lustre/lmv/lmv_obd.c

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