tgt_exp))
2106                 RETURN(PTR_ERR(tgt_exp));
2107
2108         rc = md_sync(tgt_exp, fid, oc, request);
2109         RETURN(rc);
2110 }
2111
2112 /* main purpose of LMV blocking ast is to remove split directory LMV
2113  * presentation object (struct lmv_obj) attached to the lock being revoked. */
2114 int lmv_blocking_ast(struct ldlm_lock *lock,
2115                      struct ldlm_lock_desc *desc,
2116                      void *data, int flag)
2117 {
2118         struct lustre_handle lockh;
2119         struct lmv_obj *obj;
2120         int rc;
2121         ENTRY;
2122
2123         switch (flag) {
2124         case LDLM_CB_BLOCKING:
2125                 ldlm_lock2handle(lock, &lockh);
2126                 rc = ldlm_cli_cancel(&lockh);
2127                 if (rc < 0) {
2128                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2129                         RETURN(rc);
2130                 }
2131                 break;
2132         case LDLM_CB_CANCELING:
2133                 /* time to drop cached attrs for dirobj */
2134                 obj = lock->l_ast_data;
2135                 if (obj) {
2136                         CDEBUG(D_OTHER, "cancel %s on "LPU64"/"LPU64
2137                                ", master "DFID"\n",
2138                                lock->l_resource->lr_name.name[3] == 1 ?
2139                                "LOOKUP" : "UPDATE",
2140                                lock->l_resource->lr_name.name[0],
2141                                lock->l_resource->lr_name.name[1],
2142                                PFID(&obj->lo_fid));
2143                         lmv_obj_put(obj);
2144                 }
2145                 break;
2146         default:
2147                 LBUG();
2148         }
2149         RETURN(0);
2150 }
2151
2152 static void lmv_hash_adjust(__u32 *hash, __u32 hash_adj)
2153 {
2154         __u32 val;
2155
2156         val = le32_to_cpu(*hash);
2157         if (val < hash_adj)
2158                 val += MAX_HASH_SIZE;
2159         if (val != DIR_END_OFF)
2160                 *hash = cpu_to_le32(val - hash_adj);
2161 }
2162
2163 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2164 {
2165         __u64 id;
2166         struct obd_import *imp;
2167
2168         /*
2169          * XXX Hack: to get nid we assume that underlying obd device is mdc.
2170          */
2171         imp  = class_exp2cliimp(exp);
2172         id   = imp->imp_connection->c_self + fid_flatten(fid);
2173
2174         CDEBUG(D_INFO, "node rank: %llx "DFID" %llx %llx\n",
2175                imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2176
2177         return id ^ (id >> 32);
2178 }
2179
2180 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2181                         struct obd_capa *oc, __u64 offset64, struct page *page,
2182                         struct ptlrpc_request **request)
2183 {
2184         struct obd_device *obd = exp->exp_obd;
2185         struct lmv_obd *lmv = &obd->u.lmv;
2186         struct obd_export *tgt_exp;
2187         struct lu_fid rid = *fid;
2188         struct lmv_obj *obj;
2189         __u32 offset0;
2190         __u32 offset;
2191         __u32 hash_adj = 0;
2192         __u32 rank = 0;
2193         __u32 seg_size = 0;
2194         int tgt = 0;
2195         int tgt0 = 0;
2196         int rc;
2197         int nr = 0;
2198         ENTRY;
2199
2200         offset0 = offset = offset64;
2201         /*
2202          * Check that offset is representable by 32bit number.
2203          */
2204         LASSERT((__u64)offset == offset64);
2205
2206         rc = lmv_check_connect(obd);
2207         if (rc)
2208                 RETURN(rc);
2209
2210         CDEBUG(D_INFO, "READPAGE at %x from "DFID"\n", offset, PFID(&rid));
2211
2212         obj = lmv_obj_grab(obd, fid);
2213         if (obj) {
2214                 struct lmv_inode *loi;
2215
2216                 lmv_obj_lock(obj);
2217
2218                 nr       = obj->lo_objcount;
2219                 LASSERT(nr > 0);
2220                 seg_size = MAX_HASH_SIZE / nr;
2221                 loi      = obj->lo_inodes;
2222                 rank     = lmv_node_rank(lmv_get_export(lmv, loi[0].li_mds),
2223                                          fid) % nr;
2224                 tgt0     = (offset / seg_size) % nr;
2225                 tgt      = (tgt0 + rank) % nr;
2226
2227                 if (tgt < tgt0)
2228                         /*
2229                          * Wrap around.
2230                          *
2231                          * Last segment has unusual length due to division
2232                          * rounding.
2233                          */
2234                         hash_adj = MAX_HASH_SIZE - seg_size * nr;
2235                 else
2236                         hash_adj = 0;
2237
2238                 hash_adj += rank * seg_size;
2239
2240                 CDEBUG(D_INFO, "hash_adj: %x %x %x/%x -> %x/%x\n",
2241                        rank, hash_adj, offset, tgt0, offset + hash_adj, tgt);
2242
2243                 offset = (offset + hash_adj) % MAX_HASH_SIZE;
2244                 rid = obj->lo_inodes[tgt].li_fid;
2245                 tgt_exp = lmv_get_export(lmv, loi[tgt].li_mds);
2246
2247                 CDEBUG(D_INFO, "forward to "DFID" with offset %lu i %d\n",
2248                        PFID(&rid), (unsigned long)offset, tgt);
2249         } else
2250                 tgt_exp = lmv_find_export(lmv, &rid);
2251
2252         if (IS_ERR(tgt_exp))
2253                 GOTO(cleanup, rc = PTR_ERR(tgt_exp));
2254
2255         rc = md_readpage(tgt_exp, &rid, oc, offset, page, request);
2256         if (rc)
2257                 GOTO(cleanup, rc);
2258         if (obj) {
2259                 struct lu_dirpage *dp;
2260                 struct lu_dirent  *ent;
2261
2262                 dp = cfs_kmap(page);
2263
2264                 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2265                 lmv_hash_adjust(&dp->ldp_hash_end,   hash_adj);
2266                 LASSERT(cpu_to_le32(dp->ldp_hash_start) <= offset0);
2267
2268                 for (ent = lu_dirent_start(dp); ent != NULL;
2269                      ent = lu_dirent_next(ent))
2270                         lmv_hash_adjust(&ent->lde_hash, hash_adj);
2271
2272                 if (tgt0 != nr - 1) {
2273                         __u32 end;
2274
2275                         end = le32_to_cpu(dp->ldp_hash_end);
2276                         if (end == DIR_END_OFF) {
2277                                 dp->ldp_hash_end = cpu_to_le32(seg_size *
2278                                                                (tgt0 + 1));
2279                                 CDEBUG(D_INFO, ""DFID" reset end %x tgt %d\n",
2280                                        PFID(&rid),
2281                                        le32_to_cpu(dp->ldp_hash_end), tgt);
2282                         }
2283                 }
2284                 cfs_kunmap(page);
2285         }
2286         /*
2287          * Here we could remove "." and ".." from all pages which at not from
2288          * master. But MDS has only "." and ".." for master dir.
2289          */
2290         EXIT;
2291 cleanup:
2292         if (obj) {
2293                 lmv_obj_unlock(obj);
2294                 lmv_obj_put(obj);
2295         }
2296         return rc;
2297 }
2298
2299 static int lmv_unlink_slaves(struct obd_export *exp,
2300                              struct md_op_data *op_data,
2301                              struct ptlrpc_request **req)
2302 {
2303         struct obd_device *obd = exp->exp_obd;
2304         struct lmv_obd *lmv = &obd->u.lmv;
2305         struct lmv_stripe_md *mea = op_data->op_mea1;
2306         struct md_op_data *op_data2;
2307         struct obd_export *tgt_exp;
2308         int i, rc = 0;
2309         ENTRY;
2310
2311         OBD_ALLOC_PTR(op_data2);
2312         if (op_data2 == NULL)
2313                 RETURN(-ENOMEM);
2314
2315         op_data2->op_mode = S_IFDIR;
2316         op_data2->op_fsuid = current->fsuid;
2317         op_data2->op_fsgid = current->fsgid;
2318         op_data2->op_bias = 0;
2319
2320         LASSERT(mea != NULL);
2321         for (i = 0; i < mea->mea_count; i++) {
2322                 memset(op_data2, 0, sizeof(*op_data2));
2323                 op_data2->op_fid1 = mea->mea_ids[i];
2324                 tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1);
2325                 if (IS_ERR(tgt_exp))
2326                         GOTO(out_free_op_data2, rc = PTR_ERR(tgt_exp));
2327
2328                 if (tgt_exp == NULL)
2329                         continue;
2330
2331                 rc = md_unlink(tgt_exp, op_data2, req);
2332
2333                 CDEBUG(D_OTHER, "unlink slave "DFID" -> %d\n",
2334                        PFID(&mea->mea_ids[i]), rc);
2335
2336                 if (*req) {
2337                         ptlrpc_req_finished(*req);
2338                         *req = NULL;
2339                 }
2340                 if (rc)
2341                         GOTO(out_free_op_data2, rc);
2342         }
2343
2344         EXIT;
2345 out_free_op_data2:
2346         OBD_FREE_PTR(op_data2);
2347         return rc;
2348 }
2349
2350 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2351                       struct ptlrpc_request **request)
2352 {
2353         struct obd_device *obd = exp->exp_obd;
2354         struct lmv_obd *lmv = &obd->u.lmv;
2355         struct obd_export *tgt_exp = NULL;
2356         struct lmv_obj *obj;
2357         int rc, loop = 0;
2358         ENTRY;
2359
2360         rc = lmv_check_connect(obd);
2361         if (rc)
2362                 RETURN(rc);
2363
2364         if (op_data->op_namelen == 0 && op_data->op_mea1 != NULL) {
2365                 /* mds asks to remove slave objects */
2366                 rc = lmv_unlink_slaves(exp, op_data, request);
2367                 RETURN(rc);
2368         }
2369
2370 repeat:
2371         ++loop;
2372         LASSERT(loop <= 2);
2373         if (op_data->op_namelen != 0) {
2374                 int mea_idx;
2375
2376                 obj = lmv_obj_grab(obd, &op_data->op_fid1);
2377                 if (obj) {
2378                         mea_idx = raw_name2idx(obj->lo_hashtype,
2379                                                obj->lo_objcount,
2380                                                op_data->op_name,
2381                                                op_data->op_namelen);
2382                         op_data->op_bias &= ~MDS_CHECK_SPLIT;
2383                         op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
2384                         tgt_exp = lmv_get_export(lmv,
2385                                                  obj->lo_inodes[mea_idx].li_mds);
2386                         lmv_obj_put(obj);
2387                         CDEBUG(D_OTHER, "unlink '%*s' in "DFID" -> %u\n",
2388                                op_data->op_namelen, op_data->op_name,
2389                                PFID(&op_data->op_fid1), mea_idx);
2390                 }
2391         } else {
2392                 CDEBUG(D_OTHER, "drop i_nlink on "DFID"\n",
2393                        PFID(&op_data->op_fid1));
2394         }
2395         if (tgt_exp == NULL) {
2396                 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
2397                 if (IS_ERR(tgt_exp))
2398                         RETURN(PTR_ERR(tgt_exp));
2399                 op_data->op_bias |= MDS_CHECK_SPLIT;
2400         }
2401
2402         op_data->op_fsuid = current->fsuid;
2403         op_data->op_fsgid = current->fsgid;
2404         op_data->op_cap   = current->cap_effective;
2405
2406         /* If child's fid is given, cancel unused locks for it if it is from
2407          * another export than parent. */
2408         if (op_data->op_namelen) {
2409                 /* LOOKUP lock for child (fid3) should also be cancelled on
2410                  * parent tgt_exp in mdc_unlink(). */
2411                 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2412
2413                 /* XXX: the case when child is a striped dir is not supported.
2414                  * Only the master stripe has all locks cancelled early. */
2415                 /* Cancel FULL locks on child (fid3). */
2416                 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX,
2417                                       MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2418         }
2419         if (rc == 0)
2420                 rc = md_unlink(tgt_exp, op_data, request);
2421         if (rc == -ERESTART) {
2422                 LASSERT(*request != NULL);
2423                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2424                           "Got -ERESTART during unlink!\n");
2425                 ptlrpc_req_finished(*request);
2426                 *request = NULL;
2427
2428                 /*
2429                  * Directory got split. Time to update local object and repeat
2430                  * the request with proper MDS.
2431                  */
2432                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2433                 if (rc == 0)
2434                         goto repeat;
2435         }
2436         RETURN(rc);
2437 }
2438
2439 static int lmv_llog_init(struct obd_device *obd, struct obd_llogs* llogs,
2440                          struct obd_device *tgt, int count,
2441                          struct llog_catid *logid, struct obd_uuid *uuid)
2442 {
2443         struct llog_ctxt *ctxt;
2444         int rc;
2445         ENTRY;
2446
2447         rc = llog_setup(obd, llogs, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
2448                         &llog_client_ops);
2449         if (rc == 0) {
2450                 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
2451                 ctxt->loc_imp = tgt->u.cli.cl_import;
2452         }
2453
2454         RETURN(rc);
2455 }
2456
2457 static int lmv_llog_finish(struct obd_device *obd, int count)
2458 {
2459         int rc;
2460         ENTRY;
2461
2462         rc = llog_cleanup(llog_get_context(obd, LLOG_CONFIG_REPL_CTXT));
2463         RETURN(rc);
2464 }
2465
2466 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2467 {
2468         int rc = 0;
2469
2470         switch (stage) {
2471         case OBD_CLEANUP_EARLY:
2472                 /* XXX: here should be calling obd_precleanup() down to
2473                  * stack. */
2474                 break;
2475         case OBD_CLEANUP_SELF_EXP:
2476                 rc = obd_llog_finish(obd, 0);
2477                 if (rc != 0)
2478                         CERROR("failed to cleanup llogging subsystems\n");
2479                 break;
2480         default:
2481                 break;
2482         }
2483         RETURN(rc);
2484 }
2485
2486 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2487                         void *key, __u32 *vallen, void *val)
2488 {
2489         struct obd_device *obd;
2490         struct lmv_obd *lmv;
2491         int rc = 0;
2492         ENTRY;
2493
2494         obd = class_exp2obd(exp);
2495         if (obd == NULL) {
2496                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2497                        exp->exp_handle.h_cookie);
2498                 RETURN(-EINVAL);
2499         }
2500
2501         lmv = &obd->u.lmv;
2502         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2503                 struct lmv_tgt_desc *tgts;
2504                 int i;
2505
2506                 rc = lmv_check_connect(obd);
2507                 if (rc)
2508                         RETURN(rc);
2509
2510                 LASSERT(*vallen == sizeof(__u32));
2511                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2512                      i++, tgts++) {
2513
2514                         /* all tgts should be connected when this get called. */
2515                         if (!tgts || !tgts->ltd_exp) {
2516                                 CERROR("target not setup?\n");
2517                                 continue;
2518                         }
2519
2520                         if (!obd_get_info(tgts->ltd_exp, keylen, key,
2521                                           vallen, val))
2522                                 RETURN(0);
2523                 }
2524                 RETURN(-EINVAL);
2525         } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2526                 rc = lmv_check_connect(obd);
2527                 if (rc)
2528                         RETURN(rc);
2529
2530                 /* forwarding this request to first MDS, it should know LOV
2531                  * desc. */
2532                 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2533                                   vallen, val);
2534                 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2535                         exp->exp_connect_flags =
2536                         ((struct obd_connect_data *)val)->ocd_connect_flags;
2537                 }
2538                 RETURN(rc);
2539         }
2540
2541         CDEBUG(D_IOCTL, "invalid key\n");
2542         RETURN(-EINVAL);
2543 }
2544
2545 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2546                        void *key, obd_count vallen, void *val,
2547                        struct ptlrpc_request_set *set)
2548 {
2549         struct lmv_tgt_desc    *tgt;
2550         struct obd_device      *obd;
2551         struct lmv_obd         *lmv;
2552         int rc = 0;
2553         ENTRY;
2554
2555         obd = class_exp2obd(exp);
2556         if (obd == NULL) {
2557                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2558                        exp->exp_handle.h_cookie);
2559                 RETURN(-EINVAL);
2560         }
2561         lmv = &obd->u.lmv;
2562
2563         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2564             KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2565                 int i, err = 0;
2566
2567                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2568                         tgt = &lmv->tgts[i];
2569
2570                         if (!tgt->ltd_exp)
2571                                 continue;
2572
2573                         err = obd_set_info_async(tgt->ltd_exp,
2574                                                  keylen, key, vallen, val, set);
2575                         if (err && rc == 0)
2576                                 rc = err;
2577                 }
2578
2579                 RETURN(rc);
2580         }
2581
2582         RETURN(-EINVAL);
2583 }
2584
2585 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2586                struct lov_stripe_md *lsm)
2587 {
2588         struct obd_device *obd = class_exp2obd(exp);
2589         struct lmv_obd *lmv = &obd->u.lmv;
2590         struct lmv_stripe_md *meap, *lsmp;
2591         int mea_size, i;
2592         ENTRY;
2593
2594         mea_size = lmv_get_easize(lmv);
2595         if (!lmmp)
2596                 RETURN(mea_size);
2597
2598         if (*lmmp && !lsm) {
2599                 OBD_FREE(*lmmp, mea_size);
2600                 *lmmp = NULL;
2601                 RETURN(0);
2602         }
2603
2604         if (*lmmp == NULL) {
2605                 OBD_ALLOC(*lmmp, mea_size);
2606                 if (*lmmp == NULL)
2607                         RETURN(-ENOMEM);
2608         }
2609
2610         if (!lsm)
2611                 RETURN(mea_size);
2612
2613         lsmp = (struct lmv_stripe_md *)lsm;
2614         meap = (struct lmv_stripe_md *)*lmmp;
2615
2616         if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2617             lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2618                 RETURN(-EINVAL);
2619
2620         meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2621         meap->mea_count = cpu_to_le32(lsmp->mea_count);
2622         meap->mea_master = cpu_to_le32(lsmp->mea_master);
2623
2624         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2625                 meap->mea_ids[i] = meap->mea_ids[i];
2626                 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2627         }
2628
2629         RETURN(mea_size);
2630 }
2631
2632 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2633                  struct lov_mds_md *lmm, int lmm_size)
2634 {
2635         struct obd_device *obd = class_exp2obd(exp);
2636         struct lmv_stripe_md **tmea = (struct lmv_stripe_md **)lsmp;
2637         struct lmv_stripe_md *mea = (struct lmv_stripe_md *)lmm;
2638         struct lmv_obd *lmv = &obd->u.lmv;
2639         int mea_size, i;
2640         __u32 magic;
2641         ENTRY;
2642
2643         mea_size = lmv_get_easize(lmv);
2644         if (lsmp == NULL)
2645                 return mea_size;
2646
2647         if (*lsmp != NULL && lmm == NULL) {
2648                 OBD_FREE(*tmea, mea_size);
2649                 *lsmp = NULL;
2650                 RETURN(0);
2651         }
2652
2653         LASSERT(mea_size == lmm_size);
2654
2655         OBD_ALLOC(*tmea, mea_size);
2656         if (*tmea == NULL)
2657                 RETURN(-ENOMEM);
2658
2659         if (!lmm)
2660                 RETURN(mea_size);
2661
2662         if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2663             mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2664             mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2665         {
2666                 magic = le32_to_cpu(mea->mea_magic);
2667         } else {
2668                 /* old mea is not handled here */
2669                 LBUG();
2670         }
2671
2672         (*tmea)->mea_magic = magic;
2673         (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2674         (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2675
2676         for (i = 0; i < (*tmea)->mea_count; i++) {
2677                 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2678                 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2679         }
2680         RETURN(mea_size);
2681 }
2682
2683 static int lmv_cancel_unused(struct obd_export *exp,
2684                              const struct lu_fid *fid,
2685                              ldlm_policy_data_t *policy,
2686                              ldlm_mode_t mode, int flags, void *opaque)
2687 {
2688         struct obd_device *obd = exp->exp_obd;
2689         struct lmv_obd *lmv = &obd->u.lmv;
2690         int rc = 0, err, i;
2691         ENTRY;
2692
2693         LASSERT(fid != NULL);
2694
2695         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2696                 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2697                         continue;
2698
2699                 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2700                                        policy, mode, flags, opaque);
2701                 if (!rc)
2702                         rc = err;
2703         }
2704         RETURN(rc);
2705 }
2706
2707 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data)
2708 {
2709         struct obd_device *obd = exp->exp_obd;
2710         struct lmv_obd *lmv = &obd->u.lmv;
2711
2712         ENTRY;
2713         RETURN(md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data));
2714 }
2715
2716 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2717                            const struct lu_fid *fid, ldlm_type_t type,
2718                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
2719                            struct lustre_handle *lockh)
2720 {
2721         struct obd_device *obd = exp->exp_obd;
2722         struct lmv_obd *lmv = &obd->u.lmv;
2723         ldlm_mode_t rc;
2724         int i;
2725         ENTRY;
2726
2727         CDEBUG(D_OTHER, "lock match for "DFID"\n", PFID(fid));
2728
2729         /* with CMD every object can have two locks in different namespaces:
2730          * lookup lock in space of mds storing direntry and update/open lock in
2731          * space of mds storing inode. Thus we check all targets, not only that
2732          * one fid was created in. */
2733         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2734                 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2735                                    type, policy, mode, lockh);
2736                 if (rc)
2737                         RETURN(rc);
2738         }
2739
2740         RETURN(0);
2741 }
2742
2743 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2744                       int offset, struct obd_export *dt_exp,
2745                       struct obd_export *md_exp, struct lustre_md *md)
2746 {
2747         struct obd_device *obd = exp->exp_obd;
2748         struct lmv_obd *lmv = &obd->u.lmv;
2749         int rc;
2750
2751         ENTRY;
2752         rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, offset, dt_exp, md_exp,
2753                               md);
2754         RETURN(rc);
2755 }
2756
2757 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2758 {
2759         struct obd_device *obd = exp->exp_obd;
2760         struct lmv_obd *lmv = &obd->u.lmv;
2761
2762         ENTRY;
2763         if (md->mea)
2764                 obd_free_memmd(exp, (struct lov_stripe_md**)&md->mea);
2765         RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2766 }
2767
2768 int lmv_set_open_replay_data(struct obd_export *exp,
2769                              struct obd_client_handle *och,
2770                              struct ptlrpc_request *open_req)
2771 {
2772         struct obd_device *obd = exp->exp_obd;
2773         struct lmv_obd *lmv = &obd->u.lmv;
2774         struct obd_export *tgt_exp;
2775
2776         ENTRY;
2777
2778         tgt_exp = lmv_find_export(lmv, &och->och_fid);
2779         if (IS_ERR(tgt_exp))
2780                 RETURN(PTR_ERR(tgt_exp));
2781
2782         RETURN(md_set_open_replay_data(tgt_exp, och, open_req));
2783 }
2784
2785 int lmv_clear_open_replay_data(struct obd_export *exp,
2786                                struct obd_client_handle *och)
2787 {
2788         struct obd_device *obd = exp->exp_obd;
2789         struct lmv_obd *lmv = &obd->u.lmv;
2790         struct obd_export *tgt_exp;
2791         ENTRY;
2792
2793         tgt_exp = lmv_find_export(lmv, &och->och_fid);
2794         if (IS_ERR(tgt_exp))
2795                 RETURN(PTR_ERR(tgt_exp));
2796
2797         RETURN(md_clear_open_replay_data(tgt_exp, och));
2798 }
2799
2800 static int lmv_get_remote_perm(struct obd_export *exp,
2801                                const struct lu_fid *fid,
2802                                struct obd_capa *oc, __u32 suppgid,
2803                                struct ptlrpc_request **request)
2804 {
2805         struct obd_device *obd = exp->exp_obd;
2806         struct lmv_obd *lmv = &obd->u.lmv;
2807         struct obd_export *tgt_exp;
2808         int rc;
2809
2810         ENTRY;
2811
2812         rc = lmv_check_connect(obd);
2813         if (rc)
2814                 RETURN(rc);
2815
2816         tgt_exp = lmv_find_export(lmv, fid);
2817         if (IS_ERR(tgt_exp))
2818                 RETURN(PTR_ERR(tgt_exp));
2819
2820         rc = md_get_remote_perm(tgt_exp, fid, oc, suppgid, request);
2821
2822         RETURN(rc);
2823 }
2824
2825 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2826                           renew_capa_cb_t cb)
2827 {
2828         struct obd_device *obd = exp->exp_obd;
2829         struct lmv_obd *lmv = &obd->u.lmv;
2830         struct obd_export *tgt_exp;
2831         int rc;
2832         ENTRY;
2833
2834         rc = lmv_check_connect(obd);
2835         if (rc)
2836                 RETURN(rc);
2837
2838         tgt_exp = lmv_find_export(lmv, &oc->c_capa.lc_fid);
2839         if (IS_ERR(tgt_exp))
2840                 RETURN(PTR_ERR(tgt_exp));
2841
2842         rc = md_renew_capa(tgt_exp, oc, cb);
2843         RETURN(rc);
2844 }
2845
2846 struct obd_ops lmv_obd_ops = {
2847         .o_owner                = THIS_MODULE,
2848         .o_setup                = lmv_setup,
2849         .o_cleanup              = lmv_cleanup,
2850         .o_precleanup           = lmv_precleanup,
2851         .o_process_config       = lmv_process_config,
2852         .o_connect              = lmv_connect,
2853         .o_disconnect           = lmv_disconnect,
2854         .o_statfs               = lmv_statfs,
2855         .o_llog_init            = lmv_llog_init,
2856         .o_llog_finish          = lmv_llog_finish,
2857         .o_get_info             = lmv_get_info,
2858         .o_set_info_async       = lmv_set_info_async,
2859         .o_packmd               = lmv_packmd,
2860         .o_unpackmd             = lmv_unpackmd,
2861         .o_notify               = lmv_notify,
2862         .o_iocontrol            = lmv_iocontrol,
2863         .o_fid_delete           = lmv_fid_delete
2864 };
2865
2866 struct md_ops lmv_md_ops = {
2867         .m_getstatus            = lmv_getstatus,
2868         .m_change_cbdata        = lmv_change_cbdata,
2869         .m_close                = lmv_close,
2870         .m_create               = lmv_create,
2871         .m_done_writing         = lmv_done_writing,
2872         .m_enqueue              = lmv_enqueue,
2873         .m_getattr              = lmv_getattr,
2874         .m_getxattr             = lmv_getxattr,
2875         .m_getattr_name         = lmv_getattr_name,
2876         .m_intent_lock          = lmv_intent_lock,
2877         .m_link                 = lmv_link,
2878         .m_rename               = lmv_rename,
2879         .m_setattr              = lmv_setattr,
2880         .m_setxattr             = lmv_setxattr,
2881         .m_sync                 = lmv_sync,
2882         .m_readpage             = lmv_readpage,
2883         .m_unlink               = lmv_unlink,
2884         .m_init_ea_size         = lmv_init_ea_size,
2885         .m_cancel_unused        = lmv_cancel_unused,
2886         .m_set_lock_data        = lmv_set_lock_data,
2887         .m_lock_match           = lmv_lock_match,
2888         .m_get_lustre_md        = lmv_get_lustre_md,
2889         .m_free_lustre_md       = lmv_free_lustre_md,
2890         .m_set_open_replay_data = lmv_set_open_replay_data,
2891         .m_clear_open_replay_data = lmv_clear_open_replay_data,
2892         .m_get_remote_perm      = lmv_get_remote_perm,
2893         .m_renew_capa           = lmv_renew_capa
2894 };
2895
2896 int __init lmv_init(void)
2897 {
2898         struct lprocfs_static_vars lvars;
2899         int rc;
2900
2901         obj_cache = cfs_mem_cache_create("lmv_objects",
2902                                       sizeof(struct lmv_obj),
2903                                       0, 0);
2904         if (!obj_cache) {
2905                 CERROR("error allocating lmv objects cache\n");
2906                 return -ENOMEM;
2907         }
2908
2909         lprocfs_lmv_init_vars(&lvars);
2910         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
2911                                  lvars.module_vars, LUSTRE_LMV_NAME, NULL);
2912         if (rc)
2913                 cfs_mem_cache_destroy(obj_cache);
2914
2915         return rc;
2916 }
2917
2918 #ifdef __KERNEL__
2919 static void lmv_exit(void)
2920 {
2921         int rc;
2922
2923         class_unregister_type(LUSTRE_LMV_NAME);
2924
2925         rc = cfs_mem_cache_destroy(obj_cache);
2926         LASSERTF(rc == 0,
2927                  "can't free lmv objects cache, %d object(s)"
2928                  "still in use\n", atomic_read(&obj_cache_count));
2929 }
2930
2931 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2932 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
2933 MODULE_LICENSE("GPL");
2934
2935 module_init(lmv_init);
2936 module_exit(lmv_exit);
2937 #endif