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                kref_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 void lmv_statfs_update(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
 567                               struct obd_statfs *osfs)
 568 {
 569         spin_lock(&lmv->lmv_lock);
 570         tgt->ltd_statfs = *osfs;
 571         tgt->ltd_statfs_age = ktime_get_seconds();
 572         spin_unlock(&lmv->lmv_lock);
 573         set_bit(LQ_DIRTY, &lmv->lmv_qos.lq_flags);
 574 }
 575
 576 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
 577                         void __user *uarg)
 578 {
 579         struct obd_device *obd = class_exp2obd(exp);
 580         struct lmv_obd *lmv = &obd->u.lmv;
 581         struct getinfo_fid2path *gf;
 582         struct lmv_tgt_desc *tgt;
 583         struct getinfo_fid2path *remote_gf = NULL;
 584         struct lu_fid root_fid;
 585         int remote_gf_size = 0;
 586         int currentisenc = 0;
 587         int globalisenc = 0;
 588         int rc;
 589
 590         gf = karg;
 591         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 592         if (IS_ERR(tgt))
 593                 RETURN(PTR_ERR(tgt));
 594
 595         root_fid = *gf->gf_u.gf_root_fid;
 596         LASSERT(fid_is_sane(&root_fid));
 597
 598 repeat_fid2path:
 599         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
 600         if (rc != 0 && rc != -EREMOTE)
 601                 GOTO(out_fid2path, rc);
 602
 603         if (gf->gf_u.gf_path[0] == '/') {
 604                 /* by convention, server side (mdt_path_current()) puts
 605                  * a leading '/' to tell client that we are dealing with
 606                  * an encrypted file
 607                  */
 608                 currentisenc = 1;
 609                 globalisenc = 1;
 610         } else {
 611                 currentisenc = 0;
 612         }
 613
 614         /* If remote_gf != NULL, it means just building the
 615          * path on the remote MDT, copy this path segment to gf.
 616          */
 617         if (remote_gf != NULL) {
 618                 struct getinfo_fid2path *ori_gf;
 619                 int oldisenc = 0;
 620                 char *ptr;
 621                 int len;
 622
 623                 ori_gf = (struct getinfo_fid2path *)karg;
 624                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
 625                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
 626                         GOTO(out_fid2path, rc = -EOVERFLOW);
 627
 628                 ptr = ori_gf->gf_u.gf_path;
 629                 oldisenc = ptr[0] == '/';
 630
 631                 len = strlen(gf->gf_u.gf_path);
 632                 if (len) {
 633                         /* move the current path to the right to release space
 634                          * for closer-to-root part
 635                          */
 636                         memmove(ptr + len - currentisenc + 1 + globalisenc,
 637                                 ptr + oldisenc,
 638                                 strlen(ori_gf->gf_u.gf_path) - oldisenc + 1);
 639                         if (globalisenc)
 640                                 *(ptr++) = '/';
 641                         memcpy(ptr, gf->gf_u.gf_path + currentisenc,
 642                                len - currentisenc);
 643                         ptr[len - currentisenc] = '/';
 644                 }
 645         }
 646
 647         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
 648                tgt->ltd_exp->exp_obd->obd_name,
 649                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
 650                gf->gf_linkno);
 651
 652         if (rc == 0)
 653                 GOTO(out_fid2path, rc);
 654
 655         /* sigh, has to go to another MDT to do path building further */
 656         if (remote_gf == NULL) {
 657                 remote_gf_size = sizeof(*remote_gf) + len - sizeof(*gf);
 658                 OBD_ALLOC(remote_gf, remote_gf_size);
 659                 if (remote_gf == NULL)
 660                         GOTO(out_fid2path, rc = -ENOMEM);
 661                 remote_gf->gf_pathlen = len - sizeof(*gf);
 662         }
 663
 664         if (!fid_is_sane(&gf->gf_fid)) {
 665                 CERROR("%s: invalid FID "DFID": rc = %d\n",
 666                        tgt->ltd_exp->exp_obd->obd_name,
 667                        PFID(&gf->gf_fid), -EINVAL);
 668                 GOTO(out_fid2path, rc = -EINVAL);
 669         }
 670
 671         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 672         if (IS_ERR(tgt))
 673                 GOTO(out_fid2path, rc = -EINVAL);
 674
 675         remote_gf->gf_fid = gf->gf_fid;
 676         remote_gf->gf_recno = -1;
 677         remote_gf->gf_linkno = -1;
 678         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
 679         *remote_gf->gf_u.gf_root_fid = root_fid;
 680         gf = remote_gf;
 681         goto repeat_fid2path;
 682
 683 out_fid2path:
 684         if (remote_gf != NULL)
 685                 OBD_FREE(remote_gf, remote_gf_size);
 686         RETURN(rc);
 687 }
 688
 689 static int lmv_hsm_req_count(struct lmv_obd *lmv,
 690                              const struct hsm_user_request *hur,
 691                              const struct lmv_tgt_desc *tgt_mds)
 692 {
 693         struct lmv_tgt_desc *curr_tgt;
 694         __u32 i;
 695         int nr = 0;
 696
 697         /* count how many requests must be sent to the given target */
 698         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
 699                 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
 700                 if (IS_ERR(curr_tgt))
 701                         RETURN(PTR_ERR(curr_tgt));
 702                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
 703                         nr++;
 704         }
 705         return nr;
 706 }
 707
 708 static int lmv_hsm_req_build(struct lmv_obd *lmv,
 709                               struct hsm_user_request *hur_in,
 710                               const struct lmv_tgt_desc *tgt_mds,
 711                               struct hsm_user_request *hur_out)
 712 {
 713         __u32 i, nr_out;
 714         struct lmv_tgt_desc *curr_tgt;
 715
 716         /* build the hsm_user_request for the given target */
 717         hur_out->hur_request = hur_in->hur_request;
 718         nr_out = 0;
 719         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
 720                 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
 721                 if (IS_ERR(curr_tgt))
 722                         RETURN(PTR_ERR(curr_tgt));
 723                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
 724                         hur_out->hur_user_item[nr_out] =
 725                                                 hur_in->hur_user_item[i];
 726                         nr_out++;
 727                 }
 728         }
 729         hur_out->hur_request.hr_itemcount = nr_out;
 730         memcpy(hur_data(hur_out), hur_data(hur_in),
 731                hur_in->hur_request.hr_data_len);
 732
 733         RETURN(0);
 734 }
 735
 736 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
 737                                  int len, struct lustre_kernelcomm *lk,
 738                                  void __user *uarg)
 739 {
 740         struct lmv_obd *lmv = &obd->u.lmv;
 741         struct lu_tgt_desc *tgt;
 742         int rc;
 743
 744         ENTRY;
 745
 746         /* unregister request (call from llapi_hsm_copytool_fini) */
 747         lmv_foreach_connected_tgt(lmv, tgt)
 748                 /* best effort: try to clean as much as possible
 749                  * (continue on error) */
 750                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 751
 752         /* Whatever the result, remove copytool from kuc groups.
 753          * Unreached coordinators will get EPIPE on next requests
 754          * and will unregister automatically.
 755          */
 756         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 757
 758         RETURN(rc);
 759 }
 760
 761 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
 762                                int len, struct lustre_kernelcomm *lk,
 763                                void __user *uarg)
 764 {
 765         struct lmv_obd *lmv = &obd->u.lmv;
 766         struct file *filp;
 767         bool any_set = false;
 768         struct kkuc_ct_data *kcd;
 769         size_t kcd_size;
 770         struct lu_tgt_desc *tgt;
 771         __u32 i;
 772         int err;
 773         int rc = 0;
 774
 775         ENTRY;
 776
 777         filp = fget(lk->lk_wfd);
 778         if (!filp)
 779                 RETURN(-EBADF);
 780
 781         if (lk->lk_flags & LK_FLG_DATANR)
 782                 kcd_size = offsetof(struct kkuc_ct_data,
 783                                     kcd_archives[lk->lk_data_count]);
 784         else
 785                 kcd_size = sizeof(*kcd);
 786
 787         OBD_ALLOC(kcd, kcd_size);
 788         if (kcd == NULL)
 789                 GOTO(err_fput, rc = -ENOMEM);
 790
 791         kcd->kcd_nr_archives = lk->lk_data_count;
 792         if (lk->lk_flags & LK_FLG_DATANR) {
 793                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
 794                 if (lk->lk_data_count > 0)
 795                         memcpy(kcd->kcd_archives, lk->lk_data,
 796                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
 797         } else {
 798                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
 799         }
 800
 801         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
 802                                    lk->lk_group, kcd, kcd_size);
 803         OBD_FREE(kcd, kcd_size);
 804         if (rc)
 805                 GOTO(err_fput, rc);
 806
 807         /* All or nothing: try to register to all MDS.
 808          * In case of failure, unregister from previous MDS,
 809          * except if it because of inactive target. */
 810         lmv_foreach_connected_tgt(lmv, tgt) {
 811                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 812                 if (err) {
 813                         if (tgt->ltd_active) {
 814                                 /* permanent error */
 815                                 CERROR("%s: iocontrol MDC %s on MDT"
 816                                        " idx %d cmd %x: err = %d\n",
 817                                        lmv2obd_dev(lmv)->obd_name,
 818                                        tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
 819                                        err);
 820                                 rc = err;
 821                                 lk->lk_flags |= LK_FLG_STOP;
 822                                 i = tgt->ltd_index;
 823                                 /* unregister from previous MDS */
 824                                 lmv_foreach_connected_tgt(lmv, tgt) {
 825                                         if (tgt->ltd_index >= i)
 826                                                 break;
 827
 828                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
 829                                                       lk, uarg);
 830                                 }
 831                                 GOTO(err_kkuc_rem, rc);
 832                         }
 833                         /* else: transient error.
 834                          * kuc will register to the missing MDT
 835                          * when it is back */
 836                 } else {
 837                         any_set = true;
 838                 }
 839         }
 840
 841         if (!any_set)
 842                 /* no registration done: return error */
 843                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
 844
 845         RETURN(0);
 846
 847 err_kkuc_rem:
 848         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 849
 850 err_fput:
 851         fput(filp);
 852         return rc;
 853 }
 854
 855 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 856                          int len, void *karg, void __user *uarg)
 857 {
 858         struct obd_device *obd = class_exp2obd(exp);
 859         struct lmv_obd *lmv = &obd->u.lmv;
 860         struct lu_tgt_desc *tgt = NULL;
 861         int set = 0;
 862         __u32 count = lmv->lmv_mdt_count;
 863         int rc = 0;
 864
 865         ENTRY;
 866         CDEBUG(D_IOCTL, "%s: cmd=%x len=%u karg=%pK uarg=%pK\n",
 867                exp->exp_obd->obd_name, cmd, len, karg, uarg);
 868         if (count == 0)
 869                 RETURN(-ENOTTY);
 870
 871         /* exit early for unknown ioctl types */
 872         if (unlikely(_IOC_TYPE(cmd) != 'f' && !IOC_OSC_SET_ACTIVE_ALLOW(cmd)))
 873                 RETURN(OBD_IOC_ERROR(obd->obd_name, cmd, "unknown", -ENOTTY));
 874
 875         /* handle commands that don't use @karg first */
 876         switch (cmd) {
 877         case LL_IOC_GET_CONNECT_FLAGS:
 878                 tgt = lmv_tgt(lmv, 0);
 879                 rc = -ENODATA;
 880                 if (tgt && tgt->ltd_exp)
 881                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, NULL, uarg);
 882                 RETURN(rc);
 883         }
 884
 885         if (unlikely(karg == NULL))
 886                 RETURN(OBD_IOC_ERROR(obd->obd_name, cmd, "karg=NULL", -EINVAL));
 887
 888         switch (cmd) {
 889         case IOC_OBD_STATFS: {
 890                 struct obd_ioctl_data *data = karg;
 891                 struct obd_device *mdc_obd;
 892                 struct obd_statfs stat_buf = {0};
 893                 __u32 index;
 894
 895                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 896
 897                 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
 898                         RETURN(-ENODEV);
 899
 900                 tgt = lmv_tgt(lmv, index);
 901                 if (!tgt)
 902                         RETURN(-EAGAIN);
 903
 904                 if (!tgt->ltd_active)
 905                         RETURN(-ENODATA);
 906
 907                 mdc_obd = class_exp2obd(tgt->ltd_exp);
 908                 if (!mdc_obd)
 909                         RETURN(-EINVAL);
 910
 911                 /* copy UUID */
 912                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 913                                  min((int) data->ioc_plen2,
 914                                      (int) sizeof(struct obd_uuid))))
 915                         RETURN(-EFAULT);
 916
 917                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
 918                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
 919                                 0);
 920                 if (rc)
 921                         RETURN(rc);
 922                 lmv_statfs_update(lmv, tgt, &stat_buf);
 923                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
 924                                  min_t(int, data->ioc_plen1, sizeof(stat_buf))))
 925                         RETURN(-EFAULT);
 926                 break;
 927         }
 928         case OBD_IOC_QUOTACTL: {
 929                 struct if_quotactl *qctl = karg;
 930                 struct obd_quotactl *oqctl;
 931                 struct obd_import *imp;
 932
 933                 if (qctl->qc_valid == QC_MDTIDX) {
 934                         tgt = lmv_tgt(lmv, qctl->qc_idx);
 935                 } else if (qctl->qc_valid == QC_UUID) {
 936                         lmv_foreach_tgt(lmv, tgt) {
 937                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 938                                                      &qctl->obd_uuid))
 939                                         continue;
 940
 941                                 if (!tgt->ltd_exp)
 942                                         RETURN(-EINVAL);
 943
 944                                 break;
 945                         }
 946                 } else {
 947                         RETURN(-EINVAL);
 948                 }
 949
 950                 if (!tgt)
 951                         RETURN(-ENODEV);
 952
 953                 if (!tgt->ltd_exp)
 954                         RETURN(-EINVAL);
 955
 956                 imp = class_exp2cliimp(tgt->ltd_exp);
 957                 if (!tgt->ltd_active && imp->imp_state != LUSTRE_IMP_IDLE) {
 958                         qctl->qc_valid = QC_MDTIDX;
 959                         qctl->obd_uuid = tgt->ltd_uuid;
 960                         RETURN(-ENODATA);
 961                 }
 962
 963                 OBD_ALLOC_PTR(oqctl);
 964                 if (!oqctl)
 965                         RETURN(-ENOMEM);
 966
 967                 QCTL_COPY(oqctl, qctl);
 968                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 969                 if (rc == 0) {
 970                         QCTL_COPY_NO_PNAME(qctl, oqctl);
 971                         qctl->qc_valid = QC_MDTIDX;
 972                         qctl->obd_uuid = tgt->ltd_uuid;
 973                 }
 974                 OBD_FREE_PTR(oqctl);
 975                 break;
 976         }
 977         case LL_IOC_FID2MDTIDX: {
 978                 struct lu_fid *fid = karg;
 979                 int             mdt_index;
 980
 981                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
 982                 if (rc != 0)
 983                         RETURN(rc);
 984
 985                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
 986                  * point to user space memory for FID2MDTIDX. */
 987                 *(__u32 *)uarg = mdt_index;
 988                 break;
 989         }
 990         case OBD_IOC_FID2PATH: {
 991                 rc = lmv_fid2path(exp, len, karg, uarg);
 992                 break;
 993         }
 994         case LL_IOC_HSM_STATE_GET:
 995         case LL_IOC_HSM_STATE_SET:
 996         case LL_IOC_HSM_ACTION: {
 997                 struct md_op_data *op_data = karg;
 998
 999                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1000                 if (IS_ERR(tgt))
1001                         RETURN(PTR_ERR(tgt));
1002
1003                 if (tgt->ltd_exp == NULL)
1004                         RETURN(-EINVAL);
1005
1006                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1007                 break;
1008         }
1009         case LL_IOC_HSM_PROGRESS: {
1010                 const struct hsm_progress_kernel *hpk = karg;
1011
1012                 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
1013                 if (IS_ERR(tgt))
1014                         RETURN(PTR_ERR(tgt));
1015                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1016                 break;
1017         }
1018         case LL_IOC_HSM_REQUEST: {
1019                 struct hsm_user_request *hur = karg;
1020                 unsigned int reqcount = hur->hur_request.hr_itemcount;
1021
1022                 if (reqcount == 0)
1023                         RETURN(0);
1024
1025                 /* if the request is about a single fid
1026                  * or if there is a single MDS, no need to split
1027                  * the request. */
1028                 if (reqcount == 1 || count == 1) {
1029                         tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
1030                         if (IS_ERR(tgt))
1031                                 RETURN(PTR_ERR(tgt));
1032                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1033                 } else {
1034                         /* split fid list to their respective MDS */
1035                         lmv_foreach_connected_tgt(lmv, tgt) {
1036                                 int nr, rc1;
1037                                 size_t reqlen;
1038                                 struct hsm_user_request *req;
1039
1040                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
1041                                 if (nr < 0)
1042                                         RETURN(nr);
1043                                 if (nr == 0) /* nothing for this MDS */
1044                                         continue;
1045
1046                                 /* build a request with fids for this MDS */
1047                                 reqlen = offsetof(typeof(*hur),
1048                                                   hur_user_item[nr])
1049                                                 + hur->hur_request.hr_data_len;
1050                                 OBD_ALLOC_LARGE(req, reqlen);
1051                                 if (req == NULL)
1052                                         RETURN(-ENOMEM);
1053                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
1054                                 if (rc1 < 0)
1055                                         GOTO(hsm_req_err, rc1);
1056                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
1057                                                     req, uarg);
1058 hsm_req_err:
1059                                 if (rc1 != 0 && rc == 0)
1060                                         rc = rc1;
1061                                 OBD_FREE_LARGE(req, reqlen);
1062                         }
1063                 }
1064                 break;
1065         }
1066         case LL_IOC_LOV_SWAP_LAYOUTS: {
1067                 struct md_op_data *op_data = karg;
1068                 struct lmv_tgt_desc *tgt1, *tgt2;
1069
1070                 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1071                 if (IS_ERR(tgt1))
1072                         RETURN(PTR_ERR(tgt1));
1073
1074                 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1075                 if (IS_ERR(tgt2))
1076                         RETURN(PTR_ERR(tgt2));
1077
1078                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1079                         RETURN(-EINVAL);
1080
1081                 /* only files on same MDT can have their layouts swapped */
1082                 if (tgt1->ltd_index != tgt2->ltd_index)
1083                         RETURN(-EPERM);
1084
1085                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1086                 break;
1087         }
1088         case LL_IOC_HSM_CT_START: {
1089                 struct lustre_kernelcomm *lk = karg;
1090
1091                 if (lk->lk_flags & LK_FLG_STOP)
1092                         rc = lmv_hsm_ct_unregister(obd, cmd, len, lk, uarg);
1093                 else
1094                         rc = lmv_hsm_ct_register(obd, cmd, len, lk, uarg);
1095                 break;
1096         }
1097         default:
1098                 lmv_foreach_connected_tgt(lmv, tgt) {
1099                         struct obd_device *mdc_obd;
1100                         int err;
1101
1102                         /* ll_umount_begin() sets force flag but for lmv, not
1103                          * mdc. Let's pass it through */
1104                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1105                         mdc_obd->obd_force = obd->obd_force;
1106                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1107                         if (err) {
1108                                 if (tgt->ltd_active) {
1109                                         OBD_IOC_ERROR(obd->obd_name, cmd,
1110                                                       tgt->ltd_uuid.uuid, err);
1111                                         if (!rc)
1112                                                 rc = err;
1113                                         if (unlikely(err == -ENOTTY))
1114                                                 break;
1115                                 }
1116                         } else {
1117                                 set = 1;
1118                         }
1119                 }
1120                 if (!set && !rc)
1121                         rc = -EIO;
1122                 break;
1123         }
1124         RETURN(rc);
1125 }
1126
1127 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1128                   struct lu_fid *fid, struct md_op_data *op_data)
1129 {
1130         struct obd_device *obd = class_exp2obd(exp);
1131         struct lmv_obd *lmv = &obd->u.lmv;
1132         struct lmv_tgt_desc *tgt;
1133         int rc;
1134
1135         ENTRY;
1136
1137         LASSERT(op_data);
1138         LASSERT(fid);
1139
1140         tgt = lmv_tgt(lmv, op_data->op_mds);
1141         if (!tgt)
1142                 RETURN(-ENODEV);
1143
1144         if (!tgt->ltd_active || !tgt->ltd_exp)
1145                 RETURN(-ENODEV);
1146
1147         /*
1148          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1149          * on server that seq in new allocated fid is not yet known.
1150          */
1151         mutex_lock(&tgt->ltd_fid_mutex);
1152         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1153         mutex_unlock(&tgt->ltd_fid_mutex);
1154         if (rc > 0) {
1155                 LASSERT(fid_is_sane(fid));
1156                 rc = 0;
1157         }
1158
1159         RETURN(rc);
1160 }
1161
1162 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1163 {
1164         struct lmv_obd *lmv = &obd->u.lmv;
1165         struct lmv_desc *desc;
1166         struct lnet_processid lnet_id;
1167         int i = 0;
1168         int rc;
1169
1170         ENTRY;
1171
1172         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1173                 CERROR("LMV setup requires a descriptor\n");
1174                 RETURN(-EINVAL);
1175         }
1176
1177         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1178         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1179                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1180                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1181                 RETURN(-EINVAL);
1182         }
1183
1184         obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1185                      desc->ld_uuid.uuid);
1186         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1187         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1188         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1189                 LMV_DESC_QOS_MAXAGE_DEFAULT;
1190         lmv->max_def_easize = 0;
1191         lmv->max_easize = 0;
1192
1193         spin_lock_init(&lmv->lmv_lock);
1194
1195         /*
1196          * initialize rr_index to lower 32bit of netid, so that client
1197          * can distribute subdirs evenly from the beginning.
1198          */
1199         while (LNetGetId(i++, &lnet_id, true) != -ENOENT) {
1200                 if (!nid_is_lo0(&lnet_id.nid)) {
1201                         lmv->lmv_qos_rr_index = nidhash(&lnet_id.nid);
1202                         break;
1203                 }
1204         }
1205
1206         rc = lmv_tunables_init(obd);
1207         if (rc)
1208                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1209                       obd->obd_name, rc);
1210
1211         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1212                              LUSTRE_CLI_FLD_HASH_DHT);
1213         if (rc)
1214                 CERROR("Can't init FLD, err %d\n", rc);
1215
1216         rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1217         if (rc)
1218                 CWARN("%s: error initialize target table: rc = %d\n",
1219                       obd->obd_name, rc);
1220
1221         RETURN(rc);
1222 }
1223
1224 static int lmv_cleanup(struct obd_device *obd)
1225 {
1226         struct lmv_obd *lmv = &obd->u.lmv;
1227         struct lu_tgt_desc *tgt;
1228         struct lu_tgt_desc *tmp;
1229
1230         ENTRY;
1231
1232         fld_client_fini(&lmv->lmv_fld);
1233         fld_client_debugfs_fini(&lmv->lmv_fld);
1234
1235         lprocfs_obd_cleanup(obd);
1236         lprocfs_free_md_stats(obd);
1237
1238         lmv_foreach_tgt_safe(lmv, tgt, tmp)
1239                 lmv_del_target(lmv, tgt);
1240         lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1241
1242         RETURN(0);
1243 }
1244
1245 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1246 {
1247         struct lustre_cfg       *lcfg = buf;
1248         struct obd_uuid         obd_uuid;
1249         int                     gen;
1250         __u32                   index;
1251         int                     rc;
1252         ENTRY;
1253
1254         switch (lcfg->lcfg_command) {
1255         case LCFG_ADD_MDC:
1256                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1257                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1258                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1259                         GOTO(out, rc = -EINVAL);
1260
1261                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1262
1263                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1264                         GOTO(out, rc = -EINVAL);
1265                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1266                         GOTO(out, rc = -EINVAL);
1267                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1268                 GOTO(out, rc);
1269         default:
1270                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1271                 GOTO(out, rc = -EINVAL);
1272         }
1273 out:
1274         RETURN(rc);
1275 }
1276
1277 static int lmv_select_statfs_mdt(struct obd_export *exp, struct lmv_obd *lmv,
1278                                  u32 flags)
1279 {
1280         bool large_nid = exp_connect_flags2(exp) & OBD_CONNECT2_LARGE_NID;
1281         int i;
1282
1283         if (flags & OBD_STATFS_FOR_MDT0)
1284                 return 0;
1285
1286         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1287                 return lmv->lmv_statfs_start;
1288
1289         /* choose initial MDT for this client */
1290         for (i = 0;; i++) {
1291                 struct lnet_processid lnet_id;
1292
1293                 if (LNetGetId(i, &lnet_id, large_nid) == -ENOENT)
1294                         break;
1295
1296                 if (!nid_is_lo0(&lnet_id.nid)) {
1297                         /* We dont need a full 64-bit modulus, just enough
1298                          * to distribute the requests across MDTs evenly.
1299                          */
1300                         lmv->lmv_statfs_start = nidhash(&lnet_id.nid) %
1301                                                 lmv->lmv_mdt_count;
1302                         break;
1303                 }
1304         }
1305
1306         return lmv->lmv_statfs_start;
1307 }
1308
1309 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1310                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1311 {
1312         struct obd_device *obd = class_exp2obd(exp);
1313         struct lmv_obd *lmv = &obd->u.lmv;
1314         struct obd_statfs *temp;
1315         struct lu_tgt_desc *tgt;
1316         __u32 i;
1317         __u32 idx;
1318         int rc = 0;
1319         int err = 0;
1320
1321         ENTRY;
1322
1323         OBD_ALLOC(temp, sizeof(*temp));
1324         if (temp == NULL)
1325                 RETURN(-ENOMEM);
1326
1327         /* distribute statfs among MDTs */
1328         idx = lmv_select_statfs_mdt(exp, lmv, flags);
1329
1330         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1331                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1332                 tgt = lmv_tgt(lmv, idx);
1333                 if (!tgt || !tgt->ltd_exp)
1334                         continue;
1335
1336                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age,
1337                                 flags | OBD_STATFS_NESTED);
1338                 if (rc) {
1339                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1340                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1341                         err = rc;
1342                         /* Try another MDT */
1343                         if (flags & OBD_STATFS_SUM)
1344                                 continue;
1345                         GOTO(out_free_temp, rc);
1346                 }
1347
1348                 if (temp->os_state & OS_STATFS_SUM ||
1349                     flags == OBD_STATFS_FOR_MDT0) {
1350                         /* reset to the last aggregated values
1351                          * and don't sum with non-aggrated data */
1352                         /* If the statfs is from mount, it needs to retrieve
1353                          * necessary information from MDT0. i.e. mount does
1354                          * not need the merged osfs from all of MDT. Also
1355                          * clients can be mounted as long as MDT0 is in
1356                          * service */
1357                         *osfs = *temp;
1358                         GOTO(out_free_temp, rc);
1359                 }
1360
1361                 if (i == 0) {
1362                         *osfs = *temp;
1363                 } else {
1364                         osfs->os_bavail += temp->os_bavail;
1365                         osfs->os_blocks += temp->os_blocks;
1366                         osfs->os_ffree += temp->os_ffree;
1367                         osfs->os_files += temp->os_files;
1368                         osfs->os_granted += temp->os_granted;
1369                 }
1370         }
1371         /* There is no stats from some MDTs, data incomplete */
1372         if (err)
1373                 rc = err;
1374 out_free_temp:
1375         OBD_FREE(temp, sizeof(*temp));
1376         RETURN(rc);
1377 }
1378
1379 static int lmv_statfs_cb(void *cookie, int rc)
1380 {
1381         struct obd_info *oinfo = cookie;
1382         struct obd_device *obd = oinfo->oi_obd;
1383         struct lmv_obd *lmv = &obd->u.lmv;
1384         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1385         struct obd_statfs *osfs = oinfo->oi_osfs;
1386
1387         /*
1388          * NB: don't deactivate TGT upon error, because we may not trigger async
1389          * statfs any longer, then there is no chance to activate TGT.
1390          */
1391         if (!rc)
1392                 lmv_statfs_update(lmv, tgt, osfs);
1393
1394         return rc;
1395 }
1396
1397 /* update tgt statfs async if it's ld_qos_maxage old */
1398 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1399 {
1400         struct obd_info oinfo = {
1401                 .oi_obd = obd,
1402                 .oi_tgt = tgt,
1403                 .oi_cb_up = lmv_statfs_cb,
1404         };
1405         int rc;
1406
1407         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1408             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1409                 return 0;
1410
1411         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1412
1413         return rc;
1414 }
1415
1416 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1417                         struct lu_fid *fid)
1418 {
1419         struct obd_device *obd = exp->exp_obd;
1420         struct lmv_obd *lmv = &obd->u.lmv;
1421         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1422         int rc;
1423
1424         ENTRY;
1425
1426         if (!tgt)
1427                 RETURN(-ENODEV);
1428
1429         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1430         RETURN(rc);
1431 }
1432
1433 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1434                         u64 obd_md_valid, const char *name, size_t buf_size,
1435                         struct ptlrpc_request **req)
1436 {
1437         struct obd_device *obd = exp->exp_obd;
1438         struct lmv_obd *lmv = &obd->u.lmv;
1439         struct lmv_tgt_desc *tgt;
1440         int rc;
1441
1442         ENTRY;
1443
1444         tgt = lmv_fid2tgt(lmv, fid);
1445         if (IS_ERR(tgt))
1446                 RETURN(PTR_ERR(tgt));
1447
1448         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1449
1450         RETURN(rc);
1451 }
1452
1453 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1454                         u64 obd_md_valid, const char *name,
1455                         const void *value, size_t value_size,
1456                         unsigned int xattr_flags, u32 suppgid,
1457                         struct ptlrpc_request **req)
1458 {
1459         struct obd_device *obd = exp->exp_obd;
1460         struct lmv_obd *lmv = &obd->u.lmv;
1461         struct lmv_tgt_desc *tgt;
1462         int rc;
1463
1464         ENTRY;
1465
1466         tgt = lmv_fid2tgt(lmv, fid);
1467         if (IS_ERR(tgt))
1468                 RETURN(PTR_ERR(tgt));
1469
1470         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1471                          value, value_size, xattr_flags, suppgid, req);
1472
1473         RETURN(rc);
1474 }
1475
1476 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1477                        struct ptlrpc_request **request)
1478 {
1479         struct obd_device *obd = exp->exp_obd;
1480         struct lmv_obd *lmv = &obd->u.lmv;
1481         struct lmv_tgt_desc *tgt;
1482         int rc;
1483
1484         ENTRY;
1485
1486         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1487         if (IS_ERR(tgt))
1488                 RETURN(PTR_ERR(tgt));
1489
1490         if (op_data->op_flags & MF_GET_MDT_IDX) {
1491                 op_data->op_mds = tgt->ltd_index;
1492                 RETURN(0);
1493         }
1494
1495         rc = md_getattr(tgt->ltd_exp, op_data, request);
1496
1497         RETURN(rc);
1498 }
1499
1500 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1501 {
1502         struct obd_device *obd = exp->exp_obd;
1503         struct lmv_obd *lmv = &obd->u.lmv;
1504         struct lu_tgt_desc *tgt;
1505
1506         ENTRY;
1507
1508         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1509
1510         /*
1511          * With DNE every object can have two locks in different namespaces:
1512          * lookup lock in space of MDT storing direntry and update/open lock in
1513          * space of MDT storing inode.
1514          */
1515         lmv_foreach_connected_tgt(lmv, tgt)
1516                 md_null_inode(tgt->ltd_exp, fid);
1517
1518         RETURN(0);
1519 }
1520
1521 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1522                      struct md_open_data *mod, struct ptlrpc_request **request)
1523 {
1524         struct obd_device *obd = exp->exp_obd;
1525         struct lmv_obd *lmv = &obd->u.lmv;
1526         struct lmv_tgt_desc *tgt;
1527         int rc;
1528
1529         ENTRY;
1530
1531         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1532         if (IS_ERR(tgt))
1533                 RETURN(PTR_ERR(tgt));
1534
1535         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1536         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1537         RETURN(rc);
1538 }
1539
1540 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv,
1541                                               struct md_op_data *op_data)
1542 {
1543         struct lu_tgt_desc *tgt, *cur = NULL;
1544         __u64 total_avail = 0;
1545         __u64 total_weight = 0;
1546         __u64 cur_weight = 0;
1547         int total_usable = 0;
1548         __u64 rand;
1549         int rc;
1550
1551         ENTRY;
1552
1553         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1554                 RETURN(ERR_PTR(-EAGAIN));
1555
1556         down_write(&lmv->lmv_qos.lq_rw_sem);
1557
1558         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1559                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1560
1561         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1562         if (rc)
1563                 GOTO(unlock, tgt = ERR_PTR(rc));
1564
1565         lmv_foreach_tgt(lmv, tgt) {
1566                 if (!tgt->ltd_exp || !tgt->ltd_active ||
1567                     (tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE)) {
1568                         tgt->ltd_qos.ltq_usable = 0;
1569                         continue;
1570                 }
1571                 /* update one hour overdue statfs */
1572                 if (ktime_get_seconds() - tgt->ltd_statfs_age >
1573                     60 * lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1574                         lmv_statfs_check_update(lmv2obd_dev(lmv), tgt);
1575                 tgt->ltd_qos.ltq_usable = 1;
1576                 lu_tgt_qos_weight_calc(tgt, true);
1577                 if (tgt->ltd_index == op_data->op_mds)
1578                         cur = tgt;
1579                 total_avail += tgt->ltd_qos.ltq_avail;
1580                 total_weight += tgt->ltd_qos.ltq_weight;
1581                 total_usable++;
1582         }
1583
1584         /* If current MDT has above-average space and dir is not already using
1585          * round-robin to spread across more MDTs, stay on the parent MDT
1586          * to avoid creating needless remote MDT directories.  Remote dirs
1587          * close to the root balance space more effectively than bottom dirs,
1588          * so prefer to create remote dirs at top level of directory tree.
1589          * "16 / (dir_depth + 10)" is the factor to make it less likely
1590          * for top-level directories to stay local unless they have more than
1591          * average free space, while deep dirs prefer local until more full.
1592          *    depth=0 -> 160%, depth=3 -> 123%, depth=6 -> 100%,
1593          *    depth=9 -> 84%, depth=12 -> 73%, depth=15 -> 64%
1594          */
1595         if (!lmv_op_default_rr_mkdir(op_data)) {
1596                 rand = total_avail * 16 /
1597                         (total_usable * (op_data->op_dir_depth + 10));
1598                 if (cur && cur->ltd_qos.ltq_avail >= rand) {
1599                         tgt = cur;
1600                         GOTO(unlock, tgt);
1601                 }
1602         }
1603
1604         rand = lu_prandom_u64_max(total_weight);
1605
1606         lmv_foreach_connected_tgt(lmv, tgt) {
1607                 if (!tgt->ltd_qos.ltq_usable)
1608                         continue;
1609
1610                 cur_weight += tgt->ltd_qos.ltq_weight;
1611                 if (cur_weight < rand)
1612                         continue;
1613
1614                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1615                 GOTO(unlock, tgt);
1616         }
1617
1618         /* no proper target found */
1619         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1620 unlock:
1621         up_write(&lmv->lmv_qos.lq_rw_sem);
1622
1623         return tgt;
1624 }
1625
1626 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv)
1627 {
1628         struct lu_tgt_desc *tgt;
1629         int i;
1630         int index;
1631
1632         ENTRY;
1633
1634         spin_lock(&lmv->lmv_lock);
1635         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1636                 index = (i + lmv->lmv_qos_rr_index) %
1637                         lmv->lmv_mdt_descs.ltd_tgts_size;
1638                 tgt = lmv_tgt(lmv, index);
1639                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active ||
1640                     (tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE))
1641                         continue;
1642
1643                 lmv->lmv_qos_rr_index = (tgt->ltd_index + 1) %
1644                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1645                 spin_unlock(&lmv->lmv_lock);
1646
1647                 RETURN(tgt);
1648         }
1649         spin_unlock(&lmv->lmv_lock);
1650
1651         RETURN(ERR_PTR(-ENODEV));
1652 }
1653
1654 /* locate MDT which is less full (avoid the most full MDT) */
1655 static struct lu_tgt_desc *lmv_locate_tgt_lf(struct lmv_obd *lmv)
1656 {
1657         struct lu_tgt_desc *min = NULL;
1658         struct lu_tgt_desc *tgt;
1659         __u64 avail = 0;
1660         __u64 rand;
1661
1662         ENTRY;
1663
1664         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1665                 RETURN(ERR_PTR(-EAGAIN));
1666
1667         down_write(&lmv->lmv_qos.lq_rw_sem);
1668
1669         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1670                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1671
1672         lmv_foreach_tgt(lmv, tgt) {
1673                 if (!tgt->ltd_exp || !tgt->ltd_active ||
1674                     (tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE)) {
1675                         tgt->ltd_qos.ltq_usable = 0;
1676                         continue;
1677                 }
1678
1679                 tgt->ltd_qos.ltq_usable = 1;
1680                 lu_tgt_qos_weight_calc(tgt, true);
1681                 avail += tgt->ltd_qos.ltq_avail;
1682                 if (!min || min->ltd_qos.ltq_avail > tgt->ltd_qos.ltq_avail)
1683                         min = tgt;
1684         }
1685
1686         /* avoid the most full MDT */
1687         if (min)
1688                 avail -= min->ltd_qos.ltq_avail;
1689
1690         rand = lu_prandom_u64_max(avail);
1691         avail = 0;
1692         lmv_foreach_connected_tgt(lmv, tgt) {
1693                 if (!tgt->ltd_qos.ltq_usable)
1694                         continue;
1695
1696                 if (tgt == min)
1697                         continue;
1698
1699                 avail += tgt->ltd_qos.ltq_avail;
1700                 if (avail < rand)
1701                         continue;
1702
1703                 GOTO(unlock, tgt);
1704         }
1705
1706         /* no proper target found */
1707         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1708 unlock:
1709         up_write(&lmv->lmv_qos.lq_rw_sem);
1710
1711         RETURN(tgt);
1712 }
1713
1714 /* locate MDT by file name, for striped directory, the file name hash decides
1715  * which stripe its dirent is stored.
1716  */
1717 static struct lmv_tgt_desc *
1718 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_object *lso,
1719                        const char *name, int namelen, struct lu_fid *fid,
1720                        __u32 *mds, bool new_layout)
1721 {
1722         struct lmv_tgt_desc *tgt;
1723         const struct lmv_oinfo *oinfo;
1724
1725         if (!lmv_dir_striped(lso) || !namelen) {
1726                 tgt = lmv_fid2tgt(lmv, fid);
1727                 if (IS_ERR(tgt))
1728                         return tgt;
1729
1730                 *mds = tgt->ltd_index;
1731                 return tgt;
1732         }
1733
1734         if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1735                 if (cfs_fail_val >= lso->lso_lsm.lsm_md_stripe_count)
1736                         return ERR_PTR(-EBADF);
1737                 oinfo = &lso->lso_lsm.lsm_md_oinfo[cfs_fail_val];
1738         } else {
1739                 oinfo = lsm_name_to_stripe_info(lso, name, namelen, new_layout);
1740                 if (IS_ERR(oinfo))
1741                         return ERR_CAST(oinfo);
1742         }
1743
1744         /* check stripe FID is sane */
1745         if (!fid_is_sane(&oinfo->lmo_fid))
1746                 return ERR_PTR(-ENODEV);
1747
1748         *fid = oinfo->lmo_fid;
1749         *mds = oinfo->lmo_mds;
1750         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1751
1752         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1753
1754         return tgt ? tgt : ERR_PTR(-ENODEV);
1755 }
1756
1757 /**
1758  * Locate MDT of op_data->op_fid1
1759  *
1760  * For striped directory, it will locate the stripe by name hash, if hash_type
1761  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1762  * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1763  * indicates whether old or new layout is used to locate.
1764  *
1765  * For plain directory, it just locate the MDT of op_data->op_fid1.
1766  *
1767  * \param[in] lmv               LMV device
1768  * \param[in/out] op_data       client MD stack parameters, name, namelen etc,
1769  *                              op_mds and op_fid1 will be updated if op_lso1
1770  *                              indicates fid1 represents a striped directory.
1771  *
1772  * retval               pointer to the lmv_tgt_desc if succeed.
1773  *                      ERR_PTR(errno) if failed.
1774  */
1775 struct lmv_tgt_desc *
1776 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1777 {
1778         struct lmv_stripe_md *lsm;
1779         struct lmv_oinfo *oinfo;
1780         struct lmv_tgt_desc *tgt;
1781
1782         if (lmv_dir_foreign(op_data->op_lso1))
1783                 return ERR_PTR(-ENODATA);
1784
1785         /* During creating VOLATILE file, it should honor the mdt
1786          * index if the file under striped dir is being restored, see
1787          * ct_restore(). */
1788         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1789             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1790                 tgt = lmv_tgt(lmv, op_data->op_mds);
1791                 if (!tgt)
1792                         return ERR_PTR(-ENODEV);
1793
1794                 if (lmv_dir_striped(op_data->op_lso1)) {
1795                         int i;
1796
1797                         /* refill the right parent fid */
1798                         lsm = &op_data->op_lso1->lso_lsm;
1799                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1800                                 oinfo = &lsm->lsm_md_oinfo[i];
1801                                 if (oinfo->lmo_mds == op_data->op_mds) {
1802                                         op_data->op_fid1 = oinfo->lmo_fid;
1803                                         break;
1804                                 }
1805                         }
1806
1807                         if (i == lsm->lsm_md_stripe_count)
1808                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1809                 }
1810         } else if (lmv_dir_bad_hash(op_data->op_lso1)) {
1811                 lsm = &op_data->op_lso1->lso_lsm;
1812
1813                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1814                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1815
1816                 op_data->op_fid1 = oinfo->lmo_fid;
1817                 op_data->op_mds = oinfo->lmo_mds;
1818                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1819                 if (!tgt)
1820                         return ERR_PTR(-ENODEV);
1821         } else {
1822                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_lso1,
1823                                 op_data->op_name, op_data->op_namelen,
1824                                 &op_data->op_fid1, &op_data->op_mds,
1825                                 op_data->op_new_layout);
1826         }
1827
1828         return tgt;
1829 }
1830
1831 /* Locate MDT of op_data->op_fid2 for link/rename */
1832 static struct lmv_tgt_desc *
1833 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1834 {
1835         struct lmv_tgt_desc *tgt;
1836         int rc;
1837
1838         LASSERT(op_data->op_name);
1839         if (lmv_dir_layout_changing(op_data->op_lso2)) {
1840                 struct lu_fid fid1 = op_data->op_fid1;
1841                 struct lmv_stripe_object *lso1 = op_data->op_lso1;
1842                 struct ptlrpc_request *request = NULL;
1843
1844                 /*
1845                  * avoid creating new file under old layout of migrating
1846                  * directory, check it here.
1847                  */
1848                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_lso2,
1849                                 op_data->op_name, op_data->op_namelen,
1850                                 &op_data->op_fid2, &op_data->op_mds, false);
1851                 if (IS_ERR(tgt))
1852                         RETURN(tgt);
1853
1854                 op_data->op_fid1 = op_data->op_fid2;
1855                 op_data->op_lso1 = op_data->op_lso2;
1856                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1857                 op_data->op_fid1 = fid1;
1858                 op_data->op_lso1 = lso1;
1859                 if (!rc) {
1860                         ptlrpc_req_finished(request);
1861                         RETURN(ERR_PTR(-EEXIST));
1862                 }
1863
1864                 if (rc != -ENOENT)
1865                         RETURN(ERR_PTR(rc));
1866         }
1867
1868         return lmv_locate_tgt_by_name(lmv, op_data->op_lso2,
1869                                       op_data->op_name, op_data->op_namelen,
1870                                       &op_data->op_fid2, &op_data->op_mds,
1871                                       true);
1872 }
1873
1874 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1875 {
1876         struct lu_tgt_desc *tgt;
1877         struct ptlrpc_request *request;
1878         int rc;
1879
1880         LASSERT(lmv_dir_layout_changing(op_data->op_lso1));
1881         LASSERT(!op_data->op_new_layout);
1882
1883         tgt = lmv_locate_tgt(lmv, op_data);
1884         if (IS_ERR(tgt))
1885                 return PTR_ERR(tgt);
1886
1887         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1888         if (!rc) {
1889                 ptlrpc_req_finished(request);
1890                 return -EEXIST;
1891         }
1892
1893         return rc;
1894 }
1895
1896 /* mkdir by QoS upon 'lfs mkdir -i -1'.
1897  *
1898  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1899  * directories under striped directory.
1900  */
1901 static inline bool lmv_op_user_qos_mkdir(const struct md_op_data *op_data)
1902 {
1903         const struct lmv_user_md *lum = op_data->op_data;
1904
1905         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1906                 return false;
1907
1908         if (lmv_dir_striped(op_data->op_lso1))
1909                 return false;
1910
1911         return (op_data->op_cli_flags & CLI_SET_MEA) && lum &&
1912                le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC &&
1913                le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
1914 }
1915
1916 /* mkdir by QoS if either ROOT or parent default LMV is space balanced. */
1917 static inline bool lmv_op_default_qos_mkdir(const struct md_op_data *op_data)
1918 {
1919         const struct lmv_stripe_object *lso = op_data->op_default_lso1;
1920
1921         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1922                 return false;
1923
1924         if (lmv_dir_striped(op_data->op_lso1))
1925                 return false;
1926
1927         return (op_data->op_flags & MF_QOS_MKDIR) ||
1928                (lso && lso->lso_lsm.lsm_md_master_mdt_index ==
1929                 LMV_OFFSET_DEFAULT);
1930 }
1931
1932 /* if parent default LMV is space balanced, and
1933  * 1. max_inherit_rr is set
1934  * 2. or parent is ROOT
1935  * mkdir roundrobin. Or if parent doesn't have default LMV, while ROOT default
1936  * LMV requests roundrobin mkdir, do the same.
1937  * NB, this needs to check server is balanced, which is done by caller.
1938  */
1939 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data)
1940 {
1941         const struct lmv_stripe_object *lso = op_data->op_default_lso1;
1942
1943         return (op_data->op_flags & MF_RR_MKDIR) ||
1944                (lso && lso->lso_lsm.lsm_md_max_inherit_rr !=
1945                 LMV_INHERIT_RR_NONE) || fid_is_root(&op_data->op_fid1);
1946 }
1947
1948 /* 'lfs mkdir -i <specific_MDT>' */
1949 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1950 {
1951         const struct lmv_user_md *lum = op_data->op_data;
1952
1953         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1954                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1955                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1956                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1957                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1958 }
1959
1960 /* parent default LMV master_mdt_index is not -1. */
1961 static inline bool
1962 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1963 {
1964         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1965                op_data->op_default_lso1 &&
1966                op_data->op_default_lso1->lso_lsm.lsm_md_master_mdt_index !=
1967                         LMV_OFFSET_DEFAULT;
1968 }
1969
1970 /* locate MDT by space usage */
1971 static struct lu_tgt_desc *lmv_locate_tgt_by_space(struct lmv_obd *lmv,
1972                                                    struct md_op_data *op_data,
1973                                                    struct lmv_tgt_desc *tgt)
1974 {
1975         struct lmv_tgt_desc *tmp = tgt;
1976
1977         tgt = lmv_locate_tgt_qos(lmv, op_data);
1978         if (tgt == ERR_PTR(-EAGAIN)) {
1979                 if (ltd_qos_is_balanced(&lmv->lmv_mdt_descs) &&
1980                     !lmv_op_default_rr_mkdir(op_data) &&
1981                     !lmv_op_user_qos_mkdir(op_data) &&
1982                     !(tmp->ltd_statfs.os_state & OS_STATFS_NOCREATE))
1983                         /* if not necessary, don't create remote directory. */
1984                         tgt = tmp;
1985                 else
1986                         tgt = lmv_locate_tgt_rr(lmv);
1987                 if (!IS_ERR(tgt))
1988                         lmv_statfs_check_update(lmv2obd_dev(lmv), tgt);
1989         }
1990
1991         if (!IS_ERR(tgt))
1992                 op_data->op_mds = tgt->ltd_index;
1993
1994         /* If space balance was called because the original target was marked
1995          * NOCREATE, periodically check whether the state has changed.
1996          */
1997         if (tmp != tgt && tmp->ltd_statfs.os_state & OS_STATFS_NOCREATE)
1998                 lmv_statfs_check_update(lmv2obd_dev(lmv), tmp);
1999
2000         return tgt;
2001 }
2002
2003 static int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
2004                       const void *data, size_t datalen, umode_t mode, uid_t uid,
2005                       gid_t gid, kernel_cap_t cap_effective, __u64 rdev,
2006                       struct ptlrpc_request **request)
2007 {
2008         struct obd_device *obd = exp->exp_obd;
2009         struct lmv_obd *lmv = &obd->u.lmv;
2010         struct lmv_tgt_desc *tgt;
2011         struct mdt_body *repbody;
2012         int rc;
2013
2014         ENTRY;
2015
2016         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
2017                 RETURN(-EIO);
2018
2019         if (lmv_dir_bad_hash(op_data->op_lso1))
2020                 RETURN(-EBADF);
2021
2022         if (lmv_dir_layout_changing(op_data->op_lso1)) {
2023                 /*
2024                  * if parent is migrating, create() needs to lookup existing
2025                  * name in both old and new layout, check old layout on client.
2026                  */
2027                 rc = lmv_old_layout_lookup(lmv, op_data);
2028                 if (rc != -ENOENT)
2029                         RETURN(rc);
2030
2031                 op_data->op_new_layout = true;
2032         }
2033
2034         tgt = lmv_locate_tgt(lmv, op_data);
2035         if (IS_ERR(tgt))
2036                 RETURN(PTR_ERR(tgt));
2037
2038         /* the order to apply policy in mkdir:
2039          * 1. is "lfs mkdir -i N"? mkdir on MDT N.
2040          * 2. is "lfs mkdir -i -1"? mkdir by space usage.
2041          * 3. is starting MDT specified in default LMV? mkdir on MDT N.
2042          * 4. is default LMV space balanced? mkdir by space usage.
2043          *
2044          * If the existing parent or specific MDT selected is deactivated
2045          * with OS_STATFS_NOCREATE then select a different MDT by QOS.
2046          */
2047         if (lmv_op_user_specific_mkdir(op_data)) {
2048                 struct lmv_user_md *lum = op_data->op_data;
2049
2050                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2051                 tgt = lmv_tgt(lmv, op_data->op_mds);
2052                 if (!tgt)
2053                         RETURN(-ENODEV);
2054                 if (unlikely(tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE))
2055                         GOTO(new_tgt, -EAGAIN);
2056         } else if (lmv_op_user_qos_mkdir(op_data)) {
2057                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
2058                 if (IS_ERR(tgt))
2059                         RETURN(PTR_ERR(tgt));
2060         } else if (lmv_op_default_specific_mkdir(op_data)) {
2061                 struct lmv_stripe_md *lsm = &op_data->op_default_lso1->lso_lsm;
2062
2063                 op_data->op_mds = lsm->lsm_md_master_mdt_index;
2064                 tgt = lmv_tgt(lmv, op_data->op_mds);
2065                 if (!tgt)
2066                         RETURN(-ENODEV);
2067                 if (unlikely(tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE))
2068                         GOTO(new_tgt, -EAGAIN);
2069         } else if (lmv_op_default_qos_mkdir(op_data) ||
2070                    tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE) {
2071 new_tgt:
2072                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
2073                 if (IS_ERR(tgt))
2074                         RETURN(PTR_ERR(tgt));
2075         }
2076
2077 retry:
2078         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
2079         if (rc)
2080                 RETURN(rc);
2081
2082         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
2083                 (int)op_data->op_namelen, op_data->op_name,
2084                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
2085                 op_data->op_mds);
2086
2087         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2088         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
2089                        cap_effective, rdev, request);
2090         if (rc == 0) {
2091                 if (*request == NULL)
2092                         RETURN(rc);
2093                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
2094         }
2095
2096         /* dir restripe needs to send to MDT where dir is located */
2097         if (rc != -EREMOTE ||
2098             !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
2099                 RETURN(rc);
2100
2101         repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2102         if (repbody == NULL)
2103                 RETURN(-EPROTO);
2104
2105         /* Not cross-ref case, just get out of here. */
2106         if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
2107                 RETURN(rc);
2108
2109         op_data->op_fid2 = repbody->mbo_fid1;
2110         ptlrpc_req_finished(*request);
2111         *request = NULL;
2112
2113         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2114         if (IS_ERR(tgt))
2115                 RETURN(PTR_ERR(tgt));
2116
2117         op_data->op_mds = tgt->ltd_index;
2118         goto retry;
2119 }
2120
2121 static int
2122 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
2123             const union ldlm_policy_data *policy, struct md_op_data *op_data,
2124             struct lustre_handle *lockh, __u64 extra_lock_flags)
2125 {
2126         struct obd_device *obd = exp->exp_obd;
2127         struct lmv_obd *lmv = &obd->u.lmv;
2128         struct lmv_tgt_desc *tgt;
2129         int rc;
2130
2131         ENTRY;
2132
2133         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
2134
2135         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2136         if (IS_ERR(tgt))
2137                 RETURN(PTR_ERR(tgt));
2138
2139         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
2140                PFID(&op_data->op_fid1), tgt->ltd_index);
2141
2142         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
2143                         extra_lock_flags);
2144
2145         RETURN(rc);
2146 }
2147
2148 int
2149 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
2150                  struct ptlrpc_request **preq)
2151 {
2152         struct obd_device *obd = exp->exp_obd;
2153         struct lmv_obd *lmv = &obd->u.lmv;
2154         struct lmv_tgt_desc *tgt;
2155         struct mdt_body *body;
2156         int rc;
2157
2158         ENTRY;
2159
2160 retry:
2161         if (op_data->op_namelen == 2 &&
2162             op_data->op_name[0] == '.' && op_data->op_name[1] == '.')
2163                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2164         else
2165                 tgt = lmv_locate_tgt(lmv, op_data);
2166         if (IS_ERR(tgt))
2167                 RETURN(PTR_ERR(tgt));
2168
2169         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
2170                 (int)op_data->op_namelen, op_data->op_name,
2171                 PFID(&op_data->op_fid1), tgt->ltd_index);
2172
2173         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
2174         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2175                 ptlrpc_req_finished(*preq);
2176                 *preq = NULL;
2177                 goto retry;
2178         }
2179
2180         if (rc)
2181                 RETURN(rc);
2182
2183         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
2184         LASSERT(body != NULL);
2185
2186         if (body->mbo_valid & OBD_MD_MDS) {
2187                 op_data->op_fid1 = body->mbo_fid1;
2188                 op_data->op_valid |= OBD_MD_FLCROSSREF;
2189                 op_data->op_namelen = 0;
2190                 op_data->op_name = NULL;
2191
2192                 ptlrpc_req_finished(*preq);
2193                 *preq = NULL;
2194
2195                 goto retry;
2196         }
2197
2198         RETURN(rc);
2199 }
2200
2201 #define md_op_data_fid(op_data, fl)                     \
2202         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2203          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2204          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2205          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2206          NULL)
2207
2208 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2209                             struct md_op_data *op_data, __u32 op_tgt,
2210                             enum ldlm_mode mode, int bits, int flag)
2211 {
2212         struct lu_fid *fid = md_op_data_fid(op_data, flag);
2213         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
2214         union ldlm_policy_data policy = { { 0 } };
2215         int rc = 0;
2216         ENTRY;
2217
2218         if (!fid_is_sane(fid))
2219                 RETURN(0);
2220
2221         if (tgt == NULL) {
2222                 tgt = lmv_fid2tgt(lmv, fid);
2223                 if (IS_ERR(tgt))
2224                         RETURN(PTR_ERR(tgt));
2225         }
2226
2227         if (tgt->ltd_index != op_tgt) {
2228                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2229                 policy.l_inodebits.bits = bits;
2230                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2231                                       mode, LCF_ASYNC, NULL);
2232         } else {
2233                 CDEBUG(D_INODE,
2234                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2235                        op_tgt, PFID(fid));
2236                 op_data->op_flags |= flag;
2237                 rc = 0;
2238         }
2239
2240         RETURN(rc);
2241 }
2242
2243 /*
2244  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2245  * op_data->op_fid2
2246  */
2247 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2248                     struct ptlrpc_request **request)
2249 {
2250         struct obd_device       *obd = exp->exp_obd;
2251         struct lmv_obd          *lmv = &obd->u.lmv;
2252         struct lmv_tgt_desc     *tgt;
2253         int                      rc;
2254         ENTRY;
2255
2256         LASSERT(op_data->op_namelen != 0);
2257
2258         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2259                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2260                op_data->op_name, PFID(&op_data->op_fid1));
2261
2262         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2263         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2264         op_data->op_cap = current_cap();
2265
2266         tgt = lmv_locate_tgt2(lmv, op_data);
2267         if (IS_ERR(tgt))
2268                 RETURN(PTR_ERR(tgt));
2269
2270         /*
2271          * Cancel UPDATE lock on child (fid1).
2272          */
2273         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2274         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2275                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2276         if (rc != 0)
2277                 RETURN(rc);
2278
2279         rc = md_link(tgt->ltd_exp, op_data, request);
2280
2281         RETURN(rc);
2282 }
2283
2284 /* migrate the top directory */
2285 static inline bool lmv_op_topdir_migrate(const struct md_op_data *op_data)
2286 {
2287         if (!S_ISDIR(op_data->op_mode))
2288                 return false;
2289
2290         if (lmv_dir_layout_changing(op_data->op_lso1))
2291                 return false;
2292
2293         return true;
2294 }
2295
2296 /* migrate top dir to specific MDTs */
2297 static inline bool lmv_topdir_specific_migrate(const struct md_op_data *op_data)
2298 {
2299         const struct lmv_user_md *lum = op_data->op_data;
2300
2301         if (!lmv_op_topdir_migrate(op_data))
2302                 return false;
2303
2304         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2305 }
2306
2307 /* migrate top dir in QoS mode if user issued "lfs migrate -m -1..." */
2308 static inline bool lmv_topdir_qos_migrate(const struct md_op_data *op_data)
2309 {
2310         const struct lmv_user_md *lum = op_data->op_data;
2311
2312         if (!lmv_op_topdir_migrate(op_data))
2313                 return false;
2314
2315         return le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
2316 }
2317
2318 static inline bool lmv_subdir_specific_migrate(const struct md_op_data *op_data)
2319 {
2320         const struct lmv_user_md *lum = op_data->op_data;
2321
2322         if (!S_ISDIR(op_data->op_mode))
2323                 return false;
2324
2325         if (!lmv_dir_layout_changing(op_data->op_lso1))
2326                 return false;
2327
2328         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2329 }
2330
2331 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2332                         const char *name, size_t namelen,
2333                         struct ptlrpc_request **request)
2334 {
2335         struct obd_device *obd = exp->exp_obd;
2336         struct lmv_obd *lmv = &obd->u.lmv;
2337         struct lmv_stripe_object *lso = op_data->op_lso1;
2338         struct lmv_tgt_desc *parent_tgt;
2339         struct lmv_tgt_desc *sp_tgt;
2340         struct lmv_tgt_desc *tp_tgt = NULL;
2341         struct lmv_tgt_desc *child_tgt;
2342         struct lmv_tgt_desc *tgt;
2343         struct lu_fid target_fid = { 0 };
2344         int rc;
2345
2346         ENTRY;
2347
2348         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2349
2350         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2351                PFID(&op_data->op_fid1), (int)namelen, name);
2352
2353         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2354         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2355         op_data->op_cap = current_cap();
2356
2357         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2358         if (IS_ERR(parent_tgt))
2359                 RETURN(PTR_ERR(parent_tgt));
2360
2361         if (lmv_dir_striped(lso)) {
2362                 const struct lmv_oinfo *oinfo;
2363
2364                 oinfo = lsm_name_to_stripe_info(lso, name, namelen, false);
2365                 if (IS_ERR(oinfo))
2366                         RETURN(PTR_ERR(oinfo));
2367
2368                 /* save source stripe FID in fid4 temporarily for ELC */
2369                 op_data->op_fid4 = oinfo->lmo_fid;
2370                 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2371                 if (!sp_tgt)
2372                         RETURN(-ENODEV);
2373
2374                 /*
2375                  * if parent is being migrated too, fill op_fid2 with target
2376                  * stripe fid, otherwise the target stripe is not created yet.
2377                  */
2378                 if (lmv_dir_layout_changing(lso)) {
2379                         oinfo = lsm_name_to_stripe_info(lso, name, namelen,
2380                                                         true);
2381                         if (IS_ERR(oinfo))
2382                                 RETURN(PTR_ERR(oinfo));
2383
2384                         op_data->op_fid2 = oinfo->lmo_fid;
2385                         tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2386                         if (!tp_tgt)
2387                                 RETURN(-ENODEV);
2388
2389                         /* parent unchanged and update namespace only */
2390                         if (lu_fid_eq(&op_data->op_fid4, &op_data->op_fid2) &&
2391                             op_data->op_bias & MDS_MIGRATE_NSONLY)
2392                                 RETURN(-EALREADY);
2393                 }
2394         } else {
2395                 sp_tgt = parent_tgt;
2396         }
2397
2398         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2399         if (IS_ERR(child_tgt))
2400                 RETURN(PTR_ERR(child_tgt));
2401
2402         if (lmv_topdir_specific_migrate(op_data)) {
2403                 struct lmv_user_md *lum = op_data->op_data;
2404
2405                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2406         } else if (lmv_topdir_qos_migrate(op_data)) {
2407                 tgt = lmv_locate_tgt_lf(lmv);
2408                 if (tgt == ERR_PTR(-EAGAIN))
2409                         tgt = lmv_locate_tgt_rr(lmv);
2410                 if (IS_ERR(tgt))
2411                         RETURN(PTR_ERR(tgt));
2412
2413                 op_data->op_mds = tgt->ltd_index;
2414         } else if (lmv_subdir_specific_migrate(op_data)) {
2415                 struct lmv_user_md *lum = op_data->op_data;
2416                 __u32 i;
2417
2418                 LASSERT(tp_tgt);
2419                 if (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) {
2420                         /* adjust MDTs in lum, since subdir is located on where
2421                          * its parent stripe is, not the first specified MDT.
2422                          */
2423                         for (i = 0; i < le32_to_cpu(lum->lum_stripe_count);
2424                              i++) {
2425                                 if (le32_to_cpu(lum->lum_objects[i].lum_mds) ==
2426                                     tp_tgt->ltd_index)
2427                                         break;
2428                         }
2429
2430                         if (i == le32_to_cpu(lum->lum_stripe_count))
2431                                 RETURN(-ENODEV);
2432
2433                         lum->lum_objects[i].lum_mds =
2434                                 lum->lum_objects[0].lum_mds;
2435                         lum->lum_objects[0].lum_mds =
2436                                 cpu_to_le32(tp_tgt->ltd_index);
2437                 }
2438                 /* NB, the above adjusts subdir migration for command like
2439                  * "lfs migrate -m 0,1,2 ...", but for migration like
2440                  * "lfs migrate -m 0 -c 2 ...", the top dir is migrated to MDT0
2441                  * and MDT1, however its subdir may be migrated to MDT1 and MDT2
2442                  */
2443
2444                 lum->lum_stripe_offset = cpu_to_le32(tp_tgt->ltd_index);
2445                 op_data->op_mds = tp_tgt->ltd_index;
2446         } else if (tp_tgt) {
2447                 op_data->op_mds = tp_tgt->ltd_index;
2448         } else {
2449                 op_data->op_mds = sp_tgt->ltd_index;
2450         }
2451
2452         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2453         if (rc)
2454                 RETURN(rc);
2455
2456         /*
2457          * for directory, send migrate request to the MDT where the object will
2458          * be migrated to, because we can't create a striped directory remotely.
2459          *
2460          * otherwise, send to the MDT where source is located because regular
2461          * file may open lease.
2462          *
2463          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2464          * backward compatibility.
2465          */
2466         if (S_ISDIR(op_data->op_mode) &&
2467             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2468                 tgt = lmv_fid2tgt(lmv, &target_fid);
2469                 if (IS_ERR(tgt))
2470                         RETURN(PTR_ERR(tgt));
2471         } else {
2472                 tgt = child_tgt;
2473         }
2474
2475         /* cancel UPDATE lock of parent master object */
2476         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2477                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2478         if (rc)
2479                 RETURN(rc);
2480
2481         /* cancel UPDATE lock of source parent */
2482         if (sp_tgt != parent_tgt) {
2483                 /*
2484                  * migrate RPC packs master object FID, because we can only pack
2485                  * two FIDs in reint RPC, but MDS needs to know both source
2486                  * parent and target parent, and it will obtain them from master
2487                  * FID and LMV, the other FID in RPC is kept for target.
2488                  *
2489                  * since this FID is not passed to MDC, cancel it anyway.
2490                  */
2491                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2492                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2493                 if (rc)
2494                         RETURN(rc);
2495
2496                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2497         }
2498         op_data->op_fid4 = target_fid;
2499
2500         /* cancel UPDATE locks of target parent */
2501         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2502                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2503         if (rc)
2504                 RETURN(rc);
2505
2506         /* cancel LOOKUP lock of source if source is remote object */
2507         if (child_tgt != sp_tgt) {
2508                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2509                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2510                                       MF_MDC_CANCEL_FID3);
2511                 if (rc)
2512                         RETURN(rc);
2513         }
2514
2515         /* cancel ELC locks of source */
2516         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2517                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2518         if (rc)
2519                 RETURN(rc);
2520
2521         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2522
2523         RETURN(rc);
2524 }
2525
2526 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2527                       const char *old, size_t oldlen,
2528                       const char *new, size_t newlen,
2529                       struct ptlrpc_request **request)
2530 {
2531         struct obd_device *obd = exp->exp_obd;
2532         struct lmv_obd *lmv = &obd->u.lmv;
2533         struct lmv_tgt_desc *sp_tgt;
2534         struct lmv_tgt_desc *tp_tgt = NULL;
2535         struct lmv_tgt_desc *src_tgt = NULL;
2536         struct lmv_tgt_desc *tgt;
2537         struct mdt_body *body;
2538         int rc;
2539
2540         ENTRY;
2541
2542         LASSERT(oldlen != 0);
2543
2544         if (op_data->op_cli_flags & CLI_MIGRATE) {
2545                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2546                 RETURN(rc);
2547         }
2548
2549         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2550         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2551         op_data->op_cap = current_cap();
2552
2553         op_data->op_name = new;
2554         op_data->op_namelen = newlen;
2555
2556         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2557         if (IS_ERR(tp_tgt))
2558                 RETURN(PTR_ERR(tp_tgt));
2559
2560         /* Since the target child might be destroyed, and it might become
2561          * orphan, and we can only check orphan on the local MDT right now, so
2562          * we send rename request to the MDT where target child is located. If
2563          * target child does not exist, then it will send the request to the
2564          * target parent */
2565         if (fid_is_sane(&op_data->op_fid4)) {
2566                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2567                 if (IS_ERR(tgt))
2568                         RETURN(PTR_ERR(tgt));
2569         } else {
2570                 tgt = tp_tgt;
2571         }
2572
2573         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2574
2575         /* cancel UPDATE locks of target parent */
2576         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2577                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2578         if (rc != 0)
2579                 RETURN(rc);
2580
2581         if (fid_is_sane(&op_data->op_fid4)) {
2582                 /* cancel LOOKUP lock of target on target parent */
2583                 if (tgt != tp_tgt) {
2584                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2585                                               tgt->ltd_index, LCK_EX,
2586                                               MDS_INODELOCK_LOOKUP,
2587                                               MF_MDC_CANCEL_FID4);
2588                         if (rc != 0)
2589                                 RETURN(rc);
2590                 }
2591         }
2592
2593         if (fid_is_sane(&op_data->op_fid3)) {
2594                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2595                 if (IS_ERR(src_tgt))
2596                         RETURN(PTR_ERR(src_tgt));
2597
2598                 /* cancel ELC locks of source */
2599                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2600                                       LCK_EX, MDS_INODELOCK_ELC,
2601                                       MF_MDC_CANCEL_FID3);
2602                 if (rc != 0)
2603                         RETURN(rc);
2604         }
2605
2606         op_data->op_name = old;
2607         op_data->op_namelen = oldlen;
2608 retry:
2609         sp_tgt = lmv_locate_tgt(lmv, op_data);
2610         if (IS_ERR(sp_tgt))
2611                 RETURN(PTR_ERR(sp_tgt));
2612
2613         /* cancel UPDATE locks of source parent */
2614         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2615                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2616         if (rc != 0)
2617                 RETURN(rc);
2618
2619         if (fid_is_sane(&op_data->op_fid3)) {
2620                 /* cancel LOOKUP lock of source on source parent */
2621                 if (src_tgt != sp_tgt) {
2622                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2623                                               tgt->ltd_index, LCK_EX,
2624                                               MDS_INODELOCK_LOOKUP,
2625                                               MF_MDC_CANCEL_FID3);
2626                         if (rc != 0)
2627                                 RETURN(rc);
2628                 }
2629         }
2630
2631 rename:
2632         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2633                 PFID(&op_data->op_fid1), (int)oldlen, old,
2634                 PFID(&op_data->op_fid2), (int)newlen, new);
2635
2636         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2637                         request);
2638         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2639                 ptlrpc_req_finished(*request);
2640                 *request = NULL;
2641                 goto retry;
2642         }
2643
2644         if (rc && rc != -EXDEV)
2645                 RETURN(rc);
2646
2647         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2648         if (body == NULL)
2649                 RETURN(-EPROTO);
2650
2651         /* Not cross-ref case, just get out of here. */
2652         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2653                 RETURN(rc);
2654
2655         op_data->op_fid4 = body->mbo_fid1;
2656
2657         ptlrpc_req_finished(*request);
2658         *request = NULL;
2659
2660         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2661         if (IS_ERR(tgt))
2662                 RETURN(PTR_ERR(tgt));
2663
2664         if (fid_is_sane(&op_data->op_fid4)) {
2665                 /* cancel LOOKUP lock of target on target parent */
2666                 if (tgt != tp_tgt) {
2667                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2668                                               tgt->ltd_index, LCK_EX,
2669                                               MDS_INODELOCK_LOOKUP,
2670                                               MF_MDC_CANCEL_FID4);
2671                         if (rc != 0)
2672                                 RETURN(rc);
2673                 }
2674         }
2675
2676         goto rename;
2677 }
2678
2679 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2680                        void *ea, size_t ealen, 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 = 0;
2686
2687         ENTRY;
2688
2689         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2690                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2691                op_data->op_xvalid);
2692
2693         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2694         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2695         if (IS_ERR(tgt))
2696                 RETURN(PTR_ERR(tgt));
2697
2698         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2699
2700         RETURN(rc);
2701 }
2702
2703 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2704                      struct ptlrpc_request **request)
2705 {
2706         struct obd_device *obd = exp->exp_obd;
2707         struct lmv_obd *lmv = &obd->u.lmv;
2708         struct lmv_tgt_desc *tgt;
2709         int rc;
2710
2711         ENTRY;
2712
2713         tgt = lmv_fid2tgt(lmv, fid);
2714         if (IS_ERR(tgt))
2715                 RETURN(PTR_ERR(tgt));
2716
2717         rc = md_fsync(tgt->ltd_exp, fid, request);
2718         RETURN(rc);
2719 }
2720
2721 struct stripe_dirent {
2722         struct page             *sd_page;
2723         struct lu_dirpage       *sd_dp;
2724         struct lu_dirent        *sd_ent;
2725         bool                     sd_eof;
2726 };
2727
2728 struct lmv_dir_ctxt {
2729         struct lmv_obd          *ldc_lmv;
2730         struct md_op_data       *ldc_op_data;
2731         struct md_readdir_info  *ldc_mrinfo;
2732         __u64                    ldc_hash;
2733         int                      ldc_count;
2734         struct stripe_dirent     ldc_stripes[0];
2735 };
2736
2737 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2738 {
2739         if (stripe->sd_page) {
2740                 kunmap(stripe->sd_page);
2741                 put_page(stripe->sd_page);
2742                 stripe->sd_page = NULL;
2743                 stripe->sd_ent = NULL;
2744         }
2745 }
2746
2747 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2748 {
2749         int i;
2750
2751         for (i = 0; i < ctxt->ldc_count; i++)
2752                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2753 }
2754
2755 /* if @ent is dummy, or . .., get next */
2756 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2757                                            struct lu_dirent *ent,
2758                                            int stripe_index)
2759 {
2760         for (; ent; ent = lu_dirent_next(ent)) {
2761                 /* Skip dummy entry */
2762                 if (le16_to_cpu(ent->lde_namelen) == 0)
2763                         continue;
2764
2765                 /* skip . and .. for other stripes */
2766                 if (stripe_index &&
2767                     (strncmp(ent->lde_name, ".",
2768                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2769                      strncmp(ent->lde_name, "..",
2770                              le16_to_cpu(ent->lde_namelen)) == 0))
2771                         continue;
2772
2773                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2774                         break;
2775         }
2776
2777         return ent;
2778 }
2779
2780 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2781                                             struct stripe_dirent *stripe,
2782                                             int stripe_index)
2783 {
2784         struct md_op_data *op_data = ctxt->ldc_op_data;
2785         struct lmv_oinfo *oinfo;
2786         struct lu_fid fid = op_data->op_fid1;
2787         struct inode *inode = op_data->op_data;
2788         struct lmv_tgt_desc *tgt;
2789         struct lu_dirent *ent = stripe->sd_ent;
2790         __u64 hash = ctxt->ldc_hash;
2791         int rc = 0;
2792
2793         ENTRY;
2794
2795         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2796         LASSERT(!ent);
2797
2798         do {
2799                 if (stripe->sd_page) {
2800                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2801
2802                         /* @hash should be the last dirent hash */
2803                         LASSERTF(hash <= end,
2804                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2805                                  ctxt, stripe, hash, end);
2806                         /* unload last page */
2807                         stripe_dirent_unload(stripe);
2808                         /* eof */
2809                         if (end == MDS_DIR_END_OFF) {
2810                                 stripe->sd_eof = true;
2811                                 break;
2812                         }
2813                         hash = end;
2814                 }
2815
2816                 oinfo = &op_data->op_lso1->lso_lsm.lsm_md_oinfo[stripe_index];
2817                 if (!oinfo->lmo_root) {
2818                         rc = -ENOENT;
2819                         break;
2820                 }
2821
2822                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2823                 if (!tgt) {
2824                         rc = -ENODEV;
2825                         break;
2826                 }
2827
2828                 /* op_data is shared by stripes, reset after use */
2829                 op_data->op_fid1 = oinfo->lmo_fid;
2830                 op_data->op_fid2 = oinfo->lmo_fid;
2831                 op_data->op_data = oinfo->lmo_root;
2832
2833                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_mrinfo, hash,
2834                                   &stripe->sd_page);
2835
2836                 op_data->op_fid1 = fid;
2837                 op_data->op_fid2 = fid;
2838                 op_data->op_data = inode;
2839
2840                 if (rc)
2841                         break;
2842
2843                 stripe->sd_dp = page_address(stripe->sd_page);
2844                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2845                                         stripe_index);
2846                 /* in case a page filled with ., .. and dummy, read next */
2847         } while (!ent);
2848
2849         stripe->sd_ent = ent;
2850         if (rc) {
2851                 LASSERT(!ent);
2852                 /* treat error as eof, so dir can be partially accessed */
2853                 stripe->sd_eof = true;
2854                 ctxt->ldc_mrinfo->mr_partial_readdir_rc = rc;
2855                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2856                               "directory is partially accessed!\n",
2857                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2858                               rc);
2859         }
2860
2861         RETURN(ent);
2862 }
2863
2864 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2865 {
2866         struct obd_device *obd = exp->exp_obd;
2867         struct lmv_obd *lmv = &obd->u.lmv;
2868         struct lmv_tgt_desc *tgt;
2869         int rc;
2870
2871         ENTRY;
2872
2873         rc = lmv_check_connect(obd);
2874         if (rc != 0)
2875                 RETURN(rc);
2876
2877         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2878         if (IS_ERR(tgt))
2879                 RETURN(PTR_ERR(tgt));
2880
2881         data->op_flags |= MF_MDC_CANCEL_FID1;
2882         rc = md_file_resync(tgt->ltd_exp, data);
2883         RETURN(rc);
2884 }
2885
2886 /**
2887  * Get dirent with the closest hash for striped directory
2888  *
2889  * This function will search the dir entry, whose hash value is the
2890  * closest(>=) to hash from all of sub-stripes, and it is only being called
2891  * for striped directory.
2892  *
2893  * \param[in] ctxt              dir read context
2894  *
2895  * \retval                      dirent get the entry successfully
2896  *                              NULL does not get the entry, normally it means
2897  *                              it reaches the end of the directory, while read
2898  *                              stripe dirent error is ignored to allow partial
2899  *                              access.
2900  */
2901 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2902 {
2903         struct stripe_dirent *stripe;
2904         struct lu_dirent *ent = NULL;
2905         int i;
2906         int min = -1;
2907
2908         /* TODO: optimize with k-way merge sort */
2909         for (i = 0; i < ctxt->ldc_count; i++) {
2910                 stripe = &ctxt->ldc_stripes[i];
2911                 if (stripe->sd_eof)
2912                         continue;
2913
2914                 if (!stripe->sd_ent) {
2915                         stripe_dirent_load(ctxt, stripe, i);
2916                         if (!stripe->sd_ent) {
2917                                 LASSERT(stripe->sd_eof);
2918                                 continue;
2919                         }
2920                 }
2921
2922                 if (min == -1 ||
2923                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2924                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2925                         min = i;
2926                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2927                             ctxt->ldc_hash)
2928                                 break;
2929                 }
2930         }
2931
2932         if (min != -1) {
2933                 stripe = &ctxt->ldc_stripes[min];
2934                 ent = stripe->sd_ent;
2935                 /* pop found dirent */
2936                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2937                                                    min);
2938         }
2939
2940         return ent;
2941 }
2942
2943 /**
2944  * Build dir entry page for striped directory
2945  *
2946  * This function gets one entry by @offset from a striped directory. It will
2947  * read entries from all of stripes, and choose one closest to the required
2948  * offset(&offset). A few notes
2949  * 1. skip . and .. for non-zero stripes, because there can only have one .
2950  * and .. in a directory.
2951  * 2. op_data will be shared by all of stripes, instead of allocating new
2952  * one, so need to restore before reusing.
2953  *
2954  * \param[in] exp       obd export refer to LMV
2955  * \param[in] op_data   hold those MD parameters of read_entry
2956  * \param[in] mrinfo    ldlm callback being used in enqueue in mdc_read_entry,
2957  *                      and partial readdir result will be stored in it.
2958  * \param[in] offset    starting hash offset
2959  * \param[out] ppage    the page holding the entry. Note: because the entry
2960  *                      will be accessed in upper layer, so we need hold the
2961  *                      page until the usages of entry is finished, see
2962  *                      ll_dir_entry_next.
2963  *
2964  * retval               =0 if get entry successfully
2965  *                      <0 cannot get entry
2966  */
2967 static int lmv_striped_read_page(struct obd_export *exp,
2968                                  struct md_op_data *op_data,
2969                                  struct md_readdir_info *mrinfo, __u64 offset,
2970                                  struct page **ppage)
2971 {
2972         struct page *page = NULL;
2973         struct lu_dirpage *dp;
2974         void *start;
2975         struct lu_dirent *ent;
2976         struct lu_dirent *last_ent;
2977         int stripe_count;
2978         struct lmv_dir_ctxt *ctxt;
2979         struct lu_dirent *next = NULL;
2980         __u16 ent_size;
2981         size_t left_bytes;
2982         int rc = 0;
2983         ENTRY;
2984
2985         /* Allocate a page and read entries from all of stripes and fill
2986          * the page by hash order */
2987         page = alloc_page(GFP_KERNEL);
2988         if (!page)
2989                 RETURN(-ENOMEM);
2990
2991         /* Initialize the entry page */
2992         dp = kmap(page);
2993         memset(dp, 0, sizeof(*dp));
2994         dp->ldp_hash_start = cpu_to_le64(offset);
2995
2996         start = dp + 1;
2997         left_bytes = PAGE_SIZE - sizeof(*dp);
2998         ent = start;
2999         last_ent = ent;
3000
3001         /* initalize dir read context */
3002         stripe_count = op_data->op_lso1->lso_lsm.lsm_md_stripe_count;
3003         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
3004         if (!ctxt)
3005                 GOTO(free_page, rc = -ENOMEM);
3006         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
3007         ctxt->ldc_op_data = op_data;
3008         ctxt->ldc_mrinfo = mrinfo;
3009         ctxt->ldc_hash = offset;
3010         ctxt->ldc_count = stripe_count;
3011
3012         while (1) {
3013                 next = lmv_dirent_next(ctxt);
3014
3015                 /* end of directory */
3016                 if (!next) {
3017                         ctxt->ldc_hash = MDS_DIR_END_OFF;
3018                         break;
3019                 }
3020                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
3021
3022                 ent_size = le16_to_cpu(next->lde_reclen);
3023
3024                 /* the last entry lde_reclen is 0, but it might not be the last
3025                  * one of this temporay dir page */
3026                 if (!ent_size)
3027                         ent_size = lu_dirent_calc_size(
3028                                         le16_to_cpu(next->lde_namelen),
3029                                         le32_to_cpu(next->lde_attrs));
3030                 /* page full */
3031                 if (ent_size > left_bytes)
3032                         break;
3033
3034                 memcpy(ent, next, ent_size);
3035
3036                 /* Replace . with master FID and Replace .. with the parent FID
3037                  * of master object */
3038                 if (strncmp(ent->lde_name, ".",
3039                             le16_to_cpu(ent->lde_namelen)) == 0 &&
3040                     le16_to_cpu(ent->lde_namelen) == 1)
3041                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
3042                 else if (strncmp(ent->lde_name, "..",
3043                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
3044                            le16_to_cpu(ent->lde_namelen) == 2)
3045                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
3046
3047                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
3048                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
3049                        le64_to_cpu(ent->lde_hash));
3050
3051                 left_bytes -= ent_size;
3052                 ent->lde_reclen = cpu_to_le16(ent_size);
3053                 last_ent = ent;
3054                 ent = (void *)ent + ent_size;
3055         };
3056
3057         last_ent->lde_reclen = 0;
3058
3059         if (ent == start)
3060                 dp->ldp_flags |= LDF_EMPTY;
3061         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
3062                 dp->ldp_flags |= LDF_COLLIDE;
3063         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
3064         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
3065
3066         put_lmv_dir_ctxt(ctxt);
3067         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
3068
3069         *ppage = page;
3070
3071         RETURN(0);
3072
3073 free_page:
3074         kunmap(page);
3075         __free_page(page);
3076
3077         return rc;
3078 }
3079
3080 static int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
3081                          struct md_readdir_info *mrinfo, __u64 offset,
3082                          struct page **ppage)
3083 {
3084         struct obd_device *obd = exp->exp_obd;
3085         struct lmv_obd *lmv = &obd->u.lmv;
3086         struct lmv_tgt_desc *tgt;
3087         int rc;
3088
3089         ENTRY;
3090
3091         if (unlikely(lmv_dir_foreign(op_data->op_lso1)))
3092                 RETURN(-ENODATA);
3093
3094         if (unlikely(lmv_dir_striped(op_data->op_lso1))) {
3095                 rc = lmv_striped_read_page(exp, op_data, mrinfo, offset, ppage);
3096                 RETURN(rc);
3097         }
3098
3099         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
3100         if (IS_ERR(tgt))
3101                 RETURN(PTR_ERR(tgt));
3102
3103         rc = md_read_page(tgt->ltd_exp, op_data, mrinfo, offset, ppage);
3104
3105         RETURN(rc);
3106 }
3107
3108 /**
3109  * Unlink a file/directory
3110  *
3111  * Unlink a file or directory under the parent dir. The unlink request
3112  * usually will be sent to the MDT where the child is located, but if
3113  * the client does not have the child FID then request will be sent to the
3114  * MDT where the parent is located.
3115  *
3116  * If the parent is a striped directory then it also needs to locate which
3117  * stripe the name of the child is located, and replace the parent FID
3118  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
3119  * it will walk through all of sub-stripes until the child is being
3120  * unlinked finally.
3121  *
3122  * \param[in] exp       export refer to LMV
3123  * \param[in] op_data   different parameters transferred beween client
3124  *                      MD stacks, name, namelen, FIDs etc.
3125  *                      op_fid1 is the parent FID, op_fid2 is the child
3126  *                      FID.
3127  * \param[out] request  point to the request of unlink.
3128  *
3129  * retval               0 if succeed
3130  *                      negative errno if failed.
3131  */
3132 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
3133                       struct ptlrpc_request **request)
3134 {
3135         struct obd_device *obd = exp->exp_obd;
3136         struct lmv_obd *lmv = &obd->u.lmv;
3137         struct lmv_tgt_desc *tgt;
3138         struct lmv_tgt_desc *parent_tgt;
3139         struct mdt_body *body;
3140         int rc;
3141
3142         ENTRY;
3143
3144         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
3145         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
3146         op_data->op_cap = current_cap();
3147
3148 retry:
3149         parent_tgt = lmv_locate_tgt(lmv, op_data);
3150         if (IS_ERR(parent_tgt))
3151                 RETURN(PTR_ERR(parent_tgt));
3152
3153         if (likely(!fid_is_zero(&op_data->op_fid2))) {
3154                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3155                 if (IS_ERR(tgt))
3156                         RETURN(PTR_ERR(tgt));
3157         } else {
3158                 tgt = parent_tgt;
3159         }
3160
3161         /*
3162          * If child's fid is given, cancel unused locks for it if it is from
3163          * another export than parent.
3164          *
3165          * LOOKUP lock for child (fid3) should also be cancelled on parent
3166          * tgt_tgt in mdc_unlink().
3167          */
3168         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
3169
3170         if (parent_tgt != tgt)
3171                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
3172                                       LCK_EX, MDS_INODELOCK_LOOKUP,
3173                                       MF_MDC_CANCEL_FID3);
3174
3175         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
3176                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
3177         if (rc)
3178                 RETURN(rc);
3179
3180         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
3181                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
3182                tgt->ltd_index);
3183
3184         rc = md_unlink(tgt->ltd_exp, op_data, request);
3185         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
3186                 ptlrpc_req_finished(*request);
3187                 *request = NULL;
3188                 goto retry;
3189         }
3190
3191         if (rc != -EREMOTE)
3192                 RETURN(rc);
3193
3194         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
3195         if (body == NULL)
3196                 RETURN(-EPROTO);
3197
3198         /* Not cross-ref case, just get out of here. */
3199         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
3200                 RETURN(rc);
3201
3202         /* This is a remote object, try remote MDT. */
3203         op_data->op_fid2 = body->mbo_fid1;
3204         ptlrpc_req_finished(*request);
3205         *request = NULL;
3206
3207         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3208         if (IS_ERR(tgt))
3209                 RETURN(PTR_ERR(tgt));
3210
3211         goto retry;
3212 }
3213
3214 static int lmv_precleanup(struct obd_device *obd)
3215 {
3216         ENTRY;
3217         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
3218         RETURN(0);
3219 }
3220
3221 /**
3222  * Get by key a value associated with a LMV device.
3223  *
3224  * Dispatch request to lower-layer devices as needed.
3225  *
3226  * \param[in] env               execution environment for this thread
3227  * \param[in] exp               export for the LMV device
3228  * \param[in] keylen            length of key identifier
3229  * \param[in] key               identifier of key to get value for
3230  * \param[in] vallen            size of \a val
3231  * \param[out] val              pointer to storage location for value
3232  * \param[in] lsm               optional striping metadata of object
3233  *
3234  * \retval 0            on success
3235  * \retval negative     negated errno on failure
3236  */
3237 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
3238                         __u32 keylen, void *key, __u32 *vallen, void *val)
3239 {
3240         struct obd_device *obd;
3241         struct lmv_obd *lmv;
3242         struct lu_tgt_desc *tgt;
3243         int rc = 0;
3244
3245         ENTRY;
3246
3247         obd = class_exp2obd(exp);
3248         if (obd == NULL) {
3249                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3250                        exp->exp_handle.h_cookie);
3251                 RETURN(-EINVAL);
3252         }
3253
3254         lmv = &obd->u.lmv;
3255         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
3256                 LASSERT(*vallen == sizeof(__u32));
3257                 lmv_foreach_connected_tgt(lmv, tgt) {
3258                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
3259                                           vallen, val))
3260                                 RETURN(0);
3261                 }
3262                 RETURN(-EINVAL);
3263         } else if (KEY_IS(KEY_MAX_EASIZE) ||
3264                    KEY_IS(KEY_DEFAULT_EASIZE) ||
3265                    KEY_IS(KEY_CONN_DATA)) {
3266                 /*
3267                  * Forwarding this request to first MDS, it should know LOV
3268                  * desc.
3269                  */
3270                 tgt = lmv_tgt(lmv, 0);
3271                 if (!tgt)
3272                         RETURN(-ENODEV);
3273
3274                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
3275                 if (!rc && KEY_IS(KEY_CONN_DATA))
3276                         exp->exp_connect_data = *(struct obd_connect_data *)val;
3277                 RETURN(rc);
3278         } else if (KEY_IS(KEY_TGT_COUNT)) {
3279                 *((int *)val) = lmv->lmv_mdt_descs.ltd_tgts_size;
3280                 RETURN(0);
3281         }
3282
3283         CDEBUG(D_IOCTL, "Invalid key\n");
3284         RETURN(-EINVAL);
3285 }
3286
3287 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
3288                      int *__rcs, struct ptlrpc_request_set *_set)
3289 {
3290         struct obd_device *obd = class_exp2obd(exp);
3291         struct ptlrpc_request_set *set = _set;
3292         struct lmv_obd *lmv = &obd->u.lmv;
3293         int tgt_count = lmv->lmv_mdt_count;
3294         struct lu_tgt_desc *tgt;
3295         struct fid_array *fat, **fas = NULL;
3296         int i, rc, **rcs = NULL;
3297
3298         if (!set) {
3299                 set = ptlrpc_prep_set();
3300                 if (!set)
3301                         RETURN(-ENOMEM);
3302         }
3303
3304         /* split FIDs by targets */
3305         OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
3306         if (fas == NULL)
3307                 GOTO(out, rc = -ENOMEM);
3308         OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
3309         if (rcs == NULL)
3310                 GOTO(out_fas, rc = -ENOMEM);
3311
3312         for (i = 0; i < fa->fa_nr; i++) {
3313                 unsigned int idx;
3314
3315                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
3316                 if (rc) {
3317                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
3318                                PFID(&fa->fa_fids[i]), rc);
3319                         continue;
3320                 }
3321                 LASSERT(idx < tgt_count);
3322                 if (!fas[idx])
3323                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
3324                                   fa_fids[fa->fa_nr]));
3325                 if (!fas[idx])
3326                         GOTO(out, rc = -ENOMEM);
3327                 if (!rcs[idx])
3328                         OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
3329                 if (!rcs[idx])
3330                         GOTO(out, rc = -ENOMEM);
3331
3332                 fat = fas[idx];
3333                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
3334         }
3335
3336         lmv_foreach_connected_tgt(lmv, tgt) {
3337                 fat = fas[tgt->ltd_index];
3338                 if (!fat || fat->fa_nr == 0)
3339                         continue;
3340                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3341         }
3342
3343         rc = ptlrpc_set_wait(NULL, set);
3344         if (rc == 0) {
3345                 int j = 0;
3346                 for (i = 0; i < tgt_count; i++) {
3347                         fat = fas[i];
3348                         if (!fat || fat->fa_nr == 0)
3349                                 continue;
3350                         /* copy FIDs back */
3351                         memcpy(fa->fa_fids + j, fat->fa_fids,
3352                                fat->fa_nr * sizeof(struct lu_fid));
3353                         /* copy rcs back */
3354                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3355                         j += fat->fa_nr;
3356                 }
3357         }
3358         if (set != _set)
3359                 ptlrpc_set_destroy(set);
3360
3361 out:
3362         for (i = 0; i < tgt_count; i++) {
3363                 if (fas && fas[i])
3364                         OBD_FREE(fas[i], offsetof(struct fid_array,
3365                                                 fa_fids[fa->fa_nr]));
3366                 if (rcs && rcs[i])
3367                         OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3368         }
3369         if (rcs)
3370                 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3371 out_fas:
3372         if (fas)
3373                 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3374
3375         RETURN(rc);
3376 }
3377
3378 /**
3379  * Asynchronously set by key a value associated with a LMV device.
3380  *
3381  * Dispatch request to lower-layer devices as needed.
3382  *
3383  * \param[in] env       execution environment for this thread
3384  * \param[in] exp       export for the LMV device
3385  * \param[in] keylen    length of key identifier
3386  * \param[in] key       identifier of key to store value for
3387  * \param[in] vallen    size of value to store
3388  * \param[in] val       pointer to data to be stored
3389  * \param[in] set       optional list of related ptlrpc requests
3390  *
3391  * \retval 0            on success
3392  * \retval negative     negated errno on failure
3393  */
3394 static int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3395                               __u32 keylen, void *key, __u32 vallen, void *val,
3396                               struct ptlrpc_request_set *set)
3397 {
3398         struct lmv_tgt_desc *tgt;
3399         struct obd_device *obd;
3400         struct lmv_obd *lmv;
3401         int rc = 0;
3402         ENTRY;
3403
3404         obd = class_exp2obd(exp);
3405         if (obd == NULL) {
3406                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3407                        exp->exp_handle.h_cookie);
3408                 RETURN(-EINVAL);
3409         }
3410         lmv = &obd->u.lmv;
3411
3412         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3413             KEY_IS(KEY_DEFAULT_EASIZE)) {
3414                 int err = 0;
3415
3416                 lmv_foreach_connected_tgt(lmv, tgt) {
3417                         err = obd_set_info_async(env, tgt->ltd_exp,
3418                                                  keylen, key, vallen, val, set);
3419                         if (err && rc == 0)
3420                                 rc = err;
3421                 }
3422
3423                 RETURN(rc);
3424         }
3425
3426         RETURN(-EINVAL);
3427 }
3428
3429 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3430                             const struct lmv_mds_md_v1 *lmm1)
3431 {
3432         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3433         int             stripe_count;
3434         int             cplen;
3435         int             i;
3436         int             rc = 0;
3437         ENTRY;
3438
3439         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3440         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3441         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3442         if (CFS_FAIL_CHECK(OBD_FAIL_LMV_UNKNOWN_STRIPE))
3443                 lsm->lsm_md_hash_type = cfs_fail_val ?: LMV_HASH_TYPE_UNKNOWN;
3444         else
3445                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3446         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3447         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3448         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3449         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3450                         sizeof(lsm->lsm_md_pool_name));
3451
3452         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3453                 RETURN(-E2BIG);
3454
3455         CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3456                "layout_version %d\n", lsm->lsm_md_stripe_count,
3457                lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3458                lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3459                lsm->lsm_md_layout_version);
3460
3461         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3462         for (i = 0; i < stripe_count; i++) {
3463                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3464                               &lmm1->lmv_stripe_fids[i]);
3465                 /*
3466                  * set default value -1, so lmv_locate_tgt() knows this stripe
3467                  * target is not initialized.
3468                  */
3469                 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3470                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3471                         continue;
3472
3473                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3474                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3475                 if (rc == -ENOENT)
3476                         continue;
3477
3478                 if (rc)
3479                         RETURN(rc);
3480
3481                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3482                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3483         }
3484
3485         RETURN(rc);
3486 }
3487
3488 static inline int lmv_unpack_user_md(struct obd_export *exp,
3489                                      struct lmv_stripe_md *lsm,
3490                                      const struct lmv_user_md *lmu)
3491 {
3492         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3493         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3494         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3495         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3496         lsm->lsm_md_max_inherit = lmu->lum_max_inherit;
3497         lsm->lsm_md_max_inherit_rr = lmu->lum_max_inherit_rr;
3498         lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3499
3500         return 0;
3501 }
3502
3503 struct lmv_stripe_object *lmv_stripe_object_alloc(__u32 magic,
3504                                                   const union lmv_mds_md *lmm,
3505                                                   size_t lmm_size)
3506 {
3507         struct lmv_stripe_object *lsm_obj;
3508         int size;
3509
3510         if (magic == LMV_MAGIC_FOREIGN) {
3511                 struct lmv_foreign_md *lfm;
3512
3513                 size = offsetof(typeof(*lfm), lfm_value[0]);
3514                 if (lmm_size < size)
3515                         RETURN(ERR_PTR(-EPROTO));
3516
3517                 size += le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3518                 if (lmm_size < size)
3519                         RETURN(ERR_PTR(-EPROTO));
3520
3521                 OBD_ALLOC_LARGE(lsm_obj, lmm_size +
3522                                 offsetof(typeof(*lsm_obj), lso_lfm));
3523         } else {
3524                 if (magic == LMV_MAGIC_V1) {
3525                         int count;
3526
3527                         size = offsetof(struct lmv_mds_md_v1,
3528                                         lmv_stripe_fids[0]);
3529                         if (lmm_size < size)
3530                                 RETURN(ERR_PTR(-EPROTO));
3531
3532                         count = lmv_mds_md_stripe_count_get(lmm);
3533                         size += count * sizeof(struct lu_fid);
3534                         if (lmm_size < size)
3535                                 RETURN(ERR_PTR(-EPROTO));
3536
3537                         size = lmv_stripe_md_size(count);
3538                 } else {
3539                         if (lmm && lmm_size < sizeof(struct lmv_user_md))
3540                                 RETURN(ERR_PTR(-EPROTO));
3541
3542                         /**
3543                          * Unpack default dirstripe(lmv_user_md) to
3544                          * lmv_stripe_md, stripecount should be 0 then.
3545                          */
3546                         size = lmv_stripe_md_size(0);
3547                 }
3548                 size += offsetof(typeof(*lsm_obj), lso_lsm);
3549                 OBD_ALLOC(lsm_obj, size);
3550         }
3551
3552         if (lsm_obj) {
3553                 atomic_set(&lsm_obj->lso_refs, 1);
3554                 RETURN(lsm_obj);
3555         }
3556
3557         RETURN(ERR_PTR(-ENOMEM));
3558 }
3559 EXPORT_SYMBOL(lmv_stripe_object_alloc);
3560
3561 static int lmv_stripe_object_create(struct obd_export *exp,
3562                                     struct lmv_stripe_object **lsop,
3563                                     const union lmv_mds_md *lmm,
3564                                     size_t lmm_size)
3565 {
3566         struct lmv_stripe_object *lsm_obj;
3567         __u32 magic;
3568         int rc;
3569         ENTRY;
3570
3571         LASSERT(lsop != NULL && *lsop == NULL);
3572
3573         if (lmm_size == 0)
3574                 RETURN(-EPROTO);
3575
3576         magic = le32_to_cpu(lmm->lmv_magic);
3577         if (magic == LMV_MAGIC_STRIPE)
3578                 RETURN(-EPERM);
3579
3580         if (magic != LMV_MAGIC_V1 && magic != LMV_USER_MAGIC &&
3581             magic != LMV_MAGIC_FOREIGN) {
3582                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3583                        exp->exp_obd->obd_name, magic, -EIO);
3584                 RETURN(-EIO);
3585         }
3586
3587         /* foreign lmv case */
3588         if (magic == LMV_MAGIC_FOREIGN) {
3589                 struct lmv_foreign_md *lfm;
3590
3591                 lsm_obj = lmv_stripe_object_alloc(magic, lmm, lmm_size);
3592                 if (IS_ERR(lsm_obj))
3593                         RETURN(PTR_ERR(lsm_obj));
3594
3595                 *lsop = lsm_obj;
3596                 lfm = &lsm_obj->lso_lfm;
3597                 lfm->lfm_magic = magic;
3598                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3599                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3600                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3601                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3602                        lfm->lfm_length);
3603                 RETURN(0);
3604         }
3605
3606         /* Unpack memmd */
3607         lsm_obj = lmv_stripe_object_alloc(magic, lmm, lmm_size);
3608         if (IS_ERR(lsm_obj))
3609                 RETURN(PTR_ERR(lsm_obj));
3610
3611         switch (magic) {
3612         case LMV_MAGIC_V1:
3613                 rc = lmv_unpack_md_v1(exp, &lsm_obj->lso_lsm, &lmm->lmv_md_v1);
3614                 break;
3615         case LMV_USER_MAGIC:
3616                 rc = lmv_unpack_user_md(exp, &lsm_obj->lso_lsm,
3617                                         &lmm->lmv_user_md);
3618                 break;
3619         default:
3620                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3621                        magic);
3622                 rc = -EINVAL;
3623                 break;
3624         }
3625
3626         if (rc != 0)
3627                 lmv_stripe_object_put(&lsm_obj);
3628
3629         *lsop = lsm_obj;
3630         RETURN(rc);
3631 }
3632
3633 struct lmv_stripe_object *
3634 lmv_stripe_object_get(struct lmv_stripe_object *lsm_obj)
3635 {
3636         if (lsm_obj == NULL)
3637                 return NULL;
3638
3639         atomic_inc(&lsm_obj->lso_refs);
3640         CDEBUG(D_INODE, "get %p %u\n", lsm_obj,
3641                atomic_read(&lsm_obj->lso_refs));
3642         return lsm_obj;
3643 }
3644 EXPORT_SYMBOL(lmv_stripe_object_get);
3645
3646 void lmv_stripe_object_put(struct lmv_stripe_object **lsop)
3647 {
3648         struct lmv_stripe_object *lsm_obj;
3649         size_t size;
3650         int i;
3651
3652         LASSERT(lsop != NULL);
3653
3654         lsm_obj = *lsop;
3655         if (lsm_obj == NULL)
3656                 return;
3657
3658         *lsop = NULL;
3659         CDEBUG(D_INODE, "put %p %u\n", lsm_obj,
3660                atomic_read(&lsm_obj->lso_refs) - 1);
3661
3662         if (!atomic_dec_and_test(&lsm_obj->lso_refs))
3663                 return;
3664
3665         if (lmv_dir_foreign(lsm_obj)) {
3666                 size = lsm_obj->lso_lfm.lfm_length +
3667                         offsetof(typeof(lsm_obj->lso_lfm), lfm_value[0]) +
3668                         offsetof(typeof(*lsm_obj), lso_lsm);
3669                 OBD_FREE_LARGE(lsm_obj, size);
3670                 return;
3671         }
3672
3673         if (lmv_dir_striped(lsm_obj)) {
3674                 struct lmv_stripe_md *lsm = &lsm_obj->lso_lsm;
3675
3676                 for (i = 0; i < lsm->lsm_md_stripe_count; i++)
3677                         iput(lsm->lsm_md_oinfo[i].lmo_root);
3678                 size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3679         } else {
3680                 size = lmv_stripe_md_size(0);
3681         }
3682         OBD_FREE(lsm_obj, size + offsetof(typeof(*lsm_obj), lso_lsm));
3683 }
3684 EXPORT_SYMBOL(lmv_stripe_object_put);
3685
3686 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3687                              union ldlm_policy_data *policy,
3688                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3689                              void *opaque)
3690 {
3691         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3692         struct lu_tgt_desc *tgt;
3693         int err;
3694         int rc = 0;
3695
3696         ENTRY;
3697
3698         LASSERT(fid != NULL);
3699
3700         lmv_foreach_connected_tgt(lmv, tgt) {
3701                 if (!tgt->ltd_active)
3702                         continue;
3703
3704                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3705                                        opaque);
3706                 if (!rc)
3707                         rc = err;
3708         }
3709         RETURN(rc);
3710 }
3711
3712 static int lmv_set_lock_data(struct obd_export *exp,
3713                              const struct lustre_handle *lockh,
3714                              void *data, __u64 *bits)
3715 {
3716         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3717         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3718         int rc;
3719
3720         ENTRY;
3721
3722         if (tgt == NULL || tgt->ltd_exp == NULL)
3723                 RETURN(-EINVAL);
3724         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3725         RETURN(rc);
3726 }
3727
3728 static enum ldlm_mode
3729 lmv_lock_match(struct obd_export *exp, __u64 flags,
3730                const struct lu_fid *fid, enum ldlm_type type,
3731                union ldlm_policy_data *policy,
3732                enum ldlm_mode mode, struct lustre_handle *lockh)
3733 {
3734         struct obd_device *obd = exp->exp_obd;
3735         struct lmv_obd *lmv = &obd->u.lmv;
3736         struct lu_tgt_desc *tgt;
3737         __u64 bits = policy->l_inodebits.bits;
3738         enum ldlm_mode rc = LCK_MINMODE;
3739         int index;
3740         int i;
3741
3742         /* only one bit is set */
3743         LASSERT(bits && !(bits & (bits - 1)));
3744         /* With DNE every object can have two locks in different namespaces:
3745          * lookup lock in space of MDT storing direntry and update/open lock in
3746          * space of MDT storing inode.  Try the MDT that the FID maps to first,
3747          * since this can be easily found, and only try others if that fails.
3748          */
3749         if (bits == MDS_INODELOCK_LOOKUP) {
3750                 for (i = 0, index = lmv_fid2tgt_index(lmv, fid);
3751                      i < lmv->lmv_mdt_descs.ltd_tgts_size; i++,
3752                      index = (index + 1) % lmv->lmv_mdt_descs.ltd_tgts_size) {
3753                         if (index < 0) {
3754                                 CDEBUG(D_HA,
3755                                        "%s: "DFID" is inaccessible: rc = %d\n",
3756                                        obd->obd_name, PFID(fid), index);
3757                                 index = 0;
3758                         }
3759                         tgt = lmv_tgt(lmv, index);
3760                         if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
3761                                 continue;
3762                         rc = md_lock_match(tgt->ltd_exp, flags, fid, type,
3763                                            policy, mode, lockh);
3764                         if (rc)
3765                                 break;
3766                 }
3767         } else {
3768                 tgt = lmv_fid2tgt(lmv, fid);
3769                 if (!IS_ERR(tgt) && tgt->ltd_exp && tgt->ltd_active)
3770                         rc = md_lock_match(tgt->ltd_exp, flags, fid, type,
3771                                            policy, mode, lockh);
3772         }
3773
3774         CDEBUG(D_INODE, "Lock match for "DFID": %d\n", PFID(fid), rc);
3775
3776         return rc;
3777 }
3778
3779 static int
3780 lmv_get_lustre_md(struct obd_export *exp, struct req_capsule *pill,
3781                   struct obd_export *dt_exp, struct obd_export *md_exp,
3782                   struct lustre_md *md)
3783 {
3784         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3785         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3786
3787         if (!tgt || !tgt->ltd_exp)
3788                 return -EINVAL;
3789
3790         return md_get_lustre_md(tgt->ltd_exp, pill, dt_exp, md_exp, md);
3791 }
3792
3793 static int lmv_put_lustre_md(struct obd_export *exp, struct lustre_md *md)
3794 {
3795         struct obd_device *obd = exp->exp_obd;
3796         struct lmv_obd *lmv = &obd->u.lmv;
3797         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3798         ENTRY;
3799
3800         lmv_stripe_object_put(&md->def_lsm_obj);
3801         lmv_stripe_object_put(&md->lsm_obj);
3802
3803         if (!tgt || !tgt->ltd_exp)
3804                 RETURN(-EINVAL);
3805         RETURN(0);
3806 }
3807
3808 static int lmv_set_open_replay_data(struct obd_export *exp,
3809                                     struct obd_client_handle *och,
3810                                     struct lookup_intent *it)
3811 {
3812         struct obd_device *obd = exp->exp_obd;
3813         struct lmv_obd *lmv = &obd->u.lmv;
3814         struct lmv_tgt_desc *tgt;
3815
3816         ENTRY;
3817
3818         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3819         if (IS_ERR(tgt))
3820                 RETURN(PTR_ERR(tgt));
3821
3822         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3823 }
3824
3825 static int lmv_clear_open_replay_data(struct obd_export *exp,
3826                                       struct obd_client_handle *och)
3827 {
3828         struct obd_device *obd = exp->exp_obd;
3829         struct lmv_obd *lmv = &obd->u.lmv;
3830         struct lmv_tgt_desc *tgt;
3831
3832         ENTRY;
3833
3834         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3835         if (IS_ERR(tgt))
3836                 RETURN(PTR_ERR(tgt));
3837
3838         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3839 }
3840
3841 static int lmv_intent_getattr_async(struct obd_export *exp,
3842                                     struct md_op_item *item)
3843 {
3844         struct md_op_data *op_data = &item->mop_data;
3845         struct obd_device *obd = exp->exp_obd;
3846         struct lmv_obd *lmv = &obd->u.lmv;
3847         struct lmv_tgt_desc *ptgt;
3848         struct lmv_tgt_desc *ctgt;
3849         int rc;
3850
3851         ENTRY;
3852
3853         if (!(fid_is_sane(&op_data->op_fid2) ||
3854               fid_is_zero(&op_data->op_fid2)))
3855                 RETURN(-EINVAL);
3856
3857         ptgt = lmv_locate_tgt(lmv, op_data);
3858         if (IS_ERR(ptgt))
3859                 RETURN(PTR_ERR(ptgt));
3860
3861         /*
3862          * Zeroed FID @op_fid2 means that the intent getattr() comes from
3863          * statahead by regularized file names. Currently only do statahead
3864          * for the children files located same as the parent directory.
3865          */
3866         if (!fid_is_zero(&op_data->op_fid2)) {
3867                 ctgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3868                 if (IS_ERR(ctgt))
3869                         RETURN(PTR_ERR(ctgt));
3870
3871                 /*
3872                  * remote object needs two RPCs to lookup and getattr,
3873                  * considering the complexity don't support statahead for now.
3874                  */
3875                 if (ctgt != ptgt)
3876                         RETURN(-EREMOTE);
3877         }
3878
3879         rc = md_intent_getattr_async(ptgt->ltd_exp, item);
3880
3881         RETURN(rc);
3882 }
3883
3884 static int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3885                                struct lu_fid *fid, __u64 *bits)
3886 {
3887         struct obd_device *obd = exp->exp_obd;
3888         struct lmv_obd *lmv = &obd->u.lmv;
3889         struct lmv_tgt_desc *tgt;
3890         int rc;
3891
3892         ENTRY;
3893
3894         tgt = lmv_fid2tgt(lmv, fid);
3895         if (IS_ERR(tgt))
3896                 RETURN(PTR_ERR(tgt));
3897
3898         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3899         RETURN(rc);
3900 }
3901
3902 static int lmv_get_fid_from_lsm(struct obd_export *exp,
3903                                 const struct lmv_stripe_object *lso,
3904                                 const char *name, int namelen,
3905                                 struct lu_fid *fid)
3906 {
3907         const struct lmv_oinfo *oinfo;
3908
3909         LASSERT(lmv_dir_striped(lso));
3910         oinfo = lsm_name_to_stripe_info(lso, name, namelen, false);
3911         if (IS_ERR(oinfo))
3912                 return PTR_ERR(oinfo);
3913
3914         *fid = oinfo->lmo_fid;
3915
3916         RETURN(0);
3917 }
3918
3919 /**
3920  * For lmv, only need to send request to master MDT, and the master MDT will
3921  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3922  * we directly fetch data from the slave MDTs.
3923  */
3924 static int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3925                         struct obd_quotactl *oqctl)
3926 {
3927         struct obd_device *obd = class_exp2obd(exp);
3928         struct lmv_obd *lmv = &obd->u.lmv;
3929         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3930         __u64 curspace, curinodes;
3931         int rc = 0;
3932
3933         ENTRY;
3934
3935         if (!tgt || !tgt->ltd_exp || !tgt->ltd_active) {
3936                 CERROR("master lmv inactive\n");
3937                 RETURN(-EIO);
3938         }
3939
3940         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3941                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3942                 RETURN(rc);
3943         }
3944
3945         curspace = curinodes = 0;
3946         lmv_foreach_connected_tgt(lmv, tgt) {
3947                 int err;
3948
3949                 if (!tgt->ltd_active)
3950                         continue;
3951
3952                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3953                 if (err) {
3954                         CERROR("getquota on mdt %d failed. %d\n",
3955                                tgt->ltd_index, err);
3956                         if (!rc)
3957                                 rc = err;
3958                 } else {
3959                         curspace += oqctl->qc_dqblk.dqb_curspace;
3960                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3961                 }
3962         }
3963         oqctl->qc_dqblk.dqb_curspace = curspace;
3964         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3965
3966         RETURN(rc);
3967 }
3968
3969 static int lmv_merge_attr(struct obd_export *exp,
3970                           const struct lmv_stripe_object *lso,
3971                           struct cl_attr *attr,
3972                           ldlm_blocking_callback cb_blocking)
3973 {
3974         const struct lmv_stripe_md *lsm = &lso->lso_lsm;
3975         int rc;
3976         int i;
3977
3978         if (!lmv_dir_striped(lso))
3979                 return 0;
3980
3981         rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3982         if (rc < 0)
3983                 return rc;
3984
3985         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3986                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3987
3988                 if (!inode)
3989                         continue;
3990
3991                 CDEBUG(D_INFO,
3992                        "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
3993                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3994                        i_size_read(inode), (unsigned long long)inode->i_blocks,
3995                        inode->i_nlink, (s64)inode->i_atime.tv_sec,
3996                        (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
3997
3998                 /* for slave stripe, it needs to subtract nlink for . and .. */
3999                 if (i != 0)
4000                         attr->cat_nlink += inode->i_nlink - 2;
4001                 else
4002                         attr->cat_nlink = inode->i_nlink;
4003
4004                 attr->cat_size += i_size_read(inode);
4005                 attr->cat_blocks += inode->i_blocks;
4006
4007                 if (attr->cat_atime < inode->i_atime.tv_sec)
4008                         attr->cat_atime = inode->i_atime.tv_sec;
4009
4010                 if (attr->cat_ctime < inode->i_ctime.tv_sec)
4011                         attr->cat_ctime = inode->i_ctime.tv_sec;
4012
4013                 if (attr->cat_mtime < inode->i_mtime.tv_sec)
4014                         attr->cat_mtime = inode->i_mtime.tv_sec;
4015         }
4016         return 0;
4017 }
4018
4019 static struct lu_batch *lmv_batch_create(struct obd_export *exp,
4020                                          enum lu_batch_flags flags,
4021                                          __u32 max_count)
4022 {
4023         struct lu_batch *bh;
4024         struct lmv_batch *lbh;
4025
4026         ENTRY;
4027         OBD_ALLOC_PTR(lbh);
4028         if (!lbh)
4029                 RETURN(ERR_PTR(-ENOMEM));
4030
4031         bh = &lbh->lbh_super;
4032         bh->lbt_flags = flags;
4033         bh->lbt_max_count = max_count;
4034
4035         if (flags & BATCH_FL_RQSET) {
4036                 bh->lbt_rqset = ptlrpc_prep_set();
4037                 if (bh->lbt_rqset == NULL) {
4038                         OBD_FREE_PTR(lbh);
4039                         RETURN(ERR_PTR(-ENOMEM));
4040                 }
4041         }
4042
4043         INIT_LIST_HEAD(&lbh->lbh_sub_batch_list);
4044         RETURN(bh);
4045 }
4046
4047 static int lmv_batch_stop(struct obd_export *exp, struct lu_batch *bh)
4048 {
4049         struct lmv_batch *lbh;
4050         struct lmvsub_batch *sub;
4051         struct lmvsub_batch *tmp;
4052         int rc = 0;
4053
4054         ENTRY;
4055
4056         lbh = container_of(bh, struct lmv_batch, lbh_super);
4057         list_for_each_entry_safe(sub, tmp, &lbh->lbh_sub_batch_list,
4058                                  sbh_sub_item) {
4059                 list_del(&sub->sbh_sub_item);
4060                 rc = md_batch_stop(sub->sbh_tgt->ltd_exp, sub->sbh_sub);
4061                 if (rc < 0) {
4062                         CERROR("%s: stop batch processing failed: rc = %d\n",
4063                                exp->exp_obd->obd_name, rc);
4064                         if (bh->lbt_result == 0)
4065                                 bh->lbt_result = rc;
4066                 }
4067                 OBD_FREE_PTR(sub);
4068         }
4069
4070         if (bh->lbt_flags & BATCH_FL_RQSET) {
4071                 rc = ptlrpc_set_wait(NULL, bh->lbt_rqset);
4072                 ptlrpc_set_destroy(bh->lbt_rqset);
4073         }
4074
4075         OBD_FREE_PTR(lbh);
4076         RETURN(rc);
4077 }
4078
4079 static int lmv_batch_flush(struct obd_export *exp, struct lu_batch *bh,
4080                            bool wait)
4081 {
4082         struct lmv_batch *lbh;
4083         struct lmvsub_batch *sub;
4084         int rc = 0;
4085         int rc1;
4086
4087         ENTRY;
4088
4089         lbh = container_of(bh, struct lmv_batch, lbh_super);
4090         list_for_each_entry(sub, &lbh->lbh_sub_batch_list, sbh_sub_item) {
4091                 rc1 = md_batch_flush(sub->sbh_tgt->ltd_exp, sub->sbh_sub, wait);
4092                 if (rc1 < 0) {
4093                         CERROR("%s: stop batch processing failed: rc = %d\n",
4094                                exp->exp_obd->obd_name, rc);
4095                         if (bh->lbt_result == 0)
4096                                 bh->lbt_result = rc;
4097
4098                         if (rc == 0)
4099                                 rc = rc1;
4100                 }
4101         }
4102
4103         if (wait && bh->lbt_flags & BATCH_FL_RQSET) {
4104                 rc1 = ptlrpc_set_wait(NULL, bh->lbt_rqset);
4105                 if (rc == 0)
4106                         rc = rc1;
4107         }
4108
4109         RETURN(rc);
4110 }
4111
4112 static inline struct lmv_tgt_desc *
4113 lmv_batch_locate_tgt(struct lmv_obd *lmv, struct md_op_item *item)
4114 {
4115         struct md_op_data *op_data = &item->mop_data;
4116         struct lmv_tgt_desc *tgt;
4117
4118         switch (item->mop_opc) {
4119         case MD_OP_GETATTR: {
4120                 struct lmv_tgt_desc *ptgt;
4121
4122                 if (!(fid_is_sane(&op_data->op_fid2) ||
4123                       fid_is_zero(&op_data->op_fid2)))
4124                         RETURN(ERR_PTR(-EINVAL));
4125
4126                 ptgt = lmv_locate_tgt(lmv, op_data);
4127                 if (IS_ERR(ptgt))
4128                         RETURN(ptgt);
4129
4130                 /*
4131                  * Zeroed @op_fid2 means that it is a statahead populating call
4132                  * in the file name pattern which is using file name format to
4133                  * prefetch the attributes. Thus it has no idea about the FID of
4134                  * the children file. The children file is considered to be
4135                  * located on the same storage target with the parent directory
4136                  * or the stripped directory.
4137                  */
4138                 if (fid_is_zero(&op_data->op_fid2)) {
4139                         tgt = ptgt;
4140                         break;
4141                 }
4142
4143                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
4144                 if (IS_ERR(tgt))
4145                         RETURN(tgt);
4146
4147                 /*
4148                  * Remote object needs two RPCs to lookup and getattr,
4149                  * considering the complexity don't support statahead for now.
4150                  */
4151                 if (tgt != ptgt)
4152                         RETURN(ERR_PTR(-EREMOTE));
4153
4154                 break;
4155         }
4156         default:
4157                 tgt = ERR_PTR(-ENOTSUPP);
4158         }
4159
4160         return tgt;
4161 }
4162
4163 static struct lu_batch *lmv_batch_lookup_sub(struct lmv_batch *lbh,
4164                                              struct lmv_tgt_desc *tgt)
4165 {
4166         struct lmvsub_batch *sub;
4167
4168         list_for_each_entry(sub, &lbh->lbh_sub_batch_list, sbh_sub_item) {
4169                 if (sub->sbh_tgt == tgt)
4170                         return sub->sbh_sub;
4171         }
4172
4173         return NULL;
4174 }
4175
4176 static struct lu_batch *lmv_batch_get_sub(struct lmv_batch *lbh,
4177                                           struct lmv_tgt_desc *tgt)
4178 {
4179         struct lmvsub_batch *sbh;
4180         struct lu_batch *child_bh;
4181         struct lu_batch *bh;
4182
4183         ENTRY;
4184
4185         child_bh = lmv_batch_lookup_sub(lbh, tgt);
4186         if (child_bh != NULL)
4187                 RETURN(child_bh);
4188
4189         OBD_ALLOC_PTR(sbh);
4190         if (sbh == NULL)
4191                 RETURN(ERR_PTR(-ENOMEM));
4192
4193         INIT_LIST_HEAD(&sbh->sbh_sub_item);
4194         sbh->sbh_tgt = tgt;
4195
4196         bh = &lbh->lbh_super;
4197         child_bh = md_batch_create(tgt->ltd_exp, bh->lbt_flags,
4198                                    bh->lbt_max_count);
4199         if (IS_ERR(child_bh)) {
4200                 OBD_FREE_PTR(sbh);
4201                 RETURN(child_bh);
4202         }
4203
4204         child_bh->lbt_rqset = bh->lbt_rqset;
4205         sbh->sbh_sub = child_bh;
4206         list_add(&sbh->sbh_sub_item, &lbh->lbh_sub_batch_list);
4207         RETURN(child_bh);
4208 }
4209
4210 static int lmv_batch_add(struct obd_export *exp, struct lu_batch *bh,
4211                          struct md_op_item *item)
4212 {
4213         struct obd_device *obd = exp->exp_obd;
4214         struct lmv_obd *lmv = &obd->u.lmv;
4215         struct lmv_tgt_desc *tgt;
4216         struct lmv_batch *lbh;
4217         struct lu_batch *child_bh;
4218         int rc;
4219
4220         ENTRY;
4221
4222         tgt = lmv_batch_locate_tgt(lmv, item);
4223         if (IS_ERR(tgt))
4224                 RETURN(PTR_ERR(tgt));
4225
4226         lbh = container_of(bh, struct lmv_batch, lbh_super);
4227         child_bh = lmv_batch_get_sub(lbh, tgt);
4228         if (IS_ERR(child_bh))
4229                 RETURN(PTR_ERR(child_bh));
4230
4231         rc = md_batch_add(tgt->ltd_exp, child_bh, item);
4232         RETURN(rc);
4233 }
4234
4235 static const struct obd_ops lmv_obd_ops = {
4236         .o_owner                = THIS_MODULE,
4237         .o_setup                = lmv_setup,
4238         .o_cleanup              = lmv_cleanup,
4239         .o_precleanup           = lmv_precleanup,
4240         .o_process_config       = lmv_process_config,
4241         .o_connect              = lmv_connect,
4242         .o_disconnect           = lmv_disconnect,
4243         .o_statfs               = lmv_statfs,
4244         .o_get_info             = lmv_get_info,
4245         .o_set_info_async       = lmv_set_info_async,
4246         .o_notify               = lmv_notify,
4247         .o_get_uuid             = lmv_get_uuid,
4248         .o_fid_alloc            = lmv_fid_alloc,
4249         .o_iocontrol            = lmv_iocontrol,
4250         .o_quotactl             = lmv_quotactl
4251 };
4252
4253 static const struct md_ops lmv_md_ops = {
4254         .m_get_root             = lmv_get_root,
4255         .m_null_inode           = lmv_null_inode,
4256         .m_close                = lmv_close,
4257         .m_create               = lmv_create,
4258         .m_enqueue              = lmv_enqueue,
4259         .m_getattr              = lmv_getattr,
4260         .m_getxattr             = lmv_getxattr,
4261         .m_getattr_name         = lmv_getattr_name,
4262         .m_intent_lock          = lmv_intent_lock,
4263         .m_link                 = lmv_link,
4264         .m_rename               = lmv_rename,
4265         .m_setattr              = lmv_setattr,
4266         .m_setxattr             = lmv_setxattr,
4267         .m_fsync                = lmv_fsync,
4268         .m_file_resync          = lmv_file_resync,
4269         .m_read_page            = lmv_read_page,
4270         .m_unlink               = lmv_unlink,
4271         .m_init_ea_size         = lmv_init_ea_size,
4272         .m_cancel_unused        = lmv_cancel_unused,
4273         .m_set_lock_data        = lmv_set_lock_data,
4274         .m_lock_match           = lmv_lock_match,
4275         .m_get_lustre_md        = lmv_get_lustre_md,
4276         .m_put_lustre_md        = lmv_put_lustre_md,
4277         .m_merge_attr           = lmv_merge_attr,
4278         .m_set_open_replay_data = lmv_set_open_replay_data,
4279         .m_clear_open_replay_data = lmv_clear_open_replay_data,
4280         .m_intent_getattr_async = lmv_intent_getattr_async,
4281         .m_revalidate_lock      = lmv_revalidate_lock,
4282         .m_get_fid_from_lsm     = lmv_get_fid_from_lsm,
4283         .m_stripe_object_create = lmv_stripe_object_create,
4284         .m_rmfid                = lmv_rmfid,
4285         .m_batch_create         = lmv_batch_create,
4286         .m_batch_add            = lmv_batch_add,
4287         .m_batch_stop           = lmv_batch_stop,
4288         .m_batch_flush          = lmv_batch_flush,
4289 };
4290
4291 static int __init lmv_init(void)
4292 {
4293         int rc;
4294
4295         rc  = libcfs_setup();
4296         if (rc)
4297                 return rc;
4298
4299         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true,
4300                                    LUSTRE_LMV_NAME, NULL);
4301 }
4302
4303 static void __exit lmv_exit(void)
4304 {
4305         class_unregister_type(LUSTRE_LMV_NAME);
4306 }
4307
4308 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
4309 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
4310 MODULE_VERSION(LUSTRE_VERSION_STRING);
4311 MODULE_LICENSE("GPL");
4312
4313 module_init(lmv_init);
4314 module_exit(lmv_exit);