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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  */
  31
  32 #define DEBUG_SUBSYSTEM S_LMV
  33
  34 #include <linux/file.h>
  35 #include <linux/module.h>
  36 #include <linux/init.h>
  37 #include <linux/user_namespace.h>
  38 #include <linux/uidgid.h>
  39 #include <linux/slab.h>
  40 #include <linux/pagemap.h>
  41 #include <linux/mm.h>
  42 #include <linux/math64.h>
  43 #include <linux/seq_file.h>
  44 #include <linux/namei.h>
  45
  46 #include <obd_support.h>
  47 #include <lustre_lib.h>
  48 #include <lustre_net.h>
  49 #include <obd_class.h>
  50 #include <lustre_lmv.h>
  51 #include <lprocfs_status.h>
  52 #include <cl_object.h>
  53 #include <lustre_fid.h>
  54 #include <uapi/linux/lustre/lustre_ioctl.h>
  55 #include <lustre_ioctl_old.h>
  56 #include <lustre_kernelcomm.h>
  57 #include "lmv_internal.h"
  58
  59 static int lmv_check_connect(struct obd_device *obd);
  60 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data);
  61
  62 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
  63                          int activate)
  64 {
  65         if (tgt->ltd_active == activate)
  66                 return;
  67
  68         tgt->ltd_active = activate;
  69         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
  70                 (activate ? 1 : -1);
  71
  72         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
  73 }
  74
  75 /**
  76  * Error codes:
  77  *
  78  *  -EINVAL  : UUID can't be found in the LMV's target list
  79  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  80  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  81  */
  82 static int lmv_set_mdc_active(struct lmv_obd *lmv,
  83                               const struct obd_uuid *uuid,
  84                               int activate)
  85 {
  86         struct lu_tgt_desc *tgt = NULL;
  87         struct obd_device *obd;
  88         int rc = 0;
  89
  90         ENTRY;
  91
  92         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  93                         lmv, uuid->uuid, activate);
  94
  95         spin_lock(&lmv->lmv_lock);
  96         lmv_foreach_connected_tgt(lmv, tgt) {
  97                 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
  98                        tgt->ltd_index, tgt->ltd_uuid.uuid,
  99                        tgt->ltd_exp->exp_handle.h_cookie);
 100
 101                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 102                         break;
 103         }
 104
 105         if (!tgt)
 106                 GOTO(out_lmv_lock, rc = -EINVAL);
 107
 108         obd = class_exp2obd(tgt->ltd_exp);
 109         if (obd == NULL)
 110                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 111
 112         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 113                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 114                obd->obd_type->typ_name, tgt->ltd_index);
 115         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 116
 117         if (tgt->ltd_active == activate) {
 118                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 119                        activate ? "" : "in");
 120                 GOTO(out_lmv_lock, rc);
 121         }
 122
 123         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 124                activate ? "" : "in");
 125         lmv_activate_target(lmv, tgt, activate);
 126         EXIT;
 127
 128  out_lmv_lock:
 129         spin_unlock(&lmv->lmv_lock);
 130         return rc;
 131 }
 132
 133 static struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
 134 {
 135         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
 136         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
 137
 138         return tgt ? obd_get_uuid(tgt->ltd_exp) : NULL;
 139 }
 140
 141 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 142                       enum obd_notify_event ev)
 143 {
 144         struct obd_connect_data *conn_data;
 145         struct lmv_obd          *lmv = &obd->u.lmv;
 146         struct obd_uuid         *uuid;
 147         int                      rc = 0;
 148         ENTRY;
 149
 150         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 151                 CERROR("unexpected notification of %s %s!\n",
 152                        watched->obd_type->typ_name,
 153                        watched->obd_name);
 154                 RETURN(-EINVAL);
 155         }
 156
 157         uuid = &watched->u.cli.cl_target_uuid;
 158         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 159                 /*
 160                  * Set MDC as active before notifying the observer, so the
 161                  * observer can use the MDC normally.
 162                  */
 163                 rc = lmv_set_mdc_active(lmv, uuid,
 164                                         ev == OBD_NOTIFY_ACTIVE);
 165                 if (rc) {
 166                         CERROR("%sactivation of %s failed: %d\n",
 167                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 168                                uuid->uuid, rc);
 169                         RETURN(rc);
 170                 }
 171         } else if (ev == OBD_NOTIFY_OCD) {
 172                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 173                 /*
 174                  * XXX: Make sure that ocd_connect_flags from all targets are
 175                  * the same. Otherwise one of MDTs runs wrong version or
 176                  * something like this.  --umka
 177                  */
 178                 obd->obd_self_export->exp_connect_data = *conn_data;
 179         }
 180
 181         /*
 182          * Pass the notification up the chain.
 183          */
 184         if (obd->obd_observer)
 185                 rc = obd_notify(obd->obd_observer, watched, ev);
 186
 187         RETURN(rc);
 188 }
 189
 190 static int lmv_connect(const struct lu_env *env,
 191                        struct obd_export **pexp, struct obd_device *obd,
 192                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 193                        void *localdata)
 194 {
 195         struct lmv_obd *lmv = &obd->u.lmv;
 196         struct lustre_handle conn = { 0 };
 197         struct obd_export *exp;
 198         int rc;
 199         ENTRY;
 200
 201         rc = class_connect(&conn, obd, cluuid);
 202         if (rc) {
 203                 CERROR("class_connection() returned %d\n", rc);
 204                 RETURN(rc);
 205         }
 206
 207         exp = class_conn2export(&conn);
 208
 209         lmv->connected = 0;
 210         lmv->conn_data = *data;
 211         lmv->lmv_cache = localdata;
 212
 213         lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
 214                                                     &obd->obd_kset.kobj);
 215         if (!lmv->lmv_tgts_kobj) {
 216                 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
 217                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
 218         }
 219
 220         rc = lmv_check_connect(obd);
 221         if (rc != 0)
 222                 GOTO(out_sysfs, rc);
 223
 224         *pexp = exp;
 225
 226         RETURN(rc);
 227
 228 out_sysfs:
 229         if (lmv->lmv_tgts_kobj)
 230                 kobject_put(lmv->lmv_tgts_kobj);
 231
 232         class_disconnect(exp);
 233
 234         return rc;
 235 }
 236
 237 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
 238                             __u32 def_easize)
 239 {
 240         struct obd_device *obd = exp->exp_obd;
 241         struct lmv_obd *lmv = &obd->u.lmv;
 242         struct lmv_tgt_desc *tgt;
 243         int change = 0;
 244         int rc = 0;
 245
 246         ENTRY;
 247
 248         if (lmv->max_easize < easize) {
 249                 lmv->max_easize = easize;
 250                 change = 1;
 251         }
 252         if (lmv->max_def_easize < def_easize) {
 253                 lmv->max_def_easize = def_easize;
 254                 change = 1;
 255         }
 256
 257         if (change == 0)
 258                 RETURN(0);
 259
 260         if (lmv->connected == 0)
 261                 RETURN(0);
 262
 263         lmv_foreach_connected_tgt(lmv, tgt) {
 264                 if (!tgt->ltd_active)
 265                         continue;
 266
 267                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
 268                 if (rc) {
 269                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
 270                                " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
 271                         break;
 272                 }
 273         }
 274         RETURN(rc);
 275 }
 276
 277 #define MAX_STRING_SIZE 128
 278
 279 static int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 280 {
 281         struct lmv_obd *lmv = &obd->u.lmv;
 282         struct obd_device *mdc_obd;
 283         struct obd_export *mdc_exp;
 284         struct lu_fld_target target;
 285         int  rc;
 286         ENTRY;
 287
 288         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 289                                         &obd->obd_uuid);
 290         if (!mdc_obd) {
 291                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 292                 RETURN(-EINVAL);
 293         }
 294
 295         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
 296                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 297                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
 298
 299         if (!mdc_obd->obd_set_up) {
 300                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 301                 RETURN(-EINVAL);
 302         }
 303
 304         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
 305                          &lmv->conn_data, lmv->lmv_cache);
 306         if (rc) {
 307                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 308                 RETURN(rc);
 309         }
 310
 311         /*
 312          * Init fid sequence client for this mdc and add new fld target.
 313          */
 314         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
 315         if (rc)
 316                 RETURN(rc);
 317
 318         target.ft_srv = NULL;
 319         target.ft_exp = mdc_exp;
 320         target.ft_idx = tgt->ltd_index;
 321
 322         fld_client_add_target(&lmv->lmv_fld, &target);
 323
 324         rc = obd_register_observer(mdc_obd, obd);
 325         if (rc) {
 326                 obd_disconnect(mdc_exp);
 327                 CERROR("target %s register_observer error %d\n",
 328                        tgt->ltd_uuid.uuid, rc);
 329                 RETURN(rc);
 330         }
 331
 332         if (obd->obd_observer) {
 333                 /*
 334                  * Tell the observer about the new target.
 335                  */
 336                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 337                                 OBD_NOTIFY_ACTIVE);
 338                 if (rc) {
 339                         obd_disconnect(mdc_exp);
 340                         RETURN(rc);
 341                 }
 342         }
 343
 344         tgt->ltd_active = 1;
 345         tgt->ltd_exp = mdc_exp;
 346         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
 347
 348         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
 349
 350         rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
 351         if (rc) {
 352                 obd_disconnect(mdc_exp);
 353                 RETURN(rc);
 354         }
 355
 356         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 357                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 358                atomic_read(&obd->obd_refcount));
 359
 360         lmv_statfs_check_update(obd, tgt);
 361
 362         if (lmv->lmv_tgts_kobj)
 363                 /* Even if we failed to create the link, that's fine */
 364                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
 365                                        &mdc_obd->obd_kset.kobj,
 366                                        mdc_obd->obd_name);
 367         RETURN(0);
 368 }
 369
 370 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
 371 {
 372         LASSERT(tgt);
 373         ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
 374         OBD_FREE_PTR(tgt);
 375 }
 376
 377 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 378                            __u32 index, int gen)
 379 {
 380         struct obd_device *mdc_obd;
 381         struct lmv_obd *lmv = &obd->u.lmv;
 382         struct lmv_tgt_desc *tgt;
 383         struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
 384         int rc = 0;
 385
 386         ENTRY;
 387
 388         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
 389         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
 390                                         &obd->obd_uuid);
 391         if (!mdc_obd) {
 392                 CERROR("%s: Target %s not attached: rc = %d\n",
 393                        obd->obd_name, uuidp->uuid, -EINVAL);
 394                 RETURN(-EINVAL);
 395         }
 396
 397         OBD_ALLOC_PTR(tgt);
 398         if (!tgt)
 399                 RETURN(-ENOMEM);
 400
 401         mutex_init(&tgt->ltd_fid_mutex);
 402         tgt->ltd_index = index;
 403         tgt->ltd_uuid = *uuidp;
 404         tgt->ltd_active = 0;
 405
 406         mutex_lock(&ltd->ltd_mutex);
 407         rc = ltd_add_tgt(ltd, tgt);
 408         mutex_unlock(&ltd->ltd_mutex);
 409
 410         if (rc)
 411                 GOTO(out_tgt, rc);
 412
 413         if (!lmv->connected)
 414                 /* lmv_check_connect() will connect this target. */
 415                 RETURN(0);
 416
 417         rc = lmv_connect_mdc(obd, tgt);
 418         if (!rc) {
 419                 int easize = sizeof(struct lmv_stripe_md) +
 420                         lmv->lmv_mdt_count * sizeof(struct lu_fid);
 421
 422                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
 423         }
 424
 425         RETURN(rc);
 426
 427 out_tgt:
 428         OBD_FREE_PTR(tgt);
 429         return rc;
 430 }
 431
 432 static int lmv_check_connect(struct obd_device *obd)
 433 {
 434         struct lmv_obd *lmv = &obd->u.lmv;
 435         struct lmv_tgt_desc *tgt;
 436         int easize;
 437         int rc;
 438
 439         ENTRY;
 440
 441         if (lmv->connected)
 442                 RETURN(0);
 443
 444         mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
 445         if (lmv->connected)
 446                 GOTO(unlock, rc = 0);
 447
 448         if (!lmv->lmv_mdt_count) {
 449                 CERROR("%s: no targets configured: rc = -EINVAL\n",
 450                        obd->obd_name);
 451                 GOTO(unlock, rc = -EINVAL);
 452         }
 453
 454         if (!lmv_mdt0_inited(lmv)) {
 455                 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
 456                        obd->obd_name);
 457                 GOTO(unlock, rc = -EINVAL);
 458         }
 459
 460         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 461                obd->obd_uuid.uuid, obd->obd_name);
 462
 463         lmv_foreach_tgt(lmv, tgt) {
 464                 rc = lmv_connect_mdc(obd, tgt);
 465                 if (rc)
 466                         GOTO(out_disc, rc);
 467         }
 468
 469         lmv->connected = 1;
 470         easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
 471         lmv_init_ea_size(obd->obd_self_export, easize, 0);
 472         EXIT;
 473 unlock:
 474         mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
 475
 476         return rc;
 477
 478 out_disc:
 479         lmv_foreach_tgt(lmv, tgt) {
 480                 tgt->ltd_active = 0;
 481                 if (!tgt->ltd_exp)
 482                         continue;
 483
 484                 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
 485                 obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 486                 obd_disconnect(tgt->ltd_exp);
 487         }
 488
 489         goto unlock;
 490 }
 491
 492 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 493 {
 494         struct lmv_obd *lmv = &obd->u.lmv;
 495         struct obd_device *mdc_obd;
 496         int rc;
 497         ENTRY;
 498
 499         LASSERT(tgt != NULL);
 500         LASSERT(obd != NULL);
 501
 502         mdc_obd = class_exp2obd(tgt->ltd_exp);
 503
 504         if (mdc_obd) {
 505                 mdc_obd->obd_force = obd->obd_force;
 506                 mdc_obd->obd_fail = obd->obd_fail;
 507                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 508
 509                 if (lmv->lmv_tgts_kobj)
 510                         sysfs_remove_link(lmv->lmv_tgts_kobj,
 511                                           mdc_obd->obd_name);
 512         }
 513
 514         rc = lu_qos_del_tgt(&lmv->lmv_qos, tgt);
 515         if (rc)
 516                 CERROR("%s: Can't del target from QoS table: rc = %d\n",
 517                        tgt->ltd_exp->exp_obd->obd_name, rc);
 518
 519         rc = fld_client_del_target(&lmv->lmv_fld, tgt->ltd_index);
 520         if (rc)
 521                 CERROR("%s: Can't del fld targets: rc = %d\n",
 522                        tgt->ltd_exp->exp_obd->obd_name, rc);
 523
 524         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
 525         if (rc)
 526                 CERROR("%s: Can't finalize fids factory: rc = %d\n",
 527                        tgt->ltd_exp->exp_obd->obd_name, rc);
 528
 529         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 530                tgt->ltd_exp->exp_obd->obd_name,
 531                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 532
 533         lmv_activate_target(lmv, tgt, 0);
 534         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 535         rc = obd_disconnect(tgt->ltd_exp);
 536         if (rc) {
 537                 CERROR("%s: Target %s disconnect error: rc = %d\n",
 538                        tgt->ltd_exp->exp_obd->obd_name,
 539                        tgt->ltd_uuid.uuid, rc);
 540         }
 541         tgt->ltd_exp = NULL;
 542         RETURN(0);
 543 }
 544
 545 static int lmv_disconnect(struct obd_export *exp)
 546 {
 547         struct obd_device *obd = class_exp2obd(exp);
 548         struct lmv_obd *lmv = &obd->u.lmv;
 549         struct lmv_tgt_desc *tgt;
 550         int rc;
 551
 552         ENTRY;
 553
 554         lmv_foreach_connected_tgt(lmv, tgt)
 555                 lmv_disconnect_mdc(obd, tgt);
 556
 557         if (lmv->lmv_tgts_kobj)
 558                 kobject_put(lmv->lmv_tgts_kobj);
 559
 560         lmv->connected = 0;
 561         rc = class_disconnect(exp);
 562
 563         RETURN(rc);
 564 }
 565
 566 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
 567                         void __user *uarg)
 568 {
 569         struct obd_device *obd = class_exp2obd(exp);
 570         struct lmv_obd *lmv = &obd->u.lmv;
 571         struct getinfo_fid2path *gf;
 572         struct lmv_tgt_desc *tgt;
 573         struct getinfo_fid2path *remote_gf = NULL;
 574         struct lu_fid root_fid;
 575         int remote_gf_size = 0;
 576         int currentisenc = 0;
 577         int globalisenc = 0;
 578         int rc;
 579
 580         gf = karg;
 581         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 582         if (IS_ERR(tgt))
 583                 RETURN(PTR_ERR(tgt));
 584
 585         root_fid = *gf->gf_u.gf_root_fid;
 586         LASSERT(fid_is_sane(&root_fid));
 587
 588 repeat_fid2path:
 589         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
 590         if (rc != 0 && rc != -EREMOTE)
 591                 GOTO(out_fid2path, rc);
 592
 593         if (gf->gf_u.gf_path[0] == '/') {
 594                 /* by convention, server side (mdt_path_current()) puts
 595                  * a leading '/' to tell client that we are dealing with
 596                  * an encrypted file
 597                  */
 598                 currentisenc = 1;
 599                 globalisenc = 1;
 600         } else {
 601                 currentisenc = 0;
 602         }
 603
 604         /* If remote_gf != NULL, it means just building the
 605          * path on the remote MDT, copy this path segment to gf.
 606          */
 607         if (remote_gf != NULL) {
 608                 struct getinfo_fid2path *ori_gf;
 609                 int oldisenc = 0;
 610                 char *ptr;
 611                 int len;
 612
 613                 ori_gf = (struct getinfo_fid2path *)karg;
 614                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
 615                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
 616                         GOTO(out_fid2path, rc = -EOVERFLOW);
 617
 618                 ptr = ori_gf->gf_u.gf_path;
 619                 oldisenc = ptr[0] == '/';
 620
 621                 len = strlen(gf->gf_u.gf_path);
 622                 if (len) {
 623                         /* move the current path to the right to release space
 624                          * for closer-to-root part
 625                          */
 626                         memmove(ptr + len - currentisenc + 1 + globalisenc,
 627                                 ptr + oldisenc,
 628                                 strlen(ori_gf->gf_u.gf_path) - oldisenc + 1);
 629                         if (globalisenc)
 630                                 *(ptr++) = '/';
 631                         memcpy(ptr, gf->gf_u.gf_path + currentisenc,
 632                                len - currentisenc);
 633                         ptr[len - currentisenc] = '/';
 634                 }
 635         }
 636
 637         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
 638                tgt->ltd_exp->exp_obd->obd_name,
 639                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
 640                gf->gf_linkno);
 641
 642         if (rc == 0)
 643                 GOTO(out_fid2path, rc);
 644
 645         /* sigh, has to go to another MDT to do path building further */
 646         if (remote_gf == NULL) {
 647                 remote_gf_size = sizeof(*remote_gf) + len - sizeof(*gf);
 648                 OBD_ALLOC(remote_gf, remote_gf_size);
 649                 if (remote_gf == NULL)
 650                         GOTO(out_fid2path, rc = -ENOMEM);
 651                 remote_gf->gf_pathlen = len - sizeof(*gf);
 652         }
 653
 654         if (!fid_is_sane(&gf->gf_fid)) {
 655                 CERROR("%s: invalid FID "DFID": rc = %d\n",
 656                        tgt->ltd_exp->exp_obd->obd_name,
 657                        PFID(&gf->gf_fid), -EINVAL);
 658                 GOTO(out_fid2path, rc = -EINVAL);
 659         }
 660
 661         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 662         if (IS_ERR(tgt))
 663                 GOTO(out_fid2path, rc = -EINVAL);
 664
 665         remote_gf->gf_fid = gf->gf_fid;
 666         remote_gf->gf_recno = -1;
 667         remote_gf->gf_linkno = -1;
 668         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
 669         *remote_gf->gf_u.gf_root_fid = root_fid;
 670         gf = remote_gf;
 671         goto repeat_fid2path;
 672
 673 out_fid2path:
 674         if (remote_gf != NULL)
 675                 OBD_FREE(remote_gf, remote_gf_size);
 676         RETURN(rc);
 677 }
 678
 679 static int lmv_hsm_req_count(struct lmv_obd *lmv,
 680                              const struct hsm_user_request *hur,
 681                              const struct lmv_tgt_desc *tgt_mds)
 682 {
 683         struct lmv_tgt_desc *curr_tgt;
 684         __u32 i;
 685         int nr = 0;
 686
 687         /* count how many requests must be sent to the given target */
 688         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
 689                 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
 690                 if (IS_ERR(curr_tgt))
 691                         RETURN(PTR_ERR(curr_tgt));
 692                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
 693                         nr++;
 694         }
 695         return nr;
 696 }
 697
 698 static int lmv_hsm_req_build(struct lmv_obd *lmv,
 699                               struct hsm_user_request *hur_in,
 700                               const struct lmv_tgt_desc *tgt_mds,
 701                               struct hsm_user_request *hur_out)
 702 {
 703         __u32 i, nr_out;
 704         struct lmv_tgt_desc *curr_tgt;
 705
 706         /* build the hsm_user_request for the given target */
 707         hur_out->hur_request = hur_in->hur_request;
 708         nr_out = 0;
 709         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
 710                 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
 711                 if (IS_ERR(curr_tgt))
 712                         RETURN(PTR_ERR(curr_tgt));
 713                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
 714                         hur_out->hur_user_item[nr_out] =
 715                                                 hur_in->hur_user_item[i];
 716                         nr_out++;
 717                 }
 718         }
 719         hur_out->hur_request.hr_itemcount = nr_out;
 720         memcpy(hur_data(hur_out), hur_data(hur_in),
 721                hur_in->hur_request.hr_data_len);
 722
 723         RETURN(0);
 724 }
 725
 726 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
 727                                  int len, struct lustre_kernelcomm *lk,
 728                                  void __user *uarg)
 729 {
 730         struct lmv_obd *lmv = &obd->u.lmv;
 731         struct lu_tgt_desc *tgt;
 732         int rc;
 733
 734         ENTRY;
 735
 736         /* unregister request (call from llapi_hsm_copytool_fini) */
 737         lmv_foreach_connected_tgt(lmv, tgt)
 738                 /* best effort: try to clean as much as possible
 739                  * (continue on error) */
 740                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 741
 742         /* Whatever the result, remove copytool from kuc groups.
 743          * Unreached coordinators will get EPIPE on next requests
 744          * and will unregister automatically.
 745          */
 746         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 747
 748         RETURN(rc);
 749 }
 750
 751 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
 752                                int len, struct lustre_kernelcomm *lk,
 753                                void __user *uarg)
 754 {
 755         struct lmv_obd *lmv = &obd->u.lmv;
 756         struct file *filp;
 757         bool any_set = false;
 758         struct kkuc_ct_data *kcd;
 759         size_t kcd_size;
 760         struct lu_tgt_desc *tgt;
 761         __u32 i;
 762         int err;
 763         int rc = 0;
 764
 765         ENTRY;
 766
 767         filp = fget(lk->lk_wfd);
 768         if (!filp)
 769                 RETURN(-EBADF);
 770
 771         if (lk->lk_flags & LK_FLG_DATANR)
 772                 kcd_size = offsetof(struct kkuc_ct_data,
 773                                     kcd_archives[lk->lk_data_count]);
 774         else
 775                 kcd_size = sizeof(*kcd);
 776
 777         OBD_ALLOC(kcd, kcd_size);
 778         if (kcd == NULL)
 779                 GOTO(err_fput, rc = -ENOMEM);
 780
 781         kcd->kcd_nr_archives = lk->lk_data_count;
 782         if (lk->lk_flags & LK_FLG_DATANR) {
 783                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
 784                 if (lk->lk_data_count > 0)
 785                         memcpy(kcd->kcd_archives, lk->lk_data,
 786                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
 787         } else {
 788                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
 789         }
 790
 791         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
 792                                    lk->lk_group, kcd, kcd_size);
 793         OBD_FREE(kcd, kcd_size);
 794         if (rc)
 795                 GOTO(err_fput, rc);
 796
 797         /* All or nothing: try to register to all MDS.
 798          * In case of failure, unregister from previous MDS,
 799          * except if it because of inactive target. */
 800         lmv_foreach_connected_tgt(lmv, tgt) {
 801                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 802                 if (err) {
 803                         if (tgt->ltd_active) {
 804                                 /* permanent error */
 805                                 CERROR("%s: iocontrol MDC %s on MDT"
 806                                        " idx %d cmd %x: err = %d\n",
 807                                        lmv2obd_dev(lmv)->obd_name,
 808                                        tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
 809                                        err);
 810                                 rc = err;
 811                                 lk->lk_flags |= LK_FLG_STOP;
 812                                 i = tgt->ltd_index;
 813                                 /* unregister from previous MDS */
 814                                 lmv_foreach_connected_tgt(lmv, tgt) {
 815                                         if (tgt->ltd_index >= i)
 816                                                 break;
 817
 818                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
 819                                                       lk, uarg);
 820                                 }
 821                                 GOTO(err_kkuc_rem, rc);
 822                         }
 823                         /* else: transient error.
 824                          * kuc will register to the missing MDT
 825                          * when it is back */
 826                 } else {
 827                         any_set = true;
 828                 }
 829         }
 830
 831         if (!any_set)
 832                 /* no registration done: return error */
 833                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
 834
 835         RETURN(0);
 836
 837 err_kkuc_rem:
 838         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 839
 840 err_fput:
 841         fput(filp);
 842         return rc;
 843 }
 844
 845 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 846                          int len, void *karg, void __user *uarg)
 847 {
 848         struct obd_device *obd = class_exp2obd(exp);
 849         struct lmv_obd *lmv = &obd->u.lmv;
 850         struct lu_tgt_desc *tgt = NULL;
 851         int set = 0;
 852         __u32 count = lmv->lmv_mdt_count;
 853         int rc = 0;
 854
 855         ENTRY;
 856         CDEBUG(D_IOCTL, "%s: cmd=%x len=%u karg=%pK uarg=%pK\n",
 857                exp->exp_obd->obd_name, cmd, len, karg, uarg);
 858         if (count == 0)
 859                 RETURN(-ENOTTY);
 860
 861         /* exit early for unknown ioctl types */
 862         if (unlikely(_IOC_TYPE(cmd) != 'f' && !IOC_OSC_SET_ACTIVE_ALLOW(cmd)))
 863                 RETURN(OBD_IOC_ERROR(obd->obd_name, cmd, "unknown", -ENOTTY));
 864
 865         /* handle commands that don't use @karg first */
 866         switch (cmd) {
 867         case LL_IOC_GET_CONNECT_FLAGS:
 868                 tgt = lmv_tgt(lmv, 0);
 869                 rc = -ENODATA;
 870                 if (tgt && tgt->ltd_exp)
 871                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, NULL, uarg);
 872                 RETURN(rc);
 873         }
 874
 875         if (unlikely(karg == NULL))
 876                 RETURN(OBD_IOC_ERROR(obd->obd_name, cmd, "karg=NULL", -EINVAL));
 877
 878         switch (cmd) {
 879         case IOC_OBD_STATFS: {
 880                 struct obd_ioctl_data *data = karg;
 881                 struct obd_device *mdc_obd;
 882                 struct obd_statfs stat_buf = {0};
 883                 __u32 index;
 884
 885                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 886
 887                 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
 888                         RETURN(-ENODEV);
 889
 890                 tgt = lmv_tgt(lmv, index);
 891                 if (!tgt)
 892                         RETURN(-EAGAIN);
 893
 894                 if (!tgt->ltd_active)
 895                         RETURN(-ENODATA);
 896
 897                 mdc_obd = class_exp2obd(tgt->ltd_exp);
 898                 if (!mdc_obd)
 899                         RETURN(-EINVAL);
 900
 901                 /* copy UUID */
 902                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 903                                  min((int) data->ioc_plen2,
 904                                      (int) sizeof(struct obd_uuid))))
 905                         RETURN(-EFAULT);
 906
 907                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
 908                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
 909                                 0);
 910                 if (rc)
 911                         RETURN(rc);
 912                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
 913                                  min((int) data->ioc_plen1,
 914                                      (int) sizeof(stat_buf))))
 915                         RETURN(-EFAULT);
 916                 break;
 917         }
 918         case OBD_IOC_QUOTACTL: {
 919                 struct if_quotactl *qctl = karg;
 920                 struct obd_quotactl *oqctl;
 921                 struct obd_import *imp;
 922
 923                 if (qctl->qc_valid == QC_MDTIDX) {
 924                         tgt = lmv_tgt(lmv, qctl->qc_idx);
 925                 } else if (qctl->qc_valid == QC_UUID) {
 926                         lmv_foreach_tgt(lmv, tgt) {
 927                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 928                                                      &qctl->obd_uuid))
 929                                         continue;
 930
 931                                 if (!tgt->ltd_exp)
 932                                         RETURN(-EINVAL);
 933
 934                                 break;
 935                         }
 936                 } else {
 937                         RETURN(-EINVAL);
 938                 }
 939
 940                 if (!tgt)
 941                         RETURN(-ENODEV);
 942
 943                 if (!tgt->ltd_exp)
 944                         RETURN(-EINVAL);
 945
 946                 imp = class_exp2cliimp(tgt->ltd_exp);
 947                 if (!tgt->ltd_active && imp->imp_state != LUSTRE_IMP_IDLE) {
 948                         qctl->qc_valid = QC_MDTIDX;
 949                         qctl->obd_uuid = tgt->ltd_uuid;
 950                         RETURN(-ENODATA);
 951                 }
 952
 953                 OBD_ALLOC_PTR(oqctl);
 954                 if (!oqctl)
 955                         RETURN(-ENOMEM);
 956
 957                 QCTL_COPY(oqctl, qctl);
 958                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 959                 if (rc == 0) {
 960                         QCTL_COPY(qctl, oqctl);
 961                         qctl->qc_valid = QC_MDTIDX;
 962                         qctl->obd_uuid = tgt->ltd_uuid;
 963                 }
 964                 OBD_FREE_PTR(oqctl);
 965                 break;
 966         }
 967         case LL_IOC_FID2MDTIDX: {
 968                 struct lu_fid *fid = karg;
 969                 int             mdt_index;
 970
 971                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
 972                 if (rc != 0)
 973                         RETURN(rc);
 974
 975                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
 976                  * point to user space memory for FID2MDTIDX. */
 977                 *(__u32 *)uarg = mdt_index;
 978                 break;
 979         }
 980         case OBD_IOC_FID2PATH: {
 981                 rc = lmv_fid2path(exp, len, karg, uarg);
 982                 break;
 983         }
 984         case LL_IOC_HSM_STATE_GET:
 985         case LL_IOC_HSM_STATE_SET:
 986         case LL_IOC_HSM_ACTION: {
 987                 struct md_op_data *op_data = karg;
 988
 989                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
 990                 if (IS_ERR(tgt))
 991                         RETURN(PTR_ERR(tgt));
 992
 993                 if (tgt->ltd_exp == NULL)
 994                         RETURN(-EINVAL);
 995
 996                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 997                 break;
 998         }
 999         case LL_IOC_HSM_PROGRESS: {
1000                 const struct hsm_progress_kernel *hpk = karg;
1001
1002                 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
1003                 if (IS_ERR(tgt))
1004                         RETURN(PTR_ERR(tgt));
1005                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1006                 break;
1007         }
1008         case LL_IOC_HSM_REQUEST: {
1009                 struct hsm_user_request *hur = karg;
1010                 unsigned int reqcount = hur->hur_request.hr_itemcount;
1011
1012                 if (reqcount == 0)
1013                         RETURN(0);
1014
1015                 /* if the request is about a single fid
1016                  * or if there is a single MDS, no need to split
1017                  * the request. */
1018                 if (reqcount == 1 || count == 1) {
1019                         tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
1020                         if (IS_ERR(tgt))
1021                                 RETURN(PTR_ERR(tgt));
1022                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1023                 } else {
1024                         /* split fid list to their respective MDS */
1025                         lmv_foreach_connected_tgt(lmv, tgt) {
1026                                 int nr, rc1;
1027                                 size_t reqlen;
1028                                 struct hsm_user_request *req;
1029
1030                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
1031                                 if (nr < 0)
1032                                         RETURN(nr);
1033                                 if (nr == 0) /* nothing for this MDS */
1034                                         continue;
1035
1036                                 /* build a request with fids for this MDS */
1037                                 reqlen = offsetof(typeof(*hur),
1038                                                   hur_user_item[nr])
1039                                                 + hur->hur_request.hr_data_len;
1040                                 OBD_ALLOC_LARGE(req, reqlen);
1041                                 if (req == NULL)
1042                                         RETURN(-ENOMEM);
1043                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
1044                                 if (rc1 < 0)
1045                                         GOTO(hsm_req_err, rc1);
1046                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
1047                                                     req, uarg);
1048 hsm_req_err:
1049                                 if (rc1 != 0 && rc == 0)
1050                                         rc = rc1;
1051                                 OBD_FREE_LARGE(req, reqlen);
1052                         }
1053                 }
1054                 break;
1055         }
1056         case LL_IOC_LOV_SWAP_LAYOUTS: {
1057                 struct md_op_data *op_data = karg;
1058                 struct lmv_tgt_desc *tgt1, *tgt2;
1059
1060                 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1061                 if (IS_ERR(tgt1))
1062                         RETURN(PTR_ERR(tgt1));
1063
1064                 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1065                 if (IS_ERR(tgt2))
1066                         RETURN(PTR_ERR(tgt2));
1067
1068                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1069                         RETURN(-EINVAL);
1070
1071                 /* only files on same MDT can have their layouts swapped */
1072                 if (tgt1->ltd_index != tgt2->ltd_index)
1073                         RETURN(-EPERM);
1074
1075                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1076                 break;
1077         }
1078         case LL_IOC_HSM_CT_START: {
1079                 struct lustre_kernelcomm *lk = karg;
1080
1081                 if (lk->lk_flags & LK_FLG_STOP)
1082                         rc = lmv_hsm_ct_unregister(obd, cmd, len, lk, uarg);
1083                 else
1084                         rc = lmv_hsm_ct_register(obd, cmd, len, lk, uarg);
1085                 break;
1086         }
1087         default:
1088                 lmv_foreach_connected_tgt(lmv, tgt) {
1089                         struct obd_device *mdc_obd;
1090                         int err;
1091
1092                         /* ll_umount_begin() sets force flag but for lmv, not
1093                          * mdc. Let's pass it through */
1094                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1095                         mdc_obd->obd_force = obd->obd_force;
1096                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1097                         if (err) {
1098                                 if (tgt->ltd_active) {
1099                                         OBD_IOC_ERROR(obd->obd_name, cmd,
1100                                                       tgt->ltd_uuid.uuid, err);
1101                                         if (!rc)
1102                                                 rc = err;
1103                                         if (unlikely(err == -ENOTTY))
1104                                                 break;
1105                                 }
1106                         } else {
1107                                 set = 1;
1108                         }
1109                 }
1110                 if (!set && !rc)
1111                         rc = -EIO;
1112                 break;
1113         }
1114         RETURN(rc);
1115 }
1116
1117 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1118                   struct lu_fid *fid, struct md_op_data *op_data)
1119 {
1120         struct obd_device *obd = class_exp2obd(exp);
1121         struct lmv_obd *lmv = &obd->u.lmv;
1122         struct lmv_tgt_desc *tgt;
1123         int rc;
1124
1125         ENTRY;
1126
1127         LASSERT(op_data);
1128         LASSERT(fid);
1129
1130         tgt = lmv_tgt(lmv, op_data->op_mds);
1131         if (!tgt)
1132                 RETURN(-ENODEV);
1133
1134         if (!tgt->ltd_active || !tgt->ltd_exp)
1135                 RETURN(-ENODEV);
1136
1137         /*
1138          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1139          * on server that seq in new allocated fid is not yet known.
1140          */
1141         mutex_lock(&tgt->ltd_fid_mutex);
1142         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1143         mutex_unlock(&tgt->ltd_fid_mutex);
1144         if (rc > 0) {
1145                 LASSERT(fid_is_sane(fid));
1146                 rc = 0;
1147         }
1148
1149         RETURN(rc);
1150 }
1151
1152 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1153 {
1154         struct lmv_obd *lmv = &obd->u.lmv;
1155         struct lmv_desc *desc;
1156         struct lnet_processid lnet_id;
1157         int i = 0;
1158         int rc;
1159
1160         ENTRY;
1161
1162         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1163                 CERROR("LMV setup requires a descriptor\n");
1164                 RETURN(-EINVAL);
1165         }
1166
1167         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1168         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1169                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1170                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1171                 RETURN(-EINVAL);
1172         }
1173
1174         obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1175                      desc->ld_uuid.uuid);
1176         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1177         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1178         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1179                 LMV_DESC_QOS_MAXAGE_DEFAULT;
1180         lmv->max_def_easize = 0;
1181         lmv->max_easize = 0;
1182
1183         spin_lock_init(&lmv->lmv_lock);
1184
1185         /*
1186          * initialize rr_index to lower 32bit of netid, so that client
1187          * can distribute subdirs evenly from the beginning.
1188          */
1189         while (LNetGetId(i++, &lnet_id, false) != -ENOENT) {
1190                 if (!nid_is_lo0(&lnet_id.nid)) {
1191                         lmv->lmv_qos_rr_index = ntohl(lnet_id.nid.nid_addr[0]);
1192                         break;
1193                 }
1194         }
1195
1196         rc = lmv_tunables_init(obd);
1197         if (rc)
1198                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1199                       obd->obd_name, rc);
1200
1201         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1202                              LUSTRE_CLI_FLD_HASH_DHT);
1203         if (rc)
1204                 CERROR("Can't init FLD, err %d\n", rc);
1205
1206         rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1207         if (rc)
1208                 CWARN("%s: error initialize target table: rc = %d\n",
1209                       obd->obd_name, rc);
1210
1211         RETURN(rc);
1212 }
1213
1214 static int lmv_cleanup(struct obd_device *obd)
1215 {
1216         struct lmv_obd *lmv = &obd->u.lmv;
1217         struct lu_tgt_desc *tgt;
1218         struct lu_tgt_desc *tmp;
1219
1220         ENTRY;
1221
1222         fld_client_fini(&lmv->lmv_fld);
1223         fld_client_debugfs_fini(&lmv->lmv_fld);
1224
1225         lprocfs_obd_cleanup(obd);
1226         lprocfs_free_md_stats(obd);
1227
1228         lmv_foreach_tgt_safe(lmv, tgt, tmp)
1229                 lmv_del_target(lmv, tgt);
1230         lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1231
1232         RETURN(0);
1233 }
1234
1235 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1236 {
1237         struct lustre_cfg       *lcfg = buf;
1238         struct obd_uuid         obd_uuid;
1239         int                     gen;
1240         __u32                   index;
1241         int                     rc;
1242         ENTRY;
1243
1244         switch (lcfg->lcfg_command) {
1245         case LCFG_ADD_MDC:
1246                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1247                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1248                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1249                         GOTO(out, rc = -EINVAL);
1250
1251                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1252
1253                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1254                         GOTO(out, rc = -EINVAL);
1255                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1256                         GOTO(out, rc = -EINVAL);
1257                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1258                 GOTO(out, rc);
1259         default:
1260                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1261                 GOTO(out, rc = -EINVAL);
1262         }
1263 out:
1264         RETURN(rc);
1265 }
1266
1267 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1268 {
1269         int i;
1270
1271         if (flags & OBD_STATFS_FOR_MDT0)
1272                 return 0;
1273
1274         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1275                 return lmv->lmv_statfs_start;
1276
1277         /* choose initial MDT for this client */
1278         for (i = 0;; i++) {
1279                 struct lnet_processid lnet_id;
1280                 if (LNetGetId(i, &lnet_id, false) == -ENOENT)
1281                         break;
1282
1283                 if (!nid_is_lo0(&lnet_id.nid)) {
1284                         /* We dont need a full 64-bit modulus, just enough
1285                          * to distribute the requests across MDTs evenly.
1286                          */
1287                         lmv->lmv_statfs_start = nidhash(&lnet_id.nid) %
1288                                                 lmv->lmv_mdt_count;
1289                         break;
1290                 }
1291         }
1292
1293         return lmv->lmv_statfs_start;
1294 }
1295
1296 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1297                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1298 {
1299         struct obd_device *obd = class_exp2obd(exp);
1300         struct lmv_obd *lmv = &obd->u.lmv;
1301         struct obd_statfs *temp;
1302         struct lu_tgt_desc *tgt;
1303         __u32 i;
1304         __u32 idx;
1305         int rc = 0;
1306         int err = 0;
1307
1308         ENTRY;
1309
1310         OBD_ALLOC(temp, sizeof(*temp));
1311         if (temp == NULL)
1312                 RETURN(-ENOMEM);
1313
1314         /* distribute statfs among MDTs */
1315         idx = lmv_select_statfs_mdt(lmv, flags);
1316
1317         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1318                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1319                 tgt = lmv_tgt(lmv, idx);
1320                 if (!tgt || !tgt->ltd_exp)
1321                         continue;
1322
1323                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age,
1324                                 flags | OBD_STATFS_NESTED);
1325                 if (rc) {
1326                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1327                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1328                         err = rc;
1329                         /* Try another MDT */
1330                         if (flags & OBD_STATFS_SUM)
1331                                 continue;
1332                         GOTO(out_free_temp, rc);
1333                 }
1334
1335                 if (temp->os_state & OS_STATFS_SUM ||
1336                     flags == OBD_STATFS_FOR_MDT0) {
1337                         /* reset to the last aggregated values
1338                          * and don't sum with non-aggrated data */
1339                         /* If the statfs is from mount, it needs to retrieve
1340                          * necessary information from MDT0. i.e. mount does
1341                          * not need the merged osfs from all of MDT. Also
1342                          * clients can be mounted as long as MDT0 is in
1343                          * service */
1344                         *osfs = *temp;
1345                         GOTO(out_free_temp, rc);
1346                 }
1347
1348                 if (i == 0) {
1349                         *osfs = *temp;
1350                 } else {
1351                         osfs->os_bavail += temp->os_bavail;
1352                         osfs->os_blocks += temp->os_blocks;
1353                         osfs->os_ffree += temp->os_ffree;
1354                         osfs->os_files += temp->os_files;
1355                         osfs->os_granted += temp->os_granted;
1356                 }
1357         }
1358         /* There is no stats from some MDTs, data incomplete */
1359         if (err)
1360                 rc = err;
1361 out_free_temp:
1362         OBD_FREE(temp, sizeof(*temp));
1363         RETURN(rc);
1364 }
1365
1366 static int lmv_statfs_update(void *cookie, int rc)
1367 {
1368         struct obd_info *oinfo = cookie;
1369         struct obd_device *obd = oinfo->oi_obd;
1370         struct lmv_obd *lmv = &obd->u.lmv;
1371         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1372         struct obd_statfs *osfs = oinfo->oi_osfs;
1373
1374         /*
1375          * NB: don't deactivate TGT upon error, because we may not trigger async
1376          * statfs any longer, then there is no chance to activate TGT.
1377          */
1378         if (!rc) {
1379                 spin_lock(&lmv->lmv_lock);
1380                 tgt->ltd_statfs = *osfs;
1381                 tgt->ltd_statfs_age = ktime_get_seconds();
1382                 spin_unlock(&lmv->lmv_lock);
1383                 set_bit(LQ_DIRTY, &lmv->lmv_qos.lq_flags);
1384         }
1385
1386         return rc;
1387 }
1388
1389 /* update tgt statfs async if it's ld_qos_maxage old */
1390 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1391 {
1392         struct obd_info oinfo = {
1393                 .oi_obd = obd,
1394                 .oi_tgt = tgt,
1395                 .oi_cb_up = lmv_statfs_update,
1396         };
1397         int rc;
1398
1399         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1400             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1401                 return 0;
1402
1403         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1404
1405         return rc;
1406 }
1407
1408 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1409                         struct lu_fid *fid)
1410 {
1411         struct obd_device *obd = exp->exp_obd;
1412         struct lmv_obd *lmv = &obd->u.lmv;
1413         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1414         int rc;
1415
1416         ENTRY;
1417
1418         if (!tgt)
1419                 RETURN(-ENODEV);
1420
1421         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1422         RETURN(rc);
1423 }
1424
1425 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1426                         u64 obd_md_valid, const char *name, size_t buf_size,
1427                         struct ptlrpc_request **req)
1428 {
1429         struct obd_device *obd = exp->exp_obd;
1430         struct lmv_obd *lmv = &obd->u.lmv;
1431         struct lmv_tgt_desc *tgt;
1432         int rc;
1433
1434         ENTRY;
1435
1436         tgt = lmv_fid2tgt(lmv, fid);
1437         if (IS_ERR(tgt))
1438                 RETURN(PTR_ERR(tgt));
1439
1440         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1441
1442         RETURN(rc);
1443 }
1444
1445 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1446                         u64 obd_md_valid, const char *name,
1447                         const void *value, size_t value_size,
1448                         unsigned int xattr_flags, u32 suppgid,
1449                         struct ptlrpc_request **req)
1450 {
1451         struct obd_device *obd = exp->exp_obd;
1452         struct lmv_obd *lmv = &obd->u.lmv;
1453         struct lmv_tgt_desc *tgt;
1454         int rc;
1455
1456         ENTRY;
1457
1458         tgt = lmv_fid2tgt(lmv, fid);
1459         if (IS_ERR(tgt))
1460                 RETURN(PTR_ERR(tgt));
1461
1462         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1463                          value, value_size, xattr_flags, suppgid, req);
1464
1465         RETURN(rc);
1466 }
1467
1468 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1469                        struct ptlrpc_request **request)
1470 {
1471         struct obd_device *obd = exp->exp_obd;
1472         struct lmv_obd *lmv = &obd->u.lmv;
1473         struct lmv_tgt_desc *tgt;
1474         int rc;
1475
1476         ENTRY;
1477
1478         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1479         if (IS_ERR(tgt))
1480                 RETURN(PTR_ERR(tgt));
1481
1482         if (op_data->op_flags & MF_GET_MDT_IDX) {
1483                 op_data->op_mds = tgt->ltd_index;
1484                 RETURN(0);
1485         }
1486
1487         rc = md_getattr(tgt->ltd_exp, op_data, request);
1488
1489         RETURN(rc);
1490 }
1491
1492 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1493 {
1494         struct obd_device *obd = exp->exp_obd;
1495         struct lmv_obd *lmv = &obd->u.lmv;
1496         struct lu_tgt_desc *tgt;
1497
1498         ENTRY;
1499
1500         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1501
1502         /*
1503          * With DNE every object can have two locks in different namespaces:
1504          * lookup lock in space of MDT storing direntry and update/open lock in
1505          * space of MDT storing inode.
1506          */
1507         lmv_foreach_connected_tgt(lmv, tgt)
1508                 md_null_inode(tgt->ltd_exp, fid);
1509
1510         RETURN(0);
1511 }
1512
1513 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1514                      struct md_open_data *mod, struct ptlrpc_request **request)
1515 {
1516         struct obd_device *obd = exp->exp_obd;
1517         struct lmv_obd *lmv = &obd->u.lmv;
1518         struct lmv_tgt_desc *tgt;
1519         int rc;
1520
1521         ENTRY;
1522
1523         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1524         if (IS_ERR(tgt))
1525                 RETURN(PTR_ERR(tgt));
1526
1527         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1528         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1529         RETURN(rc);
1530 }
1531
1532 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv,
1533                                               struct md_op_data *op_data)
1534 {
1535         struct lu_tgt_desc *tgt, *cur = NULL;
1536         __u64 total_avail = 0;
1537         __u64 total_weight = 0;
1538         __u64 cur_weight = 0;
1539         int total_usable = 0;
1540         __u64 rand;
1541         int rc;
1542
1543         ENTRY;
1544
1545         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1546                 RETURN(ERR_PTR(-EAGAIN));
1547
1548         down_write(&lmv->lmv_qos.lq_rw_sem);
1549
1550         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1551                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1552
1553         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1554         if (rc)
1555                 GOTO(unlock, tgt = ERR_PTR(rc));
1556
1557         lmv_foreach_tgt(lmv, tgt) {
1558                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1559                         tgt->ltd_qos.ltq_usable = 0;
1560                         continue;
1561                 }
1562
1563                 tgt->ltd_qos.ltq_usable = 1;
1564                 lu_tgt_qos_weight_calc(tgt, true);
1565                 if (tgt->ltd_index == op_data->op_mds)
1566                         cur = tgt;
1567                 total_avail += tgt->ltd_qos.ltq_avail;
1568                 total_weight += tgt->ltd_qos.ltq_weight;
1569                 total_usable++;
1570         }
1571
1572         /* If current MDT has above-average space and dir is not aleady using
1573          * round-robin to spread across more MDTs, stay on the parent MDT
1574          * to avoid creating needless remote MDT directories.  Remote dirs
1575          * close to the root balance space more effectively than bottom dirs,
1576          * so prefer to create remote dirs at top level of directory tree.
1577          * "16 / (dir_depth + 10)" is the factor to make it less likely
1578          * for top-level directories to stay local unless they have more than
1579          * average free space, while deep dirs prefer local until more full.
1580          *    depth=0 -> 160%, depth=3 -> 123%, depth=6 -> 100%,
1581          *    depth=9 -> 84%, depth=12 -> 73%, depth=15 -> 64%
1582          */
1583         if (!lmv_op_default_rr_mkdir(op_data)) {
1584                 rand = total_avail * 16 /
1585                         (total_usable * (op_data->op_dir_depth + 10));
1586                 if (cur && cur->ltd_qos.ltq_avail >= rand) {
1587                         tgt = cur;
1588                         GOTO(unlock, tgt);
1589                 }
1590         }
1591
1592         rand = lu_prandom_u64_max(total_weight);
1593
1594         lmv_foreach_connected_tgt(lmv, tgt) {
1595                 if (!tgt->ltd_qos.ltq_usable)
1596                         continue;
1597
1598                 cur_weight += tgt->ltd_qos.ltq_weight;
1599                 if (cur_weight < rand)
1600                         continue;
1601
1602                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1603                 GOTO(unlock, tgt);
1604         }
1605
1606         /* no proper target found */
1607         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1608 unlock:
1609         up_write(&lmv->lmv_qos.lq_rw_sem);
1610
1611         return tgt;
1612 }
1613
1614 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv)
1615 {
1616         struct lu_tgt_desc *tgt;
1617         int i;
1618         int index;
1619
1620         ENTRY;
1621
1622         spin_lock(&lmv->lmv_lock);
1623         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1624                 index = (i + lmv->lmv_qos_rr_index) %
1625                         lmv->lmv_mdt_descs.ltd_tgts_size;
1626                 tgt = lmv_tgt(lmv, index);
1627                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1628                         continue;
1629
1630                 lmv->lmv_qos_rr_index = (tgt->ltd_index + 1) %
1631                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1632                 spin_unlock(&lmv->lmv_lock);
1633
1634                 RETURN(tgt);
1635         }
1636         spin_unlock(&lmv->lmv_lock);
1637
1638         RETURN(ERR_PTR(-ENODEV));
1639 }
1640
1641 /* locate MDT which is less full (avoid the most full MDT) */
1642 static struct lu_tgt_desc *lmv_locate_tgt_lf(struct lmv_obd *lmv)
1643 {
1644         struct lu_tgt_desc *min = NULL;
1645         struct lu_tgt_desc *tgt;
1646         __u64 avail = 0;
1647         __u64 rand;
1648
1649         ENTRY;
1650
1651         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1652                 RETURN(ERR_PTR(-EAGAIN));
1653
1654         down_write(&lmv->lmv_qos.lq_rw_sem);
1655
1656         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1657                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1658
1659         lmv_foreach_tgt(lmv, tgt) {
1660                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1661                         tgt->ltd_qos.ltq_usable = 0;
1662                         continue;
1663                 }
1664
1665                 tgt->ltd_qos.ltq_usable = 1;
1666                 lu_tgt_qos_weight_calc(tgt, true);
1667                 avail += tgt->ltd_qos.ltq_avail;
1668                 if (!min || min->ltd_qos.ltq_avail > tgt->ltd_qos.ltq_avail)
1669                         min = tgt;
1670         }
1671
1672         /* avoid the most full MDT */
1673         if (min)
1674                 avail -= min->ltd_qos.ltq_avail;
1675
1676         rand = lu_prandom_u64_max(avail);
1677         avail = 0;
1678         lmv_foreach_connected_tgt(lmv, tgt) {
1679                 if (!tgt->ltd_qos.ltq_usable)
1680                         continue;
1681
1682                 if (tgt == min)
1683                         continue;
1684
1685                 avail += tgt->ltd_qos.ltq_avail;
1686                 if (avail < rand)
1687                         continue;
1688
1689                 GOTO(unlock, tgt);
1690         }
1691
1692         /* no proper target found */
1693         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1694 unlock:
1695         up_write(&lmv->lmv_qos.lq_rw_sem);
1696
1697         RETURN(tgt);
1698 }
1699
1700 /* locate MDT by file name, for striped directory, the file name hash decides
1701  * which stripe its dirent is stored.
1702  */
1703 static struct lmv_tgt_desc *
1704 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_object *lso,
1705                        const char *name, int namelen, struct lu_fid *fid,
1706                        __u32 *mds, bool new_layout)
1707 {
1708         struct lmv_tgt_desc *tgt;
1709         const struct lmv_oinfo *oinfo;
1710
1711         if (!lmv_dir_striped(lso) || !namelen) {
1712                 tgt = lmv_fid2tgt(lmv, fid);
1713                 if (IS_ERR(tgt))
1714                         return tgt;
1715
1716                 *mds = tgt->ltd_index;
1717                 return tgt;
1718         }
1719
1720         if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1721                 if (cfs_fail_val >= lso->lso_lsm.lsm_md_stripe_count)
1722                         return ERR_PTR(-EBADF);
1723                 oinfo = &lso->lso_lsm.lsm_md_oinfo[cfs_fail_val];
1724         } else {
1725                 oinfo = lsm_name_to_stripe_info(lso, name, namelen, new_layout);
1726                 if (IS_ERR(oinfo))
1727                         return ERR_CAST(oinfo);
1728         }
1729
1730         /* check stripe FID is sane */
1731         if (!fid_is_sane(&oinfo->lmo_fid))
1732                 return ERR_PTR(-ENODEV);
1733
1734         *fid = oinfo->lmo_fid;
1735         *mds = oinfo->lmo_mds;
1736         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1737
1738         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1739
1740         return tgt ? tgt : ERR_PTR(-ENODEV);
1741 }
1742
1743 /**
1744  * Locate MDT of op_data->op_fid1
1745  *
1746  * For striped directory, it will locate the stripe by name hash, if hash_type
1747  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1748  * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1749  * indicates whether old or new layout is used to locate.
1750  *
1751  * For plain direcotry, it just locate the MDT of op_data->op_fid1.
1752  *
1753  * \param[in] lmv               LMV device
1754  * \param[in/out] op_data       client MD stack parameters, name, namelen etc,
1755  *                              op_mds and op_fid1 will be updated if op_lso1
1756  *                              indicates fid1 represents a striped directory.
1757  *
1758  * retval               pointer to the lmv_tgt_desc if succeed.
1759  *                      ERR_PTR(errno) if failed.
1760  */
1761 struct lmv_tgt_desc *
1762 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1763 {
1764         struct lmv_stripe_md *lsm;
1765         struct lmv_oinfo *oinfo;
1766         struct lmv_tgt_desc *tgt;
1767
1768         if (lmv_dir_foreign(op_data->op_lso1))
1769                 return ERR_PTR(-ENODATA);
1770
1771         /* During creating VOLATILE file, it should honor the mdt
1772          * index if the file under striped dir is being restored, see
1773          * ct_restore(). */
1774         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1775             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1776                 tgt = lmv_tgt(lmv, op_data->op_mds);
1777                 if (!tgt)
1778                         return ERR_PTR(-ENODEV);
1779
1780                 if (lmv_dir_striped(op_data->op_lso1)) {
1781                         int i;
1782
1783                         /* refill the right parent fid */
1784                         lsm = &op_data->op_lso1->lso_lsm;
1785                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1786                                 oinfo = &lsm->lsm_md_oinfo[i];
1787                                 if (oinfo->lmo_mds == op_data->op_mds) {
1788                                         op_data->op_fid1 = oinfo->lmo_fid;
1789                                         break;
1790                                 }
1791                         }
1792
1793                         if (i == lsm->lsm_md_stripe_count)
1794                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1795                 }
1796         } else if (lmv_dir_bad_hash(op_data->op_lso1)) {
1797                 lsm = &op_data->op_lso1->lso_lsm;
1798
1799                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1800                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1801
1802                 op_data->op_fid1 = oinfo->lmo_fid;
1803                 op_data->op_mds = oinfo->lmo_mds;
1804                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1805                 if (!tgt)
1806                         return ERR_PTR(-ENODEV);
1807         } else {
1808                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_lso1,
1809                                 op_data->op_name, op_data->op_namelen,
1810                                 &op_data->op_fid1, &op_data->op_mds,
1811                                 op_data->op_new_layout);
1812         }
1813
1814         return tgt;
1815 }
1816
1817 /* Locate MDT of op_data->op_fid2 for link/rename */
1818 static struct lmv_tgt_desc *
1819 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1820 {
1821         struct lmv_tgt_desc *tgt;
1822         int rc;
1823
1824         LASSERT(op_data->op_name);
1825         if (lmv_dir_layout_changing(op_data->op_lso2)) {
1826                 struct lu_fid fid1 = op_data->op_fid1;
1827                 struct lmv_stripe_object *lso1 = op_data->op_lso1;
1828                 struct ptlrpc_request *request = NULL;
1829
1830                 /*
1831                  * avoid creating new file under old layout of migrating
1832                  * directory, check it here.
1833                  */
1834                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_lso2,
1835                                 op_data->op_name, op_data->op_namelen,
1836                                 &op_data->op_fid2, &op_data->op_mds, false);
1837                 if (IS_ERR(tgt))
1838                         RETURN(tgt);
1839
1840                 op_data->op_fid1 = op_data->op_fid2;
1841                 op_data->op_lso1 = op_data->op_lso2;
1842                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1843                 op_data->op_fid1 = fid1;
1844                 op_data->op_lso1 = lso1;
1845                 if (!rc) {
1846                         ptlrpc_req_finished(request);
1847                         RETURN(ERR_PTR(-EEXIST));
1848                 }
1849
1850                 if (rc != -ENOENT)
1851                         RETURN(ERR_PTR(rc));
1852         }
1853
1854         return lmv_locate_tgt_by_name(lmv, op_data->op_lso2,
1855                                       op_data->op_name, op_data->op_namelen,
1856                                       &op_data->op_fid2, &op_data->op_mds,
1857                                       true);
1858 }
1859
1860 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1861 {
1862         struct lu_tgt_desc *tgt;
1863         struct ptlrpc_request *request;
1864         int rc;
1865
1866         LASSERT(lmv_dir_layout_changing(op_data->op_lso1));
1867         LASSERT(!op_data->op_new_layout);
1868
1869         tgt = lmv_locate_tgt(lmv, op_data);
1870         if (IS_ERR(tgt))
1871                 return PTR_ERR(tgt);
1872
1873         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1874         if (!rc) {
1875                 ptlrpc_req_finished(request);
1876                 return -EEXIST;
1877         }
1878
1879         return rc;
1880 }
1881
1882 /* mkdir by QoS upon 'lfs mkdir -i -1'.
1883  *
1884  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1885  * directories under striped directory.
1886  */
1887 static inline bool lmv_op_user_qos_mkdir(const struct md_op_data *op_data)
1888 {
1889         const struct lmv_user_md *lum = op_data->op_data;
1890
1891         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1892                 return false;
1893
1894         if (lmv_dir_striped(op_data->op_lso1))
1895                 return false;
1896
1897         return (op_data->op_cli_flags & CLI_SET_MEA) && lum &&
1898                le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC &&
1899                le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
1900 }
1901
1902 /* mkdir by QoS if either ROOT or parent default LMV is space balanced. */
1903 static inline bool lmv_op_default_qos_mkdir(const struct md_op_data *op_data)
1904 {
1905         const struct lmv_stripe_object *lso = op_data->op_default_lso1;
1906
1907         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1908                 return false;
1909
1910         if (lmv_dir_striped(op_data->op_lso1))
1911                 return false;
1912
1913         return (op_data->op_flags & MF_QOS_MKDIR) ||
1914                (lso && lso->lso_lsm.lsm_md_master_mdt_index ==
1915                 LMV_OFFSET_DEFAULT);
1916 }
1917
1918 /* if parent default LMV is space balanced, and
1919  * 1. max_inherit_rr is set
1920  * 2. or parent is ROOT
1921  * mkdir roundrobin. Or if parent doesn't have default LMV, while ROOT default
1922  * LMV requests roundrobin mkdir, do the same.
1923  * NB, this needs to check server is balanced, which is done by caller.
1924  */
1925 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data)
1926 {
1927         const struct lmv_stripe_object *lso = op_data->op_default_lso1;
1928
1929         return (op_data->op_flags & MF_RR_MKDIR) ||
1930                (lso && lso->lso_lsm.lsm_md_max_inherit_rr !=
1931                 LMV_INHERIT_RR_NONE) || fid_is_root(&op_data->op_fid1);
1932 }
1933
1934 /* 'lfs mkdir -i <specific_MDT>' */
1935 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1936 {
1937         const struct lmv_user_md *lum = op_data->op_data;
1938
1939         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1940                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1941                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1942                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1943                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1944 }
1945
1946 /* parent default LMV master_mdt_index is not -1. */
1947 static inline bool
1948 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1949 {
1950         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1951                op_data->op_default_lso1 &&
1952                op_data->op_default_lso1->lso_lsm.lsm_md_master_mdt_index !=
1953                         LMV_OFFSET_DEFAULT;
1954 }
1955
1956 /* locate MDT by space usage */
1957 static struct lu_tgt_desc *lmv_locate_tgt_by_space(struct lmv_obd *lmv,
1958                                                    struct md_op_data *op_data,
1959                                                    struct lmv_tgt_desc *tgt)
1960 {
1961         struct lmv_tgt_desc *tmp = tgt;
1962
1963         tgt = lmv_locate_tgt_qos(lmv, op_data);
1964         if (tgt == ERR_PTR(-EAGAIN)) {
1965                 if (ltd_qos_is_balanced(&lmv->lmv_mdt_descs) &&
1966                     !lmv_op_default_rr_mkdir(op_data) &&
1967                     !lmv_op_user_qos_mkdir(op_data))
1968                         /* if not necessary, don't create remote directory. */
1969                         tgt = tmp;
1970                 else
1971                         tgt = lmv_locate_tgt_rr(lmv);
1972         }
1973
1974         /*
1975          * only update statfs after QoS mkdir, this means the cached statfs may
1976          * be stale, and current mkdir may not follow QoS accurately, but it's
1977          * not serious, and avoids periodic statfs when client doesn't mkdir by
1978          * QoS.
1979          */
1980         if (!IS_ERR(tgt)) {
1981                 op_data->op_mds = tgt->ltd_index;
1982                 lmv_statfs_check_update(lmv2obd_dev(lmv), tgt);
1983         }
1984
1985         return tgt;
1986 }
1987
1988 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1989                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1990                 gid_t gid, kernel_cap_t cap_effective, __u64 rdev,
1991                 struct ptlrpc_request **request)
1992 {
1993         struct obd_device *obd = exp->exp_obd;
1994         struct lmv_obd *lmv = &obd->u.lmv;
1995         struct lmv_tgt_desc *tgt;
1996         struct mdt_body *repbody;
1997         int rc;
1998
1999         ENTRY;
2000
2001         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
2002                 RETURN(-EIO);
2003
2004         if (lmv_dir_bad_hash(op_data->op_lso1))
2005                 RETURN(-EBADF);
2006
2007         if (lmv_dir_layout_changing(op_data->op_lso1)) {
2008                 /*
2009                  * if parent is migrating, create() needs to lookup existing
2010                  * name in both old and new layout, check old layout on client.
2011                  */
2012                 rc = lmv_old_layout_lookup(lmv, op_data);
2013                 if (rc != -ENOENT)
2014                         RETURN(rc);
2015
2016                 op_data->op_new_layout = true;
2017         }
2018
2019         tgt = lmv_locate_tgt(lmv, op_data);
2020         if (IS_ERR(tgt))
2021                 RETURN(PTR_ERR(tgt));
2022
2023         /* the order to apply policy in mkdir:
2024          * 1. is "lfs mkdir -i N"? mkdir on MDT N.
2025          * 2. is "lfs mkdir -i -1"? mkdir by space usage.
2026          * 3. is starting MDT specified in default LMV? mkdir on MDT N.
2027          * 4. is default LMV space balanced? mkdir by space usage.
2028          */
2029         if (lmv_op_user_specific_mkdir(op_data)) {
2030                 struct lmv_user_md *lum = op_data->op_data;
2031
2032                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2033                 tgt = lmv_tgt(lmv, op_data->op_mds);
2034                 if (!tgt)
2035                         RETURN(-ENODEV);
2036         } else if (lmv_op_user_qos_mkdir(op_data)) {
2037                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
2038                 if (IS_ERR(tgt))
2039                         RETURN(PTR_ERR(tgt));
2040         } else if (lmv_op_default_specific_mkdir(op_data)) {
2041                 struct lmv_stripe_md *lsm = &op_data->op_default_lso1->lso_lsm;
2042
2043                 op_data->op_mds = lsm->lsm_md_master_mdt_index;
2044                 tgt = lmv_tgt(lmv, op_data->op_mds);
2045                 if (!tgt)
2046                         RETURN(-ENODEV);
2047         } else if (lmv_op_default_qos_mkdir(op_data)) {
2048                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
2049                 if (IS_ERR(tgt))
2050                         RETURN(PTR_ERR(tgt));
2051         }
2052
2053 retry:
2054         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
2055         if (rc)
2056                 RETURN(rc);
2057
2058         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
2059                 (int)op_data->op_namelen, op_data->op_name,
2060                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
2061                 op_data->op_mds);
2062
2063         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2064         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
2065                        cap_effective, rdev, request);
2066         if (rc == 0) {
2067                 if (*request == NULL)
2068                         RETURN(rc);
2069                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
2070         }
2071
2072         /* dir restripe needs to send to MDT where dir is located */
2073         if (rc != -EREMOTE ||
2074             !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
2075                 RETURN(rc);
2076
2077         repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2078         if (repbody == NULL)
2079                 RETURN(-EPROTO);
2080
2081         /* Not cross-ref case, just get out of here. */
2082         if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
2083                 RETURN(rc);
2084
2085         op_data->op_fid2 = repbody->mbo_fid1;
2086         ptlrpc_req_finished(*request);
2087         *request = NULL;
2088
2089         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2090         if (IS_ERR(tgt))
2091                 RETURN(PTR_ERR(tgt));
2092
2093         op_data->op_mds = tgt->ltd_index;
2094         goto retry;
2095 }
2096
2097 static int
2098 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
2099             const union ldlm_policy_data *policy, struct md_op_data *op_data,
2100             struct lustre_handle *lockh, __u64 extra_lock_flags)
2101 {
2102         struct obd_device *obd = exp->exp_obd;
2103         struct lmv_obd *lmv = &obd->u.lmv;
2104         struct lmv_tgt_desc *tgt;
2105         int rc;
2106
2107         ENTRY;
2108
2109         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
2110
2111         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2112         if (IS_ERR(tgt))
2113                 RETURN(PTR_ERR(tgt));
2114
2115         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
2116                PFID(&op_data->op_fid1), tgt->ltd_index);
2117
2118         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
2119                         extra_lock_flags);
2120
2121         RETURN(rc);
2122 }
2123
2124 int
2125 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
2126                  struct ptlrpc_request **preq)
2127 {
2128         struct obd_device *obd = exp->exp_obd;
2129         struct lmv_obd *lmv = &obd->u.lmv;
2130         struct lmv_tgt_desc *tgt;
2131         struct mdt_body *body;
2132         int rc;
2133
2134         ENTRY;
2135
2136 retry:
2137         if (op_data->op_namelen == 2 &&
2138             op_data->op_name[0] == '.' && op_data->op_name[1] == '.')
2139                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2140         else
2141                 tgt = lmv_locate_tgt(lmv, op_data);
2142         if (IS_ERR(tgt))
2143                 RETURN(PTR_ERR(tgt));
2144
2145         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
2146                 (int)op_data->op_namelen, op_data->op_name,
2147                 PFID(&op_data->op_fid1), tgt->ltd_index);
2148
2149         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
2150         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2151                 ptlrpc_req_finished(*preq);
2152                 *preq = NULL;
2153                 goto retry;
2154         }
2155
2156         if (rc)
2157                 RETURN(rc);
2158
2159         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
2160         LASSERT(body != NULL);
2161
2162         if (body->mbo_valid & OBD_MD_MDS) {
2163                 op_data->op_fid1 = body->mbo_fid1;
2164                 op_data->op_valid |= OBD_MD_FLCROSSREF;
2165                 op_data->op_namelen = 0;
2166                 op_data->op_name = NULL;
2167
2168                 ptlrpc_req_finished(*preq);
2169                 *preq = NULL;
2170
2171                 goto retry;
2172         }
2173
2174         RETURN(rc);
2175 }
2176
2177 #define md_op_data_fid(op_data, fl)                     \
2178         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2179          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2180          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2181          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2182          NULL)
2183
2184 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2185                             struct md_op_data *op_data, __u32 op_tgt,
2186                             enum ldlm_mode mode, int bits, int flag)
2187 {
2188         struct lu_fid *fid = md_op_data_fid(op_data, flag);
2189         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
2190         union ldlm_policy_data policy = { { 0 } };
2191         int rc = 0;
2192         ENTRY;
2193
2194         if (!fid_is_sane(fid))
2195                 RETURN(0);
2196
2197         if (tgt == NULL) {
2198                 tgt = lmv_fid2tgt(lmv, fid);
2199                 if (IS_ERR(tgt))
2200                         RETURN(PTR_ERR(tgt));
2201         }
2202
2203         if (tgt->ltd_index != op_tgt) {
2204                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2205                 policy.l_inodebits.bits = bits;
2206                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2207                                       mode, LCF_ASYNC, NULL);
2208         } else {
2209                 CDEBUG(D_INODE,
2210                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2211                        op_tgt, PFID(fid));
2212                 op_data->op_flags |= flag;
2213                 rc = 0;
2214         }
2215
2216         RETURN(rc);
2217 }
2218
2219 /*
2220  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2221  * op_data->op_fid2
2222  */
2223 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2224                     struct ptlrpc_request **request)
2225 {
2226         struct obd_device       *obd = exp->exp_obd;
2227         struct lmv_obd          *lmv = &obd->u.lmv;
2228         struct lmv_tgt_desc     *tgt;
2229         int                      rc;
2230         ENTRY;
2231
2232         LASSERT(op_data->op_namelen != 0);
2233
2234         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2235                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2236                op_data->op_name, PFID(&op_data->op_fid1));
2237
2238         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2239         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2240         op_data->op_cap = current_cap();
2241
2242         tgt = lmv_locate_tgt2(lmv, op_data);
2243         if (IS_ERR(tgt))
2244                 RETURN(PTR_ERR(tgt));
2245
2246         /*
2247          * Cancel UPDATE lock on child (fid1).
2248          */
2249         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2250         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2251                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2252         if (rc != 0)
2253                 RETURN(rc);
2254
2255         rc = md_link(tgt->ltd_exp, op_data, request);
2256
2257         RETURN(rc);
2258 }
2259
2260 /* migrate the top directory */
2261 static inline bool lmv_op_topdir_migrate(const struct md_op_data *op_data)
2262 {
2263         if (!S_ISDIR(op_data->op_mode))
2264                 return false;
2265
2266         if (lmv_dir_layout_changing(op_data->op_lso1))
2267                 return false;
2268
2269         return true;
2270 }
2271
2272 /* migrate top dir to specific MDTs */
2273 static inline bool lmv_topdir_specific_migrate(const struct md_op_data *op_data)
2274 {
2275         const struct lmv_user_md *lum = op_data->op_data;
2276
2277         if (!lmv_op_topdir_migrate(op_data))
2278                 return false;
2279
2280         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2281 }
2282
2283 /* migrate top dir in QoS mode if user issued "lfs migrate -m -1..." */
2284 static inline bool lmv_topdir_qos_migrate(const struct md_op_data *op_data)
2285 {
2286         const struct lmv_user_md *lum = op_data->op_data;
2287
2288         if (!lmv_op_topdir_migrate(op_data))
2289                 return false;
2290
2291         return le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
2292 }
2293
2294 static inline bool lmv_subdir_specific_migrate(const struct md_op_data *op_data)
2295 {
2296         const struct lmv_user_md *lum = op_data->op_data;
2297
2298         if (!S_ISDIR(op_data->op_mode))
2299                 return false;
2300
2301         if (!lmv_dir_layout_changing(op_data->op_lso1))
2302                 return false;
2303
2304         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2305 }
2306
2307 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2308                         const char *name, size_t namelen,
2309                         struct ptlrpc_request **request)
2310 {
2311         struct obd_device *obd = exp->exp_obd;
2312         struct lmv_obd *lmv = &obd->u.lmv;
2313         struct lmv_stripe_object *lso = op_data->op_lso1;
2314         struct lmv_tgt_desc *parent_tgt;
2315         struct lmv_tgt_desc *sp_tgt;
2316         struct lmv_tgt_desc *tp_tgt = NULL;
2317         struct lmv_tgt_desc *child_tgt;
2318         struct lmv_tgt_desc *tgt;
2319         struct lu_fid target_fid = { 0 };
2320         int rc;
2321
2322         ENTRY;
2323
2324         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2325
2326         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2327                PFID(&op_data->op_fid1), (int)namelen, name);
2328
2329         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2330         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2331         op_data->op_cap = current_cap();
2332
2333         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2334         if (IS_ERR(parent_tgt))
2335                 RETURN(PTR_ERR(parent_tgt));
2336
2337         if (lmv_dir_striped(lso)) {
2338                 const struct lmv_oinfo *oinfo;
2339
2340                 oinfo = lsm_name_to_stripe_info(lso, name, namelen, false);
2341                 if (IS_ERR(oinfo))
2342                         RETURN(PTR_ERR(oinfo));
2343
2344                 /* save source stripe FID in fid4 temporarily for ELC */
2345                 op_data->op_fid4 = oinfo->lmo_fid;
2346                 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2347                 if (!sp_tgt)
2348                         RETURN(-ENODEV);
2349
2350                 /*
2351                  * if parent is being migrated too, fill op_fid2 with target
2352                  * stripe fid, otherwise the target stripe is not created yet.
2353                  */
2354                 if (lmv_dir_layout_changing(lso)) {
2355                         oinfo = lsm_name_to_stripe_info(lso, name, namelen,
2356                                                         true);
2357                         if (IS_ERR(oinfo))
2358                                 RETURN(PTR_ERR(oinfo));
2359
2360                         op_data->op_fid2 = oinfo->lmo_fid;
2361                         tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2362                         if (!tp_tgt)
2363                                 RETURN(-ENODEV);
2364
2365                         /* parent unchanged and update namespace only */
2366                         if (lu_fid_eq(&op_data->op_fid4, &op_data->op_fid2) &&
2367                             op_data->op_bias & MDS_MIGRATE_NSONLY)
2368                                 RETURN(-EALREADY);
2369                 }
2370         } else {
2371                 sp_tgt = parent_tgt;
2372         }
2373
2374         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2375         if (IS_ERR(child_tgt))
2376                 RETURN(PTR_ERR(child_tgt));
2377
2378         if (lmv_topdir_specific_migrate(op_data)) {
2379                 struct lmv_user_md *lum = op_data->op_data;
2380
2381                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2382         } else if (lmv_topdir_qos_migrate(op_data)) {
2383                 tgt = lmv_locate_tgt_lf(lmv);
2384                 if (tgt == ERR_PTR(-EAGAIN))
2385                         tgt = lmv_locate_tgt_rr(lmv);
2386                 if (IS_ERR(tgt))
2387                         RETURN(PTR_ERR(tgt));
2388
2389                 op_data->op_mds = tgt->ltd_index;
2390         } else if (lmv_subdir_specific_migrate(op_data)) {
2391                 struct lmv_user_md *lum = op_data->op_data;
2392                 __u32 i;
2393
2394                 LASSERT(tp_tgt);
2395                 if (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) {
2396                         /* adjust MDTs in lum, since subdir is located on where
2397                          * its parent stripe is, not the first specified MDT.
2398                          */
2399                         for (i = 0; i < le32_to_cpu(lum->lum_stripe_count);
2400                              i++) {
2401                                 if (le32_to_cpu(lum->lum_objects[i].lum_mds) ==
2402                                     tp_tgt->ltd_index)
2403                                         break;
2404                         }
2405
2406                         if (i == le32_to_cpu(lum->lum_stripe_count))
2407                                 RETURN(-ENODEV);
2408
2409                         lum->lum_objects[i].lum_mds =
2410                                 lum->lum_objects[0].lum_mds;
2411                         lum->lum_objects[0].lum_mds =
2412                                 cpu_to_le32(tp_tgt->ltd_index);
2413                 }
2414                 /* NB, the above adjusts subdir migration for command like
2415                  * "lfs migrate -m 0,1,2 ...", but for migration like
2416                  * "lfs migrate -m 0 -c 2 ...", the top dir is migrated to MDT0
2417                  * and MDT1, however its subdir may be migrated to MDT1 and MDT2
2418                  */
2419
2420                 lum->lum_stripe_offset = cpu_to_le32(tp_tgt->ltd_index);
2421                 op_data->op_mds = tp_tgt->ltd_index;
2422         } else if (tp_tgt) {
2423                 op_data->op_mds = tp_tgt->ltd_index;
2424         } else {
2425                 op_data->op_mds = sp_tgt->ltd_index;
2426         }
2427
2428         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2429         if (rc)
2430                 RETURN(rc);
2431
2432         /*
2433          * for directory, send migrate request to the MDT where the object will
2434          * be migrated to, because we can't create a striped directory remotely.
2435          *
2436          * otherwise, send to the MDT where source is located because regular
2437          * file may open lease.
2438          *
2439          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2440          * backward compatibility.
2441          */
2442         if (S_ISDIR(op_data->op_mode) &&
2443             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2444                 tgt = lmv_fid2tgt(lmv, &target_fid);
2445                 if (IS_ERR(tgt))
2446                         RETURN(PTR_ERR(tgt));
2447         } else {
2448                 tgt = child_tgt;
2449         }
2450
2451         /* cancel UPDATE lock of parent master object */
2452         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2453                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2454         if (rc)
2455                 RETURN(rc);
2456
2457         /* cancel UPDATE lock of source parent */
2458         if (sp_tgt != parent_tgt) {
2459                 /*
2460                  * migrate RPC packs master object FID, because we can only pack
2461                  * two FIDs in reint RPC, but MDS needs to know both source
2462                  * parent and target parent, and it will obtain them from master
2463                  * FID and LMV, the other FID in RPC is kept for target.
2464                  *
2465                  * since this FID is not passed to MDC, cancel it anyway.
2466                  */
2467                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2468                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2469                 if (rc)
2470                         RETURN(rc);
2471
2472                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2473         }
2474         op_data->op_fid4 = target_fid;
2475
2476         /* cancel UPDATE locks of target parent */
2477         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2478                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2479         if (rc)
2480                 RETURN(rc);
2481
2482         /* cancel LOOKUP lock of source if source is remote object */
2483         if (child_tgt != sp_tgt) {
2484                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2485                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2486                                       MF_MDC_CANCEL_FID3);
2487                 if (rc)
2488                         RETURN(rc);
2489         }
2490
2491         /* cancel ELC locks of source */
2492         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2493                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2494         if (rc)
2495                 RETURN(rc);
2496
2497         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2498
2499         RETURN(rc);
2500 }
2501
2502 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2503                       const char *old, size_t oldlen,
2504                       const char *new, size_t newlen,
2505                       struct ptlrpc_request **request)
2506 {
2507         struct obd_device *obd = exp->exp_obd;
2508         struct lmv_obd *lmv = &obd->u.lmv;
2509         struct lmv_tgt_desc *sp_tgt;
2510         struct lmv_tgt_desc *tp_tgt = NULL;
2511         struct lmv_tgt_desc *src_tgt = NULL;
2512         struct lmv_tgt_desc *tgt;
2513         struct mdt_body *body;
2514         int rc;
2515
2516         ENTRY;
2517
2518         LASSERT(oldlen != 0);
2519
2520         if (op_data->op_cli_flags & CLI_MIGRATE) {
2521                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2522                 RETURN(rc);
2523         }
2524
2525         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2526         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2527         op_data->op_cap = current_cap();
2528
2529         op_data->op_name = new;
2530         op_data->op_namelen = newlen;
2531
2532         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2533         if (IS_ERR(tp_tgt))
2534                 RETURN(PTR_ERR(tp_tgt));
2535
2536         /* Since the target child might be destroyed, and it might become
2537          * orphan, and we can only check orphan on the local MDT right now, so
2538          * we send rename request to the MDT where target child is located. If
2539          * target child does not exist, then it will send the request to the
2540          * target parent */
2541         if (fid_is_sane(&op_data->op_fid4)) {
2542                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2543                 if (IS_ERR(tgt))
2544                         RETURN(PTR_ERR(tgt));
2545         } else {
2546                 tgt = tp_tgt;
2547         }
2548
2549         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2550
2551         /* cancel UPDATE locks of target parent */
2552         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2553                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2554         if (rc != 0)
2555                 RETURN(rc);
2556
2557         if (fid_is_sane(&op_data->op_fid4)) {
2558                 /* cancel LOOKUP lock of target on target parent */
2559                 if (tgt != tp_tgt) {
2560                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2561                                               tgt->ltd_index, LCK_EX,
2562                                               MDS_INODELOCK_LOOKUP,
2563                                               MF_MDC_CANCEL_FID4);
2564                         if (rc != 0)
2565                                 RETURN(rc);
2566                 }
2567         }
2568
2569         if (fid_is_sane(&op_data->op_fid3)) {
2570                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2571                 if (IS_ERR(src_tgt))
2572                         RETURN(PTR_ERR(src_tgt));
2573
2574                 /* cancel ELC locks of source */
2575                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2576                                       LCK_EX, MDS_INODELOCK_ELC,
2577                                       MF_MDC_CANCEL_FID3);
2578                 if (rc != 0)
2579                         RETURN(rc);
2580         }
2581
2582         op_data->op_name = old;
2583         op_data->op_namelen = oldlen;
2584 retry:
2585         sp_tgt = lmv_locate_tgt(lmv, op_data);
2586         if (IS_ERR(sp_tgt))
2587                 RETURN(PTR_ERR(sp_tgt));
2588
2589         /* cancel UPDATE locks of source parent */
2590         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2591                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2592         if (rc != 0)
2593                 RETURN(rc);
2594
2595         if (fid_is_sane(&op_data->op_fid3)) {
2596                 /* cancel LOOKUP lock of source on source parent */
2597                 if (src_tgt != sp_tgt) {
2598                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2599                                               tgt->ltd_index, LCK_EX,
2600                                               MDS_INODELOCK_LOOKUP,
2601                                               MF_MDC_CANCEL_FID3);
2602                         if (rc != 0)
2603                                 RETURN(rc);
2604                 }
2605         }
2606
2607 rename:
2608         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2609                 PFID(&op_data->op_fid1), (int)oldlen, old,
2610                 PFID(&op_data->op_fid2), (int)newlen, new);
2611
2612         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2613                         request);
2614         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2615                 ptlrpc_req_finished(*request);
2616                 *request = NULL;
2617                 goto retry;
2618         }
2619
2620         if (rc && rc != -EXDEV)
2621                 RETURN(rc);
2622
2623         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2624         if (body == NULL)
2625                 RETURN(-EPROTO);
2626
2627         /* Not cross-ref case, just get out of here. */
2628         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2629                 RETURN(rc);
2630
2631         op_data->op_fid4 = body->mbo_fid1;
2632
2633         ptlrpc_req_finished(*request);
2634         *request = NULL;
2635
2636         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2637         if (IS_ERR(tgt))
2638                 RETURN(PTR_ERR(tgt));
2639
2640         if (fid_is_sane(&op_data->op_fid4)) {
2641                 /* cancel LOOKUP lock of target on target parent */
2642                 if (tgt != tp_tgt) {
2643                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2644                                               tgt->ltd_index, LCK_EX,
2645                                               MDS_INODELOCK_LOOKUP,
2646                                               MF_MDC_CANCEL_FID4);
2647                         if (rc != 0)
2648                                 RETURN(rc);
2649                 }
2650         }
2651
2652         goto rename;
2653 }
2654
2655 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2656                        void *ea, size_t ealen, struct ptlrpc_request **request)
2657 {
2658         struct obd_device *obd = exp->exp_obd;
2659         struct lmv_obd *lmv = &obd->u.lmv;
2660         struct lmv_tgt_desc *tgt;
2661         int rc = 0;
2662
2663         ENTRY;
2664
2665         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2666                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2667                op_data->op_xvalid);
2668
2669         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2670         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2671         if (IS_ERR(tgt))
2672                 RETURN(PTR_ERR(tgt));
2673
2674         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2675
2676         RETURN(rc);
2677 }
2678
2679 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2680                      struct ptlrpc_request **request)
2681 {
2682         struct obd_device *obd = exp->exp_obd;
2683         struct lmv_obd *lmv = &obd->u.lmv;
2684         struct lmv_tgt_desc *tgt;
2685         int rc;
2686
2687         ENTRY;
2688
2689         tgt = lmv_fid2tgt(lmv, fid);
2690         if (IS_ERR(tgt))
2691                 RETURN(PTR_ERR(tgt));
2692
2693         rc = md_fsync(tgt->ltd_exp, fid, request);
2694         RETURN(rc);
2695 }
2696
2697 struct stripe_dirent {
2698         struct page             *sd_page;
2699         struct lu_dirpage       *sd_dp;
2700         struct lu_dirent        *sd_ent;
2701         bool                     sd_eof;
2702 };
2703
2704 struct lmv_dir_ctxt {
2705         struct lmv_obd          *ldc_lmv;
2706         struct md_op_data       *ldc_op_data;
2707         struct md_readdir_info  *ldc_mrinfo;
2708         __u64                    ldc_hash;
2709         int                      ldc_count;
2710         struct stripe_dirent     ldc_stripes[0];
2711 };
2712
2713 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2714 {
2715         if (stripe->sd_page) {
2716                 kunmap(stripe->sd_page);
2717                 put_page(stripe->sd_page);
2718                 stripe->sd_page = NULL;
2719                 stripe->sd_ent = NULL;
2720         }
2721 }
2722
2723 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2724 {
2725         int i;
2726
2727         for (i = 0; i < ctxt->ldc_count; i++)
2728                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2729 }
2730
2731 /* if @ent is dummy, or . .., get next */
2732 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2733                                            struct lu_dirent *ent,
2734                                            int stripe_index)
2735 {
2736         for (; ent; ent = lu_dirent_next(ent)) {
2737                 /* Skip dummy entry */
2738                 if (le16_to_cpu(ent->lde_namelen) == 0)
2739                         continue;
2740
2741                 /* skip . and .. for other stripes */
2742                 if (stripe_index &&
2743                     (strncmp(ent->lde_name, ".",
2744                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2745                      strncmp(ent->lde_name, "..",
2746                              le16_to_cpu(ent->lde_namelen)) == 0))
2747                         continue;
2748
2749                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2750                         break;
2751         }
2752
2753         return ent;
2754 }
2755
2756 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2757                                             struct stripe_dirent *stripe,
2758                                             int stripe_index)
2759 {
2760         struct md_op_data *op_data = ctxt->ldc_op_data;
2761         struct lmv_oinfo *oinfo;
2762         struct lu_fid fid = op_data->op_fid1;
2763         struct inode *inode = op_data->op_data;
2764         struct lmv_tgt_desc *tgt;
2765         struct lu_dirent *ent = stripe->sd_ent;
2766         __u64 hash = ctxt->ldc_hash;
2767         int rc = 0;
2768
2769         ENTRY;
2770
2771         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2772         LASSERT(!ent);
2773
2774         do {
2775                 if (stripe->sd_page) {
2776                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2777
2778                         /* @hash should be the last dirent hash */
2779                         LASSERTF(hash <= end,
2780                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2781                                  ctxt, stripe, hash, end);
2782                         /* unload last page */
2783                         stripe_dirent_unload(stripe);
2784                         /* eof */
2785                         if (end == MDS_DIR_END_OFF) {
2786                                 stripe->sd_eof = true;
2787                                 break;
2788                         }
2789                         hash = end;
2790                 }
2791
2792                 oinfo = &op_data->op_lso1->lso_lsm.lsm_md_oinfo[stripe_index];
2793                 if (!oinfo->lmo_root) {
2794                         rc = -ENOENT;
2795                         break;
2796                 }
2797
2798                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2799                 if (!tgt) {
2800                         rc = -ENODEV;
2801                         break;
2802                 }
2803
2804                 /* op_data is shared by stripes, reset after use */
2805                 op_data->op_fid1 = oinfo->lmo_fid;
2806                 op_data->op_fid2 = oinfo->lmo_fid;
2807                 op_data->op_data = oinfo->lmo_root;
2808
2809                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_mrinfo, hash,
2810                                   &stripe->sd_page);
2811
2812                 op_data->op_fid1 = fid;
2813                 op_data->op_fid2 = fid;
2814                 op_data->op_data = inode;
2815
2816                 if (rc)
2817                         break;
2818
2819                 stripe->sd_dp = page_address(stripe->sd_page);
2820                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2821                                         stripe_index);
2822                 /* in case a page filled with ., .. and dummy, read next */
2823         } while (!ent);
2824
2825         stripe->sd_ent = ent;
2826         if (rc) {
2827                 LASSERT(!ent);
2828                 /* treat error as eof, so dir can be partially accessed */
2829                 stripe->sd_eof = true;
2830                 ctxt->ldc_mrinfo->mr_partial_readdir_rc = rc;
2831                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2832                               "directory is partially accessed!\n",
2833                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2834                               rc);
2835         }
2836
2837         RETURN(ent);
2838 }
2839
2840 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2841 {
2842         struct obd_device *obd = exp->exp_obd;
2843         struct lmv_obd *lmv = &obd->u.lmv;
2844         struct lmv_tgt_desc *tgt;
2845         int rc;
2846
2847         ENTRY;
2848
2849         rc = lmv_check_connect(obd);
2850         if (rc != 0)
2851                 RETURN(rc);
2852
2853         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2854         if (IS_ERR(tgt))
2855                 RETURN(PTR_ERR(tgt));
2856
2857         data->op_flags |= MF_MDC_CANCEL_FID1;
2858         rc = md_file_resync(tgt->ltd_exp, data);
2859         RETURN(rc);
2860 }
2861
2862 /**
2863  * Get dirent with the closest hash for striped directory
2864  *
2865  * This function will search the dir entry, whose hash value is the
2866  * closest(>=) to hash from all of sub-stripes, and it is only being called
2867  * for striped directory.
2868  *
2869  * \param[in] ctxt              dir read context
2870  *
2871  * \retval                      dirent get the entry successfully
2872  *                              NULL does not get the entry, normally it means
2873  *                              it reaches the end of the directory, while read
2874  *                              stripe dirent error is ignored to allow partial
2875  *                              access.
2876  */
2877 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2878 {
2879         struct stripe_dirent *stripe;
2880         struct lu_dirent *ent = NULL;
2881         int i;
2882         int min = -1;
2883
2884         /* TODO: optimize with k-way merge sort */
2885         for (i = 0; i < ctxt->ldc_count; i++) {
2886                 stripe = &ctxt->ldc_stripes[i];
2887                 if (stripe->sd_eof)
2888                         continue;
2889
2890                 if (!stripe->sd_ent) {
2891                         stripe_dirent_load(ctxt, stripe, i);
2892                         if (!stripe->sd_ent) {
2893                                 LASSERT(stripe->sd_eof);
2894                                 continue;
2895                         }
2896                 }
2897
2898                 if (min == -1 ||
2899                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2900                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2901                         min = i;
2902                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2903                             ctxt->ldc_hash)
2904                                 break;
2905                 }
2906         }
2907
2908         if (min != -1) {
2909                 stripe = &ctxt->ldc_stripes[min];
2910                 ent = stripe->sd_ent;
2911                 /* pop found dirent */
2912                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2913                                                    min);
2914         }
2915
2916         return ent;
2917 }
2918
2919 /**
2920  * Build dir entry page for striped directory
2921  *
2922  * This function gets one entry by @offset from a striped directory. It will
2923  * read entries from all of stripes, and choose one closest to the required
2924  * offset(&offset). A few notes
2925  * 1. skip . and .. for non-zero stripes, because there can only have one .
2926  * and .. in a directory.
2927  * 2. op_data will be shared by all of stripes, instead of allocating new
2928  * one, so need to restore before reusing.
2929  *
2930  * \param[in] exp       obd export refer to LMV
2931  * \param[in] op_data   hold those MD parameters of read_entry
2932  * \param[in] mrinfo    ldlm callback being used in enqueue in mdc_read_entry,
2933  *                      and partial readdir result will be stored in it.
2934  * \param[in] offset    starting hash offset
2935  * \param[out] ppage    the page holding the entry. Note: because the entry
2936  *                      will be accessed in upper layer, so we need hold the
2937  *                      page until the usages of entry is finished, see
2938  *                      ll_dir_entry_next.
2939  *
2940  * retval               =0 if get entry successfully
2941  *                      <0 cannot get entry
2942  */
2943 static int lmv_striped_read_page(struct obd_export *exp,
2944                                  struct md_op_data *op_data,
2945                                  struct md_readdir_info *mrinfo, __u64 offset,
2946                                  struct page **ppage)
2947 {
2948         struct page *page = NULL;
2949         struct lu_dirpage *dp;
2950         void *start;
2951         struct lu_dirent *ent;
2952         struct lu_dirent *last_ent;
2953         int stripe_count;
2954         struct lmv_dir_ctxt *ctxt;
2955         struct lu_dirent *next = NULL;
2956         __u16 ent_size;
2957         size_t left_bytes;
2958         int rc = 0;
2959         ENTRY;
2960
2961         /* Allocate a page and read entries from all of stripes and fill
2962          * the page by hash order */
2963         page = alloc_page(GFP_KERNEL);
2964         if (!page)
2965                 RETURN(-ENOMEM);
2966
2967         /* Initialize the entry page */
2968         dp = kmap(page);
2969         memset(dp, 0, sizeof(*dp));
2970         dp->ldp_hash_start = cpu_to_le64(offset);
2971
2972         start = dp + 1;
2973         left_bytes = PAGE_SIZE - sizeof(*dp);
2974         ent = start;
2975         last_ent = ent;
2976
2977         /* initalize dir read context */
2978         stripe_count = op_data->op_lso1->lso_lsm.lsm_md_stripe_count;
2979         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2980         if (!ctxt)
2981                 GOTO(free_page, rc = -ENOMEM);
2982         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2983         ctxt->ldc_op_data = op_data;
2984         ctxt->ldc_mrinfo = mrinfo;
2985         ctxt->ldc_hash = offset;
2986         ctxt->ldc_count = stripe_count;
2987
2988         while (1) {
2989                 next = lmv_dirent_next(ctxt);
2990
2991                 /* end of directory */
2992                 if (!next) {
2993                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2994                         break;
2995                 }
2996                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2997
2998                 ent_size = le16_to_cpu(next->lde_reclen);
2999
3000                 /* the last entry lde_reclen is 0, but it might not be the last
3001                  * one of this temporay dir page */
3002                 if (!ent_size)
3003                         ent_size = lu_dirent_calc_size(
3004                                         le16_to_cpu(next->lde_namelen),
3005                                         le32_to_cpu(next->lde_attrs));
3006                 /* page full */
3007                 if (ent_size > left_bytes)
3008                         break;
3009
3010                 memcpy(ent, next, ent_size);
3011
3012                 /* Replace . with master FID and Replace .. with the parent FID
3013                  * of master object */
3014                 if (strncmp(ent->lde_name, ".",
3015                             le16_to_cpu(ent->lde_namelen)) == 0 &&
3016                     le16_to_cpu(ent->lde_namelen) == 1)
3017                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
3018                 else if (strncmp(ent->lde_name, "..",
3019                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
3020                            le16_to_cpu(ent->lde_namelen) == 2)
3021                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
3022
3023                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
3024                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
3025                        le64_to_cpu(ent->lde_hash));
3026
3027                 left_bytes -= ent_size;
3028                 ent->lde_reclen = cpu_to_le16(ent_size);
3029                 last_ent = ent;
3030                 ent = (void *)ent + ent_size;
3031         };
3032
3033         last_ent->lde_reclen = 0;
3034
3035         if (ent == start)
3036                 dp->ldp_flags |= LDF_EMPTY;
3037         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
3038                 dp->ldp_flags |= LDF_COLLIDE;
3039         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
3040         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
3041
3042         put_lmv_dir_ctxt(ctxt);
3043         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
3044
3045         *ppage = page;
3046
3047         RETURN(0);
3048
3049 free_page:
3050         kunmap(page);
3051         __free_page(page);
3052
3053         return rc;
3054 }
3055
3056 static int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
3057                          struct md_readdir_info *mrinfo, __u64 offset,
3058                          struct page **ppage)
3059 {
3060         struct obd_device *obd = exp->exp_obd;
3061         struct lmv_obd *lmv = &obd->u.lmv;
3062         struct lmv_tgt_desc *tgt;
3063         int rc;
3064
3065         ENTRY;
3066
3067         if (unlikely(lmv_dir_foreign(op_data->op_lso1)))
3068                 RETURN(-ENODATA);
3069
3070         if (unlikely(lmv_dir_striped(op_data->op_lso1))) {
3071                 rc = lmv_striped_read_page(exp, op_data, mrinfo, offset, ppage);
3072                 RETURN(rc);
3073         }
3074
3075         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
3076         if (IS_ERR(tgt))
3077                 RETURN(PTR_ERR(tgt));
3078
3079         rc = md_read_page(tgt->ltd_exp, op_data, mrinfo, offset, ppage);
3080
3081         RETURN(rc);
3082 }
3083
3084 /**
3085  * Unlink a file/directory
3086  *
3087  * Unlink a file or directory under the parent dir. The unlink request
3088  * usually will be sent to the MDT where the child is located, but if
3089  * the client does not have the child FID then request will be sent to the
3090  * MDT where the parent is located.
3091  *
3092  * If the parent is a striped directory then it also needs to locate which
3093  * stripe the name of the child is located, and replace the parent FID
3094  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
3095  * it will walk through all of sub-stripes until the child is being
3096  * unlinked finally.
3097  *
3098  * \param[in] exp       export refer to LMV
3099  * \param[in] op_data   different parameters transferred beween client
3100  *                      MD stacks, name, namelen, FIDs etc.
3101  *                      op_fid1 is the parent FID, op_fid2 is the child
3102  *                      FID.
3103  * \param[out] request  point to the request of unlink.
3104  *
3105  * retval               0 if succeed
3106  *                      negative errno if failed.
3107  */
3108 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
3109                       struct ptlrpc_request **request)
3110 {
3111         struct obd_device *obd = exp->exp_obd;
3112         struct lmv_obd *lmv = &obd->u.lmv;
3113         struct lmv_tgt_desc *tgt;
3114         struct lmv_tgt_desc *parent_tgt;
3115         struct mdt_body *body;
3116         int rc;
3117
3118         ENTRY;
3119
3120         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
3121         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
3122         op_data->op_cap = current_cap();
3123
3124 retry:
3125         parent_tgt = lmv_locate_tgt(lmv, op_data);
3126         if (IS_ERR(parent_tgt))
3127                 RETURN(PTR_ERR(parent_tgt));
3128
3129         if (likely(!fid_is_zero(&op_data->op_fid2))) {
3130                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3131                 if (IS_ERR(tgt))
3132                         RETURN(PTR_ERR(tgt));
3133         } else {
3134                 tgt = parent_tgt;
3135         }
3136
3137         /*
3138          * If child's fid is given, cancel unused locks for it if it is from
3139          * another export than parent.
3140          *
3141          * LOOKUP lock for child (fid3) should also be cancelled on parent
3142          * tgt_tgt in mdc_unlink().
3143          */
3144         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
3145
3146         if (parent_tgt != tgt)
3147                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
3148                                       LCK_EX, MDS_INODELOCK_LOOKUP,
3149                                       MF_MDC_CANCEL_FID3);
3150
3151         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
3152                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
3153         if (rc)
3154                 RETURN(rc);
3155
3156         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
3157                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
3158                tgt->ltd_index);
3159
3160         rc = md_unlink(tgt->ltd_exp, op_data, request);
3161         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
3162                 ptlrpc_req_finished(*request);
3163                 *request = NULL;
3164                 goto retry;
3165         }
3166
3167         if (rc != -EREMOTE)
3168                 RETURN(rc);
3169
3170         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
3171         if (body == NULL)
3172                 RETURN(-EPROTO);
3173
3174         /* Not cross-ref case, just get out of here. */
3175         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
3176                 RETURN(rc);
3177
3178         /* This is a remote object, try remote MDT. */
3179         op_data->op_fid2 = body->mbo_fid1;
3180         ptlrpc_req_finished(*request);
3181         *request = NULL;
3182
3183         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3184         if (IS_ERR(tgt))
3185                 RETURN(PTR_ERR(tgt));
3186
3187         goto retry;
3188 }
3189
3190 static int lmv_precleanup(struct obd_device *obd)
3191 {
3192         ENTRY;
3193         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
3194         RETURN(0);
3195 }
3196
3197 /**
3198  * Get by key a value associated with a LMV device.
3199  *
3200  * Dispatch request to lower-layer devices as needed.
3201  *
3202  * \param[in] env               execution environment for this thread
3203  * \param[in] exp               export for the LMV device
3204  * \param[in] keylen            length of key identifier
3205  * \param[in] key               identifier of key to get value for
3206  * \param[in] vallen            size of \a val
3207  * \param[out] val              pointer to storage location for value
3208  * \param[in] lsm               optional striping metadata of object
3209  *
3210  * \retval 0            on success
3211  * \retval negative     negated errno on failure
3212  */
3213 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
3214                         __u32 keylen, void *key, __u32 *vallen, void *val)
3215 {
3216         struct obd_device *obd;
3217         struct lmv_obd *lmv;
3218         struct lu_tgt_desc *tgt;
3219         int rc = 0;
3220
3221         ENTRY;
3222
3223         obd = class_exp2obd(exp);
3224         if (obd == NULL) {
3225                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3226                        exp->exp_handle.h_cookie);
3227                 RETURN(-EINVAL);
3228         }
3229
3230         lmv = &obd->u.lmv;
3231         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
3232                 LASSERT(*vallen == sizeof(__u32));
3233                 lmv_foreach_connected_tgt(lmv, tgt) {
3234                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
3235                                           vallen, val))
3236                                 RETURN(0);
3237                 }
3238                 RETURN(-EINVAL);
3239         } else if (KEY_IS(KEY_MAX_EASIZE) ||
3240                    KEY_IS(KEY_DEFAULT_EASIZE) ||
3241                    KEY_IS(KEY_CONN_DATA)) {
3242                 /*
3243                  * Forwarding this request to first MDS, it should know LOV
3244                  * desc.
3245                  */
3246                 tgt = lmv_tgt(lmv, 0);
3247                 if (!tgt)
3248                         RETURN(-ENODEV);
3249
3250                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
3251                 if (!rc && KEY_IS(KEY_CONN_DATA))
3252                         exp->exp_connect_data = *(struct obd_connect_data *)val;
3253                 RETURN(rc);
3254         } else if (KEY_IS(KEY_TGT_COUNT)) {
3255                 *((int *)val) = lmv->lmv_mdt_descs.ltd_tgts_size;
3256                 RETURN(0);
3257         }
3258
3259         CDEBUG(D_IOCTL, "Invalid key\n");
3260         RETURN(-EINVAL);
3261 }
3262
3263 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
3264                      int *__rcs, struct ptlrpc_request_set *_set)
3265 {
3266         struct obd_device *obd = class_exp2obd(exp);
3267         struct ptlrpc_request_set *set = _set;
3268         struct lmv_obd *lmv = &obd->u.lmv;
3269         int tgt_count = lmv->lmv_mdt_count;
3270         struct lu_tgt_desc *tgt;
3271         struct fid_array *fat, **fas = NULL;
3272         int i, rc, **rcs = NULL;
3273
3274         if (!set) {
3275                 set = ptlrpc_prep_set();
3276                 if (!set)
3277                         RETURN(-ENOMEM);
3278         }
3279
3280         /* split FIDs by targets */
3281         OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
3282         if (fas == NULL)
3283                 GOTO(out, rc = -ENOMEM);
3284         OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
3285         if (rcs == NULL)
3286                 GOTO(out_fas, rc = -ENOMEM);
3287
3288         for (i = 0; i < fa->fa_nr; i++) {
3289                 unsigned int idx;
3290
3291                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
3292                 if (rc) {
3293                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
3294                                PFID(&fa->fa_fids[i]), rc);
3295                         continue;
3296                 }
3297                 LASSERT(idx < tgt_count);
3298                 if (!fas[idx])
3299                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
3300                                   fa_fids[fa->fa_nr]));
3301                 if (!fas[idx])
3302                         GOTO(out, rc = -ENOMEM);
3303                 if (!rcs[idx])
3304                         OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
3305                 if (!rcs[idx])
3306                         GOTO(out, rc = -ENOMEM);
3307
3308                 fat = fas[idx];
3309                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
3310         }
3311
3312         lmv_foreach_connected_tgt(lmv, tgt) {
3313                 fat = fas[tgt->ltd_index];
3314                 if (!fat || fat->fa_nr == 0)
3315                         continue;
3316                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3317         }
3318
3319         rc = ptlrpc_set_wait(NULL, set);
3320         if (rc == 0) {
3321                 int j = 0;
3322                 for (i = 0; i < tgt_count; i++) {
3323                         fat = fas[i];
3324                         if (!fat || fat->fa_nr == 0)
3325                                 continue;
3326                         /* copy FIDs back */
3327                         memcpy(fa->fa_fids + j, fat->fa_fids,
3328                                fat->fa_nr * sizeof(struct lu_fid));
3329                         /* copy rcs back */
3330                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3331                         j += fat->fa_nr;
3332                 }
3333         }
3334         if (set != _set)
3335                 ptlrpc_set_destroy(set);
3336
3337 out:
3338         for (i = 0; i < tgt_count; i++) {
3339                 if (fas && fas[i])
3340                         OBD_FREE(fas[i], offsetof(struct fid_array,
3341                                                 fa_fids[fa->fa_nr]));
3342                 if (rcs && rcs[i])
3343                         OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3344         }
3345         if (rcs)
3346                 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3347 out_fas:
3348         if (fas)
3349                 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3350
3351         RETURN(rc);
3352 }
3353
3354 /**
3355  * Asynchronously set by key a value associated with a LMV device.
3356  *
3357  * Dispatch request to lower-layer devices as needed.
3358  *
3359  * \param[in] env       execution environment for this thread
3360  * \param[in] exp       export for the LMV device
3361  * \param[in] keylen    length of key identifier
3362  * \param[in] key       identifier of key to store value for
3363  * \param[in] vallen    size of value to store
3364  * \param[in] val       pointer to data to be stored
3365  * \param[in] set       optional list of related ptlrpc requests
3366  *
3367  * \retval 0            on success
3368  * \retval negative     negated errno on failure
3369  */
3370 static int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3371                               __u32 keylen, void *key, __u32 vallen, void *val,
3372                               struct ptlrpc_request_set *set)
3373 {
3374         struct lmv_tgt_desc *tgt;
3375         struct obd_device *obd;
3376         struct lmv_obd *lmv;
3377         int rc = 0;
3378         ENTRY;
3379
3380         obd = class_exp2obd(exp);
3381         if (obd == NULL) {
3382                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3383                        exp->exp_handle.h_cookie);
3384                 RETURN(-EINVAL);
3385         }
3386         lmv = &obd->u.lmv;
3387
3388         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3389             KEY_IS(KEY_DEFAULT_EASIZE)) {
3390                 int err = 0;
3391
3392                 lmv_foreach_connected_tgt(lmv, tgt) {
3393                         err = obd_set_info_async(env, tgt->ltd_exp,
3394                                                  keylen, key, vallen, val, set);
3395                         if (err && rc == 0)
3396                                 rc = err;
3397                 }
3398
3399                 RETURN(rc);
3400         }
3401
3402         RETURN(-EINVAL);
3403 }
3404
3405 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3406                             const struct lmv_mds_md_v1 *lmm1)
3407 {
3408         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3409         int             stripe_count;
3410         int             cplen;
3411         int             i;
3412         int             rc = 0;
3413         ENTRY;
3414
3415         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3416         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3417         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3418         if (CFS_FAIL_CHECK(OBD_FAIL_LMV_UNKNOWN_STRIPE))
3419                 lsm->lsm_md_hash_type = cfs_fail_val ?: LMV_HASH_TYPE_UNKNOWN;
3420         else
3421                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3422         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3423         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3424         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3425         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3426                         sizeof(lsm->lsm_md_pool_name));
3427
3428         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3429                 RETURN(-E2BIG);
3430
3431         CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3432                "layout_version %d\n", lsm->lsm_md_stripe_count,
3433                lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3434                lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3435                lsm->lsm_md_layout_version);
3436
3437         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3438         for (i = 0; i < stripe_count; i++) {
3439                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3440                               &lmm1->lmv_stripe_fids[i]);
3441                 /*
3442                  * set default value -1, so lmv_locate_tgt() knows this stripe
3443                  * target is not initialized.
3444                  */
3445                 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3446                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3447                         continue;
3448
3449                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3450                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3451                 if (rc == -ENOENT)
3452                         continue;
3453
3454                 if (rc)
3455                         RETURN(rc);
3456
3457                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3458                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3459         }
3460
3461         RETURN(rc);
3462 }
3463
3464 static inline int lmv_unpack_user_md(struct obd_export *exp,
3465                                      struct lmv_stripe_md *lsm,
3466                                      const struct lmv_user_md *lmu)
3467 {
3468         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3469         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3470         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3471         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3472         lsm->lsm_md_max_inherit = lmu->lum_max_inherit;
3473         lsm->lsm_md_max_inherit_rr = lmu->lum_max_inherit_rr;
3474         lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3475
3476         return 0;
3477 }
3478
3479 struct lmv_stripe_object *lmv_stripe_object_alloc(__u32 magic,
3480                                                   const union lmv_mds_md *lmm,
3481                                                   size_t lmm_size)
3482 {
3483         struct lmv_stripe_object *lsm_obj;
3484         int size;
3485
3486         if (magic == LMV_MAGIC_FOREIGN) {
3487                 struct lmv_foreign_md *lfm;
3488
3489                 size = offsetof(typeof(*lfm), lfm_value[0]);
3490                 if (lmm_size < size)
3491                         RETURN(ERR_PTR(-EPROTO));
3492
3493                 size += le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3494                 if (lmm_size < size)
3495                         RETURN(ERR_PTR(-EPROTO));
3496
3497                 OBD_ALLOC_LARGE(lsm_obj, lmm_size +
3498                                 offsetof(typeof(*lsm_obj), lso_lfm));
3499         } else {
3500                 if (magic == LMV_MAGIC_V1) {
3501                         int count;
3502
3503                         size = offsetof(struct lmv_mds_md_v1,
3504                                         lmv_stripe_fids[0]);
3505                         if (lmm_size < size)
3506                                 RETURN(ERR_PTR(-EPROTO));
3507
3508                         count = lmv_mds_md_stripe_count_get(lmm);
3509                         size += count * sizeof(struct lu_fid);
3510                         if (lmm_size < size)
3511                                 RETURN(ERR_PTR(-EPROTO));
3512
3513                         size = lmv_stripe_md_size(count);
3514                 } else {
3515                         if (lmm && lmm_size < sizeof(struct lmv_user_md))
3516                                 RETURN(ERR_PTR(-EPROTO));
3517
3518                         /**
3519                          * Unpack default dirstripe(lmv_user_md) to
3520                          * lmv_stripe_md, stripecount should be 0 then.
3521                          */
3522                         size = lmv_stripe_md_size(0);
3523                 }
3524                 size += offsetof(typeof(*lsm_obj), lso_lsm);
3525                 OBD_ALLOC(lsm_obj, size);
3526         }
3527
3528         if (lsm_obj) {
3529                 atomic_set(&lsm_obj->lso_refs, 1);
3530                 RETURN(lsm_obj);
3531         }
3532
3533         RETURN(ERR_PTR(-ENOMEM));
3534 }
3535 EXPORT_SYMBOL(lmv_stripe_object_alloc);
3536
3537 static int lmv_stripe_object_create(struct obd_export *exp,
3538                                     struct lmv_stripe_object **lsop,
3539                                     const union lmv_mds_md *lmm,
3540                                     size_t lmm_size)
3541 {
3542         struct lmv_stripe_object *lsm_obj;
3543         __u32 magic;
3544         int rc;
3545         ENTRY;
3546
3547         LASSERT(lsop != NULL && *lsop == NULL);
3548
3549         if (lmm_size == 0)
3550                 RETURN(-EPROTO);
3551
3552         magic = le32_to_cpu(lmm->lmv_magic);
3553         if (magic == LMV_MAGIC_STRIPE)
3554                 RETURN(-EPERM);
3555
3556         if (magic != LMV_MAGIC_V1 && magic != LMV_USER_MAGIC &&
3557             magic != LMV_MAGIC_FOREIGN) {
3558                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3559                        exp->exp_obd->obd_name, magic, -EIO);
3560                 RETURN(-EIO);
3561         }
3562
3563         /* foreign lmv case */
3564         if (magic == LMV_MAGIC_FOREIGN) {
3565                 struct lmv_foreign_md *lfm;
3566
3567                 lsm_obj = lmv_stripe_object_alloc(magic, lmm, lmm_size);
3568                 if (IS_ERR(lsm_obj))
3569                         RETURN(PTR_ERR(lsm_obj));
3570
3571                 *lsop = lsm_obj;
3572                 lfm = &lsm_obj->lso_lfm;
3573                 lfm->lfm_magic = magic;
3574                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3575                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3576                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3577                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3578                        lfm->lfm_length);
3579                 RETURN(0);
3580         }
3581
3582         /* Unpack memmd */
3583         lsm_obj = lmv_stripe_object_alloc(magic, lmm, lmm_size);
3584         if (IS_ERR(lsm_obj))
3585                 RETURN(PTR_ERR(lsm_obj));
3586
3587         switch (magic) {
3588         case LMV_MAGIC_V1:
3589                 rc = lmv_unpack_md_v1(exp, &lsm_obj->lso_lsm, &lmm->lmv_md_v1);
3590                 break;
3591         case LMV_USER_MAGIC:
3592                 rc = lmv_unpack_user_md(exp, &lsm_obj->lso_lsm,
3593                                         &lmm->lmv_user_md);
3594                 break;
3595         default:
3596                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3597                        magic);
3598                 rc = -EINVAL;
3599                 break;
3600         }
3601
3602         if (rc != 0)
3603                 lmv_stripe_object_put(&lsm_obj);
3604
3605         *lsop = lsm_obj;
3606         RETURN(rc);
3607 }
3608
3609 struct lmv_stripe_object *
3610 lmv_stripe_object_get(struct lmv_stripe_object *lsm_obj)
3611 {
3612         if (lsm_obj == NULL)
3613                 return NULL;
3614
3615         atomic_inc(&lsm_obj->lso_refs);
3616         CDEBUG(D_INODE, "get %p %u\n", lsm_obj,
3617                atomic_read(&lsm_obj->lso_refs));
3618         return lsm_obj;
3619 }
3620 EXPORT_SYMBOL(lmv_stripe_object_get);
3621
3622 void lmv_stripe_object_put(struct lmv_stripe_object **lsop)
3623 {
3624         struct lmv_stripe_object *lsm_obj;
3625         size_t size;
3626         int i;
3627
3628         LASSERT(lsop != NULL);
3629
3630         lsm_obj = *lsop;
3631         if (lsm_obj == NULL)
3632                 return;
3633
3634         *lsop = NULL;
3635         CDEBUG(D_INODE, "put %p %u\n", lsm_obj,
3636                atomic_read(&lsm_obj->lso_refs) - 1);
3637
3638         if (!atomic_dec_and_test(&lsm_obj->lso_refs))
3639                 return;
3640
3641         if (lmv_dir_foreign(lsm_obj)) {
3642                 size = lsm_obj->lso_lfm.lfm_length +
3643                         offsetof(typeof(lsm_obj->lso_lfm), lfm_value[0]) +
3644                         offsetof(typeof(*lsm_obj), lso_lsm);
3645                 OBD_FREE_LARGE(lsm_obj, size);
3646                 return;
3647         }
3648
3649         if (lmv_dir_striped(lsm_obj)) {
3650                 struct lmv_stripe_md *lsm = &lsm_obj->lso_lsm;
3651
3652                 for (i = 0; i < lsm->lsm_md_stripe_count; i++)
3653                         iput(lsm->lsm_md_oinfo[i].lmo_root);
3654                 size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3655         } else {
3656                 size = lmv_stripe_md_size(0);
3657         }
3658         OBD_FREE(lsm_obj, size + offsetof(typeof(*lsm_obj), lso_lsm));
3659 }
3660 EXPORT_SYMBOL(lmv_stripe_object_put);
3661
3662 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3663                              union ldlm_policy_data *policy,
3664                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3665                              void *opaque)
3666 {
3667         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3668         struct lu_tgt_desc *tgt;
3669         int err;
3670         int rc = 0;
3671
3672         ENTRY;
3673
3674         LASSERT(fid != NULL);
3675
3676         lmv_foreach_connected_tgt(lmv, tgt) {
3677                 if (!tgt->ltd_active)
3678                         continue;
3679
3680                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3681                                        opaque);
3682                 if (!rc)
3683                         rc = err;
3684         }
3685         RETURN(rc);
3686 }
3687
3688 static int lmv_set_lock_data(struct obd_export *exp,
3689                              const struct lustre_handle *lockh,
3690                              void *data, __u64 *bits)
3691 {
3692         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3693         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3694         int rc;
3695
3696         ENTRY;
3697
3698         if (tgt == NULL || tgt->ltd_exp == NULL)
3699                 RETURN(-EINVAL);
3700         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3701         RETURN(rc);
3702 }
3703
3704 static enum ldlm_mode
3705 lmv_lock_match(struct obd_export *exp, __u64 flags,
3706                const struct lu_fid *fid, enum ldlm_type type,
3707                union ldlm_policy_data *policy,
3708                enum ldlm_mode mode, struct lustre_handle *lockh)
3709 {
3710         struct obd_device *obd = exp->exp_obd;
3711         struct lmv_obd *lmv = &obd->u.lmv;
3712         struct lu_tgt_desc *tgt;
3713         __u64 bits = policy->l_inodebits.bits;
3714         enum ldlm_mode rc = LCK_MINMODE;
3715         int index;
3716         int i;
3717
3718         /* only one bit is set */
3719         LASSERT(bits && !(bits & (bits - 1)));
3720         /* With DNE every object can have two locks in different namespaces:
3721          * lookup lock in space of MDT storing direntry and update/open lock in
3722          * space of MDT storing inode.  Try the MDT that the FID maps to first,
3723          * since this can be easily found, and only try others if that fails.
3724          */
3725         if (bits == MDS_INODELOCK_LOOKUP) {
3726                 for (i = 0, index = lmv_fid2tgt_index(lmv, fid);
3727                      i < lmv->lmv_mdt_descs.ltd_tgts_size; i++,
3728                      index = (index + 1) % lmv->lmv_mdt_descs.ltd_tgts_size) {
3729                         if (index < 0) {
3730                                 CDEBUG(D_HA,
3731                                        "%s: "DFID" is inaccessible: rc = %d\n",
3732                                        obd->obd_name, PFID(fid), index);
3733                                 index = 0;
3734                         }
3735                         tgt = lmv_tgt(lmv, index);
3736                         if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
3737                                 continue;
3738                         rc = md_lock_match(tgt->ltd_exp, flags, fid, type,
3739                                            policy, mode, lockh);
3740                         if (rc)
3741                                 break;
3742                 }
3743         } else {
3744                 tgt = lmv_fid2tgt(lmv, fid);
3745                 if (!IS_ERR(tgt) && tgt->ltd_exp && tgt->ltd_active)
3746                         rc = md_lock_match(tgt->ltd_exp, flags, fid, type,
3747                                            policy, mode, lockh);
3748         }
3749
3750         CDEBUG(D_INODE, "Lock match for "DFID": %d\n", PFID(fid), rc);
3751
3752         return rc;
3753 }
3754
3755 static int
3756 lmv_get_lustre_md(struct obd_export *exp, struct req_capsule *pill,
3757                   struct obd_export *dt_exp, struct obd_export *md_exp,
3758                   struct lustre_md *md)
3759 {
3760         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3761         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3762
3763         if (!tgt || !tgt->ltd_exp)
3764                 return -EINVAL;
3765
3766         return md_get_lustre_md(tgt->ltd_exp, pill, dt_exp, md_exp, md);
3767 }
3768
3769 static int lmv_put_lustre_md(struct obd_export *exp, struct lustre_md *md)
3770 {
3771         struct obd_device *obd = exp->exp_obd;
3772         struct lmv_obd *lmv = &obd->u.lmv;
3773         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3774         ENTRY;
3775
3776         lmv_stripe_object_put(&md->def_lsm_obj);
3777         lmv_stripe_object_put(&md->lsm_obj);
3778
3779         if (!tgt || !tgt->ltd_exp)
3780                 RETURN(-EINVAL);
3781         RETURN(0);
3782 }
3783
3784 static int lmv_set_open_replay_data(struct obd_export *exp,
3785                                     struct obd_client_handle *och,
3786                                     struct lookup_intent *it)
3787 {
3788         struct obd_device *obd = exp->exp_obd;
3789         struct lmv_obd *lmv = &obd->u.lmv;
3790         struct lmv_tgt_desc *tgt;
3791
3792         ENTRY;
3793
3794         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3795         if (IS_ERR(tgt))
3796                 RETURN(PTR_ERR(tgt));
3797
3798         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3799 }
3800
3801 static int lmv_clear_open_replay_data(struct obd_export *exp,
3802                                       struct obd_client_handle *och)
3803 {
3804         struct obd_device *obd = exp->exp_obd;
3805         struct lmv_obd *lmv = &obd->u.lmv;
3806         struct lmv_tgt_desc *tgt;
3807
3808         ENTRY;
3809
3810         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3811         if (IS_ERR(tgt))
3812                 RETURN(PTR_ERR(tgt));
3813
3814         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3815 }
3816
3817 static int lmv_intent_getattr_async(struct obd_export *exp,
3818                                     struct md_op_item *item)
3819 {
3820         struct md_op_data *op_data = &item->mop_data;
3821         struct obd_device *obd = exp->exp_obd;
3822         struct lmv_obd *lmv = &obd->u.lmv;
3823         struct lmv_tgt_desc *ptgt;
3824         struct lmv_tgt_desc *ctgt;
3825         int rc;
3826
3827         ENTRY;
3828
3829         if (!fid_is_sane(&op_data->op_fid2))
3830                 RETURN(-EINVAL);
3831
3832         ptgt = lmv_locate_tgt(lmv, op_data);
3833         if (IS_ERR(ptgt))
3834                 RETURN(PTR_ERR(ptgt));
3835
3836         ctgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3837         if (IS_ERR(ctgt))
3838                 RETURN(PTR_ERR(ctgt));
3839
3840         /*
3841          * remote object needs two RPCs to lookup and getattr, considering the
3842          * complexity don't support statahead for now.
3843          */
3844         if (ctgt != ptgt)
3845                 RETURN(-EREMOTE);
3846
3847         rc = md_intent_getattr_async(ptgt->ltd_exp, item);
3848
3849         RETURN(rc);
3850 }
3851
3852 static int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3853                                struct lu_fid *fid, __u64 *bits)
3854 {
3855         struct obd_device *obd = exp->exp_obd;
3856         struct lmv_obd *lmv = &obd->u.lmv;
3857         struct lmv_tgt_desc *tgt;
3858         int rc;
3859
3860         ENTRY;
3861
3862         tgt = lmv_fid2tgt(lmv, fid);
3863         if (IS_ERR(tgt))
3864                 RETURN(PTR_ERR(tgt));
3865
3866         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3867         RETURN(rc);
3868 }
3869
3870 static int lmv_get_fid_from_lsm(struct obd_export *exp,
3871                                 const struct lmv_stripe_object *lso,
3872                                 const char *name, int namelen,
3873                                 struct lu_fid *fid)
3874 {
3875         const struct lmv_oinfo *oinfo;
3876
3877         LASSERT(lmv_dir_striped(lso));
3878         oinfo = lsm_name_to_stripe_info(lso, name, namelen, false);
3879         if (IS_ERR(oinfo))
3880                 return PTR_ERR(oinfo);
3881
3882         *fid = oinfo->lmo_fid;
3883
3884         RETURN(0);
3885 }
3886
3887 /**
3888  * For lmv, only need to send request to master MDT, and the master MDT will
3889  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3890  * we directly fetch data from the slave MDTs.
3891  */
3892 static int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3893                         struct obd_quotactl *oqctl)
3894 {
3895         struct obd_device *obd = class_exp2obd(exp);
3896         struct lmv_obd *lmv = &obd->u.lmv;
3897         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3898         __u64 curspace, curinodes;
3899         int rc = 0;
3900
3901         ENTRY;
3902
3903         if (!tgt || !tgt->ltd_exp || !tgt->ltd_active) {
3904                 CERROR("master lmv inactive\n");
3905                 RETURN(-EIO);
3906         }
3907
3908         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3909                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3910                 RETURN(rc);
3911         }
3912
3913         curspace = curinodes = 0;
3914         lmv_foreach_connected_tgt(lmv, tgt) {
3915                 int err;
3916
3917                 if (!tgt->ltd_active)
3918                         continue;
3919
3920                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3921                 if (err) {
3922                         CERROR("getquota on mdt %d failed. %d\n",
3923                                tgt->ltd_index, err);
3924                         if (!rc)
3925                                 rc = err;
3926                 } else {
3927                         curspace += oqctl->qc_dqblk.dqb_curspace;
3928                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3929                 }
3930         }
3931         oqctl->qc_dqblk.dqb_curspace = curspace;
3932         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3933
3934         RETURN(rc);
3935 }
3936
3937 static int lmv_merge_attr(struct obd_export *exp,
3938                           const struct lmv_stripe_object *lso,
3939                           struct cl_attr *attr,
3940                           ldlm_blocking_callback cb_blocking)
3941 {
3942         const struct lmv_stripe_md *lsm = &lso->lso_lsm;
3943         int rc;
3944         int i;
3945
3946         if (!lmv_dir_striped(lso))
3947                 return 0;
3948
3949         rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3950         if (rc < 0)
3951                 return rc;
3952
3953         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3954                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3955
3956                 if (!inode)
3957                         continue;
3958
3959                 CDEBUG(D_INFO,
3960                        "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
3961                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3962                        i_size_read(inode), (unsigned long long)inode->i_blocks,
3963                        inode->i_nlink, (s64)inode->i_atime.tv_sec,
3964                        (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
3965
3966                 /* for slave stripe, it needs to subtract nlink for . and .. */
3967                 if (i != 0)
3968                         attr->cat_nlink += inode->i_nlink - 2;
3969                 else
3970                         attr->cat_nlink = inode->i_nlink;
3971
3972                 attr->cat_size += i_size_read(inode);
3973                 attr->cat_blocks += inode->i_blocks;
3974
3975                 if (attr->cat_atime < inode->i_atime.tv_sec)
3976                         attr->cat_atime = inode->i_atime.tv_sec;
3977
3978                 if (attr->cat_ctime < inode->i_ctime.tv_sec)
3979                         attr->cat_ctime = inode->i_ctime.tv_sec;
3980
3981                 if (attr->cat_mtime < inode->i_mtime.tv_sec)
3982                         attr->cat_mtime = inode->i_mtime.tv_sec;
3983         }
3984         return 0;
3985 }
3986
3987 static struct lu_batch *lmv_batch_create(struct obd_export *exp,
3988                                          enum lu_batch_flags flags,
3989                                          __u32 max_count)
3990 {
3991         struct lu_batch *bh;
3992         struct lmv_batch *lbh;
3993
3994         ENTRY;
3995         OBD_ALLOC_PTR(lbh);
3996         if (!lbh)
3997                 RETURN(ERR_PTR(-ENOMEM));
3998
3999         bh = &lbh->lbh_super;
4000         bh->lbt_flags = flags;
4001         bh->lbt_max_count = max_count;
4002
4003         if (flags & BATCH_FL_RQSET) {
4004                 bh->lbt_rqset = ptlrpc_prep_set();
4005                 if (bh->lbt_rqset == NULL) {
4006                         OBD_FREE_PTR(lbh);
4007                         RETURN(ERR_PTR(-ENOMEM));
4008                 }
4009         }
4010
4011         INIT_LIST_HEAD(&lbh->lbh_sub_batch_list);
4012         RETURN(bh);
4013 }
4014
4015 static int lmv_batch_stop(struct obd_export *exp, struct lu_batch *bh)
4016 {
4017         struct lmv_batch *lbh;
4018         struct lmvsub_batch *sub;
4019         struct lmvsub_batch *tmp;
4020         int rc = 0;
4021
4022         ENTRY;
4023
4024         lbh = container_of(bh, struct lmv_batch, lbh_super);
4025         list_for_each_entry_safe(sub, tmp, &lbh->lbh_sub_batch_list,
4026                                  sbh_sub_item) {
4027                 list_del(&sub->sbh_sub_item);
4028                 rc = md_batch_stop(sub->sbh_tgt->ltd_exp, sub->sbh_sub);
4029                 if (rc < 0) {
4030                         CERROR("%s: stop batch processing failed: rc = %d\n",
4031                                exp->exp_obd->obd_name, rc);
4032                         if (bh->lbt_result == 0)
4033                                 bh->lbt_result = rc;
4034                 }
4035                 OBD_FREE_PTR(sub);
4036         }
4037
4038         if (bh->lbt_flags & BATCH_FL_RQSET) {
4039                 rc = ptlrpc_set_wait(NULL, bh->lbt_rqset);
4040                 ptlrpc_set_destroy(bh->lbt_rqset);
4041         }
4042
4043         OBD_FREE_PTR(lbh);
4044         RETURN(rc);
4045 }
4046
4047 static int lmv_batch_flush(struct obd_export *exp, struct lu_batch *bh,
4048                            bool wait)
4049 {
4050         struct lmv_batch *lbh;
4051         struct lmvsub_batch *sub;
4052         int rc = 0;
4053         int rc1;
4054
4055         ENTRY;
4056
4057         lbh = container_of(bh, struct lmv_batch, lbh_super);
4058         list_for_each_entry(sub, &lbh->lbh_sub_batch_list, sbh_sub_item) {
4059                 rc1 = md_batch_flush(sub->sbh_tgt->ltd_exp, sub->sbh_sub, wait);
4060                 if (rc1 < 0) {
4061                         CERROR("%s: stop batch processing failed: rc = %d\n",
4062                                exp->exp_obd->obd_name, rc);
4063                         if (bh->lbt_result == 0)
4064                                 bh->lbt_result = rc;
4065
4066                         if (rc == 0)
4067                                 rc = rc1;
4068                 }
4069         }
4070
4071         if (wait && bh->lbt_flags & BATCH_FL_RQSET) {
4072                 rc1 = ptlrpc_set_wait(NULL, bh->lbt_rqset);
4073                 if (rc == 0)
4074                         rc = rc1;
4075         }
4076
4077         RETURN(rc);
4078 }
4079
4080 static inline struct lmv_tgt_desc *
4081 lmv_batch_locate_tgt(struct lmv_obd *lmv, struct md_op_item *item)
4082 {
4083         struct md_op_data *op_data = &item->mop_data;
4084         struct lmv_tgt_desc *tgt;
4085
4086         switch (item->mop_opc) {
4087         case MD_OP_GETATTR: {
4088                 struct lmv_tgt_desc *ptgt;
4089
4090                 if (!fid_is_sane(&op_data->op_fid2))
4091                         RETURN(ERR_PTR(-EINVAL));
4092
4093                 ptgt = lmv_locate_tgt(lmv, op_data);
4094                 if (IS_ERR(ptgt))
4095                         RETURN(ptgt);
4096
4097                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
4098                 if (IS_ERR(tgt))
4099                         RETURN(tgt);
4100
4101                 /*
4102                  * Remote object needs two RPCs to lookup and getattr,
4103                  * considering the complexity don't support statahead for now.
4104                  */
4105                 if (tgt != ptgt)
4106                         RETURN(ERR_PTR(-EREMOTE));
4107
4108                 break;
4109         }
4110         default:
4111                 tgt = ERR_PTR(-ENOTSUPP);
4112         }
4113
4114         return tgt;
4115 }
4116
4117 struct lu_batch *lmv_batch_lookup_sub(struct lmv_batch *lbh,
4118                                       struct lmv_tgt_desc *tgt)
4119 {
4120         struct lmvsub_batch *sub;
4121
4122         list_for_each_entry(sub, &lbh->lbh_sub_batch_list, sbh_sub_item) {
4123                 if (sub->sbh_tgt == tgt)
4124                         return sub->sbh_sub;
4125         }
4126
4127         return NULL;
4128 }
4129
4130 struct lu_batch *lmv_batch_get_sub(struct lmv_batch *lbh,
4131                                    struct lmv_tgt_desc *tgt)
4132 {
4133         struct lmvsub_batch *sbh;
4134         struct lu_batch *child_bh;
4135         struct lu_batch *bh;
4136
4137         ENTRY;
4138
4139         child_bh = lmv_batch_lookup_sub(lbh, tgt);
4140         if (child_bh != NULL)
4141                 RETURN(child_bh);
4142
4143         OBD_ALLOC_PTR(sbh);
4144         if (sbh == NULL)
4145                 RETURN(ERR_PTR(-ENOMEM));
4146
4147         INIT_LIST_HEAD(&sbh->sbh_sub_item);
4148         sbh->sbh_tgt = tgt;
4149
4150         bh = &lbh->lbh_super;
4151         child_bh = md_batch_create(tgt->ltd_exp, bh->lbt_flags,
4152                                    bh->lbt_max_count);
4153         if (IS_ERR(child_bh)) {
4154                 OBD_FREE_PTR(sbh);
4155                 RETURN(child_bh);
4156         }
4157
4158         child_bh->lbt_rqset = bh->lbt_rqset;
4159         sbh->sbh_sub = child_bh;
4160         list_add(&sbh->sbh_sub_item, &lbh->lbh_sub_batch_list);
4161         RETURN(child_bh);
4162 }
4163
4164 static int lmv_batch_add(struct obd_export *exp, struct lu_batch *bh,
4165                          struct md_op_item *item)
4166 {
4167         struct obd_device *obd = exp->exp_obd;
4168         struct lmv_obd *lmv = &obd->u.lmv;
4169         struct lmv_tgt_desc *tgt;
4170         struct lmv_batch *lbh;
4171         struct lu_batch *child_bh;
4172         int rc;
4173
4174         ENTRY;
4175
4176         tgt = lmv_batch_locate_tgt(lmv, item);
4177         if (IS_ERR(tgt))
4178                 RETURN(PTR_ERR(tgt));
4179
4180         lbh = container_of(bh, struct lmv_batch, lbh_super);
4181         child_bh = lmv_batch_get_sub(lbh, tgt);
4182         if (IS_ERR(child_bh))
4183                 RETURN(PTR_ERR(child_bh));
4184
4185         rc = md_batch_add(tgt->ltd_exp, child_bh, item);
4186         RETURN(rc);
4187 }
4188
4189 static const struct obd_ops lmv_obd_ops = {
4190         .o_owner                = THIS_MODULE,
4191         .o_setup                = lmv_setup,
4192         .o_cleanup              = lmv_cleanup,
4193         .o_precleanup           = lmv_precleanup,
4194         .o_process_config       = lmv_process_config,
4195         .o_connect              = lmv_connect,
4196         .o_disconnect           = lmv_disconnect,
4197         .o_statfs               = lmv_statfs,
4198         .o_get_info             = lmv_get_info,
4199         .o_set_info_async       = lmv_set_info_async,
4200         .o_notify               = lmv_notify,
4201         .o_get_uuid             = lmv_get_uuid,
4202         .o_fid_alloc            = lmv_fid_alloc,
4203         .o_iocontrol            = lmv_iocontrol,
4204         .o_quotactl             = lmv_quotactl
4205 };
4206
4207 static const struct md_ops lmv_md_ops = {
4208         .m_get_root             = lmv_get_root,
4209         .m_null_inode           = lmv_null_inode,
4210         .m_close                = lmv_close,
4211         .m_create               = lmv_create,
4212         .m_enqueue              = lmv_enqueue,
4213         .m_getattr              = lmv_getattr,
4214         .m_getxattr             = lmv_getxattr,
4215         .m_getattr_name         = lmv_getattr_name,
4216         .m_intent_lock          = lmv_intent_lock,
4217         .m_link                 = lmv_link,
4218         .m_rename               = lmv_rename,
4219         .m_setattr              = lmv_setattr,
4220         .m_setxattr             = lmv_setxattr,
4221         .m_fsync                = lmv_fsync,
4222         .m_file_resync          = lmv_file_resync,
4223         .m_read_page            = lmv_read_page,
4224         .m_unlink               = lmv_unlink,
4225         .m_init_ea_size         = lmv_init_ea_size,
4226         .m_cancel_unused        = lmv_cancel_unused,
4227         .m_set_lock_data        = lmv_set_lock_data,
4228         .m_lock_match           = lmv_lock_match,
4229         .m_get_lustre_md        = lmv_get_lustre_md,
4230         .m_put_lustre_md        = lmv_put_lustre_md,
4231         .m_merge_attr           = lmv_merge_attr,
4232         .m_set_open_replay_data = lmv_set_open_replay_data,
4233         .m_clear_open_replay_data = lmv_clear_open_replay_data,
4234         .m_intent_getattr_async = lmv_intent_getattr_async,
4235         .m_revalidate_lock      = lmv_revalidate_lock,
4236         .m_get_fid_from_lsm     = lmv_get_fid_from_lsm,
4237         .m_stripe_object_create = lmv_stripe_object_create,
4238         .m_rmfid                = lmv_rmfid,
4239         .m_batch_create         = lmv_batch_create,
4240         .m_batch_add            = lmv_batch_add,
4241         .m_batch_stop           = lmv_batch_stop,
4242         .m_batch_flush          = lmv_batch_flush,
4243 };
4244
4245 static int __init lmv_init(void)
4246 {
4247         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true,
4248                                    LUSTRE_LMV_NAME, NULL);
4249 }
4250
4251 static void __exit lmv_exit(void)
4252 {
4253         class_unregister_type(LUSTRE_LMV_NAME);
4254 }
4255
4256 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
4257 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
4258 MODULE_VERSION(LUSTRE_VERSION_STRING);
4259 MODULE_LICENSE("GPL");
4260
4261 module_init(lmv_init);
4262 module_exit(lmv_exit);