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 = req_capsule_server_get(&(*request)->rq_pill,
1199                                               &RMF_MDT_BODY);
1200                 LASSERT(body != NULL);
1201
1202                 lmv_obj_lock(obj);
1203
1204                 for (i = 0; i < obj->lo_objcount; i++) {
1205                         if (lmv->tgts[i].ltd_exp == NULL) {
1206                                 CWARN("%s: NULL export for %d\n",
1207                                       obd->obd_name, i);
1208                                 continue;
1209                         }
1210
1211                         /* skip master obj. */
1212                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_inodes[i].li_fid))
1213                                 continue;
1214
1215                         lmv_update_body(body, &obj->lo_inodes[i]);
1216                 }
1217
1218                 lmv_obj_unlock(obj);
1219                 lmv_obj_put(obj);
1220         }
1221
1222         RETURN(rc);
1223 }
1224
1225 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1226                              ldlm_iterator_t it, void *data)
1227 {
1228         struct obd_device *obd = exp->exp_obd;
1229         struct lmv_obd *lmv = &obd->u.lmv;
1230         int i, rc;
1231         ENTRY;
1232
1233         rc = lmv_check_connect(obd);
1234         if (rc)
1235                 RETURN(rc);
1236
1237         CDEBUG(D_OTHER, "CBDATA for "DFID"\n", PFID(fid));
1238
1239         /*
1240          * With CMD every object can have two locks in different namespaces:
1241          * lookup lock in space of mds storing direntry and update/open lock in
1242          * space of mds storing inode.
1243          */
1244         for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1245                 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1246
1247         RETURN(0);
1248 }
1249
1250 static int lmv_close(struct obd_export *exp,
1251                      struct md_op_data *op_data,
1252                      struct md_open_data *mod,
1253                      struct ptlrpc_request **request)
1254 {
1255         struct obd_device *obd = exp->exp_obd;
1256         struct lmv_obd *lmv = &obd->u.lmv;
1257         struct obd_export *tgt_exp;
1258         int rc;
1259         ENTRY;
1260
1261         rc = lmv_check_connect(obd);
1262         if (rc)
1263                 RETURN(rc);
1264
1265         tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1266         if (IS_ERR(tgt_exp))
1267                 RETURN(PTR_ERR(tgt_exp));
1268
1269         CDEBUG(D_OTHER, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1270         rc = md_close(tgt_exp, op_data, mod, request);
1271         RETURN(rc);
1272 }
1273
1274 /*
1275  * Called in the case MDS returns -ERESTART on create on open, what means that
1276  * directory is split and its LMV presentation object has to be updated.
1277  */
1278 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1279 {
1280         struct obd_device *obd = exp->exp_obd;
1281         struct lmv_obd *lmv = &obd->u.lmv;
1282         struct ptlrpc_request *req = NULL;
1283         struct obd_export *tgt_exp;
1284         struct lmv_obj *obj;
1285         struct lustre_md md;
1286         int mealen, rc;
1287         __u64 valid;
1288         ENTRY;
1289
1290         md.mea = NULL;
1291         mealen = lmv_get_easize(lmv);
1292
1293         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1294
1295         tgt_exp = lmv_find_export(lmv, fid);
1296         if (IS_ERR(tgt_exp))
1297                 RETURN(PTR_ERR(tgt_exp));
1298
1299         /* time to update mea of parent fid */
1300         rc = md_getattr(tgt_exp, fid, NULL, valid, mealen, &req);
1301         if (rc) {
1302                 CERROR("md_getattr() failed, error %d\n", rc);
1303                 GOTO(cleanup, rc);
1304         }
1305
1306         rc = md_get_lustre_md(tgt_exp, req, NULL, exp, &md);
1307         if (rc) {
1308                 CERROR("mdc_get_lustre_md() failed, error %d\n", rc);
1309                 GOTO(cleanup, rc);
1310         }
1311
1312         if (md.mea == NULL)
1313                 GOTO(cleanup, rc = -ENODATA);
1314
1315         obj = lmv_obj_create(exp, fid, md.mea);
1316         if (IS_ERR(obj))
1317                 rc = PTR_ERR(obj);
1318         else
1319                 lmv_obj_put(obj);
1320
1321         obd_free_memmd(exp, (struct lov_stripe_md **)&md.mea);
1322
1323         EXIT;
1324 cleanup:
1325         if (req)
1326                 ptlrpc_req_finished(req);
1327         return rc;
1328 }
1329
1330 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1331                const void *data, int datalen, int mode, __u32 uid,
1332                __u32 gid, __u32 cap_effective,  __u64 rdev,
1333                struct ptlrpc_request **request)
1334 {
1335         struct obd_device *obd = exp->exp_obd;
1336         struct lmv_obd *lmv = &obd->u.lmv;
1337         struct obd_export *tgt_exp;
1338         struct lmv_obj *obj;
1339         int rc, loop = 0;
1340         ENTRY;
1341
1342         rc = lmv_check_connect(obd);
1343         if (rc)
1344                 RETURN(rc);
1345
1346         if (!lmv->desc.ld_active_tgt_count)
1347                 RETURN(-EIO);
1348 repeat:
1349         ++loop;
1350         LASSERT(loop <= 2);
1351         obj = lmv_obj_grab(obd, &op_data->op_fid1);
1352         if (obj) {
1353                 int mea_idx;
1354
1355                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1356                                        op_data->op_name, op_data->op_namelen);
1357                 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1358                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1359                 op_data->op_mds = obj->lo_inodes[mea_idx].li_mds;
1360                 tgt_exp = lmv_get_export(lmv, op_data->op_mds);
1361                 lmv_obj_put(obj);
1362         } else {
1363                 struct lmv_tgt_desc *tgt;
1364
1365                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1366                 op_data->op_bias |= MDS_CHECK_SPLIT;
1367                 op_data->op_mds = tgt->ltd_idx;
1368                 tgt_exp = tgt->ltd_exp;
1369         }
1370
1371         if (IS_ERR(tgt_exp))
1372                 RETURN(PTR_ERR(tgt_exp));
1373
1374         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1375         if (rc == -ERESTART)
1376                 goto repeat;
1377         else if (rc)
1378                 RETURN(rc);
1379
1380         CDEBUG(D_OTHER, "CREATE '%*s' on "DFID"\n", op_data->op_namelen,
1381                op_data->op_name, PFID(&op_data->op_fid1));
1382
1383         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1384         rc = md_create(tgt_exp, op_data, data, datalen, mode, uid, gid,
1385                        cap_effective, rdev, request);
1386         if (rc == 0) {
1387                 if (*request == NULL)
1388                         RETURN(rc);
1389                 CDEBUG(D_OTHER, "created - "DFID"\n", PFID(&op_data->op_fid1));
1390         } else if (rc == -ERESTART) {
1391                 LASSERT(*request != NULL);
1392                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1393                           "Got -ERESTART during create!\n");
1394                 ptlrpc_req_finished(*request);
1395                 *request = NULL;
1396
1397                 /*
1398                  * Directory got split. Time to update local object and repeat
1399                  * the request with proper MDS.
1400                  */
1401                 rc = lmv_handle_split(exp, &op_data->op_fid1);
1402                 if (rc == 0) {
1403                         rc = lmv_alloc_slave_fids(obd, &op_data->op_fid1,
1404                                                   op_data, &op_data->op_fid2);
1405                         if (rc)
1406                                 RETURN(rc);
1407                         goto repeat;
1408                 }
1409         }
1410         RETURN(rc);
1411 }
1412
1413 static int lmv_done_writing(struct obd_export *exp,
1414                             struct md_op_data *op_data,
1415                             struct md_open_data *mod)
1416 {
1417         struct obd_device *obd = exp->exp_obd;
1418         struct lmv_obd *lmv = &obd->u.lmv;
1419         struct obd_export *tgt_exp;
1420         int rc;
1421         ENTRY;
1422
1423         rc = lmv_check_connect(obd);
1424         if (rc)
1425                 RETURN(rc);
1426
1427         tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1428         if (IS_ERR(tgt_exp))
1429                 RETURN(PTR_ERR(tgt_exp));
1430
1431         rc = md_done_writing(tgt_exp, op_data, mod);
1432         RETURN(rc);
1433 }
1434
1435 static int
1436 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1437                    struct lookup_intent *it, struct md_op_data *op_data,
1438                    struct lustre_handle *lockh, void *lmm, int lmmsize)
1439 {
1440         struct obd_device *obd = exp->exp_obd;
1441         struct lmv_obd *lmv = &obd->u.lmv;
1442         struct lmv_stripe_md *mea = op_data->op_mea1;
1443         struct md_op_data *op_data2;
1444         struct obd_export *tgt_exp;
1445         int i, rc = 0;
1446         ENTRY;
1447
1448         OBD_ALLOC_PTR(op_data2);
1449         if (op_data2 == NULL)
1450                 RETURN(-ENOMEM);
1451
1452         LASSERT(mea != NULL);
1453         for (i = 0; i < mea->mea_count; i++) {
1454                 memset(op_data2, 0, sizeof(*op_data2));
1455                 op_data2->op_fid1 = mea->mea_ids[i];
1456                 op_data2->op_bias = 0;
1457
1458                 tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1);
1459                 if (IS_ERR(tgt_exp))
1460                         GOTO(cleanup, rc = PTR_ERR(tgt_exp));
1461
1462                 if (tgt_exp == NULL)
1463                         continue;
1464
1465                 rc = md_enqueue(tgt_exp, einfo, it, op_data2,
1466                                 lockh + i, lmm, lmmsize, 0);
1467
1468                 CDEBUG(D_OTHER, "take lock on slave "DFID" -> %d/%d\n",
1469                        PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1470
1471                 if (rc)
1472                         GOTO(cleanup, rc);
1473
1474                 if (it->d.lustre.it_data) {
1475                         struct ptlrpc_request *req;
1476                         req = (struct ptlrpc_request *)it->d.lustre.it_data;
1477                         ptlrpc_req_finished(req);
1478                 }
1479
1480                 if (it->d.lustre.it_status)
1481                         GOTO(cleanup, rc = it->d.lustre.it_status);
1482         }
1483
1484         EXIT;
1485 cleanup:
1486         OBD_FREE_PTR(op_data2);
1487
1488         if (rc != 0) {
1489                 /* drop all taken locks */
1490                 while (--i >= 0) {
1491                         if (lockh[i].cookie)
1492                                 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1493                         lockh[i].cookie = 0;
1494                 }
1495         }
1496         return rc;
1497 }
1498
1499 static int
1500 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1501                    struct lookup_intent *it, struct md_op_data *op_data,
1502                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1503                    int extra_lock_flags)
1504 {
1505         struct ptlrpc_request *req = it->d.lustre.it_data;
1506         struct obd_device *obd = exp->exp_obd;
1507         struct lmv_obd *lmv = &obd->u.lmv;
1508         struct lustre_handle plock;
1509         struct obd_export *tgt_exp;
1510         struct md_op_data *rdata;
1511         struct lu_fid fid_copy;
1512         struct mdt_body *body;
1513         int rc = 0, pmode;
1514         ENTRY;
1515
1516         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1517         LASSERT(body != NULL);
1518
1519         if (!(body->valid & OBD_MD_MDS))
1520                 RETURN(0);
1521
1522         CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID" -> "DFID"\n",
1523                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1524
1525         /* We got LOOKUP lock, but we really need attrs */
1526         pmode = it->d.lustre.it_lock_mode;
1527         LASSERT(pmode != 0);
1528         memcpy(&plock, lockh, sizeof(plock));
1529         it->d.lustre.it_lock_mode = 0;
1530         it->d.lustre.it_data = NULL;
1531         fid_copy = body->fid1;
1532
1533         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1534         ptlrpc_req_finished(req);
1535
1536         tgt_exp = lmv_find_export(lmv, &fid_copy);
1537         if (IS_ERR(tgt_exp))
1538                 GOTO(out, rc = PTR_ERR(tgt_exp));
1539
1540         OBD_ALLOC_PTR(rdata);
1541         if (rdata == NULL)
1542                 GOTO(out, rc = -ENOMEM);
1543
1544         rdata->op_fid1 = fid_copy;
1545         rdata->op_bias = MDS_CROSS_REF;
1546
1547         rc = md_enqueue(tgt_exp, einfo, it, rdata, lockh,
1548                         lmm, lmmsize, extra_lock_flags);
1549         OBD_FREE_PTR(rdata);
1550         EXIT;
1551 out:
1552         ldlm_lock_decref(&plock, pmode);
1553         return rc;
1554 }
1555
1556 static int
1557 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1558             struct lookup_intent *it, struct md_op_data *op_data,
1559             struct lustre_handle *lockh, void *lmm, int lmmsize,
1560             int extra_lock_flags)
1561 {
1562         struct obd_device *obd = exp->exp_obd;
1563         struct lmv_obd *lmv = &obd->u.lmv;
1564         struct obd_export *tgt_exp = NULL;
1565         struct lmv_obj *obj;
1566         int rc;
1567         ENTRY;
1568
1569         rc = lmv_check_connect(obd);
1570         if (rc)
1571                 RETURN(rc);
1572
1573         if (op_data->op_mea1 && it->it_op == IT_UNLINK) {
1574                 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1575                                         lockh, lmm, lmmsize);
1576                 RETURN(rc);
1577         }
1578
1579         if (op_data->op_namelen) {
1580                 obj = lmv_obj_grab(obd, &op_data->op_fid1);
1581                 if (obj) {
1582                         int mea_idx;
1583
1584                         /* directory is split. look for right mds for this
1585                          * name */
1586                         mea_idx = raw_name2idx(obj->lo_hashtype,
1587                                                obj->lo_objcount,
1588                                                (char *)op_data->op_name,
1589                                                op_data->op_namelen);
1590                         op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1591                         tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds);
1592                         lmv_obj_put(obj);
1593                 }
1594         }
1595
1596         if (tgt_exp == NULL)
1597                 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1598         if (IS_ERR(tgt_exp))
1599                 RETURN(PTR_ERR(tgt_exp));
1600
1601         CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID"\n", LL_IT2STR(it),
1602                PFID(&op_data->op_fid1));
1603
1604         rc = md_enqueue(tgt_exp, einfo, it, op_data, lockh,
1605                         lmm, lmmsize, extra_lock_flags);
1606
1607         if (rc == 0 && it->it_op == IT_OPEN)
1608                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1609                                         lmm, lmmsize, extra_lock_flags);
1610         RETURN(rc);
1611 }
1612
1613 static int
1614 lmv_getattr_name(struct obd_export *exp, const struct lu_fid *fid,
1615                  struct obd_capa *oc, const char *filename, int namelen,
1616                  obd_valid valid, int ea_size, __u32 suppgid,
1617                  struct ptlrpc_request **request)
1618 {
1619         struct obd_device *obd = exp->exp_obd;
1620         struct lmv_obd *lmv = &obd->u.lmv;
1621         struct lu_fid rid = *fid;
1622         struct obd_export *tgt_exp;
1623         struct mdt_body *body;
1624         struct lmv_obj *obj;
1625         int rc, loop = 0;
1626         ENTRY;
1627
1628         rc = lmv_check_connect(obd);
1629         if (rc)
1630                 RETURN(rc);
1631
1632 repeat:
1633         ++loop;
1634         LASSERT(loop <= 2);
1635         obj = lmv_obj_grab(obd, &rid);
1636         if (obj) {
1637                 int mea_idx;
1638
1639                 /* Directory is split. Look for right mds for this name */
1640                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1641                                        filename, namelen - 1);
1642                 rid = obj->lo_inodes[mea_idx].li_fid;
1643                 tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds);
1644                 lmv_obj_put(obj);
1645                 valid &= ~OBD_MD_FLCKSPLIT;
1646         } else {
1647                 tgt_exp = lmv_find_export(lmv, &rid);
1648                 valid |= OBD_MD_FLCKSPLIT;
1649         }
1650         if (IS_ERR(tgt_exp))
1651                 RETURN(PTR_ERR(tgt_exp));
1652
1653         CDEBUG(D_OTHER, "getattr_name for %*s on "DFID" -> "DFID"\n",
1654                namelen, filename, PFID(fid), PFID(&rid));
1655
1656         rc = md_getattr_name(tgt_exp, &rid, oc, filename, namelen, valid,
1657                              ea_size, suppgid, request);
1658         if (rc == 0) {
1659                 body = req_capsule_server_get(&(*request)->rq_pill,
1660                                               &RMF_MDT_BODY);
1661                 LASSERT(body != NULL);
1662
1663                 if (body->valid & OBD_MD_MDS) {
1664                         struct ptlrpc_request *req = NULL;
1665
1666                         rid = body->fid1;
1667                         CDEBUG(D_OTHER, "request attrs for "DFID"\n",
1668                                PFID(&rid));
1669
1670                         tgt_exp = lmv_find_export(lmv, &rid);
1671                         if (IS_ERR(tgt_exp)) {
1672                                 ptlrpc_req_finished(*request);
1673                                 RETURN(PTR_ERR(tgt_exp));
1674                         }
1675
1676                         rc = md_getattr_name(tgt_exp, &rid, NULL, NULL, 1,
1677                                              valid | OBD_MD_FLCROSSREF,
1678                                              ea_size, suppgid, &req);
1679                         ptlrpc_req_finished(*request);
1680                         *request = req;
1681                 }
1682         } else if (rc == -ERESTART) {
1683                 LASSERT(*request != NULL);
1684                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1685                           "Got -ERESTART during getattr!\n");
1686                 ptlrpc_req_finished(*request);
1687                 *request = NULL;
1688
1689                 /*
1690                  * Directory got split. Time to update local object and repeat
1691                  * the request with proper MDS.
1692                  */
1693                 rc = lmv_handle_split(exp, &rid);
1694                 if (rc == 0)
1695                         goto repeat;
1696         }
1697         RETURN(rc);
1698 }
1699
1700 #define md_op_data_fid(op_data, fl)                     \
1701         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1702          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1703          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1704          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1705          NULL)
1706
1707 /* @tgt_exp is the export the metadata request is sent.
1708  * @fid_exp is the export the cancel should be sent for the current fid.
1709  * if @fid_exp is NULL, the export is found for the current fid.
1710  * @op_data keeps the current fid, which is pointed through @flag.
1711  * @mode, @bits -- lock match parameters. */
1712 static int lmv_early_cancel(struct lmv_obd *lmv, struct obd_export *tgt_exp,
1713                             struct obd_export *fid_exp,
1714                             struct md_op_data *op_data,
1715                             ldlm_mode_t mode, int bits, int flag)
1716 {
1717         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1718         ldlm_policy_data_t policy = {{0}};
1719         int rc = 0;
1720         ENTRY;
1721
1722         if (!fid_is_sane(fid))
1723                 RETURN(0);
1724
1725         if (fid_exp == NULL)
1726                 fid_exp = lmv_find_export(lmv, fid);
1727
1728         if (tgt_exp == fid_exp) {
1729                 /* The export is the same as on the target server, cancel
1730                  * will be sent along with the main metadata operation. */
1731                 op_data->op_flags |= flag;
1732                 RETURN(0);
1733         }
1734
1735         policy.l_inodebits.bits = bits;
1736         rc = md_cancel_unused(fid_exp, fid, &policy, mode, LDLM_FL_ASYNC, NULL);
1737         RETURN(rc);
1738 }
1739
1740 #ifdef EARLY_CANCEL_FOR_STRIPED_DIR_IS_READY
1741 /* Check if the fid in @op_data pointed to by flag is of the same export(s)
1742  * as @tgt_exp. Early cancels will be sent later by mdc code, otherwise, call
1743  * md_cancel_unused for child export(s). */
1744 static int lmv_early_cancel_stripes(struct obd_export *exp,
1745                                     struct obd_export *tgt_exp,
1746                                     struct md_op_data *op_data,
1747                                     ldlm_mode_t mode, int bits, int flag)
1748 {
1749         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1750         struct obd_device *obd = exp->exp_obd;
1751         struct lmv_obd *lmv = &obd->u.lmv;
1752         struct obd_export *st_exp;
1753         struct lmv_obj *obj;
1754         int rc = 0;
1755         ENTRY;
1756
1757         if (!fid_is_sane(fid))
1758                 RETURN(0);
1759
1760         obj = lmv_obj_grab(obd, fid);
1761         if (obj) {
1762                 ldlm_policy_data_t policy = {{0}};
1763                 struct lu_fid *st_fid;
1764                 int i;
1765
1766                 policy.l_inodebits.bits = bits;
1767                 for (i = 0; i < obj->lo_objcount; i++) {
1768                         st_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds);
1769                         st_fid = &obj->lo_inodes[i].li_fid;
1770                         if (tgt_exp != st_exp) {
1771                                 rc = md_cancel_unused(st_exp, st_fid, &policy,
1772                                                       mode, LDLM_FL_ASYNC,
1773                                                       NULL);
1774                                 if (rc)
1775                                         break;
1776                         } else {
1777                                 /* Some export matches to @tgt_exp, do cancel
1778                                  * for its fid in mdc */
1779                                 *fid = *st_fid;
1780                                 op_data->op_flags |= flag;
1781                         }
1782                 }
1783                 lmv_obj_put(obj);
1784         } else {
1785                 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data,
1786                                       mode, bits, flag);
1787         }
1788         RETURN(rc);
1789 }
1790 #endif
1791
1792 /*
1793  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1794  * op_data->op_fid2
1795  */
1796 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1797                     struct ptlrpc_request **request)
1798 {
1799         struct obd_device *obd = exp->exp_obd;
1800         struct lmv_obd *lmv = &obd->u.lmv;
1801         struct obd_export *tgt_exp;
1802         struct lmv_obj *obj;
1803         int rc, loop = 0;
1804         mdsno_t mds;
1805         ENTRY;
1806
1807         rc = lmv_check_connect(obd);
1808         if (rc)
1809                 RETURN(rc);
1810
1811 repeat:
1812         ++loop;
1813         LASSERT(loop <= 2);
1814         if (op_data->op_namelen != 0) {
1815                 int mea_idx;
1816
1817                 /* Usual link request */
1818                 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1819                 if (obj) {
1820                         mea_idx = raw_name2idx(obj->lo_hashtype,
1821                                                obj->lo_objcount,
1822                                                op_data->op_name,
1823                                                op_data->op_namelen);
1824                         op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1825                         mds = obj->lo_inodes[mea_idx].li_mds;
1826                         lmv_obj_put(obj);
1827                 } else {
1828                         rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1829                         if (rc)
1830                                 RETURN(rc);
1831                 }
1832
1833                 CDEBUG(D_OTHER,"link "DFID":%*s to "DFID"\n",
1834                        PFID(&op_data->op_fid2), op_data->op_namelen,
1835                        op_data->op_name, PFID(&op_data->op_fid1));
1836         } else {
1837                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds);
1838                 if (rc)
1839                         RETURN(rc);
1840
1841                 /* request from MDS to acquire i_links for inode by fid1 */
1842                 CDEBUG(D_OTHER, "inc i_nlinks for "DFID"\n",
1843                        PFID(&op_data->op_fid1));
1844         }
1845
1846         CDEBUG(D_OTHER, "forward to MDS #"LPU64" ("DFID")\n",
1847                mds, PFID(&op_data->op_fid1));
1848
1849         op_data->op_fsuid = current->fsuid;
1850         op_data->op_fsgid = current->fsgid;
1851         op_data->op_cap   = current->cap_effective;
1852
1853         tgt_exp = lmv->tgts[mds].ltd_exp;
1854         if (op_data->op_namelen) {
1855                 op_data->op_flags |= MF_MDC_CANCEL_FID2;
1856                 /* Cancel UPDATE lock on child (fid1). */
1857                 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX,
1858                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1859         }
1860         if (rc == 0)
1861                 rc = md_link(tgt_exp, op_data, request);
1862         if (rc == -ERESTART) {
1863                 LASSERT(*request != NULL);
1864                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1865                           "Got -ERESTART during link!\n");
1866                 ptlrpc_req_finished(*request);
1867                 *request = NULL;
1868
1869                 /*
1870                  * Directory got split. Time to update local object and repeat
1871                  * the request with proper MDS.
1872                  */
1873                 rc = lmv_handle_split(exp, &op_data->op_fid2);
1874                 if (rc == 0)
1875                         goto repeat;
1876         }
1877
1878         RETURN(rc);
1879 }
1880
1881 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
1882                       const char *old, int oldlen, const char *new, int newlen,
1883                       struct ptlrpc_request **request)
1884 {
1885         struct obd_export *tgt_exp = NULL, *src_exp;
1886         struct obd_device *obd = exp->exp_obd;
1887         struct lmv_obd *lmv = &obd->u.lmv;
1888         int rc, mea_idx, loop = 0;
1889         struct lmv_obj *obj;
1890         mdsno_t mds1, mds2;
1891         ENTRY;
1892
1893         CDEBUG(D_OTHER, "rename %*s in "DFID" to %*s in "DFID"\n",
1894                oldlen, old, PFID(&op_data->op_fid1),
1895                newlen, new, PFID(&op_data->op_fid2));
1896
1897         rc = lmv_check_connect(obd);
1898         if (rc)
1899                 RETURN(rc);
1900
1901         if (oldlen == 0) {
1902                 /*
1903                  * MDS with old dir entry is asking another MDS to create name
1904                  * there.
1905                  */
1906                 CDEBUG(D_OTHER,
1907                        "create %*s(%d/%d) in "DFID" pointing "
1908                        "to "DFID"\n", newlen, new, oldlen, newlen,
1909                        PFID(&op_data->op_fid2), PFID(&op_data->op_fid1));
1910
1911                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds1);
1912                 if (rc)
1913                         RETURN(rc);
1914
1915                 /*
1916                  * Target directory can be split, sowe should forward request to
1917                  * the right MDS.
1918                  */
1919                 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1920                 if (obj) {
1921                         mea_idx = raw_name2idx(obj->lo_hashtype,
1922                                                obj->lo_objcount,
1923                                                (char *)new, newlen);
1924                         op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1925                         CDEBUG(D_OTHER, "Parent obj "DFID"\n",
1926                                PFID(&op_data->op_fid2));
1927                         lmv_obj_put(obj);
1928                 }
1929                 goto request;
1930         }
1931
1932 repeat:
1933         ++loop;
1934         LASSERT(loop <= 2);
1935         obj = lmv_obj_grab(obd, &op_data->op_fid1);
1936         if (obj) {
1937                 /*
1938                  * directory is already split, so we have to forward request to
1939                  * the right MDS.
1940                  */
1941                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1942                                        (char *)old, oldlen);
1943                 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1944                 mds1 = obj->lo_inodes[mea_idx].li_mds;
1945                 CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
1946                 lmv_obj_put(obj);
1947         } else {
1948                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
1949                 if (rc)
1950                         RETURN(rc);
1951         }
1952
1953         obj = lmv_obj_grab(obd, &op_data->op_fid2);
1954         if (obj) {
1955                 /*
1956                  * Directory is already split, so we have to forward request to
1957                  * the right MDS.
1958                  */
1959                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1960                                        (char *)new, newlen);
1961
1962                 mds2 = obj->lo_inodes[mea_idx].li_mds;
1963                 op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1964                 CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
1965                 lmv_obj_put(obj);
1966         } else {
1967                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
1968                 if (rc)
1969                         RETURN(rc);
1970         }
1971
1972 request:
1973         op_data->op_fsuid = current->fsuid;
1974         op_data->op_fsgid = current->fsgid;
1975         op_data->op_cap   = current->cap_effective;
1976
1977         src_exp = lmv_get_export(lmv, mds1);
1978         tgt_exp = lmv_get_export(lmv, mds2);
1979         if (oldlen) {
1980                 /* LOOKUP lock on src child (fid3) should also be cancelled for
1981                  * src_exp in mdc_rename. */
1982                 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
1983
1984                 /* Cancel UPDATE locks on tgt parent (fid2), tgt_exp is its
1985                  * own export. */
1986                 rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data, LCK_EX,
1987                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
1988
1989                 /* Cancel LOOKUP locks on tgt child (fid4) for parent tgt_exp.*/
1990                 if (rc == 0)
1991                         rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data,
1992                                               LCK_EX, MDS_INODELOCK_LOOKUP,
1993                                               MF_MDC_CANCEL_FID4);
1994
1995                 /* XXX: the case when child is a striped dir is not supported.
1996                  * Only the master stripe has all locks cancelled early. */
1997                 /* Cancel all the locks on tgt child (fid4). */
1998                 if (rc == 0)
1999                         rc = lmv_early_cancel(lmv, src_exp, NULL, op_data,
2000                                               LCK_EX, MDS_INODELOCK_FULL,
2001                                               MF_MDC_CANCEL_FID4);
2002         }
2003
2004         if (rc == 0)
2005                 rc = md_rename(src_exp, op_data, old, oldlen,
2006                                new, newlen, request);
2007         if (rc == -ERESTART) {
2008                 LASSERT(*request != NULL);
2009                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2010                           "Got -ERESTART during rename!\n");
2011                 ptlrpc_req_finished(*request);
2012                 *request = NULL;
2013
2014                 /*
2015                  * Directory got split. Time to update local object and repeat
2016                  * the request with proper MDS.
2017                  */
2018                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2019                 if (rc == 0)
2020                         goto repeat;
2021         }
2022         RETURN(rc);
2023 }
2024
2025 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2026                        void *ea, int ealen, void *ea2, int ea2len,
2027                        struct ptlrpc_request **request,
2028                        struct md_open_data **mod)
2029 {
2030         struct obd_device *obd = exp->exp_obd;
2031         struct lmv_obd *lmv = &obd->u.lmv;
2032         struct ptlrpc_request *req;
2033         struct obd_export *tgt_exp;
2034         struct lmv_obj *obj;
2035         int rc = 0, i;
2036         ENTRY;
2037
2038         rc = lmv_check_connect(obd);
2039         if (rc)
2040                 RETURN(rc);
2041
2042         obj = lmv_obj_grab(obd, &op_data->op_fid1);
2043
2044         CDEBUG(D_OTHER, "SETATTR for "DFID", valid 0x%x%s\n",
2045                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2046                obj ? ", split" : "");
2047
2048         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2049         if (obj) {
2050                 for (i = 0; i < obj->lo_objcount; i++) {
2051                         op_data->op_fid1 = obj->lo_inodes[i].li_fid;
2052
2053                         tgt_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds);
2054                         if (IS_ERR(tgt_exp)) {
2055                                 rc = PTR_ERR(tgt_exp);
2056                                 break;
2057                         }
2058
2059                         rc = md_setattr(tgt_exp, op_data, ea, ealen,
2060                                         ea2, ea2len, &req, mod);
2061
2062                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_inodes[i].li_fid)) {
2063                                 /*
2064                                  * this is master object and this request should
2065                                  * be returned back to llite.
2066                                  */
2067                                 *request = req;
2068                         } else {
2069                                 ptlrpc_req_finished(req);
2070                         }
2071
2072                         if (rc)
2073                                 break;
2074                 }
2075                 lmv_obj_put(obj);
2076         } else {
2077                 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
2078                 if (IS_ERR(tgt_exp))
2079                         RETURN(PTR_ERR(tgt_exp));
2080
2081                 rc = md_setattr(tgt_exp, op_data, ea, ealen, ea2,
2082                                 ea2len, request, mod);
2083         }
2084         RETURN(rc);
2085 }
2086
2087 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2088                     struct obd_capa *oc, struct ptlrpc_request **request)
2089 {
2090         struct obd_device *obd = exp->exp_obd;
2091         struct lmv_obd *lmv = &obd->u.lmv;
2092         struct obd_export *tgt_exp;
2093         int rc;
2094         ENTRY;
2095
2096         rc = lmv_check_connect(obd);
2097         if (rc)
2098                 RETURN(rc);
2099
2100         tgt_exp = lmv_find_export(lmv, fid);
2101         if (IS_ERR(tgt_exp))
2102                 RETURN(PTR_ERR(tgt_exp));
2103
2104         rc = md_sync(tgt_exp, fid, oc, request);
2105         RETURN(rc);
2106 }
2107
2108 /* main purpose of LMV blocking ast is to remove split directory LMV
2109  * presentation object (struct lmv_obj) attached to the lock being revoked. */
2110 int lmv_blocking_ast(struct ldlm_lock *lock,
2111                      struct ldlm_lock_desc *desc,
2112                      void *data, int flag)
2113 {
2114         struct lustre_handle lockh;
2115         struct lmv_obj *obj;
2116         int rc;
2117         ENTRY;
2118
2119         switch (flag) {
2120         case LDLM_CB_BLOCKING:
2121                 ldlm_lock2handle(lock, &lockh);
2122                 rc = ldlm_cli_cancel(&lockh);
2123                 if (rc < 0) {
2124                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2125                         RETURN(rc);
2126                 }
2127                 break;
2128         case LDLM_CB_CANCELING:
2129                 /* time to drop cached attrs for dirobj */
2130                 obj = lock->l_ast_data;
2131                 if (obj) {
2132                         CDEBUG(D_OTHER, "cancel %s on "LPU64"/"LPU64
2133                                ", master "DFID"\n",
2134                                lock->l_resource->lr_name.name[3] == 1 ?
2135                                "LOOKUP" : "UPDATE",
2136                                lock->l_resource->lr_name.name[0],
2137                                lock->l_resource->lr_name.name[1],
2138                                PFID(&obj->lo_fid));
2139                         lmv_obj_put(obj);
2140                 }
2141                 break;
2142         default:
2143                 LBUG();
2144         }
2145         RETURN(0);
2146 }
2147
2148 static void lmv_hash_adjust(__u32 *hash, __u32 hash_adj)
2149 {
2150         __u32 val;
2151
2152         val = le32_to_cpu(*hash);
2153         if (val < hash_adj)
2154                 val += MAX_HASH_SIZE;
2155         if (val != DIR_END_OFF)
2156                 *hash = cpu_to_le32(val - hash_adj);
2157 }
2158
2159 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2160 {
2161         __u64 id;
2162         struct obd_import *imp;
2163
2164         /*
2165          * XXX Hack: to get nid we assume that underlying obd device is mdc.
2166          */
2167         imp  = class_exp2cliimp(exp);
2168         id   = imp->imp_connection->c_self + fid_flatten(fid);
2169
2170         CDEBUG(D_INFO, "node rank: %llx "DFID" %llx %llx\n",
2171                imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2172
2173         return id ^ (id >> 32);
2174 }
2175
2176 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2177                         struct obd_capa *oc, __u64 offset64, struct page *page,
2178                         struct ptlrpc_request **request)
2179 {
2180         struct obd_device *obd = exp->exp_obd;
2181         struct lmv_obd *lmv = &obd->u.lmv;
2182         struct obd_export *tgt_exp;
2183         struct lu_fid rid = *fid;
2184         struct lmv_obj *obj;
2185         __u32 offset0;
2186         __u32 offset;
2187         __u32 hash_adj = 0;
2188         __u32 rank = 0;
2189         __u32 seg_size = 0;
2190         int tgt = 0;
2191         int tgt0 = 0;
2192         int rc;
2193         int nr = 0;
2194         ENTRY;
2195
2196         offset0 = offset = offset64;
2197         /*
2198          * Check that offset is representable by 32bit number.
2199          */
2200         LASSERT((__u64)offset == offset64);
2201
2202         rc = lmv_check_connect(obd);
2203         if (rc)
2204                 RETURN(rc);
2205
2206         CDEBUG(D_INFO, "READPAGE at %x from "DFID"\n", offset, PFID(&rid));
2207
2208         obj = lmv_obj_grab(obd, fid);
2209         if (obj) {
2210                 struct lmv_inode *loi;
2211
2212                 lmv_obj_lock(obj);
2213
2214                 nr       = obj->lo_objcount;
2215                 LASSERT(nr > 0);
2216                 seg_size = MAX_HASH_SIZE / nr;
2217                 loi      = obj->lo_inodes;
2218                 rank     = lmv_node_rank(lmv_get_export(lmv, loi[0].li_mds),
2219                                          fid) % nr;
2220                 tgt0     = (offset / seg_size) % nr;
2221                 tgt      = (tgt0 + rank) % nr;
2222
2223                 if (tgt < tgt0)
2224                         /*
2225                          * Wrap around.
2226                          *
2227                          * Last segment has unusual length due to division
2228                          * rounding.
2229                          */
2230                         hash_adj = MAX_HASH_SIZE - seg_size * nr;
2231                 else
2232                         hash_adj = 0;
2233
2234                 hash_adj += rank * seg_size;
2235
2236                 CDEBUG(D_INFO, "hash_adj: %x %x %x/%x -> %x/%x\n",
2237                        rank, hash_adj, offset, tgt0, offset + hash_adj, tgt);
2238
2239                 offset = (offset + hash_adj) % MAX_HASH_SIZE;
2240                 rid = obj->lo_inodes[tgt].li_fid;
2241                 tgt_exp = lmv_get_export(lmv, loi[tgt].li_mds);
2242
2243                 CDEBUG(D_INFO, "forward to "DFID" with offset %lu i %d\n",
2244                        PFID(&rid), (unsigned long)offset, tgt);
2245         } else
2246                 tgt_exp = lmv_find_export(lmv, &rid);
2247
2248         if (IS_ERR(tgt_exp))
2249                 GOTO(cleanup, rc = PTR_ERR(tgt_exp));
2250
2251         rc = md_readpage(tgt_exp, &rid, oc, offset, page, request);
2252         if (rc)
2253                 GOTO(cleanup, rc);
2254         if (obj) {
2255                 struct lu_dirpage *dp;
2256                 struct lu_dirent  *ent;
2257
2258                 dp = cfs_kmap(page);
2259
2260                 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2261                 lmv_hash_adjust(&dp->ldp_hash_end,   hash_adj);
2262                 LASSERT(cpu_to_le32(dp->ldp_hash_start) <= offset0);
2263
2264                 for (ent = lu_dirent_start(dp); ent != NULL;
2265                      ent = lu_dirent_next(ent))
2266                         lmv_hash_adjust(&ent->lde_hash, hash_adj);
2267
2268                 if (tgt0 != nr - 1) {
2269                         __u32 end;
2270
2271                         end = le32_to_cpu(dp->ldp_hash_end);
2272                         if (end == DIR_END_OFF) {
2273                                 dp->ldp_hash_end = cpu_to_le32(seg_size *
2274                                                                (tgt0 + 1));
2275                                 CDEBUG(D_INFO, ""DFID" reset end %x tgt %d\n",
2276                                        PFID(&rid),
2277                                        le32_to_cpu(dp->ldp_hash_end), tgt);
2278                         }
2279                 }
2280                 cfs_kunmap(page);
2281         }
2282         /*
2283          * Here we could remove "." and ".." from all pages which at not from
2284          * master. But MDS has only "." and ".." for master dir.
2285          */
2286         EXIT;
2287 cleanup:
2288         if (obj) {
2289                 lmv_obj_unlock(obj);
2290                 lmv_obj_put(obj);
2291         }
2292         return rc;
2293 }
2294
2295 static int lmv_unlink_slaves(struct obd_export *exp,
2296                              struct md_op_data *op_data,
2297                              struct ptlrpc_request **req)
2298 {
2299         struct obd_device *obd = exp->exp_obd;
2300         struct lmv_obd *lmv = &obd->u.lmv;
2301         struct lmv_stripe_md *mea = op_data->op_mea1;
2302         struct md_op_data *op_data2;
2303         struct obd_export *tgt_exp;
2304         int i, rc = 0;
2305         ENTRY;
2306
2307         OBD_ALLOC_PTR(op_data2);
2308         if (op_data2 == NULL)
2309                 RETURN(-ENOMEM);
2310
2311         op_data2->op_mode = S_IFDIR;
2312         op_data2->op_fsuid = current->fsuid;
2313         op_data2->op_fsgid = current->fsgid;
2314         op_data2->op_bias = 0;
2315
2316         LASSERT(mea != NULL);
2317         for (i = 0; i < mea->mea_count; i++) {
2318                 memset(op_data2, 0, sizeof(*op_data2));
2319                 op_data2->op_fid1 = mea->mea_ids[i];
2320                 tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1);
2321                 if (IS_ERR(tgt_exp))
2322                         GOTO(out_free_op_data2, rc = PTR_ERR(tgt_exp));
2323
2324                 if (tgt_exp == NULL)
2325                         continue;
2326
2327                 rc = md_unlink(tgt_exp, op_data2, req);
2328
2329                 CDEBUG(D_OTHER, "unlink slave "DFID" -> %d\n",
2330                        PFID(&mea->mea_ids[i]), rc);
2331
2332                 if (*req) {
2333                         ptlrpc_req_finished(*req);
2334                         *req = NULL;
2335                 }
2336                 if (rc)
2337                         GOTO(out_free_op_data2, rc);
2338         }
2339
2340         EXIT;
2341 out_free_op_data2:
2342         OBD_FREE_PTR(op_data2);
2343         return rc;
2344 }
2345
2346 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2347                       struct ptlrpc_request **request)
2348 {
2349         struct obd_device *obd = exp->exp_obd;
2350         struct lmv_obd *lmv = &obd->u.lmv;
2351         struct obd_export *tgt_exp = NULL;
2352         struct lmv_obj *obj;
2353         int rc, loop = 0;
2354         ENTRY;
2355
2356         rc = lmv_check_connect(obd);
2357         if (rc)
2358                 RETURN(rc);
2359
2360         if (op_data->op_namelen == 0 && op_data->op_mea1 != NULL) {
2361                 /* mds asks to remove slave objects */
2362                 rc = lmv_unlink_slaves(exp, op_data, request);
2363                 RETURN(rc);
2364         }
2365
2366 repeat:
2367         ++loop;
2368         LASSERT(loop <= 2);
2369         if (op_data->op_namelen != 0) {
2370                 int mea_idx;
2371
2372                 obj = lmv_obj_grab(obd, &op_data->op_fid1);
2373                 if (obj) {
2374                         mea_idx = raw_name2idx(obj->lo_hashtype,
2375                                                obj->lo_objcount,
2376                                                op_data->op_name,
2377                                                op_data->op_namelen);
2378                         op_data->op_bias &= ~MDS_CHECK_SPLIT;
2379                         op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
2380                         tgt_exp = lmv_get_export(lmv,
2381                                                  obj->lo_inodes[mea_idx].li_mds);
2382                         lmv_obj_put(obj);
2383                         CDEBUG(D_OTHER, "unlink '%*s' in "DFID" -> %u\n",
2384                                op_data->op_namelen, op_data->op_name,
2385                                PFID(&op_data->op_fid1), mea_idx);
2386                 }
2387         } else {
2388                 CDEBUG(D_OTHER, "drop i_nlink on "DFID"\n",
2389                        PFID(&op_data->op_fid1));
2390         }
2391         if (tgt_exp == NULL) {
2392                 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
2393                 if (IS_ERR(tgt_exp))
2394                         RETURN(PTR_ERR(tgt_exp));
2395                 op_data->op_bias |= MDS_CHECK_SPLIT;
2396         }
2397
2398         op_data->op_fsuid = current->fsuid;
2399         op_data->op_fsgid = current->fsgid;
2400         op_data->op_cap   = current->cap_effective;
2401
2402         /* If child's fid is given, cancel unused locks for it if it is from
2403          * another export than parent. */
2404         if (op_data->op_namelen) {
2405                 /* LOOKUP lock for child (fid3) should also be cancelled on
2406                  * parent tgt_exp in mdc_unlink(). */
2407                 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2408
2409                 /* XXX: the case when child is a striped dir is not supported.
2410                  * Only the master stripe has all locks cancelled early. */
2411                 /* Cancel FULL locks on child (fid3). */
2412                 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX,
2413                                       MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2414         }
2415         if (rc == 0)
2416                 rc = md_unlink(tgt_exp, op_data, request);
2417         if (rc == -ERESTART) {
2418                 LASSERT(*request != NULL);
2419                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2420                           "Got -ERESTART during unlink!\n");
2421                 ptlrpc_req_finished(*request);
2422                 *request = NULL;
2423
2424                 /*
2425                  * Directory got split. Time to update local object and repeat
2426                  * the request with proper MDS.
2427                  */
2428                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2429                 if (rc == 0)
2430                         goto repeat;
2431         }
2432         RETURN(rc);
2433 }
2434
2435 static int lmv_llog_init(struct obd_device *obd, struct obd_llogs* llogs,
2436                          struct obd_device *tgt, int count,
2437                          struct llog_catid *logid, struct obd_uuid *uuid)
2438 {
2439         struct llog_ctxt *ctxt;
2440         int rc;
2441         ENTRY;
2442
2443         rc = llog_setup(obd, llogs, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
2444                         &llog_client_ops);
2445         if (rc == 0) {
2446                 ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
2447                 ctxt->loc_imp = tgt->u.cli.cl_import;
2448         }
2449
2450         RETURN(rc);
2451 }
2452
2453 static int lmv_llog_finish(struct obd_device *obd, int count)
2454 {
2455         int rc;
2456         ENTRY;
2457
2458         rc = llog_cleanup(llog_get_context(obd, LLOG_CONFIG_REPL_CTXT));
2459         RETURN(rc);
2460 }
2461
2462 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2463 {
2464         int rc = 0;
2465
2466         switch (stage) {
2467         case OBD_CLEANUP_EARLY:
2468                 /* XXX: here should be calling obd_precleanup() down to
2469                  * stack. */
2470                 break;
2471         case OBD_CLEANUP_SELF_EXP:
2472                 rc = obd_llog_finish(obd, 0);
2473                 if (rc != 0)
2474                         CERROR("failed to cleanup llogging subsystems\n");
2475                 break;
2476         default:
2477                 break;
2478         }
2479         RETURN(rc);
2480 }
2481
2482 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2483                         void *key, __u32 *vallen, void *val)
2484 {
2485         struct obd_device *obd;
2486         struct lmv_obd *lmv;
2487         int rc = 0;
2488         ENTRY;
2489
2490         obd = class_exp2obd(exp);
2491         if (obd == NULL) {
2492                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2493                        exp->exp_handle.h_cookie);
2494                 RETURN(-EINVAL);
2495         }
2496
2497         lmv = &obd->u.lmv;
2498         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2499                 struct lmv_tgt_desc *tgts;
2500                 int i;
2501
2502                 rc = lmv_check_connect(obd);
2503                 if (rc)
2504                         RETURN(rc);
2505
2506                 LASSERT(*vallen == sizeof(__u32));
2507                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2508                      i++, tgts++) {
2509
2510                         /* all tgts should be connected when this get called. */
2511                         if (!tgts || !tgts->ltd_exp) {
2512                                 CERROR("target not setup?\n");
2513                                 continue;
2514                         }
2515
2516                         if (!obd_get_info(tgts->ltd_exp, keylen, key,
2517                                           vallen, val))
2518                                 RETURN(0);
2519                 }
2520                 RETURN(-EINVAL);
2521         } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2522                 rc = lmv_check_connect(obd);
2523                 if (rc)
2524                         RETURN(rc);
2525
2526                 /* forwarding this request to first MDS, it should know LOV
2527                  * desc. */
2528                 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2529                                   vallen, val);
2530                 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2531                         exp->exp_connect_flags =
2532                         ((struct obd_connect_data *)val)->ocd_connect_flags;
2533                 }
2534                 RETURN(rc);
2535         }
2536
2537         CDEBUG(D_IOCTL, "invalid key\n");
2538         RETURN(-EINVAL);
2539 }
2540
2541 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2542                        void *key, obd_count vallen, void *val,
2543                        struct ptlrpc_request_set *set)
2544 {
2545         struct lmv_tgt_desc    *tgt;
2546         struct obd_device      *obd;
2547         struct lmv_obd         *lmv;
2548         int rc = 0;
2549         ENTRY;
2550
2551         obd = class_exp2obd(exp);
2552         if (obd == NULL) {
2553                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2554                        exp->exp_handle.h_cookie);
2555                 RETURN(-EINVAL);
2556         }
2557         lmv = &obd->u.lmv;
2558
2559         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2560             KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2561                 int i, err = 0;
2562
2563                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2564                         tgt = &lmv->tgts[i];
2565
2566                         if (!tgt->ltd_exp)
2567                                 continue;
2568
2569                         err = obd_set_info_async(tgt->ltd_exp,
2570                                                  keylen, key, vallen, val, set);
2571                         if (err && rc == 0)
2572                                 rc = err;
2573                 }
2574
2575                 RETURN(rc);
2576         }
2577
2578         RETURN(-EINVAL);
2579 }
2580
2581 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2582                struct lov_stripe_md *lsm)
2583 {
2584         struct obd_device *obd = class_exp2obd(exp);
2585         struct lmv_obd *lmv = &obd->u.lmv;
2586         struct lmv_stripe_md *meap, *lsmp;
2587         int mea_size, i;
2588         ENTRY;
2589
2590         mea_size = lmv_get_easize(lmv);
2591         if (!lmmp)
2592                 RETURN(mea_size);
2593
2594         if (*lmmp && !lsm) {
2595                 OBD_FREE(*lmmp, mea_size);
2596                 *lmmp = NULL;
2597                 RETURN(0);
2598         }
2599
2600         if (*lmmp == NULL) {
2601                 OBD_ALLOC(*lmmp, mea_size);
2602                 if (*lmmp == NULL)
2603                         RETURN(-ENOMEM);
2604         }
2605
2606         if (!lsm)
2607                 RETURN(mea_size);
2608
2609         lsmp = (struct lmv_stripe_md *)lsm;
2610         meap = (struct lmv_stripe_md *)*lmmp;
2611
2612         if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2613             lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2614                 RETURN(-EINVAL);
2615
2616         meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2617         meap->mea_count = cpu_to_le32(lsmp->mea_count);
2618         meap->mea_master = cpu_to_le32(lsmp->mea_master);
2619
2620         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2621                 meap->mea_ids[i] = meap->mea_ids[i];
2622                 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2623         }
2624
2625         RETURN(mea_size);
2626 }
2627
2628 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2629                  struct lov_mds_md *lmm, int lmm_size)
2630 {
2631         struct obd_device *obd = class_exp2obd(exp);
2632         struct lmv_stripe_md **tmea = (struct lmv_stripe_md **)lsmp;
2633         struct lmv_stripe_md *mea = (struct lmv_stripe_md *)lmm;
2634         struct lmv_obd *lmv = &obd->u.lmv;
2635         int mea_size, i;
2636         __u32 magic;
2637         ENTRY;
2638
2639         mea_size = lmv_get_easize(lmv);
2640         if (lsmp == NULL)
2641                 return mea_size;
2642
2643         if (*lsmp != NULL && lmm == NULL) {
2644                 OBD_FREE(*tmea, mea_size);
2645                 *lsmp = NULL;
2646                 RETURN(0);
2647         }
2648
2649         LASSERT(mea_size == lmm_size);
2650
2651         OBD_ALLOC(*tmea, mea_size);
2652         if (*tmea == NULL)
2653                 RETURN(-ENOMEM);
2654
2655         if (!lmm)
2656                 RETURN(mea_size);
2657
2658         if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2659             mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2660             mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2661         {
2662                 magic = le32_to_cpu(mea->mea_magic);
2663         } else {
2664                 /* old mea is not handled here */
2665                 LBUG();
2666         }
2667
2668         (*tmea)->mea_magic = magic;
2669         (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2670         (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2671
2672         for (i = 0; i < (*tmea)->mea_count; i++) {
2673                 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2674                 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2675         }
2676         RETURN(mea_size);
2677 }
2678
2679 static int lmv_cancel_unused(struct obd_export *exp,
2680                              const struct lu_fid *fid,
2681                              ldlm_policy_data_t *policy,
2682                              ldlm_mode_t mode, int flags, void *opaque)
2683 {
2684         struct obd_device *obd = exp->exp_obd;
2685         struct lmv_obd *lmv = &obd->u.lmv;
2686         int rc = 0, err, i;
2687         ENTRY;
2688
2689         LASSERT(fid != NULL);
2690
2691         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2692                 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2693                         continue;
2694
2695                 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2696                                        policy, mode, flags, opaque);
2697                 if (!rc)
2698                         rc = err;
2699         }
2700         RETURN(rc);
2701 }
2702
2703 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data)
2704 {
2705         struct obd_device *obd = exp->exp_obd;
2706         struct lmv_obd *lmv = &obd->u.lmv;
2707
2708         ENTRY;
2709         RETURN(md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data));
2710 }
2711
2712 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2713                            const struct lu_fid *fid, ldlm_type_t type,
2714                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
2715                            struct lustre_handle *lockh)
2716 {
2717         struct obd_device *obd = exp->exp_obd;
2718         struct lmv_obd *lmv = &obd->u.lmv;
2719         ldlm_mode_t rc;
2720         int i;
2721         ENTRY;
2722
2723         CDEBUG(D_OTHER, "lock match for "DFID"\n", PFID(fid));
2724
2725         /* with CMD every object can have two locks in different namespaces:
2726          * lookup lock in space of mds storing direntry and update/open lock in
2727          * space of mds storing inode. Thus we check all targets, not only that
2728          * one fid was created in. */
2729         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2730                 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2731                                    type, policy, mode, lockh);
2732                 if (rc)
2733                         RETURN(rc);
2734         }
2735
2736         RETURN(0);
2737 }
2738
2739 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2740                       struct obd_export *dt_exp, struct obd_export *md_exp,
2741                       struct lustre_md *md)
2742 {
2743         struct obd_device *obd = exp->exp_obd;
2744         struct lmv_obd *lmv = &obd->u.lmv;
2745         int rc;
2746
2747         ENTRY;
2748         rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2749         RETURN(rc);
2750 }
2751
2752 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2753 {
2754         struct obd_device *obd = exp->exp_obd;
2755         struct lmv_obd *lmv = &obd->u.lmv;
2756
2757         ENTRY;
2758         if (md->mea)
2759                 obd_free_memmd(exp, (struct lov_stripe_md**)&md->mea);
2760         RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2761 }
2762
2763 int lmv_set_open_replay_data(struct obd_export *exp,
2764                              struct obd_client_handle *och,
2765                              struct ptlrpc_request *open_req)
2766 {
2767         struct obd_device *obd = exp->exp_obd;
2768         struct lmv_obd *lmv = &obd->u.lmv;
2769         struct obd_export *tgt_exp;
2770
2771         ENTRY;
2772
2773         tgt_exp = lmv_find_export(lmv, &och->och_fid);
2774         if (IS_ERR(tgt_exp))
2775                 RETURN(PTR_ERR(tgt_exp));
2776
2777         RETURN(md_set_open_replay_data(tgt_exp, och, open_req));
2778 }
2779
2780 int lmv_clear_open_replay_data(struct obd_export *exp,
2781                                struct obd_client_handle *och)
2782 {
2783         struct obd_device *obd = exp->exp_obd;
2784         struct lmv_obd *lmv = &obd->u.lmv;
2785         struct obd_export *tgt_exp;
2786         ENTRY;
2787
2788         tgt_exp = lmv_find_export(lmv, &och->och_fid);
2789         if (IS_ERR(tgt_exp))
2790                 RETURN(PTR_ERR(tgt_exp));
2791
2792         RETURN(md_clear_open_replay_data(tgt_exp, och));
2793 }
2794
2795 static int lmv_get_remote_perm(struct obd_export *exp,
2796                                const struct lu_fid *fid,
2797                                struct obd_capa *oc, __u32 suppgid,
2798                                struct ptlrpc_request **request)
2799 {
2800         struct obd_device *obd = exp->exp_obd;
2801         struct lmv_obd *lmv = &obd->u.lmv;
2802         struct obd_export *tgt_exp;
2803         int rc;
2804
2805         ENTRY;
2806
2807         rc = lmv_check_connect(obd);
2808         if (rc)
2809                 RETURN(rc);
2810
2811         tgt_exp = lmv_find_export(lmv, fid);
2812         if (IS_ERR(tgt_exp))
2813                 RETURN(PTR_ERR(tgt_exp));
2814
2815         rc = md_get_remote_perm(tgt_exp, fid, oc, suppgid, request);
2816
2817         RETURN(rc);
2818 }
2819
2820 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2821                           renew_capa_cb_t cb)
2822 {
2823         struct obd_device *obd = exp->exp_obd;
2824         struct lmv_obd *lmv = &obd->u.lmv;
2825         struct obd_export *tgt_exp;
2826         int rc;
2827         ENTRY;
2828
2829         rc = lmv_check_connect(obd);
2830         if (rc)
2831                 RETURN(rc);
2832
2833         tgt_exp = lmv_find_export(lmv, &oc->c_capa.lc_fid);
2834         if (IS_ERR(tgt_exp))
2835                 RETURN(PTR_ERR(tgt_exp));
2836
2837         rc = md_renew_capa(tgt_exp, oc, cb);
2838         RETURN(rc);
2839 }
2840
2841 struct obd_ops lmv_obd_ops = {
2842         .o_owner                = THIS_MODULE,
2843         .o_setup                = lmv_setup,
2844         .o_cleanup              = lmv_cleanup,
2845         .o_precleanup           = lmv_precleanup,
2846         .o_process_config       = lmv_process_config,
2847         .o_connect              = lmv_connect,
2848         .o_disconnect           = lmv_disconnect,
2849         .o_statfs               = lmv_statfs,
2850         .o_llog_init            = lmv_llog_init,
2851         .o_llog_finish          = lmv_llog_finish,
2852         .o_get_info             = lmv_get_info,
2853         .o_set_info_async       = lmv_set_info_async,
2854         .o_packmd               = lmv_packmd,
2855         .o_unpackmd             = lmv_unpackmd,
2856         .o_notify               = lmv_notify,
2857         .o_iocontrol            = lmv_iocontrol,
2858         .o_fid_delete           = lmv_fid_delete
2859 };
2860
2861 struct md_ops lmv_md_ops = {
2862         .m_getstatus            = lmv_getstatus,
2863         .m_change_cbdata        = lmv_change_cbdata,
2864         .m_close                = lmv_close,
2865         .m_create               = lmv_create,
2866         .m_done_writing         = lmv_done_writing,
2867         .m_enqueue              = lmv_enqueue,
2868         .m_getattr              = lmv_getattr,
2869         .m_getxattr             = lmv_getxattr,
2870         .m_getattr_name         = lmv_getattr_name,
2871         .m_intent_lock          = lmv_intent_lock,
2872         .m_link                 = lmv_link,
2873         .m_rename               = lmv_rename,
2874         .m_setattr              = lmv_setattr,
2875         .m_setxattr             = lmv_setxattr,
2876         .m_sync                 = lmv_sync,
2877         .m_readpage             = lmv_readpage,
2878         .m_unlink               = lmv_unlink,
2879         .m_init_ea_size         = lmv_init_ea_size,
2880         .m_cancel_unused        = lmv_cancel_unused,
2881         .m_set_lock_data        = lmv_set_lock_data,
2882         .m_lock_match           = lmv_lock_match,
2883         .m_get_lustre_md        = lmv_get_lustre_md,
2884         .m_free_lustre_md       = lmv_free_lustre_md,
2885         .m_set_open_replay_data = lmv_set_open_replay_data,
2886         .m_clear_open_replay_data = lmv_clear_open_replay_data,
2887         .m_get_remote_perm      = lmv_get_remote_perm,
2888         .m_renew_capa           = lmv_renew_capa
2889 };
2890
2891 int __init lmv_init(void)
2892 {
2893         struct lprocfs_static_vars lvars;
2894         int rc;
2895
2896         obj_cache = cfs_mem_cache_create("lmv_objects",
2897                                       sizeof(struct lmv_obj),
2898                                       0, 0);
2899         if (!obj_cache) {
2900                 CERROR("error allocating lmv objects cache\n");
2901                 return -ENOMEM;
2902         }
2903
2904         lprocfs_lmv_init_vars(&lvars);
2905         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
2906                                  lvars.module_vars, LUSTRE_LMV_NAME, NULL);
2907         if (rc)
2908                 cfs_mem_cache_destroy(obj_cache);
2909
2910         return rc;
2911 }
2912
2913 #ifdef __KERNEL__
2914 static void lmv_exit(void)
2915 {
2916         int rc;
2917
2918         class_unregister_type(LUSTRE_LMV_NAME);
2919
2920         rc = cfs_mem_cache_destroy(obj_cache);
2921         LASSERTF(rc == 0,
2922                  "can't free lmv objects cache, %d object(s)"
2923                  "still in use\n", atomic_read(&obj_cache_count));
2924 }
2925
2926 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2927 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
2928 MODULE_LICENSE("GPL");
2929
2930 module_init(lmv_init);
2931 module_exit(lmv_exit);
2932 #endif