lustre/lmv/lmv_obd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LMV
  41 #ifdef __KERNEL__
  42 #include <linux/slab.h>
  43 #include <linux/module.h>
  44 #include <linux/init.h>
  45 #include <linux/slab.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/mm.h>
  48 #include <asm/div64.h>
  49 #include <linux/seq_file.h>
  50 #include <linux/namei.h>
  51 #else
  52 #include <liblustre.h>
  53 #endif
  54
  55 #include <lustre/lustre_idl.h>
  56 #include <lustre_log.h>
  57 #include <obd_support.h>
  58 #include <lustre_lib.h>
  59 #include <lustre_net.h>
  60 #include <obd_class.h>
  61 #include <lprocfs_status.h>
  62 #include <lustre_lite.h>
  63 #include <lustre_fid.h>
  64 #include "lmv_internal.h"
  65
  66 /* object cache. */
  67 cfs_mem_cache_t *lmv_object_cache;
  68 atomic_t lmv_object_count = ATOMIC_INIT(0);
  69
  70 static void lmv_activate_target(struct lmv_obd *lmv,
  71                                 struct lmv_tgt_desc *tgt,
  72                                 int activate)
  73 {
  74         if (tgt->ltd_active == activate)
  75                 return;
  76
  77         tgt->ltd_active = activate;
  78         lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
  79 }
  80
  81 /**
  82  * Error codes:
  83  *
  84  *  -EINVAL  : UUID can't be found in the LMV's target list
  85  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  86  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  87  */
  88 static int lmv_set_mdc_active(struct lmv_obd *lmv, struct obd_uuid *uuid,
  89                               int activate)
  90 {
  91         struct lmv_tgt_desc    *tgt;
  92         struct obd_device      *obd;
  93         int                     i;
  94         int                     rc = 0;
  95         ENTRY;
  96
  97         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  98                lmv, uuid->uuid, activate);
  99
 100         spin_lock(&lmv->lmv_lock);
 101         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 102                 if (tgt->ltd_exp == NULL)
 103                         continue;
 104
 105                 CDEBUG(D_INFO, "Target idx %d is %s conn "LPX64"\n",
 106                        i, tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
 107
 108                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 109                         break;
 110         }
 111
 112         if (i == lmv->desc.ld_tgt_count)
 113                 GOTO(out_lmv_lock, rc = -EINVAL);
 114
 115         obd = class_exp2obd(tgt->ltd_exp);
 116         if (obd == NULL)
 117                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 118
 119         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 120                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 121                obd->obd_type->typ_name, i);
 122         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 123
 124         if (tgt->ltd_active == activate) {
 125                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 126                        activate ? "" : "in");
 127                 GOTO(out_lmv_lock, rc);
 128         }
 129
 130         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 131                activate ? "" : "in");
 132         lmv_activate_target(lmv, tgt, activate);
 133         EXIT;
 134
 135  out_lmv_lock:
 136         spin_unlock(&lmv->lmv_lock);
 137         return rc;
 138 }
 139
 140 static int lmv_set_mdc_data(struct lmv_obd *lmv, struct obd_uuid *uuid,
 141                             struct obd_connect_data *data)
 142 {
 143         struct lmv_tgt_desc    *tgt;
 144         int                     i;
 145         ENTRY;
 146
 147         LASSERT(data != NULL);
 148
 149         spin_lock(&lmv->lmv_lock);
 150         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 151                 if (tgt->ltd_exp == NULL)
 152                         continue;
 153
 154                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
 155                         lmv->datas[tgt->ltd_idx] = *data;
 156                         break;
 157                 }
 158         }
 159         spin_unlock(&lmv->lmv_lock);
 160         RETURN(0);
 161 }
 162
 163 struct obd_uuid *lmv_get_uuid(struct obd_export *exp) {
 164         struct obd_device *obd = exp->exp_obd;
 165         struct lmv_obd *lmv = &obd->u.lmv;
 166         return obd_get_uuid(lmv->tgts[0].ltd_exp);
 167 }
 168
 169 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 170                       enum obd_notify_event ev, void *data)
 171 {
 172         struct obd_connect_data *conn_data;
 173         struct lmv_obd          *lmv = &obd->u.lmv;
 174         struct obd_uuid         *uuid;
 175         int                      rc = 0;
 176         ENTRY;
 177
 178         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 179                 CERROR("unexpected notification of %s %s!\n",
 180                        watched->obd_type->typ_name,
 181                        watched->obd_name);
 182                 RETURN(-EINVAL);
 183         }
 184
 185         uuid = &watched->u.cli.cl_target_uuid;
 186         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 187                 /*
 188                  * Set MDC as active before notifying the observer, so the
 189                  * observer can use the MDC normally.
 190                  */
 191                 rc = lmv_set_mdc_active(lmv, uuid,
 192                                         ev == OBD_NOTIFY_ACTIVE);
 193                 if (rc) {
 194                         CERROR("%sactivation of %s failed: %d\n",
 195                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 196                                uuid->uuid, rc);
 197                         RETURN(rc);
 198                 }
 199         } else if (ev == OBD_NOTIFY_OCD) {
 200                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 201
 202                 /*
 203                  * Set connect data to desired target, update exp_connect_flags.
 204                  */
 205                 rc = lmv_set_mdc_data(lmv, uuid, conn_data);
 206                 if (rc) {
 207                         CERROR("can't set connect data to target %s, rc %d\n",
 208                                uuid->uuid, rc);
 209                         RETURN(rc);
 210                 }
 211
 212                 /*
 213                  * XXX: Make sure that ocd_connect_flags from all targets are
 214                  * the same. Otherwise one of MDTs runs wrong version or
 215                  * something like this.  --umka
 216                  */
 217                 obd->obd_self_export->exp_connect_flags =
 218                         conn_data->ocd_connect_flags;
 219         }
 220 #if 0
 221         else if (ev == OBD_NOTIFY_DISCON) {
 222                 /*
 223                  * For disconnect event, flush fld cache for failout MDS case.
 224                  */
 225                 fld_client_flush(&lmv->lmv_fld);
 226         }
 227 #endif
 228         /*
 229          * Pass the notification up the chain.
 230          */
 231         if (obd->obd_observer)
 232                 rc = obd_notify(obd->obd_observer, watched, ev, data);
 233
 234         RETURN(rc);
 235 }
 236
 237 /**
 238  * This is fake connect function. Its purpose is to initialize lmv and say
 239  * caller that everything is okay. Real connection will be performed later.
 240  */
 241 static int lmv_connect(const struct lu_env *env,
 242                        struct obd_export **exp, struct obd_device *obd,
 243                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 244                        void *localdata)
 245 {
 246 #ifdef __KERNEL__
 247         struct proc_dir_entry *lmv_proc_dir;
 248 #endif
 249         struct lmv_obd        *lmv = &obd->u.lmv;
 250         struct lustre_handle  conn = { 0 };
 251         int                    rc = 0;
 252         ENTRY;
 253
 254         /*
 255          * We don't want to actually do the underlying connections more than
 256          * once, so keep track.
 257          */
 258         lmv->refcount++;
 259         if (lmv->refcount > 1) {
 260                 *exp = NULL;
 261                 RETURN(0);
 262         }
 263
 264         rc = class_connect(&conn, obd, cluuid);
 265         if (rc) {
 266                 CERROR("class_connection() returned %d\n", rc);
 267                 RETURN(rc);
 268         }
 269
 270         *exp = class_conn2export(&conn);
 271         class_export_get(*exp);
 272
 273         lmv->exp = *exp;
 274         lmv->connected = 0;
 275         lmv->cluuid = *cluuid;
 276
 277         if (data)
 278                 lmv->conn_data = *data;
 279
 280 #ifdef __KERNEL__
 281         lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry,
 282                                         NULL, NULL);
 283         if (IS_ERR(lmv_proc_dir)) {
 284                 CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.",
 285                        obd->obd_type->typ_name, obd->obd_name);
 286                 lmv_proc_dir = NULL;
 287         }
 288 #endif
 289
 290         /*
 291          * All real clients should perform actual connection right away, because
 292          * it is possible, that LMV will not have opportunity to connect targets
 293          * and MDC stuff will be called directly, for instance while reading
 294          * ../mdc/../kbytesfree procfs file, etc.
 295          */
 296         if (data->ocd_connect_flags & OBD_CONNECT_REAL)
 297                 rc = lmv_check_connect(obd);
 298
 299 #ifdef __KERNEL__
 300         if (rc) {
 301                 if (lmv_proc_dir)
 302                         lprocfs_remove(&lmv_proc_dir);
 303         }
 304 #endif
 305
 306         RETURN(rc);
 307 }
 308
 309 static void lmv_set_timeouts(struct obd_device *obd)
 310 {
 311         struct lmv_tgt_desc   *tgts;
 312         struct lmv_obd        *lmv;
 313         int                    i;
 314
 315         lmv = &obd->u.lmv;
 316         if (lmv->server_timeout == 0)
 317                 return;
 318
 319         if (lmv->connected == 0)
 320                 return;
 321
 322         for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
 323                 if (tgts->ltd_exp == NULL)
 324                         continue;
 325
 326                 obd_set_info_async(tgts->ltd_exp, sizeof(KEY_INTERMDS),
 327                                    KEY_INTERMDS, 0, NULL, NULL);
 328         }
 329 }
 330
 331 static int lmv_init_ea_size(struct obd_export *exp, int easize,
 332                             int def_easize, int cookiesize)
 333 {
 334         struct obd_device   *obd = exp->exp_obd;
 335         struct lmv_obd      *lmv = &obd->u.lmv;
 336         int                  i;
 337         int                  rc = 0;
 338         int                  change = 0;
 339         ENTRY;
 340
 341         if (lmv->max_easize < easize) {
 342                 lmv->max_easize = easize;
 343                 change = 1;
 344         }
 345         if (lmv->max_def_easize < def_easize) {
 346                 lmv->max_def_easize = def_easize;
 347                 change = 1;
 348         }
 349         if (lmv->max_cookiesize < cookiesize) {
 350                 lmv->max_cookiesize = cookiesize;
 351                 change = 1;
 352         }
 353         if (change == 0)
 354                 RETURN(0);
 355
 356         if (lmv->connected == 0)
 357                 RETURN(0);
 358
 359         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 360                 if (lmv->tgts[i].ltd_exp == NULL) {
 361                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
 362                         continue;
 363                 }
 364
 365                 rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize,
 366                                      cookiesize);
 367                 if (rc) {
 368                         CERROR("obd_init_ea_size() failed on MDT target %d, "
 369                                "error %d.\n", i, rc);
 370                         break;
 371                 }
 372         }
 373         RETURN(rc);
 374 }
 375
 376 #define MAX_STRING_SIZE 128
 377
 378 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 379 {
 380 #ifdef __KERNEL__
 381         struct proc_dir_entry   *lmv_proc_dir;
 382 #endif
 383         struct lmv_obd          *lmv = &obd->u.lmv;
 384         struct obd_uuid         *cluuid = &lmv->cluuid;
 385         struct obd_connect_data *mdc_data = NULL;
 386         struct obd_uuid          lmv_mdc_uuid = { "LMV_MDC_UUID" };
 387         struct obd_device       *mdc_obd;
 388         struct obd_export       *mdc_exp;
 389         struct lu_fld_target     target;
 390         int                      rc;
 391         ENTRY;
 392
 393         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 394                                         &obd->obd_uuid);
 395         if (!mdc_obd) {
 396                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 397                 RETURN(-EINVAL);
 398         }
 399
 400         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
 401                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 402                 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
 403                 cluuid->uuid);
 404
 405         if (!mdc_obd->obd_set_up) {
 406                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 407                 RETURN(-EINVAL);
 408         }
 409
 410         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &lmv_mdc_uuid,
 411                          &lmv->conn_data, NULL);
 412         if (rc) {
 413                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 414                 RETURN(rc);
 415         }
 416
 417         /*
 418          * Init fid sequence client for this mdc and add new fld target.
 419          */
 420         rc = obd_fid_init(mdc_exp);
 421         if (rc)
 422                 RETURN(rc);
 423
 424         target.ft_srv = NULL;
 425         target.ft_exp = mdc_exp;
 426         target.ft_idx = tgt->ltd_idx;
 427
 428         fld_client_add_target(&lmv->lmv_fld, &target);
 429
 430         mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data;
 431
 432         rc = obd_register_observer(mdc_obd, obd);
 433         if (rc) {
 434                 obd_disconnect(mdc_exp);
 435                 CERROR("target %s register_observer error %d\n",
 436                        tgt->ltd_uuid.uuid, rc);
 437                 RETURN(rc);
 438         }
 439
 440         if (obd->obd_observer) {
 441                 /*
 442                  * Tell the observer about the new target.
 443                  */
 444                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 445                                 OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts));
 446                 if (rc) {
 447                         obd_disconnect(mdc_exp);
 448                         RETURN(rc);
 449                 }
 450         }
 451
 452         tgt->ltd_active = 1;
 453         tgt->ltd_exp = mdc_exp;
 454         lmv->desc.ld_active_tgt_count++;
 455
 456         /*
 457          * Copy connect data, it may be used later.
 458          */
 459         lmv->datas[tgt->ltd_idx] = *mdc_data;
 460
 461         md_init_ea_size(tgt->ltd_exp, lmv->max_easize,
 462                         lmv->max_def_easize, lmv->max_cookiesize);
 463
 464         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 465                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 466                 atomic_read(&obd->obd_refcount));
 467
 468 #ifdef __KERNEL__
 469         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 470         if (lmv_proc_dir) {
 471                 struct proc_dir_entry *mdc_symlink;
 472                 char name[MAX_STRING_SIZE + 1];
 473
 474                 LASSERT(mdc_obd->obd_type != NULL);
 475                 LASSERT(mdc_obd->obd_type->typ_name != NULL);
 476                 name[MAX_STRING_SIZE] = '\0';
 477                 snprintf(name, MAX_STRING_SIZE, "../../../%s/%s",
 478                          mdc_obd->obd_type->typ_name,
 479                          mdc_obd->obd_name);
 480                 mdc_symlink = proc_symlink(mdc_obd->obd_name,
 481                                            lmv_proc_dir, name);
 482                 if (mdc_symlink == NULL) {
 483                         CERROR("Could not register LMV target "
 484                                "/proc/fs/lustre/%s/%s/target_obds/%s.",
 485                                obd->obd_type->typ_name, obd->obd_name,
 486                                mdc_obd->obd_name);
 487                         lprocfs_remove(&lmv_proc_dir);
 488                         lmv_proc_dir = NULL;
 489                 }
 490         }
 491 #endif
 492         RETURN(0);
 493 }
 494
 495 int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid)
 496 {
 497         struct lmv_obd      *lmv = &obd->u.lmv;
 498         struct lmv_tgt_desc *tgt;
 499         int                  rc = 0;
 500         ENTRY;
 501
 502         CDEBUG(D_CONFIG, "Target uuid: %s.\n", tgt_uuid->uuid);
 503
 504         lmv_init_lock(lmv);
 505
 506         if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) {
 507                 lmv_init_unlock(lmv);
 508                 CERROR("Can't add %s, LMV module compiled for %d MDCs. "
 509                        "That many MDCs already configured.\n",
 510                        tgt_uuid->uuid, LMV_MAX_TGT_COUNT);
 511                 RETURN(-EINVAL);
 512         }
 513         if (lmv->desc.ld_tgt_count == 0) {
 514                 struct obd_device *mdc_obd;
 515
 516                 mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME,
 517                                                 &obd->obd_uuid);
 518                 if (!mdc_obd) {
 519                         lmv_init_unlock(lmv);
 520                         CERROR("Target %s not attached\n", tgt_uuid->uuid);
 521                         RETURN(-EINVAL);
 522                 }
 523
 524                 rc = obd_llog_init(obd, &obd->obd_olg, mdc_obd, 0, NULL, tgt_uuid);
 525                 if (rc) {
 526                         lmv_init_unlock(lmv);
 527                         CERROR("lmv failed to setup llogging subsystems\n");
 528                 }
 529         }
 530         spin_lock(&lmv->lmv_lock);
 531         tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
 532         tgt->ltd_uuid = *tgt_uuid;
 533         spin_unlock(&lmv->lmv_lock);
 534
 535         if (lmv->connected) {
 536                 rc = lmv_connect_mdc(obd, tgt);
 537                 if (rc) {
 538                         spin_lock(&lmv->lmv_lock);
 539                         lmv->desc.ld_tgt_count--;
 540                         memset(tgt, 0, sizeof(*tgt));
 541                         spin_unlock(&lmv->lmv_lock);
 542                 } else {
 543                         int easize = sizeof(struct lmv_stripe_md) +
 544                                      lmv->desc.ld_tgt_count *
 545                                      sizeof(struct lu_fid);
 546                         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 547                 }
 548         }
 549
 550         lmv_init_unlock(lmv);
 551         RETURN(rc);
 552 }
 553
 554 int lmv_check_connect(struct obd_device *obd)
 555 {
 556         struct lmv_obd       *lmv = &obd->u.lmv;
 557         struct lmv_tgt_desc  *tgt;
 558         int                   i;
 559         int                   rc;
 560         int                   easize;
 561         ENTRY;
 562
 563         if (lmv->connected)
 564                 RETURN(0);
 565
 566         lmv_init_lock(lmv);
 567         if (lmv->connected) {
 568                 lmv_init_unlock(lmv);
 569                 RETURN(0);
 570         }
 571
 572         if (lmv->desc.ld_tgt_count == 0) {
 573                 CERROR("%s: no targets configured.\n", obd->obd_name);
 574                 RETURN(-EINVAL);
 575         }
 576
 577         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 578                lmv->cluuid.uuid, obd->obd_name);
 579
 580         LASSERT(lmv->tgts != NULL);
 581
 582         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
 583                 rc = lmv_connect_mdc(obd, tgt);
 584                 if (rc)
 585                         GOTO(out_disc, rc);
 586         }
 587
 588         lmv_set_timeouts(obd);
 589         class_export_put(lmv->exp);
 590         lmv->connected = 1;
 591         easize = lmv_get_easize(lmv);
 592         lmv_init_ea_size(obd->obd_self_export, easize, 0, 0);
 593         lmv_init_unlock(lmv);
 594         RETURN(0);
 595
 596  out_disc:
 597         while (i-- > 0) {
 598                 int rc2;
 599                 --tgt;
 600                 tgt->ltd_active = 0;
 601                 if (tgt->ltd_exp) {
 602                         --lmv->desc.ld_active_tgt_count;
 603                         rc2 = obd_disconnect(tgt->ltd_exp);
 604                         if (rc2) {
 605                                 CERROR("LMV target %s disconnect on "
 606                                        "MDC idx %d: error %d\n",
 607                                        tgt->ltd_uuid.uuid, i, rc2);
 608                         }
 609                 }
 610         }
 611         class_disconnect(lmv->exp);
 612         lmv_init_unlock(lmv);
 613         RETURN(rc);
 614 }
 615
 616 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 617 {
 618 #ifdef __KERNEL__
 619         struct proc_dir_entry  *lmv_proc_dir;
 620 #endif
 621         struct lmv_obd         *lmv = &obd->u.lmv;
 622         struct obd_device      *mdc_obd;
 623         int                     rc;
 624         ENTRY;
 625
 626         LASSERT(tgt != NULL);
 627         LASSERT(obd != NULL);
 628
 629         mdc_obd = class_exp2obd(tgt->ltd_exp);
 630
 631         if (mdc_obd)
 632                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 633
 634 #ifdef __KERNEL__
 635         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 636         if (lmv_proc_dir) {
 637                 struct proc_dir_entry *mdc_symlink;
 638
 639                 mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name);
 640                 if (mdc_symlink) {
 641                         lprocfs_remove(&mdc_symlink);
 642                 } else {
 643                         CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n",
 644                                obd->obd_type->typ_name, obd->obd_name,
 645                                mdc_obd->obd_name);
 646                 }
 647         }
 648 #endif
 649         rc = obd_fid_fini(tgt->ltd_exp);
 650         if (rc)
 651                 CERROR("Can't finanize fids factory\n");
 652
 653         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 654                tgt->ltd_exp->exp_obd->obd_name,
 655                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 656
 657         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 658         rc = obd_disconnect(tgt->ltd_exp);
 659         if (rc) {
 660                 if (tgt->ltd_active) {
 661                         CERROR("Target %s disconnect error %d\n",
 662                                tgt->ltd_uuid.uuid, rc);
 663                 }
 664         }
 665
 666         lmv_activate_target(lmv, tgt, 0);
 667         tgt->ltd_exp = NULL;
 668         RETURN(0);
 669 }
 670
 671 static int lmv_disconnect(struct obd_export *exp)
 672 {
 673         struct obd_device     *obd = class_exp2obd(exp);
 674 #ifdef __KERNEL__
 675         struct proc_dir_entry *lmv_proc_dir;
 676 #endif
 677         struct lmv_obd        *lmv = &obd->u.lmv;
 678         int                    rc;
 679         int                    i;
 680         ENTRY;
 681
 682         if (!lmv->tgts)
 683                 goto out_local;
 684
 685         /*
 686          * Only disconnect the underlying layers on the final disconnect.
 687          */
 688         lmv->refcount--;
 689         if (lmv->refcount != 0)
 690                 goto out_local;
 691
 692         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 693                 if (lmv->tgts[i].ltd_exp == NULL)
 694                         continue;
 695                 lmv_disconnect_mdc(obd, &lmv->tgts[i]);
 696         }
 697
 698 #ifdef __KERNEL__
 699         lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 700         if (lmv_proc_dir) {
 701                 lprocfs_remove(&lmv_proc_dir);
 702         } else {
 703                 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
 704                        obd->obd_type->typ_name, obd->obd_name);
 705         }
 706 #endif
 707
 708 out_local:
 709         /*
 710          * This is the case when no real connection is established by
 711          * lmv_check_connect().
 712          */
 713         if (!lmv->connected)
 714                 class_export_put(exp);
 715         rc = class_disconnect(exp);
 716         if (lmv->refcount == 0)
 717                 lmv->connected = 0;
 718         RETURN(rc);
 719 }
 720
 721 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 722                          int len, void *karg, void *uarg)
 723 {
 724         struct obd_device    *obddev = class_exp2obd(exp);
 725         struct lmv_obd       *lmv = &obddev->u.lmv;
 726         int                   i = 0;
 727         int                   rc = 0;
 728         int                   set = 0;
 729         int                   count = lmv->desc.ld_tgt_count;
 730         ENTRY;
 731
 732         if (count == 0)
 733                 RETURN(-ENOTTY);
 734
 735         switch (cmd) {
 736         case IOC_OBD_STATFS: {
 737                 struct obd_ioctl_data *data = karg;
 738                 struct obd_device *mdc_obd;
 739                 struct obd_statfs stat_buf = {0};
 740                 __u32 index;
 741
 742                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 743                 LASSERT(data->ioc_plen1 == sizeof(struct obd_statfs));
 744
 745                 if ((index >= count))
 746                         RETURN(-ENODEV);
 747
 748                 if (!lmv->tgts[index].ltd_active)
 749                         RETURN(-ENODATA);
 750
 751                 mdc_obd = class_exp2obd(lmv->tgts[index].ltd_exp);
 752                 if (!mdc_obd)
 753                         RETURN(-EINVAL);
 754
 755                 rc = obd_statfs(mdc_obd, &stat_buf,
 756                                 cfs_time_current_64() - HZ, 0);
 757                 if (rc)
 758                         RETURN(rc);
 759                 if (copy_to_user(data->ioc_pbuf1, &stat_buf, data->ioc_plen1))
 760                         RETURN(-EFAULT);
 761                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 762                                   data->ioc_plen2))
 763                         RETURN(-EFAULT);
 764                 break;
 765         }
 766         case OBD_IOC_QUOTACTL: {
 767                 struct if_quotactl *qctl = karg;
 768                 struct lmv_tgt_desc *tgt = NULL;
 769                 struct obd_quotactl *oqctl;
 770
 771                 if (qctl->qc_valid == QC_MDTIDX) {
 772                         if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
 773                                 RETURN(-EINVAL);
 774
 775                         tgt = &lmv->tgts[qctl->qc_idx];
 776                         if (!tgt->ltd_exp)
 777                                 RETURN(-EINVAL);
 778                 } else if (qctl->qc_valid == QC_UUID) {
 779                         for (i = 0; i < count; i++) {
 780                                 tgt = &lmv->tgts[i];
 781                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 782                                                      &qctl->obd_uuid))
 783                                         continue;
 784
 785                                 if (tgt->ltd_exp == NULL)
 786                                         RETURN(-EINVAL);
 787
 788                                 break;
 789                         }
 790                 } else {
 791                         RETURN(-EINVAL);
 792                 }
 793
 794                 if (i >= count)
 795                         RETURN(-EAGAIN);
 796
 797                 LASSERT(tgt && tgt->ltd_exp);
 798                 OBD_ALLOC_PTR(oqctl);
 799                 if (!oqctl)
 800                         RETURN(-ENOMEM);
 801
 802                 QCTL_COPY(oqctl, qctl);
 803                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 804                 if (rc == 0) {
 805                         QCTL_COPY(qctl, oqctl);
 806                         qctl->qc_valid = QC_MDTIDX;
 807                         qctl->obd_uuid = tgt->ltd_uuid;
 808                 }
 809                 OBD_FREE_PTR(oqctl);
 810                 break;
 811         }
 812         default : {
 813                 for (i = 0; i < count; i++) {
 814                         int err;
 815
 816                         if (lmv->tgts[i].ltd_exp == NULL)
 817                                 continue;
 818
 819                         err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
 820                                             karg, uarg);
 821                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
 822                                 RETURN(err);
 823                         } else if (err) {
 824                                 if (lmv->tgts[i].ltd_active) {
 825                                         CERROR("error: iocontrol MDC %s on MDT"
 826                                                "idx %d cmd %x: err = %d\n",
 827                                                 lmv->tgts[i].ltd_uuid.uuid,
 828                                                 i, cmd, err);
 829                                         if (!rc)
 830                                                 rc = err;
 831                                 }
 832                         } else
 833                                 set = 1;
 834                 }
 835                 if (!set && !rc)
 836                         rc = -EIO;
 837         }
 838         }
 839         RETURN(rc);
 840 }
 841
 842 static int lmv_all_chars_policy(int count, const char *name,
 843                                 int len)
 844 {
 845         unsigned int c = 0;
 846
 847         while (len > 0)
 848                 c += name[--len];
 849         c = c % count;
 850         return c;
 851 }
 852
 853 static int lmv_nid_policy(struct lmv_obd *lmv)
 854 {
 855         struct obd_import *imp;
 856         __u32              id;
 857
 858         /*
 859          * XXX: To get nid we assume that underlying obd device is mdc.
 860          */
 861         imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
 862         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
 863         return id % lmv->desc.ld_tgt_count;
 864 }
 865
 866 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
 867                           placement_policy_t placement)
 868 {
 869         switch (placement) {
 870         case PLACEMENT_CHAR_POLICY:
 871                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
 872                                             op_data->op_name,
 873                                             op_data->op_namelen);
 874         case PLACEMENT_NID_POLICY:
 875                 return lmv_nid_policy(lmv);
 876
 877         default:
 878                 break;
 879         }
 880
 881         CERROR("Unsupported placement policy %x\n", placement);
 882         return -EINVAL;
 883 }
 884
 885 /**
 886  * This is _inode_ placement policy function (not name).
 887  */
 888 static int lmv_placement_policy(struct obd_device *obd,
 889                                 struct md_op_data *op_data,
 890                                 mdsno_t *mds)
 891 {
 892         struct lmv_obd          *lmv = &obd->u.lmv;
 893         struct lmv_object       *obj;
 894         int                      rc;
 895         ENTRY;
 896
 897         LASSERT(mds != NULL);
 898
 899         if (lmv->desc.ld_tgt_count == 1) {
 900                 *mds = 0;
 901                 RETURN(0);
 902         }
 903
 904         /*
 905          * Allocate new fid on target according to operation type and parent
 906          * home mds.
 907          */
 908         obj = lmv_object_find(obd, &op_data->op_fid1);
 909         if (obj != NULL || op_data->op_name == NULL ||
 910             op_data->op_opc != LUSTRE_OPC_MKDIR) {
 911                 /*
 912                  * Allocate fid for non-dir or for null name or for case parent
 913                  * dir is split.
 914                  */
 915                 if (obj) {
 916                         lmv_object_put(obj);
 917
 918                         /*
 919                          * If we have this flag turned on, and we see that
 920                          * parent dir is split, this means, that caller did not
 921                          * notice split yet. This is race and we would like to
 922                          * let caller know that.
 923                          */
 924                         if (op_data->op_bias & MDS_CHECK_SPLIT)
 925                                 RETURN(-ERESTART);
 926                 }
 927
 928                 /*
 929                  * Allocate new fid on same mds where parent fid is located and
 930                  * where operation will be sent. In case of split dir, ->op_fid1
 931                  * and ->op_mds here will contain fid and mds of slave directory
 932                  * object (assigned by caller).
 933                  */
 934                 *mds = op_data->op_mds;
 935                 rc = 0;
 936         } else {
 937                 /*
 938                  * Parent directory is not split and we want to create a
 939                  * directory in it. Let's calculate where to place it according
 940                  * to operation data @op_data.
 941                  */
 942                 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
 943                 rc = 0;
 944         }
 945
 946         if (rc) {
 947                 CERROR("Can't choose MDS, err = %d\n", rc);
 948         } else {
 949                 LASSERT(*mds < lmv->desc.ld_tgt_count);
 950         }
 951
 952         RETURN(rc);
 953 }
 954
 955 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
 956                     mdsno_t mds)
 957 {
 958         struct lmv_tgt_desc *tgt;
 959         int                  rc;
 960         ENTRY;
 961
 962         tgt = lmv_get_target(lmv, mds);
 963
 964         /*
 965          * New seq alloc and FLD setup should be atomic. Otherwise we may find
 966          * on server that seq in new allocated fid is not yet known.
 967          */
 968         down(&tgt->ltd_fid_sem);
 969
 970         if (!tgt->ltd_active)
 971                 GOTO(out, rc = -ENODEV);
 972
 973         /*
 974          * Asking underlaying tgt layer to allocate new fid.
 975          */
 976         rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
 977         if (rc > 0) {
 978                 LASSERT(fid_is_sane(fid));
 979                 rc = 0;
 980         }
 981
 982         EXIT;
 983 out:
 984         up(&tgt->ltd_fid_sem);
 985         return rc;
 986 }
 987
 988 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
 989                   struct md_op_data *op_data)
 990 {
 991         struct obd_device     *obd = class_exp2obd(exp);
 992         struct lmv_obd        *lmv = &obd->u.lmv;
 993         mdsno_t                mds;
 994         int                    rc;
 995         ENTRY;
 996
 997         LASSERT(op_data != NULL);
 998         LASSERT(fid != NULL);
 999
1000         rc = lmv_placement_policy(obd, op_data, &mds);
1001         if (rc) {
1002                 CERROR("Can't get target for allocating fid, "
1003                        "rc %d\n", rc);
1004                 RETURN(rc);
1005         }
1006
1007         rc = __lmv_fid_alloc(lmv, fid, mds);
1008         if (rc) {
1009                 CERROR("Can't alloc new fid, rc %d\n", rc);
1010                 RETURN(rc);
1011         }
1012
1013         RETURN(rc);
1014 }
1015
1016 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1017 {
1018         ENTRY;
1019         LASSERT(exp != NULL && fid != NULL);
1020         if (lmv_object_delete(exp, fid)) {
1021                 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1022                        PFID(fid));
1023         }
1024         RETURN(0);
1025 }
1026
1027 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1028 {
1029         struct lmv_obd             *lmv = &obd->u.lmv;
1030         struct lprocfs_static_vars  lvars;
1031         struct lmv_desc            *desc;
1032         int                         rc;
1033         int                         i = 0;
1034         ENTRY;
1035
1036         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1037                 CERROR("LMV setup requires a descriptor\n");
1038                 RETURN(-EINVAL);
1039         }
1040
1041         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1042         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1043                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1044                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1045                 RETURN(-EINVAL);
1046         }
1047
1048         lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1049
1050         OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1051         if (lmv->tgts == NULL)
1052                 RETURN(-ENOMEM);
1053
1054         for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1055                 sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
1056                 lmv->tgts[i].ltd_idx = i;
1057         }
1058
1059         lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1060
1061         OBD_ALLOC(lmv->datas, lmv->datas_size);
1062         if (lmv->datas == NULL)
1063                 GOTO(out_free_tgts, rc = -ENOMEM);
1064
1065         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1066         lmv->desc.ld_tgt_count = 0;
1067         lmv->desc.ld_active_tgt_count = 0;
1068         lmv->max_cookiesize = 0;
1069         lmv->max_def_easize = 0;
1070         lmv->max_easize = 0;
1071         lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1072
1073         spin_lock_init(&lmv->lmv_lock);
1074         sema_init(&lmv->init_sem, 1);
1075
1076         rc = lmv_object_setup(obd);
1077         if (rc) {
1078                 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1079                 GOTO(out_free_datas, rc);
1080         }
1081
1082         lprocfs_lmv_init_vars(&lvars);
1083         lprocfs_obd_setup(obd, lvars.obd_vars);
1084 #ifdef LPROCFS
1085         {
1086                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1087                                         0444, &lmv_proc_target_fops, obd);
1088                 if (rc)
1089                         CWARN("Error adding target_obd_stats file (%d)\n", rc);
1090        }
1091 #endif
1092         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1093                              LUSTRE_CLI_FLD_HASH_DHT);
1094         if (rc) {
1095                 CERROR("Can't init FLD, err %d\n", rc);
1096                 GOTO(out_free_datas, rc);
1097         }
1098
1099         RETURN(0);
1100
1101 out_free_datas:
1102         OBD_FREE(lmv->datas, lmv->datas_size);
1103         lmv->datas = NULL;
1104 out_free_tgts:
1105         OBD_FREE(lmv->tgts, lmv->tgts_size);
1106         lmv->tgts = NULL;
1107         return rc;
1108 }
1109
1110 static int lmv_cleanup(struct obd_device *obd)
1111 {
1112         struct lmv_obd   *lmv = &obd->u.lmv;
1113         ENTRY;
1114
1115         fld_client_fini(&lmv->lmv_fld);
1116         lprocfs_obd_cleanup(obd);
1117         lmv_object_cleanup(obd);
1118         OBD_FREE(lmv->datas, lmv->datas_size);
1119         OBD_FREE(lmv->tgts, lmv->tgts_size);
1120
1121         RETURN(0);
1122 }
1123
1124 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1125 {
1126         struct lustre_cfg     *lcfg = buf;
1127         struct obd_uuid        tgt_uuid;
1128         int                    rc;
1129         ENTRY;
1130
1131         switch(lcfg->lcfg_command) {
1132         case LCFG_ADD_MDC:
1133                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1134                         GOTO(out, rc = -EINVAL);
1135
1136                 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1137                 rc = lmv_add_target(obd, &tgt_uuid);
1138                 GOTO(out, rc);
1139         default: {
1140                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1141                 GOTO(out, rc = -EINVAL);
1142         }
1143         }
1144 out:
1145         RETURN(rc);
1146 }
1147
1148 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1149                       __u64 max_age, __u32 flags)
1150 {
1151         struct lmv_obd        *lmv = &obd->u.lmv;
1152         struct obd_statfs     *temp;
1153         int                    rc = 0;
1154         int                    i;
1155         ENTRY;
1156
1157         rc = lmv_check_connect(obd);
1158         if (rc)
1159                 RETURN(rc);
1160
1161         OBD_ALLOC(temp, sizeof(*temp));
1162         if (temp == NULL)
1163                 RETURN(-ENOMEM);
1164
1165         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1166                 if (lmv->tgts[i].ltd_exp == NULL)
1167                         continue;
1168
1169                 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1170                                 max_age, flags);
1171                 if (rc) {
1172                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1173                                lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1174                                rc);
1175                         GOTO(out_free_temp, rc);
1176                 }
1177                 if (i == 0) {
1178                         *osfs = *temp;
1179                 } else {
1180                         osfs->os_bavail += temp->os_bavail;
1181                         osfs->os_blocks += temp->os_blocks;
1182                         osfs->os_ffree += temp->os_ffree;
1183                         osfs->os_files += temp->os_files;
1184                 }
1185         }
1186
1187         EXIT;
1188 out_free_temp:
1189         OBD_FREE(temp, sizeof(*temp));
1190         return rc;
1191 }
1192
1193 static int lmv_getstatus(struct obd_export *exp,
1194                          struct lu_fid *fid,
1195                          struct obd_capa **pc)
1196 {
1197         struct obd_device    *obd = exp->exp_obd;
1198         struct lmv_obd       *lmv = &obd->u.lmv;
1199         int                   rc;
1200         ENTRY;
1201
1202         rc = lmv_check_connect(obd);
1203         if (rc)
1204                 RETURN(rc);
1205
1206         rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1207         RETURN(rc);
1208 }
1209
1210 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1211                         struct obd_capa *oc, obd_valid valid, const char *name,
1212                         const char *input, int input_size, int output_size,
1213                         int flags, struct ptlrpc_request **request)
1214 {
1215         struct obd_device      *obd = exp->exp_obd;
1216         struct lmv_obd         *lmv = &obd->u.lmv;
1217         struct lmv_tgt_desc    *tgt;
1218         int                     rc;
1219         ENTRY;
1220
1221         rc = lmv_check_connect(obd);
1222         if (rc)
1223                 RETURN(rc);
1224
1225         tgt = lmv_find_target(lmv, fid);
1226         if (IS_ERR(tgt))
1227                 RETURN(PTR_ERR(tgt));
1228
1229         rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1230                          input_size, output_size, flags, request);
1231
1232         RETURN(rc);
1233 }
1234
1235 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1236                         struct obd_capa *oc, obd_valid valid, const char *name,
1237                         const char *input, int input_size, int output_size,
1238                         int flags, __u32 suppgid,
1239                         struct ptlrpc_request **request)
1240 {
1241         struct obd_device      *obd = exp->exp_obd;
1242         struct lmv_obd         *lmv = &obd->u.lmv;
1243         struct lmv_tgt_desc    *tgt;
1244         int                     rc;
1245         ENTRY;
1246
1247         rc = lmv_check_connect(obd);
1248         if (rc)
1249                 RETURN(rc);
1250
1251         tgt = lmv_find_target(lmv, fid);
1252         if (IS_ERR(tgt))
1253                 RETURN(PTR_ERR(tgt));
1254
1255         rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1256                          input_size, output_size, flags, suppgid,
1257                          request);
1258
1259         RETURN(rc);
1260 }
1261
1262 static int lmv_getattr(struct obd_export *exp, const struct lu_fid *fid,
1263                        struct obd_capa *oc, obd_valid valid, int ea_size,
1264                        struct ptlrpc_request **request)
1265 {
1266         struct obd_device       *obd = exp->exp_obd;
1267         struct lmv_obd          *lmv = &obd->u.lmv;
1268         struct lmv_tgt_desc     *tgt;
1269         struct lmv_object       *obj;
1270         int                      rc;
1271         int                      i;
1272         ENTRY;
1273
1274         rc = lmv_check_connect(obd);
1275         if (rc)
1276                 RETURN(rc);
1277
1278         tgt = lmv_find_target(lmv, fid);
1279         if (IS_ERR(tgt))
1280                 RETURN(PTR_ERR(tgt));
1281
1282         rc = md_getattr(tgt->ltd_exp, fid, oc, valid, ea_size, request);
1283         if (rc)
1284                 RETURN(rc);
1285
1286         obj = lmv_object_find_lock(obd, fid);
1287
1288         CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(fid),
1289                obj ? "(split)" : "");
1290
1291         /*
1292          * If object is split, then we loop over all the slaves and gather size
1293          * attribute. In ideal world we would have to gather also mds field from
1294          * all slaves, as object is spread over the cluster and this is
1295          * definitely interesting information and it is not good to loss it,
1296          * but...
1297          */
1298         if (obj) {
1299                 struct mdt_body *body;
1300
1301                 if (*request == NULL) {
1302                         lmv_object_put(obj);
1303                         RETURN(rc);
1304                 }
1305
1306                 body = req_capsule_server_get(&(*request)->rq_pill,
1307                                               &RMF_MDT_BODY);
1308                 LASSERT(body != NULL);
1309
1310                 for (i = 0; i < obj->lo_objcount; i++) {
1311                         if (lmv->tgts[i].ltd_exp == NULL) {
1312                                 CWARN("%s: NULL export for %d\n",
1313                                       obd->obd_name, i);
1314                                 continue;
1315                         }
1316
1317                         /*
1318                          * Skip master object.
1319                          */
1320                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1321                                 continue;
1322
1323                         body->size += obj->lo_stripes[i].ls_size;
1324                 }
1325
1326                 lmv_object_put_unlock(obj);
1327         }
1328
1329         RETURN(rc);
1330 }
1331
1332 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1333                              ldlm_iterator_t it, void *data)
1334 {
1335         struct obd_device   *obd = exp->exp_obd;
1336         struct lmv_obd      *lmv = &obd->u.lmv;
1337         int                  i;
1338         int                  rc;
1339         ENTRY;
1340
1341         rc = lmv_check_connect(obd);
1342         if (rc)
1343                 RETURN(rc);
1344
1345         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1346
1347         /*
1348          * With CMD every object can have two locks in different namespaces:
1349          * lookup lock in space of mds storing direntry and update/open lock in
1350          * space of mds storing inode.
1351          */
1352         for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1353                 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1354
1355         RETURN(0);
1356 }
1357
1358 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1359                      struct md_open_data *mod, struct ptlrpc_request **request)
1360 {
1361         struct obd_device     *obd = exp->exp_obd;
1362         struct lmv_obd        *lmv = &obd->u.lmv;
1363         struct lmv_tgt_desc   *tgt;
1364         int                    rc;
1365         ENTRY;
1366
1367         rc = lmv_check_connect(obd);
1368         if (rc)
1369                 RETURN(rc);
1370
1371         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1372         if (IS_ERR(tgt))
1373                 RETURN(PTR_ERR(tgt));
1374
1375         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1376         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1377         RETURN(rc);
1378 }
1379
1380 /**
1381  * Called in the case MDS returns -ERESTART on create on open, what means that
1382  * directory is split and its LMV presentation object has to be updated.
1383  */
1384 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1385 {
1386         struct obd_device       *obd = exp->exp_obd;
1387         struct lmv_obd          *lmv = &obd->u.lmv;
1388         struct ptlrpc_request   *req = NULL;
1389         struct lmv_tgt_desc     *tgt;
1390         struct lmv_object       *obj;
1391         struct lustre_md         md;
1392         int                      mealen;
1393         int                      rc;
1394         __u64                    valid;
1395         ENTRY;
1396
1397         md.mea = NULL;
1398         mealen = lmv_get_easize(lmv);
1399
1400         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1401
1402         tgt = lmv_find_target(lmv, fid);
1403         if (IS_ERR(tgt))
1404                 RETURN(PTR_ERR(tgt));
1405
1406         /*
1407          * Time to update mea of parent fid.
1408          */
1409         rc = md_getattr(tgt->ltd_exp, fid, NULL, valid, mealen, &req);
1410         if (rc) {
1411                 CERROR("md_getattr() failed, error %d\n", rc);
1412                 GOTO(cleanup, rc);
1413         }
1414
1415         rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1416         if (rc) {
1417                 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1418                 GOTO(cleanup, rc);
1419         }
1420
1421         if (md.mea == NULL)
1422                 GOTO(cleanup, rc = -ENODATA);
1423
1424         obj = lmv_object_create(exp, fid, md.mea);
1425         if (IS_ERR(obj))
1426                 rc = PTR_ERR(obj);
1427         else
1428                 lmv_object_put(obj);
1429
1430         obd_free_memmd(exp, (void *)&md.mea);
1431         EXIT;
1432 cleanup:
1433         if (req)
1434                 ptlrpc_req_finished(req);
1435         return rc;
1436 }
1437
1438 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1439                const void *data, int datalen, int mode, __u32 uid,
1440                __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1441                struct ptlrpc_request **request)
1442 {
1443         struct obd_device       *obd = exp->exp_obd;
1444         struct lmv_obd          *lmv = &obd->u.lmv;
1445         struct lmv_tgt_desc     *tgt;
1446         struct lmv_object       *obj;
1447         int                      rc;
1448         int                      loop = 0;
1449         int                      sidx;
1450         ENTRY;
1451
1452         rc = lmv_check_connect(obd);
1453         if (rc)
1454                 RETURN(rc);
1455
1456         if (!lmv->desc.ld_active_tgt_count)
1457                 RETURN(-EIO);
1458 repeat:
1459         ++loop;
1460         LASSERT(loop <= 2);
1461
1462         obj = lmv_object_find(obd, &op_data->op_fid1);
1463         if (obj) {
1464                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1465                                        op_data->op_name, op_data->op_namelen);
1466                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1467                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1468                 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1469                 tgt = lmv_get_target(lmv, op_data->op_mds);
1470                 lmv_object_put(obj);
1471         } else {
1472                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1473                 op_data->op_bias |= MDS_CHECK_SPLIT;
1474                 op_data->op_mds = tgt->ltd_idx;
1475         }
1476
1477         if (IS_ERR(tgt))
1478                 RETURN(PTR_ERR(tgt));
1479
1480         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1481         if (rc == -ERESTART)
1482                 goto repeat;
1483         else if (rc)
1484                 RETURN(rc);
1485
1486         CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #%x\n",
1487                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1488                op_data->op_mds);
1489
1490         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1491         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1492                        cap_effective, rdev, request);
1493         if (rc == 0) {
1494                 if (*request == NULL)
1495                         RETURN(rc);
1496                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1497         } else if (rc == -ERESTART) {
1498                 LASSERT(*request != NULL);
1499                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1500                           "Got -ERESTART during create!\n");
1501                 ptlrpc_req_finished(*request);
1502                 *request = NULL;
1503
1504                 /*
1505                  * Directory got split. Time to update local object and repeat
1506                  * the request with proper MDS.
1507                  */
1508                 rc = lmv_handle_split(exp, &op_data->op_fid1);
1509                 if (rc == 0) {
1510                         rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1511                                                  op_data, &op_data->op_fid2);
1512                         if (rc)
1513                                 RETURN(rc);
1514                         goto repeat;
1515                 }
1516         }
1517         RETURN(rc);
1518 }
1519
1520 static int lmv_done_writing(struct obd_export *exp,
1521                             struct md_op_data *op_data,
1522                             struct md_open_data *mod)
1523 {
1524         struct obd_device     *obd = exp->exp_obd;
1525         struct lmv_obd        *lmv = &obd->u.lmv;
1526         struct lmv_tgt_desc   *tgt;
1527         int                    rc;
1528         ENTRY;
1529
1530         rc = lmv_check_connect(obd);
1531         if (rc)
1532                 RETURN(rc);
1533
1534         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1535         if (IS_ERR(tgt))
1536                 RETURN(PTR_ERR(tgt));
1537
1538         rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1539         RETURN(rc);
1540 }
1541
1542 static int
1543 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1544                    struct lookup_intent *it, struct md_op_data *op_data,
1545                    struct lustre_handle *lockh, void *lmm, int lmmsize)
1546 {
1547         struct obd_device     *obd = exp->exp_obd;
1548         struct lmv_obd        *lmv = &obd->u.lmv;
1549         struct lmv_stripe_md  *mea = op_data->op_mea1;
1550         struct md_op_data     *op_data2;
1551         struct lmv_tgt_desc   *tgt;
1552         int                    i;
1553         int                    rc = 0;
1554         ENTRY;
1555
1556         OBD_ALLOC_PTR(op_data2);
1557         if (op_data2 == NULL)
1558                 RETURN(-ENOMEM);
1559
1560         LASSERT(mea != NULL);
1561         for (i = 0; i < mea->mea_count; i++) {
1562                 memset(op_data2, 0, sizeof(*op_data2));
1563                 op_data2->op_fid1 = mea->mea_ids[i];
1564                 op_data2->op_bias = 0;
1565
1566                 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1567                 if (IS_ERR(tgt))
1568                         GOTO(cleanup, rc = PTR_ERR(tgt));
1569
1570                 if (tgt->ltd_exp == NULL)
1571                         continue;
1572
1573                 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1574                                 lockh + i, lmm, lmmsize, NULL, 0);
1575
1576                 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1577                        PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1578
1579                 if (rc)
1580                         GOTO(cleanup, rc);
1581
1582                 if (it->d.lustre.it_data) {
1583                         struct ptlrpc_request *req;
1584                         req = (struct ptlrpc_request *)it->d.lustre.it_data;
1585                         ptlrpc_req_finished(req);
1586                 }
1587
1588                 if (it->d.lustre.it_status)
1589                         GOTO(cleanup, rc = it->d.lustre.it_status);
1590         }
1591
1592         EXIT;
1593 cleanup:
1594         OBD_FREE_PTR(op_data2);
1595
1596         if (rc != 0) {
1597                 /*
1598                  * Drop all taken locks.
1599                  */
1600                 while (--i >= 0) {
1601                         if (lockh[i].cookie)
1602                                 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1603                         lockh[i].cookie = 0;
1604                 }
1605         }
1606         return rc;
1607 }
1608
1609 static int
1610 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1611                    struct lookup_intent *it, struct md_op_data *op_data,
1612                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1613                    int extra_lock_flags)
1614 {
1615         struct ptlrpc_request      *req = it->d.lustre.it_data;
1616         struct obd_device          *obd = exp->exp_obd;
1617         struct lmv_obd             *lmv = &obd->u.lmv;
1618         struct lustre_handle        plock;
1619         struct lmv_tgt_desc        *tgt;
1620         struct md_op_data          *rdata;
1621         struct lu_fid               fid1;
1622         struct mdt_body            *body;
1623         int                         rc = 0;
1624         int                         pmode;
1625         ENTRY;
1626
1627         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1628         LASSERT(body != NULL);
1629
1630         if (!(body->valid & OBD_MD_MDS))
1631                 RETURN(0);
1632
1633         CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1634                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1635
1636         /*
1637          * We got LOOKUP lock, but we really need attrs.
1638          */
1639         pmode = it->d.lustre.it_lock_mode;
1640         LASSERT(pmode != 0);
1641         memcpy(&plock, lockh, sizeof(plock));
1642         it->d.lustre.it_lock_mode = 0;
1643         it->d.lustre.it_data = NULL;
1644         fid1 = body->fid1;
1645
1646         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1647         ptlrpc_req_finished(req);
1648
1649         tgt = lmv_find_target(lmv, &fid1);
1650         if (IS_ERR(tgt))
1651                 GOTO(out, rc = PTR_ERR(tgt));
1652
1653         OBD_ALLOC_PTR(rdata);
1654         if (rdata == NULL)
1655                 GOTO(out, rc = -ENOMEM);
1656
1657         rdata->op_fid1 = fid1;
1658         rdata->op_bias = MDS_CROSS_REF;
1659
1660         rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1661                         lmm, lmmsize, NULL, extra_lock_flags);
1662         OBD_FREE_PTR(rdata);
1663         EXIT;
1664 out:
1665         ldlm_lock_decref(&plock, pmode);
1666         return rc;
1667 }
1668
1669 static int
1670 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1671             struct lookup_intent *it, struct md_op_data *op_data,
1672             struct lustre_handle *lockh, void *lmm, int lmmsize,
1673             struct ptlrpc_request **req, int extra_lock_flags)
1674 {
1675         struct obd_device        *obd = exp->exp_obd;
1676         struct lmv_obd           *lmv = &obd->u.lmv;
1677         struct lmv_tgt_desc      *tgt;
1678         struct lmv_object        *obj;
1679         int                       sidx;
1680         int                       rc;
1681         ENTRY;
1682
1683         rc = lmv_check_connect(obd);
1684         if (rc)
1685                 RETURN(rc);
1686
1687         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1688                LL_IT2STR(it), PFID(&op_data->op_fid1));
1689
1690         if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1691                 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1692                                         lockh, lmm, lmmsize);
1693                 RETURN(rc);
1694         }
1695
1696         obj = lmv_object_find(obd, &op_data->op_fid1);
1697         if (obj && op_data->op_namelen) {
1698                 sidx = raw_name2idx(obj->lo_hashtype,
1699                                        obj->lo_objcount,
1700                                        (char *)op_data->op_name,
1701                                        op_data->op_namelen);
1702                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1703                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1704         } else {
1705                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1706         }
1707         if (obj)
1708                 lmv_object_put(obj);
1709
1710         if (IS_ERR(tgt))
1711                 RETURN(PTR_ERR(tgt));
1712
1713         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1714                LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1715
1716         rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1717                         lmm, lmmsize, req, extra_lock_flags);
1718
1719         if (rc == 0 && it && it->it_op == IT_OPEN) {
1720                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1721                                         lmm, lmmsize, extra_lock_flags);
1722         }
1723         RETURN(rc);
1724 }
1725
1726 static int
1727 lmv_getattr_name(struct obd_export *exp, const struct lu_fid *fid,
1728                  struct obd_capa *oc, const char *name, int namelen,
1729                  obd_valid valid, int ea_size, __u32 suppgid,
1730                  struct ptlrpc_request **request)
1731 {
1732         struct ptlrpc_request   *req = NULL;
1733         struct obd_device       *obd = exp->exp_obd;
1734         struct lmv_obd          *lmv = &obd->u.lmv;
1735         struct lu_fid            rid = *fid;
1736         struct lmv_tgt_desc     *tgt;
1737         struct mdt_body         *body;
1738         struct lmv_object       *obj;
1739         int                      rc;
1740         int                      loop = 0;
1741         int                      sidx;
1742         ENTRY;
1743
1744         rc = lmv_check_connect(obd);
1745         if (rc)
1746                 RETURN(rc);
1747
1748 repeat:
1749         ++loop;
1750         LASSERT(loop <= 2);
1751         obj = lmv_object_find(obd, &rid);
1752         if (obj) {
1753                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1754                                        name, namelen - 1);
1755                 rid = obj->lo_stripes[sidx].ls_fid;
1756                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1757                 valid &= ~OBD_MD_FLCKSPLIT;
1758                 lmv_object_put(obj);
1759         } else {
1760                 tgt = lmv_find_target(lmv, &rid);
1761                 valid |= OBD_MD_FLCKSPLIT;
1762         }
1763         if (IS_ERR(tgt))
1764                 RETURN(PTR_ERR(tgt));
1765
1766         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1767                namelen, name, PFID(fid), PFID(&rid), tgt->ltd_idx);
1768
1769         rc = md_getattr_name(tgt->ltd_exp, &rid, oc, name, namelen, valid,
1770                              ea_size, suppgid, request);
1771         if (rc == 0) {
1772                 body = req_capsule_server_get(&(*request)->rq_pill,
1773                                               &RMF_MDT_BODY);
1774                 LASSERT(body != NULL);
1775
1776                 if (body->valid & OBD_MD_MDS) {
1777                         rid = body->fid1;
1778                         CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1779                                PFID(&rid));
1780
1781                         tgt = lmv_find_target(lmv, &rid);
1782                         if (IS_ERR(tgt)) {
1783                                 ptlrpc_req_finished(*request);
1784                                 RETURN(PTR_ERR(tgt));
1785                         }
1786
1787                         rc = md_getattr_name(tgt->ltd_exp, &rid, NULL, NULL,
1788                                              1, valid | OBD_MD_FLCROSSREF,
1789                                              ea_size, suppgid, &req);
1790                         ptlrpc_req_finished(*request);
1791                         *request = req;
1792                 }
1793         } else if (rc == -ERESTART) {
1794                 LASSERT(*request != NULL);
1795                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1796                           "Got -ERESTART during getattr!\n");
1797                 ptlrpc_req_finished(*request);
1798                 *request = NULL;
1799
1800                 /*
1801                  * Directory got split. Time to update local object and repeat
1802                  * the request with proper MDS.
1803                  */
1804                 rc = lmv_handle_split(exp, &rid);
1805                 if (rc == 0)
1806                         goto repeat;
1807         }
1808         RETURN(rc);
1809 }
1810
1811 #define md_op_data_fid(op_data, fl)                     \
1812         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1813          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1814          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1815          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1816          NULL)
1817
1818 static int lmv_early_cancel_slaves(struct obd_export *exp,
1819                                    struct md_op_data *op_data, int op_tgt,
1820                                    ldlm_mode_t mode, int bits, int flag)
1821 {
1822         struct obd_device      *obd = exp->exp_obd;
1823         struct lmv_obd         *lmv = &obd->u.lmv;
1824         ldlm_policy_data_t      policy = {{0}};
1825         struct lu_fid          *op_fid;
1826         struct lu_fid          *st_fid;
1827         struct lmv_tgt_desc    *tgt;
1828         struct lmv_object      *obj;
1829         int                     rc = 0;
1830         int                     i;
1831         ENTRY;
1832
1833         op_fid = md_op_data_fid(op_data, flag);
1834         if (!fid_is_sane(op_fid))
1835                 RETURN(0);
1836
1837         obj = lmv_object_find(obd, op_fid);
1838         if (obj == NULL)
1839                 RETURN(-EALREADY);
1840
1841         policy.l_inodebits.bits = bits;
1842         for (i = 0; i < obj->lo_objcount; i++) {
1843                 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1844                 st_fid = &obj->lo_stripes[i].ls_fid;
1845                 if (op_tgt != tgt->ltd_idx) {
1846                         CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1847                                PFID(st_fid), tgt->ltd_idx);
1848                         rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1849                                               mode, LDLM_FL_ASYNC, NULL);
1850                         if (rc)
1851                                 GOTO(out_put_obj, rc);
1852                 } else {
1853                         CDEBUG(D_INODE,
1854                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1855                                op_tgt, PFID(st_fid));
1856                         /*
1857                          * Do not cancel locks for operation target, they will
1858                          * be handled later in underlaying layer when calling
1859                          * function we run on behalf of.
1860                          */
1861                         *op_fid = *st_fid;
1862                         op_data->op_flags |= flag;
1863                 }
1864         }
1865         EXIT;
1866 out_put_obj:
1867         lmv_object_put(obj);
1868         return rc;
1869 }
1870
1871 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1872                             int op_tgt, ldlm_mode_t mode, int bits, int flag)
1873 {
1874         struct lu_fid          *fid = md_op_data_fid(op_data, flag);
1875         struct obd_device      *obd = exp->exp_obd;
1876         struct lmv_obd         *lmv = &obd->u.lmv;
1877         struct lmv_tgt_desc    *tgt;
1878         ldlm_policy_data_t      policy = {{0}};
1879         struct lmv_object      *obj;
1880         int                     rc = 0;
1881         ENTRY;
1882
1883         if (!fid_is_sane(fid))
1884                 RETURN(0);
1885
1886         obj = lmv_object_find(obd, fid);
1887         if (obj) {
1888                 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1889                                              bits, flag);
1890                 lmv_object_put(obj);
1891         } else {
1892                 tgt = lmv_find_target(lmv, fid);
1893                 if (IS_ERR(tgt))
1894                         RETURN(PTR_ERR(tgt));
1895
1896                 if (tgt->ltd_idx != op_tgt) {
1897                         CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1898                         policy.l_inodebits.bits = bits;
1899                         rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1900                                               mode, LDLM_FL_ASYNC, NULL);
1901                 } else {
1902                         CDEBUG(D_INODE,
1903                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1904                                op_tgt, PFID(fid));
1905                         op_data->op_flags |= flag;
1906                         rc = 0;
1907                 }
1908
1909         }
1910         RETURN(rc);
1911 }
1912
1913 /*
1914  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1915  * op_data->op_fid2
1916  */
1917 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1918                     struct ptlrpc_request **request)
1919 {
1920         struct obd_device       *obd = exp->exp_obd;
1921         struct lmv_obd          *lmv = &obd->u.lmv;
1922         struct lmv_tgt_desc     *tgt;
1923         struct lmv_object       *obj;
1924         int                      rc;
1925         int                      loop = 0;
1926         mdsno_t                  mds;
1927         int                      sidx;
1928         ENTRY;
1929
1930         rc = lmv_check_connect(obd);
1931         if (rc)
1932                 RETURN(rc);
1933
1934 repeat:
1935         ++loop;
1936         LASSERT(loop <= 2);
1937         LASSERT(op_data->op_namelen != 0);
1938
1939         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
1940                PFID(&op_data->op_fid2), op_data->op_namelen,
1941                op_data->op_name, PFID(&op_data->op_fid1));
1942
1943         obj = lmv_object_find(obd, &op_data->op_fid2);
1944         if (obj) {
1945                 sidx = raw_name2idx(obj->lo_hashtype,
1946                                     obj->lo_objcount,
1947                                     op_data->op_name,
1948                                     op_data->op_namelen);
1949                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
1950                 mds = obj->lo_stripes[sidx].ls_mds;
1951                 lmv_object_put(obj);
1952         } else {
1953                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1954                 if (rc)
1955                         RETURN(rc);
1956         }
1957
1958         CDEBUG(D_INODE, "Forward to mds #%x ("DFID")\n",
1959                mds, PFID(&op_data->op_fid1));
1960
1961         op_data->op_fsuid = current->fsuid;
1962         op_data->op_fsgid = current->fsgid;
1963         op_data->op_cap = cfs_curproc_cap_pack();
1964         tgt = lmv_get_target(lmv, mds);
1965
1966         /*
1967          * Cancel UPDATE lock on child (fid1).
1968          */
1969         op_data->op_flags |= MF_MDC_CANCEL_FID2;
1970         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
1971                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1972         if (rc == 0)
1973                 rc = md_link(tgt->ltd_exp, op_data, request);
1974         if (rc == -ERESTART) {
1975                 LASSERT(*request != NULL);
1976                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1977                           "Got -ERESTART during link!\n");
1978                 ptlrpc_req_finished(*request);
1979                 *request = NULL;
1980
1981                 /*
1982                  * Directory got split. Time to update local object and repeat
1983                  * the request with proper MDS.
1984                  */
1985                 rc = lmv_handle_split(exp, &op_data->op_fid2);
1986                 if (rc == 0)
1987                         goto repeat;
1988         }
1989
1990         RETURN(rc);
1991 }
1992
1993 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
1994                       const char *old, int oldlen, const char *new, int newlen,
1995                       struct ptlrpc_request **request)
1996 {
1997         struct obd_device       *obd = exp->exp_obd;
1998         struct lmv_obd          *lmv = &obd->u.lmv;
1999         struct lmv_tgt_desc     *src_tgt;
2000         struct lmv_tgt_desc     *tgt_tgt;
2001         int                      rc;
2002         int                      sidx;
2003         int                      loop = 0;
2004         struct lmv_object       *obj;
2005         mdsno_t                  mds1;
2006         mdsno_t                  mds2;
2007         ENTRY;
2008
2009         LASSERT(oldlen != 0);
2010
2011         CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2012                oldlen, old, PFID(&op_data->op_fid1),
2013                newlen, new, PFID(&op_data->op_fid2));
2014
2015         rc = lmv_check_connect(obd);
2016         if (rc)
2017                 RETURN(rc);
2018
2019 repeat:
2020         ++loop;
2021         LASSERT(loop <= 2);
2022         obj = lmv_object_find(obd, &op_data->op_fid1);
2023         if (obj) {
2024                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2025                                     (char *)old, oldlen);
2026                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2027                 mds1 = obj->lo_stripes[sidx].ls_mds;
2028                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2029                 lmv_object_put(obj);
2030         } else {
2031                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2032                 if (rc)
2033                         RETURN(rc);
2034         }
2035
2036         obj = lmv_object_find(obd, &op_data->op_fid2);
2037         if (obj) {
2038                 /*
2039                  * Directory is already split, so we have to forward request to
2040                  * the right MDS.
2041                  */
2042                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2043                                        (char *)new, newlen);
2044
2045                 mds2 = obj->lo_stripes[sidx].ls_mds;
2046                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2047                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2048                 lmv_object_put(obj);
2049         } else {
2050                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2051                 if (rc)
2052                         RETURN(rc);
2053         }
2054
2055         op_data->op_fsuid = current->fsuid;
2056         op_data->op_fsgid = current->fsgid;
2057         op_data->op_cap = cfs_curproc_cap_pack();
2058
2059         src_tgt = lmv_get_target(lmv, mds1);
2060         tgt_tgt = lmv_get_target(lmv, mds2);
2061
2062         /*
2063          * LOOKUP lock on src child (fid3) should also be cancelled for
2064          * src_tgt in mdc_rename.
2065          */
2066         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2067
2068         /*
2069          * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2070          * own target.
2071          */
2072         rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2073                               LCK_EX, MDS_INODELOCK_UPDATE,
2074                               MF_MDC_CANCEL_FID2);
2075
2076         /*
2077          * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2078          */
2079         if (rc == 0) {
2080                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2081                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2082                                       MF_MDC_CANCEL_FID4);
2083         }
2084
2085         /*
2086          * Cancel all the locks on tgt child (fid4).
2087          */
2088         if (rc == 0)
2089                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2090                                       LCK_EX, MDS_INODELOCK_FULL,
2091                                       MF_MDC_CANCEL_FID4);
2092
2093         if (rc == 0)
2094                 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2095                                new, newlen, request);
2096
2097         if (rc == -ERESTART) {
2098                 LASSERT(*request != NULL);
2099                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2100                           "Got -ERESTART during rename!\n");
2101                 ptlrpc_req_finished(*request);
2102                 *request = NULL;
2103
2104                 /*
2105                  * Directory got split. Time to update local object and repeat
2106                  * the request with proper MDS.
2107                  */
2108                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2109                 if (rc == 0)
2110                         goto repeat;
2111         }
2112         RETURN(rc);
2113 }
2114
2115 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2116                        void *ea, int ealen, void *ea2, int ea2len,
2117                        struct ptlrpc_request **request,
2118                        struct md_open_data **mod)
2119 {
2120         struct obd_device       *obd = exp->exp_obd;
2121         struct lmv_obd          *lmv = &obd->u.lmv;
2122         struct ptlrpc_request   *req;
2123         struct lmv_tgt_desc     *tgt;
2124         struct lmv_object       *obj;
2125         int                      rc = 0;
2126         int                      i;
2127         ENTRY;
2128
2129         rc = lmv_check_connect(obd);
2130         if (rc)
2131                 RETURN(rc);
2132
2133         obj = lmv_object_find(obd, &op_data->op_fid1);
2134
2135         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2136                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2137                obj ? ", split" : "");
2138
2139         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2140         if (obj) {
2141                 for (i = 0; i < obj->lo_objcount; i++) {
2142                         op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2143
2144                         tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2145                         if (IS_ERR(tgt)) {
2146                                 rc = PTR_ERR(tgt);
2147                                 break;
2148                         }
2149
2150                         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2151                                         ea2, ea2len, &req, mod);
2152
2153                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2154                                 /*
2155                                  * This is master object and this request should
2156                                  * be returned back to llite.
2157                                  */
2158                                 *request = req;
2159                         } else {
2160                                 ptlrpc_req_finished(req);
2161                         }
2162
2163                         if (rc)
2164                                 break;
2165                 }
2166                 lmv_object_put(obj);
2167         } else {
2168                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2169                 if (IS_ERR(tgt))
2170                         RETURN(PTR_ERR(tgt));
2171
2172                 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2173                                 ea2len, request, mod);
2174         }
2175         RETURN(rc);
2176 }
2177
2178 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2179                     struct obd_capa *oc, struct ptlrpc_request **request)
2180 {
2181         struct obd_device         *obd = exp->exp_obd;
2182         struct lmv_obd            *lmv = &obd->u.lmv;
2183         struct lmv_tgt_desc       *tgt;
2184         int                        rc;
2185         ENTRY;
2186
2187         rc = lmv_check_connect(obd);
2188         if (rc)
2189                 RETURN(rc);
2190
2191         tgt = lmv_find_target(lmv, fid);
2192         if (IS_ERR(tgt))
2193                 RETURN(PTR_ERR(tgt));
2194
2195         rc = md_sync(tgt->ltd_exp, fid, oc, request);
2196         RETURN(rc);
2197 }
2198
2199 /**
2200  * Main purpose of LMV blocking ast is to remove split directory LMV
2201  * presentation object (struct lmv_object) attached to the lock being revoked.
2202  */
2203 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2204                      void *data, int flag)
2205 {
2206         struct lustre_handle    lockh;
2207         struct lmv_object      *obj;
2208         int                     rc;
2209         ENTRY;
2210
2211         switch (flag) {
2212         case LDLM_CB_BLOCKING:
2213                 ldlm_lock2handle(lock, &lockh);
2214                 rc = ldlm_cli_cancel(&lockh);
2215                 if (rc < 0) {
2216                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2217                         RETURN(rc);
2218                 }
2219                 break;
2220         case LDLM_CB_CANCELING:
2221                 /*
2222                  * Time to drop cached attrs for split directory object
2223                  */
2224                 obj = lock->l_ast_data;
2225                 if (obj) {
2226                         CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2227                                ", master "DFID"\n",
2228                                lock->l_resource->lr_name.name[3] == 1 ?
2229                                "LOOKUP" : "UPDATE",
2230                                lock->l_resource->lr_name.name[0],
2231                                lock->l_resource->lr_name.name[1],
2232                                PFID(&obj->lo_fid));
2233                         lmv_object_put(obj);
2234                 }
2235                 break;
2236         default:
2237                 LBUG();
2238         }
2239         RETURN(0);
2240 }
2241
2242 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2243 {
2244         __u64         val;
2245
2246         val = le64_to_cpu(*hash);
2247         if (val < hash_adj)
2248                 val += MAX_HASH_SIZE;
2249         if (val != DIR_END_OFF)
2250                 *hash = cpu_to_le64(val - hash_adj);
2251 }
2252
2253 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2254 {
2255         __u64              id;
2256         struct obd_import *imp;
2257
2258         /*
2259          * XXX: to get nid we assume that underlying obd device is mdc.
2260          */
2261         imp  = class_exp2cliimp(exp);
2262         id   = imp->imp_connection->c_self + fid_flatten(fid);
2263
2264         CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2265                imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2266
2267         return id ^ (id >> 32);
2268 }
2269
2270 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2271                         struct obd_capa *oc, __u64 offset64, struct page *page,
2272                         struct ptlrpc_request **request)
2273 {
2274         struct obd_device       *obd = exp->exp_obd;
2275         struct lmv_obd          *lmv = &obd->u.lmv;
2276         struct lu_fid            rid = *fid;
2277         struct lmv_object       *obj;
2278         __u64                    offset;
2279         __u64                    hash_adj = 0;
2280         __u32                    rank = 0;
2281         __u64                    seg_size = 0;
2282         __u64                    tgt_tmp = 0;
2283         int                      tgt_idx = 0;
2284         int                      tgt0_idx = 0;
2285         int                      rc;
2286         int                      nr = 0;
2287         struct lmv_stripe       *los;
2288         struct lmv_tgt_desc     *tgt;
2289         struct lu_dirpage       *dp;
2290         struct lu_dirent        *ent;
2291         ENTRY;
2292
2293         offset = offset64;
2294
2295         rc = lmv_check_connect(obd);
2296         if (rc)
2297                 RETURN(rc);
2298
2299         CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2300
2301         /*
2302          * This case handle directory lookup in clustered metadata case (i.e.
2303          * split directory is located on multiple md servers.)
2304          * each server keeps directory entries for certain range of hashes.
2305          * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2306          * first server will keep records with hashes [ 0 ... MAX_HASH / N  - 1],
2307          * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2308          * so on....
2309          *      readdir can simply start reading entries from 0 - N server in
2310          * order but that will not scale well as all client will request dir in
2311          * to server in same order.
2312          * Following algorithm does optimization:
2313          * Instead of doing readdir in 1, 2, ...., N order, client with a
2314          * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2315          * (every client has rank R)
2316          *      But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2317          * since client ask dir from MDS{R} client has pages with offsets
2318          * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2319          * on hash  values that we get.
2320          */
2321         obj = lmv_object_find_lock(obd, fid);
2322         if (obj) {
2323                 nr       = obj->lo_objcount;
2324                 LASSERT(nr > 0);
2325                 seg_size = MAX_HASH_SIZE;
2326                 do_div(seg_size, nr);
2327                 los      = obj->lo_stripes;
2328                 tgt      = lmv_get_target(lmv, los[0].ls_mds);
2329                 rank     = lmv_node_rank(tgt->ltd_exp, fid) % nr;
2330                 tgt_tmp  = offset;
2331                 do_div(tgt_tmp, seg_size);
2332                 tgt0_idx = do_div(tgt_tmp,  nr);
2333                 tgt_idx  = (tgt0_idx + rank) % nr;
2334
2335                 if (tgt_idx < tgt0_idx)
2336                         /*
2337                          * Wrap around.
2338                          *
2339                          * Last segment has unusual length due to division
2340                          * rounding.
2341                          */
2342                         hash_adj = MAX_HASH_SIZE - seg_size * nr;
2343                 else
2344                         hash_adj = 0;
2345
2346                 hash_adj += rank * seg_size;
2347
2348                 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2349                        LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2350                        offset, tgt0_idx, offset + hash_adj, tgt_idx);
2351
2352                 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2353                 rid = obj->lo_stripes[tgt_idx].ls_fid;
2354                 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2355
2356                 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2357                        PFID(&rid), (unsigned long)offset, tgt_idx);
2358         } else
2359                 tgt = lmv_find_target(lmv, &rid);
2360
2361         if (IS_ERR(tgt))
2362                 GOTO(cleanup, rc = PTR_ERR(tgt));
2363
2364         rc = md_readpage(tgt->ltd_exp, &rid, oc, offset, page, request);
2365         if (rc)
2366                 GOTO(cleanup, rc);
2367         if (obj) {
2368                 dp = cfs_kmap(page);
2369
2370                 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2371                 lmv_hash_adjust(&dp->ldp_hash_end,   hash_adj);
2372                 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2373
2374                 for (ent = lu_dirent_start(dp); ent != NULL;
2375                      ent = lu_dirent_next(ent))
2376                         lmv_hash_adjust(&ent->lde_hash, hash_adj);
2377
2378                 if (tgt0_idx != nr - 1) {
2379                         __u64 end;
2380
2381                         end = le64_to_cpu(dp->ldp_hash_end);
2382                         if (end == DIR_END_OFF) {
2383                                 dp->ldp_hash_end = cpu_to_le32(seg_size *
2384                                                                (tgt0_idx + 1));
2385                                 CDEBUG(D_INODE,
2386                                        ""DFID" reset end "LPX64" tgt %d\n",
2387                                        PFID(&rid),
2388                                        le64_to_cpu(dp->ldp_hash_end), tgt_idx);
2389                         }
2390                 }
2391                 cfs_kunmap(page);
2392         }
2393         EXIT;
2394 cleanup:
2395         if (obj)
2396                 lmv_object_put_unlock(obj);
2397         return rc;
2398 }
2399
2400 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2401                       struct ptlrpc_request **request)
2402 {
2403         struct obd_device       *obd = exp->exp_obd;
2404         struct lmv_obd          *lmv = &obd->u.lmv;
2405         struct lmv_tgt_desc     *tgt = NULL;
2406         struct lmv_object       *obj;
2407         int                      rc;
2408         int                      sidx;
2409         int                      loop = 0;
2410         ENTRY;
2411
2412         rc = lmv_check_connect(obd);
2413         if (rc)
2414                 RETURN(rc);
2415
2416 repeat:
2417         ++loop;
2418         LASSERT(loop <= 2);
2419         LASSERT(op_data->op_namelen != 0);
2420
2421         obj = lmv_object_find(obd, &op_data->op_fid1);
2422         if (obj) {
2423                 sidx = raw_name2idx(obj->lo_hashtype,
2424                                     obj->lo_objcount,
2425                                     op_data->op_name,
2426                                     op_data->op_namelen);
2427                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2428                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2429                 tgt = lmv_get_target(lmv,
2430                                      obj->lo_stripes[sidx].ls_mds);
2431                 lmv_object_put(obj);
2432                 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2433                        op_data->op_namelen, op_data->op_name,
2434                        PFID(&op_data->op_fid1), sidx);
2435         }
2436
2437         if (tgt == NULL) {
2438                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2439                 if (IS_ERR(tgt))
2440                         RETURN(PTR_ERR(tgt));
2441                 op_data->op_bias |= MDS_CHECK_SPLIT;
2442         }
2443
2444         op_data->op_fsuid = current->fsuid;
2445         op_data->op_fsgid = current->fsgid;
2446         op_data->op_cap = cfs_curproc_cap_pack();
2447
2448         /*
2449          * If child's fid is given, cancel unused locks for it if it is from
2450          * another export than parent.
2451          *
2452          * LOOKUP lock for child (fid3) should also be cancelled on parent
2453          * tgt_tgt in mdc_unlink().
2454          */
2455         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2456
2457         /*
2458          * Cancel FULL locks on child (fid3).
2459          */
2460         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2461                               MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2462
2463         if (rc == 0)
2464                 rc = md_unlink(tgt->ltd_exp, op_data, request);
2465
2466         if (rc == -ERESTART) {
2467                 LASSERT(*request != NULL);
2468                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2469                           "Got -ERESTART during unlink!\n");
2470                 ptlrpc_req_finished(*request);
2471                 *request = NULL;
2472
2473                 /*
2474                  * Directory got split. Time to update local object and repeat
2475                  * the request with proper MDS.
2476                  */
2477                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2478                 if (rc == 0)
2479                         goto repeat;
2480         }
2481         RETURN(rc);
2482 }
2483
2484 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2485 {
2486         int        rc = 0;
2487
2488         switch (stage) {
2489         case OBD_CLEANUP_EARLY:
2490                 /* XXX: here should be calling obd_precleanup() down to
2491                  * stack. */
2492                 break;
2493         case OBD_CLEANUP_EXPORTS:
2494                 rc = obd_llog_finish(obd, 0);
2495                 if (rc != 0)
2496                         CERROR("failed to cleanup llogging subsystems\n");
2497                 break;
2498         default:
2499                 break;
2500         }
2501         RETURN(rc);
2502 }
2503
2504 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2505                         void *key, __u32 *vallen, void *val,
2506                         struct lov_stripe_md *lsm)
2507 {
2508         struct obd_device       *obd;
2509         struct lmv_obd          *lmv;
2510         int                      rc = 0;
2511         ENTRY;
2512
2513         obd = class_exp2obd(exp);
2514         if (obd == NULL) {
2515                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2516                        exp->exp_handle.h_cookie);
2517                 RETURN(-EINVAL);
2518         }
2519
2520         lmv = &obd->u.lmv;
2521         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2522                 struct lmv_tgt_desc *tgts;
2523                 int i;
2524
2525                 rc = lmv_check_connect(obd);
2526                 if (rc)
2527                         RETURN(rc);
2528
2529                 LASSERT(*vallen == sizeof(__u32));
2530                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2531                      i++, tgts++) {
2532
2533                         /*
2534                          * All tgts should be connected when this gets called.
2535                          */
2536                         if (!tgts || !tgts->ltd_exp) {
2537                                 CERROR("target not setup?\n");
2538                                 continue;
2539                         }
2540
2541                         if (!obd_get_info(tgts->ltd_exp, keylen, key,
2542                                           vallen, val, NULL))
2543                                 RETURN(0);
2544                 }
2545                 RETURN(-EINVAL);
2546         } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2547                 rc = lmv_check_connect(obd);
2548                 if (rc)
2549                         RETURN(rc);
2550
2551                 /*
2552                  * Forwarding this request to first MDS, it should know LOV
2553                  * desc.
2554                  */
2555                 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2556                                   vallen, val, NULL);
2557                 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2558                         exp->exp_connect_flags =
2559                         ((struct obd_connect_data *)val)->ocd_connect_flags;
2560                 }
2561                 RETURN(rc);
2562         }
2563
2564         CDEBUG(D_IOCTL, "Invalid key\n");
2565         RETURN(-EINVAL);
2566 }
2567
2568 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2569                        void *key, obd_count vallen, void *val,
2570                        struct ptlrpc_request_set *set)
2571 {
2572         struct lmv_tgt_desc    *tgt;
2573         struct obd_device      *obd;
2574         struct lmv_obd         *lmv;
2575         int rc = 0;
2576         ENTRY;
2577
2578         obd = class_exp2obd(exp);
2579         if (obd == NULL) {
2580                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2581                        exp->exp_handle.h_cookie);
2582                 RETURN(-EINVAL);
2583         }
2584         lmv = &obd->u.lmv;
2585
2586         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2587             KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2588                 int i, err = 0;
2589
2590                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2591                         tgt = &lmv->tgts[i];
2592
2593                         if (!tgt->ltd_exp)
2594                                 continue;
2595
2596                         err = obd_set_info_async(tgt->ltd_exp,
2597                                                  keylen, key, vallen, val, set);
2598                         if (err && rc == 0)
2599                                 rc = err;
2600                 }
2601
2602                 RETURN(rc);
2603         }
2604
2605         RETURN(-EINVAL);
2606 }
2607
2608 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2609                struct lov_stripe_md *lsm)
2610 {
2611         struct obd_device         *obd = class_exp2obd(exp);
2612         struct lmv_obd            *lmv = &obd->u.lmv;
2613         struct lmv_stripe_md      *meap;
2614         struct lmv_stripe_md      *lsmp;
2615         int                        mea_size;
2616         int                        i;
2617         ENTRY;
2618
2619         mea_size = lmv_get_easize(lmv);
2620         if (!lmmp)
2621                 RETURN(mea_size);
2622
2623         if (*lmmp && !lsm) {
2624                 OBD_FREE(*lmmp, mea_size);
2625                 *lmmp = NULL;
2626                 RETURN(0);
2627         }
2628
2629         if (*lmmp == NULL) {
2630                 OBD_ALLOC(*lmmp, mea_size);
2631                 if (*lmmp == NULL)
2632                         RETURN(-ENOMEM);
2633         }
2634
2635         if (!lsm)
2636                 RETURN(mea_size);
2637
2638         lsmp = (struct lmv_stripe_md *)lsm;
2639         meap = (struct lmv_stripe_md *)*lmmp;
2640
2641         if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2642             lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2643                 RETURN(-EINVAL);
2644
2645         meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2646         meap->mea_count = cpu_to_le32(lsmp->mea_count);
2647         meap->mea_master = cpu_to_le32(lsmp->mea_master);
2648
2649         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2650                 meap->mea_ids[i] = meap->mea_ids[i];
2651                 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2652         }
2653
2654         RETURN(mea_size);
2655 }
2656
2657 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2658                  struct lov_mds_md *lmm, int lmm_size)
2659 {
2660         struct obd_device          *obd = class_exp2obd(exp);
2661         struct lmv_stripe_md      **tmea = (struct lmv_stripe_md **)lsmp;
2662         struct lmv_stripe_md       *mea = (struct lmv_stripe_md *)lmm;
2663         struct lmv_obd             *lmv = &obd->u.lmv;
2664         int                         mea_size;
2665         int                         i;
2666         __u32                       magic;
2667         ENTRY;
2668
2669         mea_size = lmv_get_easize(lmv);
2670         if (lsmp == NULL)
2671                 return mea_size;
2672
2673         if (*lsmp != NULL && lmm == NULL) {
2674                 OBD_FREE(*tmea, mea_size);
2675                 *lsmp = NULL;
2676                 RETURN(0);
2677         }
2678
2679         LASSERT(mea_size == lmm_size);
2680
2681         OBD_ALLOC(*tmea, mea_size);
2682         if (*tmea == NULL)
2683                 RETURN(-ENOMEM);
2684
2685         if (!lmm)
2686                 RETURN(mea_size);
2687
2688         if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2689             mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2690             mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2691         {
2692                 magic = le32_to_cpu(mea->mea_magic);
2693         } else {
2694                 /*
2695                  * Old mea is not handled here.
2696                  */
2697                 CERROR("Old not supportable EA is found\n");
2698                 LBUG();
2699         }
2700
2701         (*tmea)->mea_magic = magic;
2702         (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2703         (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2704
2705         for (i = 0; i < (*tmea)->mea_count; i++) {
2706                 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2707                 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2708         }
2709         RETURN(mea_size);
2710 }
2711
2712 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2713                              ldlm_policy_data_t *policy, ldlm_mode_t mode,
2714                              int flags, void *opaque)
2715 {
2716         struct obd_device       *obd = exp->exp_obd;
2717         struct lmv_obd          *lmv = &obd->u.lmv;
2718         int                      rc = 0;
2719         int                      err;
2720         int                      i;
2721         ENTRY;
2722
2723         LASSERT(fid != NULL);
2724
2725         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2726                 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2727                         continue;
2728
2729                 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2730                                        policy, mode, flags, opaque);
2731                 if (!rc)
2732                         rc = err;
2733         }
2734         RETURN(rc);
2735 }
2736
2737 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data)
2738 {
2739         struct obd_device       *obd = exp->exp_obd;
2740         struct lmv_obd          *lmv = &obd->u.lmv;
2741         int                      rc;
2742         ENTRY;
2743         rc =  md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data);
2744         RETURN(rc);
2745 }
2746
2747 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2748                            const struct lu_fid *fid, ldlm_type_t type,
2749                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
2750                            struct lustre_handle *lockh)
2751 {
2752         struct obd_device       *obd = exp->exp_obd;
2753         struct lmv_obd          *lmv = &obd->u.lmv;
2754         ldlm_mode_t              rc;
2755         int                      i;
2756         ENTRY;
2757
2758         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2759
2760         /*
2761          * With CMD every object can have two locks in different namespaces:
2762          * lookup lock in space of mds storing direntry and update/open lock in
2763          * space of mds storing inode. Thus we check all targets, not only that
2764          * one fid was created in.
2765          */
2766         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2767                 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2768                                    type, policy, mode, lockh);
2769                 if (rc)
2770                         RETURN(rc);
2771         }
2772
2773         RETURN(0);
2774 }
2775
2776 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2777                       struct obd_export *dt_exp, struct obd_export *md_exp,
2778                       struct lustre_md *md)
2779 {
2780         struct obd_device       *obd = exp->exp_obd;
2781         struct lmv_obd          *lmv = &obd->u.lmv;
2782         int                      rc;
2783         ENTRY;
2784         rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2785         RETURN(rc);
2786 }
2787
2788 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2789 {
2790         struct obd_device       *obd = exp->exp_obd;
2791         struct lmv_obd          *lmv = &obd->u.lmv;
2792         ENTRY;
2793
2794         if (md->mea)
2795                 obd_free_memmd(exp, (void *)&md->mea);
2796         RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2797 }
2798
2799 int lmv_set_open_replay_data(struct obd_export *exp,
2800                              struct obd_client_handle *och,
2801                              struct ptlrpc_request *open_req)
2802 {
2803         struct obd_device       *obd = exp->exp_obd;
2804         struct lmv_obd          *lmv = &obd->u.lmv;
2805         struct lmv_tgt_desc     *tgt;
2806         ENTRY;
2807
2808         tgt = lmv_find_target(lmv, &och->och_fid);
2809         if (IS_ERR(tgt))
2810                 RETURN(PTR_ERR(tgt));
2811
2812         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2813 }
2814
2815 int lmv_clear_open_replay_data(struct obd_export *exp,
2816                                struct obd_client_handle *och)
2817 {
2818         struct obd_device       *obd = exp->exp_obd;
2819         struct lmv_obd          *lmv = &obd->u.lmv;
2820         struct lmv_tgt_desc     *tgt;
2821         ENTRY;
2822
2823         tgt = lmv_find_target(lmv, &och->och_fid);
2824         if (IS_ERR(tgt))
2825                 RETURN(PTR_ERR(tgt));
2826
2827         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2828 }
2829
2830 static int lmv_get_remote_perm(struct obd_export *exp,
2831                                const struct lu_fid *fid,
2832                                struct obd_capa *oc, __u32 suppgid,
2833                                struct ptlrpc_request **request)
2834 {
2835         struct obd_device       *obd = exp->exp_obd;
2836         struct lmv_obd          *lmv = &obd->u.lmv;
2837         struct lmv_tgt_desc     *tgt;
2838         int                      rc;
2839         ENTRY;
2840
2841         rc = lmv_check_connect(obd);
2842         if (rc)
2843                 RETURN(rc);
2844
2845         tgt = lmv_find_target(lmv, fid);
2846         if (IS_ERR(tgt))
2847                 RETURN(PTR_ERR(tgt));
2848
2849         rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
2850         RETURN(rc);
2851 }
2852
2853 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2854                           renew_capa_cb_t cb)
2855 {
2856         struct obd_device       *obd = exp->exp_obd;
2857         struct lmv_obd          *lmv = &obd->u.lmv;
2858         struct lmv_tgt_desc     *tgt;
2859         int                      rc;
2860         ENTRY;
2861
2862         rc = lmv_check_connect(obd);
2863         if (rc)
2864                 RETURN(rc);
2865
2866         tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
2867         if (IS_ERR(tgt))
2868                 RETURN(PTR_ERR(tgt));
2869
2870         rc = md_renew_capa(tgt->ltd_exp, oc, cb);
2871         RETURN(rc);
2872 }
2873
2874 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
2875                     const struct req_msg_field *field, struct obd_capa **oc)
2876 {
2877         struct obd_device *obd = exp->exp_obd;
2878         struct lmv_obd *lmv = &obd->u.lmv;
2879         int rc;
2880
2881         ENTRY;
2882         rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
2883         RETURN(rc);
2884 }
2885
2886 int lmv_intent_getattr_async(struct obd_export *exp,
2887                              struct md_enqueue_info *minfo,
2888                              struct ldlm_enqueue_info *einfo)
2889 {
2890         struct md_op_data       *op_data = &minfo->mi_data;
2891         struct obd_device       *obd = exp->exp_obd;
2892         struct lmv_obd          *lmv = &obd->u.lmv;
2893         struct lmv_object       *obj;
2894         struct lmv_tgt_desc     *tgt;
2895         int                      rc;
2896         int                      sidx;
2897         ENTRY;
2898
2899         rc = lmv_check_connect(obd);
2900         if (rc)
2901                 RETURN(rc);
2902
2903         if (!fid_is_sane(&op_data->op_fid2)) {
2904                 obj = lmv_object_find(obd, &op_data->op_fid1);
2905                 if (obj && op_data->op_namelen) {
2906                         sidx = raw_name2idx(obj->lo_hashtype,
2907                                             obj->lo_objcount,
2908                                             (char *)op_data->op_name,
2909                                             op_data->op_namelen);
2910                         op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2911                         tgt = lmv_get_target(lmv,
2912                                              obj->lo_stripes[sidx].ls_mds);
2913                         CDEBUG(D_INODE,
2914                                "Choose slave dir ("DFID") -> mds #%d\n",
2915                                PFID(&op_data->op_fid1), tgt->ltd_idx);
2916                 } else {
2917                         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2918                 }
2919                 if (obj)
2920                         lmv_object_put(obj);
2921         } else {
2922                 op_data->op_fid1 = op_data->op_fid2;
2923                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2924                 op_data->op_bias = MDS_CROSS_REF;
2925                 /*
2926                  * Unfortunately, we have to lie to MDC/MDS to retrieve
2927                  * attributes llite needs.
2928                 */
2929                 if (minfo->mi_it.it_op & IT_LOOKUP)
2930                         minfo->mi_it.it_op = IT_GETATTR;
2931         }
2932
2933         if (IS_ERR(tgt))
2934                 RETURN(PTR_ERR(tgt));
2935
2936         rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
2937         RETURN(rc);
2938 }
2939
2940 int lmv_revalidate_lock(struct obd_export *exp,
2941                         struct lookup_intent *it,
2942                         struct lu_fid *fid)
2943 {
2944         struct obd_device       *obd = exp->exp_obd;
2945         struct lmv_obd          *lmv = &obd->u.lmv;
2946         struct lmv_tgt_desc     *tgt;
2947         int                      rc;
2948         ENTRY;
2949
2950         rc = lmv_check_connect(obd);
2951         if (rc)
2952                 RETURN(rc);
2953
2954         tgt = lmv_find_target(lmv, fid);
2955         if (IS_ERR(tgt))
2956                 RETURN(PTR_ERR(tgt));
2957
2958         rc = md_revalidate_lock(tgt->ltd_exp, it, fid);
2959         RETURN(rc);
2960 }
2961
2962
2963 struct obd_ops lmv_obd_ops = {
2964         .o_owner                = THIS_MODULE,
2965         .o_setup                = lmv_setup,
2966         .o_cleanup              = lmv_cleanup,
2967         .o_precleanup           = lmv_precleanup,
2968         .o_process_config       = lmv_process_config,
2969         .o_connect              = lmv_connect,
2970         .o_disconnect           = lmv_disconnect,
2971         .o_statfs               = lmv_statfs,
2972         .o_get_info             = lmv_get_info,
2973         .o_set_info_async       = lmv_set_info_async,
2974         .o_packmd               = lmv_packmd,
2975         .o_unpackmd             = lmv_unpackmd,
2976         .o_notify               = lmv_notify,
2977         .o_get_uuid             = lmv_get_uuid,
2978         .o_iocontrol            = lmv_iocontrol,
2979         .o_fid_delete           = lmv_fid_delete
2980 };
2981
2982 struct md_ops lmv_md_ops = {
2983         .m_getstatus            = lmv_getstatus,
2984         .m_change_cbdata        = lmv_change_cbdata,
2985         .m_close                = lmv_close,
2986         .m_create               = lmv_create,
2987         .m_done_writing         = lmv_done_writing,
2988         .m_enqueue              = lmv_enqueue,
2989         .m_getattr              = lmv_getattr,
2990         .m_getxattr             = lmv_getxattr,
2991         .m_getattr_name         = lmv_getattr_name,
2992         .m_intent_lock          = lmv_intent_lock,
2993         .m_link                 = lmv_link,
2994         .m_rename               = lmv_rename,
2995         .m_setattr              = lmv_setattr,
2996         .m_setxattr             = lmv_setxattr,
2997         .m_sync                 = lmv_sync,
2998         .m_readpage             = lmv_readpage,
2999         .m_unlink               = lmv_unlink,
3000         .m_init_ea_size         = lmv_init_ea_size,
3001         .m_cancel_unused        = lmv_cancel_unused,
3002         .m_set_lock_data        = lmv_set_lock_data,
3003         .m_lock_match           = lmv_lock_match,
3004         .m_get_lustre_md        = lmv_get_lustre_md,
3005         .m_free_lustre_md       = lmv_free_lustre_md,
3006         .m_set_open_replay_data = lmv_set_open_replay_data,
3007         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3008         .m_renew_capa           = lmv_renew_capa,
3009         .m_unpack_capa          = lmv_unpack_capa,
3010         .m_get_remote_perm      = lmv_get_remote_perm,
3011         .m_intent_getattr_async = lmv_intent_getattr_async,
3012         .m_revalidate_lock      = lmv_revalidate_lock
3013 };
3014
3015 static quota_interface_t *quota_interface;
3016 extern quota_interface_t lmv_quota_interface;
3017
3018 int __init lmv_init(void)
3019 {
3020         struct lprocfs_static_vars lvars;
3021         int                        rc;
3022
3023         lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3024                                                 sizeof(struct lmv_object),
3025                                                 0, 0);
3026         if (!lmv_object_cache) {
3027                 CERROR("Error allocating lmv objects cache\n");
3028                 return -ENOMEM;
3029         }
3030
3031         lprocfs_lmv_init_vars(&lvars);
3032
3033         request_module("lquota");
3034         quota_interface = PORTAL_SYMBOL_GET(lmv_quota_interface);
3035         init_obd_quota_ops(quota_interface, &lmv_obd_ops);
3036
3037         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3038                                  lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3039         if (rc) {
3040                 if (quota_interface)
3041                         PORTAL_SYMBOL_PUT(lmv_quota_interface);
3042                 cfs_mem_cache_destroy(lmv_object_cache);
3043         }
3044
3045         return rc;
3046 }
3047
3048 #ifdef __KERNEL__
3049 static void lmv_exit(void)
3050 {
3051         if (quota_interface)
3052                 PORTAL_SYMBOL_PUT(lmv_quota_interface);
3053
3054         class_unregister_type(LUSTRE_LMV_NAME);
3055
3056         LASSERTF(atomic_read(&lmv_object_count) == 0,
3057                  "Can't free lmv objects cache, %d object(s) busy\n",
3058                  atomic_read(&lmv_object_count));
3059         cfs_mem_cache_destroy(lmv_object_cache);
3060 }
3061
3062 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3063 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3064 MODULE_LICENSE("GPL");
3065
3066 module_init(lmv_init);
3067 module_exit(lmv_exit);
3068 #endif