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, strlen("inter_mds"),
 301                                    "inter_mds", 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_LLOG_GROUP, 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         obd_free_memmd(exp, (struct lov_stripe_md **)&md.mea);
1356
1357         EXIT;
1358 cleanup:
1359         if (req)
1360                 ptlrpc_req_finished(req);
1361         return rc;
1362 }
1363
1364 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1365                const void *data, int datalen, int mode, __u32 uid,
1366                __u32 gid, __u32 cap_effective,  __u64 rdev,
1367                struct ptlrpc_request **request)
1368 {
1369         struct obd_device *obd = exp->exp_obd;
1370         struct lmv_obd *lmv = &obd->u.lmv;
1371         struct obd_export *tgt_exp;
1372         struct lmv_obj *obj;
1373         int rc, loop = 0;
1374         ENTRY;
1375
1376         rc = lmv_check_connect(obd);
1377         if (rc)
1378                 RETURN(rc);
1379
1380         if (!lmv->desc.ld_active_tgt_count)
1381                 RETURN(-EIO);
1382 repeat:
1383         ++loop;
1384         LASSERT(loop <= 2);
1385         obj = lmv_obj_grab(obd, &op_data->op_fid1);
1386         if (obj) {
1387                 int mea_idx;
1388
1389                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1390                                        op_data->op_name, op_data->op_namelen);
1391                 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1392                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1393                 op_data->op_mds = obj->lo_inodes[mea_idx].li_mds;
1394                 tgt_exp = lmv_get_export(lmv, op_data->op_mds);
1395                 lmv_obj_put(obj);
1396         } else {
1397                 struct lmv_tgt_desc *tgt;
1398
1399                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1400                 op_data->op_bias |= MDS_CHECK_SPLIT;
1401                 op_data->op_mds = tgt->ltd_idx;
1402                 tgt_exp = tgt->ltd_exp;
1403         }
1404
1405         if (IS_ERR(tgt_exp))
1406                 RETURN(PTR_ERR(tgt_exp));
1407
1408         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1409         if (rc == -ERESTART)
1410                 goto repeat;
1411         else if (rc)
1412                 RETURN(rc);
1413
1414         CDEBUG(D_OTHER, "CREATE '%*s' on "DFID"\n", op_data->op_namelen,
1415                op_data->op_name, PFID(&op_data->op_fid1));
1416
1417         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1418         rc = md_create(tgt_exp, op_data, data, datalen, mode, uid, gid,
1419                        cap_effective, rdev, request);
1420         if (rc == 0) {
1421                 if (*request == NULL)
1422                         RETURN(rc);
1423                 CDEBUG(D_OTHER, "created - "DFID"\n", PFID(&op_data->op_fid1));
1424         } else if (rc == -ERESTART) {
1425                 LASSERT(*request != NULL);
1426                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1427                           "Got -ERESTART during create!\n");
1428                 ptlrpc_req_finished(*request);
1429                 *request = NULL;
1430
1431                 /*
1432                  * Directory got split. Time to update local object and repeat
1433                  * the request with proper MDS.
1434                  */
1435                 rc = lmv_handle_split(exp, &op_data->op_fid1);
1436                 if (rc == 0) {
1437                         rc = lmv_alloc_slave_fids(obd, &op_data->op_fid1,
1438                                                   op_data, &op_data->op_fid2);
1439                         if (rc)
1440                                 RETURN(rc);
1441                         goto repeat;
1442                 }
1443         }
1444         RETURN(rc);
1445 }
1446
1447 static int lmv_done_writing(struct obd_export *exp,
1448                             struct md_op_data *op_data,
1449                             struct md_open_data *mod)
1450 {
1451         struct obd_device *obd = exp->exp_obd;
1452         struct lmv_obd *lmv = &obd->u.lmv;
1453         struct obd_export *tgt_exp;
1454         int rc;
1455         ENTRY;
1456
1457         rc = lmv_check_connect(obd);
1458         if (rc)
1459                 RETURN(rc);
1460
1461         tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1462         if (IS_ERR(tgt_exp))
1463                 RETURN(PTR_ERR(tgt_exp));
1464
1465         rc = md_done_writing(tgt_exp, op_data, mod);
1466         RETURN(rc);
1467 }
1468
1469 static int
1470 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1471                    struct lookup_intent *it, struct md_op_data *op_data,
1472                    struct lustre_handle *lockh, void *lmm, int lmmsize)
1473 {
1474         struct obd_device *obd = exp->exp_obd;
1475         struct lmv_obd *lmv = &obd->u.lmv;
1476         struct lmv_stripe_md *mea = op_data->op_mea1;
1477         struct md_op_data *op_data2;
1478         struct obd_export *tgt_exp;
1479         int i, rc = 0;
1480         ENTRY;
1481
1482         OBD_ALLOC_PTR(op_data2);
1483         if (op_data2 == NULL)
1484                 RETURN(-ENOMEM);
1485
1486         LASSERT(mea != NULL);
1487         for (i = 0; i < mea->mea_count; i++) {
1488                 memset(op_data2, 0, sizeof(*op_data2));
1489                 op_data2->op_fid1 = mea->mea_ids[i];
1490                 op_data2->op_bias = 0;
1491
1492                 tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1);
1493                 if (IS_ERR(tgt_exp))
1494                         GOTO(cleanup, rc = PTR_ERR(tgt_exp));
1495
1496                 if (tgt_exp == NULL)
1497                         continue;
1498
1499                 rc = md_enqueue(tgt_exp, einfo, it, op_data2,
1500                                 lockh + i, lmm, lmmsize, 0);
1501
1502                 CDEBUG(D_OTHER, "take lock on slave "DFID" -> %d/%d\n",
1503                        PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1504
1505                 if (rc)
1506                         GOTO(cleanup, rc);
1507
1508                 if (it->d.lustre.it_data) {
1509                         struct ptlrpc_request *req;
1510                         req = (struct ptlrpc_request *)it->d.lustre.it_data;
1511                         ptlrpc_req_finished(req);
1512                 }
1513
1514                 if (it->d.lustre.it_status)
1515                         GOTO(cleanup, rc = it->d.lustre.it_status);
1516         }
1517
1518         EXIT;
1519 cleanup:
1520         OBD_FREE_PTR(op_data2);
1521
1522         if (rc != 0) {
1523                 /* drop all taken locks */
1524                 while (--i >= 0) {
1525                         if (lockh[i].cookie)
1526                                 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1527                         lockh[i].cookie = 0;
1528                 }
1529         }
1530         return rc;
1531 }
1532
1533 static int
1534 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1535                    struct lookup_intent *it, struct md_op_data *op_data,
1536                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1537                    int extra_lock_flags)
1538 {
1539         struct ptlrpc_request *req = it->d.lustre.it_data;
1540         struct obd_device *obd = exp->exp_obd;
1541         struct lmv_obd *lmv = &obd->u.lmv;
1542         struct lustre_handle plock;
1543         struct obd_export *tgt_exp;
1544         struct md_op_data *rdata;
1545         struct lu_fid fid_copy;
1546         struct mdt_body *body;
1547         int rc = 0, pmode;
1548         ENTRY;
1549
1550         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1551         LASSERT(body != NULL);
1552
1553         if (!(body->valid & OBD_MD_MDS))
1554                 RETURN(0);
1555
1556         CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID" -> "DFID"\n",
1557                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1558
1559         /* We got LOOKUP lock, but we really need attrs */
1560         pmode = it->d.lustre.it_lock_mode;
1561         LASSERT(pmode != 0);
1562         memcpy(&plock, lockh, sizeof(plock));
1563         it->d.lustre.it_lock_mode = 0;
1564         it->d.lustre.it_data = NULL;
1565         fid_copy = body->fid1;
1566
1567         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1568         ptlrpc_req_finished(req);
1569
1570         tgt_exp = lmv_find_export(lmv, &fid_copy);
1571         if (IS_ERR(tgt_exp))
1572                 GOTO(out, rc = PTR_ERR(tgt_exp));
1573
1574         OBD_ALLOC_PTR(rdata);
1575         if (rdata == NULL)
1576                 GOTO(out, rc = -ENOMEM);
1577
1578         rdata->op_fid1 = fid_copy;
1579         rdata->op_bias = MDS_CROSS_REF;
1580
1581         rc = md_enqueue(tgt_exp, einfo, it, rdata, lockh,
1582                         lmm, lmmsize, extra_lock_flags);
1583         OBD_FREE_PTR(rdata);
1584         EXIT;
1585 out:
1586         ldlm_lock_decref(&plock, pmode);
1587         return rc;
1588 }
1589
1590 static int
1591 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1592             struct lookup_intent *it, struct md_op_data *op_data,
1593             struct lustre_handle *lockh, void *lmm, int lmmsize,
1594             int extra_lock_flags)
1595 {
1596         struct obd_device *obd = exp->exp_obd;
1597         struct lmv_obd *lmv = &obd->u.lmv;
1598         struct obd_export *tgt_exp = NULL;
1599         struct lmv_obj *obj;
1600         int rc;
1601         ENTRY;
1602
1603         rc = lmv_check_connect(obd);
1604         if (rc)
1605                 RETURN(rc);
1606
1607         if (op_data->op_mea1 && it->it_op == IT_UNLINK) {
1608                 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1609                                         lockh, lmm, lmmsize);
1610                 RETURN(rc);
1611         }
1612
1613         if (op_data->op_namelen) {
1614                 obj = lmv_obj_grab(obd, &op_data->op_fid1);
1615                 if (obj) {
1616                         int mea_idx;
1617
1618                         /* directory is split. look for right mds for this
1619                          * name */
1620                         mea_idx = raw_name2idx(obj->lo_hashtype,
1621                                                obj->lo_objcount,
1622                                                (char *)op_data->op_name,
1623                                                op_data->op_namelen);
1624                         op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1625                         tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds);
1626                         lmv_obj_put(obj);
1627                 }
1628         }
1629
1630         if (tgt_exp == NULL)
1631                 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
1632         if (IS_ERR(tgt_exp))
1633                 RETURN(PTR_ERR(tgt_exp));
1634
1635         CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID"\n", LL_IT2STR(it),
1636                PFID(&op_data->op_fid1));
1637
1638         rc = md_enqueue(tgt_exp, einfo, it, op_data, lockh,
1639                         lmm, lmmsize, extra_lock_flags);
1640
1641         if (rc == 0 && it->it_op == IT_OPEN)
1642                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1643                                         lmm, lmmsize, extra_lock_flags);
1644         RETURN(rc);
1645 }
1646
1647 static int
1648 lmv_getattr_name(struct obd_export *exp, const struct lu_fid *fid,
1649                  struct obd_capa *oc, const char *filename, int namelen,
1650                  obd_valid valid, int ea_size, __u32 suppgid,
1651                  struct ptlrpc_request **request)
1652 {
1653         struct obd_device *obd = exp->exp_obd;
1654         struct lmv_obd *lmv = &obd->u.lmv;
1655         struct lu_fid rid = *fid;
1656         struct obd_export *tgt_exp;
1657         struct mdt_body *body;
1658         struct lmv_obj *obj;
1659         int rc, loop = 0;
1660         ENTRY;
1661
1662         rc = lmv_check_connect(obd);
1663         if (rc)
1664                 RETURN(rc);
1665
1666 repeat:
1667         ++loop;
1668         LASSERT(loop <= 2);
1669         obj = lmv_obj_grab(obd, &rid);
1670         if (obj) {
1671                 int mea_idx;
1672
1673                 /* Directory is split. Look for right mds for this name */
1674                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1675                                        filename, namelen - 1);
1676                 rid = obj->lo_inodes[mea_idx].li_fid;
1677                 tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds);
1678                 lmv_obj_put(obj);
1679                 valid &= ~OBD_MD_FLCKSPLIT;
1680         } else {
1681                 tgt_exp = lmv_find_export(lmv, &rid);
1682                 valid |= OBD_MD_FLCKSPLIT;
1683         }
1684         if (IS_ERR(tgt_exp))
1685                 RETURN(PTR_ERR(tgt_exp));
1686
1687         CDEBUG(D_OTHER, "getattr_name for %*s on "DFID" -> "DFID"\n",
1688                namelen, filename, PFID(fid), PFID(&rid));
1689
1690         rc = md_getattr_name(tgt_exp, &rid, oc, filename, namelen, valid,
1691                              ea_size, suppgid, request);
1692         if (rc == 0) {
1693                 body = req_capsule_server_get(&(*request)->rq_pill,
1694                                               &RMF_MDT_BODY);
1695                 LASSERT(body != NULL);
1696
1697                 if (body->valid & OBD_MD_MDS) {
1698                         struct ptlrpc_request *req = NULL;
1699
1700                         rid = body->fid1;
1701                         CDEBUG(D_OTHER, "request attrs for "DFID"\n",
1702                                PFID(&rid));
1703
1704                         tgt_exp = lmv_find_export(lmv, &rid);
1705                         if (IS_ERR(tgt_exp)) {
1706                                 ptlrpc_req_finished(*request);
1707                                 RETURN(PTR_ERR(tgt_exp));
1708                         }
1709
1710                         rc = md_getattr_name(tgt_exp, &rid, NULL, NULL, 1,
1711                                              valid | OBD_MD_FLCROSSREF,
1712                                              ea_size, suppgid, &req);
1713                         ptlrpc_req_finished(*request);
1714                         *request = req;
1715                 }
1716         } else if (rc == -ERESTART) {
1717                 LASSERT(*request != NULL);
1718                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1719                           "Got -ERESTART during getattr!\n");
1720                 ptlrpc_req_finished(*request);
1721                 *request = NULL;
1722
1723                 /*
1724                  * Directory got split. Time to update local object and repeat
1725                  * the request with proper MDS.
1726                  */
1727                 rc = lmv_handle_split(exp, &rid);
1728                 if (rc == 0)
1729                         goto repeat;
1730         }
1731         RETURN(rc);
1732 }
1733
1734 #define md_op_data_fid(op_data, fl)                     \
1735         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1736          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1737          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1738          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1739          NULL)
1740
1741 /* @tgt_exp is the export the metadata request is sent.
1742  * @fid_exp is the export the cancel should be sent for the current fid.
1743  * if @fid_exp is NULL, the export is found for the current fid.
1744  * @op_data keeps the current fid, which is pointed through @flag.
1745  * @mode, @bits -- lock match parameters. */
1746 static int lmv_early_cancel(struct lmv_obd *lmv, struct obd_export *tgt_exp,
1747                             struct obd_export *fid_exp,
1748                             struct md_op_data *op_data,
1749                             ldlm_mode_t mode, int bits, int flag)
1750 {
1751         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1752         ldlm_policy_data_t policy = {{0}};
1753         int rc = 0;
1754         ENTRY;
1755
1756         if (!fid_is_sane(fid))
1757                 RETURN(0);
1758
1759         if (fid_exp == NULL)
1760                 fid_exp = lmv_find_export(lmv, fid);
1761
1762         if (tgt_exp == fid_exp) {
1763                 /* The export is the same as on the target server, cancel
1764                  * will be sent along with the main metadata operation. */
1765                 op_data->op_flags |= flag;
1766                 RETURN(0);
1767         }
1768
1769         policy.l_inodebits.bits = bits;
1770         rc = md_cancel_unused(fid_exp, fid, &policy, mode, LDLM_FL_ASYNC, NULL);
1771         RETURN(rc);
1772 }
1773
1774 #ifdef EARLY_CANCEL_FOR_STRIPED_DIR_IS_READY
1775 /* Check if the fid in @op_data pointed to by flag is of the same export(s)
1776  * as @tgt_exp. Early cancels will be sent later by mdc code, otherwise, call
1777  * md_cancel_unused for child export(s). */
1778 static int lmv_early_cancel_stripes(struct obd_export *exp,
1779                                     struct obd_export *tgt_exp,
1780                                     struct md_op_data *op_data,
1781                                     ldlm_mode_t mode, int bits, int flag)
1782 {
1783         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1784         struct obd_device *obd = exp->exp_obd;
1785         struct lmv_obd *lmv = &obd->u.lmv;
1786         struct obd_export *st_exp;
1787         struct lmv_obj *obj;
1788         int rc = 0;
1789         ENTRY;
1790
1791         if (!fid_is_sane(fid))
1792                 RETURN(0);
1793
1794         obj = lmv_obj_grab(obd, fid);
1795         if (obj) {
1796                 ldlm_policy_data_t policy = {{0}};
1797                 struct lu_fid *st_fid;
1798                 int i;
1799
1800                 policy.l_inodebits.bits = bits;
1801                 for (i = 0; i < obj->lo_objcount; i++) {
1802                         st_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds);
1803                         st_fid = &obj->lo_inodes[i].li_fid;
1804                         if (tgt_exp != st_exp) {
1805                                 rc = md_cancel_unused(st_exp, st_fid, &policy,
1806                                                       mode, LDLM_FL_ASYNC,
1807                                                       NULL);
1808                                 if (rc)
1809                                         break;
1810                         } else {
1811                                 /* Some export matches to @tgt_exp, do cancel
1812                                  * for its fid in mdc */
1813                                 *fid = *st_fid;
1814                                 op_data->op_flags |= flag;
1815                         }
1816                 }
1817                 lmv_obj_put(obj);
1818         } else {
1819                 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data,
1820                                       mode, bits, flag);
1821         }
1822         RETURN(rc);
1823 }
1824 #endif
1825
1826 /*
1827  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1828  * op_data->op_fid2
1829  */
1830 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1831                     struct ptlrpc_request **request)
1832 {
1833         struct obd_device *obd = exp->exp_obd;
1834         struct lmv_obd *lmv = &obd->u.lmv;
1835         struct obd_export *tgt_exp;
1836         struct lmv_obj *obj;
1837         int rc, loop = 0;
1838         mdsno_t mds;
1839         ENTRY;
1840
1841         rc = lmv_check_connect(obd);
1842         if (rc)
1843                 RETURN(rc);
1844
1845 repeat:
1846         ++loop;
1847         LASSERT(loop <= 2);
1848         if (op_data->op_namelen != 0) {
1849                 int mea_idx;
1850
1851                 /* Usual link request */
1852                 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1853                 if (obj) {
1854                         mea_idx = raw_name2idx(obj->lo_hashtype,
1855                                                obj->lo_objcount,
1856                                                op_data->op_name,
1857                                                op_data->op_namelen);
1858                         op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1859                         mds = obj->lo_inodes[mea_idx].li_mds;
1860                         lmv_obj_put(obj);
1861                 } else {
1862                         rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1863                         if (rc)
1864                                 RETURN(rc);
1865                 }
1866
1867                 CDEBUG(D_OTHER,"link "DFID":%*s to "DFID"\n",
1868                        PFID(&op_data->op_fid2), op_data->op_namelen,
1869                        op_data->op_name, PFID(&op_data->op_fid1));
1870         } else {
1871                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds);
1872                 if (rc)
1873                         RETURN(rc);
1874
1875                 /* request from MDS to acquire i_links for inode by fid1 */
1876                 CDEBUG(D_OTHER, "inc i_nlinks for "DFID"\n",
1877                        PFID(&op_data->op_fid1));
1878         }
1879
1880         CDEBUG(D_OTHER, "forward to MDS #"LPU64" ("DFID")\n",
1881                mds, PFID(&op_data->op_fid1));
1882
1883         op_data->op_fsuid = current->fsuid;
1884         op_data->op_fsgid = current->fsgid;
1885         op_data->op_cap   = current->cap_effective;
1886
1887         tgt_exp = lmv->tgts[mds].ltd_exp;
1888         if (op_data->op_namelen) {
1889                 op_data->op_flags |= MF_MDC_CANCEL_FID2;
1890                 /* Cancel UPDATE lock on child (fid1). */
1891                 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX,
1892                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1893         }
1894         if (rc == 0)
1895                 rc = md_link(tgt_exp, op_data, request);
1896         if (rc == -ERESTART) {
1897                 LASSERT(*request != NULL);
1898                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1899                           "Got -ERESTART during link!\n");
1900                 ptlrpc_req_finished(*request);
1901                 *request = NULL;
1902
1903                 /*
1904                  * Directory got split. Time to update local object and repeat
1905                  * the request with proper MDS.
1906                  */
1907                 rc = lmv_handle_split(exp, &op_data->op_fid2);
1908                 if (rc == 0)
1909                         goto repeat;
1910         }
1911
1912         RETURN(rc);
1913 }
1914
1915 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
1916                       const char *old, int oldlen, const char *new, int newlen,
1917                       struct ptlrpc_request **request)
1918 {
1919         struct obd_export *tgt_exp = NULL, *src_exp;
1920         struct obd_device *obd = exp->exp_obd;
1921         struct lmv_obd *lmv = &obd->u.lmv;
1922         int rc, mea_idx, loop = 0;
1923         struct lmv_obj *obj;
1924         mdsno_t mds1, mds2;
1925         ENTRY;
1926
1927         CDEBUG(D_OTHER, "rename %*s in "DFID" to %*s in "DFID"\n",
1928                oldlen, old, PFID(&op_data->op_fid1),
1929                newlen, new, PFID(&op_data->op_fid2));
1930
1931         rc = lmv_check_connect(obd);
1932         if (rc)
1933                 RETURN(rc);
1934
1935         if (oldlen == 0) {
1936                 /*
1937                  * MDS with old dir entry is asking another MDS to create name
1938                  * there.
1939                  */
1940                 CDEBUG(D_OTHER,
1941                        "create %*s(%d/%d) in "DFID" pointing "
1942                        "to "DFID"\n", newlen, new, oldlen, newlen,
1943                        PFID(&op_data->op_fid2), PFID(&op_data->op_fid1));
1944
1945                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds1);
1946                 if (rc)
1947                         RETURN(rc);
1948
1949                 /*
1950                  * Target directory can be split, sowe should forward request to
1951                  * the right MDS.
1952                  */
1953                 obj = lmv_obj_grab(obd, &op_data->op_fid2);
1954                 if (obj) {
1955                         mea_idx = raw_name2idx(obj->lo_hashtype,
1956                                                obj->lo_objcount,
1957                                                (char *)new, newlen);
1958                         op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1959                         CDEBUG(D_OTHER, "Parent obj "DFID"\n",
1960                                PFID(&op_data->op_fid2));
1961                         lmv_obj_put(obj);
1962                 }
1963                 goto request;
1964         }
1965
1966 repeat:
1967         ++loop;
1968         LASSERT(loop <= 2);
1969         obj = lmv_obj_grab(obd, &op_data->op_fid1);
1970         if (obj) {
1971                 /*
1972                  * directory is already split, so we have to forward request to
1973                  * the right MDS.
1974                  */
1975                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1976                                        (char *)old, oldlen);
1977                 op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
1978                 mds1 = obj->lo_inodes[mea_idx].li_mds;
1979                 CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
1980                 lmv_obj_put(obj);
1981         } else {
1982                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
1983                 if (rc)
1984                         RETURN(rc);
1985         }
1986
1987         obj = lmv_obj_grab(obd, &op_data->op_fid2);
1988         if (obj) {
1989                 /*
1990                  * Directory is already split, so we have to forward request to
1991                  * the right MDS.
1992                  */
1993                 mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1994                                        (char *)new, newlen);
1995
1996                 mds2 = obj->lo_inodes[mea_idx].li_mds;
1997                 op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid;
1998                 CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
1999                 lmv_obj_put(obj);
2000         } else {
2001                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2002                 if (rc)
2003                         RETURN(rc);
2004         }
2005
2006 request:
2007         op_data->op_fsuid = current->fsuid;
2008         op_data->op_fsgid = current->fsgid;
2009         op_data->op_cap   = current->cap_effective;
2010
2011         src_exp = lmv_get_export(lmv, mds1);
2012         tgt_exp = lmv_get_export(lmv, mds2);
2013         if (oldlen) {
2014                 /* LOOKUP lock on src child (fid3) should also be cancelled for
2015                  * src_exp in mdc_rename. */
2016                 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2017
2018                 /* Cancel UPDATE locks on tgt parent (fid2), tgt_exp is its
2019                  * own export. */
2020                 rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data, LCK_EX,
2021                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2022
2023                 /* Cancel LOOKUP locks on tgt child (fid4) for parent tgt_exp.*/
2024                 if (rc == 0)
2025                         rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data,
2026                                               LCK_EX, MDS_INODELOCK_LOOKUP,
2027                                               MF_MDC_CANCEL_FID4);
2028
2029                 /* XXX: the case when child is a striped dir is not supported.
2030                  * Only the master stripe has all locks cancelled early. */
2031                 /* Cancel all the locks on tgt child (fid4). */
2032                 if (rc == 0)
2033                         rc = lmv_early_cancel(lmv, src_exp, NULL, op_data,
2034                                               LCK_EX, MDS_INODELOCK_FULL,
2035                                               MF_MDC_CANCEL_FID4);
2036         }
2037
2038         if (rc == 0)
2039                 rc = md_rename(src_exp, op_data, old, oldlen,
2040                                new, newlen, request);
2041         if (rc == -ERESTART) {
2042                 LASSERT(*request != NULL);
2043                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2044                           "Got -ERESTART during rename!\n");
2045                 ptlrpc_req_finished(*request);
2046                 *request = NULL;
2047
2048                 /*
2049                  * Directory got split. Time to update local object and repeat
2050                  * the request with proper MDS.
2051                  */
2052                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2053                 if (rc == 0)
2054                         goto repeat;
2055         }
2056         RETURN(rc);
2057 }
2058
2059 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2060                        void *ea, int ealen, void *ea2, int ea2len,
2061                        struct ptlrpc_request **request,
2062                        struct md_open_data **mod)
2063 {
2064         struct obd_device *obd = exp->exp_obd;
2065         struct lmv_obd *lmv = &obd->u.lmv;
2066         struct ptlrpc_request *req;
2067         struct obd_export *tgt_exp;
2068         struct lmv_obj *obj;
2069         int rc = 0, i;
2070         ENTRY;
2071
2072         rc = lmv_check_connect(obd);
2073         if (rc)
2074                 RETURN(rc);
2075
2076         obj = lmv_obj_grab(obd, &op_data->op_fid1);
2077
2078         CDEBUG(D_OTHER, "SETATTR for "DFID", valid 0x%x%s\n",
2079                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2080                obj ? ", split" : "");
2081
2082         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2083         if (obj) {
2084                 for (i = 0; i < obj->lo_objcount; i++) {
2085                         op_data->op_fid1 = obj->lo_inodes[i].li_fid;
2086
2087                         tgt_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds);
2088                         if (IS_ERR(tgt_exp)) {
2089                                 rc = PTR_ERR(tgt_exp);
2090                                 break;
2091                         }
2092
2093                         rc = md_setattr(tgt_exp, op_data, ea, ealen,
2094                                         ea2, ea2len, &req, mod);
2095
2096                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_inodes[i].li_fid)) {
2097                                 /*
2098                                  * this is master object and this request should
2099                                  * be returned back to llite.
2100                                  */
2101                                 *request = req;
2102                         } else {
2103                                 ptlrpc_req_finished(req);
2104                         }
2105
2106                         if (rc)
2107                                 break;
2108                 }
2109                 lmv_obj_put(obj);
2110         } else {
2111                 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
2112                 if (IS_ERR(tgt_exp))
2113                         RETURN(PTR_ERR(tgt_exp));
2114
2115                 rc = md_setattr(tgt_exp, op_data, ea, ealen, ea2,
2116                                 ea2len, request, mod);
2117         }
2118         RETURN(rc);
2119 }
2120
2121 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2122                     struct obd_capa *oc, struct ptlrpc_request **request)
2123 {
2124         struct obd_device *obd = exp->exp_obd;
2125         struct lmv_obd *lmv = &obd->u.lmv;
2126         struct obd_export *tgt_exp;
2127         int rc;
2128         ENTRY;
2129
2130         rc = lmv_check_connect(obd);
2131         if (rc)
2132                 RETURN(rc);
2133
2134         tgt_exp = lmv_find_export(lmv, fid);
2135         if (IS_ERR(tgt_exp))
2136                 RETURN(PTR_ERR(tgt_exp));
2137
2138         rc = md_sync(tgt_exp, fid, oc, request);
2139         RETURN(rc);
2140 }
2141
2142 /* main purpose of LMV blocking ast is to remove split directory LMV
2143  * presentation object (struct lmv_obj) attached to the lock being revoked. */
2144 int lmv_blocking_ast(struct ldlm_lock *lock,
2145                      struct ldlm_lock_desc *desc,
2146                      void *data, int flag)
2147 {
2148         struct lustre_handle lockh;
2149         struct lmv_obj *obj;
2150         int rc;
2151         ENTRY;
2152
2153         switch (flag) {
2154         case LDLM_CB_BLOCKING:
2155                 ldlm_lock2handle(lock, &lockh);
2156                 rc = ldlm_cli_cancel(&lockh);
2157                 if (rc < 0) {
2158                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2159                         RETURN(rc);
2160                 }
2161                 break;
2162         case LDLM_CB_CANCELING:
2163                 /* time to drop cached attrs for dirobj */
2164                 obj = lock->l_ast_data;
2165                 if (obj) {
2166                         CDEBUG(D_OTHER, "cancel %s on "LPU64"/"LPU64
2167                                ", master "DFID"\n",
2168                                lock->l_resource->lr_name.name[3] == 1 ?
2169                                "LOOKUP" : "UPDATE",
2170                                lock->l_resource->lr_name.name[0],
2171                                lock->l_resource->lr_name.name[1],
2172                                PFID(&obj->lo_fid));
2173                         lmv_obj_put(obj);
2174                 }
2175                 break;
2176         default:
2177                 LBUG();
2178         }
2179         RETURN(0);
2180 }
2181
2182 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2183 {
2184         __u64 val;
2185
2186         val = le64_to_cpu(*hash);
2187         if (val < hash_adj)
2188                 val += MAX_HASH_SIZE;
2189         if (val != DIR_END_OFF)
2190                 *hash = cpu_to_le64(val - hash_adj);
2191 }
2192
2193 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2194 {
2195         __u64 id;
2196         struct obd_import *imp;
2197
2198         /*
2199          * XXX Hack: to get nid we assume that underlying obd device is mdc.
2200          */
2201         imp  = class_exp2cliimp(exp);
2202         id   = imp->imp_connection->c_self + fid_flatten(fid);
2203
2204         CDEBUG(D_INFO, "node rank: %llx "DFID" %llx %llx\n",
2205                imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2206
2207         return id ^ (id >> 32);
2208 }
2209
2210 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2211                         struct obd_capa *oc, __u64 offset64, struct page *page,
2212                         struct ptlrpc_request **request)
2213 {
2214         struct obd_device *obd = exp->exp_obd;
2215         struct lmv_obd *lmv = &obd->u.lmv;
2216         struct obd_export *tgt_exp;
2217         struct lu_fid rid = *fid;
2218         struct lmv_obj *obj;
2219         __u64 offset;
2220         __u64 hash_adj = 0;
2221         __u32 rank = 0;
2222         __u64 seg_size = 0;
2223         __u64 tgt_tmp = 0;
2224         int tgt = 0;
2225         int tgt0 = 0;
2226         int rc;
2227         int nr = 0;
2228         ENTRY;
2229
2230         offset = offset64;
2231
2232         rc = lmv_check_connect(obd);
2233         if (rc)
2234                 RETURN(rc);
2235
2236         CDEBUG(D_INFO, "READPAGE at %llx from "DFID"\n", offset, PFID(&rid));
2237
2238         obj = lmv_obj_grab(obd, fid);
2239         if (obj) {
2240
2241         /*
2242          * This case handle directory lookup in clustered metadata case (i.e.
2243          * split directory is located on multiple md servers.)
2244          * each server keeps directory entries for certain range of hashes.
2245          * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2246          * first server will keep records with hashes [ 0 ... MAX_HASH / N  - 1],
2247          * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2248          * so on....
2249          *      readdir can simply start reading entries from 0 - N server in
2250          * order but that will not scale well as all client will request dir in
2251          * to server in same order.
2252          * Following algorithm does optimization:
2253          * Instead of doing readdir in 1, 2, ...., N order, client with a
2254          * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2255          * (every client has rank R)
2256          *      But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2257          * since client ask dir from MDS{R} client has pages with offsets
2258          * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2259          * on hash  values that we get.
2260          */
2261
2262                 struct lmv_inode *loi;
2263
2264                 lmv_obj_lock(obj);
2265
2266                 nr       = obj->lo_objcount;
2267                 LASSERT(nr > 0);
2268                 seg_size = MAX_HASH_SIZE;
2269                 do_div(seg_size, nr);
2270                 loi      = obj->lo_inodes;
2271                 rank     = lmv_node_rank(lmv_get_export(lmv, loi[0].li_mds),
2272                                          fid) % nr;
2273                 tgt_tmp = offset;
2274                 do_div(tgt_tmp, seg_size);
2275                 tgt0     = do_div(tgt_tmp,  nr);
2276                 tgt      = (tgt0 + rank) % nr;
2277
2278                 if (tgt < tgt0)
2279                         /*
2280                          * Wrap around.
2281                          *
2282                          * Last segment has unusual length due to division
2283                          * rounding.
2284                          */
2285                         hash_adj = MAX_HASH_SIZE - seg_size * nr;
2286                 else
2287                         hash_adj = 0;
2288
2289                 hash_adj += rank * seg_size;
2290
2291                 CDEBUG(D_INFO, "hash_adj: %x %llx %llx/%x -> %llx/%x\n",
2292                        rank, hash_adj, offset, tgt0, offset + hash_adj, tgt);
2293
2294                 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2295                 rid = obj->lo_inodes[tgt].li_fid;
2296                 tgt_exp = lmv_get_export(lmv, loi[tgt].li_mds);
2297
2298                 CDEBUG(D_INFO, "forward to "DFID" with offset %lu i %d\n",
2299                        PFID(&rid), (unsigned long)offset, tgt);
2300         } else
2301                 tgt_exp = lmv_find_export(lmv, &rid);
2302
2303         if (IS_ERR(tgt_exp))
2304                 GOTO(cleanup, rc = PTR_ERR(tgt_exp));
2305
2306         rc = md_readpage(tgt_exp, &rid, oc, offset, page, request);
2307         if (rc)
2308                 GOTO(cleanup, rc);
2309         if (obj) {
2310                 struct lu_dirpage *dp;
2311                 struct lu_dirent  *ent;
2312
2313                 dp = cfs_kmap(page);
2314
2315                 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2316                 lmv_hash_adjust(&dp->ldp_hash_end,   hash_adj);
2317                 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2318
2319                 for (ent = lu_dirent_start(dp); ent != NULL;
2320                      ent = lu_dirent_next(ent))
2321                         lmv_hash_adjust(&ent->lde_hash, hash_adj);
2322
2323                 if (tgt0 != nr - 1) {
2324                         __u64 end;
2325
2326                         end = le64_to_cpu(dp->ldp_hash_end);
2327                         if (end == DIR_END_OFF) {
2328                                 dp->ldp_hash_end = cpu_to_le32(seg_size *
2329                                                                (tgt0 + 1));
2330                                 CDEBUG(D_INFO, ""DFID" reset end %llx tgt %d\n",
2331                                        PFID(&rid),
2332                                        le64_to_cpu(dp->ldp_hash_end), tgt);
2333                         }
2334                 }
2335                 cfs_kunmap(page);
2336         }
2337         /*
2338          * Here we could remove "." and ".." from all pages which at not from
2339          * master. But MDS has only "." and ".." for master dir.
2340          */
2341         EXIT;
2342 cleanup:
2343         if (obj) {
2344                 lmv_obj_unlock(obj);
2345                 lmv_obj_put(obj);
2346         }
2347         return rc;
2348 }
2349
2350 static int lmv_unlink_slaves(struct obd_export *exp,
2351                              struct md_op_data *op_data,
2352                              struct ptlrpc_request **req)
2353 {
2354         struct obd_device *obd = exp->exp_obd;
2355         struct lmv_obd *lmv = &obd->u.lmv;
2356         struct lmv_stripe_md *mea = op_data->op_mea1;
2357         struct md_op_data *op_data2;
2358         struct obd_export *tgt_exp;
2359         int i, rc = 0;
2360         ENTRY;
2361
2362         OBD_ALLOC_PTR(op_data2);
2363         if (op_data2 == NULL)
2364                 RETURN(-ENOMEM);
2365
2366         op_data2->op_mode = S_IFDIR;
2367         op_data2->op_fsuid = current->fsuid;
2368         op_data2->op_fsgid = current->fsgid;
2369         op_data2->op_bias = 0;
2370
2371         LASSERT(mea != NULL);
2372         for (i = 0; i < mea->mea_count; i++) {
2373                 memset(op_data2, 0, sizeof(*op_data2));
2374                 op_data2->op_fid1 = mea->mea_ids[i];
2375                 tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1);
2376                 if (IS_ERR(tgt_exp))
2377                         GOTO(out_free_op_data2, rc = PTR_ERR(tgt_exp));
2378
2379                 if (tgt_exp == NULL)
2380                         continue;
2381
2382                 rc = md_unlink(tgt_exp, op_data2, req);
2383
2384                 CDEBUG(D_OTHER, "unlink slave "DFID" -> %d\n",
2385                        PFID(&mea->mea_ids[i]), rc);
2386
2387                 if (*req) {
2388                         ptlrpc_req_finished(*req);
2389                         *req = NULL;
2390                 }
2391                 if (rc)
2392                         GOTO(out_free_op_data2, rc);
2393         }
2394
2395         EXIT;
2396 out_free_op_data2:
2397         OBD_FREE_PTR(op_data2);
2398         return rc;
2399 }
2400
2401 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2402                       struct ptlrpc_request **request)
2403 {
2404         struct obd_device *obd = exp->exp_obd;
2405         struct lmv_obd *lmv = &obd->u.lmv;
2406         struct obd_export *tgt_exp = NULL;
2407         struct lmv_obj *obj;
2408         int rc, loop = 0;
2409         ENTRY;
2410
2411         rc = lmv_check_connect(obd);
2412         if (rc)
2413                 RETURN(rc);
2414
2415         if (op_data->op_namelen == 0 && op_data->op_mea1 != NULL) {
2416                 /* mds asks to remove slave objects */
2417                 rc = lmv_unlink_slaves(exp, op_data, request);
2418                 RETURN(rc);
2419         }
2420
2421 repeat:
2422         ++loop;
2423         LASSERT(loop <= 2);
2424         if (op_data->op_namelen != 0) {
2425                 int mea_idx;
2426
2427                 obj = lmv_obj_grab(obd, &op_data->op_fid1);
2428                 if (obj) {
2429                         mea_idx = raw_name2idx(obj->lo_hashtype,
2430                                                obj->lo_objcount,
2431                                                op_data->op_name,
2432                                                op_data->op_namelen);
2433                         op_data->op_bias &= ~MDS_CHECK_SPLIT;
2434                         op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid;
2435                         tgt_exp = lmv_get_export(lmv,
2436                                                  obj->lo_inodes[mea_idx].li_mds);
2437                         lmv_obj_put(obj);
2438                         CDEBUG(D_OTHER, "unlink '%*s' in "DFID" -> %u\n",
2439                                op_data->op_namelen, op_data->op_name,
2440                                PFID(&op_data->op_fid1), mea_idx);
2441                 }
2442         } else {
2443                 CDEBUG(D_OTHER, "drop i_nlink on "DFID"\n",
2444                        PFID(&op_data->op_fid1));
2445         }
2446         if (tgt_exp == NULL) {
2447                 tgt_exp = lmv_find_export(lmv, &op_data->op_fid1);
2448                 if (IS_ERR(tgt_exp))
2449                         RETURN(PTR_ERR(tgt_exp));
2450                 op_data->op_bias |= MDS_CHECK_SPLIT;
2451         }
2452
2453         op_data->op_fsuid = current->fsuid;
2454         op_data->op_fsgid = current->fsgid;
2455         op_data->op_cap   = current->cap_effective;
2456
2457         /* If child's fid is given, cancel unused locks for it if it is from
2458          * another export than parent. */
2459         if (op_data->op_namelen) {
2460                 /* LOOKUP lock for child (fid3) should also be cancelled on
2461                  * parent tgt_exp in mdc_unlink(). */
2462                 op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2463
2464                 /* XXX: the case when child is a striped dir is not supported.
2465                  * Only the master stripe has all locks cancelled early. */
2466                 /* Cancel FULL locks on child (fid3). */
2467                 rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX,
2468                                       MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2469         }
2470         if (rc == 0)
2471                 rc = md_unlink(tgt_exp, op_data, request);
2472         if (rc == -ERESTART) {
2473                 LASSERT(*request != NULL);
2474                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2475                           "Got -ERESTART during unlink!\n");
2476                 ptlrpc_req_finished(*request);
2477                 *request = NULL;
2478
2479                 /*
2480                  * Directory got split. Time to update local object and repeat
2481                  * the request with proper MDS.
2482                  */
2483                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2484                 if (rc == 0)
2485                         goto repeat;
2486         }
2487         RETURN(rc);
2488 }
2489
2490 static int lmv_llog_init(struct obd_device *obd, int group,
2491                          struct obd_device *tgt, int count,
2492                          struct llog_catid *logid, struct obd_uuid *uuid)
2493 {
2494 #if 0
2495         struct llog_ctxt *ctxt;
2496         int rc;
2497         ENTRY;
2498
2499         LASSERT(group == OBD_LLOG_GROUP);
2500         rc = llog_setup(obd, &obd->obd_olg, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL,
2501                         &llog_client_ops);
2502         if (rc == 0) {
2503                 ctxt = llog_group_get_ctxt(&obd->obd_olg, LLOG_CONFIG_REPL_CTXT);
2504                 llog_initiator_connect(ctxt, tgt);
2505                 llog_ctxt_put(ctxt);
2506         }
2507         RETURN(rc);
2508 #else
2509         return 0;
2510 #endif
2511 }
2512
2513 static int lmv_llog_finish(struct obd_device *obd, int count)
2514 {
2515         struct llog_ctxt *ctxt;
2516         int rc = 0;
2517         ENTRY;
2518
2519         ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT);
2520         if (ctxt)
2521                 rc = llog_cleanup(ctxt);
2522
2523         RETURN(rc);
2524 }
2525
2526 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2527 {
2528         int rc = 0;
2529
2530         switch (stage) {
2531         case OBD_CLEANUP_EARLY:
2532                 /* XXX: here should be calling obd_precleanup() down to
2533                  * stack. */
2534                 break;
2535         case OBD_CLEANUP_SELF_EXP:
2536                 rc = obd_llog_finish(obd, 0);
2537                 if (rc != 0)
2538                         CERROR("failed to cleanup llogging subsystems\n");
2539                 break;
2540         default:
2541                 break;
2542         }
2543         RETURN(rc);
2544 }
2545
2546 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2547                         void *key, __u32 *vallen, void *val)
2548 {
2549         struct obd_device *obd;
2550         struct lmv_obd *lmv;
2551         int rc = 0;
2552         ENTRY;
2553
2554         obd = class_exp2obd(exp);
2555         if (obd == NULL) {
2556                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2557                        exp->exp_handle.h_cookie);
2558                 RETURN(-EINVAL);
2559         }
2560
2561         lmv = &obd->u.lmv;
2562         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2563                 struct lmv_tgt_desc *tgts;
2564                 int i;
2565
2566                 rc = lmv_check_connect(obd);
2567                 if (rc)
2568                         RETURN(rc);
2569
2570                 LASSERT(*vallen == sizeof(__u32));
2571                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2572                      i++, tgts++) {
2573
2574                         /* all tgts should be connected when this get called. */
2575                         if (!tgts || !tgts->ltd_exp) {
2576                                 CERROR("target not setup?\n");
2577                                 continue;
2578                         }
2579
2580                         if (!obd_get_info(tgts->ltd_exp, keylen, key,
2581                                           vallen, val))
2582                                 RETURN(0);
2583                 }
2584                 RETURN(-EINVAL);
2585         } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2586                 rc = lmv_check_connect(obd);
2587                 if (rc)
2588                         RETURN(rc);
2589
2590                 /* forwarding this request to first MDS, it should know LOV
2591                  * desc. */
2592                 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2593                                   vallen, val);
2594                 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2595                         exp->exp_connect_flags =
2596                         ((struct obd_connect_data *)val)->ocd_connect_flags;
2597                 }
2598                 RETURN(rc);
2599         }
2600
2601         CDEBUG(D_IOCTL, "invalid key\n");
2602         RETURN(-EINVAL);
2603 }
2604
2605 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2606                        void *key, obd_count vallen, void *val,
2607                        struct ptlrpc_request_set *set)
2608 {
2609         struct lmv_tgt_desc    *tgt;
2610         struct obd_device      *obd;
2611         struct lmv_obd         *lmv;
2612         int rc = 0;
2613         ENTRY;
2614
2615         obd = class_exp2obd(exp);
2616         if (obd == NULL) {
2617                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2618                        exp->exp_handle.h_cookie);
2619                 RETURN(-EINVAL);
2620         }
2621         lmv = &obd->u.lmv;
2622
2623         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2624             KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2625                 int i, err = 0;
2626
2627                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2628                         tgt = &lmv->tgts[i];
2629
2630                         if (!tgt->ltd_exp)
2631                                 continue;
2632
2633                         err = obd_set_info_async(tgt->ltd_exp,
2634                                                  keylen, key, vallen, val, set);
2635                         if (err && rc == 0)
2636                                 rc = err;
2637                 }
2638
2639                 RETURN(rc);
2640         }
2641
2642         RETURN(-EINVAL);
2643 }
2644
2645 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2646                struct lov_stripe_md *lsm)
2647 {
2648         struct obd_device *obd = class_exp2obd(exp);
2649         struct lmv_obd *lmv = &obd->u.lmv;
2650         struct lmv_stripe_md *meap, *lsmp;
2651         int mea_size, i;
2652         ENTRY;
2653
2654         mea_size = lmv_get_easize(lmv);
2655         if (!lmmp)
2656                 RETURN(mea_size);
2657
2658         if (*lmmp && !lsm) {
2659                 OBD_FREE(*lmmp, mea_size);
2660                 *lmmp = NULL;
2661                 RETURN(0);
2662         }
2663
2664         if (*lmmp == NULL) {
2665                 OBD_ALLOC(*lmmp, mea_size);
2666                 if (*lmmp == NULL)
2667                         RETURN(-ENOMEM);
2668         }
2669
2670         if (!lsm)
2671                 RETURN(mea_size);
2672
2673         lsmp = (struct lmv_stripe_md *)lsm;
2674         meap = (struct lmv_stripe_md *)*lmmp;
2675
2676         if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2677             lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2678                 RETURN(-EINVAL);
2679
2680         meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2681         meap->mea_count = cpu_to_le32(lsmp->mea_count);
2682         meap->mea_master = cpu_to_le32(lsmp->mea_master);
2683
2684         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2685                 meap->mea_ids[i] = meap->mea_ids[i];
2686                 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2687         }
2688
2689         RETURN(mea_size);
2690 }
2691
2692 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2693                  struct lov_mds_md *lmm, int lmm_size)
2694 {
2695         struct obd_device *obd = class_exp2obd(exp);
2696         struct lmv_stripe_md **tmea = (struct lmv_stripe_md **)lsmp;
2697         struct lmv_stripe_md *mea = (struct lmv_stripe_md *)lmm;
2698         struct lmv_obd *lmv = &obd->u.lmv;
2699         int mea_size, i;
2700         __u32 magic;
2701         ENTRY;
2702
2703         mea_size = lmv_get_easize(lmv);
2704         if (lsmp == NULL)
2705                 return mea_size;
2706
2707         if (*lsmp != NULL && lmm == NULL) {
2708                 OBD_FREE(*tmea, mea_size);
2709                 *lsmp = NULL;
2710                 RETURN(0);
2711         }
2712
2713         LASSERT(mea_size == lmm_size);
2714
2715         OBD_ALLOC(*tmea, mea_size);
2716         if (*tmea == NULL)
2717                 RETURN(-ENOMEM);
2718
2719         if (!lmm)
2720                 RETURN(mea_size);
2721
2722         if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2723             mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2724             mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2725         {
2726                 magic = le32_to_cpu(mea->mea_magic);
2727         } else {
2728                 /* old mea is not handled here */
2729                 LBUG();
2730         }
2731
2732         (*tmea)->mea_magic = magic;
2733         (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2734         (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2735
2736         for (i = 0; i < (*tmea)->mea_count; i++) {
2737                 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2738                 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2739         }
2740         RETURN(mea_size);
2741 }
2742
2743 static int lmv_cancel_unused(struct obd_export *exp,
2744                              const struct lu_fid *fid,
2745                              ldlm_policy_data_t *policy,
2746                              ldlm_mode_t mode, int flags, void *opaque)
2747 {
2748         struct obd_device *obd = exp->exp_obd;
2749         struct lmv_obd *lmv = &obd->u.lmv;
2750         int rc = 0, err, i;
2751         ENTRY;
2752
2753         LASSERT(fid != NULL);
2754
2755         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2756                 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2757                         continue;
2758
2759                 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2760                                        policy, mode, flags, opaque);
2761                 if (!rc)
2762                         rc = err;
2763         }
2764         RETURN(rc);
2765 }
2766
2767 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data)
2768 {
2769         struct obd_device *obd = exp->exp_obd;
2770         struct lmv_obd *lmv = &obd->u.lmv;
2771
2772         ENTRY;
2773         RETURN(md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data));
2774 }
2775
2776 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2777                            const struct lu_fid *fid, ldlm_type_t type,
2778                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
2779                            struct lustre_handle *lockh)
2780 {
2781         struct obd_device *obd = exp->exp_obd;
2782         struct lmv_obd *lmv = &obd->u.lmv;
2783         ldlm_mode_t rc;
2784         int i;
2785         ENTRY;
2786
2787         CDEBUG(D_OTHER, "lock match for "DFID"\n", PFID(fid));
2788
2789         /* with CMD every object can have two locks in different namespaces:
2790          * lookup lock in space of mds storing direntry and update/open lock in
2791          * space of mds storing inode. Thus we check all targets, not only that
2792          * one fid was created in. */
2793         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2794                 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2795                                    type, policy, mode, lockh);
2796                 if (rc)
2797                         RETURN(rc);
2798         }
2799
2800         RETURN(0);
2801 }
2802
2803 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2804                       struct obd_export *dt_exp, struct obd_export *md_exp,
2805                       struct lustre_md *md)
2806 {
2807         struct obd_device *obd = exp->exp_obd;
2808         struct lmv_obd *lmv = &obd->u.lmv;
2809         int rc;
2810
2811         ENTRY;
2812         rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2813         RETURN(rc);
2814 }
2815
2816 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2817 {
2818         struct obd_device *obd = exp->exp_obd;
2819         struct lmv_obd *lmv = &obd->u.lmv;
2820
2821         ENTRY;
2822         if (md->mea)
2823                 obd_free_memmd(exp, (struct lov_stripe_md**)&md->mea);
2824         RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2825 }
2826
2827 int lmv_set_open_replay_data(struct obd_export *exp,
2828                              struct obd_client_handle *och,
2829                              struct ptlrpc_request *open_req)
2830 {
2831         struct obd_device *obd = exp->exp_obd;
2832         struct lmv_obd *lmv = &obd->u.lmv;
2833         struct obd_export *tgt_exp;
2834
2835         ENTRY;
2836
2837         tgt_exp = lmv_find_export(lmv, &och->och_fid);
2838         if (IS_ERR(tgt_exp))
2839                 RETURN(PTR_ERR(tgt_exp));
2840
2841         RETURN(md_set_open_replay_data(tgt_exp, och, open_req));
2842 }
2843
2844 int lmv_clear_open_replay_data(struct obd_export *exp,
2845                                struct obd_client_handle *och)
2846 {
2847         struct obd_device *obd = exp->exp_obd;
2848         struct lmv_obd *lmv = &obd->u.lmv;
2849         struct obd_export *tgt_exp;
2850         ENTRY;
2851
2852         tgt_exp = lmv_find_export(lmv, &och->och_fid);
2853         if (IS_ERR(tgt_exp))
2854                 RETURN(PTR_ERR(tgt_exp));
2855
2856         RETURN(md_clear_open_replay_data(tgt_exp, och));
2857 }
2858
2859 static int lmv_get_remote_perm(struct obd_export *exp,
2860                                const struct lu_fid *fid,
2861                                struct obd_capa *oc, __u32 suppgid,
2862                                struct ptlrpc_request **request)
2863 {
2864         struct obd_device *obd = exp->exp_obd;
2865         struct lmv_obd *lmv = &obd->u.lmv;
2866         struct obd_export *tgt_exp;
2867         int rc;
2868
2869         ENTRY;
2870
2871         rc = lmv_check_connect(obd);
2872         if (rc)
2873                 RETURN(rc);
2874
2875         tgt_exp = lmv_find_export(lmv, fid);
2876         if (IS_ERR(tgt_exp))
2877                 RETURN(PTR_ERR(tgt_exp));
2878
2879         rc = md_get_remote_perm(tgt_exp, fid, oc, suppgid, request);
2880
2881         RETURN(rc);
2882 }
2883
2884 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2885                           renew_capa_cb_t cb)
2886 {
2887         struct obd_device *obd = exp->exp_obd;
2888         struct lmv_obd *lmv = &obd->u.lmv;
2889         struct obd_export *tgt_exp;
2890         int rc;
2891         ENTRY;
2892
2893         rc = lmv_check_connect(obd);
2894         if (rc)
2895                 RETURN(rc);
2896
2897         tgt_exp = lmv_find_export(lmv, &oc->c_capa.lc_fid);
2898         if (IS_ERR(tgt_exp))
2899                 RETURN(PTR_ERR(tgt_exp));
2900
2901         rc = md_renew_capa(tgt_exp, oc, cb);
2902         RETURN(rc);
2903 }
2904
2905 struct obd_ops lmv_obd_ops = {
2906         .o_owner                = THIS_MODULE,
2907         .o_setup                = lmv_setup,
2908         .o_cleanup              = lmv_cleanup,
2909         .o_precleanup           = lmv_precleanup,
2910         .o_process_config       = lmv_process_config,
2911         .o_connect              = lmv_connect,
2912         .o_disconnect           = lmv_disconnect,
2913         .o_statfs               = lmv_statfs,
2914         .o_llog_init            = lmv_llog_init,
2915         .o_llog_finish          = lmv_llog_finish,
2916         .o_get_info             = lmv_get_info,
2917         .o_set_info_async       = lmv_set_info_async,
2918         .o_packmd               = lmv_packmd,
2919         .o_unpackmd             = lmv_unpackmd,
2920         .o_notify               = lmv_notify,
2921         .o_iocontrol            = lmv_iocontrol,
2922         .o_fid_delete           = lmv_fid_delete
2923 };
2924
2925 struct md_ops lmv_md_ops = {
2926         .m_getstatus            = lmv_getstatus,
2927         .m_change_cbdata        = lmv_change_cbdata,
2928         .m_close                = lmv_close,
2929         .m_create               = lmv_create,
2930         .m_done_writing         = lmv_done_writing,
2931         .m_enqueue              = lmv_enqueue,
2932         .m_getattr              = lmv_getattr,
2933         .m_getxattr             = lmv_getxattr,
2934         .m_getattr_name         = lmv_getattr_name,
2935         .m_intent_lock          = lmv_intent_lock,
2936         .m_link                 = lmv_link,
2937         .m_rename               = lmv_rename,
2938         .m_setattr              = lmv_setattr,
2939         .m_setxattr             = lmv_setxattr,
2940         .m_sync                 = lmv_sync,
2941         .m_readpage             = lmv_readpage,
2942         .m_unlink               = lmv_unlink,
2943         .m_init_ea_size         = lmv_init_ea_size,
2944         .m_cancel_unused        = lmv_cancel_unused,
2945         .m_set_lock_data        = lmv_set_lock_data,
2946         .m_lock_match           = lmv_lock_match,
2947         .m_get_lustre_md        = lmv_get_lustre_md,
2948         .m_free_lustre_md       = lmv_free_lustre_md,
2949         .m_set_open_replay_data = lmv_set_open_replay_data,
2950         .m_clear_open_replay_data = lmv_clear_open_replay_data,
2951         .m_get_remote_perm      = lmv_get_remote_perm,
2952         .m_renew_capa           = lmv_renew_capa
2953 };
2954
2955 int __init lmv_init(void)
2956 {
2957         struct lprocfs_static_vars lvars;
2958         int rc;
2959
2960         obj_cache = cfs_mem_cache_create("lmv_objects",
2961                                       sizeof(struct lmv_obj),
2962                                       0, 0);
2963         if (!obj_cache) {
2964                 CERROR("error allocating lmv objects cache\n");
2965                 return -ENOMEM;
2966         }
2967
2968         lprocfs_lmv_init_vars(&lvars);
2969         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
2970                                  lvars.module_vars, LUSTRE_LMV_NAME, NULL);
2971         if (rc)
2972                 cfs_mem_cache_destroy(obj_cache);
2973
2974         return rc;
2975 }
2976
2977 #ifdef __KERNEL__
2978 static void lmv_exit(void)
2979 {
2980         int rc;
2981
2982         class_unregister_type(LUSTRE_LMV_NAME);
2983
2984         rc = cfs_mem_cache_destroy(obj_cache);
2985         LASSERTF(rc == 0,
2986                  "can't free lmv objects cache, %d object(s)"
2987                  "still in use\n", atomic_read(&obj_cache_count));
2988 }
2989
2990 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2991 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
2992 MODULE_LICENSE("GPL");
2993
2994 module_init(lmv_init);
2995 module_exit(lmv_exit);
2996 #endif