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 cfs_atomic_t lmv_object_count = CFS_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         cfs_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         cfs_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         cfs_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         cfs_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                 cfs_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, NULL);
 525                 if (rc) {
 526                         lmv_init_unlock(lmv);
 527                         CERROR("lmv failed to setup llogging subsystems\n");
 528                 }
 529         }
 530         cfs_spin_lock(&lmv->lmv_lock);
 531         tgt = lmv->tgts + lmv->desc.ld_tgt_count++;
 532         tgt->ltd_uuid = *tgt_uuid;
 533         cfs_spin_unlock(&lmv->lmv_lock);
 534
 535         if (lmv->connected) {
 536                 rc = lmv_connect_mdc(obd, tgt);
 537                 if (rc) {
 538                         cfs_spin_lock(&lmv->lmv_lock);
 539                         lmv->desc.ld_tgt_count--;
 540                         memset(tgt, 0, sizeof(*tgt));
 541                         cfs_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() - CFS_HZ, 0);
 757                 if (rc)
 758                         RETURN(rc);
 759                 if (cfs_copy_to_user(data->ioc_pbuf1, &stat_buf,
 760                                      data->ioc_plen1))
 761                         RETURN(-EFAULT);
 762                 if (cfs_copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 763                                      data->ioc_plen2))
 764                         RETURN(-EFAULT);
 765                 break;
 766         }
 767         case OBD_IOC_QUOTACTL: {
 768                 struct if_quotactl *qctl = karg;
 769                 struct lmv_tgt_desc *tgt = NULL;
 770                 struct obd_quotactl *oqctl;
 771
 772                 if (qctl->qc_valid == QC_MDTIDX) {
 773                         if (qctl->qc_idx < 0 || count <= qctl->qc_idx)
 774                                 RETURN(-EINVAL);
 775
 776                         tgt = &lmv->tgts[qctl->qc_idx];
 777                         if (!tgt->ltd_exp)
 778                                 RETURN(-EINVAL);
 779                 } else if (qctl->qc_valid == QC_UUID) {
 780                         for (i = 0; i < count; i++) {
 781                                 tgt = &lmv->tgts[i];
 782                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 783                                                      &qctl->obd_uuid))
 784                                         continue;
 785
 786                                 if (tgt->ltd_exp == NULL)
 787                                         RETURN(-EINVAL);
 788
 789                                 break;
 790                         }
 791                 } else {
 792                         RETURN(-EINVAL);
 793                 }
 794
 795                 if (i >= count)
 796                         RETURN(-EAGAIN);
 797
 798                 LASSERT(tgt && tgt->ltd_exp);
 799                 OBD_ALLOC_PTR(oqctl);
 800                 if (!oqctl)
 801                         RETURN(-ENOMEM);
 802
 803                 QCTL_COPY(oqctl, qctl);
 804                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 805                 if (rc == 0) {
 806                         QCTL_COPY(qctl, oqctl);
 807                         qctl->qc_valid = QC_MDTIDX;
 808                         qctl->obd_uuid = tgt->ltd_uuid;
 809                 }
 810                 OBD_FREE_PTR(oqctl);
 811                 break;
 812         }
 813         case OBD_IOC_CHANGELOG_CLEAR: {
 814                 struct ioc_changelog_clear *icc = karg;
 815
 816                 if (icc->icc_mdtindex >= count)
 817                         RETURN(-ENODEV);
 818
 819                 rc = obd_iocontrol(cmd, lmv->tgts[icc->icc_mdtindex].ltd_exp,
 820                                    sizeof(*icc), icc, NULL);
 821                 break;
 822         }
 823
 824         default : {
 825                 for (i = 0; i < count; i++) {
 826                         int err;
 827
 828                         if (lmv->tgts[i].ltd_exp == NULL)
 829                                 continue;
 830
 831                         err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len,
 832                                             karg, uarg);
 833                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
 834                                 RETURN(err);
 835                         } else if (err) {
 836                                 if (lmv->tgts[i].ltd_active) {
 837                                         CERROR("error: iocontrol MDC %s on MDT"
 838                                                "idx %d cmd %x: err = %d\n",
 839                                                 lmv->tgts[i].ltd_uuid.uuid,
 840                                                 i, cmd, err);
 841                                         if (!rc)
 842                                                 rc = err;
 843                                 }
 844                         } else
 845                                 set = 1;
 846                 }
 847                 if (!set && !rc)
 848                         rc = -EIO;
 849         }
 850         }
 851         RETURN(rc);
 852 }
 853
 854 static int lmv_all_chars_policy(int count, const char *name,
 855                                 int len)
 856 {
 857         unsigned int c = 0;
 858
 859         while (len > 0)
 860                 c += name[--len];
 861         c = c % count;
 862         return c;
 863 }
 864
 865 static int lmv_nid_policy(struct lmv_obd *lmv)
 866 {
 867         struct obd_import *imp;
 868         __u32              id;
 869
 870         /*
 871          * XXX: To get nid we assume that underlying obd device is mdc.
 872          */
 873         imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
 874         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
 875         return id % lmv->desc.ld_tgt_count;
 876 }
 877
 878 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
 879                           placement_policy_t placement)
 880 {
 881         switch (placement) {
 882         case PLACEMENT_CHAR_POLICY:
 883                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
 884                                             op_data->op_name,
 885                                             op_data->op_namelen);
 886         case PLACEMENT_NID_POLICY:
 887                 return lmv_nid_policy(lmv);
 888
 889         default:
 890                 break;
 891         }
 892
 893         CERROR("Unsupported placement policy %x\n", placement);
 894         return -EINVAL;
 895 }
 896
 897 /**
 898  * This is _inode_ placement policy function (not name).
 899  */
 900 static int lmv_placement_policy(struct obd_device *obd,
 901                                 struct md_op_data *op_data,
 902                                 mdsno_t *mds)
 903 {
 904         struct lmv_obd          *lmv = &obd->u.lmv;
 905         struct lmv_object       *obj;
 906         int                      rc;
 907         ENTRY;
 908
 909         LASSERT(mds != NULL);
 910
 911         if (lmv->desc.ld_tgt_count == 1) {
 912                 *mds = 0;
 913                 RETURN(0);
 914         }
 915
 916         /*
 917          * Allocate new fid on target according to operation type and parent
 918          * home mds.
 919          */
 920         obj = lmv_object_find(obd, &op_data->op_fid1);
 921         if (obj != NULL || op_data->op_name == NULL ||
 922             op_data->op_opc != LUSTRE_OPC_MKDIR) {
 923                 /*
 924                  * Allocate fid for non-dir or for null name or for case parent
 925                  * dir is split.
 926                  */
 927                 if (obj) {
 928                         lmv_object_put(obj);
 929
 930                         /*
 931                          * If we have this flag turned on, and we see that
 932                          * parent dir is split, this means, that caller did not
 933                          * notice split yet. This is race and we would like to
 934                          * let caller know that.
 935                          */
 936                         if (op_data->op_bias & MDS_CHECK_SPLIT)
 937                                 RETURN(-ERESTART);
 938                 }
 939
 940                 /*
 941                  * Allocate new fid on same mds where parent fid is located and
 942                  * where operation will be sent. In case of split dir, ->op_fid1
 943                  * and ->op_mds here will contain fid and mds of slave directory
 944                  * object (assigned by caller).
 945                  */
 946                 *mds = op_data->op_mds;
 947                 rc = 0;
 948         } else {
 949                 /*
 950                  * Parent directory is not split and we want to create a
 951                  * directory in it. Let's calculate where to place it according
 952                  * to operation data @op_data.
 953                  */
 954                 *mds = lmv_choose_mds(lmv, op_data, lmv->lmv_placement);
 955                 rc = 0;
 956         }
 957
 958         if (rc) {
 959                 CERROR("Can't choose MDS, err = %d\n", rc);
 960         } else {
 961                 LASSERT(*mds < lmv->desc.ld_tgt_count);
 962         }
 963
 964         RETURN(rc);
 965 }
 966
 967 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid,
 968                     mdsno_t mds)
 969 {
 970         struct lmv_tgt_desc *tgt;
 971         int                  rc;
 972         ENTRY;
 973
 974         tgt = lmv_get_target(lmv, mds);
 975
 976         /*
 977          * New seq alloc and FLD setup should be atomic. Otherwise we may find
 978          * on server that seq in new allocated fid is not yet known.
 979          */
 980         cfs_down(&tgt->ltd_fid_sem);
 981
 982         if (!tgt->ltd_active)
 983                 GOTO(out, rc = -ENODEV);
 984
 985         /*
 986          * Asking underlaying tgt layer to allocate new fid.
 987          */
 988         rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL);
 989         if (rc > 0) {
 990                 LASSERT(fid_is_sane(fid));
 991                 rc = 0;
 992         }
 993
 994         EXIT;
 995 out:
 996         cfs_up(&tgt->ltd_fid_sem);
 997         return rc;
 998 }
 999
1000 int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid,
1001                   struct md_op_data *op_data)
1002 {
1003         struct obd_device     *obd = class_exp2obd(exp);
1004         struct lmv_obd        *lmv = &obd->u.lmv;
1005         mdsno_t                mds;
1006         int                    rc;
1007         ENTRY;
1008
1009         LASSERT(op_data != NULL);
1010         LASSERT(fid != NULL);
1011
1012         rc = lmv_placement_policy(obd, op_data, &mds);
1013         if (rc) {
1014                 CERROR("Can't get target for allocating fid, "
1015                        "rc %d\n", rc);
1016                 RETURN(rc);
1017         }
1018
1019         rc = __lmv_fid_alloc(lmv, fid, mds);
1020         if (rc) {
1021                 CERROR("Can't alloc new fid, rc %d\n", rc);
1022                 RETURN(rc);
1023         }
1024
1025         RETURN(rc);
1026 }
1027
1028 static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid)
1029 {
1030         ENTRY;
1031         LASSERT(exp != NULL && fid != NULL);
1032         if (lmv_object_delete(exp, fid)) {
1033                 CDEBUG(D_INODE, "Object "DFID" is destroyed.\n",
1034                        PFID(fid));
1035         }
1036         RETURN(0);
1037 }
1038
1039 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1040 {
1041         struct lmv_obd             *lmv = &obd->u.lmv;
1042         struct lprocfs_static_vars  lvars;
1043         struct lmv_desc            *desc;
1044         int                         rc;
1045         int                         i = 0;
1046         ENTRY;
1047
1048         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1049                 CERROR("LMV setup requires a descriptor\n");
1050                 RETURN(-EINVAL);
1051         }
1052
1053         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1054         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1055                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1056                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1057                 RETURN(-EINVAL);
1058         }
1059
1060         lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc);
1061
1062         OBD_ALLOC(lmv->tgts, lmv->tgts_size);
1063         if (lmv->tgts == NULL)
1064                 RETURN(-ENOMEM);
1065
1066         for (i = 0; i < LMV_MAX_TGT_COUNT; i++) {
1067                 cfs_sema_init(&lmv->tgts[i].ltd_fid_sem, 1);
1068                 lmv->tgts[i].ltd_idx = i;
1069         }
1070
1071         lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data);
1072
1073         OBD_ALLOC(lmv->datas, lmv->datas_size);
1074         if (lmv->datas == NULL)
1075                 GOTO(out_free_tgts, rc = -ENOMEM);
1076
1077         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1078         lmv->desc.ld_tgt_count = 0;
1079         lmv->desc.ld_active_tgt_count = 0;
1080         lmv->max_cookiesize = 0;
1081         lmv->max_def_easize = 0;
1082         lmv->max_easize = 0;
1083         lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1084
1085         cfs_spin_lock_init(&lmv->lmv_lock);
1086         cfs_sema_init(&lmv->init_sem, 1);
1087
1088         rc = lmv_object_setup(obd);
1089         if (rc) {
1090                 CERROR("Can't setup LMV object manager, error %d.\n", rc);
1091                 GOTO(out_free_datas, rc);
1092         }
1093
1094         lprocfs_lmv_init_vars(&lvars);
1095         lprocfs_obd_setup(obd, lvars.obd_vars);
1096 #ifdef LPROCFS
1097         {
1098                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd_status",
1099                                         0444, &lmv_proc_target_fops, obd);
1100                 if (rc)
1101                         CWARN("Error adding target_obd_stats file (%d)\n", rc);
1102        }
1103 #endif
1104         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1105                              LUSTRE_CLI_FLD_HASH_DHT);
1106         if (rc) {
1107                 CERROR("Can't init FLD, err %d\n", rc);
1108                 GOTO(out_free_datas, rc);
1109         }
1110
1111         RETURN(0);
1112
1113 out_free_datas:
1114         OBD_FREE(lmv->datas, lmv->datas_size);
1115         lmv->datas = NULL;
1116 out_free_tgts:
1117         OBD_FREE(lmv->tgts, lmv->tgts_size);
1118         lmv->tgts = NULL;
1119         return rc;
1120 }
1121
1122 static int lmv_cleanup(struct obd_device *obd)
1123 {
1124         struct lmv_obd   *lmv = &obd->u.lmv;
1125         ENTRY;
1126
1127         fld_client_fini(&lmv->lmv_fld);
1128         lprocfs_obd_cleanup(obd);
1129         lmv_object_cleanup(obd);
1130         OBD_FREE(lmv->datas, lmv->datas_size);
1131         OBD_FREE(lmv->tgts, lmv->tgts_size);
1132
1133         RETURN(0);
1134 }
1135
1136 static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf)
1137 {
1138         struct lustre_cfg     *lcfg = buf;
1139         struct obd_uuid        tgt_uuid;
1140         int                    rc;
1141         ENTRY;
1142
1143         switch(lcfg->lcfg_command) {
1144         case LCFG_ADD_MDC:
1145                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid))
1146                         GOTO(out, rc = -EINVAL);
1147
1148                 obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1));
1149                 rc = lmv_add_target(obd, &tgt_uuid);
1150                 GOTO(out, rc);
1151         default: {
1152                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1153                 GOTO(out, rc = -EINVAL);
1154         }
1155         }
1156 out:
1157         RETURN(rc);
1158 }
1159
1160 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1161                       __u64 max_age, __u32 flags)
1162 {
1163         struct lmv_obd        *lmv = &obd->u.lmv;
1164         struct obd_statfs     *temp;
1165         int                    rc = 0;
1166         int                    i;
1167         ENTRY;
1168
1169         rc = lmv_check_connect(obd);
1170         if (rc)
1171                 RETURN(rc);
1172
1173         OBD_ALLOC(temp, sizeof(*temp));
1174         if (temp == NULL)
1175                 RETURN(-ENOMEM);
1176
1177         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1178                 if (lmv->tgts[i].ltd_exp == NULL)
1179                         continue;
1180
1181                 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp,
1182                                 max_age, flags);
1183                 if (rc) {
1184                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1185                                lmv->tgts[i].ltd_exp->exp_obd->obd_name,
1186                                rc);
1187                         GOTO(out_free_temp, rc);
1188                 }
1189                 if (i == 0) {
1190                         *osfs = *temp;
1191                 } else {
1192                         osfs->os_bavail += temp->os_bavail;
1193                         osfs->os_blocks += temp->os_blocks;
1194                         osfs->os_ffree += temp->os_ffree;
1195                         osfs->os_files += temp->os_files;
1196                 }
1197         }
1198
1199         EXIT;
1200 out_free_temp:
1201         OBD_FREE(temp, sizeof(*temp));
1202         return rc;
1203 }
1204
1205 static int lmv_getstatus(struct obd_export *exp,
1206                          struct lu_fid *fid,
1207                          struct obd_capa **pc)
1208 {
1209         struct obd_device    *obd = exp->exp_obd;
1210         struct lmv_obd       *lmv = &obd->u.lmv;
1211         int                   rc;
1212         ENTRY;
1213
1214         rc = lmv_check_connect(obd);
1215         if (rc)
1216                 RETURN(rc);
1217
1218         rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc);
1219         RETURN(rc);
1220 }
1221
1222 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1223                         struct obd_capa *oc, obd_valid valid, const char *name,
1224                         const char *input, int input_size, int output_size,
1225                         int flags, struct ptlrpc_request **request)
1226 {
1227         struct obd_device      *obd = exp->exp_obd;
1228         struct lmv_obd         *lmv = &obd->u.lmv;
1229         struct lmv_tgt_desc    *tgt;
1230         int                     rc;
1231         ENTRY;
1232
1233         rc = lmv_check_connect(obd);
1234         if (rc)
1235                 RETURN(rc);
1236
1237         tgt = lmv_find_target(lmv, fid);
1238         if (IS_ERR(tgt))
1239                 RETURN(PTR_ERR(tgt));
1240
1241         rc = md_getxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1242                          input_size, output_size, flags, request);
1243
1244         RETURN(rc);
1245 }
1246
1247 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1248                         struct obd_capa *oc, obd_valid valid, const char *name,
1249                         const char *input, int input_size, int output_size,
1250                         int flags, __u32 suppgid,
1251                         struct ptlrpc_request **request)
1252 {
1253         struct obd_device      *obd = exp->exp_obd;
1254         struct lmv_obd         *lmv = &obd->u.lmv;
1255         struct lmv_tgt_desc    *tgt;
1256         int                     rc;
1257         ENTRY;
1258
1259         rc = lmv_check_connect(obd);
1260         if (rc)
1261                 RETURN(rc);
1262
1263         tgt = lmv_find_target(lmv, fid);
1264         if (IS_ERR(tgt))
1265                 RETURN(PTR_ERR(tgt));
1266
1267         rc = md_setxattr(tgt->ltd_exp, fid, oc, valid, name, input,
1268                          input_size, output_size, flags, suppgid,
1269                          request);
1270
1271         RETURN(rc);
1272 }
1273
1274 static int lmv_getattr(struct obd_export *exp, const struct lu_fid *fid,
1275                        struct obd_capa *oc, obd_valid valid, int ea_size,
1276                        struct ptlrpc_request **request)
1277 {
1278         struct obd_device       *obd = exp->exp_obd;
1279         struct lmv_obd          *lmv = &obd->u.lmv;
1280         struct lmv_tgt_desc     *tgt;
1281         struct lmv_object       *obj;
1282         int                      rc;
1283         int                      i;
1284         ENTRY;
1285
1286         rc = lmv_check_connect(obd);
1287         if (rc)
1288                 RETURN(rc);
1289
1290         tgt = lmv_find_target(lmv, fid);
1291         if (IS_ERR(tgt))
1292                 RETURN(PTR_ERR(tgt));
1293
1294         rc = md_getattr(tgt->ltd_exp, fid, oc, valid, ea_size, request);
1295         if (rc)
1296                 RETURN(rc);
1297
1298         obj = lmv_object_find_lock(obd, fid);
1299
1300         CDEBUG(D_INODE, "GETATTR for "DFID" %s\n", PFID(fid),
1301                obj ? "(split)" : "");
1302
1303         /*
1304          * If object is split, then we loop over all the slaves and gather size
1305          * attribute. In ideal world we would have to gather also mds field from
1306          * all slaves, as object is spread over the cluster and this is
1307          * definitely interesting information and it is not good to loss it,
1308          * but...
1309          */
1310         if (obj) {
1311                 struct mdt_body *body;
1312
1313                 if (*request == NULL) {
1314                         lmv_object_put(obj);
1315                         RETURN(rc);
1316                 }
1317
1318                 body = req_capsule_server_get(&(*request)->rq_pill,
1319                                               &RMF_MDT_BODY);
1320                 LASSERT(body != NULL);
1321
1322                 for (i = 0; i < obj->lo_objcount; i++) {
1323                         if (lmv->tgts[i].ltd_exp == NULL) {
1324                                 CWARN("%s: NULL export for %d\n",
1325                                       obd->obd_name, i);
1326                                 continue;
1327                         }
1328
1329                         /*
1330                          * Skip master object.
1331                          */
1332                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid))
1333                                 continue;
1334
1335                         body->size += obj->lo_stripes[i].ls_size;
1336                 }
1337
1338                 lmv_object_put_unlock(obj);
1339         }
1340
1341         RETURN(rc);
1342 }
1343
1344 static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1345                              ldlm_iterator_t it, void *data)
1346 {
1347         struct obd_device   *obd = exp->exp_obd;
1348         struct lmv_obd      *lmv = &obd->u.lmv;
1349         int                  i;
1350         int                  rc;
1351         ENTRY;
1352
1353         rc = lmv_check_connect(obd);
1354         if (rc)
1355                 RETURN(rc);
1356
1357         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1358
1359         /*
1360          * With CMD every object can have two locks in different namespaces:
1361          * lookup lock in space of mds storing direntry and update/open lock in
1362          * space of mds storing inode.
1363          */
1364         for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1365                 md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data);
1366
1367         RETURN(0);
1368 }
1369
1370 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1371                      struct md_open_data *mod, struct ptlrpc_request **request)
1372 {
1373         struct obd_device     *obd = exp->exp_obd;
1374         struct lmv_obd        *lmv = &obd->u.lmv;
1375         struct lmv_tgt_desc   *tgt;
1376         int                    rc;
1377         ENTRY;
1378
1379         rc = lmv_check_connect(obd);
1380         if (rc)
1381                 RETURN(rc);
1382
1383         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1384         if (IS_ERR(tgt))
1385                 RETURN(PTR_ERR(tgt));
1386
1387         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1388         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1389         RETURN(rc);
1390 }
1391
1392 /**
1393  * Called in the case MDS returns -ERESTART on create on open, what means that
1394  * directory is split and its LMV presentation object has to be updated.
1395  */
1396 int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid)
1397 {
1398         struct obd_device       *obd = exp->exp_obd;
1399         struct lmv_obd          *lmv = &obd->u.lmv;
1400         struct ptlrpc_request   *req = NULL;
1401         struct lmv_tgt_desc     *tgt;
1402         struct lmv_object       *obj;
1403         struct lustre_md         md;
1404         int                      mealen;
1405         int                      rc;
1406         __u64                    valid;
1407         ENTRY;
1408
1409         md.mea = NULL;
1410         mealen = lmv_get_easize(lmv);
1411
1412         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA;
1413
1414         tgt = lmv_find_target(lmv, fid);
1415         if (IS_ERR(tgt))
1416                 RETURN(PTR_ERR(tgt));
1417
1418         /*
1419          * Time to update mea of parent fid.
1420          */
1421         rc = md_getattr(tgt->ltd_exp, fid, NULL, valid, mealen, &req);
1422         if (rc) {
1423                 CERROR("md_getattr() failed, error %d\n", rc);
1424                 GOTO(cleanup, rc);
1425         }
1426
1427         rc = md_get_lustre_md(tgt->ltd_exp, req, NULL, exp, &md);
1428         if (rc) {
1429                 CERROR("md_get_lustre_md() failed, error %d\n", rc);
1430                 GOTO(cleanup, rc);
1431         }
1432
1433         if (md.mea == NULL)
1434                 GOTO(cleanup, rc = -ENODATA);
1435
1436         obj = lmv_object_create(exp, fid, md.mea);
1437         if (IS_ERR(obj))
1438                 rc = PTR_ERR(obj);
1439         else
1440                 lmv_object_put(obj);
1441
1442         obd_free_memmd(exp, (void *)&md.mea);
1443         EXIT;
1444 cleanup:
1445         if (req)
1446                 ptlrpc_req_finished(req);
1447         return rc;
1448 }
1449
1450 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1451                const void *data, int datalen, int mode, __u32 uid,
1452                __u32 gid, cfs_cap_t cap_effective, __u64 rdev,
1453                struct ptlrpc_request **request)
1454 {
1455         struct obd_device       *obd = exp->exp_obd;
1456         struct lmv_obd          *lmv = &obd->u.lmv;
1457         struct lmv_tgt_desc     *tgt;
1458         struct lmv_object       *obj;
1459         int                      rc;
1460         int                      loop = 0;
1461         int                      sidx;
1462         ENTRY;
1463
1464         rc = lmv_check_connect(obd);
1465         if (rc)
1466                 RETURN(rc);
1467
1468         if (!lmv->desc.ld_active_tgt_count)
1469                 RETURN(-EIO);
1470 repeat:
1471         ++loop;
1472         LASSERT(loop <= 2);
1473
1474         obj = lmv_object_find(obd, &op_data->op_fid1);
1475         if (obj) {
1476                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1477                                        op_data->op_name, op_data->op_namelen);
1478                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1479                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
1480                 op_data->op_mds = obj->lo_stripes[sidx].ls_mds;
1481                 tgt = lmv_get_target(lmv, op_data->op_mds);
1482                 lmv_object_put(obj);
1483         } else {
1484                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1485                 op_data->op_bias |= MDS_CHECK_SPLIT;
1486                 op_data->op_mds = tgt->ltd_idx;
1487         }
1488
1489         if (IS_ERR(tgt))
1490                 RETURN(PTR_ERR(tgt));
1491
1492         rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data);
1493         if (rc == -ERESTART)
1494                 goto repeat;
1495         else if (rc)
1496                 RETURN(rc);
1497
1498         CDEBUG(D_INODE, "CREATE '%*s' on "DFID" -> mds #%x\n",
1499                op_data->op_namelen, op_data->op_name, PFID(&op_data->op_fid1),
1500                op_data->op_mds);
1501
1502         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1503         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1504                        cap_effective, rdev, request);
1505         if (rc == 0) {
1506                 if (*request == NULL)
1507                         RETURN(rc);
1508                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1509         } else if (rc == -ERESTART) {
1510                 LASSERT(*request != NULL);
1511                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1512                           "Got -ERESTART during create!\n");
1513                 ptlrpc_req_finished(*request);
1514                 *request = NULL;
1515
1516                 /*
1517                  * Directory got split. Time to update local object and repeat
1518                  * the request with proper MDS.
1519                  */
1520                 rc = lmv_handle_split(exp, &op_data->op_fid1);
1521                 if (rc == 0) {
1522                         rc = lmv_allocate_slaves(obd, &op_data->op_fid1,
1523                                                  op_data, &op_data->op_fid2);
1524                         if (rc)
1525                                 RETURN(rc);
1526                         goto repeat;
1527                 }
1528         }
1529         RETURN(rc);
1530 }
1531
1532 static int lmv_done_writing(struct obd_export *exp,
1533                             struct md_op_data *op_data,
1534                             struct md_open_data *mod)
1535 {
1536         struct obd_device     *obd = exp->exp_obd;
1537         struct lmv_obd        *lmv = &obd->u.lmv;
1538         struct lmv_tgt_desc   *tgt;
1539         int                    rc;
1540         ENTRY;
1541
1542         rc = lmv_check_connect(obd);
1543         if (rc)
1544                 RETURN(rc);
1545
1546         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1547         if (IS_ERR(tgt))
1548                 RETURN(PTR_ERR(tgt));
1549
1550         rc = md_done_writing(tgt->ltd_exp, op_data, mod);
1551         RETURN(rc);
1552 }
1553
1554 static int
1555 lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1556                    struct lookup_intent *it, struct md_op_data *op_data,
1557                    struct lustre_handle *lockh, void *lmm, int lmmsize)
1558 {
1559         struct obd_device     *obd = exp->exp_obd;
1560         struct lmv_obd        *lmv = &obd->u.lmv;
1561         struct lmv_stripe_md  *mea = op_data->op_mea1;
1562         struct md_op_data     *op_data2;
1563         struct lmv_tgt_desc   *tgt;
1564         int                    i;
1565         int                    rc = 0;
1566         ENTRY;
1567
1568         OBD_ALLOC_PTR(op_data2);
1569         if (op_data2 == NULL)
1570                 RETURN(-ENOMEM);
1571
1572         LASSERT(mea != NULL);
1573         for (i = 0; i < mea->mea_count; i++) {
1574                 memset(op_data2, 0, sizeof(*op_data2));
1575                 op_data2->op_fid1 = mea->mea_ids[i];
1576                 op_data2->op_bias = 0;
1577
1578                 tgt = lmv_find_target(lmv, &op_data2->op_fid1);
1579                 if (IS_ERR(tgt))
1580                         GOTO(cleanup, rc = PTR_ERR(tgt));
1581
1582                 if (tgt->ltd_exp == NULL)
1583                         continue;
1584
1585                 rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data2,
1586                                 lockh + i, lmm, lmmsize, NULL, 0);
1587
1588                 CDEBUG(D_INODE, "Take lock on slave "DFID" -> %d/%d\n",
1589                        PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status);
1590
1591                 if (rc)
1592                         GOTO(cleanup, rc);
1593
1594                 if (it->d.lustre.it_data) {
1595                         struct ptlrpc_request *req;
1596                         req = (struct ptlrpc_request *)it->d.lustre.it_data;
1597                         ptlrpc_req_finished(req);
1598                 }
1599
1600                 if (it->d.lustre.it_status)
1601                         GOTO(cleanup, rc = it->d.lustre.it_status);
1602         }
1603
1604         EXIT;
1605 cleanup:
1606         OBD_FREE_PTR(op_data2);
1607
1608         if (rc != 0) {
1609                 /*
1610                  * Drop all taken locks.
1611                  */
1612                 while (--i >= 0) {
1613                         if (lockh[i].cookie)
1614                                 ldlm_lock_decref(lockh + i, einfo->ei_mode);
1615                         lockh[i].cookie = 0;
1616                 }
1617         }
1618         return rc;
1619 }
1620
1621 static int
1622 lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1623                    struct lookup_intent *it, struct md_op_data *op_data,
1624                    struct lustre_handle *lockh, void *lmm, int lmmsize,
1625                    int extra_lock_flags)
1626 {
1627         struct ptlrpc_request      *req = it->d.lustre.it_data;
1628         struct obd_device          *obd = exp->exp_obd;
1629         struct lmv_obd             *lmv = &obd->u.lmv;
1630         struct lustre_handle        plock;
1631         struct lmv_tgt_desc        *tgt;
1632         struct md_op_data          *rdata;
1633         struct lu_fid               fid1;
1634         struct mdt_body            *body;
1635         int                         rc = 0;
1636         int                         pmode;
1637         ENTRY;
1638
1639         body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
1640         LASSERT(body != NULL);
1641
1642         if (!(body->valid & OBD_MD_MDS))
1643                 RETURN(0);
1644
1645         CDEBUG(D_INODE, "REMOTE_ENQUEUE '%s' on "DFID" -> "DFID"\n",
1646                LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1));
1647
1648         /*
1649          * We got LOOKUP lock, but we really need attrs.
1650          */
1651         pmode = it->d.lustre.it_lock_mode;
1652         LASSERT(pmode != 0);
1653         memcpy(&plock, lockh, sizeof(plock));
1654         it->d.lustre.it_lock_mode = 0;
1655         it->d.lustre.it_data = NULL;
1656         fid1 = body->fid1;
1657
1658         it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE;
1659         ptlrpc_req_finished(req);
1660
1661         tgt = lmv_find_target(lmv, &fid1);
1662         if (IS_ERR(tgt))
1663                 GOTO(out, rc = PTR_ERR(tgt));
1664
1665         OBD_ALLOC_PTR(rdata);
1666         if (rdata == NULL)
1667                 GOTO(out, rc = -ENOMEM);
1668
1669         rdata->op_fid1 = fid1;
1670         rdata->op_bias = MDS_CROSS_REF;
1671
1672         rc = md_enqueue(tgt->ltd_exp, einfo, it, rdata, lockh,
1673                         lmm, lmmsize, NULL, extra_lock_flags);
1674         OBD_FREE_PTR(rdata);
1675         EXIT;
1676 out:
1677         ldlm_lock_decref(&plock, pmode);
1678         return rc;
1679 }
1680
1681 static int
1682 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1683             struct lookup_intent *it, struct md_op_data *op_data,
1684             struct lustre_handle *lockh, void *lmm, int lmmsize,
1685             struct ptlrpc_request **req, int extra_lock_flags)
1686 {
1687         struct obd_device        *obd = exp->exp_obd;
1688         struct lmv_obd           *lmv = &obd->u.lmv;
1689         struct lmv_tgt_desc      *tgt;
1690         struct lmv_object        *obj;
1691         int                       sidx;
1692         int                       rc;
1693         ENTRY;
1694
1695         rc = lmv_check_connect(obd);
1696         if (rc)
1697                 RETURN(rc);
1698
1699         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1700                LL_IT2STR(it), PFID(&op_data->op_fid1));
1701
1702         if (op_data->op_mea1 && it && it->it_op == IT_UNLINK) {
1703                 rc = lmv_enqueue_slaves(exp, einfo, it, op_data,
1704                                         lockh, lmm, lmmsize);
1705                 RETURN(rc);
1706         }
1707
1708         obj = lmv_object_find(obd, &op_data->op_fid1);
1709         if (obj && op_data->op_namelen) {
1710                 sidx = raw_name2idx(obj->lo_hashtype,
1711                                        obj->lo_objcount,
1712                                        (char *)op_data->op_name,
1713                                        op_data->op_namelen);
1714                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
1715                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1716         } else {
1717                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1718         }
1719         if (obj)
1720                 lmv_object_put(obj);
1721
1722         if (IS_ERR(tgt))
1723                 RETURN(PTR_ERR(tgt));
1724
1725         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%d\n",
1726                LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1727
1728         rc = md_enqueue(tgt->ltd_exp, einfo, it, op_data, lockh,
1729                         lmm, lmmsize, req, extra_lock_flags);
1730
1731         if (rc == 0 && it && it->it_op == IT_OPEN) {
1732                 rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh,
1733                                         lmm, lmmsize, extra_lock_flags);
1734         }
1735         RETURN(rc);
1736 }
1737
1738 static int
1739 lmv_getattr_name(struct obd_export *exp, const struct lu_fid *fid,
1740                  struct obd_capa *oc, const char *name, int namelen,
1741                  obd_valid valid, int ea_size, __u32 suppgid,
1742                  struct ptlrpc_request **request)
1743 {
1744         struct ptlrpc_request   *req = NULL;
1745         struct obd_device       *obd = exp->exp_obd;
1746         struct lmv_obd          *lmv = &obd->u.lmv;
1747         struct lu_fid            rid = *fid;
1748         struct lmv_tgt_desc     *tgt;
1749         struct mdt_body         *body;
1750         struct lmv_object       *obj;
1751         int                      rc;
1752         int                      loop = 0;
1753         int                      sidx;
1754         ENTRY;
1755
1756         rc = lmv_check_connect(obd);
1757         if (rc)
1758                 RETURN(rc);
1759
1760 repeat:
1761         ++loop;
1762         LASSERT(loop <= 2);
1763         obj = lmv_object_find(obd, &rid);
1764         if (obj) {
1765                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
1766                                        name, namelen - 1);
1767                 rid = obj->lo_stripes[sidx].ls_fid;
1768                 tgt = lmv_get_target(lmv, obj->lo_stripes[sidx].ls_mds);
1769                 valid &= ~OBD_MD_FLCKSPLIT;
1770                 lmv_object_put(obj);
1771         } else {
1772                 tgt = lmv_find_target(lmv, &rid);
1773                 valid |= OBD_MD_FLCKSPLIT;
1774         }
1775         if (IS_ERR(tgt))
1776                 RETURN(PTR_ERR(tgt));
1777
1778         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" - "DFID" -> mds #%d\n",
1779                namelen, name, PFID(fid), PFID(&rid), tgt->ltd_idx);
1780
1781         rc = md_getattr_name(tgt->ltd_exp, &rid, oc, name, namelen, valid,
1782                              ea_size, suppgid, request);
1783         if (rc == 0) {
1784                 body = req_capsule_server_get(&(*request)->rq_pill,
1785                                               &RMF_MDT_BODY);
1786                 LASSERT(body != NULL);
1787
1788                 if (body->valid & OBD_MD_MDS) {
1789                         rid = body->fid1;
1790                         CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1791                                PFID(&rid));
1792
1793                         tgt = lmv_find_target(lmv, &rid);
1794                         if (IS_ERR(tgt)) {
1795                                 ptlrpc_req_finished(*request);
1796                                 RETURN(PTR_ERR(tgt));
1797                         }
1798
1799                         rc = md_getattr_name(tgt->ltd_exp, &rid, NULL, NULL,
1800                                              1, valid | OBD_MD_FLCROSSREF,
1801                                              ea_size, suppgid, &req);
1802                         ptlrpc_req_finished(*request);
1803                         *request = req;
1804                 }
1805         } else if (rc == -ERESTART) {
1806                 LASSERT(*request != NULL);
1807                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1808                           "Got -ERESTART during getattr!\n");
1809                 ptlrpc_req_finished(*request);
1810                 *request = NULL;
1811
1812                 /*
1813                  * Directory got split. Time to update local object and repeat
1814                  * the request with proper MDS.
1815                  */
1816                 rc = lmv_handle_split(exp, &rid);
1817                 if (rc == 0)
1818                         goto repeat;
1819         }
1820         RETURN(rc);
1821 }
1822
1823 #define md_op_data_fid(op_data, fl)                     \
1824         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1825          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1826          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1827          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1828          NULL)
1829
1830 static int lmv_early_cancel_slaves(struct obd_export *exp,
1831                                    struct md_op_data *op_data, int op_tgt,
1832                                    ldlm_mode_t mode, int bits, int flag)
1833 {
1834         struct obd_device      *obd = exp->exp_obd;
1835         struct lmv_obd         *lmv = &obd->u.lmv;
1836         ldlm_policy_data_t      policy = {{0}};
1837         struct lu_fid          *op_fid;
1838         struct lu_fid          *st_fid;
1839         struct lmv_tgt_desc    *tgt;
1840         struct lmv_object      *obj;
1841         int                     rc = 0;
1842         int                     i;
1843         ENTRY;
1844
1845         op_fid = md_op_data_fid(op_data, flag);
1846         if (!fid_is_sane(op_fid))
1847                 RETURN(0);
1848
1849         obj = lmv_object_find(obd, op_fid);
1850         if (obj == NULL)
1851                 RETURN(-EALREADY);
1852
1853         policy.l_inodebits.bits = bits;
1854         for (i = 0; i < obj->lo_objcount; i++) {
1855                 tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
1856                 st_fid = &obj->lo_stripes[i].ls_fid;
1857                 if (op_tgt != tgt->ltd_idx) {
1858                         CDEBUG(D_INODE, "EARLY_CANCEL slave "DFID" -> mds #%d\n",
1859                                PFID(st_fid), tgt->ltd_idx);
1860                         rc = md_cancel_unused(tgt->ltd_exp, st_fid, &policy,
1861                                               mode, LDLM_FL_ASYNC, NULL);
1862                         if (rc)
1863                                 GOTO(out_put_obj, rc);
1864                 } else {
1865                         CDEBUG(D_INODE,
1866                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1867                                op_tgt, PFID(st_fid));
1868                         /*
1869                          * Do not cancel locks for operation target, they will
1870                          * be handled later in underlaying layer when calling
1871                          * function we run on behalf of.
1872                          */
1873                         *op_fid = *st_fid;
1874                         op_data->op_flags |= flag;
1875                 }
1876         }
1877         EXIT;
1878 out_put_obj:
1879         lmv_object_put(obj);
1880         return rc;
1881 }
1882
1883 static int lmv_early_cancel(struct obd_export *exp, struct md_op_data *op_data,
1884                             int op_tgt, ldlm_mode_t mode, int bits, int flag)
1885 {
1886         struct lu_fid          *fid = md_op_data_fid(op_data, flag);
1887         struct obd_device      *obd = exp->exp_obd;
1888         struct lmv_obd         *lmv = &obd->u.lmv;
1889         struct lmv_tgt_desc    *tgt;
1890         ldlm_policy_data_t      policy = {{0}};
1891         struct lmv_object      *obj;
1892         int                     rc = 0;
1893         ENTRY;
1894
1895         if (!fid_is_sane(fid))
1896                 RETURN(0);
1897
1898         obj = lmv_object_find(obd, fid);
1899         if (obj) {
1900                 rc = lmv_early_cancel_slaves(exp, op_data, op_tgt, mode,
1901                                              bits, flag);
1902                 lmv_object_put(obj);
1903         } else {
1904                 tgt = lmv_find_target(lmv, fid);
1905                 if (IS_ERR(tgt))
1906                         RETURN(PTR_ERR(tgt));
1907
1908                 if (tgt->ltd_idx != op_tgt) {
1909                         CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1910                         policy.l_inodebits.bits = bits;
1911                         rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1912                                               mode, LDLM_FL_ASYNC, NULL);
1913                 } else {
1914                         CDEBUG(D_INODE,
1915                                "EARLY_CANCEL skip operation target %d on "DFID"\n",
1916                                op_tgt, PFID(fid));
1917                         op_data->op_flags |= flag;
1918                         rc = 0;
1919                 }
1920
1921         }
1922         RETURN(rc);
1923 }
1924
1925 /*
1926  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1927  * op_data->op_fid2
1928  */
1929 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1930                     struct ptlrpc_request **request)
1931 {
1932         struct obd_device       *obd = exp->exp_obd;
1933         struct lmv_obd          *lmv = &obd->u.lmv;
1934         struct lmv_tgt_desc     *tgt;
1935         struct lmv_object       *obj;
1936         int                      rc;
1937         int                      loop = 0;
1938         mdsno_t                  mds;
1939         int                      sidx;
1940         ENTRY;
1941
1942         rc = lmv_check_connect(obd);
1943         if (rc)
1944                 RETURN(rc);
1945
1946 repeat:
1947         ++loop;
1948         LASSERT(loop <= 2);
1949         LASSERT(op_data->op_namelen != 0);
1950
1951         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
1952                PFID(&op_data->op_fid2), op_data->op_namelen,
1953                op_data->op_name, PFID(&op_data->op_fid1));
1954
1955         obj = lmv_object_find(obd, &op_data->op_fid2);
1956         if (obj) {
1957                 sidx = raw_name2idx(obj->lo_hashtype,
1958                                     obj->lo_objcount,
1959                                     op_data->op_name,
1960                                     op_data->op_namelen);
1961                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
1962                 mds = obj->lo_stripes[sidx].ls_mds;
1963                 lmv_object_put(obj);
1964         } else {
1965                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds);
1966                 if (rc)
1967                         RETURN(rc);
1968         }
1969
1970         CDEBUG(D_INODE, "Forward to mds #%x ("DFID")\n",
1971                mds, PFID(&op_data->op_fid1));
1972
1973         op_data->op_fsuid = current->fsuid;
1974         op_data->op_fsgid = current->fsgid;
1975         op_data->op_cap = cfs_curproc_cap_pack();
1976         tgt = lmv_get_target(lmv, mds);
1977
1978         /*
1979          * Cancel UPDATE lock on child (fid1).
1980          */
1981         op_data->op_flags |= MF_MDC_CANCEL_FID2;
1982         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
1983                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1984         if (rc == 0)
1985                 rc = md_link(tgt->ltd_exp, op_data, request);
1986         if (rc == -ERESTART) {
1987                 LASSERT(*request != NULL);
1988                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
1989                           "Got -ERESTART during link!\n");
1990                 ptlrpc_req_finished(*request);
1991                 *request = NULL;
1992
1993                 /*
1994                  * Directory got split. Time to update local object and repeat
1995                  * the request with proper MDS.
1996                  */
1997                 rc = lmv_handle_split(exp, &op_data->op_fid2);
1998                 if (rc == 0)
1999                         goto repeat;
2000         }
2001
2002         RETURN(rc);
2003 }
2004
2005 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2006                       const char *old, int oldlen, const char *new, int newlen,
2007                       struct ptlrpc_request **request)
2008 {
2009         struct obd_device       *obd = exp->exp_obd;
2010         struct lmv_obd          *lmv = &obd->u.lmv;
2011         struct lmv_tgt_desc     *src_tgt;
2012         struct lmv_tgt_desc     *tgt_tgt;
2013         int                      rc;
2014         int                      sidx;
2015         int                      loop = 0;
2016         struct lmv_object       *obj;
2017         mdsno_t                  mds1;
2018         mdsno_t                  mds2;
2019         ENTRY;
2020
2021         LASSERT(oldlen != 0);
2022
2023         CDEBUG(D_INODE, "RENAME %*s in "DFID" to %*s in "DFID"\n",
2024                oldlen, old, PFID(&op_data->op_fid1),
2025                newlen, new, PFID(&op_data->op_fid2));
2026
2027         rc = lmv_check_connect(obd);
2028         if (rc)
2029                 RETURN(rc);
2030
2031 repeat:
2032         ++loop;
2033         LASSERT(loop <= 2);
2034         obj = lmv_object_find(obd, &op_data->op_fid1);
2035         if (obj) {
2036                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2037                                     (char *)old, oldlen);
2038                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2039                 mds1 = obj->lo_stripes[sidx].ls_mds;
2040                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid1));
2041                 lmv_object_put(obj);
2042         } else {
2043                 rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1);
2044                 if (rc)
2045                         RETURN(rc);
2046         }
2047
2048         obj = lmv_object_find(obd, &op_data->op_fid2);
2049         if (obj) {
2050                 /*
2051                  * Directory is already split, so we have to forward request to
2052                  * the right MDS.
2053                  */
2054                 sidx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount,
2055                                        (char *)new, newlen);
2056
2057                 mds2 = obj->lo_stripes[sidx].ls_mds;
2058                 op_data->op_fid2 = obj->lo_stripes[sidx].ls_fid;
2059                 CDEBUG(D_INODE, "Parent obj "DFID"\n", PFID(&op_data->op_fid2));
2060                 lmv_object_put(obj);
2061         } else {
2062                 rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2);
2063                 if (rc)
2064                         RETURN(rc);
2065         }
2066
2067         op_data->op_fsuid = current->fsuid;
2068         op_data->op_fsgid = current->fsgid;
2069         op_data->op_cap = cfs_curproc_cap_pack();
2070
2071         src_tgt = lmv_get_target(lmv, mds1);
2072         tgt_tgt = lmv_get_target(lmv, mds2);
2073
2074         /*
2075          * LOOKUP lock on src child (fid3) should also be cancelled for
2076          * src_tgt in mdc_rename.
2077          */
2078         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2079
2080         /*
2081          * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2082          * own target.
2083          */
2084         rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2085                               LCK_EX, MDS_INODELOCK_UPDATE,
2086                               MF_MDC_CANCEL_FID2);
2087
2088         /*
2089          * Cancel LOOKUP locks on tgt child (fid4) for parent tgt_tgt.
2090          */
2091         if (rc == 0) {
2092                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2093                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2094                                       MF_MDC_CANCEL_FID4);
2095         }
2096
2097         /*
2098          * Cancel all the locks on tgt child (fid4).
2099          */
2100         if (rc == 0)
2101                 rc = lmv_early_cancel(exp, op_data, src_tgt->ltd_idx,
2102                                       LCK_EX, MDS_INODELOCK_FULL,
2103                                       MF_MDC_CANCEL_FID4);
2104
2105         if (rc == 0)
2106                 rc = md_rename(src_tgt->ltd_exp, op_data, old, oldlen,
2107                                new, newlen, request);
2108
2109         if (rc == -ERESTART) {
2110                 LASSERT(*request != NULL);
2111                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2112                           "Got -ERESTART during rename!\n");
2113                 ptlrpc_req_finished(*request);
2114                 *request = NULL;
2115
2116                 /*
2117                  * Directory got split. Time to update local object and repeat
2118                  * the request with proper MDS.
2119                  */
2120                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2121                 if (rc == 0)
2122                         goto repeat;
2123         }
2124         RETURN(rc);
2125 }
2126
2127 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2128                        void *ea, int ealen, void *ea2, int ea2len,
2129                        struct ptlrpc_request **request,
2130                        struct md_open_data **mod)
2131 {
2132         struct obd_device       *obd = exp->exp_obd;
2133         struct lmv_obd          *lmv = &obd->u.lmv;
2134         struct ptlrpc_request   *req;
2135         struct lmv_tgt_desc     *tgt;
2136         struct lmv_object       *obj;
2137         int                      rc = 0;
2138         int                      i;
2139         ENTRY;
2140
2141         rc = lmv_check_connect(obd);
2142         if (rc)
2143                 RETURN(rc);
2144
2145         obj = lmv_object_find(obd, &op_data->op_fid1);
2146
2147         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x%s\n",
2148                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2149                obj ? ", split" : "");
2150
2151         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2152         if (obj) {
2153                 for (i = 0; i < obj->lo_objcount; i++) {
2154                         op_data->op_fid1 = obj->lo_stripes[i].ls_fid;
2155
2156                         tgt = lmv_get_target(lmv, obj->lo_stripes[i].ls_mds);
2157                         if (IS_ERR(tgt)) {
2158                                 rc = PTR_ERR(tgt);
2159                                 break;
2160                         }
2161
2162                         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen,
2163                                         ea2, ea2len, &req, mod);
2164
2165                         if (lu_fid_eq(&obj->lo_fid, &obj->lo_stripes[i].ls_fid)) {
2166                                 /*
2167                                  * This is master object and this request should
2168                                  * be returned back to llite.
2169                                  */
2170                                 *request = req;
2171                         } else {
2172                                 ptlrpc_req_finished(req);
2173                         }
2174
2175                         if (rc)
2176                                 break;
2177                 }
2178                 lmv_object_put(obj);
2179         } else {
2180                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2181                 if (IS_ERR(tgt))
2182                         RETURN(PTR_ERR(tgt));
2183
2184                 rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, ea2,
2185                                 ea2len, request, mod);
2186         }
2187         RETURN(rc);
2188 }
2189
2190 static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid,
2191                     struct obd_capa *oc, struct ptlrpc_request **request)
2192 {
2193         struct obd_device         *obd = exp->exp_obd;
2194         struct lmv_obd            *lmv = &obd->u.lmv;
2195         struct lmv_tgt_desc       *tgt;
2196         int                        rc;
2197         ENTRY;
2198
2199         rc = lmv_check_connect(obd);
2200         if (rc)
2201                 RETURN(rc);
2202
2203         tgt = lmv_find_target(lmv, fid);
2204         if (IS_ERR(tgt))
2205                 RETURN(PTR_ERR(tgt));
2206
2207         rc = md_sync(tgt->ltd_exp, fid, oc, request);
2208         RETURN(rc);
2209 }
2210
2211 /**
2212  * Main purpose of LMV blocking ast is to remove split directory LMV
2213  * presentation object (struct lmv_object) attached to the lock being revoked.
2214  */
2215 int lmv_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
2216                      void *data, int flag)
2217 {
2218         struct lustre_handle    lockh;
2219         struct lmv_object      *obj;
2220         int                     rc;
2221         ENTRY;
2222
2223         switch (flag) {
2224         case LDLM_CB_BLOCKING:
2225                 ldlm_lock2handle(lock, &lockh);
2226                 rc = ldlm_cli_cancel(&lockh);
2227                 if (rc < 0) {
2228                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
2229                         RETURN(rc);
2230                 }
2231                 break;
2232         case LDLM_CB_CANCELING:
2233                 /*
2234                  * Time to drop cached attrs for split directory object
2235                  */
2236                 obj = lock->l_ast_data;
2237                 if (obj) {
2238                         CDEBUG(D_INODE, "Cancel %s on "LPU64"/"LPU64
2239                                ", master "DFID"\n",
2240                                lock->l_resource->lr_name.name[3] == 1 ?
2241                                "LOOKUP" : "UPDATE",
2242                                lock->l_resource->lr_name.name[0],
2243                                lock->l_resource->lr_name.name[1],
2244                                PFID(&obj->lo_fid));
2245                         lmv_object_put(obj);
2246                 }
2247                 break;
2248         default:
2249                 LBUG();
2250         }
2251         RETURN(0);
2252 }
2253
2254 static void lmv_hash_adjust(__u64 *hash, __u64 hash_adj)
2255 {
2256         __u64         val;
2257
2258         val = le64_to_cpu(*hash);
2259         if (val < hash_adj)
2260                 val += MAX_HASH_SIZE;
2261         if (val != DIR_END_OFF)
2262                 *hash = cpu_to_le64(val - hash_adj);
2263 }
2264
2265 static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid)
2266 {
2267         __u64              id;
2268         struct obd_import *imp;
2269
2270         /*
2271          * XXX: to get nid we assume that underlying obd device is mdc.
2272          */
2273         imp  = class_exp2cliimp(exp);
2274         id   = imp->imp_connection->c_self + fid_flatten(fid);
2275
2276         CDEBUG(D_INODE, "Readpage node rank: "LPX64" "DFID" "LPX64" "LPX64"\n",
2277                imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32));
2278
2279         return id ^ (id >> 32);
2280 }
2281
2282 static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid,
2283                         struct obd_capa *oc, __u64 offset64, struct page *page,
2284                         struct ptlrpc_request **request)
2285 {
2286         struct obd_device       *obd = exp->exp_obd;
2287         struct lmv_obd          *lmv = &obd->u.lmv;
2288         struct lu_fid            rid = *fid;
2289         struct lmv_object       *obj;
2290         __u64                    offset;
2291         __u64                    hash_adj = 0;
2292         __u32                    rank = 0;
2293         __u64                    seg_size = 0;
2294         __u64                    tgt_tmp = 0;
2295         int                      tgt_idx = 0;
2296         int                      tgt0_idx = 0;
2297         int                      rc;
2298         int                      nr = 0;
2299         struct lmv_stripe       *los;
2300         struct lmv_tgt_desc     *tgt;
2301         struct lu_dirpage       *dp;
2302         struct lu_dirent        *ent;
2303         ENTRY;
2304
2305         offset = offset64;
2306
2307         rc = lmv_check_connect(obd);
2308         if (rc)
2309                 RETURN(rc);
2310
2311         CDEBUG(D_INODE, "READPAGE at "LPX64" from "DFID"\n", offset, PFID(&rid));
2312
2313         /*
2314          * This case handle directory lookup in clustered metadata case (i.e.
2315          * split directory is located on multiple md servers.)
2316          * each server keeps directory entries for certain range of hashes.
2317          * E.g. we have N server and suppose hash range is 0 to MAX_HASH.
2318          * first server will keep records with hashes [ 0 ... MAX_HASH / N  - 1],
2319          * second one with hashes [MAX_HASH / N ... 2 * MAX_HASH / N] and
2320          * so on....
2321          *      readdir can simply start reading entries from 0 - N server in
2322          * order but that will not scale well as all client will request dir in
2323          * to server in same order.
2324          * Following algorithm does optimization:
2325          * Instead of doing readdir in 1, 2, ...., N order, client with a
2326          * rank R does readdir in R, R + 1, ..., N, 1, ... R - 1 order.
2327          * (every client has rank R)
2328          *      But ll_readdir() expect offset range [0 to MAX_HASH/N) but
2329          * since client ask dir from MDS{R} client has pages with offsets
2330          * [R*MAX_HASH/N ... (R + 1)*MAX_HASH/N] there for we do hash_adj
2331          * on hash  values that we get.
2332          */
2333         obj = lmv_object_find_lock(obd, fid);
2334         if (obj) {
2335                 nr       = obj->lo_objcount;
2336                 LASSERT(nr > 0);
2337                 seg_size = MAX_HASH_SIZE;
2338                 do_div(seg_size, nr);
2339                 los      = obj->lo_stripes;
2340                 tgt      = lmv_get_target(lmv, los[0].ls_mds);
2341                 rank     = lmv_node_rank(tgt->ltd_exp, fid) % nr;
2342                 tgt_tmp  = offset;
2343                 do_div(tgt_tmp, seg_size);
2344                 tgt0_idx = do_div(tgt_tmp,  nr);
2345                 tgt_idx  = (tgt0_idx + rank) % nr;
2346
2347                 if (tgt_idx < tgt0_idx)
2348                         /*
2349                          * Wrap around.
2350                          *
2351                          * Last segment has unusual length due to division
2352                          * rounding.
2353                          */
2354                         hash_adj = MAX_HASH_SIZE - seg_size * nr;
2355                 else
2356                         hash_adj = 0;
2357
2358                 hash_adj += rank * seg_size;
2359
2360                 CDEBUG(D_INODE, "Readpage hash adjustment: %x "LPX64" "
2361                        LPX64"/%x -> "LPX64"/%x\n", rank, hash_adj,
2362                        offset, tgt0_idx, offset + hash_adj, tgt_idx);
2363
2364                 offset = (offset + hash_adj) & MAX_HASH_SIZE;
2365                 rid = obj->lo_stripes[tgt_idx].ls_fid;
2366                 tgt = lmv_get_target(lmv, los[tgt_idx].ls_mds);
2367
2368                 CDEBUG(D_INODE, "Forward to "DFID" with offset %lu i %d\n",
2369                        PFID(&rid), (unsigned long)offset, tgt_idx);
2370         } else
2371                 tgt = lmv_find_target(lmv, &rid);
2372
2373         if (IS_ERR(tgt))
2374                 GOTO(cleanup, rc = PTR_ERR(tgt));
2375
2376         rc = md_readpage(tgt->ltd_exp, &rid, oc, offset, page, request);
2377         if (rc)
2378                 GOTO(cleanup, rc);
2379         if (obj) {
2380                 dp = cfs_kmap(page);
2381
2382                 lmv_hash_adjust(&dp->ldp_hash_start, hash_adj);
2383                 lmv_hash_adjust(&dp->ldp_hash_end,   hash_adj);
2384                 LASSERT(le64_to_cpu(dp->ldp_hash_start) <= offset64);
2385
2386                 for (ent = lu_dirent_start(dp); ent != NULL;
2387                      ent = lu_dirent_next(ent))
2388                         lmv_hash_adjust(&ent->lde_hash, hash_adj);
2389
2390                 if (tgt0_idx != nr - 1) {
2391                         __u64 end;
2392
2393                         end = le64_to_cpu(dp->ldp_hash_end);
2394                         if (end == DIR_END_OFF) {
2395                                 dp->ldp_hash_end = cpu_to_le32(seg_size *
2396                                                                (tgt0_idx + 1));
2397                                 CDEBUG(D_INODE,
2398                                        ""DFID" reset end "LPX64" tgt %d\n",
2399                                        PFID(&rid),
2400                                        le64_to_cpu(dp->ldp_hash_end), tgt_idx);
2401                         }
2402                 }
2403                 cfs_kunmap(page);
2404         }
2405         EXIT;
2406 cleanup:
2407         if (obj)
2408                 lmv_object_put_unlock(obj);
2409         return rc;
2410 }
2411
2412 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2413                       struct ptlrpc_request **request)
2414 {
2415         struct obd_device       *obd = exp->exp_obd;
2416         struct lmv_obd          *lmv = &obd->u.lmv;
2417         struct lmv_tgt_desc     *tgt = NULL;
2418         struct lmv_object       *obj;
2419         int                      rc;
2420         int                      sidx;
2421         int                      loop = 0;
2422         ENTRY;
2423
2424         rc = lmv_check_connect(obd);
2425         if (rc)
2426                 RETURN(rc);
2427
2428 repeat:
2429         ++loop;
2430         LASSERT(loop <= 2);
2431         LASSERT(op_data->op_namelen != 0);
2432
2433         obj = lmv_object_find(obd, &op_data->op_fid1);
2434         if (obj) {
2435                 sidx = raw_name2idx(obj->lo_hashtype,
2436                                     obj->lo_objcount,
2437                                     op_data->op_name,
2438                                     op_data->op_namelen);
2439                 op_data->op_bias &= ~MDS_CHECK_SPLIT;
2440                 op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2441                 tgt = lmv_get_target(lmv,
2442                                      obj->lo_stripes[sidx].ls_mds);
2443                 lmv_object_put(obj);
2444                 CDEBUG(D_INODE, "UNLINK '%*s' in "DFID" -> %u\n",
2445                        op_data->op_namelen, op_data->op_name,
2446                        PFID(&op_data->op_fid1), sidx);
2447         }
2448
2449         if (tgt == NULL) {
2450                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2451                 if (IS_ERR(tgt))
2452                         RETURN(PTR_ERR(tgt));
2453                 op_data->op_bias |= MDS_CHECK_SPLIT;
2454         }
2455
2456         op_data->op_fsuid = current->fsuid;
2457         op_data->op_fsgid = current->fsgid;
2458         op_data->op_cap = cfs_curproc_cap_pack();
2459
2460         /*
2461          * If child's fid is given, cancel unused locks for it if it is from
2462          * another export than parent.
2463          *
2464          * LOOKUP lock for child (fid3) should also be cancelled on parent
2465          * tgt_tgt in mdc_unlink().
2466          */
2467         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2468
2469         /*
2470          * Cancel FULL locks on child (fid3).
2471          */
2472         rc = lmv_early_cancel(exp, op_data, tgt->ltd_idx, LCK_EX,
2473                               MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2474
2475         if (rc == 0)
2476                 rc = md_unlink(tgt->ltd_exp, op_data, request);
2477
2478         if (rc == -ERESTART) {
2479                 LASSERT(*request != NULL);
2480                 DEBUG_REQ(D_WARNING|D_RPCTRACE, *request,
2481                           "Got -ERESTART during unlink!\n");
2482                 ptlrpc_req_finished(*request);
2483                 *request = NULL;
2484
2485                 /*
2486                  * Directory got split. Time to update local object and repeat
2487                  * the request with proper MDS.
2488                  */
2489                 rc = lmv_handle_split(exp, &op_data->op_fid1);
2490                 if (rc == 0)
2491                         goto repeat;
2492         }
2493         RETURN(rc);
2494 }
2495
2496 static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2497 {
2498         int        rc = 0;
2499
2500         switch (stage) {
2501         case OBD_CLEANUP_EARLY:
2502                 /* XXX: here should be calling obd_precleanup() down to
2503                  * stack. */
2504                 break;
2505         case OBD_CLEANUP_EXPORTS:
2506                 rc = obd_llog_finish(obd, 0);
2507                 if (rc != 0)
2508                         CERROR("failed to cleanup llogging subsystems\n");
2509                 break;
2510         default:
2511                 break;
2512         }
2513         RETURN(rc);
2514 }
2515
2516 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
2517                         void *key, __u32 *vallen, void *val,
2518                         struct lov_stripe_md *lsm)
2519 {
2520         struct obd_device       *obd;
2521         struct lmv_obd          *lmv;
2522         int                      rc = 0;
2523         ENTRY;
2524
2525         obd = class_exp2obd(exp);
2526         if (obd == NULL) {
2527                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2528                        exp->exp_handle.h_cookie);
2529                 RETURN(-EINVAL);
2530         }
2531
2532         lmv = &obd->u.lmv;
2533         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2534                 struct lmv_tgt_desc *tgts;
2535                 int i;
2536
2537                 rc = lmv_check_connect(obd);
2538                 if (rc)
2539                         RETURN(rc);
2540
2541                 LASSERT(*vallen == sizeof(__u32));
2542                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count;
2543                      i++, tgts++) {
2544
2545                         /*
2546                          * All tgts should be connected when this gets called.
2547                          */
2548                         if (!tgts || !tgts->ltd_exp) {
2549                                 CERROR("target not setup?\n");
2550                                 continue;
2551                         }
2552
2553                         if (!obd_get_info(tgts->ltd_exp, keylen, key,
2554                                           vallen, val, NULL))
2555                                 RETURN(0);
2556                 }
2557                 RETURN(-EINVAL);
2558         } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) {
2559                 rc = lmv_check_connect(obd);
2560                 if (rc)
2561                         RETURN(rc);
2562
2563                 /*
2564                  * Forwarding this request to first MDS, it should know LOV
2565                  * desc.
2566                  */
2567                 rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key,
2568                                   vallen, val, NULL);
2569                 if (!rc && KEY_IS(KEY_CONN_DATA)) {
2570                         exp->exp_connect_flags =
2571                         ((struct obd_connect_data *)val)->ocd_connect_flags;
2572                 }
2573                 RETURN(rc);
2574         }
2575
2576         CDEBUG(D_IOCTL, "Invalid key\n");
2577         RETURN(-EINVAL);
2578 }
2579
2580 int lmv_set_info_async(struct obd_export *exp, obd_count keylen,
2581                        void *key, obd_count vallen, void *val,
2582                        struct ptlrpc_request_set *set)
2583 {
2584         struct lmv_tgt_desc    *tgt;
2585         struct obd_device      *obd;
2586         struct lmv_obd         *lmv;
2587         int rc = 0;
2588         ENTRY;
2589
2590         obd = class_exp2obd(exp);
2591         if (obd == NULL) {
2592                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2593                        exp->exp_handle.h_cookie);
2594                 RETURN(-EINVAL);
2595         }
2596         lmv = &obd->u.lmv;
2597
2598         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2599             KEY_IS(KEY_INIT_RECOV_BACKUP)) {
2600                 int i, err = 0;
2601
2602                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2603                         tgt = &lmv->tgts[i];
2604
2605                         if (!tgt->ltd_exp)
2606                                 continue;
2607
2608                         err = obd_set_info_async(tgt->ltd_exp,
2609                                                  keylen, key, vallen, val, set);
2610                         if (err && rc == 0)
2611                                 rc = err;
2612                 }
2613
2614                 RETURN(rc);
2615         }
2616
2617         RETURN(-EINVAL);
2618 }
2619
2620 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
2621                struct lov_stripe_md *lsm)
2622 {
2623         struct obd_device         *obd = class_exp2obd(exp);
2624         struct lmv_obd            *lmv = &obd->u.lmv;
2625         struct lmv_stripe_md      *meap;
2626         struct lmv_stripe_md      *lsmp;
2627         int                        mea_size;
2628         int                        i;
2629         ENTRY;
2630
2631         mea_size = lmv_get_easize(lmv);
2632         if (!lmmp)
2633                 RETURN(mea_size);
2634
2635         if (*lmmp && !lsm) {
2636                 OBD_FREE(*lmmp, mea_size);
2637                 *lmmp = NULL;
2638                 RETURN(0);
2639         }
2640
2641         if (*lmmp == NULL) {
2642                 OBD_ALLOC(*lmmp, mea_size);
2643                 if (*lmmp == NULL)
2644                         RETURN(-ENOMEM);
2645         }
2646
2647         if (!lsm)
2648                 RETURN(mea_size);
2649
2650         lsmp = (struct lmv_stripe_md *)lsm;
2651         meap = (struct lmv_stripe_md *)*lmmp;
2652
2653         if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR &&
2654             lsmp->mea_magic != MEA_MAGIC_ALL_CHARS)
2655                 RETURN(-EINVAL);
2656
2657         meap->mea_magic = cpu_to_le32(lsmp->mea_magic);
2658         meap->mea_count = cpu_to_le32(lsmp->mea_count);
2659         meap->mea_master = cpu_to_le32(lsmp->mea_master);
2660
2661         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2662                 meap->mea_ids[i] = meap->mea_ids[i];
2663                 fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]);
2664         }
2665
2666         RETURN(mea_size);
2667 }
2668
2669 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
2670                  struct lov_mds_md *lmm, int lmm_size)
2671 {
2672         struct obd_device          *obd = class_exp2obd(exp);
2673         struct lmv_stripe_md      **tmea = (struct lmv_stripe_md **)lsmp;
2674         struct lmv_stripe_md       *mea = (struct lmv_stripe_md *)lmm;
2675         struct lmv_obd             *lmv = &obd->u.lmv;
2676         int                         mea_size;
2677         int                         i;
2678         __u32                       magic;
2679         ENTRY;
2680
2681         mea_size = lmv_get_easize(lmv);
2682         if (lsmp == NULL)
2683                 return mea_size;
2684
2685         if (*lsmp != NULL && lmm == NULL) {
2686                 OBD_FREE(*tmea, mea_size);
2687                 *lsmp = NULL;
2688                 RETURN(0);
2689         }
2690
2691         LASSERT(mea_size == lmm_size);
2692
2693         OBD_ALLOC(*tmea, mea_size);
2694         if (*tmea == NULL)
2695                 RETURN(-ENOMEM);
2696
2697         if (!lmm)
2698                 RETURN(mea_size);
2699
2700         if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
2701             mea->mea_magic == MEA_MAGIC_ALL_CHARS ||
2702             mea->mea_magic == MEA_MAGIC_HASH_SEGMENT)
2703         {
2704                 magic = le32_to_cpu(mea->mea_magic);
2705         } else {
2706                 /*
2707                  * Old mea is not handled here.
2708                  */
2709                 CERROR("Old not supportable EA is found\n");
2710                 LBUG();
2711         }
2712
2713         (*tmea)->mea_magic = magic;
2714         (*tmea)->mea_count = le32_to_cpu(mea->mea_count);
2715         (*tmea)->mea_master = le32_to_cpu(mea->mea_master);
2716
2717         for (i = 0; i < (*tmea)->mea_count; i++) {
2718                 (*tmea)->mea_ids[i] = mea->mea_ids[i];
2719                 fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]);
2720         }
2721         RETURN(mea_size);
2722 }
2723
2724 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
2725                              ldlm_policy_data_t *policy, ldlm_mode_t mode,
2726                              int flags, void *opaque)
2727 {
2728         struct obd_device       *obd = exp->exp_obd;
2729         struct lmv_obd          *lmv = &obd->u.lmv;
2730         int                      rc = 0;
2731         int                      err;
2732         int                      i;
2733         ENTRY;
2734
2735         LASSERT(fid != NULL);
2736
2737         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2738                 if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active)
2739                         continue;
2740
2741                 err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid,
2742                                        policy, mode, flags, opaque);
2743                 if (!rc)
2744                         rc = err;
2745         }
2746         RETURN(rc);
2747 }
2748
2749 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
2750                       __u32 *bits)
2751 {
2752         struct obd_device       *obd = exp->exp_obd;
2753         struct lmv_obd          *lmv = &obd->u.lmv;
2754         int                      rc;
2755         ENTRY;
2756
2757         rc =  md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data, bits);
2758         RETURN(rc);
2759 }
2760
2761 ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags,
2762                            const struct lu_fid *fid, ldlm_type_t type,
2763                            ldlm_policy_data_t *policy, ldlm_mode_t mode,
2764                            struct lustre_handle *lockh)
2765 {
2766         struct obd_device       *obd = exp->exp_obd;
2767         struct lmv_obd          *lmv = &obd->u.lmv;
2768         ldlm_mode_t              rc;
2769         int                      i;
2770         ENTRY;
2771
2772         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
2773
2774         /*
2775          * With CMD every object can have two locks in different namespaces:
2776          * lookup lock in space of mds storing direntry and update/open lock in
2777          * space of mds storing inode. Thus we check all targets, not only that
2778          * one fid was created in.
2779          */
2780         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2781                 rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid,
2782                                    type, policy, mode, lockh);
2783                 if (rc)
2784                         RETURN(rc);
2785         }
2786
2787         RETURN(0);
2788 }
2789
2790 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
2791                       struct obd_export *dt_exp, struct obd_export *md_exp,
2792                       struct lustre_md *md)
2793 {
2794         struct obd_device       *obd = exp->exp_obd;
2795         struct lmv_obd          *lmv = &obd->u.lmv;
2796         int                      rc;
2797         ENTRY;
2798         rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, dt_exp, md_exp, md);
2799         RETURN(rc);
2800 }
2801
2802 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
2803 {
2804         struct obd_device       *obd = exp->exp_obd;
2805         struct lmv_obd          *lmv = &obd->u.lmv;
2806         ENTRY;
2807
2808         if (md->mea)
2809                 obd_free_memmd(exp, (void *)&md->mea);
2810         RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md));
2811 }
2812
2813 int lmv_set_open_replay_data(struct obd_export *exp,
2814                              struct obd_client_handle *och,
2815                              struct ptlrpc_request *open_req)
2816 {
2817         struct obd_device       *obd = exp->exp_obd;
2818         struct lmv_obd          *lmv = &obd->u.lmv;
2819         struct lmv_tgt_desc     *tgt;
2820         ENTRY;
2821
2822         tgt = lmv_find_target(lmv, &och->och_fid);
2823         if (IS_ERR(tgt))
2824                 RETURN(PTR_ERR(tgt));
2825
2826         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, open_req));
2827 }
2828
2829 int lmv_clear_open_replay_data(struct obd_export *exp,
2830                                struct obd_client_handle *och)
2831 {
2832         struct obd_device       *obd = exp->exp_obd;
2833         struct lmv_obd          *lmv = &obd->u.lmv;
2834         struct lmv_tgt_desc     *tgt;
2835         ENTRY;
2836
2837         tgt = lmv_find_target(lmv, &och->och_fid);
2838         if (IS_ERR(tgt))
2839                 RETURN(PTR_ERR(tgt));
2840
2841         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
2842 }
2843
2844 static int lmv_get_remote_perm(struct obd_export *exp,
2845                                const struct lu_fid *fid,
2846                                struct obd_capa *oc, __u32 suppgid,
2847                                struct ptlrpc_request **request)
2848 {
2849         struct obd_device       *obd = exp->exp_obd;
2850         struct lmv_obd          *lmv = &obd->u.lmv;
2851         struct lmv_tgt_desc     *tgt;
2852         int                      rc;
2853         ENTRY;
2854
2855         rc = lmv_check_connect(obd);
2856         if (rc)
2857                 RETURN(rc);
2858
2859         tgt = lmv_find_target(lmv, fid);
2860         if (IS_ERR(tgt))
2861                 RETURN(PTR_ERR(tgt));
2862
2863         rc = md_get_remote_perm(tgt->ltd_exp, fid, oc, suppgid, request);
2864         RETURN(rc);
2865 }
2866
2867 static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc,
2868                           renew_capa_cb_t cb)
2869 {
2870         struct obd_device       *obd = exp->exp_obd;
2871         struct lmv_obd          *lmv = &obd->u.lmv;
2872         struct lmv_tgt_desc     *tgt;
2873         int                      rc;
2874         ENTRY;
2875
2876         rc = lmv_check_connect(obd);
2877         if (rc)
2878                 RETURN(rc);
2879
2880         tgt = lmv_find_target(lmv, &oc->c_capa.lc_fid);
2881         if (IS_ERR(tgt))
2882                 RETURN(PTR_ERR(tgt));
2883
2884         rc = md_renew_capa(tgt->ltd_exp, oc, cb);
2885         RETURN(rc);
2886 }
2887
2888 int lmv_unpack_capa(struct obd_export *exp, struct ptlrpc_request *req,
2889                     const struct req_msg_field *field, struct obd_capa **oc)
2890 {
2891         struct obd_device *obd = exp->exp_obd;
2892         struct lmv_obd *lmv = &obd->u.lmv;
2893         int rc;
2894
2895         ENTRY;
2896         rc = md_unpack_capa(lmv->tgts[0].ltd_exp, req, field, oc);
2897         RETURN(rc);
2898 }
2899
2900 int lmv_intent_getattr_async(struct obd_export *exp,
2901                              struct md_enqueue_info *minfo,
2902                              struct ldlm_enqueue_info *einfo)
2903 {
2904         struct md_op_data       *op_data = &minfo->mi_data;
2905         struct obd_device       *obd = exp->exp_obd;
2906         struct lmv_obd          *lmv = &obd->u.lmv;
2907         struct lmv_object       *obj;
2908         struct lmv_tgt_desc     *tgt;
2909         int                      rc;
2910         int                      sidx;
2911         ENTRY;
2912
2913         rc = lmv_check_connect(obd);
2914         if (rc)
2915                 RETURN(rc);
2916
2917         if (!fid_is_sane(&op_data->op_fid2)) {
2918                 obj = lmv_object_find(obd, &op_data->op_fid1);
2919                 if (obj && op_data->op_namelen) {
2920                         sidx = raw_name2idx(obj->lo_hashtype,
2921                                             obj->lo_objcount,
2922                                             (char *)op_data->op_name,
2923                                             op_data->op_namelen);
2924                         op_data->op_fid1 = obj->lo_stripes[sidx].ls_fid;
2925                         tgt = lmv_get_target(lmv,
2926                                              obj->lo_stripes[sidx].ls_mds);
2927                         CDEBUG(D_INODE,
2928                                "Choose slave dir ("DFID") -> mds #%d\n",
2929                                PFID(&op_data->op_fid1), tgt->ltd_idx);
2930                 } else {
2931                         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2932                 }
2933                 if (obj)
2934                         lmv_object_put(obj);
2935         } else {
2936                 op_data->op_fid1 = op_data->op_fid2;
2937                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2938                 op_data->op_bias = MDS_CROSS_REF;
2939                 /*
2940                  * Unfortunately, we have to lie to MDC/MDS to retrieve
2941                  * attributes llite needs.
2942                 */
2943                 if (minfo->mi_it.it_op & IT_LOOKUP)
2944                         minfo->mi_it.it_op = IT_GETATTR;
2945         }
2946
2947         if (IS_ERR(tgt))
2948                 RETURN(PTR_ERR(tgt));
2949
2950         rc = md_intent_getattr_async(tgt->ltd_exp, minfo, einfo);
2951         RETURN(rc);
2952 }
2953
2954 int lmv_revalidate_lock(struct obd_export *exp,
2955                         struct lookup_intent *it,
2956                         struct lu_fid *fid)
2957 {
2958         struct obd_device       *obd = exp->exp_obd;
2959         struct lmv_obd          *lmv = &obd->u.lmv;
2960         struct lmv_tgt_desc     *tgt;
2961         int                      rc;
2962         ENTRY;
2963
2964         rc = lmv_check_connect(obd);
2965         if (rc)
2966                 RETURN(rc);
2967
2968         tgt = lmv_find_target(lmv, fid);
2969         if (IS_ERR(tgt))
2970                 RETURN(PTR_ERR(tgt));
2971
2972         rc = md_revalidate_lock(tgt->ltd_exp, it, fid);
2973         RETURN(rc);
2974 }
2975
2976
2977 struct obd_ops lmv_obd_ops = {
2978         .o_owner                = THIS_MODULE,
2979         .o_setup                = lmv_setup,
2980         .o_cleanup              = lmv_cleanup,
2981         .o_precleanup           = lmv_precleanup,
2982         .o_process_config       = lmv_process_config,
2983         .o_connect              = lmv_connect,
2984         .o_disconnect           = lmv_disconnect,
2985         .o_statfs               = lmv_statfs,
2986         .o_get_info             = lmv_get_info,
2987         .o_set_info_async       = lmv_set_info_async,
2988         .o_packmd               = lmv_packmd,
2989         .o_unpackmd             = lmv_unpackmd,
2990         .o_notify               = lmv_notify,
2991         .o_get_uuid             = lmv_get_uuid,
2992         .o_iocontrol            = lmv_iocontrol,
2993         .o_fid_delete           = lmv_fid_delete
2994 };
2995
2996 struct md_ops lmv_md_ops = {
2997         .m_getstatus            = lmv_getstatus,
2998         .m_change_cbdata        = lmv_change_cbdata,
2999         .m_close                = lmv_close,
3000         .m_create               = lmv_create,
3001         .m_done_writing         = lmv_done_writing,
3002         .m_enqueue              = lmv_enqueue,
3003         .m_getattr              = lmv_getattr,
3004         .m_getxattr             = lmv_getxattr,
3005         .m_getattr_name         = lmv_getattr_name,
3006         .m_intent_lock          = lmv_intent_lock,
3007         .m_link                 = lmv_link,
3008         .m_rename               = lmv_rename,
3009         .m_setattr              = lmv_setattr,
3010         .m_setxattr             = lmv_setxattr,
3011         .m_sync                 = lmv_sync,
3012         .m_readpage             = lmv_readpage,
3013         .m_unlink               = lmv_unlink,
3014         .m_init_ea_size         = lmv_init_ea_size,
3015         .m_cancel_unused        = lmv_cancel_unused,
3016         .m_set_lock_data        = lmv_set_lock_data,
3017         .m_lock_match           = lmv_lock_match,
3018         .m_get_lustre_md        = lmv_get_lustre_md,
3019         .m_free_lustre_md       = lmv_free_lustre_md,
3020         .m_set_open_replay_data = lmv_set_open_replay_data,
3021         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3022         .m_renew_capa           = lmv_renew_capa,
3023         .m_unpack_capa          = lmv_unpack_capa,
3024         .m_get_remote_perm      = lmv_get_remote_perm,
3025         .m_intent_getattr_async = lmv_intent_getattr_async,
3026         .m_revalidate_lock      = lmv_revalidate_lock
3027 };
3028
3029 static quota_interface_t *quota_interface;
3030 extern quota_interface_t lmv_quota_interface;
3031
3032 int __init lmv_init(void)
3033 {
3034         struct lprocfs_static_vars lvars;
3035         int                        rc;
3036
3037         lmv_object_cache = cfs_mem_cache_create("lmv_objects",
3038                                                 sizeof(struct lmv_object),
3039                                                 0, 0);
3040         if (!lmv_object_cache) {
3041                 CERROR("Error allocating lmv objects cache\n");
3042                 return -ENOMEM;
3043         }
3044
3045         lprocfs_lmv_init_vars(&lvars);
3046
3047         cfs_request_module("lquota");
3048         quota_interface = PORTAL_SYMBOL_GET(lmv_quota_interface);
3049         init_obd_quota_ops(quota_interface, &lmv_obd_ops);
3050
3051         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
3052                                  lvars.module_vars, LUSTRE_LMV_NAME, NULL);
3053         if (rc) {
3054                 if (quota_interface)
3055                         PORTAL_SYMBOL_PUT(lmv_quota_interface);
3056                 cfs_mem_cache_destroy(lmv_object_cache);
3057         }
3058
3059         return rc;
3060 }
3061
3062 #ifdef __KERNEL__
3063 static void lmv_exit(void)
3064 {
3065         if (quota_interface)
3066                 PORTAL_SYMBOL_PUT(lmv_quota_interface);
3067
3068         class_unregister_type(LUSTRE_LMV_NAME);
3069
3070         LASSERTF(cfs_atomic_read(&lmv_object_count) == 0,
3071                  "Can't free lmv objects cache, %d object(s) busy\n",
3072                  cfs_atomic_read(&lmv_object_count));
3073         cfs_mem_cache_destroy(lmv_object_cache);
3074 }
3075
3076 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
3077 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
3078 MODULE_LICENSE("GPL");
3079
3080 module_init(lmv_init);
3081 module_exit(lmv_exit);
3082 #endif