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                        void *localdata)
 224 {
 225 #ifdef __KERNEL__
 226         struct proc_dir_entry *lmv_proc_dir;
 227 #endif
 228         struct lmv_obd *lmv = &obd->u.lmv;
 229         struct obd_export *exp;
 230         int rc = 0;
 231         ENTRY;
 232
 233         rc = class_connect(conn, obd, cluuid);
 234         if (rc) {
 235                 CERROR("class_connection() returned %d\n", rc);
 236                 RETURN(rc);
 237         }
 238
 239         exp = class_conn2export(conn);
 240
 241         /* we don't want to actually do the underlying connections more than
 242          * once, so keep track. */
 243         lmv->refcount++;
 244         if (lmv->refcount > 1) {
 245                 class_export_put(exp);
 246                 RETURN(0);
 247         }
 248
 249         lmv->exp = exp;
 250         lmv->connected = 0;
 251         lmv->cluuid = *cluuid;
 252
 253         if (data)
 254                 lmv->conn_data = *data;
 255
 256 #ifdef __KERNEL__
 257         lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
 258                                         NULL, NULL);
 259         if (IS_ERR(lmv_proc_dir)) {
 260                 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
 261                        obd->obd_type->typ_name, obd->obd_name);
 262                 lmv_proc_dir = NULL;
 263         }
 264 #endif
 265
 266         /* all real clients should perform actual connection right away, because
 267          * it is possible, that LMV will not have opportunity to connect targets
 268          * and MDC stuff will be called directly, for instance while reading
 269          * ../mdc/../kbytesfree procfs file, etc. */
 270         if (data->ocd_connect_flags & OBD_CONNECT_REAL)
 271                 rc = lmv_check_connect(obd);
 272
 273 #ifdef __KERNEL__
 274         if (rc) {
 275                 if (lmv_proc_dir)
 276                         lprocfs_remove(&lmv_proc_dir);
 277         }
 278 #endif
 279
 280         RETURN(rc);
 281 }
 282
 283 static void lmv_set_timeouts(struct obd_device *obd)
 284 {
 285         struct lmv_tgt_desc *tgts;
 286         struct lmv_obd *lmv;
 287         int i;
 288
 289         lmv = &obd->u.lmv;
 290         if (lmv->server_timeout == 0)
 291                 return;
 292
 293         if (lmv->connected == 0)
 294                 return;
 295
 296         for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
 297                 if (tgts->ltd_exp == NULL)
 298                         continue;
 299
 300                 obd_set_info_async(tgts->ltd_exp, sizeof(KEY_INTERMDS),
 301                                    KEY_INTERMDS, 0, NULL, NULL);
 302         }
 303 }
 304
 305 static int lmv_init_ea_size(struct obd_export *exp, int easize,
 306                             int def_easize, int cookiesize)
 307 {
 308         struct obd_device *obd = exp->exp_obd;
 309         struct lmv_obd *lmv = &obd->u.lmv;
 310         int i, rc = 0, change = 0;
 311         ENTRY;
 312
 313         if (lmv->max_easize < easize) {
 314                 lmv->max_easize = easize;
 315                 change = 1;
 316         }
 317         if (lmv->max_def_easize < def_easize) {
 318                 lmv->max_def_easize = def_easize;
 319                 change = 1;
 320         }
 321         if (lmv->max_cookiesize < cookiesize) {
 322                 lmv->max_cookiesize = cookiesize;
 323                 change = 1;
 324         }
 325         if (change == 0)
 326                 RETURN(0);
 327
 328         if (lmv->connected == 0)
 329                 RETURN(0);
 330
 331         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 332                 if (lmv->tgts[i].ltd_exp == NULL) {
 333                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
 334                         continue;
 335                 }
 336
 337                 rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize,
 338                                      cookiesize);
 339                 if (rc) {
 340                         CERROR("obd_init_ea_size() failed on MDT target %d, "
 341                                "error %d.\n", i, rc);
 342                         break;
 343                 }
 344         }
 345         RETURN(rc);
 346 }
 347
 348 #define MAX_STRING_SIZE 128
 349
 350 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 351 {
 352         struct lmv_obd *lmv = &obd->u.lmv;
 353         struct obd_uuid *cluuid = &lmv->cluuid;
 354         struct obd_connect_data *mdc_data = NULL;
 355         struct obd_uuid lmv_mdc_uuid = { "LMV_MDC_UUID" };
 356         struct lustre_handle conn = {0, };
 357         struct obd_device *mdc_obd;
 358         struct obd_export *mdc_exp;
 359         struct lu_fld_target target;
 360         int rc;
 361 #ifdef __KERNEL__
 362         struct proc_dir_entry *lmv_proc_dir;
 363 #endif
 364         ENTRY;
 365
 366         /* for MDS: don't connect to yourself */
 367         if (obd_uuid_equals(&tgt->ltd_uuid, cluuid)) {
 368                 CDEBUG(D_CONFIG, "don't connect back to %s\n", cluuid->uuid);
 369                 /* XXX - the old code didn't increment active tgt count.
 370                  *       should we ? */
 371                 RETURN(0);
 372         }
 373
 374         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 375                                         &obd->obd_uuid);
 376         if (!mdc_obd) {
 377                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 378                 RETURN(-EINVAL);
 379         }
 380
 381         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
 382                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 383                 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
 384                 cluuid->uuid);
 385
 386         if (!mdc_obd->obd_set_up) {
 387                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 388                 RETURN(-EINVAL);
 389         }
 390
 391         rc = obd_connect(NULL, &conn, mdc_obd, &lmv_mdc_uuid,
 392                          &lmv->conn_data, NULL);
 393         if (rc) {
 394                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 395                 RETURN(rc);
 396         }
 397
 398         mdc_exp = class_conn2export(&conn);
 399
 400         /* Init fid sequence client for this mdc. */
 401         rc = obd_fid_init(mdc_exp);
 402         if (rc)
 403                 RETURN(rc);
 404
 405         /* Add new FLD target. */
 406         target.ft_srv = NULL;
 407         target.ft_exp = mdc_exp;
 408         target.ft_idx = tgt->ltd_idx;
 409
 410         fld_client_add_target(&lmv->lmv_fld, &target);
 411
 412         mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data;
 413
 414         rc = obd_register_observer(mdc_obd, obd);
 415         if (rc) {
 416                 obd_disconnect(mdc_exp);
 417                 CERROR("target %s register_observer error %d\n",
 418                        tgt->ltd_uuid.uuid, rc);
 419                 RETURN(rc);
 420         }
 421
 422         if (obd->obd_observer) {
 423                 /* tell the mds_lmv about the new target */
 424                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 425                                 OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts));
 426                 if (rc) {
 427                         obd_disconnect(mdc_exp);
 428                         RETURN(rc);
 429                 }
 430         }
 431
 432         tgt->ltd_active = 1;
 433         tgt->ltd_exp = mdc_exp;
 434         lmv->desc.ld_active_tgt_count++;
 435
 436         /* copy connect data, it may be used later */
 437         lmv->datas[tgt->ltd_idx] = *mdc_data;
 438
 439         md_init_ea_size(tgt->ltd_exp, lmv->max_easize,
 440                         lmv->max_def_easize, lmv->max_cookiesize);
 441
 442         CDEBUG(D_CONFIG, "connected to %s(%s) successfully (%d)\n",
 443                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 444                 atomic_read(&obd->obd_refcount));
 445
 446 #ifdef __KERNEL__
 447         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 448         if (lmv_proc_dir) {
 449                 struct proc_dir_entry *mdc_symlink;
 450                 char name[MAX_STRING_SIZE + 1];
 451
 452                 LASSERT(mdc_obd->obd_type != NULL);
 453                 LASSERT(mdc_obd->obd_type->typ_name != NULL);
 454                 name[MAX_STRING_SIZE] = '\0';
 455                 snprintf(name, MAX_STRING_SIZE, "../../../%s/%s",
 456                          mdc_obd->obd_type->typ_name,
 457                          mdc_obd->obd_name);
 458                 mdc_symlink = proc_symlink(mdc_obd->obd_name,
 459                                            lmv_proc_dir, name);
 460                 if (mdc_symlink == NULL) {
 461                         CERROR("could not register LMV target "
 462                                "/proc/fs/lustre/%s/%s/target_obds/%s.",
 463                                obd->obd_type->typ_name, obd->obd_name,
 464                                mdc_obd->obd_name);
 465                         lprocfs_remove(&lmv_proc_dir);
 466                         lmv_proc_dir = NULL;
 467                 }
 468         }
 469 #endif
 470         RETURN(0);
 471 }
 472
 473 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
 474 {
 475         struct lmv_obd *lmv = &obd->u.lmv;
 476         struct lmv_tgt_desc *tgt;
 477         int rc = 0;
 478         ENTRY;
 479
 480         CDEBUG(D_CONFIG, "tgt_uuid: %s.\n", tgt_uuid->uuid);
 481
 482         lmv_init_lock(lmv);
 483
 484         if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
 485                 lmv_init_unlock(lmv);
 486                 CERROR("can't add %s, LMV module compiled for %d MDCs. "
 487                        "That many MDCs already configured.\n",
 488                        tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
 489                 RETURN(-EINVAL);
 490         }
 491         if (lmv->desc.ld_tgt_count == 0) {
 492                 struct obd_device *mdc_obd;
 493
 494                 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
 495                                                 &obd->obd_uuid);
 496                 if (!mdc_obd) {
 497                         lmv_init_unlock(lmv);
 498                         CERROR("Target %s not attached\n", tgt_uuid->uuid);
 499                         RETURN(-EINVAL);
 500                 }
 501
 502                 rc = obd_llog_init(obd, &obd->obd_olg, mdc_obd, 0, NULL, tgt_uuid);
 503                 if (rc) {
 504                         lmv_init_unlock(lmv);
 505                         CERROR("lmv failed to setup llogging subsystems\n");
 506                 }
 507         }
 508         spin_lock(&lmv->lmv_lock);
 509         tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
 510         tgt->ltd_uuid = *tgt_uuid;
 511         spin_unlock(&lmv->lmv_lock);
 512
 513         if (lmv->connected) {
 514                 rc = lmv_connect_mdc(obd, tgt);
 515                 if (rc) {
 516                         spin_lock(&lmv->lmv_lock);
 517                         lmv->desc.ld_tgt_count--;
 518                         memset(tgt, 0, sizeof(*tgt));
 519                         spin_unlock(&lmv->lmv_lock);
 520                 } else {
 521                         int easize = sizeof(struct lmv_stripe_md) +
 522                                      lmv->desc.ld_tgt_count *
 523                                      sizeof(struct lu_fid);
 524                         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 525                 }
 526         }
 527
 528         lmv_init_unlock(lmv);
 529         RETURN(rc);
 530 }
 531
 532 /* performs a check if passed obd is connected. If no - connect it. */
 533 int lmv_check_connect(struct obd_device *obd)
 534 {
 535         struct lmv_obd *lmv = &obd->u.lmv;
 536         struct lmv_tgt_desc *tgt;
 537         int i, rc, easize;
 538         ENTRY;
 539
 540         if (lmv->connected)
 541                 RETURN(0);
 542
 543         lmv_init_lock(lmv);
 544         if (lmv->connected) {
 545                 lmv_init_unlock(lmv);
 546                 RETURN(0);
 547         }
 548
 549         if (lmv->desc.ld_tgt_count == 0) {
 550                 CERROR("%s: no targets configured.\n", obd->obd_name);
 551                 RETURN(-EINVAL);
 552         }
 553
 554         CDEBUG(D_CONFIG, "time to connect %s to %s\n",
 555                lmv->cluuid.uuid, obd->obd_name);
 556
 557         LASSERT(lmv->tgts != NULL);
 558
 559         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 560                 rc = lmv_connect_mdc(obd, tgt);
 561                 if (rc)
 562                         GOTO(out_disc, rc);
 563         }
 564
 565         lmv_set_timeouts(obd);
 566         class_export_put(lmv->exp);
 567         lmv->connected = 1;
 568         easize = lmv_get_easize(lmv);
 569         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 570         lmv_init_unlock(lmv);
 571         RETURN(0);
 572
 573  out_disc:
 574         while (i-- > 0) {
 575                 int rc2;
 576                 --tgt;
 577                 tgt->ltd_active = 0;
 578                 if (tgt->ltd_exp) {
 579                         --lmv->desc.ld_active_tgt_count;
 580                         rc2 = obd_disconnect(tgt->ltd_exp);
 581                         if (rc2) {
 582                                 CERROR("error: LMV target %s disconnect on "
 583                                        "MDC idx %d: error %d\n",
 584                                        tgt->ltd_uuid.uuid, i, rc2);
 585                         }
 586                 }
 587         }
 588         class_disconnect(lmv->exp);
 589         lmv_init_unlock(lmv);
 590         RETURN(rc);
 591 }
 592
 593 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 594 {
 595 #ifdef __KERNEL__
 596         struct proc_dir_entry *lmv_proc_dir;
 597 #endif
 598         struct lmv_obd *lmv = &obd->u.lmv;
 599         struct obd_device *mdc_obd;
 600         int rc;
 601         ENTRY;
 602
 603         LASSERT(tgt != NULL);
 604         LASSERT(obd != NULL);
 605
 606         mdc_obd = class_exp2obd(tgt->ltd_exp);
 607
 608         if (mdc_obd)
 609                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 610
 611 #ifdef __KERNEL__
 612         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 613         if (lmv_proc_dir) {
 614                 struct proc_dir_entry *mdc_symlink;
 615
 616                 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
 617                 if (mdc_symlink) {
 618                         lprocfs_remove(&mdc_symlink);
 619                 } else {
 620                         CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
 621                                obd->obd_type->typ_name, obd->obd_name,
 622                                mdc_obd->obd_name);
 623                 }
 624         }
 625 #endif
 626         rc = obd_fid_fini(tgt->ltd_exp);
 627         if (rc)
 628                 CERROR("Can't finanize fids factory\n");
 629
 630         CDEBUG(D_OTHER, "Disconnected from %s(%s) successfully\n",
 631                tgt->ltd_exp->exp_obd->obd_name,
 632                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 633
 634         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 635         rc = obd_disconnect(tgt->ltd_exp);
 636         if (rc) {
 637                 if (tgt->ltd_active) {
 638                         CERROR("Target %s disconnect error %d\n",
 639                                tgt->ltd_uuid.uuid, rc);
 640                 }
 641         }
 642
 643         lmv_activate_target(lmv, tgt, 0);
 644         tgt->ltd_exp = NULL;
 645         RETURN(0);
 646 }
 647
 648 static int lmv_disconnect(struct obd_export *exp)
 649 {
 650         struct obd_device *obd = class_exp2obd(exp);
 651 #ifdef __KERNEL__
 652         struct proc_dir_entry *lmv_proc_dir;
 653 #endif
 654         struct lmv_obd *lmv = &obd->u.lmv;
 655         int rc, i;
 656         ENTRY;
 657
 658         if (!lmv->tgts)
 659                 goto out_local;
 660
 661         /* Only disconnect the underlying layers on the final disconnect. */
 662         lmv->refcount--;
 663         if (lmv->refcount != 0)
 664                 goto out_local;
 665
 666         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 667                 if (lmv->tgts[i].ltd_exp == NULL)
 668                         continue;
 669                 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
 670         }
 671
 672 #ifdef __KERNEL__
 673         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 674         if (lmv_proc_dir) {
 675                 lprocfs_remove(&lmv_proc_dir);
 676         } else {
 677                 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
 678                        obd->obd_type->typ_name, obd->obd_name);
 679         }
 680 #endif
 681
 682 out_local:
 683         /*
 684          * This is the case when no real connection is established by
 685          * lmv_check_connect().
 686          */
 687         if (!lmv->connected)
 688                 class_export_put(exp);
 689         rc = class_disconnect(exp);
 690         if (lmv->refcount == 0)
 691                 lmv->connected = 0;
 692         RETURN(rc);
 693 }
 694
 695 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 696                          int len, void *karg, void *uarg)
 697 {
 698         struct obd_device *obddev = class_exp2obd(exp);
 699         struct lmv_obd *lmv = &obddev->u.lmv;
 700         int i, rc = 0, set = 0;
 701         ENTRY;
 702
 703         if (lmv->desc.ld_tgt_count == 0)
 704                 RETURN(-ENOTTY);
 705
 706         switch (cmd) {
 707         case IOC_OBD_STATFS: {
 708                 struct obd_ioctl_data *data = karg;
 709                 struct obd_device *mdc_obd;
 710                 struct obd_statfs stat_buf = {0};
 711                 __u32 index;
 712
 713                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 714                 LASSERT(data->ioc_plen1 == sizeof(struct obd_statfs));
 715
 716                 if ((index >= lmv->desc.ld_tgt_count))
 717                         RETURN(-ENODEV);
 718
 719                 if (!lmv->tgts[index].ltd_active)
 720                         RETURN(-ENODATA);
 721
 722                 mdc_obd = class_exp2obd(lmv->tgts[index].ltd_exp);
 723                 if (!mdc_obd)
 724                         RETURN(-EINVAL);
 725
 726                 /* got statfs data */
 727                 rc = obd_statfs(mdc_obd, &stat_buf,
 728                                 cfs_time_current_64() - HZ, 0);
 729                 if (rc)
 730                         RETURN(rc);
 731                 if (copy_to_user(data->ioc_pbuf1, &stat_buf, data->ioc_plen1))
 732                         RETURN(rc);
 733                 /* copy UUID */
 734                 rc = copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 735                                   data->ioc_plen2);
 736                 break;
 737         }
 738         default : {
 739                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 740                         int err;
 741
 742                         if (lmv->tgts[i].ltd_exp == NULL)
 743                                 continue;
 744
 745                         err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
 746                                             karg, uarg);
 747                         if (err) {
 748                                 if (lmv->tgts[i].ltd_active) {
 749                                         CERROR("error: iocontrol MDC %s on MDT"
 750                                                "idx %d cmd %x: err = %d\n",
 751                                                 lmv->tgts[i].ltd_uuid.uuid,
 752                                                 i, cmd, err);
 753                                         if (!rc)
 754                                                 rc = err;
 755                                 }
 756                         } else
 757                                 set = 1;
 758                 }
 759                 if (!set && !rc)
 760                         rc = -EIO;
 761         }
 762         }
 763         RETURN(rc);
 764 }
 765
 766 enum MDS_POLICY {
 767      CHAR_TYPE,
 768      NID_TYPE
 769 };
 770
 771 static int lmv_all_chars_policy(int count, const char *name,
 772                                 int len)
 773 {
 774         unsigned int c = 0;
 775
 776         while (len > 0)
 777                 c += name[--len];
 778         c = c % count;
 779         return c;
 780 }
 781
 782 static int lmv_nid_policy(struct lmv_obd *lmv)
 783 {
 784         struct obd_import *imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
 785         __u32 id;
 786         /*
 787          * XXX Hack: to get nid we assume that underlying obd device is mdc.
 788          */
 789         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
 790         return id % lmv->desc.ld_tgt_count;
 791 }
 792
 793 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
 794                           int type)
 795 {
 796         switch (type) {
 797         case CHAR_TYPE:
 798                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
 799                                             op_data->op_name,
 800                                             op_data->op_namelen);
 801         case NID_TYPE:
 802                 return lmv_nid_policy(lmv);
 803
 804         default:
 805                 break;
 806         }
 807
 808         CERROR("unsupport type %d \n", type);
 809         return -EINVAL;
 810 }
 811
 812 /* This is _inode_ placement policy function (not name). */
 813 static int lmv_placement_policy(struct obd_device *obd,
 814                                 struct md_op_data *op_data,
 815                                 mdsno_t *mds)
 816 {
 817         struct lmv_obd *lmv = &obd->u.lmv;
 818         struct lmv_obj *obj;
 819         int rc;
 820         ENTRY;
 821
 822         LASSERT(mds != NULL);
 823
 824         /*
 825          * Allocate new fid on target according to operation type and parent
 826          * home mds.
 827          */
 828         obj = lmv_obj_grab(obd, &op_data->op_fid1);
 829         if (obj != NULL || op_data->op_name == NULL ||
 830             op_data->op_opc != LUSTRE_OPC_MKDIR) {
 831                 /*
 832                  * Allocate fid for non-dir or for null name or for case parent
 833                  * dir is split.
 834                  */
 835                 if (obj) {
 836                         lmv_obj_put(obj);
 837
 838                         /*
 839                          * If we have this flag turned on, and we see that
 840                          * parent dir is split, this means, that caller did not
 841                          * notice split yet. This is race and we would like to
 842                          * let caller know that.
 843                          */
 844                         if (op_data->op_bias & MDS_CHECK_SPLIT)
 845                                 RETURN(-ERESTART);
 846                 }
 847
 848                 /*
 849                  * Allocate new fid on same mds where parent fid is located and
 850                  * where operation will be sent. In case of split dir, ->op_fid1
 851                  * and ->op_mds here will contain fid and mds of slave directory
 852                  * object (assigned by caller).
 853                  */
 854                 *mds = op_data->op_mds;
 855                 rc = 0;
 856
 857 #if 0
 858                 /* XXX: This should be removed later wehn we sure it is not
 859                  * needed. */
 860                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, mds);
 861                 if (rc)
 862                         GOTO(out, rc);
 863 #endif
 864         } else {
 865                 /*
 866                  * Parent directory is not split and we want to create a
 867                  * directory in it. Let's calculate where to place it according
 868                  * to name.
 869                  */
 870                 *mds = lmv_choose_mds(lmv, op_data, NID_TYPE);
 871                 rc = 0;
 872         }
 873         EXIT;
 874 #if 0
 875 out:
 876 #endif
 877         if (rc) {
 878                 CERROR("Can't choose MDS, err = %d\n", rc);
 879         } else {
 880                 LASSERT(*mds < lmv->desc.ld_tgt_count);
 881         }
 882
 883         return rc;
 884 }
 885
 886 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
 887                     mdsno_t mds)
 888 {
 889         struct lmv_tgt_desc *tgt = &lmv->tgts[mds];
 890         int rc;
 891         ENTRY;
 892
 893         /* New seq alloc and FLD setup should be atomic. */
 894         down(&tgt->ltd_fid_sem);
 895
 896         /* Asking underlaying tgt layer to allocate new fid. */
 897         rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
 898         if (rc > 0) {
 899                 LASSERT(fid_is_sane(fid));
 900
 901                 /* Client switches to new sequence, setup FLD. */
 902                 rc = fld_client_create(&lmv->lmv_fld, fid_seq(fid),
 903                                        mds, NULL);
 904                 if (rc) {
 905                         CERROR("Can't create fld entry, rc %d\n", rc);
 906                         /* Delete just allocated fid sequence */
 907                         obd_fid_delete(tgt->ltd_exp, NULL);
 908                 }
 909         }
 910         up(&tgt->ltd_fid_sem);
 911         RETURN(rc);
 912 }
 913
 914 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
 915                   struct md_op_data *op_data)
 916 {
 917         struct obd_device *obd = class_exp2obd(exp);
 918         struct lmv_obd *lmv = &obd->u.lmv;
 919         mdsno_t mds;
 920         int rc;
 921         ENTRY;
 922
 923         LASSERT(op_data != NULL);
 924         LASSERT(fid != NULL);
 925
 926         rc = lmv_placement_policy(obd, op_data, &mds);
 927         if (rc) {
 928                 CERROR("Can't get target for allocating fid, "
 929                        "rc %d\n", rc);
 930                 RETURN(rc);
 931         }
 932
 933         rc = __lmv_fid_alloc(lmv, fid, mds);
 934         if (rc) {
 935                 CERROR("Can't alloc new fid, rc %d\n", rc);
 936                 RETURN(rc);
 937         }
 938
 939         RETURN(rc);
 940 }
 941
 942 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
 943 {
 944         ENTRY;
 945
 946         LASSERT(exp && fid);
 947         if (lmv_obj_delete(exp, fid)) {
 948                 CDEBUG(D_OTHER, "lmv object "DFID" is destroyed.\n",
 949                        PFID(fid));
 950         }
 951         RETURN(0);
 952 }
 953
 954 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 955 {
 956         struct lmv_obd *lmv = &obd->u.lmv;
 957         struct lprocfs_static_vars lvars;
 958         struct lmv_desc *desc;
 959         int rc, i = 0;
 960         ENTRY;
 961
 962         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 963                 CERROR("LMV setup requires a descriptor\n");
 964                 RETURN(-EINVAL);
 965         }
 966
 967         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
 968         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
 969                 CERROR("descriptor size wrong: %d > %d\n",
 970                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
 971                 RETURN(-EINVAL);
 972         }
 973
 974         lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
 975
 976         OBD_ALLOC(lmv->tgts, lmv->tgts_size);
 977         if (lmv->tgts == NULL)
 978                 RETURN(-ENOMEM);
 979
 980         for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
 981                 sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
 982                 lmv->tgts[i].ltd_idx = i;
 983         }
 984
 985         lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
 986
 987         OBD_ALLOC(lmv->datas, lmv->datas_size);
 988         if (lmv->datas == NULL)
 989                 GOTO(out_free_tgts, rc = -ENOMEM);
 990
 991         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
 992         lmv->desc.ld_tgt_count = 0;
 993         lmv->desc.ld_active_tgt_count = 0;
 994         lmv->max_cookiesize = 0;
 995         lmv->max_def_easize = 0;
 996         lmv->max_easize = 0;
 997
 998         spin_lock_init(&lmv->lmv_lock);
 999         sema_init(&lmv->init_sem, 1);
1000
1001         rc = lmv_obj_setup(obd);
1002         if (rc) {
1003                 CERROR("Can't setup LMV object manager, "
1004                        "error %d.\n", rc);
1005                 GOTO(out_free_datas, rc);
1006         }
1007
1008         lprocfs_lmv_init_vars(&lvars);
1009         lprocfs_obd_setup(obd, lvars.obd_vars);
1010 #ifdef LPROCFS
1011         {
1012                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1013                                         0444, &lmv_proc_target_fops, obd);
1014                 if (rc)
1015                         CWARN("Error adding the target_obd_status file\n");
1016        }
1017 #endif
1018         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1019                              LUSTRE_CLI_FLD_HASH_DHT);
1020         if (rc) {
1021                 CERROR("can't init FLD, err %d\n",
1022                        rc);
1023                 GOTO(out_free_datas, rc);
1024         }
1025
1026         RETURN(0);
1027
1028 out_free_datas:
1029         OBD_FREE(lmv->datas, lmv->datas_size);
1030         lmv->datas = NULL;
1031 out_free_tgts:
1032         OBD_FREE(lmv->tgts, lmv->tgts_size);
1033         lmv->tgts = NULL;
1034         return rc;
1035 }
1036
1037 static int lmv_cleanup(struct obd_device *obd)
1038 {
1039         struct lmv_obd *lmv = &obd->u.lmv;
1040         ENTRY;
1041
1042         fld_client_fini(&lmv->lmv_fld);
1043         lprocfs_obd_cleanup(obd);
1044         lmv_obj_cleanup(obd);
1045         OBD_FREE(lmv->datas, lmv->datas_size);
1046         OBD_FREE(lmv->tgts, lmv->tgts_size);
1047
1048         RETURN(0);
1049 }
1050
1051 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1052 {
1053         struct lustre_cfg *lcfg = buf;
1054         struct obd_uuid tgt_uuid;
1055         int rc;
1056         ENTRY;
1057
1058         switch(lcfg->lcfg_command) {
1059         case LCFG_ADD_MDC:
1060                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1061                         GOTO(out, rc = -EINVAL);
1062
1063                 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1064                 rc = lmv_add_target(obd, &tgt_uuid);
1065                 GOTO(out, rc);
1066         default: {
1067                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1068                 GOTO(out, rc = -EINVAL);
1069         }
1070         }
1071 out:
1072         RETURN(rc);
1073 }
1074
1075 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1076                       __u64 max_age, __u32 flags)
1077 {
1078         struct lmv_obd *lmv = &obd->u.lmv;
1079         struct obd_statfs *temp;
1080         int rc = 0, i;
1081         ENTRY;
1082
1083         rc = lmv_check_connect(obd);
1084         if (rc)
1085                 RETURN(rc);
1086
1087         OBD_ALLOC(temp, sizeof(*temp));
1088         if (temp == NULL)
1089                 RETURN(-ENOMEM);
1090
1091         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1092                 if (lmv->tgts[i].ltd_exp == NULL)
1093                         continue;
1094
1095                 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1096                                 max_age, flags);
1097                 if (rc) {
1098                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1099                                lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1100                                rc);
1101                         GOTO(out_free_temp, rc);
1102                 }
1103                 if (i == 0) {
1104                         *osfs = *temp;
1105                 } else {
1106                         osfs->os_bavail += temp->os_bavail;
1107                         osfs->os_blocks += temp->os_blocks;
1108                         osfs->os_ffree += temp->os_ffree;
1109                         osfs->os_files += temp->os_files;
1110                 }
1111         }
1112
1113         EXIT;
1114 out_free_temp:
1115         OBD_FREE(temp, sizeof(*temp));
1116         return rc;
1117 }
1118
1119 static int lmv_getstatus(struct obd_export *exp,
1120                          struct lu_fid *fid,
1121                          struct obd_capa **pc)
1122 {
1123         struct obd_device *obd = exp->exp_obd;
1124         struct lmv_obd *lmv = &obd->u.lmv;
1125         int rc;
1126         ENTRY;
1127
1128         rc = lmv_check_connect(obd);
1129         if (rc)
1130                 RETURN(rc);
1131
1132         rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1133
1134         RETURN(rc);
1135 }
1136
1137 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1138                         struct obd_capa *oc, obd_valid valid, const char *name,
1139                         const char *input, int input_size, int output_size,
1140                         int flags, struct ptlrpc_request **request)
1141 {
1142         struct obd_device *obd = exp->exp_obd;
1143         struct lmv_obd *lmv = &obd->u.lmv;
1144         struct obd_export *tgt_exp;
1145         int rc;
1146         ENTRY;
1147
1148         rc = lmv_check_connect(obd);
1149         if (rc)
1150                 RETURN(rc);
1151
1152         tgt_exp = lmv_find_export(lmv, fid);
1153         if (IS_ERR(tgt_exp))
1154                 RETURN(PTR_ERR(tgt_exp));
1155
1156         rc = md_getxattr(tgt_exp, fid, oc, valid, name, input, input_size,
1157                          output_size, flags, request);
1158
1159         RETURN(rc);
1160 }
1161
1162 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1163                         struct obd_capa *oc, obd_valid valid, const char *name,
1164                         const char *input, int input_size, int output_size,
1165                         int flags, __u32 suppgid,
1166                         struct ptlrpc_request **request)
1167 {
1168         struct obd_device *obd = exp->exp_obd;
1169         struct lmv_obd *lmv = &obd->u.lmv;
1170         struct obd_export *tgt_exp;
1171         int rc;
1172         ENTRY;
1173
1174         rc = lmv_check_connect(obd);
1175         if (rc)
1176                 RETURN(rc);
1177
1178         tgt_exp = lmv_find_export(lmv, fid);
1179         if (IS_ERR(tgt_exp))
1180                 RETURN(PTR_ERR(tgt_exp));
1181
1182         rc = md_setxattr(tgt_exp, fid, oc, valid, name,
1183                          input, input_size, output_size, flags, suppgid,
1184                          request);
1185
1186         RETURN(rc);
1187 }
1188
1189 static int lmv_getattr(struct obd_export *exp, const struct lu_fid *fid,
1190                        struct obd_capa *oc, obd_valid valid, int ea_size,
1191                        struct ptlrpc_request **request)
1192 {
1193         struct obd_device *obd = exp->exp_obd;
1194         struct lmv_obd *lmv = &obd->u.lmv;
1195         struct obd_export *tgt_exp;
1196         struct lmv_obj *obj;
1197         int rc, i;
1198         ENTRY;
1199
1200         rc = lmv_check_connect(obd);
1201         if (rc)
1202                 RETURN(rc);
1203
1204         tgt_exp = lmv_find_export(lmv, fid);
1205         if (IS_ERR(tgt_exp))
1206                 RETURN(PTR_ERR(tgt_exp));
1207
1208         rc = md_getattr(tgt_exp, fid, oc, valid, ea_size, request);
1209         if (rc)
1210                 RETURN(rc);
1211
1212         obj = lmv_obj_grab(obd, fid);
1213
1214         CDEBUG(D_OTHER, "GETATTR for "DFID" %s\n", PFID(fid),
1215                obj ? "(split)" : "");
1216
1217         /*
1218          * If object is split, then we loop over all the slaves and gather size
1219          * attribute. In ideal world we would have to gather also mds field from
1220          * all slaves, as object is spread over the cluster and this is
1221          * definitely interesting information and it is not good to loss it,
1222          * but...
1223          */
1224         if (obj) {
1225                 struct mdt_body *body;
1226
1227                 if (*request == NULL) {
1228                         lmv_obj_put(obj);
1229                         RETURN(rc);
1230                 }
1231
1232                 body = req_capsule_server_get(&(*request)->rq_pill,
1233                                               &RMF_MDT_BODY);
1234                 LASSERT(body != NULL);
1235
1236                 lmv_obj_lock(obj);
1237
1238                 for (i = 0; i < obj->lo_objcount; i++) {
1239                         if (lmv->tgts[i].ltd_exp == NULL) {
1240                                 CWARN("%s: NULL export for %d\n",
1241                                       obd->obd_name, i);
1242                                 continue;
1243                         }
1244
1245                         /* skip master obj. */
1246                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_inodes[i].li_fid))
1247                                 continue;
1248
1249                         lmv_update_body(body, &obj->lo_inodes[i]);
1250                 }
1251
1252                 lmv_obj_unlock(obj);
1253                 lmv_obj_put(obj);
1254         }
1255
1256         RETURN(rc);
1257 }
1258
1259 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1260                              ldlm_iterator_t it, void *data)
1261 {
1262         struct obd_device *obd = exp->exp_obd;
1263         struct lmv_obd *lmv = &obd->u.lmv;
1264         int i, rc;
1265         ENTRY;
1266
1267         rc = lmv_check_connect(obd);
1268         if (rc)
1269                 RETURN(rc);
1270
1271         CDEBUG(D_OTHER, "CBDATA for "DFID"\n", PFID(fid));
1272
1273         /*
1274          * With CMD every object can have two locks in different namespaces:
1275          * lookup lock in space of mds storing direntry and update/open lock in
1276          * space of mds storing inode.
1277          */
1278         for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1279                 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1280
1281         RETURN(0);
1282 }
1283
1284 static int lmv_close(struct obd_export *exp,
1285                      struct md_op_data *op_data,
1286                      struct md_open_data *mod,
1287                      struct ptlrpc_request **request)
1288 {
1289         struct obd_device *obd = exp->exp_obd;
1290         struct lmv_obd *lmv = &obd->u.lmv;
1291         struct obd_export *tgt_exp;
1292         int rc;
1293         ENTRY;
1294
1295         rc = lmv_check_connect(obd);
1296         if (rc)
1297                 RETURN(rc);
1298
1299         tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1300         if (IS_ERR(tgt_exp))
1301                 RETURN(PTR_ERR(tgt_exp));
1302
1303         CDEBUG(D_OTHER, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1304         rc = md_close(tgt_exp, op_data, mod, request);
1305         RETURN(rc);
1306 }
1307
1308 /*
1309  * Called in the case MDS returns -ERESTART on create on open, what means that
1310  * directory is split and its LMV presentation object has to be updated.
1311  */
1312 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1313 {
1314         struct obd_device *obd = exp->exp_obd;
1315         struct lmv_obd *lmv = &obd->u.lmv;
1316         struct ptlrpc_request *req = NULL;
1317         struct obd_export *tgt_exp;
1318         struct lmv_obj *obj;
1319         struct lustre_md md;
1320         int mealen, rc;
1321         __u64 valid;
1322         ENTRY;
1323
1324         md.mea = NULL;
1325         mealen = lmv_get_easize(lmv);
1326
1327         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1328
1329         tgt_exp = lmv_find_export(lmv, fid);
1330         if (IS_ERR(tgt_exp))
1331                 RETURN(PTR_ERR(tgt_exp));
1332
1333         /* time to update mea of parent fid */
1334         rc = md_getattr(tgt_exp, fid, NULL, valid, mealen, &req);
1335         if (rc) {
1336                 CERROR("md_getattr() failed, error %d\n", rc);
1337                 GOTO(cleanup, rc);
1338         }
1339
1340         rc = md_get_lustre_md(tgt_exp, req, NULL, exp, &md);
1341         if (rc) {
1342                 CERROR("mdc_get_lustre_md() failed, error %d\n", rc);
1343                 GOTO(cleanup, rc);
1344         }
1345
1346         if (md.mea == NULL)
1347                 GOTO(cleanup, rc = -ENODATA);
1348
1349         obj = lmv_obj_create(exp, fid, md.mea);
1350         if (IS_ERR(obj))
1351                 rc = PTR_ERR(obj);
1352         else
1353                 lmv_obj_put(obj);
1354
1355         /* XXX LOV STACKING */
1356         obd_free_memmd(exp, (void *)&md.mea);
1357
1358         EXIT;
1359 cleanup:
1360         if (req)
1361                 ptlrpc_req_finished(req);
1362         return rc;
1363 }
1364
1365 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1366                const void *data, int datalen, int mode, __u32 uid,
1367                __u32 gid, __u32 cap_effective,  __u64 rdev,
1368                struct ptlrpc_request **request)
1369 {
1370         struct obd_device *obd = exp->exp_obd;
1371         struct lmv_obd *lmv = &obd->u.lmv;
1372         struct obd_export *tgt_exp;
1373         struct lmv_obj *obj;
1374         int rc, loop = 0;
1375         ENTRY;
1376
1377         rc = lmv_check_connect(obd);
1378         if (rc)
1379                 RETURN(rc);
1380
1381         if (!lmv->desc.ld_active_tgt_count)
1382                 RETURN(-EIO);
1383 repeat:
1384         ++loop;
1385         LASSERT(loop <= 2);
1386         obj = lmv_obj_grab(obd, &op_data->op_fid1);
1387         if (obj) {
1388                 int mea_idx;
1389
1390                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1391                                        op_data->op_name, op_data->op_namelen);
1392                 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1393                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1394                 op_data->op_mds = obj->lo_inodes[mea_idx].li_mds;
1395                 tgt_exp = lmv_get_export(lmv, op_data->op_mds);
1396                 lmv_obj_put(obj);
1397         } else {
1398                 struct lmv_tgt_desc *tgt;
1399
1400                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1401                 op_data->op_bias |= MDS_CHECK_SPLIT;
1402                 op_data->op_mds = tgt->ltd_idx;
1403                 tgt_exp = tgt->ltd_exp;
1404         }
1405
1406         if (IS_ERR(tgt_exp))
1407                 RETURN(PTR_ERR(tgt_exp));
1408
1409         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1410         if (rc == -ERESTART)
1411                 goto repeat;
1412         else if (rc)
1413                 RETURN(rc);
1414
1415         CDEBUG(D_OTHER, "CREATE '%*s' on "DFID"\n", op_data->op_namelen,
1416                op_data->op_name, PFID(&op_data->op_fid1));
1417
1418         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1419         rc = md_create(tgt_exp, op_data, data, datalen, mode, uid, gid,
1420                        cap_effective, rdev, request);
1421         if (rc == 0) {
1422                 if (*request == NULL)
1423                         RETURN(rc);
1424                 CDEBUG(D_OTHER, "created - "DFID"\n", PFID(&op_data->op_fid1));
1425         } else if (rc == -ERESTART) {
1426                 LASSERT(*request != NULL);
1427                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1428                           "Got -ERESTART during create!\n");
1429                 ptlrpc_req_finished(*request);
1430                 *request = NULL;
1431
1432                 /*
1433                  * Directory got split. Time to update local object and repeat
1434                  * the request with proper MDS.
1435                  */
1436                 rc = lmv_handle_split(exp, &op_data->op_fid1);
1437                 if (rc == 0) {
1438                         rc = lmv_alloc_slave_fids(obd, &op_data->op_fid1,
1439                                                   op_data, &op_data->op_fid2);
1440                         if (rc)
1441                                 RETURN(rc);
1442                         goto repeat;
1443                 }
1444         }
1445         RETURN(rc);
1446 }
1447
1448 static int lmv_done_writing(struct obd_export *exp,
1449                             struct md_op_data *op_data,
1450                             struct md_open_data *mod)
1451 {
1452         struct obd_device *obd = exp->exp_obd;
1453         struct lmv_obd *lmv = &obd->u.lmv;
1454         struct obd_export *tgt_exp;
1455         int rc;
1456         ENTRY;
1457
1458         rc = lmv_check_connect(obd);
1459         if (rc)
1460                 RETURN(rc);
1461
1462         tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1463         if (IS_ERR(tgt_exp))
1464                 RETURN(PTR_ERR(tgt_exp));
1465
1466         rc = md_done_writing(tgt_exp, op_data, mod);
1467         RETURN(rc);
1468 }
1469
1470 static int
1471 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1472                    struct lookup_intent *it, struct md_op_data *op_data,
1473                    struct lustre_handle *lockh, void *lmm, int lmmsize)
1474 {
1475         struct obd_device *obd = exp->exp_obd;
1476         struct lmv_obd *lmv = &obd->u.lmv;
1477         struct lmv_stripe_md *mea = op_data->op_mea1;
1478         struct md_op_data *op_data2;
1479         struct obd_export *tgt_exp;
1480         int i, rc = 0;
1481         ENTRY;
1482
1483         OBD_ALLOC_PTR(op_data2);
1484         if (op_data2 == NULL)
1485                 RETURN(-ENOMEM);
1486
1487         LASSERT(mea != NULL);
1488         for (i = 0; i < mea->mea_count; i++) {
1489                 memset(op_data2, 0, sizeof(*op_data2));
1490                 op_data2->op_fid1 = mea->mea_ids[i];
1491                 op_data2->op_bias = 0;
1492
1493                 tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1);
1494                 if (IS_ERR(tgt_exp))
1495                         GOTO(cleanup, rc = PTR_ERR(tgt_exp));
1496
1497                 if (tgt_exp == NULL)
1498                         continue;
1499
1500                 rc = md_enqueue(tgt_exp, einfo, it, op_data2,
1501                                 lockh + i, lmm, lmmsize, 0);
1502
1503                 CDEBUG(D_OTHER, "take lock on slave "DFID" -> %d/%d\n",
1504                        PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1505
1506                 if (rc)
1507                         GOTO(cleanup, rc);
1508
1509                 if (it->d.lustre.it_data) {
1510                         struct ptlrpc_request *req;
1511                         req = (struct ptlrpc_request *)it->d.lustre.it_data;
1512                         ptlrpc_req_finished(req);
1513                 }
1514
1515                 if (it->d.lustre.it_status)
1516                         GOTO(cleanup, rc = it->d.lustre.it_status);
1517         }
1518
1519         EXIT;
1520 cleanup:
1521         OBD_FREE_PTR(op_data2);
1522
1523         if (rc != 0) {
1524                 /* drop all taken locks */
1525                 while (--i >= 0) {
1526                         if (lockh[i].cookie)
1527                                 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1528                         lockh[i].cookie = 0;
1529                 }
1530         }
1531         return rc;
1532 }
1533
1534 static int
1535 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1536                    struct lookup_intent *it, struct md_op_data *op_data,
1537                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1538                    int extra_lock_flags)
1539 {
1540         struct ptlrpc_request *req = it->d.lustre.it_data;
1541         struct obd_device *obd = exp->exp_obd;
1542         struct lmv_obd *lmv = &obd->u.lmv;
1543         struct lustre_handle plock;
1544         struct obd_export *tgt_exp;
1545         struct md_op_data *rdata;
1546         struct lu_fid fid_copy;
1547         struct mdt_body *body;
1548         int rc = 0, pmode;
1549         ENTRY;
1550
1551         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1552         LASSERT(body != NULL);
1553
1554         if (!(body->valid & OBD_MD_MDS))
1555                 RETURN(0);
1556
1557         CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID" -> "DFID"\n",
1558                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1559
1560         /* We got LOOKUP lock, but we really need attrs */
1561         pmode = it->d.lustre.it_lock_mode;
1562         LASSERT(pmode != 0);
1563         memcpy(&plock, lockh, sizeof(plock));
1564         it->d.lustre.it_lock_mode = 0;
1565         it->d.lustre.it_data = NULL;
1566         fid_copy = body->fid1;
1567
1568         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1569         ptlrpc_req_finished(req);
1570
1571         tgt_exp = lmv_find_export(lmv, &fid_copy);
1572         if (IS_ERR(tgt_exp))
1573                 GOTO(out, rc = PTR_ERR(tgt_exp));
1574
1575         OBD_ALLOC_PTR(rdata);
1576         if (rdata == NULL)
1577                 GOTO(out, rc = -ENOMEM);
1578
1579         rdata->op_fid1 = fid_copy;
1580         rdata->op_bias = MDS_CROSS_REF;
1581
1582         rc = md_enqueue(tgt_exp, einfo, it, rdata, lockh,
1583                         lmm, lmmsize, extra_lock_flags);
1584         OBD_FREE_PTR(rdata);
1585         EXIT;
1586 out:
1587         ldlm_lock_decref(&plock, pmode);
1588         return rc;
1589 }
1590
1591 static int
1592 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1593             struct lookup_intent *it, struct md_op_data *op_data,
1594             struct lustre_handle *lockh, void *lmm, int lmmsize,
1595             int extra_lock_flags)
1596 {
1597         struct obd_device *obd = exp->exp_obd;
1598         struct lmv_obd *lmv = &obd->u.lmv;
1599         struct obd_export *tgt_exp = NULL;
1600         struct lmv_obj *obj;
1601         int rc;
1602         ENTRY;
1603
1604         rc = lmv_check_connect(obd);
1605         if (rc)
1606                 RETURN(rc);
1607
1608         if (op_data->op_mea1 && it->it_op == IT_UNLINK) {
1609                 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1610                                         lockh, lmm, lmmsize);
1611                 RETURN(rc);
1612         }
1613
1614         if (op_data->op_namelen) {
1615                 obj = lmv_obj_grab(obd, &op_data->op_fid1);
1616                 if (obj) {
1617                         int mea_idx;
1618
1619                         /* directory is split. look for right mds for this
1620                          * name */
1621                         mea_idx = raw_name2idx(obj->lo_hashtype,
1622                                                obj->lo_objcount,
1623                                                (char *)op_data->op_name,
1624                                                op_data->op_namelen);
1625                         op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1626                         tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds);
1627                         lmv_obj_put(obj);
1628                 }
1629         }
1630
1631         if (tgt_exp == NULL)
1632                 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1633         if (IS_ERR(tgt_exp))
1634                 RETURN(PTR_ERR(tgt_exp));
1635
1636         CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID"\n", LL_IT2STR(it),
1637                PFID(&op_data->op_fid1));
1638
1639         rc = md_enqueue(tgt_exp, einfo, it, op_data, lockh,
1640                         lmm, lmmsize, extra_lock_flags);
1641
1642         if (rc == 0 && it->it_op == IT_OPEN)
1643                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1644                                         lmm, lmmsize, extra_lock_flags);
1645         RETURN(rc);
1646 }
1647
1648 static int
1649 lmv_getattr_name(struct obd_export *exp, const struct lu_fid *fid,
1650                  struct obd_capa *oc, const char *filename, int namelen,
1651                  obd_valid valid, int ea_size, __u32 suppgid,
1652                  struct ptlrpc_request **request)
1653 {
1654         struct obd_device *obd = exp->exp_obd;
1655         struct lmv_obd *lmv = &obd->u.lmv;
1656         struct lu_fid rid = *fid;
1657         struct obd_export *tgt_exp;
1658         struct mdt_body *body;
1659         struct lmv_obj *obj;
1660         int rc, loop = 0;
1661         ENTRY;
1662
1663         rc = lmv_check_connect(obd);
1664         if (rc)
1665                 RETURN(rc);
1666
1667 repeat:
1668         ++loop;
1669         LASSERT(loop <= 2);
1670         obj = lmv_obj_grab(obd, &rid);
1671         if (obj) {
1672                 int mea_idx;
1673
1674                 /* Directory is split. Look for right mds for this name */
1675                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1676                                        filename, namelen - 1);
1677                 rid = obj->lo_inodes[mea_idx].li_fid;
1678                 tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds);
1679                 lmv_obj_put(obj);
1680                 valid &= ~OBD_MD_FLCKSPLIT;
1681         } else {
1682                 tgt_exp = lmv_find_export(lmv, &rid);
1683                 valid |= OBD_MD_FLCKSPLIT;
1684         }
1685         if (IS_ERR(tgt_exp))
1686                 RETURN(PTR_ERR(tgt_exp));
1687
1688         CDEBUG(D_OTHER, "getattr_name for %*s on "DFID" -> "DFID"\n",
1689                namelen, filename, PFID(fid), PFID(&rid));
1690
1691         rc = md_getattr_name(tgt_exp, &rid, oc, filename, namelen, valid,
1692                              ea_size, suppgid, request);
1693         if (rc == 0) {
1694                 body = req_capsule_server_get(&(*request)->rq_pill,
1695                                               &RMF_MDT_BODY);
1696                 LASSERT(body != NULL);
1697
1698                 if (body->valid & OBD_MD_MDS) {
1699                         struct ptlrpc_request *req = NULL;
1700
1701                         rid = body->fid1;
1702                         CDEBUG(D_OTHER, "request attrs for "DFID"\n",
1703                                PFID(&rid));
1704
1705                         tgt_exp = lmv_find_export(lmv, &rid);
1706                         if (IS_ERR(tgt_exp)) {
1707                                 ptlrpc_req_finished(*request);
1708                                 RETURN(PTR_ERR(tgt_exp));
1709                         }
1710
1711                         rc = md_getattr_name(tgt_exp, &rid, NULL, NULL, 1,
1712                                              valid | OBD_MD_FLCROSSREF,
1713                                              ea_size, suppgid, &req);
1714                         ptlrpc_req_finished(*request);
1715                         *request = req;
1716                 }
1717         } else if (rc == -ERESTART) {
1718                 LASSERT(*request != NULL);
1719                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1720                           "Got -ERESTART during getattr!\n");
1721                 ptlrpc_req_finished(*request);
1722                 *request = NULL;
1723
1724                 /*
1725                  * Directory got split. Time to update local object and repeat
1726                  * the request with proper MDS.
1727                  */
1728                 rc = lmv_handle_split(exp, &rid);
1729                 if (rc == 0)
1730                         goto repeat;
1731         }
1732         RETURN(rc);
1733 }
1734
1735 #define md_op_data_fid(op_data, fl)                     \
1736         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1737          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1738          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1739          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1740          NULL)
1741
1742 /* @tgt_exp is the export the metadata request is sent.
1743  * @fid_exp is the export the cancel should be sent for the current fid.
1744  * if @fid_exp is NULL, the export is found for the current fid.
1745  * @op_data keeps the current fid, which is pointed through @flag.
1746  * @mode, @bits -- lock match parameters. */
1747 static int lmv_early_cancel(struct lmv_obd *lmv, struct obd_export *tgt_exp,
1748                             struct obd_export *fid_exp,
1749                             struct md_op_data *op_data,
1750                             ldlm_mode_t mode, int bits, int flag)
1751 {
1752         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1753         ldlm_policy_data_t policy = {{0}};
1754         int rc = 0;
1755         ENTRY;
1756
1757         if (!fid_is_sane(fid))
1758                 RETURN(0);
1759
1760         if (fid_exp == NULL)
1761                 fid_exp = lmv_find_export(lmv, fid);
1762
1763         if (tgt_exp == fid_exp) {
1764                 /* The export is the same as on the target server, cancel
1765                  * will be sent along with the main metadata operation. */
1766                 op_data->op_flags |= flag;
1767                 RETURN(0);
1768         }
1769
1770         policy.l_inodebits.bits = bits;
1771         rc = md_cancel_unused(fid_exp, fid, &policy, mode, LDLM_FL_ASYNC, NULL);
1772         RETURN(rc);
1773 }
1774
1775 #ifdef EARLY_CANCEL_FOR_STRIPED_DIR_IS_READY
1776 /* Check if the fid in @op_data pointed to by flag is of the same export(s)
1777  * as @tgt_exp. Early cancels will be sent later by mdc code, otherwise, call
1778  * md_cancel_unused for child export(s). */
1779 static int lmv_early_cancel_stripes(struct obd_export *exp,
1780                                     struct obd_export *tgt_exp,
1781                                     struct md_op_data *op_data,
1782                                     ldlm_mode_t mode, int bits, int flag)
1783 {
1784         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1785         struct obd_device *obd = exp->exp_obd;
1786         struct lmv_obd *lmv = &obd->u.lmv;
1787         struct obd_export *st_exp;
1788         struct lmv_obj *obj;
1789         int rc = 0;
1790         ENTRY;
1791
1792         if (!fid_is_sane(fid))
1793                 RETURN(0);
1794
1795         obj = lmv_obj_grab(obd, fid);
1796         if (obj) {
1797                 ldlm_policy_data_t policy = {{0}};
1798                 struct lu_fid *st_fid;
1799                 int i;
1800
1801                 policy.l_inodebits.bits = bits;
1802                 for (i = 0; i < obj->lo_objcount; i++) {
1803                         st_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds);
1804                         st_fid = &obj->lo_inodes[i].li_fid;
1805                         if (tgt_exp != st_exp) {
1806                                 rc = md_cancel_unused(st_exp, st_fid, &policy,
1807                                                       mode, LDLM_FL_ASYNC,
1808                                                       NULL);
1809                                 if (rc)
1810                                         break;
1811                         } else {
1812                                 /* Some export matches to @tgt_exp, do cancel
1813                                  * for its fid in mdc */
1814                                 *fid = *st_fid;
1815                                 op_data->op_flags |= flag;
1816                         }
1817                 }
1818                 lmv_obj_put(obj);
1819         } else {
1820                 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data,
1821                                       mode, bits, flag);
1822         }
1823         RETURN(rc);
1824 }
1825 #endif
1826
1827 /*
1828  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1829  * op_data->op_fid2
1830  */
1831 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1832                     struct ptlrpc_request **request)
1833 {
1834         struct obd_device *obd = exp->exp_obd;
1835         struct lmv_obd *lmv = &obd->u.lmv;
1836         struct obd_export *tgt_exp;
1837         struct lmv_obj *obj;
1838         int rc, loop = 0;
1839         mdsno_t mds;
1840         ENTRY;
1841
1842         rc = lmv_check_connect(obd);
1843         if (rc)
1844                 RETURN(rc);
1845
1846 repeat:
1847         ++loop;
1848         LASSERT(loop <= 2);
1849         if (op_data->op_namelen != 0) {
1850                 int mea_idx;
1851
1852                 /* Usual link request */
1853                 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1854                 if (obj) {
1855                         mea_idx = raw_name2idx(obj->lo_hashtype,
1856                                                obj->lo_objcount,
1857                                                op_data->op_name,
1858                                                op_data->op_namelen);
1859                         op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1860                         mds = obj->lo_inodes[mea_idx].li_mds;
1861                         lmv_obj_put(obj);
1862                 } else {
1863                         rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1864                         if (rc)
1865                                 RETURN(rc);
1866                 }
1867
1868                 CDEBUG(D_OTHER,"link "DFID":%*s to "DFID"\n",
1869                        PFID(&op_data->op_fid2), op_data->op_namelen,
1870                        op_data->op_name, PFID(&op_data->op_fid1));
1871         } else {
1872                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds);
1873                 if (rc)
1874                         RETURN(rc);
1875
1876                 /* request from MDS to acquire i_links for inode by fid1 */
1877                 CDEBUG(D_OTHER, "inc i_nlinks for "DFID"\n",
1878                        PFID(&op_data->op_fid1));
1879         }
1880
1881         CDEBUG(D_OTHER, "forward to MDS #"LPU64" ("DFID")\n",
1882                mds, PFID(&op_data->op_fid1));
1883
1884         op_data->op_fsuid = current->fsuid;
1885         op_data->op_fsgid = current->fsgid;
1886         op_data->op_cap   = current->cap_effective;
1887
1888         tgt_exp = lmv->tgts[mds].ltd_exp;
1889         if (op_data->op_namelen) {
1890                 op_data->op_flags |= MF_MDC_CANCEL_FID2;
1891                 /* Cancel UPDATE lock on child (fid1). */
1892                 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX,
1893                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1894         }
1895         if (rc == 0)
1896                 rc = md_link(tgt_exp, op_data, request);
1897         if (rc == -ERESTART) {
1898                 LASSERT(*request != NULL);
1899                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1900                           "Got -ERESTART during link!\n");
1901                 ptlrpc_req_finished(*request);
1902                 *request = NULL;
1903
1904                 /*
1905                  * Directory got split. Time to update local object and repeat
1906                  * the request with proper MDS.
1907                  */
1908                 rc = lmv_handle_split(exp, &op_data->op_fid2);
1909                 if (rc == 0)
1910                         goto repeat;
1911         }
1912
1913         RETURN(rc);
1914 }
1915
1916 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
1917                       const char *old, int oldlen, const char *new, int newlen,
1918                       struct ptlrpc_request **request)
1919 {
1920         struct obd_export *tgt_exp = NULL, *src_exp;
1921         struct obd_device *obd = exp->exp_obd;
1922         struct lmv_obd *lmv = &obd->u.lmv;
1923         int rc, mea_idx, loop = 0;
1924         struct lmv_obj *obj;
1925         mdsno_t mds1, mds2;
1926         ENTRY;
1927
1928         CDEBUG(D_OTHER, "rename %*s in "DFID" to %*s in "DFID"\n",
1929                oldlen, old, PFID(&op_data->op_fid1),
1930                newlen, new, PFID(&op_data->op_fid2));
1931
1932         rc = lmv_check_connect(obd);
1933         if (rc)
1934                 RETURN(rc);
1935
1936         if (oldlen == 0) {
1937                 /*
1938                  * MDS with old dir entry is asking another MDS to create name
1939                  * there.
1940                  */
1941                 CDEBUG(D_OTHER,
1942                        "create %*s(%d/%d) in "DFID" pointing "
1943                        "to "DFID"\n", newlen, new, oldlen, newlen,
1944                        PFID(&op_data->op_fid2), PFID(&op_data->op_fid1));
1945
1946                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds1);
1947                 if (rc)
1948                         RETURN(rc);
1949
1950                 /*
1951                  * Target directory can be split, sowe should forward request to
1952                  * the right MDS.
1953                  */
1954                 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1955                 if (obj) {
1956                         mea_idx = raw_name2idx(obj->lo_hashtype,
1957                                                obj->lo_objcount,
1958                                                (char *)new, newlen);
1959                         op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1960                         CDEBUG(D_OTHER, "Parent obj "DFID"\n",
1961                                PFID(&op_data->op_fid2));
1962                         lmv_obj_put(obj);
1963                 }
1964                 goto request;
1965         }
1966
1967 repeat:
1968         ++loop;
1969         LASSERT(loop <= 2);
1970         obj = lmv_obj_grab(obd, &op_data->op_fid1);
1971         if (obj) {
1972                 /*
1973                  * directory is already split, so we have to forward request to
1974                  * the right MDS.
1975                  */
1976                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1977                                        (char *)old, oldlen);
1978                 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1979                 mds1 = obj->lo_inodes[mea_idx].li_mds;
1980                 CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
1981                 lmv_obj_put(obj);
1982         } else {
1983                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
1984                 if (rc)
1985                         RETURN(rc);
1986         }
1987
1988         obj = lmv_obj_grab(obd, &op_data->op_fid2);
1989         if (obj) {
1990                 /*
1991                  * Directory is already split, so we have to forward request to
1992                  * the right MDS.
1993                  */
1994                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1995                                        (char *)new, newlen);
1996
1997                 mds2 = obj->lo_inodes[mea_idx].li_mds;
1998                 op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1999                 CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2000                 lmv_obj_put(obj);
2001         } else {
2002                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2003                 if (rc)
2004                         RETURN(rc);
2005         }
2006
2007 request:
2008         op_data->op_fsuid = current->fsuid;
2009         op_data->op_fsgid = current->fsgid;
2010         op_data->op_cap   = current->cap_effective;
2011
2012         src_exp = lmv_get_export(lmv, mds1);
2013         tgt_exp = lmv_get_export(lmv, mds2);
2014         if (oldlen) {
2015                 /* LOOKUP lock on src child (fid3) should also be cancelled for
2016                  * src_exp in mdc_rename. */
2017                 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2018
2019                 /* Cancel UPDATE locks on tgt parent (fid2), tgt_exp is its
2020                  * own export. */
2021                 rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data, LCK_EX,
2022                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2023
2024                 /* Cancel LOOKUP locks on tgt child (fid4) for parent tgt_exp.*/
2025                 if (rc == 0)
2026                         rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data,
2027                                               LCK_EX, MDS_INODELOCK_LOOKUP,
2028                                               MF_MDC_CANCEL_FID4);
2029
2030                 /* XXX: the case when child is a striped dir is not supported.
2031                  * Only the master stripe has all locks cancelled early. */
2032                 /* Cancel all the locks on tgt child (fid4). */
2033                 if (rc == 0)
2034                         rc = lmv_early_cancel(lmv, src_exp, NULL, op_data,
2035                                               LCK_EX, MDS_INODELOCK_FULL,
2036                                               MF_MDC_CANCEL_FID4);
2037         }
2038
2039         if (rc == 0)
2040                 rc = md_rename(src_exp, op_data, old, oldlen,
2041                                new, newlen, request);
2042         if (rc == -ERESTART) {
2043                 LASSERT(*request != NULL);
2044                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2045                           "Got -ERESTART during rename!\n");
2046                 ptlrpc_req_finished(*request);
2047                 *request = NULL;
2048
2049                 /*
2050                  * Directory got split. Time to update local object and repeat
2051                  * the request with proper MDS.
2052                  */
2053                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2054                 if (rc == 0)
2055                         goto repeat;
2056         }
2057         RETURN(rc);
2058 }
2059
2060 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2061                        void *ea, int ealen, void *ea2, int ea2len,
2062                        struct ptlrpc_request **request,
2063                        struct md_open_data **mod)
2064 {
2065         struct obd_device *obd = exp->exp_obd;
2066         struct lmv_obd *lmv = &obd->u.lmv;
2067         struct ptlrpc_request *req;
2068         struct obd_export *tgt_exp;
2069         struct lmv_obj *obj;
2070         int rc = 0, i;
2071         ENTRY;
2072
2073         rc = lmv_check_connect(obd);
2074         if (rc)
2075                 RETURN(rc);
2076
2077         obj = lmv_obj_grab(obd, &op_data->op_fid1);
2078
2079         CDEBUG(D_OTHER, "SETATTR for "DFID", valid 0x%x%s\n",
2080                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2081                obj ? ", split" : "");
2082
2083         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2084         if (obj) {
2085                 for (i = 0; i < obj->lo_objcount; i++) {
2086                         op_data->op_fid1 = obj->lo_inodes[i].li_fid;
2087
2088                         tgt_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds);
2089                         if (IS_ERR(tgt_exp)) {
2090                                 rc = PTR_ERR(tgt_exp);
2091                                 break;
2092                         }
2093
2094                         rc = md_setattr(tgt_exp, op_data, ea, ealen,
2095                                         ea2, ea2len, &req, mod);
2096
2097                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_inodes[i].li_fid)) {
2098                                 /*
2099                                  * this is master object and this request should
2100                                  * be returned back to llite.
2101                                  */
2102                                 *request = req;
2103                         } else {
2104                                 ptlrpc_req_finished(req);
2105                         }
2106
2107                         if (rc)
2108                                 break;
2109                 }
2110                 lmv_obj_put(obj);
2111         } else {
2112                 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
2113                 if (IS_ERR(tgt_exp))
2114                         RETURN(PTR_ERR(tgt_exp));
2115
2116                 rc = md_setattr(tgt_exp, op_data, ea, ealen, ea2,
2117                                 ea2len, request, mod);
2118         }
2119         RETURN(rc);
2120 }
2121
2122 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2123                     struct obd_capa *oc, struct ptlrpc_request **request)
2124 {
2125         struct obd_device *obd = exp->exp_obd;
2126         struct lmv_obd *lmv = &obd->u.lmv;
2127         struct obd_export *tgt_exp;
2128         int rc;
2129         ENTRY;
2130
2131         rc = lmv_check_connect(obd);
2132         if (rc)
2133                 RETURN(rc);
2134
2135         tgt_exp = lmv_find_export(lmv, fid);
2136         if (IS_ERR(tgt_exp))
2137                 RETURN(PTR_ERR(tgt_exp));
2138
2139         rc = md_sync(tgt_exp, fid, oc, request);
2140         RETURN(rc);
2141 }
2142
2143 /* main purpose of LMV blocking ast is to remove split directory LMV
2144  * presentation object (struct lmv_obj) attached to the lock being revoked. */
2145 int lmv_blocking_ast(struct ldlm_lock *lock,
2146                      struct ldlm_lock_desc *desc,
2147                      void *data, int flag)
2148 {
2149         struct lustre_handle lockh;
2150         struct lmv_obj *obj;
2151         int rc;
2152         ENTRY;
2153
2154         switch (flag) {
2155         case LDLM_CB_BLOCKING:
2156                 ldlm_lock2handle(lock, &lockh);
2157                 rc = ldlm_cli_cancel(&lockh);
2158                 if (rc < 0) {
2159                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2160                         RETURN(rc);
2161                 }
2162                 break;
2163         case LDLM_CB_CANCELING:
2164                 /* time to drop cached attrs for dirobj */
2165                 obj = lock->l_ast_data;
2166                 if (obj) {
2167                         CDEBUG(D_OTHER, "cancel %s on "LPU64"/"LPU64
2168                                ", master "DFID"\n",
2169                                lock->l_resource->lr_name.name[3] == 1 ?
2170                                "LOOKUP" : "UPDATE",
2171                                lock->l_resource->lr_name.name[0],
2172                                lock->l_resource->lr_name.name[1],
2173                                PFID(&obj->lo_fid));
2174                         lmv_obj_put(obj);
2175                 }
2176                 break;
2177         default:
2178                 LBUG();
2179         }
2180         RETURN(0);
2181 }
2182
2183 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2184 {
2185         __u64 val;
2186
2187         val = le64_to_cpu(*hash);
2188         if (val < hash_adj)
2189                 val += MAX_HASH_SIZE;
2190         if (val != DIR_END_OFF)
2191                 *hash = cpu_to_le64(val - hash_adj);
2192 }
2193
2194 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2195 {
2196         __u64 id;
2197         struct obd_import *imp;
2198
2199         /*
2200          * XXX Hack: to get nid we assume that underlying obd device is mdc.
2201          */
2202         imp  = class_exp2cliimp(exp);
2203         id   = imp->imp_connection->c_self + fid_flatten(fid);
2204
2205         CDEBUG(D_INFO, "node rank: %llx "DFID" %llx %llx\n",
2206                imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2207
2208         return id ^ (id >> 32);
2209 }
2210
2211 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2212                         struct obd_capa *oc, __u64 offset64, struct page *page,
2213                         struct ptlrpc_request **request)
2214 {
2215         struct obd_device *obd = exp->exp_obd;
2216         struct lmv_obd *lmv = &obd->u.lmv;
2217         struct obd_export *tgt_exp;
2218         struct lu_fid rid = *fid;
2219         struct lmv_obj *obj;
2220         __u64 offset;
2221         __u64 hash_adj = 0;
2222         __u32 rank = 0;
2223         __u64 seg_size = 0;
2224         __u64 tgt_tmp = 0;
2225         int tgt = 0;
2226         int tgt0 = 0;
2227         int rc;
2228         int nr = 0;
2229         ENTRY;
2230
2231         offset = offset64;
2232
2233         rc = lmv_check_connect(obd);
2234         if (rc)
2235                 RETURN(rc);
2236
2237         CDEBUG(D_INFO, "READPAGE at %llx from "DFID"\n", offset, PFID(&rid));
2238
2239         obj = lmv_obj_grab(obd, fid);
2240         if (obj) {
2241
2242         /*
2243          * This case handle directory lookup in clustered metadata case (i.e.
2244          * split directory is located on multiple md servers.)
2245          * each server keeps directory entries for certain range of hashes.
2246          * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2247          * first server will keep records with hashes [ 0 ... MAX_HASH / N  - 1],
2248          * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2249          * so on....
2250          *      readdir can simply start reading entries from 0 - N server in
2251          * order but that will not scale well as all client will request dir in
2252          * to server in same order.
2253          * Following algorithm does optimization:
2254          * Instead of doing readdir in 1, 2, ...., N order, client with a
2255          * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2256          * (every client has rank R)
2257          *      But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2258          * since client ask dir from MDS{R} client has pages with offsets
2259          * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2260          * on hash  values that we get.
2261          */
2262
2263                 struct lmv_inode *loi;
2264
2265                 lmv_obj_lock(obj);
2266
2267                 nr       = obj->lo_objcount;
2268                 LASSERT(nr > 0);
2269                 seg_size = MAX_HASH_SIZE;
2270                 do_div(seg_size, nr);
2271                 loi      = obj->lo_inodes;
2272                 rank     = lmv_node_rank(lmv_get_export(lmv, loi[0].li_mds),
2273                                          fid) % nr;
2274                 tgt_tmp = offset;
2275                 do_div(tgt_tmp, seg_size);
2276                 tgt0     = do_div(tgt_tmp,  nr);
2277                 tgt      = (tgt0 + rank) % nr;
2278
2279                 if (tgt < tgt0)
2280                         /*
2281                          * Wrap around.
2282                          *
2283                          * Last segment has unusual length due to division
2284                          * rounding.
2285                          */
2286                         hash_adj = MAX_HASH_SIZE - seg_size * nr;
2287                 else
2288                         hash_adj = 0;
2289
2290                 hash_adj += rank * seg_size;
2291
2292                 CDEBUG(D_INFO, "hash_adj: %x %llx %llx/%x -> %llx/%x\n",
2293                        rank, hash_adj, offset, tgt0, offset + hash_adj, tgt);
2294
2295                 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2296                 rid = obj->lo_inodes[tgt].li_fid;
2297                 tgt_exp = lmv_get_export(lmv, loi[tgt].li_mds);
2298
2299                 CDEBUG(D_INFO, "forward to "DFID" with offset %lu i %d\n",
2300                        PFID(&rid), (unsigned long)offset, tgt);
2301         } else
2302                 tgt_exp = lmv_find_export(lmv, &rid);
2303
2304         if (IS_ERR(tgt_exp))
2305                 GOTO(cleanup, rc = PTR_ERR(tgt_exp));
2306
2307         rc = md_readpage(tgt_exp, &rid, oc, offset, page, request);
2308         if (rc)
2309                 GOTO(cleanup, rc);
2310         if (obj) {
2311                 struct lu_dirpage *dp;
2312                 struct lu_dirent  *ent;
2313
2314                 dp = cfs_kmap(page);
2315
2316                 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2317                 lmv_hash_adjust(&dp->ldp_hash_end,   hash_adj);
2318                 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2319
2320                 for (ent = lu_dirent_start(dp); ent != NULL;
2321                      ent = lu_dirent_next(ent))
2322                         lmv_hash_adjust(&ent->lde_hash, hash_adj);
2323
2324                 if (tgt0 != nr - 1) {
2325                         __u64 end;
2326
2327                         end = le64_to_cpu(dp->ldp_hash_end);
2328                         if (end == DIR_END_OFF) {
2329                                 dp->ldp_hash_end = cpu_to_le32(seg_size *
2330                                                                (tgt0 + 1));
2331                                 CDEBUG(D_INFO, ""DFID" reset end %llx tgt %d\n",
2332                                        PFID(&rid),
2333                                        le64_to_cpu(dp->ldp_hash_end), tgt);
2334                         }
2335                 }
2336                 cfs_kunmap(page);
2337         }
2338         /*
2339          * Here we could remove "." and ".." from all pages which at not from
2340          * master. But MDS has only "." and ".." for master dir.
2341          */
2342         EXIT;
2343 cleanup:
2344         if (obj) {
2345                 lmv_obj_unlock(obj);
2346                 lmv_obj_put(obj);
2347         }
2348         return rc;
2349 }
2350
2351 static int lmv_unlink_slaves(struct obd_export *exp,
2352                              struct md_op_data *op_data,
2353                              struct ptlrpc_request **req)
2354 {
2355         struct obd_device *obd = exp->exp_obd;
2356         struct lmv_obd *lmv = &obd->u.lmv;
2357         struct lmv_stripe_md *mea = op_data->op_mea1;
2358         struct md_op_data *op_data2;
2359         struct obd_export *tgt_exp;
2360         int i, rc = 0;
2361         ENTRY;
2362
2363         OBD_ALLOC_PTR(op_data2);
2364         if (op_data2 == NULL)
2365                 RETURN(-ENOMEM);
2366
2367         op_data2->op_mode = S_IFDIR;
2368         op_data2->op_fsuid = current->fsuid;
2369         op_data2->op_fsgid = current->fsgid;
2370         op_data2->op_bias = 0;
2371
2372         LASSERT(mea != NULL);
2373         for (i = 0; i < mea->mea_count; i++) {
2374                 memset(op_data2, 0, sizeof(*op_data2));
2375                 op_data2->op_fid1 = mea->mea_ids[i];
2376                 tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1);
2377                 if (IS_ERR(tgt_exp))
2378                         GOTO(out_free_op_data2, rc = PTR_ERR(tgt_exp));
2379
2380                 if (tgt_exp == NULL)
2381                         continue;
2382
2383                 rc = md_unlink(tgt_exp, op_data2, req);
2384
2385                 CDEBUG(D_OTHER, "unlink slave "DFID" -> %d\n",
2386                        PFID(&mea->mea_ids[i]), rc);
2387
2388                 if (*req) {
2389                         ptlrpc_req_finished(*req);
2390                         *req = NULL;
2391                 }
2392                 if (rc)
2393                         GOTO(out_free_op_data2, rc);
2394         }
2395
2396         EXIT;
2397 out_free_op_data2:
2398         OBD_FREE_PTR(op_data2);
2399         return rc;
2400 }
2401
2402 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2403                       struct ptlrpc_request **request)
2404 {
2405         struct obd_device *obd = exp->exp_obd;
2406         struct lmv_obd *lmv = &obd->u.lmv;
2407         struct obd_export *tgt_exp = NULL;
2408         struct lmv_obj *obj;
2409         int rc, loop = 0;
2410         ENTRY;
2411
2412         rc = lmv_check_connect(obd);
2413         if (rc)
2414                 RETURN(rc);
2415
2416         if (op_data->op_namelen == 0 && op_data->op_mea1 != NULL) {
2417                 /* mds asks to remove slave objects */
2418                 rc = lmv_unlink_slaves(exp, op_data, request);
2419                 RETURN(rc);
2420         }
2421
2422 repeat:
2423         ++loop;
2424         LASSERT(loop <= 2);
2425         if (op_data->op_namelen != 0) {
2426                 int mea_idx;
2427
2428                 obj = lmv_obj_grab(obd, &op_data->op_fid1);
2429                 if (obj) {
2430                         mea_idx = raw_name2idx(obj->lo_hashtype,
2431                                                obj->lo_objcount,
2432                                                op_data->op_name,
2433                                                op_data->op_namelen);
2434                         op_data->op_bias &= ~MDS_CHECK_SPLIT;
2435                         op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
2436                         tgt_exp = lmv_get_export(lmv,
2437                                                  obj->lo_inodes[mea_idx].li_mds);
2438                         lmv_obj_put(obj);
2439                         CDEBUG(D_OTHER, "unlink '%*s' in "DFID" -> %u\n",
2440                                op_data->op_namelen, op_data->op_name,
2441                                PFID(&op_data->op_fid1), mea_idx);
2442                 }
2443         } else {
2444                 CDEBUG(D_OTHER, "drop i_nlink on "DFID"\n",
2445                        PFID(&op_data->op_fid1));
2446         }
2447         if (tgt_exp == NULL) {
2448                 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
2449                 if (IS_ERR(tgt_exp))
2450                         RETURN(PTR_ERR(tgt_exp));
2451                 op_data->op_bias |= MDS_CHECK_SPLIT;
2452         }
2453
2454         op_data->op_fsuid = current->fsuid;
2455         op_data->op_fsgid = current->fsgid;
2456         op_data->op_cap   = current->cap_effective;
2457
2458         /* If child's fid is given, cancel unused locks for it if it is from
2459          * another export than parent. */
2460         if (op_data->op_namelen) {
2461                 /* LOOKUP lock for child (fid3) should also be cancelled on
2462                  * parent tgt_exp in mdc_unlink(). */
2463                 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2464
2465                 /* XXX: the case when child is a striped dir is not supported.
2466                  * Only the master stripe has all locks cancelled early. */
2467                 /* Cancel FULL locks on child (fid3). */
2468                 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX,
2469                                       MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2470         }
2471         if (rc == 0)
2472                 rc = md_unlink(tgt_exp, op_data, request);
2473         if (rc == -ERESTART) {
2474                 LASSERT(*request != NULL);
2475                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2476                           "Got -ERESTART during unlink!\n");
2477                 ptlrpc_req_finished(*request);
2478                 *request = NULL;
2479
2480                 /*
2481                  * Directory got split. Time to update local object and repeat
2482                  * the request with proper MDS.
2483                  */
2484                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2485                 if (rc == 0)
2486                         goto repeat;
2487         }
2488         RETURN(rc);
2489 }
2490
2491 static int lmv_llog_init(struct obd_device *obd, struct obd_llog_group *olg,
2492                          struct obd_device *tgt, int count,
2493                          struct llog_catid *logid, struct obd_uuid *uuid)
2494 {
2495 #if 0
2496         struct llog_ctxt *ctxt;
2497         int rc;
2498         ENTRY;
2499
2500         LASSERT(group == OBD_LLOG_GROUP);
2501         rc = llog_setup(obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
2502                         &llog_client_ops);
2503         if (rc == 0) {
2504                 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_CONFIG_REPL_CTXT);
2505                 llog_initiator_connect(ctxt, tgt);
2506                 llog_ctxt_put(ctxt);
2507         }
2508         RETURN(rc);
2509 #else
2510         return 0;
2511 #endif
2512 }
2513
2514 static int lmv_llog_finish(struct obd_device *obd, int count)
2515 {
2516         struct llog_ctxt *ctxt;
2517         int rc = 0;
2518         ENTRY;
2519
2520         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
2521         if (ctxt)
2522                 rc = llog_cleanup(ctxt);
2523
2524         RETURN(rc);
2525 }
2526
2527 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2528 {
2529         int rc = 0;
2530
2531         switch (stage) {
2532         case OBD_CLEANUP_EARLY:
2533                 /* XXX: here should be calling obd_precleanup() down to
2534                  * stack. */
2535                 break;
2536         case OBD_CLEANUP_EXPORTS:
2537                 rc = obd_llog_finish(obd, 0);
2538                 if (rc != 0)
2539                         CERROR("failed to cleanup llogging subsystems\n");
2540                 break;
2541         default:
2542                 break;
2543         }
2544         RETURN(rc);
2545 }
2546
2547 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2548                         void *key, __u32 *vallen, void *val)
2549 {
2550         struct obd_device *obd;
2551         struct lmv_obd *lmv;
2552         int rc = 0;
2553         ENTRY;
2554
2555         obd = class_exp2obd(exp);
2556         if (obd == NULL) {
2557                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2558                        exp->exp_handle.h_cookie);
2559                 RETURN(-EINVAL);
2560         }
2561
2562         lmv = &obd->u.lmv;
2563         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2564                 struct lmv_tgt_desc *tgts;
2565                 int i;
2566
2567                 rc = lmv_check_connect(obd);
2568                 if (rc)
2569                         RETURN(rc);
2570
2571                 LASSERT(*vallen == sizeof(__u32));
2572                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2573                      i++, tgts++) {
2574
2575                         /* all tgts should be connected when this get called. */
2576                         if (!tgts || !tgts->ltd_exp) {
2577                                 CERROR("target not setup?\n");
2578                                 continue;
2579                         }
2580
2581                         if (!obd_get_info(tgts->ltd_exp, keylen, key,
2582                                           vallen, val))
2583                                 RETURN(0);
2584                 }
2585                 RETURN(-EINVAL);
2586         } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2587                 rc = lmv_check_connect(obd);
2588                 if (rc)
2589                         RETURN(rc);
2590
2591                 /* forwarding this request to first MDS, it should know LOV
2592                  * desc. */
2593                 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2594                                   vallen, val);
2595                 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2596                         exp->exp_connect_flags =
2597                         ((struct obd_connect_data *)val)->ocd_connect_flags;
2598                 }
2599                 RETURN(rc);
2600         }
2601
2602         CDEBUG(D_IOCTL, "invalid key\n");
2603         RETURN(-EINVAL);
2604 }
2605
2606 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2607                        void *key, obd_count vallen, void *val,
2608                        struct ptlrpc_request_set *set)
2609 {
2610         struct lmv_tgt_desc    *tgt;
2611         struct obd_device      *obd;
2612         struct lmv_obd         *lmv;
2613         int rc = 0;
2614         ENTRY;
2615
2616         obd = class_exp2obd(exp);
2617         if (obd == NULL) {
2618                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2619                        exp->exp_handle.h_cookie);
2620                 RETURN(-EINVAL);
2621         }
2622         lmv = &obd->u.lmv;
2623
2624         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2625             KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2626                 int i, err = 0;
2627
2628                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2629                         tgt = &lmv->tgts[i];
2630
2631                         if (!tgt->ltd_exp)
2632                                 continue;
2633
2634                         err = obd_set_info_async(tgt->ltd_exp,
2635                                                  keylen, key, vallen, val, set);
2636                         if (err && rc == 0)
2637                                 rc = err;
2638                 }
2639
2640                 RETURN(rc);
2641         }
2642
2643         RETURN(-EINVAL);
2644 }
2645
2646 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2647                struct lov_stripe_md *lsm)
2648 {
2649         struct obd_device *obd = class_exp2obd(exp);
2650         struct lmv_obd *lmv = &obd->u.lmv;
2651         struct lmv_stripe_md *meap, *lsmp;
2652         int mea_size, i;
2653         ENTRY;
2654
2655         mea_size = lmv_get_easize(lmv);
2656         if (!lmmp)
2657                 RETURN(mea_size);
2658
2659         if (*lmmp && !lsm) {
2660                 OBD_FREE(*lmmp, mea_size);
2661                 *lmmp = NULL;
2662                 RETURN(0);
2663         }
2664
2665         if (*lmmp == NULL) {
2666                 OBD_ALLOC(*lmmp, mea_size);
2667                 if (*lmmp == NULL)
2668                         RETURN(-ENOMEM);
2669         }
2670
2671         if (!lsm)
2672                 RETURN(mea_size);
2673
2674         lsmp = (struct lmv_stripe_md *)lsm;
2675         meap = (struct lmv_stripe_md *)*lmmp;
2676
2677         if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2678             lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2679                 RETURN(-EINVAL);
2680
2681         meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2682         meap->mea_count = cpu_to_le32(lsmp->mea_count);
2683         meap->mea_master = cpu_to_le32(lsmp->mea_master);
2684
2685         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2686                 meap->mea_ids[i] = meap->mea_ids[i];
2687                 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2688         }
2689
2690         RETURN(mea_size);
2691 }
2692
2693 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2694                  struct lov_mds_md *lmm, int lmm_size)
2695 {
2696         struct obd_device *obd = class_exp2obd(exp);
2697         struct lmv_stripe_md **tmea = (struct lmv_stripe_md **)lsmp;
2698         struct lmv_stripe_md *mea = (struct lmv_stripe_md *)lmm;
2699         struct lmv_obd *lmv = &obd->u.lmv;
2700         int mea_size, i;
2701         __u32 magic;
2702         ENTRY;
2703
2704         mea_size = lmv_get_easize(lmv);
2705         if (lsmp == NULL)
2706                 return mea_size;
2707
2708         if (*lsmp != NULL && lmm == NULL) {
2709                 OBD_FREE(*tmea, mea_size);
2710                 *lsmp = NULL;
2711                 RETURN(0);
2712         }
2713
2714         LASSERT(mea_size == lmm_size);
2715
2716         OBD_ALLOC(*tmea, mea_size);
2717         if (*tmea == NULL)
2718                 RETURN(-ENOMEM);
2719
2720         if (!lmm)
2721                 RETURN(mea_size);
2722
2723         if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2724             mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2725             mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2726         {
2727                 magic = le32_to_cpu(mea->mea_magic);
2728         } else {
2729                 /* old mea is not handled here */
2730                 LBUG();
2731         }
2732
2733         (*tmea)->mea_magic = magic;
2734         (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2735         (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2736
2737         for (i = 0; i < (*tmea)->mea_count; i++) {
2738                 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2739                 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2740         }
2741         RETURN(mea_size);
2742 }
2743
2744 static int lmv_cancel_unused(struct obd_export *exp,
2745                              const struct lu_fid *fid,
2746                              ldlm_policy_data_t *policy,
2747                              ldlm_mode_t mode, int flags, void *opaque)
2748 {
2749         struct obd_device *obd = exp->exp_obd;
2750         struct lmv_obd *lmv = &obd->u.lmv;
2751         int rc = 0, err, i;
2752         ENTRY;
2753
2754         LASSERT(fid != NULL);
2755
2756         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2757                 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2758                         continue;
2759
2760                 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2761                                        policy, mode, flags, opaque);
2762                 if (!rc)
2763                         rc = err;
2764         }
2765         RETURN(rc);
2766 }
2767
2768 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data)
2769 {
2770         struct obd_device *obd = exp->exp_obd;
2771         struct lmv_obd *lmv = &obd->u.lmv;
2772
2773         ENTRY;
2774         RETURN(md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data));
2775 }
2776
2777 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2778                            const struct lu_fid *fid, ldlm_type_t type,
2779                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
2780                            struct lustre_handle *lockh)
2781 {
2782         struct obd_device *obd = exp->exp_obd;
2783         struct lmv_obd *lmv = &obd->u.lmv;
2784         ldlm_mode_t rc;
2785         int i;
2786         ENTRY;
2787
2788         CDEBUG(D_OTHER, "lock match for "DFID"\n", PFID(fid));
2789
2790         /* with CMD every object can have two locks in different namespaces:
2791          * lookup lock in space of mds storing direntry and update/open lock in
2792          * space of mds storing inode. Thus we check all targets, not only that
2793          * one fid was created in. */
2794         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2795                 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2796                                    type, policy, mode, lockh);
2797                 if (rc)
2798                         RETURN(rc);
2799         }
2800
2801         RETURN(0);
2802 }
2803
2804 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2805                       struct obd_export *dt_exp, struct obd_export *md_exp,
2806                       struct lustre_md *md)
2807 {
2808         struct obd_device *obd = exp->exp_obd;
2809         struct lmv_obd *lmv = &obd->u.lmv;
2810         int rc;
2811
2812         ENTRY;
2813         rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2814         RETURN(rc);
2815 }
2816
2817 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2818 {
2819         struct obd_device *obd = exp->exp_obd;
2820         struct lmv_obd *lmv = &obd->u.lmv;
2821
2822         ENTRY;
2823
2824         /* XXX LOV STACKING */
2825         if (md->mea)
2826                 obd_free_memmd(exp, (void *)&md->mea);
2827         RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2828 }
2829
2830 int lmv_set_open_replay_data(struct obd_export *exp,
2831                              struct obd_client_handle *och,
2832                              struct ptlrpc_request *open_req)
2833 {
2834         struct obd_device *obd = exp->exp_obd;
2835         struct lmv_obd *lmv = &obd->u.lmv;
2836         struct obd_export *tgt_exp;
2837
2838         ENTRY;
2839
2840         tgt_exp = lmv_find_export(lmv, &och->och_fid);
2841         if (IS_ERR(tgt_exp))
2842                 RETURN(PTR_ERR(tgt_exp));
2843
2844         RETURN(md_set_open_replay_data(tgt_exp, och, open_req));
2845 }
2846
2847 int lmv_clear_open_replay_data(struct obd_export *exp,
2848                                struct obd_client_handle *och)
2849 {
2850         struct obd_device *obd = exp->exp_obd;
2851         struct lmv_obd *lmv = &obd->u.lmv;
2852         struct obd_export *tgt_exp;
2853         ENTRY;
2854
2855         tgt_exp = lmv_find_export(lmv, &och->och_fid);
2856         if (IS_ERR(tgt_exp))
2857                 RETURN(PTR_ERR(tgt_exp));
2858
2859         RETURN(md_clear_open_replay_data(tgt_exp, och));
2860 }
2861
2862 static int lmv_get_remote_perm(struct obd_export *exp,
2863                                const struct lu_fid *fid,
2864                                struct obd_capa *oc, __u32 suppgid,
2865                                struct ptlrpc_request **request)
2866 {
2867         struct obd_device *obd = exp->exp_obd;
2868         struct lmv_obd *lmv = &obd->u.lmv;
2869         struct obd_export *tgt_exp;
2870         int rc;
2871
2872         ENTRY;
2873
2874         rc = lmv_check_connect(obd);
2875         if (rc)
2876                 RETURN(rc);
2877
2878         tgt_exp = lmv_find_export(lmv, fid);
2879         if (IS_ERR(tgt_exp))
2880                 RETURN(PTR_ERR(tgt_exp));
2881
2882         rc = md_get_remote_perm(tgt_exp, fid, oc, suppgid, request);
2883
2884         RETURN(rc);
2885 }
2886
2887 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2888                           renew_capa_cb_t cb)
2889 {
2890         struct obd_device *obd = exp->exp_obd;
2891         struct lmv_obd *lmv = &obd->u.lmv;
2892         struct obd_export *tgt_exp;
2893         int rc;
2894         ENTRY;
2895
2896         rc = lmv_check_connect(obd);
2897         if (rc)
2898                 RETURN(rc);
2899
2900         tgt_exp = lmv_find_export(lmv, &oc->c_capa.lc_fid);
2901         if (IS_ERR(tgt_exp))
2902                 RETURN(PTR_ERR(tgt_exp));
2903
2904         rc = md_renew_capa(tgt_exp, oc, cb);
2905         RETURN(rc);
2906 }
2907
2908 int lmv_intent_getattr_async(struct obd_export *exp,
2909                              struct md_enqueue_info *minfo,
2910                              struct ldlm_enqueue_info *einfo)
2911 {
2912         struct obd_device *obd = exp->exp_obd;
2913         struct lmv_obd    *lmv = &obd->u.lmv;
2914         struct obd_export *tgt_exp;
2915         int rc;
2916         ENTRY;
2917
2918         rc = lmv_check_connect(obd);
2919         if (rc)
2920                 RETURN(rc);
2921
2922         if (fid_is_zero(&minfo->mi_data.op_fid2))
2923                 tgt_exp = lmv_find_export(lmv, &minfo->mi_data.op_fid1);
2924         else
2925                 tgt_exp = lmv_find_export(lmv, &minfo->mi_data.op_fid2);
2926         if (IS_ERR(tgt_exp))
2927                 RETURN(PTR_ERR(tgt_exp));
2928
2929         rc = md_intent_getattr_async(tgt_exp, minfo, einfo);
2930         RETURN(rc);
2931 }
2932
2933 int lmv_revalidate_lock(struct obd_export *exp,
2934                         struct lookup_intent *it,
2935                         struct lu_fid *fid)
2936 {
2937         struct obd_device *obd = exp->exp_obd;
2938         struct lmv_obd    *lmv = &obd->u.lmv;
2939         struct obd_export *tgt_exp;
2940         int rc;
2941         ENTRY;
2942
2943         rc = lmv_check_connect(obd);
2944         if (rc)
2945                 RETURN(rc);
2946
2947         tgt_exp = lmv_find_export(lmv, fid);
2948         if (IS_ERR(tgt_exp))
2949                 RETURN(PTR_ERR(tgt_exp));
2950
2951         rc = md_revalidate_lock(tgt_exp, it, fid);
2952         RETURN(rc);
2953 }
2954
2955
2956 struct obd_ops lmv_obd_ops = {
2957         .o_owner                = THIS_MODULE,
2958         .o_setup                = lmv_setup,
2959         .o_cleanup              = lmv_cleanup,
2960         .o_precleanup           = lmv_precleanup,
2961         .o_process_config       = lmv_process_config,
2962         .o_connect              = lmv_connect,
2963         .o_disconnect           = lmv_disconnect,
2964         .o_statfs               = lmv_statfs,
2965         .o_llog_init            = lmv_llog_init,
2966         .o_llog_finish          = lmv_llog_finish,
2967         .o_get_info             = lmv_get_info,
2968         .o_set_info_async       = lmv_set_info_async,
2969         .o_packmd               = lmv_packmd,
2970         .o_unpackmd             = lmv_unpackmd,
2971         .o_notify               = lmv_notify,
2972         .o_iocontrol            = lmv_iocontrol,
2973         .o_fid_delete           = lmv_fid_delete
2974 };
2975
2976 struct md_ops lmv_md_ops = {
2977         .m_getstatus            = lmv_getstatus,
2978         .m_change_cbdata        = lmv_change_cbdata,
2979         .m_close                = lmv_close,
2980         .m_create               = lmv_create,
2981         .m_done_writing         = lmv_done_writing,
2982         .m_enqueue              = lmv_enqueue,
2983         .m_getattr              = lmv_getattr,
2984         .m_getxattr             = lmv_getxattr,
2985         .m_getattr_name         = lmv_getattr_name,
2986         .m_intent_lock          = lmv_intent_lock,
2987         .m_link                 = lmv_link,
2988         .m_rename               = lmv_rename,
2989         .m_setattr              = lmv_setattr,
2990         .m_setxattr             = lmv_setxattr,
2991         .m_sync                 = lmv_sync,
2992         .m_readpage             = lmv_readpage,
2993         .m_unlink               = lmv_unlink,
2994         .m_init_ea_size         = lmv_init_ea_size,
2995         .m_cancel_unused        = lmv_cancel_unused,
2996         .m_set_lock_data        = lmv_set_lock_data,
2997         .m_lock_match           = lmv_lock_match,
2998         .m_get_lustre_md        = lmv_get_lustre_md,
2999         .m_free_lustre_md       = lmv_free_lustre_md,
3000         .m_set_open_replay_data = lmv_set_open_replay_data,
3001         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3002         .m_renew_capa           = lmv_renew_capa,
3003         .m_get_remote_perm      = lmv_get_remote_perm,
3004         .m_intent_getattr_async = lmv_intent_getattr_async,
3005         .m_revalidate_lock      = lmv_revalidate_lock
3006 };
3007
3008 int __init lmv_init(void)
3009 {
3010         struct lprocfs_static_vars lvars;
3011         int rc;
3012
3013         obj_cache = cfs_mem_cache_create("lmv_objects",
3014                                       sizeof(struct lmv_obj),
3015                                       0, 0);
3016         if (!obj_cache) {
3017                 CERROR("error allocating lmv objects cache\n");
3018                 return -ENOMEM;
3019         }
3020
3021         lprocfs_lmv_init_vars(&lvars);
3022         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3023                                  lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3024         if (rc)
3025                 cfs_mem_cache_destroy(obj_cache);
3026
3027         return rc;
3028 }
3029
3030 #ifdef __KERNEL__
3031 static void lmv_exit(void)
3032 {
3033         int rc;
3034
3035         class_unregister_type(LUSTRE_LMV_NAME);
3036
3037         rc = cfs_mem_cache_destroy(obj_cache);
3038         LASSERTF(rc == 0,
3039                  "can't free lmv objects cache, %d object(s)"
3040                  "still in use\n", atomic_read(&obj_cache_count));
3041 }
3042
3043 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3044 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3045 MODULE_LICENSE("GPL");
3046
3047 module_init(lmv_init);
3048 module_exit(lmv_exit);
3049 #endif