lustre/lmv/lmv_obd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  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 #define DEBUG_SUBSYSTEM S_LMV
  38 #include <linux/slab.h>
  39 #include <linux/module.h>
  40 #include <linux/init.h>
  41 #include <linux/user_namespace.h>
  42 #ifdef HAVE_UIDGID_HEADER
  43 # include <linux/uidgid.h>
  44 #endif
  45 #include <linux/slab.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/mm.h>
  48 #include <linux/math64.h>
  49 #include <linux/seq_file.h>
  50 #include <linux/namei.h>
  51
  52 #include <lustre/lustre_idl.h>
  53 #include <obd_support.h>
  54 #include <lustre_lib.h>
  55 #include <lustre_net.h>
  56 #include <obd_class.h>
  57 #include <lustre_lmv.h>
  58 #include <lprocfs_status.h>
  59 #include <cl_object.h>
  60 #include <lustre_fid.h>
  61 #include <lustre_ioctl.h>
  62 #include <lustre_kernelcomm.h>
  63 #include "lmv_internal.h"
  64
  65 static void lmv_activate_target(struct lmv_obd *lmv,
  66                                 struct lmv_tgt_desc *tgt,
  67                                 int activate)
  68 {
  69         if (tgt->ltd_active == activate)
  70                 return;
  71
  72         tgt->ltd_active = activate;
  73         lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
  74
  75         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
  76 }
  77
  78 /**
  79  * Error codes:
  80  *
  81  *  -EINVAL  : UUID can't be found in the LMV's target list
  82  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  83  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  84  */
  85 static int lmv_set_mdc_active(struct lmv_obd *lmv,
  86                               const struct obd_uuid *uuid,
  87                               int activate)
  88 {
  89         struct lmv_tgt_desc     *tgt = NULL;
  90         struct obd_device       *obd;
  91         __u32                    i;
  92         int                      rc = 0;
  93         ENTRY;
  94
  95         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  96                         lmv, uuid->uuid, activate);
  97
  98         spin_lock(&lmv->lmv_lock);
  99         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 100                 tgt = lmv->tgts[i];
 101                 if (tgt == NULL || tgt->ltd_exp == NULL)
 102                         continue;
 103
 104                 CDEBUG(D_INFO, "Target idx %d is %s conn "LPX64"\n", i,
 105                        tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
 106
 107                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 108                         break;
 109         }
 110
 111         if (i == lmv->desc.ld_tgt_count)
 112                 GOTO(out_lmv_lock, rc = -EINVAL);
 113
 114         obd = class_exp2obd(tgt->ltd_exp);
 115         if (obd == NULL)
 116                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 117
 118         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 119                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 120                obd->obd_type->typ_name, i);
 121         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 122
 123         if (tgt->ltd_active == activate) {
 124                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 125                        activate ? "" : "in");
 126                 GOTO(out_lmv_lock, rc);
 127         }
 128
 129         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 130                activate ? "" : "in");
 131         lmv_activate_target(lmv, tgt, activate);
 132         EXIT;
 133
 134  out_lmv_lock:
 135         spin_unlock(&lmv->lmv_lock);
 136         return rc;
 137 }
 138
 139 struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
 140 {
 141         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
 142         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
 143
 144         return (tgt == NULL) ? NULL : obd_get_uuid(tgt->ltd_exp);
 145 }
 146
 147 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 148                       enum obd_notify_event ev, void *data)
 149 {
 150         struct obd_connect_data *conn_data;
 151         struct lmv_obd          *lmv = &obd->u.lmv;
 152         struct obd_uuid         *uuid;
 153         int                      rc = 0;
 154         ENTRY;
 155
 156         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 157                 CERROR("unexpected notification of %s %s!\n",
 158                        watched->obd_type->typ_name,
 159                        watched->obd_name);
 160                 RETURN(-EINVAL);
 161         }
 162
 163         uuid = &watched->u.cli.cl_target_uuid;
 164         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 165                 /*
 166                  * Set MDC as active before notifying the observer, so the
 167                  * observer can use the MDC normally.
 168                  */
 169                 rc = lmv_set_mdc_active(lmv, uuid,
 170                                         ev == OBD_NOTIFY_ACTIVE);
 171                 if (rc) {
 172                         CERROR("%sactivation of %s failed: %d\n",
 173                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 174                                uuid->uuid, rc);
 175                         RETURN(rc);
 176                 }
 177         } else if (ev == OBD_NOTIFY_OCD) {
 178                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 179                 /*
 180                  * XXX: Make sure that ocd_connect_flags from all targets are
 181                  * the same. Otherwise one of MDTs runs wrong version or
 182                  * something like this.  --umka
 183                  */
 184                 obd->obd_self_export->exp_connect_data = *conn_data;
 185         }
 186 #if 0
 187         else if (ev == OBD_NOTIFY_DISCON) {
 188                 /*
 189                  * For disconnect event, flush fld cache for failout MDS case.
 190                  */
 191                 fld_client_flush(&lmv->lmv_fld);
 192         }
 193 #endif
 194         /*
 195          * Pass the notification up the chain.
 196          */
 197         if (obd->obd_observer)
 198                 rc = obd_notify(obd->obd_observer, watched, ev, data);
 199
 200         RETURN(rc);
 201 }
 202
 203 /**
 204  * This is fake connect function. Its purpose is to initialize lmv and say
 205  * caller that everything is okay. Real connection will be performed later.
 206  */
 207 static int lmv_connect(const struct lu_env *env,
 208                        struct obd_export **exp, struct obd_device *obd,
 209                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 210                        void *localdata)
 211 {
 212         struct lmv_obd        *lmv = &obd->u.lmv;
 213         struct lustre_handle  conn = { 0 };
 214         int                    rc = 0;
 215         ENTRY;
 216
 217         /*
 218          * We don't want to actually do the underlying connections more than
 219          * once, so keep track.
 220          */
 221         lmv->refcount++;
 222         if (lmv->refcount > 1) {
 223                 *exp = NULL;
 224                 RETURN(0);
 225         }
 226
 227         rc = class_connect(&conn, obd, cluuid);
 228         if (rc) {
 229                 CERROR("class_connection() returned %d\n", rc);
 230                 RETURN(rc);
 231         }
 232
 233         *exp = class_conn2export(&conn);
 234         class_export_get(*exp);
 235
 236         lmv->exp = *exp;
 237         lmv->connected = 0;
 238         lmv->cluuid = *cluuid;
 239
 240         if (data)
 241                 lmv->conn_data = *data;
 242
 243         if (lmv->targets_proc_entry == NULL) {
 244                 lmv->targets_proc_entry = lprocfs_register("target_obds",
 245                                                            obd->obd_proc_entry,
 246                                                            NULL, NULL);
 247                 if (IS_ERR(lmv->targets_proc_entry)) {
 248                         CERROR("%s: cannot register "
 249                                "/proc/fs/lustre/%s/%s/target_obds\n",
 250                                obd->obd_name, obd->obd_type->typ_name,
 251                                obd->obd_name);
 252                         lmv->targets_proc_entry = NULL;
 253                 }
 254         }
 255
 256         /*
 257          * All real clients should perform actual connection right away, because
 258          * it is possible, that LMV will not have opportunity to connect targets
 259          * and MDC stuff will be called directly, for instance while reading
 260          * ../mdc/../kbytesfree procfs file, etc.
 261          */
 262         if (data != NULL && (data->ocd_connect_flags & OBD_CONNECT_REAL))
 263                 rc = lmv_check_connect(obd);
 264
 265         if (rc && lmv->targets_proc_entry != NULL)
 266                 lprocfs_remove(&lmv->targets_proc_entry);
 267         RETURN(rc);
 268 }
 269
 270 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
 271                             __u32 def_easize)
 272 {
 273         struct obd_device       *obd = exp->exp_obd;
 274         struct lmv_obd          *lmv = &obd->u.lmv;
 275         __u32                    i;
 276         int                      rc = 0;
 277         int                      change = 0;
 278         ENTRY;
 279
 280         if (lmv->max_easize < easize) {
 281                 lmv->max_easize = easize;
 282                 change = 1;
 283         }
 284         if (lmv->max_def_easize < def_easize) {
 285                 lmv->max_def_easize = def_easize;
 286                 change = 1;
 287         }
 288
 289         if (change == 0)
 290                 RETURN(0);
 291
 292         if (lmv->connected == 0)
 293                 RETURN(0);
 294
 295         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 296                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 297
 298                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active) {
 299                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
 300                         continue;
 301                 }
 302
 303                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
 304                 if (rc) {
 305                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
 306                                " rc = %d\n", obd->obd_name, i, rc);
 307                         break;
 308                 }
 309         }
 310         RETURN(rc);
 311 }
 312
 313 #define MAX_STRING_SIZE 128
 314
 315 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 316 {
 317         struct lmv_obd          *lmv = &obd->u.lmv;
 318         struct obd_uuid         *cluuid = &lmv->cluuid;
 319         struct obd_uuid          lmv_mdc_uuid = { "LMV_MDC_UUID" };
 320         struct obd_device       *mdc_obd;
 321         struct obd_export       *mdc_exp;
 322         struct lu_fld_target     target;
 323         int                      rc;
 324         ENTRY;
 325
 326         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 327                                         &obd->obd_uuid);
 328         if (!mdc_obd) {
 329                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 330                 RETURN(-EINVAL);
 331         }
 332
 333         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n",
 334                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 335                 tgt->ltd_uuid.uuid, obd->obd_uuid.uuid,
 336                 cluuid->uuid);
 337
 338         if (!mdc_obd->obd_set_up) {
 339                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 340                 RETURN(-EINVAL);
 341         }
 342
 343         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &lmv_mdc_uuid,
 344                          &lmv->conn_data, NULL);
 345         if (rc) {
 346                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 347                 RETURN(rc);
 348         }
 349
 350         /*
 351          * Init fid sequence client for this mdc and add new fld target.
 352          */
 353         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
 354         if (rc)
 355                 RETURN(rc);
 356
 357         target.ft_srv = NULL;
 358         target.ft_exp = mdc_exp;
 359         target.ft_idx = tgt->ltd_idx;
 360
 361         fld_client_add_target(&lmv->lmv_fld, &target);
 362
 363         rc = obd_register_observer(mdc_obd, obd);
 364         if (rc) {
 365                 obd_disconnect(mdc_exp);
 366                 CERROR("target %s register_observer error %d\n",
 367                        tgt->ltd_uuid.uuid, rc);
 368                 RETURN(rc);
 369         }
 370
 371         if (obd->obd_observer) {
 372                 /*
 373                  * Tell the observer about the new target.
 374                  */
 375                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 376                                 OBD_NOTIFY_ACTIVE,
 377                                 (void *)(tgt - lmv->tgts[0]));
 378                 if (rc) {
 379                         obd_disconnect(mdc_exp);
 380                         RETURN(rc);
 381                 }
 382         }
 383
 384         tgt->ltd_active = 1;
 385         tgt->ltd_exp = mdc_exp;
 386         lmv->desc.ld_active_tgt_count++;
 387
 388         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
 389
 390         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 391                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 392                 atomic_read(&obd->obd_refcount));
 393
 394         if (lmv->targets_proc_entry != NULL) {
 395                 struct proc_dir_entry *mdc_symlink;
 396
 397                 LASSERT(mdc_obd->obd_type != NULL);
 398                 LASSERT(mdc_obd->obd_type->typ_name != NULL);
 399                 mdc_symlink = lprocfs_add_symlink(mdc_obd->obd_name,
 400                                                   lmv->targets_proc_entry,
 401                                                   "../../../%s/%s",
 402                                                   mdc_obd->obd_type->typ_name,
 403                                                   mdc_obd->obd_name);
 404                 if (mdc_symlink == NULL) {
 405                         CERROR("cannot register LMV target "
 406                                "/proc/fs/lustre/%s/%s/target_obds/%s\n",
 407                                obd->obd_type->typ_name, obd->obd_name,
 408                                mdc_obd->obd_name);
 409                 }
 410         }
 411         RETURN(0);
 412 }
 413
 414 static void lmv_del_target(struct lmv_obd *lmv, int index)
 415 {
 416         if (lmv->tgts[index] == NULL)
 417                 return;
 418
 419         OBD_FREE_PTR(lmv->tgts[index]);
 420         lmv->tgts[index] = NULL;
 421         return;
 422 }
 423
 424 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 425                            __u32 index, int gen)
 426 {
 427         struct obd_device *mdc_obd;
 428         struct lmv_obd      *lmv = &obd->u.lmv;
 429         struct lmv_tgt_desc *tgt;
 430         int                  orig_tgt_count = 0;
 431         int                  rc = 0;
 432         ENTRY;
 433
 434         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
 435         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
 436                                         &obd->obd_uuid);
 437         if (!mdc_obd) {
 438                 CERROR("%s: Target %s not attached: rc = %d\n",
 439                        obd->obd_name, uuidp->uuid, -EINVAL);
 440                 RETURN(-EINVAL);
 441         }
 442
 443         mutex_lock(&lmv->lmv_init_mutex);
 444         if ((index < lmv->tgts_size) && (lmv->tgts[index] != NULL)) {
 445                 tgt = lmv->tgts[index];
 446                 CERROR("%s: UUID %s already assigned at LOV target index %d:"
 447                        " rc = %d\n", obd->obd_name,
 448                        obd_uuid2str(&tgt->ltd_uuid), index, -EEXIST);
 449                 mutex_unlock(&lmv->lmv_init_mutex);
 450                 RETURN(-EEXIST);
 451         }
 452
 453         if (index >= lmv->tgts_size) {
 454                 /* We need to reallocate the lmv target array. */
 455                 struct lmv_tgt_desc **newtgts, **old = NULL;
 456                 __u32 newsize = 1;
 457                 __u32 oldsize = 0;
 458
 459                 while (newsize < index + 1)
 460                         newsize = newsize << 1;
 461                 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
 462                 if (newtgts == NULL) {
 463                         mutex_unlock(&lmv->lmv_init_mutex);
 464                         RETURN(-ENOMEM);
 465                 }
 466
 467                 if (lmv->tgts_size) {
 468                         memcpy(newtgts, lmv->tgts,
 469                                sizeof(*newtgts) * lmv->tgts_size);
 470                         old = lmv->tgts;
 471                         oldsize = lmv->tgts_size;
 472                 }
 473
 474                 lmv->tgts = newtgts;
 475                 lmv->tgts_size = newsize;
 476                 smp_rmb();
 477                 if (old)
 478                         OBD_FREE(old, sizeof(*old) * oldsize);
 479
 480                 CDEBUG(D_CONFIG, "tgts: %p size: %d\n", lmv->tgts,
 481                        lmv->tgts_size);
 482         }
 483
 484         OBD_ALLOC_PTR(tgt);
 485         if (!tgt) {
 486                 mutex_unlock(&lmv->lmv_init_mutex);
 487                 RETURN(-ENOMEM);
 488         }
 489
 490         mutex_init(&tgt->ltd_fid_mutex);
 491         tgt->ltd_idx = index;
 492         tgt->ltd_uuid = *uuidp;
 493         tgt->ltd_active = 0;
 494         lmv->tgts[index] = tgt;
 495         if (index >= lmv->desc.ld_tgt_count) {
 496                 orig_tgt_count = lmv->desc.ld_tgt_count;
 497                 lmv->desc.ld_tgt_count = index + 1;
 498         }
 499
 500         if (lmv->connected == 0) {
 501                 /* lmv_check_connect() will connect this target. */
 502                 mutex_unlock(&lmv->lmv_init_mutex);
 503                 RETURN(0);
 504         }
 505
 506         /* Otherwise let's connect it ourselves */
 507         mutex_unlock(&lmv->lmv_init_mutex);
 508         rc = lmv_connect_mdc(obd, tgt);
 509         if (rc != 0) {
 510                 spin_lock(&lmv->lmv_lock);
 511                 if (lmv->desc.ld_tgt_count == index + 1)
 512                         lmv->desc.ld_tgt_count = orig_tgt_count;
 513                 memset(tgt, 0, sizeof(*tgt));
 514                 spin_unlock(&lmv->lmv_lock);
 515         } else {
 516                 int easize = sizeof(struct lmv_stripe_md) +
 517                         lmv->desc.ld_tgt_count * sizeof(struct lu_fid);
 518                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
 519         }
 520
 521         RETURN(rc);
 522 }
 523
 524 int lmv_check_connect(struct obd_device *obd)
 525 {
 526         struct lmv_obd          *lmv = &obd->u.lmv;
 527         struct lmv_tgt_desc     *tgt;
 528         __u32                    i;
 529         int                      rc;
 530         int                      easize;
 531         ENTRY;
 532
 533         if (lmv->connected)
 534                 RETURN(0);
 535
 536         mutex_lock(&lmv->lmv_init_mutex);
 537         if (lmv->connected) {
 538                 mutex_unlock(&lmv->lmv_init_mutex);
 539                 RETURN(0);
 540         }
 541
 542         if (lmv->desc.ld_tgt_count == 0) {
 543                 mutex_unlock(&lmv->lmv_init_mutex);
 544                 CERROR("%s: no targets configured.\n", obd->obd_name);
 545                 RETURN(-EINVAL);
 546         }
 547
 548         LASSERT(lmv->tgts != NULL);
 549
 550         if (lmv->tgts[0] == NULL) {
 551                 mutex_unlock(&lmv->lmv_init_mutex);
 552                 CERROR("%s: no target configured for index 0.\n",
 553                        obd->obd_name);
 554                 RETURN(-EINVAL);
 555         }
 556
 557         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 558                lmv->cluuid.uuid, obd->obd_name);
 559
 560         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 561                 tgt = lmv->tgts[i];
 562                 if (tgt == NULL)
 563                         continue;
 564                 rc = lmv_connect_mdc(obd, tgt);
 565                 if (rc)
 566                         GOTO(out_disc, rc);
 567         }
 568
 569         class_export_put(lmv->exp);
 570         lmv->connected = 1;
 571         easize = lmv_mds_md_size(lmv->desc.ld_tgt_count, LMV_MAGIC);
 572         lmv_init_ea_size(obd->obd_self_export, easize, 0);
 573         mutex_unlock(&lmv->lmv_init_mutex);
 574         RETURN(0);
 575
 576  out_disc:
 577         while (i-- > 0) {
 578                 int rc2;
 579                 tgt = lmv->tgts[i];
 580                 if (tgt == NULL)
 581                         continue;
 582                 tgt->ltd_active = 0;
 583                 if (tgt->ltd_exp) {
 584                         --lmv->desc.ld_active_tgt_count;
 585                         rc2 = obd_disconnect(tgt->ltd_exp);
 586                         if (rc2) {
 587                                 CERROR("LMV target %s disconnect on "
 588                                        "MDC idx %d: error %d\n",
 589                                        tgt->ltd_uuid.uuid, i, rc2);
 590                         }
 591                 }
 592         }
 593         class_disconnect(lmv->exp);
 594         mutex_unlock(&lmv->lmv_init_mutex);
 595         RETURN(rc);
 596 }
 597
 598 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 599 {
 600         struct lmv_obd         *lmv = &obd->u.lmv;
 601         struct obd_device      *mdc_obd;
 602         int                     rc;
 603         ENTRY;
 604
 605         LASSERT(tgt != NULL);
 606         LASSERT(obd != NULL);
 607
 608         mdc_obd = class_exp2obd(tgt->ltd_exp);
 609
 610         if (mdc_obd) {
 611                 mdc_obd->obd_force = obd->obd_force;
 612                 mdc_obd->obd_fail = obd->obd_fail;
 613                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 614
 615                 if (lmv->targets_proc_entry != NULL)
 616                         lprocfs_remove_proc_entry(mdc_obd->obd_name,
 617                                                   lmv->targets_proc_entry);
 618         }
 619
 620         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
 621         if (rc)
 622                 CERROR("Can't finanize fids factory\n");
 623
 624         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 625                tgt->ltd_exp->exp_obd->obd_name,
 626                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 627
 628         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 629         rc = obd_disconnect(tgt->ltd_exp);
 630         if (rc) {
 631                 if (tgt->ltd_active) {
 632                         CERROR("Target %s disconnect error %d\n",
 633                                tgt->ltd_uuid.uuid, rc);
 634                 }
 635         }
 636
 637         lmv_activate_target(lmv, tgt, 0);
 638         tgt->ltd_exp = NULL;
 639         RETURN(0);
 640 }
 641
 642 static int lmv_disconnect(struct obd_export *exp)
 643 {
 644         struct obd_device       *obd = class_exp2obd(exp);
 645         struct lmv_obd          *lmv = &obd->u.lmv;
 646         int                      rc;
 647         __u32                    i;
 648         ENTRY;
 649
 650         if (!lmv->tgts)
 651                 goto out_local;
 652
 653         /*
 654          * Only disconnect the underlying layers on the final disconnect.
 655          */
 656         lmv->refcount--;
 657         if (lmv->refcount != 0)
 658                 goto out_local;
 659
 660         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 661                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
 662                         continue;
 663
 664                 lmv_disconnect_mdc(obd, lmv->tgts[i]);
 665         }
 666
 667         if (lmv->targets_proc_entry != NULL)
 668                 lprocfs_remove(&lmv->targets_proc_entry);
 669         else
 670                 CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n",
 671                        obd->obd_type->typ_name, obd->obd_name);
 672
 673 out_local:
 674         /*
 675          * This is the case when no real connection is established by
 676          * lmv_check_connect().
 677          */
 678         if (!lmv->connected)
 679                 class_export_put(exp);
 680         rc = class_disconnect(exp);
 681         if (lmv->refcount == 0)
 682                 lmv->connected = 0;
 683         RETURN(rc);
 684 }
 685
 686 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
 687                         void __user *uarg)
 688 {
 689         struct obd_device       *obddev = class_exp2obd(exp);
 690         struct lmv_obd          *lmv = &obddev->u.lmv;
 691         struct getinfo_fid2path *gf;
 692         struct lmv_tgt_desc     *tgt;
 693         struct getinfo_fid2path *remote_gf = NULL;
 694         int                     remote_gf_size = 0;
 695         int                     rc;
 696
 697         gf = (struct getinfo_fid2path *)karg;
 698         tgt = lmv_find_target(lmv, &gf->gf_fid);
 699         if (IS_ERR(tgt))
 700                 RETURN(PTR_ERR(tgt));
 701
 702 repeat_fid2path:
 703         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
 704         if (rc != 0 && rc != -EREMOTE)
 705                 GOTO(out_fid2path, rc);
 706
 707         /* If remote_gf != NULL, it means just building the
 708          * path on the remote MDT, copy this path segement to gf */
 709         if (remote_gf != NULL) {
 710                 struct getinfo_fid2path *ori_gf;
 711                 char *ptr;
 712
 713                 ori_gf = (struct getinfo_fid2path *)karg;
 714                 if (strlen(ori_gf->gf_path) +
 715                     strlen(gf->gf_path) > ori_gf->gf_pathlen)
 716                         GOTO(out_fid2path, rc = -EOVERFLOW);
 717
 718                 ptr = ori_gf->gf_path;
 719
 720                 memmove(ptr + strlen(gf->gf_path) + 1, ptr,
 721                         strlen(ori_gf->gf_path));
 722
 723                 strncpy(ptr, gf->gf_path, strlen(gf->gf_path));
 724                 ptr += strlen(gf->gf_path);
 725                 *ptr = '/';
 726         }
 727
 728         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: "LPU64" ln: %u\n",
 729                tgt->ltd_exp->exp_obd->obd_name,
 730                gf->gf_path, PFID(&gf->gf_fid), gf->gf_recno,
 731                gf->gf_linkno);
 732
 733         if (rc == 0)
 734                 GOTO(out_fid2path, rc);
 735
 736         /* sigh, has to go to another MDT to do path building further */
 737         if (remote_gf == NULL) {
 738                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
 739                 OBD_ALLOC(remote_gf, remote_gf_size);
 740                 if (remote_gf == NULL)
 741                         GOTO(out_fid2path, rc = -ENOMEM);
 742                 remote_gf->gf_pathlen = PATH_MAX;
 743         }
 744
 745         if (!fid_is_sane(&gf->gf_fid)) {
 746                 CERROR("%s: invalid FID "DFID": rc = %d\n",
 747                        tgt->ltd_exp->exp_obd->obd_name,
 748                        PFID(&gf->gf_fid), -EINVAL);
 749                 GOTO(out_fid2path, rc = -EINVAL);
 750         }
 751
 752         tgt = lmv_find_target(lmv, &gf->gf_fid);
 753         if (IS_ERR(tgt))
 754                 GOTO(out_fid2path, rc = -EINVAL);
 755
 756         remote_gf->gf_fid = gf->gf_fid;
 757         remote_gf->gf_recno = -1;
 758         remote_gf->gf_linkno = -1;
 759         memset(remote_gf->gf_path, 0, remote_gf->gf_pathlen);
 760         gf = remote_gf;
 761         goto repeat_fid2path;
 762
 763 out_fid2path:
 764         if (remote_gf != NULL)
 765                 OBD_FREE(remote_gf, remote_gf_size);
 766         RETURN(rc);
 767 }
 768
 769 static int lmv_hsm_req_count(struct lmv_obd *lmv,
 770                              const struct hsm_user_request *hur,
 771                              const struct lmv_tgt_desc *tgt_mds)
 772 {
 773         __u32                    i;
 774         int                      nr = 0;
 775         struct lmv_tgt_desc     *curr_tgt;
 776
 777         /* count how many requests must be sent to the given target */
 778         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
 779                 curr_tgt = lmv_find_target(lmv, &hur->hur_user_item[i].hui_fid);
 780                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
 781                         nr++;
 782         }
 783         return nr;
 784 }
 785
 786 static void lmv_hsm_req_build(struct lmv_obd *lmv,
 787                               struct hsm_user_request *hur_in,
 788                               const struct lmv_tgt_desc *tgt_mds,
 789                               struct hsm_user_request *hur_out)
 790 {
 791         __u32                    i, nr_out;
 792         struct lmv_tgt_desc     *curr_tgt;
 793
 794         /* build the hsm_user_request for the given target */
 795         hur_out->hur_request = hur_in->hur_request;
 796         nr_out = 0;
 797         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
 798                 curr_tgt = lmv_find_target(lmv,
 799                                            &hur_in->hur_user_item[i].hui_fid);
 800                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
 801                         hur_out->hur_user_item[nr_out] =
 802                                                 hur_in->hur_user_item[i];
 803                         nr_out++;
 804                 }
 805         }
 806         hur_out->hur_request.hr_itemcount = nr_out;
 807         memcpy(hur_data(hur_out), hur_data(hur_in),
 808                hur_in->hur_request.hr_data_len);
 809 }
 810
 811 static int lmv_hsm_ct_unregister(struct lmv_obd *lmv, unsigned int cmd, int len,
 812                                  struct lustre_kernelcomm *lk, void *uarg)
 813 {
 814         __u32   i;
 815         int     rc;
 816         ENTRY;
 817
 818         /* unregister request (call from llapi_hsm_copytool_fini) */
 819         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 820                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 821
 822                 if (tgt == NULL || tgt->ltd_exp == NULL)
 823                         continue;
 824                 /* best effort: try to clean as much as possible
 825                  * (continue on error) */
 826                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 827         }
 828
 829         /* Whatever the result, remove copytool from kuc groups.
 830          * Unreached coordinators will get EPIPE on next requests
 831          * and will unregister automatically.
 832          */
 833         rc = libcfs_kkuc_group_rem(lk->lk_uid, lk->lk_group);
 834
 835         RETURN(rc);
 836 }
 837
 838 static int lmv_hsm_ct_register(struct lmv_obd *lmv, unsigned int cmd, int len,
 839                                struct lustre_kernelcomm *lk, void *uarg)
 840 {
 841         struct file             *filp;
 842         __u32                    i, j;
 843         int                      err, rc;
 844         bool                     any_set = false;
 845         struct kkuc_ct_data      kcd = { 0 };
 846         ENTRY;
 847
 848         /* All or nothing: try to register to all MDS.
 849          * In case of failure, unregister from previous MDS,
 850          * except if it because of inactive target. */
 851         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 852                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 853
 854                 if (tgt == NULL || tgt->ltd_exp == NULL)
 855                         continue;
 856                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 857                 if (err) {
 858                         if (tgt->ltd_active) {
 859                                 /* permanent error */
 860                                 CERROR("%s: iocontrol MDC %s on MDT"
 861                                        " idx %d cmd %x: err = %d\n",
 862                                        class_exp2obd(lmv->exp)->obd_name,
 863                                        tgt->ltd_uuid.uuid, i, cmd, err);
 864                                 rc = err;
 865                                 lk->lk_flags |= LK_FLG_STOP;
 866                                 /* unregister from previous MDS */
 867                                 for (j = 0; j < i; j++) {
 868                                         tgt = lmv->tgts[j];
 869                                         if (tgt == NULL || tgt->ltd_exp == NULL)
 870                                                 continue;
 871                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
 872                                                       lk, uarg);
 873                                 }
 874                                 RETURN(rc);
 875                         }
 876                         /* else: transient error.
 877                          * kuc will register to the missing MDT
 878                          * when it is back */
 879                 } else {
 880                         any_set = true;
 881                 }
 882         }
 883
 884         if (!any_set)
 885                 /* no registration done: return error */
 886                 RETURN(-ENOTCONN);
 887
 888         /* at least one registration done, with no failure */
 889         filp = fget(lk->lk_wfd);
 890         if (filp == NULL)
 891                 RETURN(-EBADF);
 892
 893         kcd.kcd_magic = KKUC_CT_DATA_MAGIC;
 894         kcd.kcd_uuid = lmv->cluuid;
 895         kcd.kcd_archive = lk->lk_data;
 896
 897         rc = libcfs_kkuc_group_add(filp, lk->lk_uid, lk->lk_group,
 898                                    &kcd, sizeof(kcd));
 899         if (rc != 0)
 900                 fput(filp);
 901
 902         RETURN(rc);
 903 }
 904
 905
 906
 907
 908 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 909                          int len, void *karg, void __user *uarg)
 910 {
 911         struct obd_device       *obddev = class_exp2obd(exp);
 912         struct lmv_obd          *lmv = &obddev->u.lmv;
 913         struct lmv_tgt_desc     *tgt = NULL;
 914         __u32                    i = 0;
 915         int                      rc = 0;
 916         int                      set = 0;
 917         __u32                    count = lmv->desc.ld_tgt_count;
 918         ENTRY;
 919
 920         if (count == 0)
 921                 RETURN(-ENOTTY);
 922
 923         switch (cmd) {
 924         case IOC_OBD_STATFS: {
 925                 struct obd_ioctl_data *data = karg;
 926                 struct obd_device *mdc_obd;
 927                 struct obd_statfs stat_buf = {0};
 928                 __u32 index;
 929
 930                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 931                 if ((index >= count))
 932                         RETURN(-ENODEV);
 933
 934                 tgt = lmv->tgts[index];
 935                 if (tgt == NULL || !tgt->ltd_active)
 936                         RETURN(-ENODATA);
 937
 938                 mdc_obd = class_exp2obd(tgt->ltd_exp);
 939                 if (!mdc_obd)
 940                         RETURN(-EINVAL);
 941
 942                 /* copy UUID */
 943                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 944                                  min((int) data->ioc_plen2,
 945                                      (int) sizeof(struct obd_uuid))))
 946                         RETURN(-EFAULT);
 947
 948                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
 949                                 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
 950                                 0);
 951                 if (rc)
 952                         RETURN(rc);
 953                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
 954                                  min((int) data->ioc_plen1,
 955                                      (int) sizeof(stat_buf))))
 956                         RETURN(-EFAULT);
 957                 break;
 958         }
 959         case OBD_IOC_QUOTACTL: {
 960                 struct if_quotactl *qctl = karg;
 961                 struct obd_quotactl *oqctl;
 962
 963                 if (qctl->qc_valid == QC_MDTIDX) {
 964                         if (count <= qctl->qc_idx)
 965                                 RETURN(-EINVAL);
 966
 967                         tgt = lmv->tgts[qctl->qc_idx];
 968                         if (tgt == NULL || tgt->ltd_exp == NULL)
 969                                 RETURN(-EINVAL);
 970                 } else if (qctl->qc_valid == QC_UUID) {
 971                         for (i = 0; i < count; i++) {
 972                                 tgt = lmv->tgts[i];
 973                                 if (tgt == NULL)
 974                                         continue;
 975                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 976                                                      &qctl->obd_uuid))
 977                                         continue;
 978
 979                                 if (tgt->ltd_exp == NULL)
 980                                         RETURN(-EINVAL);
 981
 982                                 break;
 983                         }
 984                 } else {
 985                         RETURN(-EINVAL);
 986                 }
 987
 988                 if (i >= count)
 989                         RETURN(-EAGAIN);
 990
 991                 LASSERT(tgt != NULL && tgt->ltd_exp != NULL);
 992                 OBD_ALLOC_PTR(oqctl);
 993                 if (!oqctl)
 994                         RETURN(-ENOMEM);
 995
 996                 QCTL_COPY(oqctl, qctl);
 997                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 998                 if (rc == 0) {
 999                         QCTL_COPY(qctl, oqctl);
1000                         qctl->qc_valid = QC_MDTIDX;
1001                         qctl->obd_uuid = tgt->ltd_uuid;
1002                 }
1003                 OBD_FREE_PTR(oqctl);
1004                 break;
1005         }
1006         case OBD_IOC_CHANGELOG_SEND:
1007         case OBD_IOC_CHANGELOG_CLEAR: {
1008                 struct ioc_changelog *icc = karg;
1009
1010                 if (icc->icc_mdtindex >= count)
1011                         RETURN(-ENODEV);
1012
1013                 tgt = lmv->tgts[icc->icc_mdtindex];
1014                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
1015                         RETURN(-ENODEV);
1016                 rc = obd_iocontrol(cmd, tgt->ltd_exp, sizeof(*icc), icc, NULL);
1017                 break;
1018         }
1019         case LL_IOC_GET_CONNECT_FLAGS: {
1020                 tgt = lmv->tgts[0];
1021                 if (tgt == NULL || tgt->ltd_exp == NULL)
1022                         RETURN(-ENODATA);
1023                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1024                 break;
1025         }
1026         case LL_IOC_FID2MDTIDX: {
1027                 struct lu_fid *fid = karg;
1028                 int             mdt_index;
1029
1030                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
1031                 if (rc != 0)
1032                         RETURN(rc);
1033
1034                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
1035                  * point to user space memory for FID2MDTIDX. */
1036                 *(__u32 *)uarg = mdt_index;
1037                 break;
1038         }
1039         case OBD_IOC_FID2PATH: {
1040                 rc = lmv_fid2path(exp, len, karg, uarg);
1041                 break;
1042         }
1043         case LL_IOC_HSM_STATE_GET:
1044         case LL_IOC_HSM_STATE_SET:
1045         case LL_IOC_HSM_ACTION: {
1046                 struct md_op_data       *op_data = karg;
1047
1048                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1049                 if (IS_ERR(tgt))
1050                         RETURN(PTR_ERR(tgt));
1051
1052                 if (tgt->ltd_exp == NULL)
1053                         RETURN(-EINVAL);
1054
1055                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1056                 break;
1057         }
1058         case LL_IOC_HSM_PROGRESS: {
1059                 const struct hsm_progress_kernel *hpk = karg;
1060
1061                 tgt = lmv_find_target(lmv, &hpk->hpk_fid);
1062                 if (IS_ERR(tgt))
1063                         RETURN(PTR_ERR(tgt));
1064                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1065                 break;
1066         }
1067         case LL_IOC_HSM_REQUEST: {
1068                 struct hsm_user_request *hur = karg;
1069                 unsigned int reqcount = hur->hur_request.hr_itemcount;
1070
1071                 if (reqcount == 0)
1072                         RETURN(0);
1073
1074                 /* if the request is about a single fid
1075                  * or if there is a single MDS, no need to split
1076                  * the request. */
1077                 if (reqcount == 1 || count == 1) {
1078                         tgt = lmv_find_target(lmv,
1079                                               &hur->hur_user_item[0].hui_fid);
1080                         if (IS_ERR(tgt))
1081                                 RETURN(PTR_ERR(tgt));
1082                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1083                 } else {
1084                         /* split fid list to their respective MDS */
1085                         for (i = 0; i < count; i++) {
1086                                 unsigned int            nr, reqlen;
1087                                 int                     rc1;
1088                                 struct hsm_user_request *req;
1089
1090                                 tgt = lmv->tgts[i];
1091                                 if (tgt == NULL || tgt->ltd_exp == NULL)
1092                                         continue;
1093
1094                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
1095                                 if (nr == 0) /* nothing for this MDS */
1096                                         continue;
1097
1098                                 /* build a request with fids for this MDS */
1099                                 reqlen = offsetof(typeof(*hur),
1100                                                   hur_user_item[nr])
1101                                                 + hur->hur_request.hr_data_len;
1102                                 OBD_ALLOC_LARGE(req, reqlen);
1103                                 if (req == NULL)
1104                                         RETURN(-ENOMEM);
1105
1106                                 lmv_hsm_req_build(lmv, hur, tgt, req);
1107
1108                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
1109                                                     req, uarg);
1110                                 if (rc1 != 0 && rc == 0)
1111                                         rc = rc1;
1112                                 OBD_FREE_LARGE(req, reqlen);
1113                         }
1114                 }
1115                 break;
1116         }
1117         case LL_IOC_LOV_SWAP_LAYOUTS: {
1118                 struct md_op_data       *op_data = karg;
1119                 struct lmv_tgt_desc     *tgt1, *tgt2;
1120
1121                 tgt1 = lmv_find_target(lmv, &op_data->op_fid1);
1122                 if (IS_ERR(tgt1))
1123                         RETURN(PTR_ERR(tgt1));
1124
1125                 tgt2 = lmv_find_target(lmv, &op_data->op_fid2);
1126                 if (IS_ERR(tgt2))
1127                         RETURN(PTR_ERR(tgt2));
1128
1129                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1130                         RETURN(-EINVAL);
1131
1132                 /* only files on same MDT can have their layouts swapped */
1133                 if (tgt1->ltd_idx != tgt2->ltd_idx)
1134                         RETURN(-EPERM);
1135
1136                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1137                 break;
1138         }
1139         case LL_IOC_HSM_CT_START: {
1140                 struct lustre_kernelcomm *lk = karg;
1141                 if (lk->lk_flags & LK_FLG_STOP)
1142                         rc = lmv_hsm_ct_unregister(lmv, cmd, len, lk, uarg);
1143                 else
1144                         rc = lmv_hsm_ct_register(lmv, cmd, len, lk, uarg);
1145                 break;
1146         }
1147         default:
1148                 for (i = 0; i < count; i++) {
1149                         struct obd_device *mdc_obd;
1150                         int err;
1151
1152                         tgt = lmv->tgts[i];
1153                         if (tgt == NULL || tgt->ltd_exp == NULL)
1154                                 continue;
1155                         /* ll_umount_begin() sets force flag but for lmv, not
1156                          * mdc. Let's pass it through */
1157                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1158                         mdc_obd->obd_force = obddev->obd_force;
1159                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1160                         if (err) {
1161                                 if (tgt->ltd_active) {
1162                                         CERROR("error: iocontrol MDC %s on MDT"
1163                                                " idx %d cmd %x: err = %d\n",
1164                                                tgt->ltd_uuid.uuid, i, cmd, err);
1165                                         if (!rc)
1166                                                 rc = err;
1167                                 }
1168                         } else
1169                                 set = 1;
1170                 }
1171                 if (!set && !rc)
1172                         rc = -EIO;
1173         }
1174         RETURN(rc);
1175 }
1176
1177 #if 0
1178 static int lmv_all_chars_policy(int count, const char *name,
1179                                 int len)
1180 {
1181         unsigned int c = 0;
1182
1183         while (len > 0)
1184                 c += name[--len];
1185         c = c % count;
1186         return c;
1187 }
1188
1189 static int lmv_nid_policy(struct lmv_obd *lmv)
1190 {
1191         struct obd_import *imp;
1192         __u32              id;
1193
1194         /*
1195          * XXX: To get nid we assume that underlying obd device is mdc.
1196          */
1197         imp = class_exp2cliimp(lmv->tgts[0].ltd_exp);
1198         id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32);
1199         return id % lmv->desc.ld_tgt_count;
1200 }
1201
1202 static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
1203                           placement_policy_t placement)
1204 {
1205         switch (placement) {
1206         case PLACEMENT_CHAR_POLICY:
1207                 return lmv_all_chars_policy(lmv->desc.ld_tgt_count,
1208                                             op_data->op_name,
1209                                             op_data->op_namelen);
1210         case PLACEMENT_NID_POLICY:
1211                 return lmv_nid_policy(lmv);
1212
1213         default:
1214                 break;
1215         }
1216
1217         CERROR("Unsupported placement policy %x\n", placement);
1218         return -EINVAL;
1219 }
1220 #endif
1221
1222 /**
1223  * This is _inode_ placement policy function (not name).
1224  */
1225 static int lmv_placement_policy(struct obd_device *obd,
1226                                 struct md_op_data *op_data, u32 *mds)
1227 {
1228         struct lmv_obd          *lmv = &obd->u.lmv;
1229         ENTRY;
1230
1231         LASSERT(mds != NULL);
1232
1233         if (lmv->desc.ld_tgt_count == 1) {
1234                 *mds = 0;
1235                 RETURN(0);
1236         }
1237
1238         if (op_data->op_default_stripe_offset != -1) {
1239                 *mds = op_data->op_default_stripe_offset;
1240                 RETURN(0);
1241         }
1242
1243         /**
1244          * If stripe_offset is provided during setdirstripe
1245          * (setdirstripe -i xx), xx MDS will be choosen.
1246          */
1247         if (op_data->op_cli_flags & CLI_SET_MEA && op_data->op_data != NULL) {
1248                 struct lmv_user_md *lum;
1249
1250                 lum = op_data->op_data;
1251
1252                 if (le32_to_cpu(lum->lum_stripe_offset) != (__u32)-1) {
1253                         *mds = le32_to_cpu(lum->lum_stripe_offset);
1254                 } else {
1255                         /* -1 means default, which will be in the same MDT with
1256                          * the stripe */
1257                         *mds = op_data->op_mds;
1258                         lum->lum_stripe_offset = cpu_to_le32(op_data->op_mds);
1259                 }
1260         } else {
1261                 /* Allocate new fid on target according to operation type and
1262                  * parent home mds. */
1263                 *mds = op_data->op_mds;
1264         }
1265
1266         RETURN(0);
1267 }
1268
1269 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid, u32 mds)
1270 {
1271         struct lmv_tgt_desc     *tgt;
1272         int                      rc;
1273         ENTRY;
1274
1275         tgt = lmv_get_target(lmv, mds, NULL);
1276         if (IS_ERR(tgt))
1277                 RETURN(PTR_ERR(tgt));
1278
1279         /*
1280          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1281          * on server that seq in new allocated fid is not yet known.
1282          */
1283         mutex_lock(&tgt->ltd_fid_mutex);
1284
1285         if (tgt->ltd_active == 0 || tgt->ltd_exp == NULL)
1286                 GOTO(out, rc = -ENODEV);
1287
1288         /*
1289          * Asking underlying tgt layer to allocate new fid.
1290          */
1291         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1292         if (rc > 0) {
1293                 LASSERT(fid_is_sane(fid));
1294                 rc = 0;
1295         }
1296
1297         EXIT;
1298 out:
1299         mutex_unlock(&tgt->ltd_fid_mutex);
1300         return rc;
1301 }
1302
1303 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1304                   struct lu_fid *fid, struct md_op_data *op_data)
1305 {
1306         struct obd_device     *obd = class_exp2obd(exp);
1307         struct lmv_obd        *lmv = &obd->u.lmv;
1308         u32                    mds = 0;
1309         int                    rc;
1310         ENTRY;
1311
1312         LASSERT(op_data != NULL);
1313         LASSERT(fid != NULL);
1314
1315         rc = lmv_placement_policy(obd, op_data, &mds);
1316         if (rc) {
1317                 CERROR("Can't get target for allocating fid, "
1318                        "rc %d\n", rc);
1319                 RETURN(rc);
1320         }
1321
1322         rc = __lmv_fid_alloc(lmv, fid, mds);
1323         if (rc) {
1324                 CERROR("Can't alloc new fid, rc %d\n", rc);
1325                 RETURN(rc);
1326         }
1327
1328         RETURN(rc);
1329 }
1330
1331 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1332 {
1333         struct lmv_obd  *lmv = &obd->u.lmv;
1334         struct lmv_desc *desc;
1335         int             rc;
1336         ENTRY;
1337
1338         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1339                 CERROR("LMV setup requires a descriptor\n");
1340                 RETURN(-EINVAL);
1341         }
1342
1343         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1344         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1345                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1346                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1347                 RETURN(-EINVAL);
1348         }
1349
1350         lmv->tgts_size = 32U;
1351         OBD_ALLOC(lmv->tgts, sizeof(*lmv->tgts) * lmv->tgts_size);
1352         if (lmv->tgts == NULL)
1353                 RETURN(-ENOMEM);
1354
1355         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1356         lmv->desc.ld_tgt_count = 0;
1357         lmv->desc.ld_active_tgt_count = 0;
1358         lmv->max_def_easize = 0;
1359         lmv->max_easize = 0;
1360         lmv->lmv_placement = PLACEMENT_CHAR_POLICY;
1361
1362         spin_lock_init(&lmv->lmv_lock);
1363         mutex_init(&lmv->lmv_init_mutex);
1364
1365 #ifdef CONFIG_PROC_FS
1366         obd->obd_vars = lprocfs_lmv_obd_vars;
1367         lprocfs_obd_setup(obd);
1368         lprocfs_alloc_md_stats(obd, 0);
1369         rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd",
1370                                 0444, &lmv_proc_target_fops, obd);
1371         if (rc)
1372                 CWARN("%s: error adding LMV target_obd file: rc = %d\n",
1373                       obd->obd_name, rc);
1374 #endif
1375         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1376                              LUSTRE_CLI_FLD_HASH_DHT);
1377         if (rc) {
1378                 CERROR("Can't init FLD, err %d\n", rc);
1379                 GOTO(out, rc);
1380         }
1381
1382         RETURN(0);
1383
1384 out:
1385         return rc;
1386 }
1387
1388 static int lmv_cleanup(struct obd_device *obd)
1389 {
1390         struct lmv_obd   *lmv = &obd->u.lmv;
1391         ENTRY;
1392
1393         fld_client_fini(&lmv->lmv_fld);
1394         if (lmv->tgts != NULL) {
1395                 int i;
1396                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1397                         if (lmv->tgts[i] == NULL)
1398                                 continue;
1399                         lmv_del_target(lmv, i);
1400                 }
1401                 OBD_FREE(lmv->tgts, sizeof(*lmv->tgts) * lmv->tgts_size);
1402                 lmv->tgts_size = 0;
1403         }
1404         RETURN(0);
1405 }
1406
1407 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1408 {
1409         struct lustre_cfg       *lcfg = buf;
1410         struct obd_uuid         obd_uuid;
1411         int                     gen;
1412         __u32                   index;
1413         int                     rc;
1414         ENTRY;
1415
1416         switch (lcfg->lcfg_command) {
1417         case LCFG_ADD_MDC:
1418                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1419                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1420                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1421                         GOTO(out, rc = -EINVAL);
1422
1423                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1424
1425                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1426                         GOTO(out, rc = -EINVAL);
1427                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1428                         GOTO(out, rc = -EINVAL);
1429                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1430                 GOTO(out, rc);
1431         default:
1432                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1433                 GOTO(out, rc = -EINVAL);
1434         }
1435 out:
1436         RETURN(rc);
1437 }
1438
1439 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1440                       struct obd_statfs *osfs, __u64 max_age, __u32 flags)
1441 {
1442         struct obd_device       *obd = class_exp2obd(exp);
1443         struct lmv_obd          *lmv = &obd->u.lmv;
1444         struct obd_statfs       *temp;
1445         int                      rc = 0;
1446         __u32                    i;
1447         ENTRY;
1448
1449         rc = lmv_check_connect(obd);
1450         if (rc)
1451                 RETURN(rc);
1452
1453         OBD_ALLOC(temp, sizeof(*temp));
1454         if (temp == NULL)
1455                 RETURN(-ENOMEM);
1456
1457         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1458                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
1459                         continue;
1460
1461                 rc = obd_statfs(env, lmv->tgts[i]->ltd_exp, temp,
1462                                 max_age, flags);
1463                 if (rc) {
1464                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1465                                lmv->tgts[i]->ltd_exp->exp_obd->obd_name,
1466                                rc);
1467                         GOTO(out_free_temp, rc);
1468                 }
1469
1470                 if (i == 0) {
1471                         *osfs = *temp;
1472                         /* If the statfs is from mount, it will needs
1473                          * retrieve necessary information from MDT0.
1474                          * i.e. mount does not need the merged osfs
1475                          * from all of MDT.
1476                          * And also clients can be mounted as long as
1477                          * MDT0 is in service*/
1478                         if (flags & OBD_STATFS_FOR_MDT0)
1479                                 GOTO(out_free_temp, rc);
1480                 } else {
1481                         osfs->os_bavail += temp->os_bavail;
1482                         osfs->os_blocks += temp->os_blocks;
1483                         osfs->os_ffree += temp->os_ffree;
1484                         osfs->os_files += temp->os_files;
1485                 }
1486         }
1487
1488         EXIT;
1489 out_free_temp:
1490         OBD_FREE(temp, sizeof(*temp));
1491         return rc;
1492 }
1493
1494 static int lmv_getstatus(struct obd_export *exp, struct lu_fid *fid)
1495 {
1496         struct obd_device    *obd = exp->exp_obd;
1497         struct lmv_obd       *lmv = &obd->u.lmv;
1498         int                   rc;
1499         ENTRY;
1500
1501         rc = lmv_check_connect(obd);
1502         if (rc)
1503                 RETURN(rc);
1504
1505         rc = md_getstatus(lmv->tgts[0]->ltd_exp, fid);
1506         RETURN(rc);
1507 }
1508
1509 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1510                         u64 valid, const char *name,
1511                         const char *input, int input_size, int output_size,
1512                         int flags, struct ptlrpc_request **request)
1513 {
1514         struct obd_device      *obd = exp->exp_obd;
1515         struct lmv_obd         *lmv = &obd->u.lmv;
1516         struct lmv_tgt_desc    *tgt;
1517         int                     rc;
1518         ENTRY;
1519
1520         rc = lmv_check_connect(obd);
1521         if (rc)
1522                 RETURN(rc);
1523
1524         tgt = lmv_find_target(lmv, fid);
1525         if (IS_ERR(tgt))
1526                 RETURN(PTR_ERR(tgt));
1527
1528         rc = md_getxattr(tgt->ltd_exp, fid, valid, name, input,
1529                          input_size, output_size, flags, request);
1530
1531         RETURN(rc);
1532 }
1533
1534 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1535                         u64 valid, const char *name,
1536                         const char *input, int input_size, int output_size,
1537                         int flags, __u32 suppgid,
1538                         struct ptlrpc_request **request)
1539 {
1540         struct obd_device      *obd = exp->exp_obd;
1541         struct lmv_obd         *lmv = &obd->u.lmv;
1542         struct lmv_tgt_desc    *tgt;
1543         int                     rc;
1544         ENTRY;
1545
1546         rc = lmv_check_connect(obd);
1547         if (rc)
1548                 RETURN(rc);
1549
1550         tgt = lmv_find_target(lmv, fid);
1551         if (IS_ERR(tgt))
1552                 RETURN(PTR_ERR(tgt));
1553
1554         rc = md_setxattr(tgt->ltd_exp, fid, valid, name, input,
1555                          input_size, output_size, flags, suppgid,
1556                          request);
1557
1558         RETURN(rc);
1559 }
1560
1561 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1562                        struct ptlrpc_request **request)
1563 {
1564         struct obd_device       *obd = exp->exp_obd;
1565         struct lmv_obd          *lmv = &obd->u.lmv;
1566         struct lmv_tgt_desc     *tgt;
1567         int                      rc;
1568         ENTRY;
1569
1570         rc = lmv_check_connect(obd);
1571         if (rc)
1572                 RETURN(rc);
1573
1574         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1575         if (IS_ERR(tgt))
1576                 RETURN(PTR_ERR(tgt));
1577
1578         if (op_data->op_flags & MF_GET_MDT_IDX) {
1579                 op_data->op_mds = tgt->ltd_idx;
1580                 RETURN(0);
1581         }
1582
1583         rc = md_getattr(tgt->ltd_exp, op_data, request);
1584
1585         RETURN(rc);
1586 }
1587
1588 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1589 {
1590         struct obd_device   *obd = exp->exp_obd;
1591         struct lmv_obd      *lmv = &obd->u.lmv;
1592         __u32                i;
1593         int                  rc;
1594         ENTRY;
1595
1596         rc = lmv_check_connect(obd);
1597         if (rc)
1598                 RETURN(rc);
1599
1600         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1601
1602         /*
1603          * With DNE every object can have two locks in different namespaces:
1604          * lookup lock in space of MDT storing direntry and update/open lock in
1605          * space of MDT storing inode.
1606          */
1607         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1608                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
1609                         continue;
1610                 md_null_inode(lmv->tgts[i]->ltd_exp, fid);
1611         }
1612
1613         RETURN(0);
1614 }
1615
1616 static int lmv_find_cbdata(struct obd_export *exp, const struct lu_fid *fid,
1617                            ldlm_iterator_t it, void *data)
1618 {
1619         struct obd_device       *obd = exp->exp_obd;
1620         struct lmv_obd          *lmv = &obd->u.lmv;
1621         int                     i;
1622         int                     tgt;
1623         int                     rc;
1624         ENTRY;
1625
1626         rc = lmv_check_connect(obd);
1627         if (rc)
1628                 RETURN(rc);
1629
1630         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1631
1632         /*
1633          * With DNE every object can have two locks in different namespaces:
1634          * lookup lock in space of MDT storing direntry and update/open lock in
1635          * space of MDT storing inode.  Try the MDT that the FID maps to first,
1636          * since this can be easily found, and only try others if that fails.
1637          */
1638         for (i = 0, tgt = lmv_find_target_index(lmv, fid);
1639              i < lmv->desc.ld_tgt_count;
1640              i++, tgt = (tgt + 1) % lmv->desc.ld_tgt_count) {
1641                 if (tgt < 0) {
1642                         CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
1643                                obd->obd_name, PFID(fid), tgt);
1644                         tgt = 0;
1645                 }
1646
1647                 if (lmv->tgts[tgt] == NULL ||
1648                     lmv->tgts[tgt]->ltd_exp == NULL)
1649                         continue;
1650
1651                 rc = md_find_cbdata(lmv->tgts[tgt]->ltd_exp, fid, it, data);
1652                 if (rc)
1653                         RETURN(rc);
1654         }
1655
1656         RETURN(rc);
1657 }
1658
1659
1660 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1661                      struct md_open_data *mod, struct ptlrpc_request **request)
1662 {
1663         struct obd_device     *obd = exp->exp_obd;
1664         struct lmv_obd        *lmv = &obd->u.lmv;
1665         struct lmv_tgt_desc   *tgt;
1666         int                    rc;
1667         ENTRY;
1668
1669         rc = lmv_check_connect(obd);
1670         if (rc)
1671                 RETURN(rc);
1672
1673         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1674         if (IS_ERR(tgt))
1675                 RETURN(PTR_ERR(tgt));
1676
1677         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1678         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1679         RETURN(rc);
1680 }
1681
1682 /**
1683  * Choosing the MDT by name or FID in @op_data.
1684  * For non-striped directory, it will locate MDT by fid.
1685  * For striped-directory, it will locate MDT by name. And also
1686  * it will reset op_fid1 with the FID of the choosen stripe.
1687  **/
1688 struct lmv_tgt_desc *
1689 lmv_locate_target_for_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1690                            const char *name, int namelen, struct lu_fid *fid,
1691                            u32 *mds)
1692 {
1693         struct lmv_tgt_desc     *tgt;
1694         const struct lmv_oinfo  *oinfo;
1695
1696         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1697                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1698                         RETURN(ERR_PTR(-EBADF));
1699                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1700         } else {
1701                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen);
1702                 if (IS_ERR(oinfo))
1703                         RETURN(ERR_CAST(oinfo));
1704         }
1705
1706         if (fid != NULL)
1707                 *fid = oinfo->lmo_fid;
1708         if (mds != NULL)
1709                 *mds = oinfo->lmo_mds;
1710
1711         tgt = lmv_get_target(lmv, oinfo->lmo_mds, NULL);
1712
1713         CDEBUG(D_INFO, "locate on mds %u "DFID"\n", oinfo->lmo_mds,
1714                PFID(&oinfo->lmo_fid));
1715         return tgt;
1716 }
1717
1718 /**
1719  * Locate mds by fid or name
1720  *
1721  * For striped directory (lsm != NULL), it will locate the stripe
1722  * by name hash (see lsm_name_to_stripe_info()). Note: if the hash_type
1723  * is unknown, it will return -EBADFD, and lmv_intent_lookup might need
1724  * walk through all of stripes to locate the entry.
1725  *
1726  * For normal direcotry, it will locate MDS by FID directly.
1727  * \param[in] lmv       LMV device
1728  * \param[in] op_data   client MD stack parameters, name, namelen
1729  *                      mds_num etc.
1730  * \param[in] fid       object FID used to locate MDS.
1731  *
1732  * retval               pointer to the lmv_tgt_desc if succeed.
1733  *                      ERR_PTR(errno) if failed.
1734  */
1735 struct lmv_tgt_desc*
1736 lmv_locate_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
1737                struct lu_fid *fid)
1738 {
1739         struct lmv_stripe_md    *lsm = op_data->op_mea1;
1740         struct lmv_tgt_desc     *tgt;
1741
1742         /* During creating VOLATILE file, it should honor the mdt
1743          * index if the file under striped dir is being restored, see
1744          * ct_restore(). */
1745         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1746             (int)op_data->op_mds != -1) {
1747                 int i;
1748                 tgt = lmv_get_target(lmv, op_data->op_mds, NULL);
1749                 if (IS_ERR(tgt))
1750                         return tgt;
1751
1752                 if (lsm != NULL) {
1753                         /* refill the right parent fid */
1754                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1755                                 struct lmv_oinfo *oinfo;
1756
1757                                 oinfo = &lsm->lsm_md_oinfo[i];
1758                                 if (oinfo->lmo_mds == op_data->op_mds) {
1759                                         *fid = oinfo->lmo_fid;
1760                                         break;
1761                                 }
1762                         }
1763
1764                         if (i == lsm->lsm_md_stripe_count)
1765                                 *fid = lsm->lsm_md_oinfo[0].lmo_fid;
1766                 }
1767
1768                 return tgt;
1769         }
1770
1771         if (lsm == NULL || op_data->op_namelen == 0) {
1772                 tgt = lmv_find_target(lmv, fid);
1773                 if (IS_ERR(tgt))
1774                         return tgt;
1775
1776                 op_data->op_mds = tgt->ltd_idx;
1777                 return tgt;
1778         }
1779
1780         return lmv_locate_target_for_name(lmv, lsm, op_data->op_name,
1781                                           op_data->op_namelen, fid,
1782                                           &op_data->op_mds);
1783 }
1784
1785 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1786                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1787                 gid_t gid, cfs_cap_t cap_effective, __u64 rdev,
1788                 struct ptlrpc_request **request)
1789 {
1790         struct obd_device       *obd = exp->exp_obd;
1791         struct lmv_obd          *lmv = &obd->u.lmv;
1792         struct lmv_tgt_desc     *tgt;
1793         int                      rc;
1794         ENTRY;
1795
1796         rc = lmv_check_connect(obd);
1797         if (rc)
1798                 RETURN(rc);
1799
1800         if (!lmv->desc.ld_active_tgt_count)
1801                 RETURN(-EIO);
1802
1803         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
1804         if (IS_ERR(tgt))
1805                 RETURN(PTR_ERR(tgt));
1806
1807         CDEBUG(D_INODE, "CREATE name '%.*s' on "DFID" -> mds #%x\n",
1808                 (int)op_data->op_namelen, op_data->op_name,
1809                 PFID(&op_data->op_fid1), op_data->op_mds);
1810
1811         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1812         if (rc)
1813                 RETURN(rc);
1814         if (exp_connect_flags(exp) & OBD_CONNECT_DIR_STRIPE) {
1815                 /* Send the create request to the MDT where the object
1816                  * will be located */
1817                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
1818                 if (IS_ERR(tgt))
1819                         RETURN(PTR_ERR(tgt));
1820
1821                 op_data->op_mds = tgt->ltd_idx;
1822         } else {
1823                 CDEBUG(D_CONFIG, "Server doesn't support striped dirs\n");
1824         }
1825
1826         CDEBUG(D_INODE, "CREATE obj "DFID" -> mds #%x\n",
1827                PFID(&op_data->op_fid2), op_data->op_mds);
1828
1829         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1830         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1831                        cap_effective, rdev, request);
1832         if (rc == 0) {
1833                 if (*request == NULL)
1834                         RETURN(rc);
1835                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1836         }
1837         RETURN(rc);
1838 }
1839
1840 static int
1841 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1842             const union ldlm_policy_data *policy,
1843             struct lookup_intent *it, struct md_op_data *op_data,
1844             struct lustre_handle *lockh, __u64 extra_lock_flags)
1845 {
1846         struct obd_device        *obd = exp->exp_obd;
1847         struct lmv_obd           *lmv = &obd->u.lmv;
1848         struct lmv_tgt_desc      *tgt;
1849         int                       rc;
1850         ENTRY;
1851
1852         rc = lmv_check_connect(obd);
1853         if (rc)
1854                 RETURN(rc);
1855
1856         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID"\n",
1857                LL_IT2STR(it), PFID(&op_data->op_fid1));
1858
1859         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
1860         if (IS_ERR(tgt))
1861                 RETURN(PTR_ERR(tgt));
1862
1863         CDEBUG(D_INODE, "ENQUEUE '%s' on "DFID" -> mds #%u\n",
1864                LL_IT2STR(it), PFID(&op_data->op_fid1), tgt->ltd_idx);
1865
1866         rc = md_enqueue(tgt->ltd_exp, einfo, policy, it, op_data, lockh,
1867                         extra_lock_flags);
1868
1869         RETURN(rc);
1870 }
1871
1872 static int
1873 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1874                  struct ptlrpc_request **preq)
1875 {
1876         struct ptlrpc_request   *req = NULL;
1877         struct obd_device       *obd = exp->exp_obd;
1878         struct lmv_obd          *lmv = &obd->u.lmv;
1879         struct lmv_tgt_desc     *tgt;
1880         struct mdt_body         *body;
1881         int                      rc;
1882         ENTRY;
1883
1884         rc = lmv_check_connect(obd);
1885         if (rc)
1886                 RETURN(rc);
1887
1888         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
1889         if (IS_ERR(tgt))
1890                 RETURN(PTR_ERR(tgt));
1891
1892         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1893                 (int)op_data->op_namelen, op_data->op_name,
1894                 PFID(&op_data->op_fid1), tgt->ltd_idx);
1895
1896         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
1897         if (rc != 0)
1898                 RETURN(rc);
1899
1900         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
1901         LASSERT(body != NULL);
1902
1903         if (body->mbo_valid & OBD_MD_MDS) {
1904                 struct lu_fid rid = body->mbo_fid1;
1905                 CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1906                        PFID(&rid));
1907
1908                 tgt = lmv_find_target(lmv, &rid);
1909                 if (IS_ERR(tgt)) {
1910                         ptlrpc_req_finished(*preq);
1911                         preq = NULL;
1912                         RETURN(PTR_ERR(tgt));
1913                 }
1914
1915                 op_data->op_fid1 = rid;
1916                 op_data->op_valid |= OBD_MD_FLCROSSREF;
1917                 op_data->op_namelen = 0;
1918                 op_data->op_name = NULL;
1919                 rc = md_getattr_name(tgt->ltd_exp, op_data, &req);
1920                 ptlrpc_req_finished(*preq);
1921                 *preq = req;
1922         }
1923
1924         RETURN(rc);
1925 }
1926
1927 #define md_op_data_fid(op_data, fl)                     \
1928         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1929          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1930          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1931          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1932          NULL)
1933
1934 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
1935                             struct md_op_data *op_data, __u32 op_tgt,
1936                             enum ldlm_mode mode, int bits, int flag)
1937 {
1938         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1939         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
1940         union ldlm_policy_data policy = { { 0 } };
1941         int rc = 0;
1942         ENTRY;
1943
1944         if (!fid_is_sane(fid))
1945                 RETURN(0);
1946
1947         if (tgt == NULL) {
1948                 tgt = lmv_find_target(lmv, fid);
1949                 if (IS_ERR(tgt))
1950                         RETURN(PTR_ERR(tgt));
1951         }
1952
1953         if (tgt->ltd_idx != op_tgt) {
1954                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1955                 policy.l_inodebits.bits = bits;
1956                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1957                                       mode, LCF_ASYNC, NULL);
1958         } else {
1959                 CDEBUG(D_INODE,
1960                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
1961                        op_tgt, PFID(fid));
1962                 op_data->op_flags |= flag;
1963                 rc = 0;
1964         }
1965
1966         RETURN(rc);
1967 }
1968
1969 /*
1970  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1971  * op_data->op_fid2
1972  */
1973 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1974                     struct ptlrpc_request **request)
1975 {
1976         struct obd_device       *obd = exp->exp_obd;
1977         struct lmv_obd          *lmv = &obd->u.lmv;
1978         struct lmv_tgt_desc     *tgt;
1979         int                      rc;
1980         ENTRY;
1981
1982         rc = lmv_check_connect(obd);
1983         if (rc)
1984                 RETURN(rc);
1985
1986         LASSERT(op_data->op_namelen != 0);
1987
1988         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
1989                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
1990                op_data->op_name, PFID(&op_data->op_fid1));
1991
1992         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
1993         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
1994         op_data->op_cap = cfs_curproc_cap_pack();
1995         if (op_data->op_mea2 != NULL) {
1996                 struct lmv_stripe_md    *lsm = op_data->op_mea2;
1997                 const struct lmv_oinfo  *oinfo;
1998
1999                 oinfo = lsm_name_to_stripe_info(lsm, op_data->op_name,
2000                                                 op_data->op_namelen);
2001                 if (IS_ERR(oinfo))
2002                         RETURN(PTR_ERR(oinfo));
2003
2004                 op_data->op_fid2 = oinfo->lmo_fid;
2005         }
2006
2007         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid2);
2008         if (IS_ERR(tgt))
2009                 RETURN(PTR_ERR(tgt));
2010
2011         /*
2012          * Cancel UPDATE lock on child (fid1).
2013          */
2014         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2015         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_idx, LCK_EX,
2016                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2017         if (rc != 0)
2018                 RETURN(rc);
2019
2020         rc = md_link(tgt->ltd_exp, op_data, request);
2021
2022         RETURN(rc);
2023 }
2024
2025 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2026                       const char *old, size_t oldlen,
2027                       const char *new, size_t newlen,
2028                       struct ptlrpc_request **request)
2029 {
2030         struct obd_device       *obd = exp->exp_obd;
2031         struct lmv_obd          *lmv = &obd->u.lmv;
2032         struct lmv_tgt_desc     *src_tgt;
2033         struct lmv_tgt_desc     *tgt_tgt;
2034         struct obd_export       *target_exp;
2035         struct mdt_body         *body;
2036         int                     rc;
2037         ENTRY;
2038
2039         LASSERT(oldlen != 0);
2040
2041         CDEBUG(D_INODE, "RENAME %.*s in "DFID":%d to %.*s in "DFID":%d\n",
2042                (int)oldlen, old, PFID(&op_data->op_fid1),
2043                op_data->op_mea1 ? op_data->op_mea1->lsm_md_stripe_count : 0,
2044                (int)newlen, new, PFID(&op_data->op_fid2),
2045                op_data->op_mea2 ? op_data->op_mea2->lsm_md_stripe_count : 0);
2046
2047         rc = lmv_check_connect(obd);
2048         if (rc)
2049                 RETURN(rc);
2050
2051         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2052         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2053         op_data->op_cap = cfs_curproc_cap_pack();
2054         if (op_data->op_cli_flags & CLI_MIGRATE) {
2055                 LASSERTF(fid_is_sane(&op_data->op_fid3), "invalid FID "DFID"\n",
2056                          PFID(&op_data->op_fid3));
2057
2058                 if (op_data->op_mea1 != NULL) {
2059                         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2060                         struct lmv_tgt_desc     *tmp;
2061
2062                         /* Fix the parent fid for striped dir */
2063                         tmp = lmv_locate_target_for_name(lmv, lsm, old,
2064                                                          oldlen,
2065                                                          &op_data->op_fid1,
2066                                                          NULL);
2067                         if (IS_ERR(tmp))
2068                                 RETURN(PTR_ERR(tmp));
2069                 }
2070
2071                 rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
2072                 if (rc != 0)
2073                         RETURN(rc);
2074
2075                 src_tgt = lmv_find_target(lmv, &op_data->op_fid3);
2076                 if (IS_ERR(src_tgt))
2077                         RETURN(PTR_ERR(src_tgt));
2078
2079                 target_exp = src_tgt->ltd_exp;
2080         } else {
2081                 if (op_data->op_mea1 != NULL) {
2082                         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2083
2084                         src_tgt = lmv_locate_target_for_name(lmv, lsm, old,
2085                                                              oldlen,
2086                                                              &op_data->op_fid1,
2087                                                              &op_data->op_mds);
2088                 } else {
2089                         src_tgt = lmv_find_target(lmv, &op_data->op_fid1);
2090                 }
2091                 if (IS_ERR(src_tgt))
2092                         RETURN(PTR_ERR(src_tgt));
2093
2094
2095                 if (op_data->op_mea2 != NULL) {
2096                         struct lmv_stripe_md    *lsm = op_data->op_mea2;
2097
2098                         tgt_tgt = lmv_locate_target_for_name(lmv, lsm, new,
2099                                                              newlen,
2100                                                              &op_data->op_fid2,
2101                                                              &op_data->op_mds);
2102                 } else {
2103                         tgt_tgt = lmv_find_target(lmv, &op_data->op_fid2);
2104
2105                 }
2106                 if (IS_ERR(tgt_tgt))
2107                         RETURN(PTR_ERR(tgt_tgt));
2108
2109                 target_exp = tgt_tgt->ltd_exp;
2110         }
2111
2112         /*
2113          * LOOKUP lock on src child (fid3) should also be cancelled for
2114          * src_tgt in mdc_rename.
2115          */
2116         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2117
2118         /*
2119          * Cancel UPDATE locks on tgt parent (fid2), tgt_tgt is its
2120          * own target.
2121          */
2122         rc = lmv_early_cancel(exp, NULL, op_data, src_tgt->ltd_idx,
2123                               LCK_EX, MDS_INODELOCK_UPDATE,
2124                               MF_MDC_CANCEL_FID2);
2125
2126         if (rc != 0)
2127                 RETURN(rc);
2128         /*
2129          * Cancel LOOKUP locks on source child (fid3) for parent tgt_tgt.
2130          */
2131         if (fid_is_sane(&op_data->op_fid3)) {
2132                 struct lmv_tgt_desc *tgt;
2133
2134                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
2135                 if (IS_ERR(tgt))
2136                         RETURN(PTR_ERR(tgt));
2137
2138                 /* Cancel LOOKUP lock on its parent */
2139                 rc = lmv_early_cancel(exp, tgt, op_data, src_tgt->ltd_idx,
2140                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2141                                       MF_MDC_CANCEL_FID3);
2142                 if (rc != 0)
2143                         RETURN(rc);
2144
2145                 rc = lmv_early_cancel(exp, NULL, op_data, src_tgt->ltd_idx,
2146                                       LCK_EX, MDS_INODELOCK_FULL,
2147                                       MF_MDC_CANCEL_FID3);
2148                 if (rc != 0)
2149                         RETURN(rc);
2150         }
2151
2152 retry_rename:
2153         /*
2154          * Cancel all the locks on tgt child (fid4).
2155          */
2156         if (fid_is_sane(&op_data->op_fid4)) {
2157                 struct lmv_tgt_desc *tgt;
2158
2159                 rc = lmv_early_cancel(exp, NULL, op_data, src_tgt->ltd_idx,
2160                                       LCK_EX, MDS_INODELOCK_FULL,
2161                                       MF_MDC_CANCEL_FID4);
2162                 if (rc != 0)
2163                         RETURN(rc);
2164
2165                 tgt = lmv_find_target(lmv, &op_data->op_fid4);
2166                 if (IS_ERR(tgt))
2167                         RETURN(PTR_ERR(tgt));
2168
2169                 /* Since the target child might be destroyed, and it might
2170                  * become orphan, and we can only check orphan on the local
2171                  * MDT right now, so we send rename request to the MDT where
2172                  * target child is located. If target child does not exist,
2173                  * then it will send the request to the target parent */
2174                 target_exp = tgt->ltd_exp;
2175         }
2176
2177         rc = md_rename(target_exp, op_data, old, oldlen, new, newlen,
2178                        request);
2179
2180         if (rc != 0 && rc != -EXDEV)
2181                 RETURN(rc);
2182
2183         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2184         if (body == NULL)
2185                 RETURN(-EPROTO);
2186
2187         /* Not cross-ref case, just get out of here. */
2188         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2189                 RETURN(rc);
2190
2191         CDEBUG(D_INODE, "%s: try rename to another MDT for "DFID"\n",
2192                exp->exp_obd->obd_name, PFID(&body->mbo_fid1));
2193
2194         op_data->op_fid4 = body->mbo_fid1;
2195         ptlrpc_req_finished(*request);
2196         *request = NULL;
2197         goto retry_rename;
2198 }
2199
2200 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2201                        void *ea, size_t ealen, struct ptlrpc_request **request)
2202 {
2203         struct obd_device       *obd = exp->exp_obd;
2204         struct lmv_obd          *lmv = &obd->u.lmv;
2205         struct lmv_tgt_desc     *tgt;
2206         int                      rc = 0;
2207         ENTRY;
2208
2209         rc = lmv_check_connect(obd);
2210         if (rc)
2211                 RETURN(rc);
2212
2213         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x\n",
2214                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid);
2215
2216         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2217         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2218         if (IS_ERR(tgt))
2219                 RETURN(PTR_ERR(tgt));
2220
2221         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2222
2223         RETURN(rc);
2224 }
2225
2226 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2227                      struct ptlrpc_request **request)
2228 {
2229         struct obd_device       *obd = exp->exp_obd;
2230         struct lmv_obd          *lmv = &obd->u.lmv;
2231         struct lmv_tgt_desc     *tgt;
2232         int                      rc;
2233         ENTRY;
2234
2235         rc = lmv_check_connect(obd);
2236         if (rc != 0)
2237                 RETURN(rc);
2238
2239         tgt = lmv_find_target(lmv, fid);
2240         if (IS_ERR(tgt))
2241                 RETURN(PTR_ERR(tgt));
2242
2243         rc = md_fsync(tgt->ltd_exp, fid, request);
2244         RETURN(rc);
2245 }
2246
2247 /**
2248  * Get current minimum entry from striped directory
2249  *
2250  * This function will search the dir entry, whose hash value is the
2251  * closest(>=) to @hash_offset, from all of sub-stripes, and it is
2252  * only being called for striped directory.
2253  *
2254  * \param[in] exp               export of LMV
2255  * \param[in] op_data           parameters transferred beween client MD stack
2256  *                              stripe_information will be included in this
2257  *                              parameter
2258  * \param[in] cb_op             ldlm callback being used in enqueue in
2259  *                              mdc_read_page
2260  * \param[in] hash_offset       the hash value, which is used to locate
2261  *                              minum(closet) dir entry
2262  * \param[in|out] stripe_offset the caller use this to indicate the stripe
2263  *                              index of last entry, so to avoid hash conflict
2264  *                              between stripes. It will also be used to
2265  *                              return the stripe index of current dir entry.
2266  * \param[in|out] entp          the minum entry and it also is being used
2267  *                              to input the last dir entry to resolve the
2268  *                              hash conflict
2269  *
2270  * \param[out] ppage            the page which holds the minum entry
2271  *
2272  * \retval                      = 0 get the entry successfully
2273  *                              negative errno (< 0) does not get the entry
2274  */
2275 static int lmv_get_min_striped_entry(struct obd_export *exp,
2276                                      struct md_op_data *op_data,
2277                                      struct md_callback *cb_op,
2278                                      __u64 hash_offset, int *stripe_offset,
2279                                      struct lu_dirent **entp,
2280                                      struct page **ppage)
2281 {
2282         struct obd_device       *obd = exp->exp_obd;
2283         struct lmv_obd          *lmv = &obd->u.lmv;
2284         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2285         struct lmv_tgt_desc     *tgt;
2286         int                     stripe_count;
2287         struct lu_dirent        *min_ent = NULL;
2288         struct page             *min_page = NULL;
2289         int                     min_idx = 0;
2290         int                     i;
2291         int                     rc = 0;
2292         ENTRY;
2293
2294         stripe_count = lsm->lsm_md_stripe_count;
2295         for (i = 0; i < stripe_count; i++) {
2296                 struct lu_dirent        *ent = NULL;
2297                 struct page             *page = NULL;
2298                 struct lu_dirpage       *dp;
2299                 __u64                   stripe_hash = hash_offset;
2300
2301                 tgt = lmv_get_target(lmv, lsm->lsm_md_oinfo[i].lmo_mds, NULL);
2302                 if (IS_ERR(tgt))
2303                         GOTO(out, rc = PTR_ERR(tgt));
2304
2305                 /* op_data will be shared by each stripe, so we need
2306                  * reset these value for each stripe */
2307                 op_data->op_fid1 = lsm->lsm_md_oinfo[i].lmo_fid;
2308                 op_data->op_fid2 = lsm->lsm_md_oinfo[i].lmo_fid;
2309                 op_data->op_data = lsm->lsm_md_oinfo[i].lmo_root;
2310 next:
2311                 rc = md_read_page(tgt->ltd_exp, op_data, cb_op, stripe_hash,
2312                                   &page);
2313                 if (rc != 0)
2314                         GOTO(out, rc);
2315
2316                 dp = page_address(page);
2317                 for (ent = lu_dirent_start(dp); ent != NULL;
2318                      ent = lu_dirent_next(ent)) {
2319                         /* Skip dummy entry */
2320                         if (le16_to_cpu(ent->lde_namelen) == 0)
2321                                 continue;
2322
2323                         if (le64_to_cpu(ent->lde_hash) < hash_offset)
2324                                 continue;
2325
2326                         if (le64_to_cpu(ent->lde_hash) == hash_offset &&
2327                             (*entp == ent || i < *stripe_offset))
2328                                 continue;
2329
2330                         /* skip . and .. for other stripes */
2331                         if (i != 0 &&
2332                             (strncmp(ent->lde_name, ".",
2333                                      le16_to_cpu(ent->lde_namelen)) == 0 ||
2334                              strncmp(ent->lde_name, "..",
2335                                      le16_to_cpu(ent->lde_namelen)) == 0))
2336                                 continue;
2337                         break;
2338                 }
2339
2340                 if (ent == NULL) {
2341                         stripe_hash = le64_to_cpu(dp->ldp_hash_end);
2342
2343                         kunmap(page);
2344                         page_cache_release(page);
2345                         page = NULL;
2346
2347                         /* reach the end of current stripe, go to next stripe */
2348                         if (stripe_hash == MDS_DIR_END_OFF)
2349                                 continue;
2350                         else
2351                                 goto next;
2352                 }
2353
2354                 if (min_ent != NULL) {
2355                         if (le64_to_cpu(min_ent->lde_hash) >
2356                             le64_to_cpu(ent->lde_hash)) {
2357                                 min_ent = ent;
2358                                 kunmap(min_page);
2359                                 page_cache_release(min_page);
2360                                 min_idx = i;
2361                                 min_page = page;
2362                         } else {
2363                                 kunmap(page);
2364                                 page_cache_release(page);
2365                                 page = NULL;
2366                         }
2367                 } else {
2368                         min_ent = ent;
2369                         min_page = page;
2370                         min_idx = i;
2371                 }
2372         }
2373
2374 out:
2375         if (*ppage != NULL) {
2376                 kunmap(*ppage);
2377                 page_cache_release(*ppage);
2378         }
2379         *stripe_offset = min_idx;
2380         *entp = min_ent;
2381         *ppage = min_page;
2382         RETURN(rc);
2383 }
2384
2385 /**
2386  * Build dir entry page from a striped directory
2387  *
2388  * This function gets one entry by @offset from a striped directory. It will
2389  * read entries from all of stripes, and choose one closest to the required
2390  * offset(&offset). A few notes
2391  * 1. skip . and .. for non-zero stripes, because there can only have one .
2392  * and .. in a directory.
2393  * 2. op_data will be shared by all of stripes, instead of allocating new
2394  * one, so need to restore before reusing.
2395  * 3. release the entry page if that is not being chosen.
2396  *
2397  * \param[in] exp       obd export refer to LMV
2398  * \param[in] op_data   hold those MD parameters of read_entry
2399  * \param[in] cb_op     ldlm callback being used in enqueue in mdc_read_entry
2400  * \param[out] ldp      the entry being read
2401  * \param[out] ppage    the page holding the entry. Note: because the entry
2402  *                      will be accessed in upper layer, so we need hold the
2403  *                      page until the usages of entry is finished, see
2404  *                      ll_dir_entry_next.
2405  *
2406  * retval               =0 if get entry successfully
2407  *                      <0 cannot get entry
2408  */
2409 static int lmv_read_striped_page(struct obd_export *exp,
2410                                  struct md_op_data *op_data,
2411                                  struct md_callback *cb_op,
2412                                  __u64 offset, struct page **ppage)
2413 {
2414         struct obd_device       *obd = exp->exp_obd;
2415         struct lu_fid           master_fid = op_data->op_fid1;
2416         struct inode            *master_inode = op_data->op_data;
2417         __u64                   hash_offset = offset;
2418         struct lu_dirpage       *dp;
2419         struct page             *min_ent_page = NULL;
2420         struct page             *ent_page = NULL;
2421         struct lu_dirent        *ent;
2422         void                    *area;
2423         int                     ent_idx = 0;
2424         struct lu_dirent        *min_ent = NULL;
2425         struct lu_dirent        *last_ent;
2426         size_t                  left_bytes;
2427         int                     rc;
2428         ENTRY;
2429
2430         rc = lmv_check_connect(obd);
2431         if (rc)
2432                 RETURN(rc);
2433
2434         /* Allocate a page and read entries from all of stripes and fill
2435          * the page by hash order */
2436         ent_page = alloc_page(GFP_KERNEL);
2437         if (ent_page == NULL)
2438                 RETURN(-ENOMEM);
2439
2440         /* Initialize the entry page */
2441         dp = kmap(ent_page);
2442         memset(dp, 0, sizeof(*dp));
2443         dp->ldp_hash_start = cpu_to_le64(offset);
2444         dp->ldp_flags |= LDF_COLLIDE;
2445
2446         area = dp + 1;
2447         left_bytes = PAGE_CACHE_SIZE - sizeof(*dp);
2448         ent = area;
2449         last_ent = ent;
2450         do {
2451                 __u16   ent_size;
2452
2453                 /* Find the minum entry from all sub-stripes */
2454                 rc = lmv_get_min_striped_entry(exp, op_data, cb_op, hash_offset,
2455                                                &ent_idx, &min_ent,
2456                                                &min_ent_page);
2457                 if (rc != 0)
2458                         GOTO(out, rc);
2459
2460                 /* If it can not get minum entry, it means it already reaches
2461                  * the end of this directory */
2462                 if (min_ent == NULL) {
2463                         last_ent->lde_reclen = 0;
2464                         hash_offset = MDS_DIR_END_OFF;
2465                         GOTO(out, rc);
2466                 }
2467
2468                 ent_size = le16_to_cpu(min_ent->lde_reclen);
2469
2470                 /* the last entry lde_reclen is 0, but it might not
2471                  * the end of this entry of this temporay entry */
2472                 if (ent_size == 0)
2473                         ent_size = lu_dirent_calc_size(
2474                                         le16_to_cpu(min_ent->lde_namelen),
2475                                         le32_to_cpu(min_ent->lde_attrs));
2476                 if (ent_size > left_bytes) {
2477                         last_ent->lde_reclen = cpu_to_le16(0);
2478                         hash_offset = le64_to_cpu(min_ent->lde_hash);
2479                         GOTO(out, rc);
2480                 }
2481
2482                 memcpy(ent, min_ent, ent_size);
2483
2484                 /* Replace . with master FID and Replace .. with the parent FID
2485                  * of master object */
2486                 if (strncmp(ent->lde_name, ".",
2487                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2488                     le16_to_cpu(ent->lde_namelen) == 1)
2489                         fid_cpu_to_le(&ent->lde_fid, &master_fid);
2490                 else if (strncmp(ent->lde_name, "..",
2491                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2492                            le16_to_cpu(ent->lde_namelen) == 2)
2493                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2494
2495                 left_bytes -= ent_size;
2496                 ent->lde_reclen = cpu_to_le16(ent_size);
2497                 last_ent = ent;
2498                 ent = (void *)ent + ent_size;
2499                 hash_offset = le64_to_cpu(min_ent->lde_hash);
2500                 if (hash_offset == MDS_DIR_END_OFF) {
2501                         last_ent->lde_reclen = 0;
2502                         break;
2503                 }
2504         } while (1);
2505 out:
2506         if (min_ent_page != NULL) {
2507                 kunmap(min_ent_page);
2508                 page_cache_release(min_ent_page);
2509         }
2510
2511         if (unlikely(rc != 0)) {
2512                 __free_page(ent_page);
2513                 ent_page = NULL;
2514         } else {
2515                 if (ent == area)
2516                         dp->ldp_flags |= LDF_EMPTY;
2517                 dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2518                 dp->ldp_hash_end = cpu_to_le64(hash_offset);
2519         }
2520
2521         /* We do not want to allocate md_op_data during each
2522          * dir entry reading, so op_data will be shared by every stripe,
2523          * then we need to restore it back to original value before
2524          * return to the upper layer */
2525         op_data->op_fid1 = master_fid;
2526         op_data->op_fid2 = master_fid;
2527         op_data->op_data = master_inode;
2528
2529         *ppage = ent_page;
2530
2531         RETURN(rc);
2532 }
2533
2534 int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2535                   struct md_callback *cb_op, __u64 offset,
2536                   struct page **ppage)
2537 {
2538         struct obd_device       *obd = exp->exp_obd;
2539         struct lmv_obd          *lmv = &obd->u.lmv;
2540         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2541         struct lmv_tgt_desc     *tgt;
2542         int                     rc;
2543         ENTRY;
2544
2545         rc = lmv_check_connect(obd);
2546         if (rc != 0)
2547                 RETURN(rc);
2548
2549         if (unlikely(lsm != NULL)) {
2550                 rc = lmv_read_striped_page(exp, op_data, cb_op, offset, ppage);
2551                 RETURN(rc);
2552         }
2553
2554         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2555         if (IS_ERR(tgt))
2556                 RETURN(PTR_ERR(tgt));
2557
2558         rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2559
2560         RETURN(rc);
2561 }
2562
2563 /**
2564  * Unlink a file/directory
2565  *
2566  * Unlink a file or directory under the parent dir. The unlink request
2567  * usually will be sent to the MDT where the child is located, but if
2568  * the client does not have the child FID then request will be sent to the
2569  * MDT where the parent is located.
2570  *
2571  * If the parent is a striped directory then it also needs to locate which
2572  * stripe the name of the child is located, and replace the parent FID
2573  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2574  * it will walk through all of sub-stripes until the child is being
2575  * unlinked finally.
2576  *
2577  * \param[in] exp       export refer to LMV
2578  * \param[in] op_data   different parameters transferred beween client
2579  *                      MD stacks, name, namelen, FIDs etc.
2580  *                      op_fid1 is the parent FID, op_fid2 is the child
2581  *                      FID.
2582  * \param[out] request  point to the request of unlink.
2583  *
2584  * retval               0 if succeed
2585  *                      negative errno if failed.
2586  */
2587 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2588                       struct ptlrpc_request **request)
2589 {
2590         struct obd_device       *obd = exp->exp_obd;
2591         struct lmv_obd          *lmv = &obd->u.lmv;
2592         struct lmv_tgt_desc     *tgt = NULL;
2593         struct lmv_tgt_desc     *parent_tgt = NULL;
2594         struct mdt_body         *body;
2595         int                     rc;
2596         int                     stripe_index = 0;
2597         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2598         ENTRY;
2599
2600         rc = lmv_check_connect(obd);
2601         if (rc)
2602                 RETURN(rc);
2603 retry_unlink:
2604         /* For striped dir, we need to locate the parent as well */
2605         if (lsm != NULL) {
2606                 struct lmv_tgt_desc *tmp;
2607
2608                 LASSERT(op_data->op_name != NULL &&
2609                         op_data->op_namelen != 0);
2610
2611                 tmp = lmv_locate_target_for_name(lmv, lsm,
2612                                                  op_data->op_name,
2613                                                  op_data->op_namelen,
2614                                                  &op_data->op_fid1,
2615                                                  &op_data->op_mds);
2616
2617                 /* return -EBADFD means unknown hash type, might
2618                  * need try all sub-stripe here */
2619                 if (IS_ERR(tmp) && PTR_ERR(tmp) != -EBADFD)
2620                         RETURN(PTR_ERR(tmp));
2621
2622                 /* Note: both migrating dir and unknown hash dir need to
2623                  * try all of sub-stripes, so we need start search the
2624                  * name from stripe 0, but migrating dir is already handled
2625                  * inside lmv_locate_target_for_name(), so we only check
2626                  * unknown hash type directory here */
2627                 if (!lmv_is_known_hash_type(lsm->lsm_md_hash_type)) {
2628                         struct lmv_oinfo *oinfo;
2629
2630                         oinfo = &lsm->lsm_md_oinfo[stripe_index];
2631
2632                         op_data->op_fid1 = oinfo->lmo_fid;
2633                         op_data->op_mds = oinfo->lmo_mds;
2634                 }
2635         }
2636
2637 try_next_stripe:
2638         /* Send unlink requests to the MDT where the child is located */
2639         if (likely(!fid_is_zero(&op_data->op_fid2)))
2640                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2641         else if (lsm != NULL)
2642                 tgt = lmv_get_target(lmv, op_data->op_mds, NULL);
2643         else
2644                 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
2645
2646         if (IS_ERR(tgt))
2647                 RETURN(PTR_ERR(tgt));
2648
2649         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2650         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2651         op_data->op_cap = cfs_curproc_cap_pack();
2652
2653         /*
2654          * If child's fid is given, cancel unused locks for it if it is from
2655          * another export than parent.
2656          *
2657          * LOOKUP lock for child (fid3) should also be cancelled on parent
2658          * tgt_tgt in mdc_unlink().
2659          */
2660         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2661
2662         /*
2663          * Cancel FULL locks on child (fid3).
2664          */
2665         parent_tgt = lmv_find_target(lmv, &op_data->op_fid1);
2666         if (IS_ERR(parent_tgt))
2667                 RETURN(PTR_ERR(parent_tgt));
2668
2669         if (parent_tgt != tgt) {
2670                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_idx,
2671                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2672                                       MF_MDC_CANCEL_FID3);
2673         }
2674
2675         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_idx, LCK_EX,
2676                               MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3);
2677         if (rc != 0)
2678                 RETURN(rc);
2679
2680         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
2681                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2), tgt->ltd_idx);
2682
2683         rc = md_unlink(tgt->ltd_exp, op_data, request);
2684         if (rc != 0 && rc != -EREMOTE && rc != -ENOENT)
2685                 RETURN(rc);
2686
2687         /* Try next stripe if it is needed. */
2688         if (rc == -ENOENT && lsm != NULL && lmv_need_try_all_stripes(lsm)) {
2689                 struct lmv_oinfo *oinfo;
2690
2691                 stripe_index++;
2692                 if (stripe_index >= lsm->lsm_md_stripe_count)
2693                         RETURN(rc);
2694
2695                 oinfo = &lsm->lsm_md_oinfo[stripe_index];
2696
2697                 op_data->op_fid1 = oinfo->lmo_fid;
2698                 op_data->op_mds = oinfo->lmo_mds;
2699
2700                 ptlrpc_req_finished(*request);
2701                 *request = NULL;
2702
2703                 goto try_next_stripe;
2704         }
2705
2706         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2707         if (body == NULL)
2708                 RETURN(-EPROTO);
2709
2710         /* Not cross-ref case, just get out of here. */
2711         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2712                 RETURN(rc);
2713
2714         CDEBUG(D_INODE, "%s: try unlink to another MDT for "DFID"\n",
2715                exp->exp_obd->obd_name, PFID(&body->mbo_fid1));
2716
2717         /* This is a remote object, try remote MDT, Note: it may
2718          * try more than 1 time here, Considering following case
2719          * /mnt/lustre is root on MDT0, remote1 is on MDT1
2720          * 1. Initially A does not know where remote1 is, it send
2721          *    unlink RPC to MDT0, MDT0 return -EREMOTE, it will
2722          *    resend unlink RPC to MDT1 (retry 1st time).
2723          *
2724          * 2. During the unlink RPC in flight,
2725          *    client B mv /mnt/lustre/remote1 /mnt/lustre/remote2
2726          *    and create new remote1, but on MDT0
2727          *
2728          * 3. MDT1 get unlink RPC(from A), then do remote lock on
2729          *    /mnt/lustre, then lookup get fid of remote1, and find
2730          *    it is remote dir again, and replay -EREMOTE again.
2731          *
2732          * 4. Then A will resend unlink RPC to MDT0. (retry 2nd times).
2733          *
2734          * In theory, it might try unlimited time here, but it should
2735          * be very rare case.  */
2736         op_data->op_fid2 = body->mbo_fid1;
2737         ptlrpc_req_finished(*request);
2738         *request = NULL;
2739
2740         goto retry_unlink;
2741 }
2742
2743 static int lmv_precleanup(struct obd_device *obd)
2744 {
2745         ENTRY;
2746         fld_client_proc_fini(&obd->u.lmv.lmv_fld);
2747         lprocfs_obd_cleanup(obd);
2748         lprocfs_free_md_stats(obd);
2749         RETURN(0);
2750 }
2751
2752 /**
2753  * Get by key a value associated with a LMV device.
2754  *
2755  * Dispatch request to lower-layer devices as needed.
2756  *
2757  * \param[in] env               execution environment for this thread
2758  * \param[in] exp               export for the LMV device
2759  * \param[in] keylen            length of key identifier
2760  * \param[in] key               identifier of key to get value for
2761  * \param[in] vallen            size of \a val
2762  * \param[out] val              pointer to storage location for value
2763  * \param[in] lsm               optional striping metadata of object
2764  *
2765  * \retval 0            on success
2766  * \retval negative     negated errno on failure
2767  */
2768 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2769                         __u32 keylen, void *key, __u32 *vallen, void *val)
2770 {
2771         struct obd_device       *obd;
2772         struct lmv_obd          *lmv;
2773         int                      rc = 0;
2774         ENTRY;
2775
2776         obd = class_exp2obd(exp);
2777         if (obd == NULL) {
2778                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2779                        exp->exp_handle.h_cookie);
2780                 RETURN(-EINVAL);
2781         }
2782
2783         lmv = &obd->u.lmv;
2784         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2785                 int i;
2786
2787                 rc = lmv_check_connect(obd);
2788                 if (rc)
2789                         RETURN(rc);
2790
2791                 LASSERT(*vallen == sizeof(__u32));
2792                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2793                         struct lmv_tgt_desc *tgt = lmv->tgts[i];
2794                         /*
2795                          * All tgts should be connected when this gets called.
2796                          */
2797                         if (tgt == NULL || tgt->ltd_exp == NULL)
2798                                 continue;
2799
2800                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2801                                           vallen, val))
2802                                 RETURN(0);
2803                 }
2804                 RETURN(-EINVAL);
2805         } else if (KEY_IS(KEY_MAX_EASIZE) ||
2806                    KEY_IS(KEY_DEFAULT_EASIZE) ||
2807                    KEY_IS(KEY_CONN_DATA)) {
2808                 rc = lmv_check_connect(obd);
2809                 if (rc)
2810                         RETURN(rc);
2811
2812                 /*
2813                  * Forwarding this request to first MDS, it should know LOV
2814                  * desc.
2815                  */
2816                 rc = obd_get_info(env, lmv->tgts[0]->ltd_exp, keylen, key,
2817                                   vallen, val);
2818                 if (!rc && KEY_IS(KEY_CONN_DATA))
2819                         exp->exp_connect_data = *(struct obd_connect_data *)val;
2820                 RETURN(rc);
2821         } else if (KEY_IS(KEY_TGT_COUNT)) {
2822                 *((int *)val) = lmv->desc.ld_tgt_count;
2823                 RETURN(0);
2824         }
2825
2826         CDEBUG(D_IOCTL, "Invalid key\n");
2827         RETURN(-EINVAL);
2828 }
2829
2830 /**
2831  * Asynchronously set by key a value associated with a LMV device.
2832  *
2833  * Dispatch request to lower-layer devices as needed.
2834  *
2835  * \param[in] env       execution environment for this thread
2836  * \param[in] exp       export for the LMV device
2837  * \param[in] keylen    length of key identifier
2838  * \param[in] key       identifier of key to store value for
2839  * \param[in] vallen    size of value to store
2840  * \param[in] val       pointer to data to be stored
2841  * \param[in] set       optional list of related ptlrpc requests
2842  *
2843  * \retval 0            on success
2844  * \retval negative     negated errno on failure
2845  */
2846 int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
2847                         __u32 keylen, void *key, __u32 vallen, void *val,
2848                         struct ptlrpc_request_set *set)
2849 {
2850         struct lmv_tgt_desc     *tgt = NULL;
2851         struct obd_device       *obd;
2852         struct lmv_obd          *lmv;
2853         int rc = 0;
2854         ENTRY;
2855
2856         obd = class_exp2obd(exp);
2857         if (obd == NULL) {
2858                 CDEBUG(D_IOCTL, "Invalid client cookie "LPX64"\n",
2859                        exp->exp_handle.h_cookie);
2860                 RETURN(-EINVAL);
2861         }
2862         lmv = &obd->u.lmv;
2863
2864         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2865             KEY_IS(KEY_DEFAULT_EASIZE)) {
2866                 int i, err = 0;
2867
2868                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2869                         tgt = lmv->tgts[i];
2870
2871                         if (tgt == NULL || tgt->ltd_exp == NULL)
2872                                 continue;
2873
2874                         err = obd_set_info_async(env, tgt->ltd_exp,
2875                                                  keylen, key, vallen, val, set);
2876                         if (err && rc == 0)
2877                                 rc = err;
2878                 }
2879
2880                 RETURN(rc);
2881         }
2882
2883         RETURN(-EINVAL);
2884 }
2885
2886 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
2887                             const struct lmv_mds_md_v1 *lmm1)
2888 {
2889         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
2890         int             stripe_count;
2891         int             cplen;
2892         int             i;
2893         int             rc = 0;
2894         ENTRY;
2895
2896         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
2897         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
2898         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
2899         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
2900                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
2901         else
2902                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
2903         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
2904         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
2905                         sizeof(lsm->lsm_md_pool_name));
2906
2907         if (cplen >= sizeof(lsm->lsm_md_pool_name))
2908                 RETURN(-E2BIG);
2909
2910         CDEBUG(D_INFO, "unpack lsm count %d, master %d hash_type %d"
2911                "layout_version %d\n", lsm->lsm_md_stripe_count,
2912                lsm->lsm_md_master_mdt_index, lsm->lsm_md_hash_type,
2913                lsm->lsm_md_layout_version);
2914
2915         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
2916         for (i = 0; i < stripe_count; i++) {
2917                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
2918                               &lmm1->lmv_stripe_fids[i]);
2919                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
2920                                     &lsm->lsm_md_oinfo[i].lmo_mds);
2921                 if (rc != 0)
2922                         RETURN(rc);
2923                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
2924                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
2925         }
2926
2927         RETURN(rc);
2928 }
2929
2930 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
2931                         const union lmv_mds_md *lmm, size_t lmm_size)
2932 {
2933         struct lmv_stripe_md     *lsm;
2934         int                      lsm_size;
2935         int                      rc;
2936         bool                     allocated = false;
2937         ENTRY;
2938
2939         LASSERT(lsmp != NULL);
2940
2941         lsm = *lsmp;
2942         /* Free memmd */
2943         if (lsm != NULL && lmm == NULL) {
2944                 int i;
2945                 for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
2946                         /* For migrating inode, the master stripe and master
2947                          * object will be the same, so do not need iput, see
2948                          * ll_update_lsm_md */
2949                         if (!(lsm->lsm_md_hash_type & LMV_HASH_FLAG_MIGRATION &&
2950                               i == 0) && lsm->lsm_md_oinfo[i].lmo_root != NULL)
2951                                 iput(lsm->lsm_md_oinfo[i].lmo_root);
2952                 }
2953                 lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
2954                 OBD_FREE(lsm, lsm_size);
2955                 *lsmp = NULL;
2956                 RETURN(0);
2957         }
2958
2959         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
2960                 RETURN(-EPERM);
2961
2962         /* Unpack memmd */
2963         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
2964             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
2965                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
2966                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
2967                        -EIO);
2968                 RETURN(-EIO);
2969         }
2970
2971         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
2972                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
2973         else
2974                 /**
2975                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
2976                  * stripecount should be 0 then.
2977                  */
2978                 lsm_size = lmv_stripe_md_size(0);
2979
2980         lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
2981         if (lsm == NULL) {
2982                 OBD_ALLOC(lsm, lsm_size);
2983                 if (lsm == NULL)
2984                         RETURN(-ENOMEM);
2985                 allocated = true;
2986                 *lsmp = lsm;
2987         }
2988
2989         switch (le32_to_cpu(lmm->lmv_magic)) {
2990         case LMV_MAGIC_V1:
2991                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
2992                 break;
2993         default:
2994                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
2995                        le32_to_cpu(lmm->lmv_magic));
2996                 rc = -EINVAL;
2997                 break;
2998         }
2999
3000         if (rc != 0 && allocated) {
3001                 OBD_FREE(lsm, lsm_size);
3002                 *lsmp = NULL;
3003                 lsm_size = rc;
3004         }
3005         RETURN(lsm_size);
3006 }
3007
3008 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3009 {
3010         lmv_unpackmd(NULL, &lsm, NULL, 0);
3011 }
3012 EXPORT_SYMBOL(lmv_free_memmd);
3013
3014 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3015                              union ldlm_policy_data *policy,
3016                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3017                              void *opaque)
3018 {
3019         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3020         int rc = 0;
3021         __u32 i;
3022         ENTRY;
3023
3024         LASSERT(fid != NULL);
3025
3026         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3027                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
3028                 int err;
3029
3030                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
3031                         continue;
3032
3033                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3034                                        opaque);
3035                 if (!rc)
3036                         rc = err;
3037         }
3038         RETURN(rc);
3039 }
3040
3041 int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data,
3042                       __u64 *bits)
3043 {
3044         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
3045         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3046         int                      rc;
3047         ENTRY;
3048
3049         if (tgt == NULL || tgt->ltd_exp == NULL)
3050                 RETURN(-EINVAL);
3051         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3052         RETURN(rc);
3053 }
3054
3055 enum ldlm_mode lmv_lock_match(struct obd_export *exp, __u64 flags,
3056                               const struct lu_fid *fid, enum ldlm_type type,
3057                               union ldlm_policy_data *policy,
3058                               enum ldlm_mode mode, struct lustre_handle *lockh)
3059 {
3060         struct obd_device       *obd = exp->exp_obd;
3061         struct lmv_obd          *lmv = &obd->u.lmv;
3062         enum ldlm_mode          rc;
3063         int                     tgt;
3064         int                     i;
3065         ENTRY;
3066
3067         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
3068
3069         /*
3070          * With DNE every object can have two locks in different namespaces:
3071          * lookup lock in space of MDT storing direntry and update/open lock in
3072          * space of MDT storing inode.  Try the MDT that the FID maps to first,
3073          * since this can be easily found, and only try others if that fails.
3074          */
3075         for (i = 0, tgt = lmv_find_target_index(lmv, fid);
3076              i < lmv->desc.ld_tgt_count;
3077              i++, tgt = (tgt + 1) % lmv->desc.ld_tgt_count) {
3078                 if (tgt < 0) {
3079                         CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
3080                                obd->obd_name, PFID(fid), tgt);
3081                         tgt = 0;
3082                 }
3083
3084                 if (lmv->tgts[tgt] == NULL ||
3085                     lmv->tgts[tgt]->ltd_exp == NULL ||
3086                     lmv->tgts[tgt]->ltd_active == 0)
3087                         continue;
3088
3089                 rc = md_lock_match(lmv->tgts[tgt]->ltd_exp, flags, fid,
3090                                    type, policy, mode, lockh);
3091                 if (rc)
3092                         RETURN(rc);
3093         }
3094
3095         RETURN(0);
3096 }
3097
3098 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
3099                       struct obd_export *dt_exp, struct obd_export *md_exp,
3100                       struct lustre_md *md)
3101 {
3102         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
3103         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3104
3105         if (tgt == NULL || tgt->ltd_exp == NULL)
3106                 RETURN(-EINVAL);
3107
3108         return md_get_lustre_md(lmv->tgts[0]->ltd_exp, req, dt_exp, md_exp, md);
3109 }
3110
3111 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
3112 {
3113         struct obd_device       *obd = exp->exp_obd;
3114         struct lmv_obd          *lmv = &obd->u.lmv;
3115         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3116         ENTRY;
3117
3118         if (md->lmv != NULL) {
3119                 lmv_free_memmd(md->lmv);
3120                 md->lmv = NULL;
3121         }
3122         if (tgt == NULL || tgt->ltd_exp == NULL)
3123                 RETURN(-EINVAL);
3124         RETURN(md_free_lustre_md(lmv->tgts[0]->ltd_exp, md));
3125 }
3126
3127 int lmv_set_open_replay_data(struct obd_export *exp,
3128                              struct obd_client_handle *och,
3129                              struct lookup_intent *it)
3130 {
3131         struct obd_device       *obd = exp->exp_obd;
3132         struct lmv_obd          *lmv = &obd->u.lmv;
3133         struct lmv_tgt_desc     *tgt;
3134         ENTRY;
3135
3136         tgt = lmv_find_target(lmv, &och->och_fid);
3137         if (IS_ERR(tgt))
3138                 RETURN(PTR_ERR(tgt));
3139
3140         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3141 }
3142
3143 int lmv_clear_open_replay_data(struct obd_export *exp,
3144                                struct obd_client_handle *och)
3145 {
3146         struct obd_device       *obd = exp->exp_obd;
3147         struct lmv_obd          *lmv = &obd->u.lmv;
3148         struct lmv_tgt_desc     *tgt;
3149         ENTRY;
3150
3151         tgt = lmv_find_target(lmv, &och->och_fid);
3152         if (IS_ERR(tgt))
3153                 RETURN(PTR_ERR(tgt));
3154
3155         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3156 }
3157
3158 static int lmv_get_remote_perm(struct obd_export *exp, const struct lu_fid *fid,
3159                                u32 suppgid, struct ptlrpc_request **request)
3160 {
3161         struct obd_device       *obd = exp->exp_obd;
3162         struct lmv_obd          *lmv = &obd->u.lmv;
3163         struct lmv_tgt_desc     *tgt;
3164         int                      rc;
3165         ENTRY;
3166
3167         rc = lmv_check_connect(obd);
3168         if (rc)
3169                 RETURN(rc);
3170
3171         tgt = lmv_find_target(lmv, fid);
3172         if (IS_ERR(tgt))
3173                 RETURN(PTR_ERR(tgt));
3174
3175         rc = md_get_remote_perm(tgt->ltd_exp, fid, suppgid, request);
3176         RETURN(rc);
3177 }
3178
3179 int lmv_intent_getattr_async(struct obd_export *exp,
3180                              struct md_enqueue_info *minfo)
3181 {
3182         struct md_op_data       *op_data = &minfo->mi_data;
3183         struct obd_device       *obd = exp->exp_obd;
3184         struct lmv_obd          *lmv = &obd->u.lmv;
3185         struct lmv_tgt_desc     *ptgt = NULL;
3186         struct lmv_tgt_desc     *ctgt = NULL;
3187         int                      rc;
3188         ENTRY;
3189
3190         if (!fid_is_sane(&op_data->op_fid2))
3191                 RETURN(-EINVAL);
3192
3193         rc = lmv_check_connect(obd);
3194         if (rc)
3195                 RETURN(rc);
3196
3197         ptgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
3198         if (IS_ERR(ptgt))
3199                 RETURN(PTR_ERR(ptgt));
3200
3201         ctgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid2);
3202         if (IS_ERR(ctgt))
3203                 RETURN(PTR_ERR(ctgt));
3204
3205         /*
3206          * if child is on remote MDT, we need 2 async RPCs to fetch both LOOKUP
3207          * lock on parent, and UPDATE lock on child MDT, which makes all
3208          * complicated. Considering remote dir is rare case, and not supporting
3209          * it in statahead won't cause any issue, drop its support for now.
3210          */
3211         if (ptgt != ctgt)
3212                 RETURN(-ENOTSUPP);
3213
3214         rc = md_intent_getattr_async(ptgt->ltd_exp, minfo);
3215         RETURN(rc);
3216 }
3217
3218 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3219                         struct lu_fid *fid, __u64 *bits)
3220 {
3221         struct obd_device       *obd = exp->exp_obd;
3222         struct lmv_obd          *lmv = &obd->u.lmv;
3223         struct lmv_tgt_desc     *tgt;
3224         int                      rc;
3225         ENTRY;
3226
3227         rc = lmv_check_connect(obd);
3228         if (rc)
3229                 RETURN(rc);
3230
3231         tgt = lmv_find_target(lmv, fid);
3232         if (IS_ERR(tgt))
3233                 RETURN(PTR_ERR(tgt));
3234
3235         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3236         RETURN(rc);
3237 }
3238
3239 int lmv_get_fid_from_lsm(struct obd_export *exp,
3240                          const struct lmv_stripe_md *lsm,
3241                          const char *name, int namelen, struct lu_fid *fid)
3242 {
3243         const struct lmv_oinfo *oinfo;
3244
3245         LASSERT(lsm != NULL);
3246         oinfo = lsm_name_to_stripe_info(lsm, name, namelen);
3247         if (IS_ERR(oinfo))
3248                 return PTR_ERR(oinfo);
3249
3250         *fid = oinfo->lmo_fid;
3251
3252         RETURN(0);
3253 }
3254
3255 /**
3256  * For lmv, only need to send request to master MDT, and the master MDT will
3257  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3258  * we directly fetch data from the slave MDTs.
3259  */
3260 int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3261                  struct obd_quotactl *oqctl)
3262 {
3263         struct obd_device   *obd = class_exp2obd(exp);
3264         struct lmv_obd      *lmv = &obd->u.lmv;
3265         struct lmv_tgt_desc *tgt = lmv->tgts[0];
3266         int                  rc = 0;
3267         __u32                i;
3268         __u64                curspace, curinodes;
3269         ENTRY;
3270
3271         if (tgt == NULL ||
3272             tgt->ltd_exp == NULL ||
3273             !tgt->ltd_active ||
3274             lmv->desc.ld_tgt_count == 0) {
3275                 CERROR("master lmv inactive\n");
3276                 RETURN(-EIO);
3277         }
3278
3279         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3280                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3281                 RETURN(rc);
3282         }
3283
3284         curspace = curinodes = 0;
3285         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3286                 int err;
3287                 tgt = lmv->tgts[i];
3288
3289                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
3290                         continue;
3291
3292                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3293                 if (err) {
3294                         CERROR("getquota on mdt %d failed. %d\n", i, err);
3295                         if (!rc)
3296                                 rc = err;
3297                 } else {
3298                         curspace += oqctl->qc_dqblk.dqb_curspace;
3299                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3300                 }
3301         }
3302         oqctl->qc_dqblk.dqb_curspace = curspace;
3303         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3304
3305         RETURN(rc);
3306 }
3307
3308 static int lmv_merge_attr(struct obd_export *exp,
3309                           const struct lmv_stripe_md *lsm,
3310                           struct cl_attr *attr,
3311                           ldlm_blocking_callback cb_blocking)
3312 {
3313         int rc;
3314         int i;
3315
3316         rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3317         if (rc < 0)
3318                 return rc;
3319
3320         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3321                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3322
3323                 CDEBUG(D_INFO, ""DFID" size %llu, blocks %llu nlink %u,"
3324                        " atime %lu ctime %lu, mtime %lu.\n",
3325                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3326                        i_size_read(inode), (unsigned long long)inode->i_blocks,
3327                        inode->i_nlink, LTIME_S(inode->i_atime),
3328                        LTIME_S(inode->i_ctime), LTIME_S(inode->i_mtime));
3329
3330                 /* for slave stripe, it needs to subtract nlink for . and .. */
3331                 if (i != 0)
3332                         attr->cat_nlink += inode->i_nlink - 2;
3333                 else
3334                         attr->cat_nlink = inode->i_nlink;
3335
3336                 attr->cat_size += i_size_read(inode);
3337                 attr->cat_blocks += inode->i_blocks;
3338
3339                 if (attr->cat_atime < LTIME_S(inode->i_atime))
3340                         attr->cat_atime = LTIME_S(inode->i_atime);
3341
3342                 if (attr->cat_ctime < LTIME_S(inode->i_ctime))
3343                         attr->cat_ctime = LTIME_S(inode->i_ctime);
3344
3345                 if (attr->cat_mtime < LTIME_S(inode->i_mtime))
3346                         attr->cat_mtime = LTIME_S(inode->i_mtime);
3347         }
3348         return 0;
3349 }
3350
3351 struct obd_ops lmv_obd_ops = {
3352         .o_owner                = THIS_MODULE,
3353         .o_setup                = lmv_setup,
3354         .o_cleanup              = lmv_cleanup,
3355         .o_precleanup           = lmv_precleanup,
3356         .o_process_config       = lmv_process_config,
3357         .o_connect              = lmv_connect,
3358         .o_disconnect           = lmv_disconnect,
3359         .o_statfs               = lmv_statfs,
3360         .o_get_info             = lmv_get_info,
3361         .o_set_info_async       = lmv_set_info_async,
3362         .o_notify               = lmv_notify,
3363         .o_get_uuid             = lmv_get_uuid,
3364         .o_iocontrol            = lmv_iocontrol,
3365         .o_quotactl             = lmv_quotactl
3366 };
3367
3368 struct md_ops lmv_md_ops = {
3369         .m_getstatus            = lmv_getstatus,
3370         .m_null_inode           = lmv_null_inode,
3371         .m_find_cbdata          = lmv_find_cbdata,
3372         .m_close                = lmv_close,
3373         .m_create               = lmv_create,
3374         .m_enqueue              = lmv_enqueue,
3375         .m_getattr              = lmv_getattr,
3376         .m_getxattr             = lmv_getxattr,
3377         .m_getattr_name         = lmv_getattr_name,
3378         .m_intent_lock          = lmv_intent_lock,
3379         .m_link                 = lmv_link,
3380         .m_rename               = lmv_rename,
3381         .m_setattr              = lmv_setattr,
3382         .m_setxattr             = lmv_setxattr,
3383         .m_fsync                = lmv_fsync,
3384         .m_read_page            = lmv_read_page,
3385         .m_unlink               = lmv_unlink,
3386         .m_init_ea_size         = lmv_init_ea_size,
3387         .m_cancel_unused        = lmv_cancel_unused,
3388         .m_set_lock_data        = lmv_set_lock_data,
3389         .m_lock_match           = lmv_lock_match,
3390         .m_get_lustre_md        = lmv_get_lustre_md,
3391         .m_free_lustre_md       = lmv_free_lustre_md,
3392         .m_merge_attr           = lmv_merge_attr,
3393         .m_set_open_replay_data = lmv_set_open_replay_data,
3394         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3395         .m_get_remote_perm      = lmv_get_remote_perm,
3396         .m_intent_getattr_async = lmv_intent_getattr_async,
3397         .m_revalidate_lock      = lmv_revalidate_lock,
3398         .m_get_fid_from_lsm     = lmv_get_fid_from_lsm,
3399         .m_unpackmd             = lmv_unpackmd,
3400 };
3401
3402 static int __init lmv_init(void)
3403 {
3404         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true, NULL,
3405                                    LUSTRE_LMV_NAME, NULL);
3406 }
3407
3408 static void __exit lmv_exit(void)
3409 {
3410         class_unregister_type(LUSTRE_LMV_NAME);
3411 }
3412
3413 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3414 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
3415 MODULE_VERSION(LUSTRE_VERSION_STRING);
3416 MODULE_LICENSE("GPL");
3417
3418 module_init(lmv_init);
3419 module_exit(lmv_exit);