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