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, false) != -ENOENT) {
1200                 if (!nid_is_lo0(&lnet_id.nid)) {
1201                         lmv->lmv_qos_rr_index = ntohl(lnet_id.nid.nid_addr[0]);
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 lmv_obd *lmv, __u32 flags)
1278 {
1279         int i;
1280
1281         if (flags & OBD_STATFS_FOR_MDT0)
1282                 return 0;
1283
1284         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1285                 return lmv->lmv_statfs_start;
1286
1287         /* choose initial MDT for this client */
1288         for (i = 0;; i++) {
1289                 struct lnet_processid lnet_id;
1290                 if (LNetGetId(i, &lnet_id, false) == -ENOENT)
1291                         break;
1292
1293                 if (!nid_is_lo0(&lnet_id.nid)) {
1294                         /* We dont need a full 64-bit modulus, just enough
1295                          * to distribute the requests across MDTs evenly.
1296                          */
1297                         lmv->lmv_statfs_start = nidhash(&lnet_id.nid) %
1298                                                 lmv->lmv_mdt_count;
1299                         break;
1300                 }
1301         }
1302
1303         return lmv->lmv_statfs_start;
1304 }
1305
1306 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1307                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1308 {
1309         struct obd_device *obd = class_exp2obd(exp);
1310         struct lmv_obd *lmv = &obd->u.lmv;
1311         struct obd_statfs *temp;
1312         struct lu_tgt_desc *tgt;
1313         __u32 i;
1314         __u32 idx;
1315         int rc = 0;
1316         int err = 0;
1317
1318         ENTRY;
1319
1320         OBD_ALLOC(temp, sizeof(*temp));
1321         if (temp == NULL)
1322                 RETURN(-ENOMEM);
1323
1324         /* distribute statfs among MDTs */
1325         idx = lmv_select_statfs_mdt(lmv, flags);
1326
1327         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1328                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1329                 tgt = lmv_tgt(lmv, idx);
1330                 if (!tgt || !tgt->ltd_exp)
1331                         continue;
1332
1333                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age,
1334                                 flags | OBD_STATFS_NESTED);
1335                 if (rc) {
1336                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1337                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1338                         err = rc;
1339                         /* Try another MDT */
1340                         if (flags & OBD_STATFS_SUM)
1341                                 continue;
1342                         GOTO(out_free_temp, rc);
1343                 }
1344
1345                 if (temp->os_state & OS_STATFS_SUM ||
1346                     flags == OBD_STATFS_FOR_MDT0) {
1347                         /* reset to the last aggregated values
1348                          * and don't sum with non-aggrated data */
1349                         /* If the statfs is from mount, it needs to retrieve
1350                          * necessary information from MDT0. i.e. mount does
1351                          * not need the merged osfs from all of MDT. Also
1352                          * clients can be mounted as long as MDT0 is in
1353                          * service */
1354                         *osfs = *temp;
1355                         GOTO(out_free_temp, rc);
1356                 }
1357
1358                 if (i == 0) {
1359                         *osfs = *temp;
1360                 } else {
1361                         osfs->os_bavail += temp->os_bavail;
1362                         osfs->os_blocks += temp->os_blocks;
1363                         osfs->os_ffree += temp->os_ffree;
1364                         osfs->os_files += temp->os_files;
1365                         osfs->os_granted += temp->os_granted;
1366                 }
1367         }
1368         /* There is no stats from some MDTs, data incomplete */
1369         if (err)
1370                 rc = err;
1371 out_free_temp:
1372         OBD_FREE(temp, sizeof(*temp));
1373         RETURN(rc);
1374 }
1375
1376 static int lmv_statfs_cb(void *cookie, int rc)
1377 {
1378         struct obd_info *oinfo = cookie;
1379         struct obd_device *obd = oinfo->oi_obd;
1380         struct lmv_obd *lmv = &obd->u.lmv;
1381         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1382         struct obd_statfs *osfs = oinfo->oi_osfs;
1383
1384         /*
1385          * NB: don't deactivate TGT upon error, because we may not trigger async
1386          * statfs any longer, then there is no chance to activate TGT.
1387          */
1388         if (!rc)
1389                 lmv_statfs_update(lmv, tgt, osfs);
1390
1391         return rc;
1392 }
1393
1394 /* update tgt statfs async if it's ld_qos_maxage old */
1395 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1396 {
1397         struct obd_info oinfo = {
1398                 .oi_obd = obd,
1399                 .oi_tgt = tgt,
1400                 .oi_cb_up = lmv_statfs_cb,
1401         };
1402         int rc;
1403
1404         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1405             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1406                 return 0;
1407
1408         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1409
1410         return rc;
1411 }
1412
1413 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1414                         struct lu_fid *fid)
1415 {
1416         struct obd_device *obd = exp->exp_obd;
1417         struct lmv_obd *lmv = &obd->u.lmv;
1418         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1419         int rc;
1420
1421         ENTRY;
1422
1423         if (!tgt)
1424                 RETURN(-ENODEV);
1425
1426         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1427         RETURN(rc);
1428 }
1429
1430 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1431                         u64 obd_md_valid, const char *name, size_t buf_size,
1432                         struct ptlrpc_request **req)
1433 {
1434         struct obd_device *obd = exp->exp_obd;
1435         struct lmv_obd *lmv = &obd->u.lmv;
1436         struct lmv_tgt_desc *tgt;
1437         int rc;
1438
1439         ENTRY;
1440
1441         tgt = lmv_fid2tgt(lmv, fid);
1442         if (IS_ERR(tgt))
1443                 RETURN(PTR_ERR(tgt));
1444
1445         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1446
1447         RETURN(rc);
1448 }
1449
1450 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1451                         u64 obd_md_valid, const char *name,
1452                         const void *value, size_t value_size,
1453                         unsigned int xattr_flags, u32 suppgid,
1454                         struct ptlrpc_request **req)
1455 {
1456         struct obd_device *obd = exp->exp_obd;
1457         struct lmv_obd *lmv = &obd->u.lmv;
1458         struct lmv_tgt_desc *tgt;
1459         int rc;
1460
1461         ENTRY;
1462
1463         tgt = lmv_fid2tgt(lmv, fid);
1464         if (IS_ERR(tgt))
1465                 RETURN(PTR_ERR(tgt));
1466
1467         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1468                          value, value_size, xattr_flags, suppgid, req);
1469
1470         RETURN(rc);
1471 }
1472
1473 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1474                        struct ptlrpc_request **request)
1475 {
1476         struct obd_device *obd = exp->exp_obd;
1477         struct lmv_obd *lmv = &obd->u.lmv;
1478         struct lmv_tgt_desc *tgt;
1479         int rc;
1480
1481         ENTRY;
1482
1483         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1484         if (IS_ERR(tgt))
1485                 RETURN(PTR_ERR(tgt));
1486
1487         if (op_data->op_flags & MF_GET_MDT_IDX) {
1488                 op_data->op_mds = tgt->ltd_index;
1489                 RETURN(0);
1490         }
1491
1492         rc = md_getattr(tgt->ltd_exp, op_data, request);
1493
1494         RETURN(rc);
1495 }
1496
1497 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1498 {
1499         struct obd_device *obd = exp->exp_obd;
1500         struct lmv_obd *lmv = &obd->u.lmv;
1501         struct lu_tgt_desc *tgt;
1502
1503         ENTRY;
1504
1505         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1506
1507         /*
1508          * With DNE every object can have two locks in different namespaces:
1509          * lookup lock in space of MDT storing direntry and update/open lock in
1510          * space of MDT storing inode.
1511          */
1512         lmv_foreach_connected_tgt(lmv, tgt)
1513                 md_null_inode(tgt->ltd_exp, fid);
1514
1515         RETURN(0);
1516 }
1517
1518 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1519                      struct md_open_data *mod, struct ptlrpc_request **request)
1520 {
1521         struct obd_device *obd = exp->exp_obd;
1522         struct lmv_obd *lmv = &obd->u.lmv;
1523         struct lmv_tgt_desc *tgt;
1524         int rc;
1525
1526         ENTRY;
1527
1528         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1529         if (IS_ERR(tgt))
1530                 RETURN(PTR_ERR(tgt));
1531
1532         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1533         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1534         RETURN(rc);
1535 }
1536
1537 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv,
1538                                               struct md_op_data *op_data)
1539 {
1540         struct lu_tgt_desc *tgt, *cur = NULL;
1541         __u64 total_avail = 0;
1542         __u64 total_weight = 0;
1543         __u64 cur_weight = 0;
1544         int total_usable = 0;
1545         __u64 rand;
1546         int rc;
1547
1548         ENTRY;
1549
1550         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1551                 RETURN(ERR_PTR(-EAGAIN));
1552
1553         down_write(&lmv->lmv_qos.lq_rw_sem);
1554
1555         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1556                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1557
1558         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1559         if (rc)
1560                 GOTO(unlock, tgt = ERR_PTR(rc));
1561
1562         lmv_foreach_tgt(lmv, tgt) {
1563                 if (!tgt->ltd_exp || !tgt->ltd_active ||
1564                     (tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE)) {
1565                         tgt->ltd_qos.ltq_usable = 0;
1566                         continue;
1567                 }
1568                 /* update one hour overdue statfs */
1569                 if (ktime_get_seconds() - tgt->ltd_statfs_age >
1570                     60 * lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1571                         lmv_statfs_check_update(lmv2obd_dev(lmv), tgt);
1572                 tgt->ltd_qos.ltq_usable = 1;
1573                 lu_tgt_qos_weight_calc(tgt, true);
1574                 if (tgt->ltd_index == op_data->op_mds)
1575                         cur = tgt;
1576                 total_avail += tgt->ltd_qos.ltq_avail;
1577                 total_weight += tgt->ltd_qos.ltq_weight;
1578                 total_usable++;
1579         }
1580
1581         /* If current MDT has above-average space and dir is not already using
1582          * round-robin to spread across more MDTs, stay on the parent MDT
1583          * to avoid creating needless remote MDT directories.  Remote dirs
1584          * close to the root balance space more effectively than bottom dirs,
1585          * so prefer to create remote dirs at top level of directory tree.
1586          * "16 / (dir_depth + 10)" is the factor to make it less likely
1587          * for top-level directories to stay local unless they have more than
1588          * average free space, while deep dirs prefer local until more full.
1589          *    depth=0 -> 160%, depth=3 -> 123%, depth=6 -> 100%,
1590          *    depth=9 -> 84%, depth=12 -> 73%, depth=15 -> 64%
1591          */
1592         if (!lmv_op_default_rr_mkdir(op_data)) {
1593                 rand = total_avail * 16 /
1594                         (total_usable * (op_data->op_dir_depth + 10));
1595                 if (cur && cur->ltd_qos.ltq_avail >= rand) {
1596                         tgt = cur;
1597                         GOTO(unlock, tgt);
1598                 }
1599         }
1600
1601         rand = lu_prandom_u64_max(total_weight);
1602
1603         lmv_foreach_connected_tgt(lmv, tgt) {
1604                 if (!tgt->ltd_qos.ltq_usable)
1605                         continue;
1606
1607                 cur_weight += tgt->ltd_qos.ltq_weight;
1608                 if (cur_weight < rand)
1609                         continue;
1610
1611                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1612                 GOTO(unlock, tgt);
1613         }
1614
1615         /* no proper target found */
1616         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1617 unlock:
1618         up_write(&lmv->lmv_qos.lq_rw_sem);
1619
1620         return tgt;
1621 }
1622
1623 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv)
1624 {
1625         struct lu_tgt_desc *tgt;
1626         int i;
1627         int index;
1628
1629         ENTRY;
1630
1631         spin_lock(&lmv->lmv_lock);
1632         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1633                 index = (i + lmv->lmv_qos_rr_index) %
1634                         lmv->lmv_mdt_descs.ltd_tgts_size;
1635                 tgt = lmv_tgt(lmv, index);
1636                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active ||
1637                     (tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE))
1638                         continue;
1639
1640                 lmv->lmv_qos_rr_index = (tgt->ltd_index + 1) %
1641                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1642                 spin_unlock(&lmv->lmv_lock);
1643
1644                 RETURN(tgt);
1645         }
1646         spin_unlock(&lmv->lmv_lock);
1647
1648         RETURN(ERR_PTR(-ENODEV));
1649 }
1650
1651 /* locate MDT which is less full (avoid the most full MDT) */
1652 static struct lu_tgt_desc *lmv_locate_tgt_lf(struct lmv_obd *lmv)
1653 {
1654         struct lu_tgt_desc *min = NULL;
1655         struct lu_tgt_desc *tgt;
1656         __u64 avail = 0;
1657         __u64 rand;
1658
1659         ENTRY;
1660
1661         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1662                 RETURN(ERR_PTR(-EAGAIN));
1663
1664         down_write(&lmv->lmv_qos.lq_rw_sem);
1665
1666         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1667                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1668
1669         lmv_foreach_tgt(lmv, tgt) {
1670                 if (!tgt->ltd_exp || !tgt->ltd_active ||
1671                     (tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE)) {
1672                         tgt->ltd_qos.ltq_usable = 0;
1673                         continue;
1674                 }
1675
1676                 tgt->ltd_qos.ltq_usable = 1;
1677                 lu_tgt_qos_weight_calc(tgt, true);
1678                 avail += tgt->ltd_qos.ltq_avail;
1679                 if (!min || min->ltd_qos.ltq_avail > tgt->ltd_qos.ltq_avail)
1680                         min = tgt;
1681         }
1682
1683         /* avoid the most full MDT */
1684         if (min)
1685                 avail -= min->ltd_qos.ltq_avail;
1686
1687         rand = lu_prandom_u64_max(avail);
1688         avail = 0;
1689         lmv_foreach_connected_tgt(lmv, tgt) {
1690                 if (!tgt->ltd_qos.ltq_usable)
1691                         continue;
1692
1693                 if (tgt == min)
1694                         continue;
1695
1696                 avail += tgt->ltd_qos.ltq_avail;
1697                 if (avail < rand)
1698                         continue;
1699
1700                 GOTO(unlock, tgt);
1701         }
1702
1703         /* no proper target found */
1704         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1705 unlock:
1706         up_write(&lmv->lmv_qos.lq_rw_sem);
1707
1708         RETURN(tgt);
1709 }
1710
1711 /* locate MDT by file name, for striped directory, the file name hash decides
1712  * which stripe its dirent is stored.
1713  */
1714 static struct lmv_tgt_desc *
1715 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_object *lso,
1716                        const char *name, int namelen, struct lu_fid *fid,
1717                        __u32 *mds, bool new_layout)
1718 {
1719         struct lmv_tgt_desc *tgt;
1720         const struct lmv_oinfo *oinfo;
1721
1722         if (!lmv_dir_striped(lso) || !namelen) {
1723                 tgt = lmv_fid2tgt(lmv, fid);
1724                 if (IS_ERR(tgt))
1725                         return tgt;
1726
1727                 *mds = tgt->ltd_index;
1728                 return tgt;
1729         }
1730
1731         if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1732                 if (cfs_fail_val >= lso->lso_lsm.lsm_md_stripe_count)
1733                         return ERR_PTR(-EBADF);
1734                 oinfo = &lso->lso_lsm.lsm_md_oinfo[cfs_fail_val];
1735         } else {
1736                 oinfo = lsm_name_to_stripe_info(lso, name, namelen, new_layout);
1737                 if (IS_ERR(oinfo))
1738                         return ERR_CAST(oinfo);
1739         }
1740
1741         /* check stripe FID is sane */
1742         if (!fid_is_sane(&oinfo->lmo_fid))
1743                 return ERR_PTR(-ENODEV);
1744
1745         *fid = oinfo->lmo_fid;
1746         *mds = oinfo->lmo_mds;
1747         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1748
1749         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1750
1751         return tgt ? tgt : ERR_PTR(-ENODEV);
1752 }
1753
1754 /**
1755  * Locate MDT of op_data->op_fid1
1756  *
1757  * For striped directory, it will locate the stripe by name hash, if hash_type
1758  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1759  * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1760  * indicates whether old or new layout is used to locate.
1761  *
1762  * For plain directory, it just locate the MDT of op_data->op_fid1.
1763  *
1764  * \param[in] lmv               LMV device
1765  * \param[in/out] op_data       client MD stack parameters, name, namelen etc,
1766  *                              op_mds and op_fid1 will be updated if op_lso1
1767  *                              indicates fid1 represents a striped directory.
1768  *
1769  * retval               pointer to the lmv_tgt_desc if succeed.
1770  *                      ERR_PTR(errno) if failed.
1771  */
1772 struct lmv_tgt_desc *
1773 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1774 {
1775         struct lmv_stripe_md *lsm;
1776         struct lmv_oinfo *oinfo;
1777         struct lmv_tgt_desc *tgt;
1778
1779         if (lmv_dir_foreign(op_data->op_lso1))
1780                 return ERR_PTR(-ENODATA);
1781
1782         /* During creating VOLATILE file, it should honor the mdt
1783          * index if the file under striped dir is being restored, see
1784          * ct_restore(). */
1785         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1786             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1787                 tgt = lmv_tgt(lmv, op_data->op_mds);
1788                 if (!tgt)
1789                         return ERR_PTR(-ENODEV);
1790
1791                 if (lmv_dir_striped(op_data->op_lso1)) {
1792                         int i;
1793
1794                         /* refill the right parent fid */
1795                         lsm = &op_data->op_lso1->lso_lsm;
1796                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1797                                 oinfo = &lsm->lsm_md_oinfo[i];
1798                                 if (oinfo->lmo_mds == op_data->op_mds) {
1799                                         op_data->op_fid1 = oinfo->lmo_fid;
1800                                         break;
1801                                 }
1802                         }
1803
1804                         if (i == lsm->lsm_md_stripe_count)
1805                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1806                 }
1807         } else if (lmv_dir_bad_hash(op_data->op_lso1)) {
1808                 lsm = &op_data->op_lso1->lso_lsm;
1809
1810                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1811                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1812
1813                 op_data->op_fid1 = oinfo->lmo_fid;
1814                 op_data->op_mds = oinfo->lmo_mds;
1815                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1816                 if (!tgt)
1817                         return ERR_PTR(-ENODEV);
1818         } else {
1819                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_lso1,
1820                                 op_data->op_name, op_data->op_namelen,
1821                                 &op_data->op_fid1, &op_data->op_mds,
1822                                 op_data->op_new_layout);
1823         }
1824
1825         return tgt;
1826 }
1827
1828 /* Locate MDT of op_data->op_fid2 for link/rename */
1829 static struct lmv_tgt_desc *
1830 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1831 {
1832         struct lmv_tgt_desc *tgt;
1833         int rc;
1834
1835         LASSERT(op_data->op_name);
1836         if (lmv_dir_layout_changing(op_data->op_lso2)) {
1837                 struct lu_fid fid1 = op_data->op_fid1;
1838                 struct lmv_stripe_object *lso1 = op_data->op_lso1;
1839                 struct ptlrpc_request *request = NULL;
1840
1841                 /*
1842                  * avoid creating new file under old layout of migrating
1843                  * directory, check it here.
1844                  */
1845                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_lso2,
1846                                 op_data->op_name, op_data->op_namelen,
1847                                 &op_data->op_fid2, &op_data->op_mds, false);
1848                 if (IS_ERR(tgt))
1849                         RETURN(tgt);
1850
1851                 op_data->op_fid1 = op_data->op_fid2;
1852                 op_data->op_lso1 = op_data->op_lso2;
1853                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1854                 op_data->op_fid1 = fid1;
1855                 op_data->op_lso1 = lso1;
1856                 if (!rc) {
1857                         ptlrpc_req_finished(request);
1858                         RETURN(ERR_PTR(-EEXIST));
1859                 }
1860
1861                 if (rc != -ENOENT)
1862                         RETURN(ERR_PTR(rc));
1863         }
1864
1865         return lmv_locate_tgt_by_name(lmv, op_data->op_lso2,
1866                                       op_data->op_name, op_data->op_namelen,
1867                                       &op_data->op_fid2, &op_data->op_mds,
1868                                       true);
1869 }
1870
1871 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1872 {
1873         struct lu_tgt_desc *tgt;
1874         struct ptlrpc_request *request;
1875         int rc;
1876
1877         LASSERT(lmv_dir_layout_changing(op_data->op_lso1));
1878         LASSERT(!op_data->op_new_layout);
1879
1880         tgt = lmv_locate_tgt(lmv, op_data);
1881         if (IS_ERR(tgt))
1882                 return PTR_ERR(tgt);
1883
1884         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1885         if (!rc) {
1886                 ptlrpc_req_finished(request);
1887                 return -EEXIST;
1888         }
1889
1890         return rc;
1891 }
1892
1893 /* mkdir by QoS upon 'lfs mkdir -i -1'.
1894  *
1895  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1896  * directories under striped directory.
1897  */
1898 static inline bool lmv_op_user_qos_mkdir(const struct md_op_data *op_data)
1899 {
1900         const struct lmv_user_md *lum = op_data->op_data;
1901
1902         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1903                 return false;
1904
1905         if (lmv_dir_striped(op_data->op_lso1))
1906                 return false;
1907
1908         return (op_data->op_cli_flags & CLI_SET_MEA) && lum &&
1909                le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC &&
1910                le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
1911 }
1912
1913 /* mkdir by QoS if either ROOT or parent default LMV is space balanced. */
1914 static inline bool lmv_op_default_qos_mkdir(const struct md_op_data *op_data)
1915 {
1916         const struct lmv_stripe_object *lso = op_data->op_default_lso1;
1917
1918         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1919                 return false;
1920
1921         if (lmv_dir_striped(op_data->op_lso1))
1922                 return false;
1923
1924         return (op_data->op_flags & MF_QOS_MKDIR) ||
1925                (lso && lso->lso_lsm.lsm_md_master_mdt_index ==
1926                 LMV_OFFSET_DEFAULT);
1927 }
1928
1929 /* if parent default LMV is space balanced, and
1930  * 1. max_inherit_rr is set
1931  * 2. or parent is ROOT
1932  * mkdir roundrobin. Or if parent doesn't have default LMV, while ROOT default
1933  * LMV requests roundrobin mkdir, do the same.
1934  * NB, this needs to check server is balanced, which is done by caller.
1935  */
1936 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data)
1937 {
1938         const struct lmv_stripe_object *lso = op_data->op_default_lso1;
1939
1940         return (op_data->op_flags & MF_RR_MKDIR) ||
1941                (lso && lso->lso_lsm.lsm_md_max_inherit_rr !=
1942                 LMV_INHERIT_RR_NONE) || fid_is_root(&op_data->op_fid1);
1943 }
1944
1945 /* 'lfs mkdir -i <specific_MDT>' */
1946 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1947 {
1948         const struct lmv_user_md *lum = op_data->op_data;
1949
1950         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1951                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1952                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1953                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1954                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1955 }
1956
1957 /* parent default LMV master_mdt_index is not -1. */
1958 static inline bool
1959 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1960 {
1961         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1962                op_data->op_default_lso1 &&
1963                op_data->op_default_lso1->lso_lsm.lsm_md_master_mdt_index !=
1964                         LMV_OFFSET_DEFAULT;
1965 }
1966
1967 /* locate MDT by space usage */
1968 static struct lu_tgt_desc *lmv_locate_tgt_by_space(struct lmv_obd *lmv,
1969                                                    struct md_op_data *op_data,
1970                                                    struct lmv_tgt_desc *tgt)
1971 {
1972         struct lmv_tgt_desc *tmp = tgt;
1973
1974         tgt = lmv_locate_tgt_qos(lmv, op_data);
1975         if (tgt == ERR_PTR(-EAGAIN)) {
1976                 if (ltd_qos_is_balanced(&lmv->lmv_mdt_descs) &&
1977                     !lmv_op_default_rr_mkdir(op_data) &&
1978                     !lmv_op_user_qos_mkdir(op_data) &&
1979                     !(tmp->ltd_statfs.os_state & OS_STATFS_NOCREATE))
1980                         /* if not necessary, don't create remote directory. */
1981                         tgt = tmp;
1982                 else
1983                         tgt = lmv_locate_tgt_rr(lmv);
1984                 if (!IS_ERR(tgt))
1985                         lmv_statfs_check_update(lmv2obd_dev(lmv), tgt);
1986         }
1987
1988         if (!IS_ERR(tgt))
1989                 op_data->op_mds = tgt->ltd_index;
1990
1991         /* If space balance was called because the original target was marked
1992          * NOCREATE, periodically check whether the state has changed.
1993          */
1994         if (tmp != tgt && tmp->ltd_statfs.os_state & OS_STATFS_NOCREATE)
1995                 lmv_statfs_check_update(lmv2obd_dev(lmv), tmp);
1996
1997         return tgt;
1998 }
1999
2000 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
2001                 const void *data, size_t datalen, umode_t mode, uid_t uid,
2002                 gid_t gid, kernel_cap_t cap_effective, __u64 rdev,
2003                 struct ptlrpc_request **request)
2004 {
2005         struct obd_device *obd = exp->exp_obd;
2006         struct lmv_obd *lmv = &obd->u.lmv;
2007         struct lmv_tgt_desc *tgt;
2008         struct mdt_body *repbody;
2009         int rc;
2010
2011         ENTRY;
2012
2013         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
2014                 RETURN(-EIO);
2015
2016         if (lmv_dir_bad_hash(op_data->op_lso1))
2017                 RETURN(-EBADF);
2018
2019         if (lmv_dir_layout_changing(op_data->op_lso1)) {
2020                 /*
2021                  * if parent is migrating, create() needs to lookup existing
2022                  * name in both old and new layout, check old layout on client.
2023                  */
2024                 rc = lmv_old_layout_lookup(lmv, op_data);
2025                 if (rc != -ENOENT)
2026                         RETURN(rc);
2027
2028                 op_data->op_new_layout = true;
2029         }
2030
2031         tgt = lmv_locate_tgt(lmv, op_data);
2032         if (IS_ERR(tgt))
2033                 RETURN(PTR_ERR(tgt));
2034
2035         /* the order to apply policy in mkdir:
2036          * 1. is "lfs mkdir -i N"? mkdir on MDT N.
2037          * 2. is "lfs mkdir -i -1"? mkdir by space usage.
2038          * 3. is starting MDT specified in default LMV? mkdir on MDT N.
2039          * 4. is default LMV space balanced? mkdir by space usage.
2040          *
2041          * If the existing parent or specific MDT selected is deactivated
2042          * with OS_STATFS_NOCREATE then select a different MDT by QOS.
2043          */
2044         if (lmv_op_user_specific_mkdir(op_data)) {
2045                 struct lmv_user_md *lum = op_data->op_data;
2046
2047                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2048                 tgt = lmv_tgt(lmv, op_data->op_mds);
2049                 if (!tgt)
2050                         RETURN(-ENODEV);
2051                 if (unlikely(tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE))
2052                         GOTO(new_tgt, -EAGAIN);
2053         } else if (lmv_op_user_qos_mkdir(op_data)) {
2054                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
2055                 if (IS_ERR(tgt))
2056                         RETURN(PTR_ERR(tgt));
2057         } else if (lmv_op_default_specific_mkdir(op_data)) {
2058                 struct lmv_stripe_md *lsm = &op_data->op_default_lso1->lso_lsm;
2059
2060                 op_data->op_mds = lsm->lsm_md_master_mdt_index;
2061                 tgt = lmv_tgt(lmv, op_data->op_mds);
2062                 if (!tgt)
2063                         RETURN(-ENODEV);
2064                 if (unlikely(tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE))
2065                         GOTO(new_tgt, -EAGAIN);
2066         } else if (lmv_op_default_qos_mkdir(op_data) ||
2067                    tgt->ltd_statfs.os_state & OS_STATFS_NOCREATE) {
2068 new_tgt:
2069                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
2070                 if (IS_ERR(tgt))
2071                         RETURN(PTR_ERR(tgt));
2072         }
2073
2074 retry:
2075         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
2076         if (rc)
2077                 RETURN(rc);
2078
2079         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
2080                 (int)op_data->op_namelen, op_data->op_name,
2081                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
2082                 op_data->op_mds);
2083
2084         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2085         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
2086                        cap_effective, rdev, request);
2087         if (rc == 0) {
2088                 if (*request == NULL)
2089                         RETURN(rc);
2090                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
2091         }
2092
2093         /* dir restripe needs to send to MDT where dir is located */
2094         if (rc != -EREMOTE ||
2095             !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
2096                 RETURN(rc);
2097
2098         repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2099         if (repbody == NULL)
2100                 RETURN(-EPROTO);
2101
2102         /* Not cross-ref case, just get out of here. */
2103         if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
2104                 RETURN(rc);
2105
2106         op_data->op_fid2 = repbody->mbo_fid1;
2107         ptlrpc_req_finished(*request);
2108         *request = NULL;
2109
2110         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2111         if (IS_ERR(tgt))
2112                 RETURN(PTR_ERR(tgt));
2113
2114         op_data->op_mds = tgt->ltd_index;
2115         goto retry;
2116 }
2117
2118 static int
2119 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
2120             const union ldlm_policy_data *policy, struct md_op_data *op_data,
2121             struct lustre_handle *lockh, __u64 extra_lock_flags)
2122 {
2123         struct obd_device *obd = exp->exp_obd;
2124         struct lmv_obd *lmv = &obd->u.lmv;
2125         struct lmv_tgt_desc *tgt;
2126         int rc;
2127
2128         ENTRY;
2129
2130         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
2131
2132         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2133         if (IS_ERR(tgt))
2134                 RETURN(PTR_ERR(tgt));
2135
2136         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
2137                PFID(&op_data->op_fid1), tgt->ltd_index);
2138
2139         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
2140                         extra_lock_flags);
2141
2142         RETURN(rc);
2143 }
2144
2145 int
2146 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
2147                  struct ptlrpc_request **preq)
2148 {
2149         struct obd_device *obd = exp->exp_obd;
2150         struct lmv_obd *lmv = &obd->u.lmv;
2151         struct lmv_tgt_desc *tgt;
2152         struct mdt_body *body;
2153         int rc;
2154
2155         ENTRY;
2156
2157 retry:
2158         if (op_data->op_namelen == 2 &&
2159             op_data->op_name[0] == '.' && op_data->op_name[1] == '.')
2160                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2161         else
2162                 tgt = lmv_locate_tgt(lmv, op_data);
2163         if (IS_ERR(tgt))
2164                 RETURN(PTR_ERR(tgt));
2165
2166         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
2167                 (int)op_data->op_namelen, op_data->op_name,
2168                 PFID(&op_data->op_fid1), tgt->ltd_index);
2169
2170         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
2171         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2172                 ptlrpc_req_finished(*preq);
2173                 *preq = NULL;
2174                 goto retry;
2175         }
2176
2177         if (rc)
2178                 RETURN(rc);
2179
2180         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
2181         LASSERT(body != NULL);
2182
2183         if (body->mbo_valid & OBD_MD_MDS) {
2184                 op_data->op_fid1 = body->mbo_fid1;
2185                 op_data->op_valid |= OBD_MD_FLCROSSREF;
2186                 op_data->op_namelen = 0;
2187                 op_data->op_name = NULL;
2188
2189                 ptlrpc_req_finished(*preq);
2190                 *preq = NULL;
2191
2192                 goto retry;
2193         }
2194
2195         RETURN(rc);
2196 }
2197
2198 #define md_op_data_fid(op_data, fl)                     \
2199         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2200          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2201          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2202          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2203          NULL)
2204
2205 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2206                             struct md_op_data *op_data, __u32 op_tgt,
2207                             enum ldlm_mode mode, int bits, int flag)
2208 {
2209         struct lu_fid *fid = md_op_data_fid(op_data, flag);
2210         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
2211         union ldlm_policy_data policy = { { 0 } };
2212         int rc = 0;
2213         ENTRY;
2214
2215         if (!fid_is_sane(fid))
2216                 RETURN(0);
2217
2218         if (tgt == NULL) {
2219                 tgt = lmv_fid2tgt(lmv, fid);
2220                 if (IS_ERR(tgt))
2221                         RETURN(PTR_ERR(tgt));
2222         }
2223
2224         if (tgt->ltd_index != op_tgt) {
2225                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2226                 policy.l_inodebits.bits = bits;
2227                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2228                                       mode, LCF_ASYNC, NULL);
2229         } else {
2230                 CDEBUG(D_INODE,
2231                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2232                        op_tgt, PFID(fid));
2233                 op_data->op_flags |= flag;
2234                 rc = 0;
2235         }
2236
2237         RETURN(rc);
2238 }
2239
2240 /*
2241  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2242  * op_data->op_fid2
2243  */
2244 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2245                     struct ptlrpc_request **request)
2246 {
2247         struct obd_device       *obd = exp->exp_obd;
2248         struct lmv_obd          *lmv = &obd->u.lmv;
2249         struct lmv_tgt_desc     *tgt;
2250         int                      rc;
2251         ENTRY;
2252
2253         LASSERT(op_data->op_namelen != 0);
2254
2255         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2256                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2257                op_data->op_name, PFID(&op_data->op_fid1));
2258
2259         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2260         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2261         op_data->op_cap = current_cap();
2262
2263         tgt = lmv_locate_tgt2(lmv, op_data);
2264         if (IS_ERR(tgt))
2265                 RETURN(PTR_ERR(tgt));
2266
2267         /*
2268          * Cancel UPDATE lock on child (fid1).
2269          */
2270         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2271         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2272                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2273         if (rc != 0)
2274                 RETURN(rc);
2275
2276         rc = md_link(tgt->ltd_exp, op_data, request);
2277
2278         RETURN(rc);
2279 }
2280
2281 /* migrate the top directory */
2282 static inline bool lmv_op_topdir_migrate(const struct md_op_data *op_data)
2283 {
2284         if (!S_ISDIR(op_data->op_mode))
2285                 return false;
2286
2287         if (lmv_dir_layout_changing(op_data->op_lso1))
2288                 return false;
2289
2290         return true;
2291 }
2292
2293 /* migrate top dir to specific MDTs */
2294 static inline bool lmv_topdir_specific_migrate(const struct md_op_data *op_data)
2295 {
2296         const struct lmv_user_md *lum = op_data->op_data;
2297
2298         if (!lmv_op_topdir_migrate(op_data))
2299                 return false;
2300
2301         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2302 }
2303
2304 /* migrate top dir in QoS mode if user issued "lfs migrate -m -1..." */
2305 static inline bool lmv_topdir_qos_migrate(const struct md_op_data *op_data)
2306 {
2307         const struct lmv_user_md *lum = op_data->op_data;
2308
2309         if (!lmv_op_topdir_migrate(op_data))
2310                 return false;
2311
2312         return le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
2313 }
2314
2315 static inline bool lmv_subdir_specific_migrate(const struct md_op_data *op_data)
2316 {
2317         const struct lmv_user_md *lum = op_data->op_data;
2318
2319         if (!S_ISDIR(op_data->op_mode))
2320                 return false;
2321
2322         if (!lmv_dir_layout_changing(op_data->op_lso1))
2323                 return false;
2324
2325         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2326 }
2327
2328 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2329                         const char *name, size_t namelen,
2330                         struct ptlrpc_request **request)
2331 {
2332         struct obd_device *obd = exp->exp_obd;
2333         struct lmv_obd *lmv = &obd->u.lmv;
2334         struct lmv_stripe_object *lso = op_data->op_lso1;
2335         struct lmv_tgt_desc *parent_tgt;
2336         struct lmv_tgt_desc *sp_tgt;
2337         struct lmv_tgt_desc *tp_tgt = NULL;
2338         struct lmv_tgt_desc *child_tgt;
2339         struct lmv_tgt_desc *tgt;
2340         struct lu_fid target_fid = { 0 };
2341         int rc;
2342
2343         ENTRY;
2344
2345         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2346
2347         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2348                PFID(&op_data->op_fid1), (int)namelen, name);
2349
2350         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2351         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2352         op_data->op_cap = current_cap();
2353
2354         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2355         if (IS_ERR(parent_tgt))
2356                 RETURN(PTR_ERR(parent_tgt));
2357
2358         if (lmv_dir_striped(lso)) {
2359                 const struct lmv_oinfo *oinfo;
2360
2361                 oinfo = lsm_name_to_stripe_info(lso, name, namelen, false);
2362                 if (IS_ERR(oinfo))
2363                         RETURN(PTR_ERR(oinfo));
2364
2365                 /* save source stripe FID in fid4 temporarily for ELC */
2366                 op_data->op_fid4 = oinfo->lmo_fid;
2367                 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2368                 if (!sp_tgt)
2369                         RETURN(-ENODEV);
2370
2371                 /*
2372                  * if parent is being migrated too, fill op_fid2 with target
2373                  * stripe fid, otherwise the target stripe is not created yet.
2374                  */
2375                 if (lmv_dir_layout_changing(lso)) {
2376                         oinfo = lsm_name_to_stripe_info(lso, name, namelen,
2377                                                         true);
2378                         if (IS_ERR(oinfo))
2379                                 RETURN(PTR_ERR(oinfo));
2380
2381                         op_data->op_fid2 = oinfo->lmo_fid;
2382                         tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2383                         if (!tp_tgt)
2384                                 RETURN(-ENODEV);
2385
2386                         /* parent unchanged and update namespace only */
2387                         if (lu_fid_eq(&op_data->op_fid4, &op_data->op_fid2) &&
2388                             op_data->op_bias & MDS_MIGRATE_NSONLY)
2389                                 RETURN(-EALREADY);
2390                 }
2391         } else {
2392                 sp_tgt = parent_tgt;
2393         }
2394
2395         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2396         if (IS_ERR(child_tgt))
2397                 RETURN(PTR_ERR(child_tgt));
2398
2399         if (lmv_topdir_specific_migrate(op_data)) {
2400                 struct lmv_user_md *lum = op_data->op_data;
2401
2402                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2403         } else if (lmv_topdir_qos_migrate(op_data)) {
2404                 tgt = lmv_locate_tgt_lf(lmv);
2405                 if (tgt == ERR_PTR(-EAGAIN))
2406                         tgt = lmv_locate_tgt_rr(lmv);
2407                 if (IS_ERR(tgt))
2408                         RETURN(PTR_ERR(tgt));
2409
2410                 op_data->op_mds = tgt->ltd_index;
2411         } else if (lmv_subdir_specific_migrate(op_data)) {
2412                 struct lmv_user_md *lum = op_data->op_data;
2413                 __u32 i;
2414
2415                 LASSERT(tp_tgt);
2416                 if (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) {
2417                         /* adjust MDTs in lum, since subdir is located on where
2418                          * its parent stripe is, not the first specified MDT.
2419                          */
2420                         for (i = 0; i < le32_to_cpu(lum->lum_stripe_count);
2421                              i++) {
2422                                 if (le32_to_cpu(lum->lum_objects[i].lum_mds) ==
2423                                     tp_tgt->ltd_index)
2424                                         break;
2425                         }
2426
2427                         if (i == le32_to_cpu(lum->lum_stripe_count))
2428                                 RETURN(-ENODEV);
2429
2430                         lum->lum_objects[i].lum_mds =
2431                                 lum->lum_objects[0].lum_mds;
2432                         lum->lum_objects[0].lum_mds =
2433                                 cpu_to_le32(tp_tgt->ltd_index);
2434                 }
2435                 /* NB, the above adjusts subdir migration for command like
2436                  * "lfs migrate -m 0,1,2 ...", but for migration like
2437                  * "lfs migrate -m 0 -c 2 ...", the top dir is migrated to MDT0
2438                  * and MDT1, however its subdir may be migrated to MDT1 and MDT2
2439                  */
2440
2441                 lum->lum_stripe_offset = cpu_to_le32(tp_tgt->ltd_index);
2442                 op_data->op_mds = tp_tgt->ltd_index;
2443         } else if (tp_tgt) {
2444                 op_data->op_mds = tp_tgt->ltd_index;
2445         } else {
2446                 op_data->op_mds = sp_tgt->ltd_index;
2447         }
2448
2449         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2450         if (rc)
2451                 RETURN(rc);
2452
2453         /*
2454          * for directory, send migrate request to the MDT where the object will
2455          * be migrated to, because we can't create a striped directory remotely.
2456          *
2457          * otherwise, send to the MDT where source is located because regular
2458          * file may open lease.
2459          *
2460          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2461          * backward compatibility.
2462          */
2463         if (S_ISDIR(op_data->op_mode) &&
2464             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2465                 tgt = lmv_fid2tgt(lmv, &target_fid);
2466                 if (IS_ERR(tgt))
2467                         RETURN(PTR_ERR(tgt));
2468         } else {
2469                 tgt = child_tgt;
2470         }
2471
2472         /* cancel UPDATE lock of parent master object */
2473         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2474                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2475         if (rc)
2476                 RETURN(rc);
2477
2478         /* cancel UPDATE lock of source parent */
2479         if (sp_tgt != parent_tgt) {
2480                 /*
2481                  * migrate RPC packs master object FID, because we can only pack
2482                  * two FIDs in reint RPC, but MDS needs to know both source
2483                  * parent and target parent, and it will obtain them from master
2484                  * FID and LMV, the other FID in RPC is kept for target.
2485                  *
2486                  * since this FID is not passed to MDC, cancel it anyway.
2487                  */
2488                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2489                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2490                 if (rc)
2491                         RETURN(rc);
2492
2493                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2494         }
2495         op_data->op_fid4 = target_fid;
2496
2497         /* cancel UPDATE locks of target parent */
2498         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2499                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2500         if (rc)
2501                 RETURN(rc);
2502
2503         /* cancel LOOKUP lock of source if source is remote object */
2504         if (child_tgt != sp_tgt) {
2505                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2506                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2507                                       MF_MDC_CANCEL_FID3);
2508                 if (rc)
2509                         RETURN(rc);
2510         }
2511
2512         /* cancel ELC locks of source */
2513         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2514                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2515         if (rc)
2516                 RETURN(rc);
2517
2518         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2519
2520         RETURN(rc);
2521 }
2522
2523 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2524                       const char *old, size_t oldlen,
2525                       const char *new, size_t newlen,
2526                       struct ptlrpc_request **request)
2527 {
2528         struct obd_device *obd = exp->exp_obd;
2529         struct lmv_obd *lmv = &obd->u.lmv;
2530         struct lmv_tgt_desc *sp_tgt;
2531         struct lmv_tgt_desc *tp_tgt = NULL;
2532         struct lmv_tgt_desc *src_tgt = NULL;
2533         struct lmv_tgt_desc *tgt;
2534         struct mdt_body *body;
2535         int rc;
2536
2537         ENTRY;
2538
2539         LASSERT(oldlen != 0);
2540
2541         if (op_data->op_cli_flags & CLI_MIGRATE) {
2542                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2543                 RETURN(rc);
2544         }
2545
2546         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2547         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2548         op_data->op_cap = current_cap();
2549
2550         op_data->op_name = new;
2551         op_data->op_namelen = newlen;
2552
2553         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2554         if (IS_ERR(tp_tgt))
2555                 RETURN(PTR_ERR(tp_tgt));
2556
2557         /* Since the target child might be destroyed, and it might become
2558          * orphan, and we can only check orphan on the local MDT right now, so
2559          * we send rename request to the MDT where target child is located. If
2560          * target child does not exist, then it will send the request to the
2561          * target parent */
2562         if (fid_is_sane(&op_data->op_fid4)) {
2563                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2564                 if (IS_ERR(tgt))
2565                         RETURN(PTR_ERR(tgt));
2566         } else {
2567                 tgt = tp_tgt;
2568         }
2569
2570         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2571
2572         /* cancel UPDATE locks of target parent */
2573         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2574                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2575         if (rc != 0)
2576                 RETURN(rc);
2577
2578         if (fid_is_sane(&op_data->op_fid4)) {
2579                 /* cancel LOOKUP lock of target on target parent */
2580                 if (tgt != tp_tgt) {
2581                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2582                                               tgt->ltd_index, LCK_EX,
2583                                               MDS_INODELOCK_LOOKUP,
2584                                               MF_MDC_CANCEL_FID4);
2585                         if (rc != 0)
2586                                 RETURN(rc);
2587                 }
2588         }
2589
2590         if (fid_is_sane(&op_data->op_fid3)) {
2591                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2592                 if (IS_ERR(src_tgt))
2593                         RETURN(PTR_ERR(src_tgt));
2594
2595                 /* cancel ELC locks of source */
2596                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2597                                       LCK_EX, MDS_INODELOCK_ELC,
2598                                       MF_MDC_CANCEL_FID3);
2599                 if (rc != 0)
2600                         RETURN(rc);
2601         }
2602
2603         op_data->op_name = old;
2604         op_data->op_namelen = oldlen;
2605 retry:
2606         sp_tgt = lmv_locate_tgt(lmv, op_data);
2607         if (IS_ERR(sp_tgt))
2608                 RETURN(PTR_ERR(sp_tgt));
2609
2610         /* cancel UPDATE locks of source parent */
2611         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2612                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2613         if (rc != 0)
2614                 RETURN(rc);
2615
2616         if (fid_is_sane(&op_data->op_fid3)) {
2617                 /* cancel LOOKUP lock of source on source parent */
2618                 if (src_tgt != sp_tgt) {
2619                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2620                                               tgt->ltd_index, LCK_EX,
2621                                               MDS_INODELOCK_LOOKUP,
2622                                               MF_MDC_CANCEL_FID3);
2623                         if (rc != 0)
2624                                 RETURN(rc);
2625                 }
2626         }
2627
2628 rename:
2629         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2630                 PFID(&op_data->op_fid1), (int)oldlen, old,
2631                 PFID(&op_data->op_fid2), (int)newlen, new);
2632
2633         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2634                         request);
2635         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2636                 ptlrpc_req_finished(*request);
2637                 *request = NULL;
2638                 goto retry;
2639         }
2640
2641         if (rc && rc != -EXDEV)
2642                 RETURN(rc);
2643
2644         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2645         if (body == NULL)
2646                 RETURN(-EPROTO);
2647
2648         /* Not cross-ref case, just get out of here. */
2649         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2650                 RETURN(rc);
2651
2652         op_data->op_fid4 = body->mbo_fid1;
2653
2654         ptlrpc_req_finished(*request);
2655         *request = NULL;
2656
2657         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2658         if (IS_ERR(tgt))
2659                 RETURN(PTR_ERR(tgt));
2660
2661         if (fid_is_sane(&op_data->op_fid4)) {
2662                 /* cancel LOOKUP lock of target on target parent */
2663                 if (tgt != tp_tgt) {
2664                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2665                                               tgt->ltd_index, LCK_EX,
2666                                               MDS_INODELOCK_LOOKUP,
2667                                               MF_MDC_CANCEL_FID4);
2668                         if (rc != 0)
2669                                 RETURN(rc);
2670                 }
2671         }
2672
2673         goto rename;
2674 }
2675
2676 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2677                        void *ea, size_t ealen, struct ptlrpc_request **request)
2678 {
2679         struct obd_device *obd = exp->exp_obd;
2680         struct lmv_obd *lmv = &obd->u.lmv;
2681         struct lmv_tgt_desc *tgt;
2682         int rc = 0;
2683
2684         ENTRY;
2685
2686         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2687                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2688                op_data->op_xvalid);
2689
2690         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2691         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2692         if (IS_ERR(tgt))
2693                 RETURN(PTR_ERR(tgt));
2694
2695         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2696
2697         RETURN(rc);
2698 }
2699
2700 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2701                      struct ptlrpc_request **request)
2702 {
2703         struct obd_device *obd = exp->exp_obd;
2704         struct lmv_obd *lmv = &obd->u.lmv;
2705         struct lmv_tgt_desc *tgt;
2706         int rc;
2707
2708         ENTRY;
2709
2710         tgt = lmv_fid2tgt(lmv, fid);
2711         if (IS_ERR(tgt))
2712                 RETURN(PTR_ERR(tgt));
2713
2714         rc = md_fsync(tgt->ltd_exp, fid, request);
2715         RETURN(rc);
2716 }
2717
2718 struct stripe_dirent {
2719         struct page             *sd_page;
2720         struct lu_dirpage       *sd_dp;
2721         struct lu_dirent        *sd_ent;
2722         bool                     sd_eof;
2723 };
2724
2725 struct lmv_dir_ctxt {
2726         struct lmv_obd          *ldc_lmv;
2727         struct md_op_data       *ldc_op_data;
2728         struct md_readdir_info  *ldc_mrinfo;
2729         __u64                    ldc_hash;
2730         int                      ldc_count;
2731         struct stripe_dirent     ldc_stripes[0];
2732 };
2733
2734 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2735 {
2736         if (stripe->sd_page) {
2737                 kunmap(stripe->sd_page);
2738                 put_page(stripe->sd_page);
2739                 stripe->sd_page = NULL;
2740                 stripe->sd_ent = NULL;
2741         }
2742 }
2743
2744 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2745 {
2746         int i;
2747
2748         for (i = 0; i < ctxt->ldc_count; i++)
2749                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2750 }
2751
2752 /* if @ent is dummy, or . .., get next */
2753 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2754                                            struct lu_dirent *ent,
2755                                            int stripe_index)
2756 {
2757         for (; ent; ent = lu_dirent_next(ent)) {
2758                 /* Skip dummy entry */
2759                 if (le16_to_cpu(ent->lde_namelen) == 0)
2760                         continue;
2761
2762                 /* skip . and .. for other stripes */
2763                 if (stripe_index &&
2764                     (strncmp(ent->lde_name, ".",
2765                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2766                      strncmp(ent->lde_name, "..",
2767                              le16_to_cpu(ent->lde_namelen)) == 0))
2768                         continue;
2769
2770                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2771                         break;
2772         }
2773
2774         return ent;
2775 }
2776
2777 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2778                                             struct stripe_dirent *stripe,
2779                                             int stripe_index)
2780 {
2781         struct md_op_data *op_data = ctxt->ldc_op_data;
2782         struct lmv_oinfo *oinfo;
2783         struct lu_fid fid = op_data->op_fid1;
2784         struct inode *inode = op_data->op_data;
2785         struct lmv_tgt_desc *tgt;
2786         struct lu_dirent *ent = stripe->sd_ent;
2787         __u64 hash = ctxt->ldc_hash;
2788         int rc = 0;
2789
2790         ENTRY;
2791
2792         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2793         LASSERT(!ent);
2794
2795         do {
2796                 if (stripe->sd_page) {
2797                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2798
2799                         /* @hash should be the last dirent hash */
2800                         LASSERTF(hash <= end,
2801                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2802                                  ctxt, stripe, hash, end);
2803                         /* unload last page */
2804                         stripe_dirent_unload(stripe);
2805                         /* eof */
2806                         if (end == MDS_DIR_END_OFF) {
2807                                 stripe->sd_eof = true;
2808                                 break;
2809                         }
2810                         hash = end;
2811                 }
2812
2813                 oinfo = &op_data->op_lso1->lso_lsm.lsm_md_oinfo[stripe_index];
2814                 if (!oinfo->lmo_root) {
2815                         rc = -ENOENT;
2816                         break;
2817                 }
2818
2819                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2820                 if (!tgt) {
2821                         rc = -ENODEV;
2822                         break;
2823                 }
2824
2825                 /* op_data is shared by stripes, reset after use */
2826                 op_data->op_fid1 = oinfo->lmo_fid;
2827                 op_data->op_fid2 = oinfo->lmo_fid;
2828                 op_data->op_data = oinfo->lmo_root;
2829
2830                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_mrinfo, hash,
2831                                   &stripe->sd_page);
2832
2833                 op_data->op_fid1 = fid;
2834                 op_data->op_fid2 = fid;
2835                 op_data->op_data = inode;
2836
2837                 if (rc)
2838                         break;
2839
2840                 stripe->sd_dp = page_address(stripe->sd_page);
2841                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2842                                         stripe_index);
2843                 /* in case a page filled with ., .. and dummy, read next */
2844         } while (!ent);
2845
2846         stripe->sd_ent = ent;
2847         if (rc) {
2848                 LASSERT(!ent);
2849                 /* treat error as eof, so dir can be partially accessed */
2850                 stripe->sd_eof = true;
2851                 ctxt->ldc_mrinfo->mr_partial_readdir_rc = rc;
2852                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2853                               "directory is partially accessed!\n",
2854                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2855                               rc);
2856         }
2857
2858         RETURN(ent);
2859 }
2860
2861 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2862 {
2863         struct obd_device *obd = exp->exp_obd;
2864         struct lmv_obd *lmv = &obd->u.lmv;
2865         struct lmv_tgt_desc *tgt;
2866         int rc;
2867
2868         ENTRY;
2869
2870         rc = lmv_check_connect(obd);
2871         if (rc != 0)
2872                 RETURN(rc);
2873
2874         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2875         if (IS_ERR(tgt))
2876                 RETURN(PTR_ERR(tgt));
2877
2878         data->op_flags |= MF_MDC_CANCEL_FID1;
2879         rc = md_file_resync(tgt->ltd_exp, data);
2880         RETURN(rc);
2881 }
2882
2883 /**
2884  * Get dirent with the closest hash for striped directory
2885  *
2886  * This function will search the dir entry, whose hash value is the
2887  * closest(>=) to hash from all of sub-stripes, and it is only being called
2888  * for striped directory.
2889  *
2890  * \param[in] ctxt              dir read context
2891  *
2892  * \retval                      dirent get the entry successfully
2893  *                              NULL does not get the entry, normally it means
2894  *                              it reaches the end of the directory, while read
2895  *                              stripe dirent error is ignored to allow partial
2896  *                              access.
2897  */
2898 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2899 {
2900         struct stripe_dirent *stripe;
2901         struct lu_dirent *ent = NULL;
2902         int i;
2903         int min = -1;
2904
2905         /* TODO: optimize with k-way merge sort */
2906         for (i = 0; i < ctxt->ldc_count; i++) {
2907                 stripe = &ctxt->ldc_stripes[i];
2908                 if (stripe->sd_eof)
2909                         continue;
2910
2911                 if (!stripe->sd_ent) {
2912                         stripe_dirent_load(ctxt, stripe, i);
2913                         if (!stripe->sd_ent) {
2914                                 LASSERT(stripe->sd_eof);
2915                                 continue;
2916                         }
2917                 }
2918
2919                 if (min == -1 ||
2920                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2921                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2922                         min = i;
2923                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2924                             ctxt->ldc_hash)
2925                                 break;
2926                 }
2927         }
2928
2929         if (min != -1) {
2930                 stripe = &ctxt->ldc_stripes[min];
2931                 ent = stripe->sd_ent;
2932                 /* pop found dirent */
2933                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2934                                                    min);
2935         }
2936
2937         return ent;
2938 }
2939
2940 /**
2941  * Build dir entry page for striped directory
2942  *
2943  * This function gets one entry by @offset from a striped directory. It will
2944  * read entries from all of stripes, and choose one closest to the required
2945  * offset(&offset). A few notes
2946  * 1. skip . and .. for non-zero stripes, because there can only have one .
2947  * and .. in a directory.
2948  * 2. op_data will be shared by all of stripes, instead of allocating new
2949  * one, so need to restore before reusing.
2950  *
2951  * \param[in] exp       obd export refer to LMV
2952  * \param[in] op_data   hold those MD parameters of read_entry
2953  * \param[in] mrinfo    ldlm callback being used in enqueue in mdc_read_entry,
2954  *                      and partial readdir result will be stored in it.
2955  * \param[in] offset    starting hash offset
2956  * \param[out] ppage    the page holding the entry. Note: because the entry
2957  *                      will be accessed in upper layer, so we need hold the
2958  *                      page until the usages of entry is finished, see
2959  *                      ll_dir_entry_next.
2960  *
2961  * retval               =0 if get entry successfully
2962  *                      <0 cannot get entry
2963  */
2964 static int lmv_striped_read_page(struct obd_export *exp,
2965                                  struct md_op_data *op_data,
2966                                  struct md_readdir_info *mrinfo, __u64 offset,
2967                                  struct page **ppage)
2968 {
2969         struct page *page = NULL;
2970         struct lu_dirpage *dp;
2971         void *start;
2972         struct lu_dirent *ent;
2973         struct lu_dirent *last_ent;
2974         int stripe_count;
2975         struct lmv_dir_ctxt *ctxt;
2976         struct lu_dirent *next = NULL;
2977         __u16 ent_size;
2978         size_t left_bytes;
2979         int rc = 0;
2980         ENTRY;
2981
2982         /* Allocate a page and read entries from all of stripes and fill
2983          * the page by hash order */
2984         page = alloc_page(GFP_KERNEL);
2985         if (!page)
2986                 RETURN(-ENOMEM);
2987
2988         /* Initialize the entry page */
2989         dp = kmap(page);
2990         memset(dp, 0, sizeof(*dp));
2991         dp->ldp_hash_start = cpu_to_le64(offset);
2992
2993         start = dp + 1;
2994         left_bytes = PAGE_SIZE - sizeof(*dp);
2995         ent = start;
2996         last_ent = ent;
2997
2998         /* initalize dir read context */
2999         stripe_count = op_data->op_lso1->lso_lsm.lsm_md_stripe_count;
3000         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
3001         if (!ctxt)
3002                 GOTO(free_page, rc = -ENOMEM);
3003         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
3004         ctxt->ldc_op_data = op_data;
3005         ctxt->ldc_mrinfo = mrinfo;
3006         ctxt->ldc_hash = offset;
3007         ctxt->ldc_count = stripe_count;
3008
3009         while (1) {
3010                 next = lmv_dirent_next(ctxt);
3011
3012                 /* end of directory */
3013                 if (!next) {
3014                         ctxt->ldc_hash = MDS_DIR_END_OFF;
3015                         break;
3016                 }
3017                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
3018
3019                 ent_size = le16_to_cpu(next->lde_reclen);
3020
3021                 /* the last entry lde_reclen is 0, but it might not be the last
3022                  * one of this temporay dir page */
3023                 if (!ent_size)
3024                         ent_size = lu_dirent_calc_size(
3025                                         le16_to_cpu(next->lde_namelen),
3026                                         le32_to_cpu(next->lde_attrs));
3027                 /* page full */
3028                 if (ent_size > left_bytes)
3029                         break;
3030
3031                 memcpy(ent, next, ent_size);
3032
3033                 /* Replace . with master FID and Replace .. with the parent FID
3034                  * of master object */
3035                 if (strncmp(ent->lde_name, ".",
3036                             le16_to_cpu(ent->lde_namelen)) == 0 &&
3037                     le16_to_cpu(ent->lde_namelen) == 1)
3038                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
3039                 else if (strncmp(ent->lde_name, "..",
3040                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
3041                            le16_to_cpu(ent->lde_namelen) == 2)
3042                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
3043
3044                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
3045                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
3046                        le64_to_cpu(ent->lde_hash));
3047
3048                 left_bytes -= ent_size;
3049                 ent->lde_reclen = cpu_to_le16(ent_size);
3050                 last_ent = ent;
3051                 ent = (void *)ent + ent_size;
3052         };
3053
3054         last_ent->lde_reclen = 0;
3055
3056         if (ent == start)
3057                 dp->ldp_flags |= LDF_EMPTY;
3058         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
3059                 dp->ldp_flags |= LDF_COLLIDE;
3060         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
3061         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
3062
3063         put_lmv_dir_ctxt(ctxt);
3064         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
3065
3066         *ppage = page;
3067
3068         RETURN(0);
3069
3070 free_page:
3071         kunmap(page);
3072         __free_page(page);
3073
3074         return rc;
3075 }
3076
3077 static int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
3078                          struct md_readdir_info *mrinfo, __u64 offset,
3079                          struct page **ppage)
3080 {
3081         struct obd_device *obd = exp->exp_obd;
3082         struct lmv_obd *lmv = &obd->u.lmv;
3083         struct lmv_tgt_desc *tgt;
3084         int rc;
3085
3086         ENTRY;
3087
3088         if (unlikely(lmv_dir_foreign(op_data->op_lso1)))
3089                 RETURN(-ENODATA);
3090
3091         if (unlikely(lmv_dir_striped(op_data->op_lso1))) {
3092                 rc = lmv_striped_read_page(exp, op_data, mrinfo, offset, ppage);
3093                 RETURN(rc);
3094         }
3095
3096         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
3097         if (IS_ERR(tgt))
3098                 RETURN(PTR_ERR(tgt));
3099
3100         rc = md_read_page(tgt->ltd_exp, op_data, mrinfo, offset, ppage);
3101
3102         RETURN(rc);
3103 }
3104
3105 /**
3106  * Unlink a file/directory
3107  *
3108  * Unlink a file or directory under the parent dir. The unlink request
3109  * usually will be sent to the MDT where the child is located, but if
3110  * the client does not have the child FID then request will be sent to the
3111  * MDT where the parent is located.
3112  *
3113  * If the parent is a striped directory then it also needs to locate which
3114  * stripe the name of the child is located, and replace the parent FID
3115  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
3116  * it will walk through all of sub-stripes until the child is being
3117  * unlinked finally.
3118  *
3119  * \param[in] exp       export refer to LMV
3120  * \param[in] op_data   different parameters transferred beween client
3121  *                      MD stacks, name, namelen, FIDs etc.
3122  *                      op_fid1 is the parent FID, op_fid2 is the child
3123  *                      FID.
3124  * \param[out] request  point to the request of unlink.
3125  *
3126  * retval               0 if succeed
3127  *                      negative errno if failed.
3128  */
3129 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
3130                       struct ptlrpc_request **request)
3131 {
3132         struct obd_device *obd = exp->exp_obd;
3133         struct lmv_obd *lmv = &obd->u.lmv;
3134         struct lmv_tgt_desc *tgt;
3135         struct lmv_tgt_desc *parent_tgt;
3136         struct mdt_body *body;
3137         int rc;
3138
3139         ENTRY;
3140
3141         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
3142         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
3143         op_data->op_cap = current_cap();
3144
3145 retry:
3146         parent_tgt = lmv_locate_tgt(lmv, op_data);
3147         if (IS_ERR(parent_tgt))
3148                 RETURN(PTR_ERR(parent_tgt));
3149
3150         if (likely(!fid_is_zero(&op_data->op_fid2))) {
3151                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3152                 if (IS_ERR(tgt))
3153                         RETURN(PTR_ERR(tgt));
3154         } else {
3155                 tgt = parent_tgt;
3156         }
3157
3158         /*
3159          * If child's fid is given, cancel unused locks for it if it is from
3160          * another export than parent.
3161          *
3162          * LOOKUP lock for child (fid3) should also be cancelled on parent
3163          * tgt_tgt in mdc_unlink().
3164          */
3165         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
3166
3167         if (parent_tgt != tgt)
3168                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
3169                                       LCK_EX, MDS_INODELOCK_LOOKUP,
3170                                       MF_MDC_CANCEL_FID3);
3171
3172         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
3173                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
3174         if (rc)
3175                 RETURN(rc);
3176
3177         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
3178                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
3179                tgt->ltd_index);
3180
3181         rc = md_unlink(tgt->ltd_exp, op_data, request);
3182         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
3183                 ptlrpc_req_finished(*request);
3184                 *request = NULL;
3185                 goto retry;
3186         }
3187
3188         if (rc != -EREMOTE)
3189                 RETURN(rc);
3190
3191         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
3192         if (body == NULL)
3193                 RETURN(-EPROTO);
3194
3195         /* Not cross-ref case, just get out of here. */
3196         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
3197                 RETURN(rc);
3198
3199         /* This is a remote object, try remote MDT. */
3200         op_data->op_fid2 = body->mbo_fid1;
3201         ptlrpc_req_finished(*request);
3202         *request = NULL;
3203
3204         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3205         if (IS_ERR(tgt))
3206                 RETURN(PTR_ERR(tgt));
3207
3208         goto retry;
3209 }
3210
3211 static int lmv_precleanup(struct obd_device *obd)
3212 {
3213         ENTRY;
3214         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
3215         RETURN(0);
3216 }
3217
3218 /**
3219  * Get by key a value associated with a LMV device.
3220  *
3221  * Dispatch request to lower-layer devices as needed.
3222  *
3223  * \param[in] env               execution environment for this thread
3224  * \param[in] exp               export for the LMV device
3225  * \param[in] keylen            length of key identifier
3226  * \param[in] key               identifier of key to get value for
3227  * \param[in] vallen            size of \a val
3228  * \param[out] val              pointer to storage location for value
3229  * \param[in] lsm               optional striping metadata of object
3230  *
3231  * \retval 0            on success
3232  * \retval negative     negated errno on failure
3233  */
3234 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
3235                         __u32 keylen, void *key, __u32 *vallen, void *val)
3236 {
3237         struct obd_device *obd;
3238         struct lmv_obd *lmv;
3239         struct lu_tgt_desc *tgt;
3240         int rc = 0;
3241
3242         ENTRY;
3243
3244         obd = class_exp2obd(exp);
3245         if (obd == NULL) {
3246                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3247                        exp->exp_handle.h_cookie);
3248                 RETURN(-EINVAL);
3249         }
3250
3251         lmv = &obd->u.lmv;
3252         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
3253                 LASSERT(*vallen == sizeof(__u32));
3254                 lmv_foreach_connected_tgt(lmv, tgt) {
3255                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
3256                                           vallen, val))
3257                                 RETURN(0);
3258                 }
3259                 RETURN(-EINVAL);
3260         } else if (KEY_IS(KEY_MAX_EASIZE) ||
3261                    KEY_IS(KEY_DEFAULT_EASIZE) ||
3262                    KEY_IS(KEY_CONN_DATA)) {
3263                 /*
3264                  * Forwarding this request to first MDS, it should know LOV
3265                  * desc.
3266                  */
3267                 tgt = lmv_tgt(lmv, 0);
3268                 if (!tgt)
3269                         RETURN(-ENODEV);
3270
3271                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
3272                 if (!rc && KEY_IS(KEY_CONN_DATA))
3273                         exp->exp_connect_data = *(struct obd_connect_data *)val;
3274                 RETURN(rc);
3275         } else if (KEY_IS(KEY_TGT_COUNT)) {
3276                 *((int *)val) = lmv->lmv_mdt_descs.ltd_tgts_size;
3277                 RETURN(0);
3278         }
3279
3280         CDEBUG(D_IOCTL, "Invalid key\n");
3281         RETURN(-EINVAL);
3282 }
3283
3284 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
3285                      int *__rcs, struct ptlrpc_request_set *_set)
3286 {
3287         struct obd_device *obd = class_exp2obd(exp);
3288         struct ptlrpc_request_set *set = _set;
3289         struct lmv_obd *lmv = &obd->u.lmv;
3290         int tgt_count = lmv->lmv_mdt_count;
3291         struct lu_tgt_desc *tgt;
3292         struct fid_array *fat, **fas = NULL;
3293         int i, rc, **rcs = NULL;
3294
3295         if (!set) {
3296                 set = ptlrpc_prep_set();
3297                 if (!set)
3298                         RETURN(-ENOMEM);
3299         }
3300
3301         /* split FIDs by targets */
3302         OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
3303         if (fas == NULL)
3304                 GOTO(out, rc = -ENOMEM);
3305         OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
3306         if (rcs == NULL)
3307                 GOTO(out_fas, rc = -ENOMEM);
3308
3309         for (i = 0; i < fa->fa_nr; i++) {
3310                 unsigned int idx;
3311
3312                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
3313                 if (rc) {
3314                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
3315                                PFID(&fa->fa_fids[i]), rc);
3316                         continue;
3317                 }
3318                 LASSERT(idx < tgt_count);
3319                 if (!fas[idx])
3320                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
3321                                   fa_fids[fa->fa_nr]));
3322                 if (!fas[idx])
3323                         GOTO(out, rc = -ENOMEM);
3324                 if (!rcs[idx])
3325                         OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
3326                 if (!rcs[idx])
3327                         GOTO(out, rc = -ENOMEM);
3328
3329                 fat = fas[idx];
3330                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
3331         }
3332
3333         lmv_foreach_connected_tgt(lmv, tgt) {
3334                 fat = fas[tgt->ltd_index];
3335                 if (!fat || fat->fa_nr == 0)
3336                         continue;
3337                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3338         }
3339
3340         rc = ptlrpc_set_wait(NULL, set);
3341         if (rc == 0) {
3342                 int j = 0;
3343                 for (i = 0; i < tgt_count; i++) {
3344                         fat = fas[i];
3345                         if (!fat || fat->fa_nr == 0)
3346                                 continue;
3347                         /* copy FIDs back */
3348                         memcpy(fa->fa_fids + j, fat->fa_fids,
3349                                fat->fa_nr * sizeof(struct lu_fid));
3350                         /* copy rcs back */
3351                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3352                         j += fat->fa_nr;
3353                 }
3354         }
3355         if (set != _set)
3356                 ptlrpc_set_destroy(set);
3357
3358 out:
3359         for (i = 0; i < tgt_count; i++) {
3360                 if (fas && fas[i])
3361                         OBD_FREE(fas[i], offsetof(struct fid_array,
3362                                                 fa_fids[fa->fa_nr]));
3363                 if (rcs && rcs[i])
3364                         OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3365         }
3366         if (rcs)
3367                 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3368 out_fas:
3369         if (fas)
3370                 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3371
3372         RETURN(rc);
3373 }
3374
3375 /**
3376  * Asynchronously set by key a value associated with a LMV device.
3377  *
3378  * Dispatch request to lower-layer devices as needed.
3379  *
3380  * \param[in] env       execution environment for this thread
3381  * \param[in] exp       export for the LMV device
3382  * \param[in] keylen    length of key identifier
3383  * \param[in] key       identifier of key to store value for
3384  * \param[in] vallen    size of value to store
3385  * \param[in] val       pointer to data to be stored
3386  * \param[in] set       optional list of related ptlrpc requests
3387  *
3388  * \retval 0            on success
3389  * \retval negative     negated errno on failure
3390  */
3391 static int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3392                               __u32 keylen, void *key, __u32 vallen, void *val,
3393                               struct ptlrpc_request_set *set)
3394 {
3395         struct lmv_tgt_desc *tgt;
3396         struct obd_device *obd;
3397         struct lmv_obd *lmv;
3398         int rc = 0;
3399         ENTRY;
3400
3401         obd = class_exp2obd(exp);
3402         if (obd == NULL) {
3403                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3404                        exp->exp_handle.h_cookie);
3405                 RETURN(-EINVAL);
3406         }
3407         lmv = &obd->u.lmv;
3408
3409         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3410             KEY_IS(KEY_DEFAULT_EASIZE)) {
3411                 int err = 0;
3412
3413                 lmv_foreach_connected_tgt(lmv, tgt) {
3414                         err = obd_set_info_async(env, tgt->ltd_exp,
3415                                                  keylen, key, vallen, val, set);
3416                         if (err && rc == 0)
3417                                 rc = err;
3418                 }
3419
3420                 RETURN(rc);
3421         }
3422
3423         RETURN(-EINVAL);
3424 }
3425
3426 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3427                             const struct lmv_mds_md_v1 *lmm1)
3428 {
3429         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3430         int             stripe_count;
3431         int             cplen;
3432         int             i;
3433         int             rc = 0;
3434         ENTRY;
3435
3436         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3437         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3438         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3439         if (CFS_FAIL_CHECK(OBD_FAIL_LMV_UNKNOWN_STRIPE))
3440                 lsm->lsm_md_hash_type = cfs_fail_val ?: LMV_HASH_TYPE_UNKNOWN;
3441         else
3442                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3443         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3444         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3445         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3446         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3447                         sizeof(lsm->lsm_md_pool_name));
3448
3449         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3450                 RETURN(-E2BIG);
3451
3452         CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3453                "layout_version %d\n", lsm->lsm_md_stripe_count,
3454                lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3455                lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3456                lsm->lsm_md_layout_version);
3457
3458         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3459         for (i = 0; i < stripe_count; i++) {
3460                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3461                               &lmm1->lmv_stripe_fids[i]);
3462                 /*
3463                  * set default value -1, so lmv_locate_tgt() knows this stripe
3464                  * target is not initialized.
3465                  */
3466                 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3467                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3468                         continue;
3469
3470                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3471                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3472                 if (rc == -ENOENT)
3473                         continue;
3474
3475                 if (rc)
3476                         RETURN(rc);
3477
3478                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3479                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3480         }
3481
3482         RETURN(rc);
3483 }
3484
3485 static inline int lmv_unpack_user_md(struct obd_export *exp,
3486                                      struct lmv_stripe_md *lsm,
3487                                      const struct lmv_user_md *lmu)
3488 {
3489         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3490         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3491         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3492         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3493         lsm->lsm_md_max_inherit = lmu->lum_max_inherit;
3494         lsm->lsm_md_max_inherit_rr = lmu->lum_max_inherit_rr;
3495         lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3496
3497         return 0;
3498 }
3499
3500 struct lmv_stripe_object *lmv_stripe_object_alloc(__u32 magic,
3501                                                   const union lmv_mds_md *lmm,
3502                                                   size_t lmm_size)
3503 {
3504         struct lmv_stripe_object *lsm_obj;
3505         int size;
3506
3507         if (magic == LMV_MAGIC_FOREIGN) {
3508                 struct lmv_foreign_md *lfm;
3509
3510                 size = offsetof(typeof(*lfm), lfm_value[0]);
3511                 if (lmm_size < size)
3512                         RETURN(ERR_PTR(-EPROTO));
3513
3514                 size += le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3515                 if (lmm_size < size)
3516                         RETURN(ERR_PTR(-EPROTO));
3517
3518                 OBD_ALLOC_LARGE(lsm_obj, lmm_size +
3519                                 offsetof(typeof(*lsm_obj), lso_lfm));
3520         } else {
3521                 if (magic == LMV_MAGIC_V1) {
3522                         int count;
3523
3524                         size = offsetof(struct lmv_mds_md_v1,
3525                                         lmv_stripe_fids[0]);
3526                         if (lmm_size < size)
3527                                 RETURN(ERR_PTR(-EPROTO));
3528
3529                         count = lmv_mds_md_stripe_count_get(lmm);
3530                         size += count * sizeof(struct lu_fid);
3531                         if (lmm_size < size)
3532                                 RETURN(ERR_PTR(-EPROTO));
3533
3534                         size = lmv_stripe_md_size(count);
3535                 } else {
3536                         if (lmm && lmm_size < sizeof(struct lmv_user_md))
3537                                 RETURN(ERR_PTR(-EPROTO));
3538
3539                         /**
3540                          * Unpack default dirstripe(lmv_user_md) to
3541                          * lmv_stripe_md, stripecount should be 0 then.
3542                          */
3543                         size = lmv_stripe_md_size(0);
3544                 }
3545                 size += offsetof(typeof(*lsm_obj), lso_lsm);
3546                 OBD_ALLOC(lsm_obj, size);
3547         }
3548
3549         if (lsm_obj) {
3550                 atomic_set(&lsm_obj->lso_refs, 1);
3551                 RETURN(lsm_obj);
3552         }
3553
3554         RETURN(ERR_PTR(-ENOMEM));
3555 }
3556 EXPORT_SYMBOL(lmv_stripe_object_alloc);
3557
3558 static int lmv_stripe_object_create(struct obd_export *exp,
3559                                     struct lmv_stripe_object **lsop,
3560                                     const union lmv_mds_md *lmm,
3561                                     size_t lmm_size)
3562 {
3563         struct lmv_stripe_object *lsm_obj;
3564         __u32 magic;
3565         int rc;
3566         ENTRY;
3567
3568         LASSERT(lsop != NULL && *lsop == NULL);
3569
3570         if (lmm_size == 0)
3571                 RETURN(-EPROTO);
3572
3573         magic = le32_to_cpu(lmm->lmv_magic);
3574         if (magic == LMV_MAGIC_STRIPE)
3575                 RETURN(-EPERM);
3576
3577         if (magic != LMV_MAGIC_V1 && magic != LMV_USER_MAGIC &&
3578             magic != LMV_MAGIC_FOREIGN) {
3579                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3580                        exp->exp_obd->obd_name, magic, -EIO);
3581                 RETURN(-EIO);
3582         }
3583
3584         /* foreign lmv case */
3585         if (magic == LMV_MAGIC_FOREIGN) {
3586                 struct lmv_foreign_md *lfm;
3587
3588                 lsm_obj = lmv_stripe_object_alloc(magic, lmm, lmm_size);
3589                 if (IS_ERR(lsm_obj))
3590                         RETURN(PTR_ERR(lsm_obj));
3591
3592                 *lsop = lsm_obj;
3593                 lfm = &lsm_obj->lso_lfm;
3594                 lfm->lfm_magic = magic;
3595                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3596                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3597                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3598                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3599                        lfm->lfm_length);
3600                 RETURN(0);
3601         }
3602
3603         /* Unpack memmd */
3604         lsm_obj = lmv_stripe_object_alloc(magic, lmm, lmm_size);
3605         if (IS_ERR(lsm_obj))
3606                 RETURN(PTR_ERR(lsm_obj));
3607
3608         switch (magic) {
3609         case LMV_MAGIC_V1:
3610                 rc = lmv_unpack_md_v1(exp, &lsm_obj->lso_lsm, &lmm->lmv_md_v1);
3611                 break;
3612         case LMV_USER_MAGIC:
3613                 rc = lmv_unpack_user_md(exp, &lsm_obj->lso_lsm,
3614                                         &lmm->lmv_user_md);
3615                 break;
3616         default:
3617                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3618                        magic);
3619                 rc = -EINVAL;
3620                 break;
3621         }
3622
3623         if (rc != 0)
3624                 lmv_stripe_object_put(&lsm_obj);
3625
3626         *lsop = lsm_obj;
3627         RETURN(rc);
3628 }
3629
3630 struct lmv_stripe_object *
3631 lmv_stripe_object_get(struct lmv_stripe_object *lsm_obj)
3632 {
3633         if (lsm_obj == NULL)
3634                 return NULL;
3635
3636         atomic_inc(&lsm_obj->lso_refs);
3637         CDEBUG(D_INODE, "get %p %u\n", lsm_obj,
3638                atomic_read(&lsm_obj->lso_refs));
3639         return lsm_obj;
3640 }
3641 EXPORT_SYMBOL(lmv_stripe_object_get);
3642
3643 void lmv_stripe_object_put(struct lmv_stripe_object **lsop)
3644 {
3645         struct lmv_stripe_object *lsm_obj;
3646         size_t size;
3647         int i;
3648
3649         LASSERT(lsop != NULL);
3650
3651         lsm_obj = *lsop;
3652         if (lsm_obj == NULL)
3653                 return;
3654
3655         *lsop = NULL;
3656         CDEBUG(D_INODE, "put %p %u\n", lsm_obj,
3657                atomic_read(&lsm_obj->lso_refs) - 1);
3658
3659         if (!atomic_dec_and_test(&lsm_obj->lso_refs))
3660                 return;
3661
3662         if (lmv_dir_foreign(lsm_obj)) {
3663                 size = lsm_obj->lso_lfm.lfm_length +
3664                         offsetof(typeof(lsm_obj->lso_lfm), lfm_value[0]) +
3665                         offsetof(typeof(*lsm_obj), lso_lsm);
3666                 OBD_FREE_LARGE(lsm_obj, size);
3667                 return;
3668         }
3669
3670         if (lmv_dir_striped(lsm_obj)) {
3671                 struct lmv_stripe_md *lsm = &lsm_obj->lso_lsm;
3672
3673                 for (i = 0; i < lsm->lsm_md_stripe_count; i++)
3674                         iput(lsm->lsm_md_oinfo[i].lmo_root);
3675                 size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3676         } else {
3677                 size = lmv_stripe_md_size(0);
3678         }
3679         OBD_FREE(lsm_obj, size + offsetof(typeof(*lsm_obj), lso_lsm));
3680 }
3681 EXPORT_SYMBOL(lmv_stripe_object_put);
3682
3683 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3684                              union ldlm_policy_data *policy,
3685                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3686                              void *opaque)
3687 {
3688         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3689         struct lu_tgt_desc *tgt;
3690         int err;
3691         int rc = 0;
3692
3693         ENTRY;
3694
3695         LASSERT(fid != NULL);
3696
3697         lmv_foreach_connected_tgt(lmv, tgt) {
3698                 if (!tgt->ltd_active)
3699                         continue;
3700
3701                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3702                                        opaque);
3703                 if (!rc)
3704                         rc = err;
3705         }
3706         RETURN(rc);
3707 }
3708
3709 static int lmv_set_lock_data(struct obd_export *exp,
3710                              const struct lustre_handle *lockh,
3711                              void *data, __u64 *bits)
3712 {
3713         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3714         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3715         int rc;
3716
3717         ENTRY;
3718
3719         if (tgt == NULL || tgt->ltd_exp == NULL)
3720                 RETURN(-EINVAL);
3721         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3722         RETURN(rc);
3723 }
3724
3725 static enum ldlm_mode
3726 lmv_lock_match(struct obd_export *exp, __u64 flags,
3727                const struct lu_fid *fid, enum ldlm_type type,
3728                union ldlm_policy_data *policy,
3729                enum ldlm_mode mode, struct lustre_handle *lockh)
3730 {
3731         struct obd_device *obd = exp->exp_obd;
3732         struct lmv_obd *lmv = &obd->u.lmv;
3733         struct lu_tgt_desc *tgt;
3734         __u64 bits = policy->l_inodebits.bits;
3735         enum ldlm_mode rc = LCK_MINMODE;
3736         int index;
3737         int i;
3738
3739         /* only one bit is set */
3740         LASSERT(bits && !(bits & (bits - 1)));
3741         /* With DNE every object can have two locks in different namespaces:
3742          * lookup lock in space of MDT storing direntry and update/open lock in
3743          * space of MDT storing inode.  Try the MDT that the FID maps to first,
3744          * since this can be easily found, and only try others if that fails.
3745          */
3746         if (bits == MDS_INODELOCK_LOOKUP) {
3747                 for (i = 0, index = lmv_fid2tgt_index(lmv, fid);
3748                      i < lmv->lmv_mdt_descs.ltd_tgts_size; i++,
3749                      index = (index + 1) % lmv->lmv_mdt_descs.ltd_tgts_size) {
3750                         if (index < 0) {
3751                                 CDEBUG(D_HA,
3752                                        "%s: "DFID" is inaccessible: rc = %d\n",
3753                                        obd->obd_name, PFID(fid), index);
3754                                 index = 0;
3755                         }
3756                         tgt = lmv_tgt(lmv, index);
3757                         if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
3758                                 continue;
3759                         rc = md_lock_match(tgt->ltd_exp, flags, fid, type,
3760                                            policy, mode, lockh);
3761                         if (rc)
3762                                 break;
3763                 }
3764         } else {
3765                 tgt = lmv_fid2tgt(lmv, fid);
3766                 if (!IS_ERR(tgt) && tgt->ltd_exp && tgt->ltd_active)
3767                         rc = md_lock_match(tgt->ltd_exp, flags, fid, type,
3768                                            policy, mode, lockh);
3769         }
3770
3771         CDEBUG(D_INODE, "Lock match for "DFID": %d\n", PFID(fid), rc);
3772
3773         return rc;
3774 }
3775
3776 static int
3777 lmv_get_lustre_md(struct obd_export *exp, struct req_capsule *pill,
3778                   struct obd_export *dt_exp, struct obd_export *md_exp,
3779                   struct lustre_md *md)
3780 {
3781         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3782         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3783
3784         if (!tgt || !tgt->ltd_exp)
3785                 return -EINVAL;
3786
3787         return md_get_lustre_md(tgt->ltd_exp, pill, dt_exp, md_exp, md);
3788 }
3789
3790 static int lmv_put_lustre_md(struct obd_export *exp, struct lustre_md *md)
3791 {
3792         struct obd_device *obd = exp->exp_obd;
3793         struct lmv_obd *lmv = &obd->u.lmv;
3794         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3795         ENTRY;
3796
3797         lmv_stripe_object_put(&md->def_lsm_obj);
3798         lmv_stripe_object_put(&md->lsm_obj);
3799
3800         if (!tgt || !tgt->ltd_exp)
3801                 RETURN(-EINVAL);
3802         RETURN(0);
3803 }
3804
3805 static int lmv_set_open_replay_data(struct obd_export *exp,
3806                                     struct obd_client_handle *och,
3807                                     struct lookup_intent *it)
3808 {
3809         struct obd_device *obd = exp->exp_obd;
3810         struct lmv_obd *lmv = &obd->u.lmv;
3811         struct lmv_tgt_desc *tgt;
3812
3813         ENTRY;
3814
3815         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3816         if (IS_ERR(tgt))
3817                 RETURN(PTR_ERR(tgt));
3818
3819         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3820 }
3821
3822 static int lmv_clear_open_replay_data(struct obd_export *exp,
3823                                       struct obd_client_handle *och)
3824 {
3825         struct obd_device *obd = exp->exp_obd;
3826         struct lmv_obd *lmv = &obd->u.lmv;
3827         struct lmv_tgt_desc *tgt;
3828
3829         ENTRY;
3830
3831         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3832         if (IS_ERR(tgt))
3833                 RETURN(PTR_ERR(tgt));
3834
3835         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3836 }
3837
3838 static int lmv_intent_getattr_async(struct obd_export *exp,
3839                                     struct md_op_item *item)
3840 {
3841         struct md_op_data *op_data = &item->mop_data;
3842         struct obd_device *obd = exp->exp_obd;
3843         struct lmv_obd *lmv = &obd->u.lmv;
3844         struct lmv_tgt_desc *ptgt;
3845         struct lmv_tgt_desc *ctgt;
3846         int rc;
3847
3848         ENTRY;
3849
3850         if (!(fid_is_sane(&op_data->op_fid2) ||
3851               fid_is_zero(&op_data->op_fid2)))
3852                 RETURN(-EINVAL);
3853
3854         ptgt = lmv_locate_tgt(lmv, op_data);
3855         if (IS_ERR(ptgt))
3856                 RETURN(PTR_ERR(ptgt));
3857
3858         /*
3859          * Zeroed FID @op_fid2 means that the intent getattr() comes from
3860          * statahead by regularized file names. Currently only do statahead
3861          * for the children files located same as the parent directory.
3862          */
3863         if (!fid_is_zero(&op_data->op_fid2)) {
3864                 ctgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3865                 if (IS_ERR(ctgt))
3866                         RETURN(PTR_ERR(ctgt));
3867
3868                 /*
3869                  * remote object needs two RPCs to lookup and getattr,
3870                  * considering the complexity don't support statahead for now.
3871                  */
3872                 if (ctgt != ptgt)
3873                         RETURN(-EREMOTE);
3874         }
3875
3876         rc = md_intent_getattr_async(ptgt->ltd_exp, item);
3877
3878         RETURN(rc);
3879 }
3880
3881 static int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3882                                struct lu_fid *fid, __u64 *bits)
3883 {
3884         struct obd_device *obd = exp->exp_obd;
3885         struct lmv_obd *lmv = &obd->u.lmv;
3886         struct lmv_tgt_desc *tgt;
3887         int rc;
3888
3889         ENTRY;
3890
3891         tgt = lmv_fid2tgt(lmv, fid);
3892         if (IS_ERR(tgt))
3893                 RETURN(PTR_ERR(tgt));
3894
3895         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3896         RETURN(rc);
3897 }
3898
3899 static int lmv_get_fid_from_lsm(struct obd_export *exp,
3900                                 const struct lmv_stripe_object *lso,
3901                                 const char *name, int namelen,
3902                                 struct lu_fid *fid)
3903 {
3904         const struct lmv_oinfo *oinfo;
3905
3906         LASSERT(lmv_dir_striped(lso));
3907         oinfo = lsm_name_to_stripe_info(lso, name, namelen, false);
3908         if (IS_ERR(oinfo))
3909                 return PTR_ERR(oinfo);
3910
3911         *fid = oinfo->lmo_fid;
3912
3913         RETURN(0);
3914 }
3915
3916 /**
3917  * For lmv, only need to send request to master MDT, and the master MDT will
3918  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3919  * we directly fetch data from the slave MDTs.
3920  */
3921 static int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3922                         struct obd_quotactl *oqctl)
3923 {
3924         struct obd_device *obd = class_exp2obd(exp);
3925         struct lmv_obd *lmv = &obd->u.lmv;
3926         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3927         __u64 curspace, curinodes;
3928         int rc = 0;
3929
3930         ENTRY;
3931
3932         if (!tgt || !tgt->ltd_exp || !tgt->ltd_active) {
3933                 CERROR("master lmv inactive\n");
3934                 RETURN(-EIO);
3935         }
3936
3937         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3938                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3939                 RETURN(rc);
3940         }
3941
3942         curspace = curinodes = 0;
3943         lmv_foreach_connected_tgt(lmv, tgt) {
3944                 int err;
3945
3946                 if (!tgt->ltd_active)
3947                         continue;
3948
3949                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3950                 if (err) {
3951                         CERROR("getquota on mdt %d failed. %d\n",
3952                                tgt->ltd_index, err);
3953                         if (!rc)
3954                                 rc = err;
3955                 } else {
3956                         curspace += oqctl->qc_dqblk.dqb_curspace;
3957                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3958                 }
3959         }
3960         oqctl->qc_dqblk.dqb_curspace = curspace;
3961         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3962
3963         RETURN(rc);
3964 }
3965
3966 static int lmv_merge_attr(struct obd_export *exp,
3967                           const struct lmv_stripe_object *lso,
3968                           struct cl_attr *attr,
3969                           ldlm_blocking_callback cb_blocking)
3970 {
3971         const struct lmv_stripe_md *lsm = &lso->lso_lsm;
3972         int rc;
3973         int i;
3974
3975         if (!lmv_dir_striped(lso))
3976                 return 0;
3977
3978         rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3979         if (rc < 0)
3980                 return rc;
3981
3982         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3983                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3984
3985                 if (!inode)
3986                         continue;
3987
3988                 CDEBUG(D_INFO,
3989                        "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
3990                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3991                        i_size_read(inode), (unsigned long long)inode->i_blocks,
3992                        inode->i_nlink, (s64)inode->i_atime.tv_sec,
3993                        (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
3994
3995                 /* for slave stripe, it needs to subtract nlink for . and .. */
3996                 if (i != 0)
3997                         attr->cat_nlink += inode->i_nlink - 2;
3998                 else
3999                         attr->cat_nlink = inode->i_nlink;
4000
4001                 attr->cat_size += i_size_read(inode);
4002                 attr->cat_blocks += inode->i_blocks;
4003
4004                 if (attr->cat_atime < inode->i_atime.tv_sec)
4005                         attr->cat_atime = inode->i_atime.tv_sec;
4006
4007                 if (attr->cat_ctime < inode->i_ctime.tv_sec)
4008                         attr->cat_ctime = inode->i_ctime.tv_sec;
4009
4010                 if (attr->cat_mtime < inode->i_mtime.tv_sec)
4011                         attr->cat_mtime = inode->i_mtime.tv_sec;
4012         }
4013         return 0;
4014 }
4015
4016 static struct lu_batch *lmv_batch_create(struct obd_export *exp,
4017                                          enum lu_batch_flags flags,
4018                                          __u32 max_count)
4019 {
4020         struct lu_batch *bh;
4021         struct lmv_batch *lbh;
4022
4023         ENTRY;
4024         OBD_ALLOC_PTR(lbh);
4025         if (!lbh)
4026                 RETURN(ERR_PTR(-ENOMEM));
4027
4028         bh = &lbh->lbh_super;
4029         bh->lbt_flags = flags;
4030         bh->lbt_max_count = max_count;
4031
4032         if (flags & BATCH_FL_RQSET) {
4033                 bh->lbt_rqset = ptlrpc_prep_set();
4034                 if (bh->lbt_rqset == NULL) {
4035                         OBD_FREE_PTR(lbh);
4036                         RETURN(ERR_PTR(-ENOMEM));
4037                 }
4038         }
4039
4040         INIT_LIST_HEAD(&lbh->lbh_sub_batch_list);
4041         RETURN(bh);
4042 }
4043
4044 static int lmv_batch_stop(struct obd_export *exp, struct lu_batch *bh)
4045 {
4046         struct lmv_batch *lbh;
4047         struct lmvsub_batch *sub;
4048         struct lmvsub_batch *tmp;
4049         int rc = 0;
4050
4051         ENTRY;
4052
4053         lbh = container_of(bh, struct lmv_batch, lbh_super);
4054         list_for_each_entry_safe(sub, tmp, &lbh->lbh_sub_batch_list,
4055                                  sbh_sub_item) {
4056                 list_del(&sub->sbh_sub_item);
4057                 rc = md_batch_stop(sub->sbh_tgt->ltd_exp, sub->sbh_sub);
4058                 if (rc < 0) {
4059                         CERROR("%s: stop batch processing failed: rc = %d\n",
4060                                exp->exp_obd->obd_name, rc);
4061                         if (bh->lbt_result == 0)
4062                                 bh->lbt_result = rc;
4063                 }
4064                 OBD_FREE_PTR(sub);
4065         }
4066
4067         if (bh->lbt_flags & BATCH_FL_RQSET) {
4068                 rc = ptlrpc_set_wait(NULL, bh->lbt_rqset);
4069                 ptlrpc_set_destroy(bh->lbt_rqset);
4070         }
4071
4072         OBD_FREE_PTR(lbh);
4073         RETURN(rc);
4074 }
4075
4076 static int lmv_batch_flush(struct obd_export *exp, struct lu_batch *bh,
4077                            bool wait)
4078 {
4079         struct lmv_batch *lbh;
4080         struct lmvsub_batch *sub;
4081         int rc = 0;
4082         int rc1;
4083
4084         ENTRY;
4085
4086         lbh = container_of(bh, struct lmv_batch, lbh_super);
4087         list_for_each_entry(sub, &lbh->lbh_sub_batch_list, sbh_sub_item) {
4088                 rc1 = md_batch_flush(sub->sbh_tgt->ltd_exp, sub->sbh_sub, wait);
4089                 if (rc1 < 0) {
4090                         CERROR("%s: stop batch processing failed: rc = %d\n",
4091                                exp->exp_obd->obd_name, rc);
4092                         if (bh->lbt_result == 0)
4093                                 bh->lbt_result = rc;
4094
4095                         if (rc == 0)
4096                                 rc = rc1;
4097                 }
4098         }
4099
4100         if (wait && bh->lbt_flags & BATCH_FL_RQSET) {
4101                 rc1 = ptlrpc_set_wait(NULL, bh->lbt_rqset);
4102                 if (rc == 0)
4103                         rc = rc1;
4104         }
4105
4106         RETURN(rc);
4107 }
4108
4109 static inline struct lmv_tgt_desc *
4110 lmv_batch_locate_tgt(struct lmv_obd *lmv, struct md_op_item *item)
4111 {
4112         struct md_op_data *op_data = &item->mop_data;
4113         struct lmv_tgt_desc *tgt;
4114
4115         switch (item->mop_opc) {
4116         case MD_OP_GETATTR: {
4117                 struct lmv_tgt_desc *ptgt;
4118
4119                 if (!(fid_is_sane(&op_data->op_fid2) ||
4120                       fid_is_zero(&op_data->op_fid2)))
4121                         RETURN(ERR_PTR(-EINVAL));
4122
4123                 ptgt = lmv_locate_tgt(lmv, op_data);
4124                 if (IS_ERR(ptgt))
4125                         RETURN(ptgt);
4126
4127                 /*
4128                  * Zeroed @op_fid2 means that it is a statahead populating call
4129                  * in the file name pattern which is using file name format to
4130                  * prefetch the attributes. Thus it has no idea about the FID of
4131                  * the children file. The children file is considered to be
4132                  * located on the same storage target with the parent directory
4133                  * or the stripped directory.
4134                  */
4135                 if (fid_is_zero(&op_data->op_fid2)) {
4136                         tgt = ptgt;
4137                         break;
4138                 }
4139
4140                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
4141                 if (IS_ERR(tgt))
4142                         RETURN(tgt);
4143
4144                 /*
4145                  * Remote object needs two RPCs to lookup and getattr,
4146                  * considering the complexity don't support statahead for now.
4147                  */
4148                 if (tgt != ptgt)
4149                         RETURN(ERR_PTR(-EREMOTE));
4150
4151                 break;
4152         }
4153         default:
4154                 tgt = ERR_PTR(-ENOTSUPP);
4155         }
4156
4157         return tgt;
4158 }
4159
4160 struct lu_batch *lmv_batch_lookup_sub(struct lmv_batch *lbh,
4161                                       struct lmv_tgt_desc *tgt)
4162 {
4163         struct lmvsub_batch *sub;
4164
4165         list_for_each_entry(sub, &lbh->lbh_sub_batch_list, sbh_sub_item) {
4166                 if (sub->sbh_tgt == tgt)
4167                         return sub->sbh_sub;
4168         }
4169
4170         return NULL;
4171 }
4172
4173 struct lu_batch *lmv_batch_get_sub(struct lmv_batch *lbh,
4174                                    struct lmv_tgt_desc *tgt)
4175 {
4176         struct lmvsub_batch *sbh;
4177         struct lu_batch *child_bh;
4178         struct lu_batch *bh;
4179
4180         ENTRY;
4181
4182         child_bh = lmv_batch_lookup_sub(lbh, tgt);
4183         if (child_bh != NULL)
4184                 RETURN(child_bh);
4185
4186         OBD_ALLOC_PTR(sbh);
4187         if (sbh == NULL)
4188                 RETURN(ERR_PTR(-ENOMEM));
4189
4190         INIT_LIST_HEAD(&sbh->sbh_sub_item);
4191         sbh->sbh_tgt = tgt;
4192
4193         bh = &lbh->lbh_super;
4194         child_bh = md_batch_create(tgt->ltd_exp, bh->lbt_flags,
4195                                    bh->lbt_max_count);
4196         if (IS_ERR(child_bh)) {
4197                 OBD_FREE_PTR(sbh);
4198                 RETURN(child_bh);
4199         }
4200
4201         child_bh->lbt_rqset = bh->lbt_rqset;
4202         sbh->sbh_sub = child_bh;
4203         list_add(&sbh->sbh_sub_item, &lbh->lbh_sub_batch_list);
4204         RETURN(child_bh);
4205 }
4206
4207 static int lmv_batch_add(struct obd_export *exp, struct lu_batch *bh,
4208                          struct md_op_item *item)
4209 {
4210         struct obd_device *obd = exp->exp_obd;
4211         struct lmv_obd *lmv = &obd->u.lmv;
4212         struct lmv_tgt_desc *tgt;
4213         struct lmv_batch *lbh;
4214         struct lu_batch *child_bh;
4215         int rc;
4216
4217         ENTRY;
4218
4219         tgt = lmv_batch_locate_tgt(lmv, item);
4220         if (IS_ERR(tgt))
4221                 RETURN(PTR_ERR(tgt));
4222
4223         lbh = container_of(bh, struct lmv_batch, lbh_super);
4224         child_bh = lmv_batch_get_sub(lbh, tgt);
4225         if (IS_ERR(child_bh))
4226                 RETURN(PTR_ERR(child_bh));
4227
4228         rc = md_batch_add(tgt->ltd_exp, child_bh, item);
4229         RETURN(rc);
4230 }
4231
4232 static const struct obd_ops lmv_obd_ops = {
4233         .o_owner                = THIS_MODULE,
4234         .o_setup                = lmv_setup,
4235         .o_cleanup              = lmv_cleanup,
4236         .o_precleanup           = lmv_precleanup,
4237         .o_process_config       = lmv_process_config,
4238         .o_connect              = lmv_connect,
4239         .o_disconnect           = lmv_disconnect,
4240         .o_statfs               = lmv_statfs,
4241         .o_get_info             = lmv_get_info,
4242         .o_set_info_async       = lmv_set_info_async,
4243         .o_notify               = lmv_notify,
4244         .o_get_uuid             = lmv_get_uuid,
4245         .o_fid_alloc            = lmv_fid_alloc,
4246         .o_iocontrol            = lmv_iocontrol,
4247         .o_quotactl             = lmv_quotactl
4248 };
4249
4250 static const struct md_ops lmv_md_ops = {
4251         .m_get_root             = lmv_get_root,
4252         .m_null_inode           = lmv_null_inode,
4253         .m_close                = lmv_close,
4254         .m_create               = lmv_create,
4255         .m_enqueue              = lmv_enqueue,
4256         .m_getattr              = lmv_getattr,
4257         .m_getxattr             = lmv_getxattr,
4258         .m_getattr_name         = lmv_getattr_name,
4259         .m_intent_lock          = lmv_intent_lock,
4260         .m_link                 = lmv_link,
4261         .m_rename               = lmv_rename,
4262         .m_setattr              = lmv_setattr,
4263         .m_setxattr             = lmv_setxattr,
4264         .m_fsync                = lmv_fsync,
4265         .m_file_resync          = lmv_file_resync,
4266         .m_read_page            = lmv_read_page,
4267         .m_unlink               = lmv_unlink,
4268         .m_init_ea_size         = lmv_init_ea_size,
4269         .m_cancel_unused        = lmv_cancel_unused,
4270         .m_set_lock_data        = lmv_set_lock_data,
4271         .m_lock_match           = lmv_lock_match,
4272         .m_get_lustre_md        = lmv_get_lustre_md,
4273         .m_put_lustre_md        = lmv_put_lustre_md,
4274         .m_merge_attr           = lmv_merge_attr,
4275         .m_set_open_replay_data = lmv_set_open_replay_data,
4276         .m_clear_open_replay_data = lmv_clear_open_replay_data,
4277         .m_intent_getattr_async = lmv_intent_getattr_async,
4278         .m_revalidate_lock      = lmv_revalidate_lock,
4279         .m_get_fid_from_lsm     = lmv_get_fid_from_lsm,
4280         .m_stripe_object_create = lmv_stripe_object_create,
4281         .m_rmfid                = lmv_rmfid,
4282         .m_batch_create         = lmv_batch_create,
4283         .m_batch_add            = lmv_batch_add,
4284         .m_batch_stop           = lmv_batch_stop,
4285         .m_batch_flush          = lmv_batch_flush,
4286 };
4287
4288 static int __init lmv_init(void)
4289 {
4290         int rc;
4291
4292         rc  = libcfs_setup();
4293         if (rc)
4294                 return rc;
4295
4296         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true,
4297                                    LUSTRE_LMV_NAME, NULL);
4298 }
4299
4300 static void __exit lmv_exit(void)
4301 {
4302         class_unregister_type(LUSTRE_LMV_NAME);
4303 }
4304
4305 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
4306 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
4307 MODULE_VERSION(LUSTRE_VERSION_STRING);
4308 MODULE_LICENSE("GPL");
4309
4310 module_init(lmv_init);
4311 module_exit(lmv_exit);