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_kernelcomm.h>
  56 #include "lmv_internal.h"
  57
  58 static int lmv_check_connect(struct obd_device *obd);
  59
  60 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
  61                          int activate)
  62 {
  63         if (tgt->ltd_active == activate)
  64                 return;
  65
  66         tgt->ltd_active = activate;
  67         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
  68                 (activate ? 1 : -1);
  69
  70         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
  71 }
  72
  73 /**
  74  * Error codes:
  75  *
  76  *  -EINVAL  : UUID can't be found in the LMV's target list
  77  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  78  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  79  */
  80 static int lmv_set_mdc_active(struct lmv_obd *lmv,
  81                               const struct obd_uuid *uuid,
  82                               int activate)
  83 {
  84         struct lu_tgt_desc *tgt = NULL;
  85         struct obd_device *obd;
  86         int rc = 0;
  87
  88         ENTRY;
  89
  90         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  91                         lmv, uuid->uuid, activate);
  92
  93         spin_lock(&lmv->lmv_lock);
  94         lmv_foreach_connected_tgt(lmv, tgt) {
  95                 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
  96                        tgt->ltd_index, tgt->ltd_uuid.uuid,
  97                        tgt->ltd_exp->exp_handle.h_cookie);
  98
  99                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 100                         break;
 101         }
 102
 103         if (!tgt)
 104                 GOTO(out_lmv_lock, rc = -EINVAL);
 105
 106         obd = class_exp2obd(tgt->ltd_exp);
 107         if (obd == NULL)
 108                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 109
 110         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 111                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 112                obd->obd_type->typ_name, tgt->ltd_index);
 113         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 114
 115         if (tgt->ltd_active == activate) {
 116                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 117                        activate ? "" : "in");
 118                 GOTO(out_lmv_lock, rc);
 119         }
 120
 121         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 122                activate ? "" : "in");
 123         lmv_activate_target(lmv, tgt, activate);
 124         EXIT;
 125
 126  out_lmv_lock:
 127         spin_unlock(&lmv->lmv_lock);
 128         return rc;
 129 }
 130
 131 static struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
 132 {
 133         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
 134         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
 135
 136         return tgt ? obd_get_uuid(tgt->ltd_exp) : NULL;
 137 }
 138
 139 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 140                       enum obd_notify_event ev)
 141 {
 142         struct obd_connect_data *conn_data;
 143         struct lmv_obd          *lmv = &obd->u.lmv;
 144         struct obd_uuid         *uuid;
 145         int                      rc = 0;
 146         ENTRY;
 147
 148         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 149                 CERROR("unexpected notification of %s %s!\n",
 150                        watched->obd_type->typ_name,
 151                        watched->obd_name);
 152                 RETURN(-EINVAL);
 153         }
 154
 155         uuid = &watched->u.cli.cl_target_uuid;
 156         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 157                 /*
 158                  * Set MDC as active before notifying the observer, so the
 159                  * observer can use the MDC normally.
 160                  */
 161                 rc = lmv_set_mdc_active(lmv, uuid,
 162                                         ev == OBD_NOTIFY_ACTIVE);
 163                 if (rc) {
 164                         CERROR("%sactivation of %s failed: %d\n",
 165                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 166                                uuid->uuid, rc);
 167                         RETURN(rc);
 168                 }
 169         } else if (ev == OBD_NOTIFY_OCD) {
 170                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 171                 /*
 172                  * XXX: Make sure that ocd_connect_flags from all targets are
 173                  * the same. Otherwise one of MDTs runs wrong version or
 174                  * something like this.  --umka
 175                  */
 176                 obd->obd_self_export->exp_connect_data = *conn_data;
 177         }
 178
 179         /*
 180          * Pass the notification up the chain.
 181          */
 182         if (obd->obd_observer)
 183                 rc = obd_notify(obd->obd_observer, watched, ev);
 184
 185         RETURN(rc);
 186 }
 187
 188 static int lmv_connect(const struct lu_env *env,
 189                        struct obd_export **pexp, struct obd_device *obd,
 190                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 191                        void *localdata)
 192 {
 193         struct lmv_obd *lmv = &obd->u.lmv;
 194         struct lustre_handle conn = { 0 };
 195         struct obd_export *exp;
 196         int rc;
 197         ENTRY;
 198
 199         rc = class_connect(&conn, obd, cluuid);
 200         if (rc) {
 201                 CERROR("class_connection() returned %d\n", rc);
 202                 RETURN(rc);
 203         }
 204
 205         exp = class_conn2export(&conn);
 206
 207         lmv->connected = 0;
 208         lmv->conn_data = *data;
 209         lmv->lmv_cache = localdata;
 210
 211         lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
 212                                                     &obd->obd_kset.kobj);
 213         if (!lmv->lmv_tgts_kobj) {
 214                 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
 215                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
 216         }
 217
 218         rc = lmv_check_connect(obd);
 219         if (rc != 0)
 220                 GOTO(out_sysfs, rc);
 221
 222         *pexp = exp;
 223
 224         RETURN(rc);
 225
 226 out_sysfs:
 227         if (lmv->lmv_tgts_kobj)
 228                 kobject_put(lmv->lmv_tgts_kobj);
 229
 230         class_disconnect(exp);
 231
 232         return rc;
 233 }
 234
 235 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
 236                             __u32 def_easize)
 237 {
 238         struct obd_device *obd = exp->exp_obd;
 239         struct lmv_obd *lmv = &obd->u.lmv;
 240         struct lmv_tgt_desc *tgt;
 241         int change = 0;
 242         int rc = 0;
 243
 244         ENTRY;
 245
 246         if (lmv->max_easize < easize) {
 247                 lmv->max_easize = easize;
 248                 change = 1;
 249         }
 250         if (lmv->max_def_easize < def_easize) {
 251                 lmv->max_def_easize = def_easize;
 252                 change = 1;
 253         }
 254
 255         if (change == 0)
 256                 RETURN(0);
 257
 258         if (lmv->connected == 0)
 259                 RETURN(0);
 260
 261         lmv_foreach_connected_tgt(lmv, tgt) {
 262                 if (!tgt->ltd_active)
 263                         continue;
 264
 265                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
 266                 if (rc) {
 267                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
 268                                " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
 269                         break;
 270                 }
 271         }
 272         RETURN(rc);
 273 }
 274
 275 #define MAX_STRING_SIZE 128
 276
 277 static int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 278 {
 279         struct lmv_obd *lmv = &obd->u.lmv;
 280         struct obd_device *mdc_obd;
 281         struct obd_export *mdc_exp;
 282         struct lu_fld_target target;
 283         int  rc;
 284         ENTRY;
 285
 286         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 287                                         &obd->obd_uuid);
 288         if (!mdc_obd) {
 289                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 290                 RETURN(-EINVAL);
 291         }
 292
 293         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
 294                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 295                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
 296
 297         if (!mdc_obd->obd_set_up) {
 298                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 299                 RETURN(-EINVAL);
 300         }
 301
 302         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
 303                          &lmv->conn_data, lmv->lmv_cache);
 304         if (rc) {
 305                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 306                 RETURN(rc);
 307         }
 308
 309         /*
 310          * Init fid sequence client for this mdc and add new fld target.
 311          */
 312         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
 313         if (rc)
 314                 RETURN(rc);
 315
 316         target.ft_srv = NULL;
 317         target.ft_exp = mdc_exp;
 318         target.ft_idx = tgt->ltd_index;
 319
 320         fld_client_add_target(&lmv->lmv_fld, &target);
 321
 322         rc = obd_register_observer(mdc_obd, obd);
 323         if (rc) {
 324                 obd_disconnect(mdc_exp);
 325                 CERROR("target %s register_observer error %d\n",
 326                        tgt->ltd_uuid.uuid, rc);
 327                 RETURN(rc);
 328         }
 329
 330         if (obd->obd_observer) {
 331                 /*
 332                  * Tell the observer about the new target.
 333                  */
 334                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 335                                 OBD_NOTIFY_ACTIVE);
 336                 if (rc) {
 337                         obd_disconnect(mdc_exp);
 338                         RETURN(rc);
 339                 }
 340         }
 341
 342         tgt->ltd_active = 1;
 343         tgt->ltd_exp = mdc_exp;
 344         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
 345
 346         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
 347
 348         rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
 349         if (rc) {
 350                 obd_disconnect(mdc_exp);
 351                 RETURN(rc);
 352         }
 353
 354         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 355                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 356                atomic_read(&obd->obd_refcount));
 357
 358         lmv_statfs_check_update(obd, tgt);
 359
 360         if (lmv->lmv_tgts_kobj)
 361                 /* Even if we failed to create the link, that's fine */
 362                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
 363                                        &mdc_obd->obd_kset.kobj,
 364                                        mdc_obd->obd_name);
 365         RETURN(0);
 366 }
 367
 368 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
 369 {
 370         LASSERT(tgt);
 371         ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
 372         OBD_FREE_PTR(tgt);
 373 }
 374
 375 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 376                            __u32 index, int gen)
 377 {
 378         struct obd_device *mdc_obd;
 379         struct lmv_obd *lmv = &obd->u.lmv;
 380         struct lmv_tgt_desc *tgt;
 381         struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
 382         int rc = 0;
 383
 384         ENTRY;
 385
 386         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
 387         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
 388                                         &obd->obd_uuid);
 389         if (!mdc_obd) {
 390                 CERROR("%s: Target %s not attached: rc = %d\n",
 391                        obd->obd_name, uuidp->uuid, -EINVAL);
 392                 RETURN(-EINVAL);
 393         }
 394
 395         OBD_ALLOC_PTR(tgt);
 396         if (!tgt)
 397                 RETURN(-ENOMEM);
 398
 399         mutex_init(&tgt->ltd_fid_mutex);
 400         tgt->ltd_index = index;
 401         tgt->ltd_uuid = *uuidp;
 402         tgt->ltd_active = 0;
 403
 404         mutex_lock(&ltd->ltd_mutex);
 405         rc = ltd_add_tgt(ltd, tgt);
 406         mutex_unlock(&ltd->ltd_mutex);
 407
 408         if (rc)
 409                 GOTO(out_tgt, rc);
 410
 411         if (!lmv->connected)
 412                 /* lmv_check_connect() will connect this target. */
 413                 RETURN(0);
 414
 415         rc = lmv_connect_mdc(obd, tgt);
 416         if (!rc) {
 417                 int easize = sizeof(struct lmv_stripe_md) +
 418                         lmv->lmv_mdt_count * sizeof(struct lu_fid);
 419
 420                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
 421         }
 422
 423         RETURN(rc);
 424
 425 out_tgt:
 426         OBD_FREE_PTR(tgt);
 427         return rc;
 428 }
 429
 430 static int lmv_check_connect(struct obd_device *obd)
 431 {
 432         struct lmv_obd *lmv = &obd->u.lmv;
 433         struct lmv_tgt_desc *tgt;
 434         int easize;
 435         int rc;
 436
 437         ENTRY;
 438
 439         if (lmv->connected)
 440                 RETURN(0);
 441
 442         mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
 443         if (lmv->connected)
 444                 GOTO(unlock, rc = 0);
 445
 446         if (!lmv->lmv_mdt_count) {
 447                 CERROR("%s: no targets configured: rc = -EINVAL\n",
 448                        obd->obd_name);
 449                 GOTO(unlock, rc = -EINVAL);
 450         }
 451
 452         if (!lmv_mdt0_inited(lmv)) {
 453                 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
 454                        obd->obd_name);
 455                 GOTO(unlock, rc = -EINVAL);
 456         }
 457
 458         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 459                obd->obd_uuid.uuid, obd->obd_name);
 460
 461         lmv_foreach_tgt(lmv, tgt) {
 462                 rc = lmv_connect_mdc(obd, tgt);
 463                 if (rc)
 464                         GOTO(out_disc, rc);
 465         }
 466
 467         lmv->connected = 1;
 468         easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
 469         lmv_init_ea_size(obd->obd_self_export, easize, 0);
 470         EXIT;
 471 unlock:
 472         mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
 473
 474         return rc;
 475
 476 out_disc:
 477         lmv_foreach_tgt(lmv, tgt) {
 478                 tgt->ltd_active = 0;
 479                 if (!tgt->ltd_exp)
 480                         continue;
 481
 482                 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
 483                 obd_disconnect(tgt->ltd_exp);
 484         }
 485
 486         goto unlock;
 487 }
 488
 489 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 490 {
 491         struct lmv_obd *lmv = &obd->u.lmv;
 492         struct obd_device *mdc_obd;
 493         int rc;
 494         ENTRY;
 495
 496         LASSERT(tgt != NULL);
 497         LASSERT(obd != NULL);
 498
 499         mdc_obd = class_exp2obd(tgt->ltd_exp);
 500
 501         if (mdc_obd) {
 502                 mdc_obd->obd_force = obd->obd_force;
 503                 mdc_obd->obd_fail = obd->obd_fail;
 504                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 505
 506                 if (lmv->lmv_tgts_kobj)
 507                         sysfs_remove_link(lmv->lmv_tgts_kobj,
 508                                           mdc_obd->obd_name);
 509         }
 510
 511         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
 512         if (rc)
 513                 CERROR("Can't finalize fids factory\n");
 514
 515         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 516                tgt->ltd_exp->exp_obd->obd_name,
 517                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 518
 519         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 520         rc = obd_disconnect(tgt->ltd_exp);
 521         if (rc) {
 522                 if (tgt->ltd_active) {
 523                         CERROR("Target %s disconnect error %d\n",
 524                                tgt->ltd_uuid.uuid, rc);
 525                 }
 526         }
 527
 528         lmv_activate_target(lmv, tgt, 0);
 529         tgt->ltd_exp = NULL;
 530         RETURN(0);
 531 }
 532
 533 static int lmv_disconnect(struct obd_export *exp)
 534 {
 535         struct obd_device *obd = class_exp2obd(exp);
 536         struct lmv_obd *lmv = &obd->u.lmv;
 537         struct lmv_tgt_desc *tgt;
 538         int rc;
 539
 540         ENTRY;
 541
 542         lmv_foreach_connected_tgt(lmv, tgt)
 543                 lmv_disconnect_mdc(obd, tgt);
 544
 545         if (lmv->lmv_tgts_kobj)
 546                 kobject_put(lmv->lmv_tgts_kobj);
 547
 548         if (!lmv->connected)
 549                 class_export_put(exp);
 550         rc = class_disconnect(exp);
 551         lmv->connected = 0;
 552
 553         RETURN(rc);
 554 }
 555
 556 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
 557                         void __user *uarg)
 558 {
 559         struct obd_device *obd = class_exp2obd(exp);
 560         struct lmv_obd *lmv = &obd->u.lmv;
 561         struct getinfo_fid2path *gf;
 562         struct lmv_tgt_desc *tgt;
 563         struct getinfo_fid2path *remote_gf = NULL;
 564         struct lu_fid root_fid;
 565         int remote_gf_size = 0;
 566         int rc;
 567
 568         gf = karg;
 569         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 570         if (IS_ERR(tgt))
 571                 RETURN(PTR_ERR(tgt));
 572
 573         root_fid = *gf->gf_u.gf_root_fid;
 574         LASSERT(fid_is_sane(&root_fid));
 575
 576 repeat_fid2path:
 577         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
 578         if (rc != 0 && rc != -EREMOTE)
 579                 GOTO(out_fid2path, rc);
 580
 581         /* If remote_gf != NULL, it means just building the
 582          * path on the remote MDT, copy this path segement to gf */
 583         if (remote_gf != NULL) {
 584                 struct getinfo_fid2path *ori_gf;
 585                 char *ptr;
 586                 int len;
 587
 588                 ori_gf = (struct getinfo_fid2path *)karg;
 589                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
 590                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
 591                         GOTO(out_fid2path, rc = -EOVERFLOW);
 592
 593                 ptr = ori_gf->gf_u.gf_path;
 594
 595                 len = strlen(gf->gf_u.gf_path);
 596                 /* move the current path to the right to release space
 597                  * for closer-to-root part */
 598                 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
 599                 memcpy(ptr, gf->gf_u.gf_path, len);
 600                 ptr[len] = '/';
 601         }
 602
 603         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
 604                tgt->ltd_exp->exp_obd->obd_name,
 605                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
 606                gf->gf_linkno);
 607
 608         if (rc == 0)
 609                 GOTO(out_fid2path, rc);
 610
 611         /* sigh, has to go to another MDT to do path building further */
 612         if (remote_gf == NULL) {
 613                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
 614                 OBD_ALLOC(remote_gf, remote_gf_size);
 615                 if (remote_gf == NULL)
 616                         GOTO(out_fid2path, rc = -ENOMEM);
 617                 remote_gf->gf_pathlen = PATH_MAX;
 618         }
 619
 620         if (!fid_is_sane(&gf->gf_fid)) {
 621                 CERROR("%s: invalid FID "DFID": rc = %d\n",
 622                        tgt->ltd_exp->exp_obd->obd_name,
 623                        PFID(&gf->gf_fid), -EINVAL);
 624                 GOTO(out_fid2path, rc = -EINVAL);
 625         }
 626
 627         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 628         if (IS_ERR(tgt))
 629                 GOTO(out_fid2path, rc = -EINVAL);
 630
 631         remote_gf->gf_fid = gf->gf_fid;
 632         remote_gf->gf_recno = -1;
 633         remote_gf->gf_linkno = -1;
 634         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
 635         *remote_gf->gf_u.gf_root_fid = root_fid;
 636         gf = remote_gf;
 637         goto repeat_fid2path;
 638
 639 out_fid2path:
 640         if (remote_gf != NULL)
 641                 OBD_FREE(remote_gf, remote_gf_size);
 642         RETURN(rc);
 643 }
 644
 645 static int lmv_hsm_req_count(struct lmv_obd *lmv,
 646                              const struct hsm_user_request *hur,
 647                              const struct lmv_tgt_desc *tgt_mds)
 648 {
 649         struct lmv_tgt_desc *curr_tgt;
 650         __u32 i;
 651         int nr = 0;
 652
 653         /* count how many requests must be sent to the given target */
 654         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
 655                 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
 656                 if (IS_ERR(curr_tgt))
 657                         RETURN(PTR_ERR(curr_tgt));
 658                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
 659                         nr++;
 660         }
 661         return nr;
 662 }
 663
 664 static int lmv_hsm_req_build(struct lmv_obd *lmv,
 665                               struct hsm_user_request *hur_in,
 666                               const struct lmv_tgt_desc *tgt_mds,
 667                               struct hsm_user_request *hur_out)
 668 {
 669         __u32 i, nr_out;
 670         struct lmv_tgt_desc *curr_tgt;
 671
 672         /* build the hsm_user_request for the given target */
 673         hur_out->hur_request = hur_in->hur_request;
 674         nr_out = 0;
 675         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
 676                 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
 677                 if (IS_ERR(curr_tgt))
 678                         RETURN(PTR_ERR(curr_tgt));
 679                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
 680                         hur_out->hur_user_item[nr_out] =
 681                                                 hur_in->hur_user_item[i];
 682                         nr_out++;
 683                 }
 684         }
 685         hur_out->hur_request.hr_itemcount = nr_out;
 686         memcpy(hur_data(hur_out), hur_data(hur_in),
 687                hur_in->hur_request.hr_data_len);
 688
 689         RETURN(0);
 690 }
 691
 692 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
 693                                  int len, struct lustre_kernelcomm *lk,
 694                                  void __user *uarg)
 695 {
 696         struct lmv_obd *lmv = &obd->u.lmv;
 697         struct lu_tgt_desc *tgt;
 698         int rc;
 699
 700         ENTRY;
 701
 702         /* unregister request (call from llapi_hsm_copytool_fini) */
 703         lmv_foreach_connected_tgt(lmv, tgt)
 704                 /* best effort: try to clean as much as possible
 705                  * (continue on error) */
 706                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 707
 708         /* Whatever the result, remove copytool from kuc groups.
 709          * Unreached coordinators will get EPIPE on next requests
 710          * and will unregister automatically.
 711          */
 712         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 713
 714         RETURN(rc);
 715 }
 716
 717 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
 718                                int len, struct lustre_kernelcomm *lk,
 719                                void __user *uarg)
 720 {
 721         struct lmv_obd *lmv = &obd->u.lmv;
 722         struct file *filp;
 723         bool any_set = false;
 724         struct kkuc_ct_data *kcd;
 725         size_t kcd_size;
 726         struct lu_tgt_desc *tgt;
 727         __u32 i;
 728         int err;
 729         int rc = 0;
 730
 731         ENTRY;
 732
 733         filp = fget(lk->lk_wfd);
 734         if (!filp)
 735                 RETURN(-EBADF);
 736
 737         if (lk->lk_flags & LK_FLG_DATANR)
 738                 kcd_size = offsetof(struct kkuc_ct_data,
 739                                     kcd_archives[lk->lk_data_count]);
 740         else
 741                 kcd_size = sizeof(*kcd);
 742
 743         OBD_ALLOC(kcd, kcd_size);
 744         if (kcd == NULL)
 745                 GOTO(err_fput, rc = -ENOMEM);
 746
 747         kcd->kcd_nr_archives = lk->lk_data_count;
 748         if (lk->lk_flags & LK_FLG_DATANR) {
 749                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
 750                 if (lk->lk_data_count > 0)
 751                         memcpy(kcd->kcd_archives, lk->lk_data,
 752                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
 753         } else {
 754                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
 755         }
 756
 757         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
 758                                    lk->lk_group, kcd, kcd_size);
 759         OBD_FREE(kcd, kcd_size);
 760         if (rc)
 761                 GOTO(err_fput, rc);
 762
 763         /* All or nothing: try to register to all MDS.
 764          * In case of failure, unregister from previous MDS,
 765          * except if it because of inactive target. */
 766         lmv_foreach_connected_tgt(lmv, tgt) {
 767                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 768                 if (err) {
 769                         if (tgt->ltd_active) {
 770                                 /* permanent error */
 771                                 CERROR("%s: iocontrol MDC %s on MDT"
 772                                        " idx %d cmd %x: err = %d\n",
 773                                        lmv2obd_dev(lmv)->obd_name,
 774                                        tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
 775                                        err);
 776                                 rc = err;
 777                                 lk->lk_flags |= LK_FLG_STOP;
 778                                 i = tgt->ltd_index;
 779                                 /* unregister from previous MDS */
 780                                 lmv_foreach_connected_tgt(lmv, tgt) {
 781                                         if (tgt->ltd_index >= i)
 782                                                 break;
 783
 784                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
 785                                                       lk, uarg);
 786                                 }
 787                                 GOTO(err_kkuc_rem, rc);
 788                         }
 789                         /* else: transient error.
 790                          * kuc will register to the missing MDT
 791                          * when it is back */
 792                 } else {
 793                         any_set = true;
 794                 }
 795         }
 796
 797         if (!any_set)
 798                 /* no registration done: return error */
 799                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
 800
 801         RETURN(0);
 802
 803 err_kkuc_rem:
 804         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 805
 806 err_fput:
 807         fput(filp);
 808         return rc;
 809 }
 810
 811 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 812                          int len, void *karg, void __user *uarg)
 813 {
 814         struct obd_device *obd = class_exp2obd(exp);
 815         struct lmv_obd *lmv = &obd->u.lmv;
 816         struct lu_tgt_desc *tgt = NULL;
 817         int set = 0;
 818         __u32 count = lmv->lmv_mdt_count;
 819         int rc = 0;
 820
 821         ENTRY;
 822
 823         if (count == 0)
 824                 RETURN(-ENOTTY);
 825
 826         switch (cmd) {
 827         case IOC_OBD_STATFS: {
 828                 struct obd_ioctl_data *data = karg;
 829                 struct obd_device *mdc_obd;
 830                 struct obd_statfs stat_buf = {0};
 831                 __u32 index;
 832
 833                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 834
 835                 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
 836                         RETURN(-ENODEV);
 837
 838                 tgt = lmv_tgt(lmv, index);
 839                 if (!tgt)
 840                         RETURN(-EAGAIN);
 841
 842                 if (!tgt->ltd_active)
 843                         RETURN(-ENODATA);
 844
 845                 mdc_obd = class_exp2obd(tgt->ltd_exp);
 846                 if (!mdc_obd)
 847                         RETURN(-EINVAL);
 848
 849                 /* copy UUID */
 850                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 851                                  min((int) data->ioc_plen2,
 852                                      (int) sizeof(struct obd_uuid))))
 853                         RETURN(-EFAULT);
 854
 855                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
 856                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
 857                                 0);
 858                 if (rc)
 859                         RETURN(rc);
 860                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
 861                                  min((int) data->ioc_plen1,
 862                                      (int) sizeof(stat_buf))))
 863                         RETURN(-EFAULT);
 864                 break;
 865         }
 866         case OBD_IOC_QUOTACTL: {
 867                 struct if_quotactl *qctl = karg;
 868                 struct obd_quotactl *oqctl;
 869
 870                 if (qctl->qc_valid == QC_MDTIDX) {
 871                         tgt = lmv_tgt(lmv, qctl->qc_idx);
 872                 } else if (qctl->qc_valid == QC_UUID) {
 873                         lmv_foreach_tgt(lmv, tgt) {
 874                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 875                                                      &qctl->obd_uuid))
 876                                         continue;
 877
 878                                 if (!tgt->ltd_exp)
 879                                         RETURN(-EINVAL);
 880
 881                                 break;
 882                         }
 883                 } else {
 884                         RETURN(-EINVAL);
 885                 }
 886
 887                 if (!tgt || !tgt->ltd_exp)
 888                         RETURN(-EINVAL);
 889
 890                 OBD_ALLOC_PTR(oqctl);
 891                 if (!oqctl)
 892                         RETURN(-ENOMEM);
 893
 894                 QCTL_COPY(oqctl, qctl);
 895                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 896                 if (rc == 0) {
 897                         QCTL_COPY(qctl, oqctl);
 898                         qctl->qc_valid = QC_MDTIDX;
 899                         qctl->obd_uuid = tgt->ltd_uuid;
 900                 }
 901                 OBD_FREE_PTR(oqctl);
 902                 break;
 903         }
 904         case LL_IOC_GET_CONNECT_FLAGS: {
 905                 tgt = lmv_tgt(lmv, 0);
 906                 rc = -ENODATA;
 907                 if (tgt && tgt->ltd_exp)
 908                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 909                 break;
 910         }
 911         case LL_IOC_FID2MDTIDX: {
 912                 struct lu_fid *fid = karg;
 913                 int             mdt_index;
 914
 915                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
 916                 if (rc != 0)
 917                         RETURN(rc);
 918
 919                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
 920                  * point to user space memory for FID2MDTIDX. */
 921                 *(__u32 *)uarg = mdt_index;
 922                 break;
 923         }
 924         case OBD_IOC_FID2PATH: {
 925                 rc = lmv_fid2path(exp, len, karg, uarg);
 926                 break;
 927         }
 928         case LL_IOC_HSM_STATE_GET:
 929         case LL_IOC_HSM_STATE_SET:
 930         case LL_IOC_HSM_ACTION: {
 931                 struct md_op_data *op_data = karg;
 932
 933                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
 934                 if (IS_ERR(tgt))
 935                         RETURN(PTR_ERR(tgt));
 936
 937                 if (tgt->ltd_exp == NULL)
 938                         RETURN(-EINVAL);
 939
 940                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 941                 break;
 942         }
 943         case LL_IOC_HSM_PROGRESS: {
 944                 const struct hsm_progress_kernel *hpk = karg;
 945
 946                 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
 947                 if (IS_ERR(tgt))
 948                         RETURN(PTR_ERR(tgt));
 949                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 950                 break;
 951         }
 952         case LL_IOC_HSM_REQUEST: {
 953                 struct hsm_user_request *hur = karg;
 954                 unsigned int reqcount = hur->hur_request.hr_itemcount;
 955
 956                 if (reqcount == 0)
 957                         RETURN(0);
 958
 959                 /* if the request is about a single fid
 960                  * or if there is a single MDS, no need to split
 961                  * the request. */
 962                 if (reqcount == 1 || count == 1) {
 963                         tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
 964                         if (IS_ERR(tgt))
 965                                 RETURN(PTR_ERR(tgt));
 966                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 967                 } else {
 968                         /* split fid list to their respective MDS */
 969                         lmv_foreach_connected_tgt(lmv, tgt) {
 970                                 int nr, rc1;
 971                                 size_t reqlen;
 972                                 struct hsm_user_request *req;
 973
 974                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
 975                                 if (nr < 0)
 976                                         RETURN(nr);
 977                                 if (nr == 0) /* nothing for this MDS */
 978                                         continue;
 979
 980                                 /* build a request with fids for this MDS */
 981                                 reqlen = offsetof(typeof(*hur),
 982                                                   hur_user_item[nr])
 983                                                 + hur->hur_request.hr_data_len;
 984                                 OBD_ALLOC_LARGE(req, reqlen);
 985                                 if (req == NULL)
 986                                         RETURN(-ENOMEM);
 987                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
 988                                 if (rc1 < 0)
 989                                         GOTO(hsm_req_err, rc1);
 990                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
 991                                                     req, uarg);
 992 hsm_req_err:
 993                                 if (rc1 != 0 && rc == 0)
 994                                         rc = rc1;
 995                                 OBD_FREE_LARGE(req, reqlen);
 996                         }
 997                 }
 998                 break;
 999         }
1000         case LL_IOC_LOV_SWAP_LAYOUTS: {
1001                 struct md_op_data *op_data = karg;
1002                 struct lmv_tgt_desc *tgt1, *tgt2;
1003
1004                 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1005                 if (IS_ERR(tgt1))
1006                         RETURN(PTR_ERR(tgt1));
1007
1008                 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1009                 if (IS_ERR(tgt2))
1010                         RETURN(PTR_ERR(tgt2));
1011
1012                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1013                         RETURN(-EINVAL);
1014
1015                 /* only files on same MDT can have their layouts swapped */
1016                 if (tgt1->ltd_index != tgt2->ltd_index)
1017                         RETURN(-EPERM);
1018
1019                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1020                 break;
1021         }
1022         case LL_IOC_HSM_CT_START: {
1023                 struct lustre_kernelcomm *lk = karg;
1024                 if (lk->lk_flags & LK_FLG_STOP)
1025                         rc = lmv_hsm_ct_unregister(obd, cmd, len, lk, uarg);
1026                 else
1027                         rc = lmv_hsm_ct_register(obd, cmd, len, lk, uarg);
1028                 break;
1029         }
1030         default:
1031                 lmv_foreach_connected_tgt(lmv, tgt) {
1032                         struct obd_device *mdc_obd;
1033                         int err;
1034
1035                         /* ll_umount_begin() sets force flag but for lmv, not
1036                          * mdc. Let's pass it through */
1037                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1038                         mdc_obd->obd_force = obd->obd_force;
1039                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1040                         if (err) {
1041                                 if (tgt->ltd_active) {
1042                                         CERROR("error: iocontrol MDC %s on MDT"
1043                                                " idx %d cmd %x: err = %d\n",
1044                                                tgt->ltd_uuid.uuid,
1045                                                tgt->ltd_index, cmd, err);
1046                                         if (!rc)
1047                                                 rc = err;
1048                                 }
1049                         } else
1050                                 set = 1;
1051                 }
1052                 if (!set && !rc)
1053                         rc = -EIO;
1054         }
1055         RETURN(rc);
1056 }
1057
1058 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1059                   struct lu_fid *fid, struct md_op_data *op_data)
1060 {
1061         struct obd_device *obd = class_exp2obd(exp);
1062         struct lmv_obd *lmv = &obd->u.lmv;
1063         struct lmv_tgt_desc *tgt;
1064         int rc;
1065
1066         ENTRY;
1067
1068         LASSERT(op_data);
1069         LASSERT(fid);
1070
1071         tgt = lmv_tgt(lmv, op_data->op_mds);
1072         if (!tgt)
1073                 RETURN(-ENODEV);
1074
1075         if (!tgt->ltd_active || !tgt->ltd_exp)
1076                 RETURN(-ENODEV);
1077
1078         /*
1079          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1080          * on server that seq in new allocated fid is not yet known.
1081          */
1082         mutex_lock(&tgt->ltd_fid_mutex);
1083         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1084         mutex_unlock(&tgt->ltd_fid_mutex);
1085         if (rc > 0) {
1086                 LASSERT(fid_is_sane(fid));
1087                 rc = 0;
1088         }
1089
1090         RETURN(rc);
1091 }
1092
1093 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1094 {
1095         struct lmv_obd *lmv = &obd->u.lmv;
1096         struct lmv_desc *desc;
1097         struct lnet_process_id lnet_id;
1098         int i = 0;
1099         int rc;
1100
1101         ENTRY;
1102
1103         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1104                 CERROR("LMV setup requires a descriptor\n");
1105                 RETURN(-EINVAL);
1106         }
1107
1108         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1109         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1110                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1111                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1112                 RETURN(-EINVAL);
1113         }
1114
1115         obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1116                      desc->ld_uuid.uuid);
1117         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1118         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1119         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1120                 LMV_DESC_QOS_MAXAGE_DEFAULT;
1121         lmv->max_def_easize = 0;
1122         lmv->max_easize = 0;
1123
1124         spin_lock_init(&lmv->lmv_lock);
1125
1126         /*
1127          * initialize rr_index to lower 32bit of netid, so that client
1128          * can distribute subdirs evenly from the beginning.
1129          */
1130         while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1131                 if (lnet_id.nid != LNET_NID_LO_0) {
1132                         lmv->lmv_qos_rr_index = (u32)lnet_id.nid;
1133                         break;
1134                 }
1135         }
1136
1137         rc = lmv_tunables_init(obd);
1138         if (rc)
1139                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1140                       obd->obd_name, rc);
1141
1142         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1143                              LUSTRE_CLI_FLD_HASH_DHT);
1144         if (rc)
1145                 CERROR("Can't init FLD, err %d\n", rc);
1146
1147         rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1148         if (rc)
1149                 CWARN("%s: error initialize target table: rc = %d\n",
1150                       obd->obd_name, rc);
1151
1152         RETURN(rc);
1153 }
1154
1155 static int lmv_cleanup(struct obd_device *obd)
1156 {
1157         struct lmv_obd *lmv = &obd->u.lmv;
1158         struct lu_tgt_desc *tgt;
1159         struct lu_tgt_desc *tmp;
1160
1161         ENTRY;
1162
1163         fld_client_fini(&lmv->lmv_fld);
1164         lmv_foreach_tgt_safe(lmv, tgt, tmp)
1165                 lmv_del_target(lmv, tgt);
1166         lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1167
1168         RETURN(0);
1169 }
1170
1171 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1172 {
1173         struct lustre_cfg       *lcfg = buf;
1174         struct obd_uuid         obd_uuid;
1175         int                     gen;
1176         __u32                   index;
1177         int                     rc;
1178         ENTRY;
1179
1180         switch (lcfg->lcfg_command) {
1181         case LCFG_ADD_MDC:
1182                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1183                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1184                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1185                         GOTO(out, rc = -EINVAL);
1186
1187                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1188
1189                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1190                         GOTO(out, rc = -EINVAL);
1191                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1192                         GOTO(out, rc = -EINVAL);
1193                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1194                 GOTO(out, rc);
1195         default:
1196                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1197                 GOTO(out, rc = -EINVAL);
1198         }
1199 out:
1200         RETURN(rc);
1201 }
1202
1203 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1204 {
1205         int i;
1206
1207         if (flags & OBD_STATFS_FOR_MDT0)
1208                 return 0;
1209
1210         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1211                 return lmv->lmv_statfs_start;
1212
1213         /* choose initial MDT for this client */
1214         for (i = 0;; i++) {
1215                 struct lnet_process_id lnet_id;
1216                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1217                         break;
1218
1219                 if (lnet_id.nid != LNET_NID_LO_0) {
1220                         /* We dont need a full 64-bit modulus, just enough
1221                          * to distribute the requests across MDTs evenly.
1222                          */
1223                         lmv->lmv_statfs_start = (u32)lnet_id.nid %
1224                                                 lmv->lmv_mdt_count;
1225                         break;
1226                 }
1227         }
1228
1229         return lmv->lmv_statfs_start;
1230 }
1231
1232 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1233                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1234 {
1235         struct obd_device *obd = class_exp2obd(exp);
1236         struct lmv_obd *lmv = &obd->u.lmv;
1237         struct obd_statfs *temp;
1238         struct lu_tgt_desc *tgt;
1239         __u32 i;
1240         __u32 idx;
1241         int rc = 0;
1242
1243         ENTRY;
1244
1245         OBD_ALLOC(temp, sizeof(*temp));
1246         if (temp == NULL)
1247                 RETURN(-ENOMEM);
1248
1249         /* distribute statfs among MDTs */
1250         idx = lmv_select_statfs_mdt(lmv, flags);
1251
1252         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1253                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1254                 tgt = lmv_tgt(lmv, idx);
1255                 if (!tgt || !tgt->ltd_exp)
1256                         continue;
1257
1258                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age,
1259                                 flags | OBD_STATFS_NESTED);
1260                 if (rc) {
1261                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1262                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1263                         GOTO(out_free_temp, rc);
1264                 }
1265
1266                 if (temp->os_state & OS_STATFS_SUM ||
1267                     flags == OBD_STATFS_FOR_MDT0) {
1268                         /* reset to the last aggregated values
1269                          * and don't sum with non-aggrated data */
1270                         /* If the statfs is from mount, it needs to retrieve
1271                          * necessary information from MDT0. i.e. mount does
1272                          * not need the merged osfs from all of MDT. Also
1273                          * clients can be mounted as long as MDT0 is in
1274                          * service */
1275                         *osfs = *temp;
1276                         break;
1277                 }
1278
1279                 if (i == 0) {
1280                         *osfs = *temp;
1281                 } else {
1282                         osfs->os_bavail += temp->os_bavail;
1283                         osfs->os_blocks += temp->os_blocks;
1284                         osfs->os_ffree += temp->os_ffree;
1285                         osfs->os_files += temp->os_files;
1286                         osfs->os_granted += temp->os_granted;
1287                 }
1288         }
1289
1290         EXIT;
1291 out_free_temp:
1292         OBD_FREE(temp, sizeof(*temp));
1293         return rc;
1294 }
1295
1296 static int lmv_statfs_update(void *cookie, int rc)
1297 {
1298         struct obd_info *oinfo = cookie;
1299         struct obd_device *obd = oinfo->oi_obd;
1300         struct lmv_obd *lmv = &obd->u.lmv;
1301         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1302         struct obd_statfs *osfs = oinfo->oi_osfs;
1303
1304         /*
1305          * NB: don't deactivate TGT upon error, because we may not trigger async
1306          * statfs any longer, then there is no chance to activate TGT.
1307          */
1308         if (!rc) {
1309                 spin_lock(&lmv->lmv_lock);
1310                 tgt->ltd_statfs = *osfs;
1311                 tgt->ltd_statfs_age = ktime_get_seconds();
1312                 spin_unlock(&lmv->lmv_lock);
1313                 set_bit(LQ_DIRTY, &lmv->lmv_qos.lq_flags);
1314         }
1315
1316         return rc;
1317 }
1318
1319 /* update tgt statfs async if it's ld_qos_maxage old */
1320 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1321 {
1322         struct obd_info oinfo = {
1323                 .oi_obd = obd,
1324                 .oi_tgt = tgt,
1325                 .oi_cb_up = lmv_statfs_update,
1326         };
1327         int rc;
1328
1329         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1330             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1331                 return 0;
1332
1333         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1334
1335         return rc;
1336 }
1337
1338 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1339                         struct lu_fid *fid)
1340 {
1341         struct obd_device *obd = exp->exp_obd;
1342         struct lmv_obd *lmv = &obd->u.lmv;
1343         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1344         int rc;
1345
1346         ENTRY;
1347
1348         if (!tgt)
1349                 RETURN(-ENODEV);
1350
1351         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1352         RETURN(rc);
1353 }
1354
1355 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1356                         u64 obd_md_valid, const char *name, size_t buf_size,
1357                         struct ptlrpc_request **req)
1358 {
1359         struct obd_device *obd = exp->exp_obd;
1360         struct lmv_obd *lmv = &obd->u.lmv;
1361         struct lmv_tgt_desc *tgt;
1362         int rc;
1363
1364         ENTRY;
1365
1366         tgt = lmv_fid2tgt(lmv, fid);
1367         if (IS_ERR(tgt))
1368                 RETURN(PTR_ERR(tgt));
1369
1370         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1371
1372         RETURN(rc);
1373 }
1374
1375 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1376                         u64 obd_md_valid, const char *name,
1377                         const void *value, size_t value_size,
1378                         unsigned int xattr_flags, u32 suppgid,
1379                         struct ptlrpc_request **req)
1380 {
1381         struct obd_device *obd = exp->exp_obd;
1382         struct lmv_obd *lmv = &obd->u.lmv;
1383         struct lmv_tgt_desc *tgt;
1384         int rc;
1385
1386         ENTRY;
1387
1388         tgt = lmv_fid2tgt(lmv, fid);
1389         if (IS_ERR(tgt))
1390                 RETURN(PTR_ERR(tgt));
1391
1392         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1393                          value, value_size, xattr_flags, suppgid, req);
1394
1395         RETURN(rc);
1396 }
1397
1398 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1399                        struct ptlrpc_request **request)
1400 {
1401         struct obd_device *obd = exp->exp_obd;
1402         struct lmv_obd *lmv = &obd->u.lmv;
1403         struct lmv_tgt_desc *tgt;
1404         int rc;
1405
1406         ENTRY;
1407
1408         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1409         if (IS_ERR(tgt))
1410                 RETURN(PTR_ERR(tgt));
1411
1412         if (op_data->op_flags & MF_GET_MDT_IDX) {
1413                 op_data->op_mds = tgt->ltd_index;
1414                 RETURN(0);
1415         }
1416
1417         rc = md_getattr(tgt->ltd_exp, op_data, request);
1418
1419         RETURN(rc);
1420 }
1421
1422 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1423 {
1424         struct obd_device *obd = exp->exp_obd;
1425         struct lmv_obd *lmv = &obd->u.lmv;
1426         struct lu_tgt_desc *tgt;
1427
1428         ENTRY;
1429
1430         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1431
1432         /*
1433          * With DNE every object can have two locks in different namespaces:
1434          * lookup lock in space of MDT storing direntry and update/open lock in
1435          * space of MDT storing inode.
1436          */
1437         lmv_foreach_connected_tgt(lmv, tgt)
1438                 md_null_inode(tgt->ltd_exp, fid);
1439
1440         RETURN(0);
1441 }
1442
1443 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1444                      struct md_open_data *mod, struct ptlrpc_request **request)
1445 {
1446         struct obd_device *obd = exp->exp_obd;
1447         struct lmv_obd *lmv = &obd->u.lmv;
1448         struct lmv_tgt_desc *tgt;
1449         int rc;
1450
1451         ENTRY;
1452
1453         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1454         if (IS_ERR(tgt))
1455                 RETURN(PTR_ERR(tgt));
1456
1457         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1458         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1459         RETURN(rc);
1460 }
1461
1462 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 *mdt)
1463 {
1464         struct lu_tgt_desc *tgt, *cur = NULL;
1465         __u64 total_avail = 0;
1466         __u64 total_weight = 0;
1467         __u64 cur_weight = 0;
1468         int total_usable = 0;
1469         __u64 rand;
1470         int rc;
1471
1472         ENTRY;
1473
1474         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1475                 RETURN(ERR_PTR(-EAGAIN));
1476
1477         down_write(&lmv->lmv_qos.lq_rw_sem);
1478
1479         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1480                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1481
1482         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1483         if (rc)
1484                 GOTO(unlock, tgt = ERR_PTR(rc));
1485
1486         lmv_foreach_tgt(lmv, tgt) {
1487                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1488                         tgt->ltd_qos.ltq_usable = 0;
1489                         continue;
1490                 }
1491
1492                 tgt->ltd_qos.ltq_usable = 1;
1493                 lu_tgt_qos_weight_calc(tgt);
1494                 if (tgt->ltd_index == *mdt)
1495                         cur = tgt;
1496                 total_avail += tgt->ltd_qos.ltq_avail;
1497                 total_weight += tgt->ltd_qos.ltq_weight;
1498                 total_usable++;
1499         }
1500
1501         /* if current MDT has above-average space, within range of the QOS
1502          * threshold, stay on the same MDT to avoid creating needless remote
1503          * MDT directories.
1504          */
1505         rand = total_avail * (256 - lmv->lmv_qos.lq_threshold_rr) /
1506                 (total_usable * 256);
1507         if (cur && cur->ltd_qos.ltq_avail >= rand) {
1508                 tgt = cur;
1509                 GOTO(unlock, rc = 0);
1510         }
1511
1512         rand = lu_prandom_u64_max(total_weight);
1513
1514         lmv_foreach_connected_tgt(lmv, tgt) {
1515                 if (!tgt->ltd_qos.ltq_usable)
1516                         continue;
1517
1518                 cur_weight += tgt->ltd_qos.ltq_weight;
1519                 if (cur_weight < rand)
1520                         continue;
1521
1522                 *mdt = tgt->ltd_index;
1523                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1524                 GOTO(unlock, rc = 0);
1525         }
1526
1527         /* no proper target found */
1528         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1529 unlock:
1530         up_write(&lmv->lmv_qos.lq_rw_sem);
1531
1532         return tgt;
1533 }
1534
1535 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv, __u32 *mdt)
1536 {
1537         struct lu_tgt_desc *tgt;
1538         int i;
1539         int index;
1540
1541         ENTRY;
1542
1543         spin_lock(&lmv->lmv_lock);
1544         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1545                 index = (i + lmv->lmv_qos_rr_index) %
1546                         lmv->lmv_mdt_descs.ltd_tgts_size;
1547                 tgt = lmv_tgt(lmv, index);
1548                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1549                         continue;
1550
1551                 *mdt = tgt->ltd_index;
1552                 lmv->lmv_qos_rr_index = (*mdt + 1) %
1553                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1554                 spin_unlock(&lmv->lmv_lock);
1555
1556                 RETURN(tgt);
1557         }
1558         spin_unlock(&lmv->lmv_lock);
1559
1560         RETURN(ERR_PTR(-ENODEV));
1561 }
1562
1563 /* locate MDT by file name, for striped directory, the file name hash decides
1564  * which stripe its dirent is stored.
1565  */
1566 static struct lmv_tgt_desc *
1567 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1568                        const char *name, int namelen, struct lu_fid *fid,
1569                        __u32 *mds, bool new_layout)
1570 {
1571         struct lmv_tgt_desc *tgt;
1572         const struct lmv_oinfo *oinfo;
1573
1574         if (!lmv_dir_striped(lsm) || !namelen) {
1575                 tgt = lmv_fid2tgt(lmv, fid);
1576                 if (IS_ERR(tgt))
1577                         return tgt;
1578
1579                 *mds = tgt->ltd_index;
1580                 return tgt;
1581         }
1582
1583         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1584                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1585                         return ERR_PTR(-EBADF);
1586                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1587         } else {
1588                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, new_layout);
1589                 if (IS_ERR(oinfo))
1590                         return ERR_CAST(oinfo);
1591         }
1592
1593         /* check stripe FID is sane */
1594         if (!fid_is_sane(&oinfo->lmo_fid))
1595                 return ERR_PTR(-ENODEV);
1596
1597         *fid = oinfo->lmo_fid;
1598         *mds = oinfo->lmo_mds;
1599         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1600
1601         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1602
1603         return tgt ? tgt : ERR_PTR(-ENODEV);
1604 }
1605
1606 /**
1607  * Locate MDT of op_data->op_fid1
1608  *
1609  * For striped directory, it will locate the stripe by name hash, if hash_type
1610  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1611  * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1612  * indicates whether old or new layout is used to locate.
1613  *
1614  * For plain direcotry, it just locate the MDT of op_data->op_fid1.
1615  *
1616  * \param[in] lmv               LMV device
1617  * \param[in/out] op_data       client MD stack parameters, name, namelen etc,
1618  *                              op_mds and op_fid1 will be updated if op_mea1
1619  *                              indicates fid1 represents a striped directory.
1620  *
1621  * retval               pointer to the lmv_tgt_desc if succeed.
1622  *                      ERR_PTR(errno) if failed.
1623  */
1624 struct lmv_tgt_desc *
1625 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1626 {
1627         struct lmv_stripe_md *lsm = op_data->op_mea1;
1628         struct lmv_oinfo *oinfo;
1629         struct lmv_tgt_desc *tgt;
1630
1631         if (lmv_dir_foreign(lsm))
1632                 return ERR_PTR(-ENODATA);
1633
1634         /* During creating VOLATILE file, it should honor the mdt
1635          * index if the file under striped dir is being restored, see
1636          * ct_restore(). */
1637         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1638             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1639                 tgt = lmv_tgt(lmv, op_data->op_mds);
1640                 if (!tgt)
1641                         return ERR_PTR(-ENODEV);
1642
1643                 if (lmv_dir_striped(lsm)) {
1644                         int i;
1645
1646                         /* refill the right parent fid */
1647                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1648                                 oinfo = &lsm->lsm_md_oinfo[i];
1649                                 if (oinfo->lmo_mds == op_data->op_mds) {
1650                                         op_data->op_fid1 = oinfo->lmo_fid;
1651                                         break;
1652                                 }
1653                         }
1654
1655                         if (i == lsm->lsm_md_stripe_count)
1656                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1657                 }
1658         } else if (lmv_dir_bad_hash(lsm)) {
1659                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1660                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1661
1662                 op_data->op_fid1 = oinfo->lmo_fid;
1663                 op_data->op_mds = oinfo->lmo_mds;
1664                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1665                 if (!tgt)
1666                         return ERR_PTR(-ENODEV);
1667         } else {
1668                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1669                                 op_data->op_name, op_data->op_namelen,
1670                                 &op_data->op_fid1, &op_data->op_mds,
1671                                 op_data->op_new_layout);
1672         }
1673
1674         return tgt;
1675 }
1676
1677 /* Locate MDT of op_data->op_fid2 for link/rename */
1678 static struct lmv_tgt_desc *
1679 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1680 {
1681         struct lmv_tgt_desc *tgt;
1682         int rc;
1683
1684         LASSERT(op_data->op_name);
1685         if (lmv_dir_layout_changing(op_data->op_mea2)) {
1686                 struct lu_fid fid1 = op_data->op_fid1;
1687                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1688                 struct ptlrpc_request *request = NULL;
1689
1690                 /*
1691                  * avoid creating new file under old layout of migrating
1692                  * directory, check it here.
1693                  */
1694                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1695                                 op_data->op_name, op_data->op_namelen,
1696                                 &op_data->op_fid2, &op_data->op_mds, false);
1697                 if (IS_ERR(tgt))
1698                         RETURN(tgt);
1699
1700                 op_data->op_fid1 = op_data->op_fid2;
1701                 op_data->op_mea1 = op_data->op_mea2;
1702                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1703                 op_data->op_fid1 = fid1;
1704                 op_data->op_mea1 = lsm1;
1705                 if (!rc) {
1706                         ptlrpc_req_finished(request);
1707                         RETURN(ERR_PTR(-EEXIST));
1708                 }
1709
1710                 if (rc != -ENOENT)
1711                         RETURN(ERR_PTR(rc));
1712         }
1713
1714         return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1715                                 op_data->op_namelen, &op_data->op_fid2,
1716                                 &op_data->op_mds, true);
1717 }
1718
1719 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1720 {
1721         struct lu_tgt_desc *tgt;
1722         struct ptlrpc_request *request;
1723         int rc;
1724
1725         LASSERT(lmv_dir_layout_changing(op_data->op_mea1));
1726         LASSERT(!op_data->op_new_layout);
1727
1728         tgt = lmv_locate_tgt(lmv, op_data);
1729         if (IS_ERR(tgt))
1730                 return PTR_ERR(tgt);
1731
1732         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1733         if (!rc) {
1734                 ptlrpc_req_finished(request);
1735                 return -EEXIST;
1736         }
1737
1738         return rc;
1739 }
1740
1741 static inline bool lmv_op_user_qos_mkdir(const struct md_op_data *op_data)
1742 {
1743         const struct lmv_user_md *lum = op_data->op_data;
1744
1745         return (op_data->op_cli_flags & CLI_SET_MEA) && lum &&
1746                le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC &&
1747                le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
1748 }
1749
1750 static inline bool lmv_op_default_qos_mkdir(const struct md_op_data *op_data)
1751 {
1752         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1753
1754         return lsm && lsm->lsm_md_master_mdt_index == LMV_OFFSET_DEFAULT;
1755 }
1756
1757 /* mkdir by QoS in two cases:
1758  * 1. 'lfs mkdir -i -1'
1759  * 2. parent default LMV master_mdt_index is -1
1760  *
1761  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1762  * directories under striped directory.
1763  */
1764 static inline bool lmv_op_qos_mkdir(const struct md_op_data *op_data)
1765 {
1766         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1767                 return false;
1768
1769         if (lmv_dir_striped(op_data->op_mea1))
1770                 return false;
1771
1772         if (lmv_op_user_qos_mkdir(op_data))
1773                 return true;
1774
1775         if (lmv_op_default_qos_mkdir(op_data))
1776                 return true;
1777
1778         return false;
1779 }
1780
1781 /* if default LMV is set, and its index is LMV_OFFSET_DEFAULT, and
1782  * 1. max_inherit_rr is set and is not LMV_INHERIT_RR_NONE
1783  * 2. or parent is ROOT
1784  * mkdir roundrobin.
1785  * NB, this also needs to check server is balanced, which is checked by caller.
1786  */
1787 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data)
1788 {
1789         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1790
1791         if (!lmv_op_default_qos_mkdir(op_data))
1792                 return false;
1793
1794         return lsm->lsm_md_max_inherit_rr != LMV_INHERIT_RR_NONE ||
1795                fid_is_root(&op_data->op_fid1);
1796 }
1797
1798 /* 'lfs mkdir -i <specific_MDT>' */
1799 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1800 {
1801         const struct lmv_user_md *lum = op_data->op_data;
1802
1803         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1804                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1805                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1806                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1807                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1808 }
1809
1810 /* parent default LMV master_mdt_index is not -1. */
1811 static inline bool
1812 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1813 {
1814         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1815                op_data->op_default_mea1 &&
1816                op_data->op_default_mea1->lsm_md_master_mdt_index !=
1817                         LMV_OFFSET_DEFAULT;
1818 }
1819
1820 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1821                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1822                 gid_t gid, cfs_cap_t cap_effective, __u64 rdev,
1823                 struct ptlrpc_request **request)
1824 {
1825         struct obd_device *obd = exp->exp_obd;
1826         struct lmv_obd *lmv = &obd->u.lmv;
1827         struct lmv_tgt_desc *tgt;
1828         struct mdt_body *repbody;
1829         int rc;
1830
1831         ENTRY;
1832
1833         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1834                 RETURN(-EIO);
1835
1836         if (lmv_dir_bad_hash(op_data->op_mea1))
1837                 RETURN(-EBADF);
1838
1839         if (lmv_dir_layout_changing(op_data->op_mea1)) {
1840                 /*
1841                  * if parent is migrating, create() needs to lookup existing
1842                  * name in both old and new layout, check old layout on client.
1843                  */
1844                 rc = lmv_old_layout_lookup(lmv, op_data);
1845                 if (rc != -ENOENT)
1846                         RETURN(rc);
1847
1848                 op_data->op_new_layout = true;
1849         }
1850
1851         tgt = lmv_locate_tgt(lmv, op_data);
1852         if (IS_ERR(tgt))
1853                 RETURN(PTR_ERR(tgt));
1854
1855         if (lmv_op_user_specific_mkdir(op_data)) {
1856                 struct lmv_user_md *lum = op_data->op_data;
1857
1858                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
1859                 tgt = lmv_tgt(lmv, op_data->op_mds);
1860                 if (!tgt)
1861                         RETURN(-ENODEV);
1862         } else if (lmv_op_default_specific_mkdir(op_data)) {
1863                 op_data->op_mds =
1864                         op_data->op_default_mea1->lsm_md_master_mdt_index;
1865                 tgt = lmv_tgt(lmv, op_data->op_mds);
1866                 if (!tgt)
1867                         RETURN(-ENODEV);
1868         } else if (lmv_op_qos_mkdir(op_data)) {
1869                 struct lmv_tgt_desc *tmp = tgt;
1870
1871                 tgt = lmv_locate_tgt_qos(lmv, &op_data->op_mds);
1872                 if (tgt == ERR_PTR(-EAGAIN)) {
1873                         if (ltd_qos_is_balanced(&lmv->lmv_mdt_descs) &&
1874                             !lmv_op_default_rr_mkdir(op_data) &&
1875                             !lmv_op_user_qos_mkdir(op_data))
1876                                 /* if it's not necessary, don't create remote
1877                                  * directory.
1878                                  */
1879                                 tgt = tmp;
1880                         else
1881                                 tgt = lmv_locate_tgt_rr(lmv, &op_data->op_mds);
1882                 }
1883                 if (IS_ERR(tgt))
1884                         RETURN(PTR_ERR(tgt));
1885
1886                 /*
1887                  * only update statfs after QoS mkdir, this means the cached
1888                  * statfs may be stale, and current mkdir may not follow QoS
1889                  * accurately, but it's not serious, and avoids periodic statfs
1890                  * when client doesn't mkdir by QoS.
1891                  */
1892                 lmv_statfs_check_update(obd, tgt);
1893         }
1894
1895 retry:
1896         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1897         if (rc)
1898                 RETURN(rc);
1899
1900         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
1901                 (int)op_data->op_namelen, op_data->op_name,
1902                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
1903                 op_data->op_mds);
1904
1905         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1906         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1907                        cap_effective, rdev, request);
1908         if (rc == 0) {
1909                 if (*request == NULL)
1910                         RETURN(rc);
1911                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1912         }
1913
1914         /* dir restripe needs to send to MDT where dir is located */
1915         if (rc != -EREMOTE ||
1916             !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
1917                 RETURN(rc);
1918
1919         repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
1920         if (repbody == NULL)
1921                 RETURN(-EPROTO);
1922
1923         /* Not cross-ref case, just get out of here. */
1924         if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
1925                 RETURN(rc);
1926
1927         op_data->op_fid2 = repbody->mbo_fid1;
1928         ptlrpc_req_finished(*request);
1929         *request = NULL;
1930
1931         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
1932         if (IS_ERR(tgt))
1933                 RETURN(PTR_ERR(tgt));
1934
1935         op_data->op_mds = tgt->ltd_index;
1936         goto retry;
1937 }
1938
1939 static int
1940 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1941             const union ldlm_policy_data *policy, struct md_op_data *op_data,
1942             struct lustre_handle *lockh, __u64 extra_lock_flags)
1943 {
1944         struct obd_device *obd = exp->exp_obd;
1945         struct lmv_obd *lmv = &obd->u.lmv;
1946         struct lmv_tgt_desc *tgt;
1947         int rc;
1948
1949         ENTRY;
1950
1951         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
1952
1953         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1954         if (IS_ERR(tgt))
1955                 RETURN(PTR_ERR(tgt));
1956
1957         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
1958                PFID(&op_data->op_fid1), tgt->ltd_index);
1959
1960         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
1961                         extra_lock_flags);
1962
1963         RETURN(rc);
1964 }
1965
1966 int
1967 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1968                  struct ptlrpc_request **preq)
1969 {
1970         struct obd_device *obd = exp->exp_obd;
1971         struct lmv_obd *lmv = &obd->u.lmv;
1972         struct lmv_tgt_desc *tgt;
1973         struct mdt_body *body;
1974         int rc;
1975
1976         ENTRY;
1977
1978 retry:
1979         tgt = lmv_locate_tgt(lmv, op_data);
1980         if (IS_ERR(tgt))
1981                 RETURN(PTR_ERR(tgt));
1982
1983         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1984                 (int)op_data->op_namelen, op_data->op_name,
1985                 PFID(&op_data->op_fid1), tgt->ltd_index);
1986
1987         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
1988         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
1989                 ptlrpc_req_finished(*preq);
1990                 *preq = NULL;
1991                 goto retry;
1992         }
1993
1994         if (rc)
1995                 RETURN(rc);
1996
1997         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
1998         LASSERT(body != NULL);
1999
2000         if (body->mbo_valid & OBD_MD_MDS) {
2001                 op_data->op_fid1 = body->mbo_fid1;
2002                 op_data->op_valid |= OBD_MD_FLCROSSREF;
2003                 op_data->op_namelen = 0;
2004                 op_data->op_name = NULL;
2005
2006                 ptlrpc_req_finished(*preq);
2007                 *preq = NULL;
2008
2009                 goto retry;
2010         }
2011
2012         RETURN(rc);
2013 }
2014
2015 #define md_op_data_fid(op_data, fl)                     \
2016         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2017          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2018          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2019          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2020          NULL)
2021
2022 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2023                             struct md_op_data *op_data, __u32 op_tgt,
2024                             enum ldlm_mode mode, int bits, int flag)
2025 {
2026         struct lu_fid *fid = md_op_data_fid(op_data, flag);
2027         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
2028         union ldlm_policy_data policy = { { 0 } };
2029         int rc = 0;
2030         ENTRY;
2031
2032         if (!fid_is_sane(fid))
2033                 RETURN(0);
2034
2035         if (tgt == NULL) {
2036                 tgt = lmv_fid2tgt(lmv, fid);
2037                 if (IS_ERR(tgt))
2038                         RETURN(PTR_ERR(tgt));
2039         }
2040
2041         if (tgt->ltd_index != op_tgt) {
2042                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2043                 policy.l_inodebits.bits = bits;
2044                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2045                                       mode, LCF_ASYNC, NULL);
2046         } else {
2047                 CDEBUG(D_INODE,
2048                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2049                        op_tgt, PFID(fid));
2050                 op_data->op_flags |= flag;
2051                 rc = 0;
2052         }
2053
2054         RETURN(rc);
2055 }
2056
2057 /*
2058  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2059  * op_data->op_fid2
2060  */
2061 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2062                     struct ptlrpc_request **request)
2063 {
2064         struct obd_device       *obd = exp->exp_obd;
2065         struct lmv_obd          *lmv = &obd->u.lmv;
2066         struct lmv_tgt_desc     *tgt;
2067         int                      rc;
2068         ENTRY;
2069
2070         LASSERT(op_data->op_namelen != 0);
2071
2072         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2073                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2074                op_data->op_name, PFID(&op_data->op_fid1));
2075
2076         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2077         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2078         op_data->op_cap = cfs_curproc_cap_pack();
2079
2080         tgt = lmv_locate_tgt2(lmv, op_data);
2081         if (IS_ERR(tgt))
2082                 RETURN(PTR_ERR(tgt));
2083
2084         /*
2085          * Cancel UPDATE lock on child (fid1).
2086          */
2087         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2088         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2089                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2090         if (rc != 0)
2091                 RETURN(rc);
2092
2093         rc = md_link(tgt->ltd_exp, op_data, request);
2094
2095         RETURN(rc);
2096 }
2097
2098 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2099                         const char *name, size_t namelen,
2100                         struct ptlrpc_request **request)
2101 {
2102         struct obd_device *obd = exp->exp_obd;
2103         struct lmv_obd *lmv = &obd->u.lmv;
2104         struct lmv_stripe_md *lsm = op_data->op_mea1;
2105         struct lmv_tgt_desc *parent_tgt;
2106         struct lmv_tgt_desc *sp_tgt;
2107         struct lmv_tgt_desc *tp_tgt = NULL;
2108         struct lmv_tgt_desc *child_tgt;
2109         struct lmv_tgt_desc *tgt;
2110         struct lu_fid target_fid = { 0 };
2111         int rc;
2112
2113         ENTRY;
2114
2115         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2116
2117         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2118                PFID(&op_data->op_fid1), (int)namelen, name);
2119
2120         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2121         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2122         op_data->op_cap = cfs_curproc_cap_pack();
2123
2124         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2125         if (IS_ERR(parent_tgt))
2126                 RETURN(PTR_ERR(parent_tgt));
2127
2128         if (lmv_dir_striped(lsm)) {
2129                 const struct lmv_oinfo *oinfo;
2130
2131                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
2132                 if (IS_ERR(oinfo))
2133                         RETURN(PTR_ERR(oinfo));
2134
2135                 /* save source stripe FID in fid4 temporarily for ELC */
2136                 op_data->op_fid4 = oinfo->lmo_fid;
2137                 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2138                 if (!sp_tgt)
2139                         RETURN(-ENODEV);
2140
2141                 /*
2142                  * if parent is being migrated too, fill op_fid2 with target
2143                  * stripe fid, otherwise the target stripe is not created yet.
2144                  */
2145                 if (lmv_dir_layout_changing(lsm)) {
2146                         oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
2147                                                         true);
2148                         if (IS_ERR(oinfo))
2149                                 RETURN(PTR_ERR(oinfo));
2150
2151                         op_data->op_fid2 = oinfo->lmo_fid;
2152                         tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2153                         if (!tp_tgt)
2154                                 RETURN(-ENODEV);
2155                 }
2156         } else {
2157                 sp_tgt = parent_tgt;
2158         }
2159
2160         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2161         if (IS_ERR(child_tgt))
2162                 RETURN(PTR_ERR(child_tgt));
2163
2164         /* for directory, migrate to MDT specified by lum_stripe_offset;
2165          * otherwise migrate to the target stripe of parent, but parent
2166          * directory may have finished migration (normally current file too),
2167          * allocate FID on MDT lum_stripe_offset, and server will check
2168          * whether file was migrated already.
2169          */
2170         if (S_ISDIR(op_data->op_mode) || !tp_tgt) {
2171                 struct lmv_user_md *lum = op_data->op_data;
2172
2173                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2174         } else  {
2175                 op_data->op_mds = tp_tgt->ltd_index;
2176         }
2177         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2178         if (rc)
2179                 RETURN(rc);
2180
2181         /*
2182          * for directory, send migrate request to the MDT where the object will
2183          * be migrated to, because we can't create a striped directory remotely.
2184          *
2185          * otherwise, send to the MDT where source is located because regular
2186          * file may open lease.
2187          *
2188          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2189          * backward compatibility.
2190          */
2191         if (S_ISDIR(op_data->op_mode) &&
2192             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2193                 tgt = lmv_fid2tgt(lmv, &target_fid);
2194                 if (IS_ERR(tgt))
2195                         RETURN(PTR_ERR(tgt));
2196         } else {
2197                 tgt = child_tgt;
2198         }
2199
2200         /* cancel UPDATE lock of parent master object */
2201         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2202                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2203         if (rc)
2204                 RETURN(rc);
2205
2206         /* cancel UPDATE lock of source parent */
2207         if (sp_tgt != parent_tgt) {
2208                 /*
2209                  * migrate RPC packs master object FID, because we can only pack
2210                  * two FIDs in reint RPC, but MDS needs to know both source
2211                  * parent and target parent, and it will obtain them from master
2212                  * FID and LMV, the other FID in RPC is kept for target.
2213                  *
2214                  * since this FID is not passed to MDC, cancel it anyway.
2215                  */
2216                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2217                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2218                 if (rc)
2219                         RETURN(rc);
2220
2221                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2222         }
2223         op_data->op_fid4 = target_fid;
2224
2225         /* cancel UPDATE locks of target parent */
2226         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2227                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2228         if (rc)
2229                 RETURN(rc);
2230
2231         /* cancel LOOKUP lock of source if source is remote object */
2232         if (child_tgt != sp_tgt) {
2233                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2234                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2235                                       MF_MDC_CANCEL_FID3);
2236                 if (rc)
2237                         RETURN(rc);
2238         }
2239
2240         /* cancel ELC locks of source */
2241         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2242                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2243         if (rc)
2244                 RETURN(rc);
2245
2246         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2247
2248         RETURN(rc);
2249 }
2250
2251 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2252                       const char *old, size_t oldlen,
2253                       const char *new, size_t newlen,
2254                       struct ptlrpc_request **request)
2255 {
2256         struct obd_device *obd = exp->exp_obd;
2257         struct lmv_obd *lmv = &obd->u.lmv;
2258         struct lmv_tgt_desc *sp_tgt;
2259         struct lmv_tgt_desc *tp_tgt = NULL;
2260         struct lmv_tgt_desc *src_tgt = NULL;
2261         struct lmv_tgt_desc *tgt;
2262         struct mdt_body *body;
2263         int rc;
2264
2265         ENTRY;
2266
2267         LASSERT(oldlen != 0);
2268
2269         if (op_data->op_cli_flags & CLI_MIGRATE) {
2270                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2271                 RETURN(rc);
2272         }
2273
2274         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2275         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2276         op_data->op_cap = cfs_curproc_cap_pack();
2277
2278         op_data->op_name = new;
2279         op_data->op_namelen = newlen;
2280
2281         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2282         if (IS_ERR(tp_tgt))
2283                 RETURN(PTR_ERR(tp_tgt));
2284
2285         /* Since the target child might be destroyed, and it might become
2286          * orphan, and we can only check orphan on the local MDT right now, so
2287          * we send rename request to the MDT where target child is located. If
2288          * target child does not exist, then it will send the request to the
2289          * target parent */
2290         if (fid_is_sane(&op_data->op_fid4)) {
2291                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2292                 if (IS_ERR(tgt))
2293                         RETURN(PTR_ERR(tgt));
2294         } else {
2295                 tgt = tp_tgt;
2296         }
2297
2298         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2299
2300         /* cancel UPDATE locks of target parent */
2301         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2302                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2303         if (rc != 0)
2304                 RETURN(rc);
2305
2306         if (fid_is_sane(&op_data->op_fid4)) {
2307                 /* cancel LOOKUP lock of target on target parent */
2308                 if (tgt != tp_tgt) {
2309                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2310                                               tgt->ltd_index, LCK_EX,
2311                                               MDS_INODELOCK_LOOKUP,
2312                                               MF_MDC_CANCEL_FID4);
2313                         if (rc != 0)
2314                                 RETURN(rc);
2315                 }
2316         }
2317
2318         if (fid_is_sane(&op_data->op_fid3)) {
2319                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2320                 if (IS_ERR(src_tgt))
2321                         RETURN(PTR_ERR(src_tgt));
2322
2323                 /* cancel ELC locks of source */
2324                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2325                                       LCK_EX, MDS_INODELOCK_ELC,
2326                                       MF_MDC_CANCEL_FID3);
2327                 if (rc != 0)
2328                         RETURN(rc);
2329         }
2330
2331         op_data->op_name = old;
2332         op_data->op_namelen = oldlen;
2333 retry:
2334         sp_tgt = lmv_locate_tgt(lmv, op_data);
2335         if (IS_ERR(sp_tgt))
2336                 RETURN(PTR_ERR(sp_tgt));
2337
2338         /* cancel UPDATE locks of source parent */
2339         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2340                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2341         if (rc != 0)
2342                 RETURN(rc);
2343
2344         if (fid_is_sane(&op_data->op_fid3)) {
2345                 /* cancel LOOKUP lock of source on source parent */
2346                 if (src_tgt != sp_tgt) {
2347                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2348                                               tgt->ltd_index, LCK_EX,
2349                                               MDS_INODELOCK_LOOKUP,
2350                                               MF_MDC_CANCEL_FID3);
2351                         if (rc != 0)
2352                                 RETURN(rc);
2353                 }
2354         }
2355
2356 rename:
2357         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2358                 PFID(&op_data->op_fid1), (int)oldlen, old,
2359                 PFID(&op_data->op_fid2), (int)newlen, new);
2360
2361         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2362                         request);
2363         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2364                 ptlrpc_req_finished(*request);
2365                 *request = NULL;
2366                 goto retry;
2367         }
2368
2369         if (rc && rc != -EXDEV)
2370                 RETURN(rc);
2371
2372         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2373         if (body == NULL)
2374                 RETURN(-EPROTO);
2375
2376         /* Not cross-ref case, just get out of here. */
2377         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2378                 RETURN(rc);
2379
2380         op_data->op_fid4 = body->mbo_fid1;
2381
2382         ptlrpc_req_finished(*request);
2383         *request = NULL;
2384
2385         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2386         if (IS_ERR(tgt))
2387                 RETURN(PTR_ERR(tgt));
2388
2389         if (fid_is_sane(&op_data->op_fid4)) {
2390                 /* cancel LOOKUP lock of target on target parent */
2391                 if (tgt != tp_tgt) {
2392                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2393                                               tgt->ltd_index, LCK_EX,
2394                                               MDS_INODELOCK_LOOKUP,
2395                                               MF_MDC_CANCEL_FID4);
2396                         if (rc != 0)
2397                                 RETURN(rc);
2398                 }
2399         }
2400
2401         goto rename;
2402 }
2403
2404 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2405                        void *ea, size_t ealen, struct ptlrpc_request **request)
2406 {
2407         struct obd_device *obd = exp->exp_obd;
2408         struct lmv_obd *lmv = &obd->u.lmv;
2409         struct lmv_tgt_desc *tgt;
2410         int rc = 0;
2411
2412         ENTRY;
2413
2414         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2415                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2416                op_data->op_xvalid);
2417
2418         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2419         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2420         if (IS_ERR(tgt))
2421                 RETURN(PTR_ERR(tgt));
2422
2423         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2424
2425         RETURN(rc);
2426 }
2427
2428 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2429                      struct ptlrpc_request **request)
2430 {
2431         struct obd_device *obd = exp->exp_obd;
2432         struct lmv_obd *lmv = &obd->u.lmv;
2433         struct lmv_tgt_desc *tgt;
2434         int rc;
2435
2436         ENTRY;
2437
2438         tgt = lmv_fid2tgt(lmv, fid);
2439         if (IS_ERR(tgt))
2440                 RETURN(PTR_ERR(tgt));
2441
2442         rc = md_fsync(tgt->ltd_exp, fid, request);
2443         RETURN(rc);
2444 }
2445
2446 struct stripe_dirent {
2447         struct page             *sd_page;
2448         struct lu_dirpage       *sd_dp;
2449         struct lu_dirent        *sd_ent;
2450         bool                     sd_eof;
2451 };
2452
2453 struct lmv_dir_ctxt {
2454         struct lmv_obd          *ldc_lmv;
2455         struct md_op_data       *ldc_op_data;
2456         struct md_callback      *ldc_cb_op;
2457         __u64                    ldc_hash;
2458         int                      ldc_count;
2459         struct stripe_dirent     ldc_stripes[0];
2460 };
2461
2462 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2463 {
2464         if (stripe->sd_page) {
2465                 kunmap(stripe->sd_page);
2466                 put_page(stripe->sd_page);
2467                 stripe->sd_page = NULL;
2468                 stripe->sd_ent = NULL;
2469         }
2470 }
2471
2472 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2473 {
2474         int i;
2475
2476         for (i = 0; i < ctxt->ldc_count; i++)
2477                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2478 }
2479
2480 /* if @ent is dummy, or . .., get next */
2481 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2482                                            struct lu_dirent *ent,
2483                                            int stripe_index)
2484 {
2485         for (; ent; ent = lu_dirent_next(ent)) {
2486                 /* Skip dummy entry */
2487                 if (le16_to_cpu(ent->lde_namelen) == 0)
2488                         continue;
2489
2490                 /* skip . and .. for other stripes */
2491                 if (stripe_index &&
2492                     (strncmp(ent->lde_name, ".",
2493                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2494                      strncmp(ent->lde_name, "..",
2495                              le16_to_cpu(ent->lde_namelen)) == 0))
2496                         continue;
2497
2498                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2499                         break;
2500         }
2501
2502         return ent;
2503 }
2504
2505 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2506                                             struct stripe_dirent *stripe,
2507                                             int stripe_index)
2508 {
2509         struct md_op_data *op_data = ctxt->ldc_op_data;
2510         struct lmv_oinfo *oinfo;
2511         struct lu_fid fid = op_data->op_fid1;
2512         struct inode *inode = op_data->op_data;
2513         struct lmv_tgt_desc *tgt;
2514         struct lu_dirent *ent = stripe->sd_ent;
2515         __u64 hash = ctxt->ldc_hash;
2516         int rc = 0;
2517
2518         ENTRY;
2519
2520         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2521         LASSERT(!ent);
2522
2523         do {
2524                 if (stripe->sd_page) {
2525                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2526
2527                         /* @hash should be the last dirent hash */
2528                         LASSERTF(hash <= end,
2529                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2530                                  ctxt, stripe, hash, end);
2531                         /* unload last page */
2532                         stripe_dirent_unload(stripe);
2533                         /* eof */
2534                         if (end == MDS_DIR_END_OFF) {
2535                                 stripe->sd_eof = true;
2536                                 break;
2537                         }
2538                         hash = end;
2539                 }
2540
2541                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2542                 if (!oinfo->lmo_root) {
2543                         rc = -ENOENT;
2544                         break;
2545                 }
2546
2547                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2548                 if (!tgt) {
2549                         rc = -ENODEV;
2550                         break;
2551                 }
2552
2553                 /* op_data is shared by stripes, reset after use */
2554                 op_data->op_fid1 = oinfo->lmo_fid;
2555                 op_data->op_fid2 = oinfo->lmo_fid;
2556                 op_data->op_data = oinfo->lmo_root;
2557
2558                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2559                                   &stripe->sd_page);
2560
2561                 op_data->op_fid1 = fid;
2562                 op_data->op_fid2 = fid;
2563                 op_data->op_data = inode;
2564
2565                 if (rc)
2566                         break;
2567
2568                 stripe->sd_dp = page_address(stripe->sd_page);
2569                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2570                                         stripe_index);
2571                 /* in case a page filled with ., .. and dummy, read next */
2572         } while (!ent);
2573
2574         stripe->sd_ent = ent;
2575         if (rc) {
2576                 LASSERT(!ent);
2577                 /* treat error as eof, so dir can be partially accessed */
2578                 stripe->sd_eof = true;
2579                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2580                               "directory is partially accessed!\n",
2581                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2582                               rc);
2583         }
2584
2585         RETURN(ent);
2586 }
2587
2588 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2589 {
2590         struct obd_device *obd = exp->exp_obd;
2591         struct lmv_obd *lmv = &obd->u.lmv;
2592         struct lmv_tgt_desc *tgt;
2593         int rc;
2594
2595         ENTRY;
2596
2597         rc = lmv_check_connect(obd);
2598         if (rc != 0)
2599                 RETURN(rc);
2600
2601         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2602         if (IS_ERR(tgt))
2603                 RETURN(PTR_ERR(tgt));
2604
2605         data->op_flags |= MF_MDC_CANCEL_FID1;
2606         rc = md_file_resync(tgt->ltd_exp, data);
2607         RETURN(rc);
2608 }
2609
2610 /**
2611  * Get dirent with the closest hash for striped directory
2612  *
2613  * This function will search the dir entry, whose hash value is the
2614  * closest(>=) to hash from all of sub-stripes, and it is only being called
2615  * for striped directory.
2616  *
2617  * \param[in] ctxt              dir read context
2618  *
2619  * \retval                      dirent get the entry successfully
2620  *                              NULL does not get the entry, normally it means
2621  *                              it reaches the end of the directory, while read
2622  *                              stripe dirent error is ignored to allow partial
2623  *                              access.
2624  */
2625 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2626 {
2627         struct stripe_dirent *stripe;
2628         struct lu_dirent *ent = NULL;
2629         int i;
2630         int min = -1;
2631
2632         /* TODO: optimize with k-way merge sort */
2633         for (i = 0; i < ctxt->ldc_count; i++) {
2634                 stripe = &ctxt->ldc_stripes[i];
2635                 if (stripe->sd_eof)
2636                         continue;
2637
2638                 if (!stripe->sd_ent) {
2639                         stripe_dirent_load(ctxt, stripe, i);
2640                         if (!stripe->sd_ent) {
2641                                 LASSERT(stripe->sd_eof);
2642                                 continue;
2643                         }
2644                 }
2645
2646                 if (min == -1 ||
2647                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2648                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2649                         min = i;
2650                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2651                             ctxt->ldc_hash)
2652                                 break;
2653                 }
2654         }
2655
2656         if (min != -1) {
2657                 stripe = &ctxt->ldc_stripes[min];
2658                 ent = stripe->sd_ent;
2659                 /* pop found dirent */
2660                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2661                                                    min);
2662         }
2663
2664         return ent;
2665 }
2666
2667 /**
2668  * Build dir entry page for striped directory
2669  *
2670  * This function gets one entry by @offset from a striped directory. It will
2671  * read entries from all of stripes, and choose one closest to the required
2672  * offset(&offset). A few notes
2673  * 1. skip . and .. for non-zero stripes, because there can only have one .
2674  * and .. in a directory.
2675  * 2. op_data will be shared by all of stripes, instead of allocating new
2676  * one, so need to restore before reusing.
2677  *
2678  * \param[in] exp       obd export refer to LMV
2679  * \param[in] op_data   hold those MD parameters of read_entry
2680  * \param[in] cb_op     ldlm callback being used in enqueue in mdc_read_entry
2681  * \param[in] offset    starting hash offset
2682  * \param[out] ppage    the page holding the entry. Note: because the entry
2683  *                      will be accessed in upper layer, so we need hold the
2684  *                      page until the usages of entry is finished, see
2685  *                      ll_dir_entry_next.
2686  *
2687  * retval               =0 if get entry successfully
2688  *                      <0 cannot get entry
2689  */
2690 static int lmv_striped_read_page(struct obd_export *exp,
2691                                  struct md_op_data *op_data,
2692                                  struct md_callback *cb_op,
2693                                  __u64 offset, struct page **ppage)
2694 {
2695         struct page *page = NULL;
2696         struct lu_dirpage *dp;
2697         void *start;
2698         struct lu_dirent *ent;
2699         struct lu_dirent *last_ent;
2700         int stripe_count;
2701         struct lmv_dir_ctxt *ctxt;
2702         struct lu_dirent *next = NULL;
2703         __u16 ent_size;
2704         size_t left_bytes;
2705         int rc = 0;
2706         ENTRY;
2707
2708         /* Allocate a page and read entries from all of stripes and fill
2709          * the page by hash order */
2710         page = alloc_page(GFP_KERNEL);
2711         if (!page)
2712                 RETURN(-ENOMEM);
2713
2714         /* Initialize the entry page */
2715         dp = kmap(page);
2716         memset(dp, 0, sizeof(*dp));
2717         dp->ldp_hash_start = cpu_to_le64(offset);
2718
2719         start = dp + 1;
2720         left_bytes = PAGE_SIZE - sizeof(*dp);
2721         ent = start;
2722         last_ent = ent;
2723
2724         /* initalize dir read context */
2725         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2726         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2727         if (!ctxt)
2728                 GOTO(free_page, rc = -ENOMEM);
2729         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2730         ctxt->ldc_op_data = op_data;
2731         ctxt->ldc_cb_op = cb_op;
2732         ctxt->ldc_hash = offset;
2733         ctxt->ldc_count = stripe_count;
2734
2735         while (1) {
2736                 next = lmv_dirent_next(ctxt);
2737
2738                 /* end of directory */
2739                 if (!next) {
2740                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2741                         break;
2742                 }
2743                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2744
2745                 ent_size = le16_to_cpu(next->lde_reclen);
2746
2747                 /* the last entry lde_reclen is 0, but it might not be the last
2748                  * one of this temporay dir page */
2749                 if (!ent_size)
2750                         ent_size = lu_dirent_calc_size(
2751                                         le16_to_cpu(next->lde_namelen),
2752                                         le32_to_cpu(next->lde_attrs));
2753                 /* page full */
2754                 if (ent_size > left_bytes)
2755                         break;
2756
2757                 memcpy(ent, next, ent_size);
2758
2759                 /* Replace . with master FID and Replace .. with the parent FID
2760                  * of master object */
2761                 if (strncmp(ent->lde_name, ".",
2762                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2763                     le16_to_cpu(ent->lde_namelen) == 1)
2764                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2765                 else if (strncmp(ent->lde_name, "..",
2766                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2767                            le16_to_cpu(ent->lde_namelen) == 2)
2768                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2769
2770                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2771                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2772                        le64_to_cpu(ent->lde_hash));
2773
2774                 left_bytes -= ent_size;
2775                 ent->lde_reclen = cpu_to_le16(ent_size);
2776                 last_ent = ent;
2777                 ent = (void *)ent + ent_size;
2778         };
2779
2780         last_ent->lde_reclen = 0;
2781
2782         if (ent == start)
2783                 dp->ldp_flags |= LDF_EMPTY;
2784         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2785                 dp->ldp_flags |= LDF_COLLIDE;
2786         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2787         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2788
2789         put_lmv_dir_ctxt(ctxt);
2790         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2791
2792         *ppage = page;
2793
2794         RETURN(0);
2795
2796 free_page:
2797         kunmap(page);
2798         __free_page(page);
2799
2800         return rc;
2801 }
2802
2803 static int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2804                          struct md_callback *cb_op, __u64 offset,
2805                          struct page **ppage)
2806 {
2807         struct obd_device *obd = exp->exp_obd;
2808         struct lmv_obd *lmv = &obd->u.lmv;
2809         struct lmv_tgt_desc *tgt;
2810         int rc;
2811
2812         ENTRY;
2813
2814         if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2815                 RETURN(-ENODATA);
2816
2817         if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2818                 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2819                 RETURN(rc);
2820         }
2821
2822         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2823         if (IS_ERR(tgt))
2824                 RETURN(PTR_ERR(tgt));
2825
2826         rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2827
2828         RETURN(rc);
2829 }
2830
2831 /**
2832  * Unlink a file/directory
2833  *
2834  * Unlink a file or directory under the parent dir. The unlink request
2835  * usually will be sent to the MDT where the child is located, but if
2836  * the client does not have the child FID then request will be sent to the
2837  * MDT where the parent is located.
2838  *
2839  * If the parent is a striped directory then it also needs to locate which
2840  * stripe the name of the child is located, and replace the parent FID
2841  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2842  * it will walk through all of sub-stripes until the child is being
2843  * unlinked finally.
2844  *
2845  * \param[in] exp       export refer to LMV
2846  * \param[in] op_data   different parameters transferred beween client
2847  *                      MD stacks, name, namelen, FIDs etc.
2848  *                      op_fid1 is the parent FID, op_fid2 is the child
2849  *                      FID.
2850  * \param[out] request  point to the request of unlink.
2851  *
2852  * retval               0 if succeed
2853  *                      negative errno if failed.
2854  */
2855 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2856                       struct ptlrpc_request **request)
2857 {
2858         struct obd_device *obd = exp->exp_obd;
2859         struct lmv_obd *lmv = &obd->u.lmv;
2860         struct lmv_tgt_desc *tgt;
2861         struct lmv_tgt_desc *parent_tgt;
2862         struct mdt_body *body;
2863         int rc;
2864
2865         ENTRY;
2866
2867         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2868         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2869         op_data->op_cap = cfs_curproc_cap_pack();
2870
2871 retry:
2872         parent_tgt = lmv_locate_tgt(lmv, op_data);
2873         if (IS_ERR(parent_tgt))
2874                 RETURN(PTR_ERR(parent_tgt));
2875
2876         if (likely(!fid_is_zero(&op_data->op_fid2))) {
2877                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2878                 if (IS_ERR(tgt))
2879                         RETURN(PTR_ERR(tgt));
2880         } else {
2881                 tgt = parent_tgt;
2882         }
2883
2884         /*
2885          * If child's fid is given, cancel unused locks for it if it is from
2886          * another export than parent.
2887          *
2888          * LOOKUP lock for child (fid3) should also be cancelled on parent
2889          * tgt_tgt in mdc_unlink().
2890          */
2891         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2892
2893         if (parent_tgt != tgt)
2894                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
2895                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2896                                       MF_MDC_CANCEL_FID3);
2897
2898         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2899                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2900         if (rc)
2901                 RETURN(rc);
2902
2903         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
2904                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
2905                tgt->ltd_index);
2906
2907         rc = md_unlink(tgt->ltd_exp, op_data, request);
2908         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2909                 ptlrpc_req_finished(*request);
2910                 *request = NULL;
2911                 goto retry;
2912         }
2913
2914         if (rc != -EREMOTE)
2915                 RETURN(rc);
2916
2917         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2918         if (body == NULL)
2919                 RETURN(-EPROTO);
2920
2921         /* Not cross-ref case, just get out of here. */
2922         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2923                 RETURN(rc);
2924
2925         /* This is a remote object, try remote MDT. */
2926         op_data->op_fid2 = body->mbo_fid1;
2927         ptlrpc_req_finished(*request);
2928         *request = NULL;
2929
2930         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2931         if (IS_ERR(tgt))
2932                 RETURN(PTR_ERR(tgt));
2933
2934         goto retry;
2935 }
2936
2937 static int lmv_precleanup(struct obd_device *obd)
2938 {
2939         ENTRY;
2940         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
2941         fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
2942         lprocfs_obd_cleanup(obd);
2943         lprocfs_free_md_stats(obd);
2944         RETURN(0);
2945 }
2946
2947 /**
2948  * Get by key a value associated with a LMV device.
2949  *
2950  * Dispatch request to lower-layer devices as needed.
2951  *
2952  * \param[in] env               execution environment for this thread
2953  * \param[in] exp               export for the LMV device
2954  * \param[in] keylen            length of key identifier
2955  * \param[in] key               identifier of key to get value for
2956  * \param[in] vallen            size of \a val
2957  * \param[out] val              pointer to storage location for value
2958  * \param[in] lsm               optional striping metadata of object
2959  *
2960  * \retval 0            on success
2961  * \retval negative     negated errno on failure
2962  */
2963 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2964                         __u32 keylen, void *key, __u32 *vallen, void *val)
2965 {
2966         struct obd_device *obd;
2967         struct lmv_obd *lmv;
2968         struct lu_tgt_desc *tgt;
2969         int rc = 0;
2970
2971         ENTRY;
2972
2973         obd = class_exp2obd(exp);
2974         if (obd == NULL) {
2975                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2976                        exp->exp_handle.h_cookie);
2977                 RETURN(-EINVAL);
2978         }
2979
2980         lmv = &obd->u.lmv;
2981         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2982                 LASSERT(*vallen == sizeof(__u32));
2983                 lmv_foreach_connected_tgt(lmv, tgt) {
2984                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2985                                           vallen, val))
2986                                 RETURN(0);
2987                 }
2988                 RETURN(-EINVAL);
2989         } else if (KEY_IS(KEY_MAX_EASIZE) ||
2990                    KEY_IS(KEY_DEFAULT_EASIZE) ||
2991                    KEY_IS(KEY_CONN_DATA)) {
2992                 /*
2993                  * Forwarding this request to first MDS, it should know LOV
2994                  * desc.
2995                  */
2996                 tgt = lmv_tgt(lmv, 0);
2997                 if (!tgt)
2998                         RETURN(-ENODEV);
2999
3000                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
3001                 if (!rc && KEY_IS(KEY_CONN_DATA))
3002                         exp->exp_connect_data = *(struct obd_connect_data *)val;
3003                 RETURN(rc);
3004         } else if (KEY_IS(KEY_TGT_COUNT)) {
3005                 *((int *)val) = lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count;
3006                 RETURN(0);
3007         }
3008
3009         CDEBUG(D_IOCTL, "Invalid key\n");
3010         RETURN(-EINVAL);
3011 }
3012
3013 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
3014                      int *__rcs, struct ptlrpc_request_set *_set)
3015 {
3016         struct obd_device *obd = class_exp2obd(exp);
3017         struct ptlrpc_request_set *set = _set;
3018         struct lmv_obd *lmv = &obd->u.lmv;
3019         int tgt_count = lmv->lmv_mdt_count;
3020         struct lu_tgt_desc *tgt;
3021         struct fid_array *fat, **fas = NULL;
3022         int i, rc, **rcs = NULL;
3023
3024         if (!set) {
3025                 set = ptlrpc_prep_set();
3026                 if (!set)
3027                         RETURN(-ENOMEM);
3028         }
3029
3030         /* split FIDs by targets */
3031         OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
3032         if (fas == NULL)
3033                 GOTO(out, rc = -ENOMEM);
3034         OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
3035         if (rcs == NULL)
3036                 GOTO(out_fas, rc = -ENOMEM);
3037
3038         for (i = 0; i < fa->fa_nr; i++) {
3039                 unsigned int idx;
3040
3041                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
3042                 if (rc) {
3043                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
3044                                PFID(&fa->fa_fids[i]), rc);
3045                         continue;
3046                 }
3047                 LASSERT(idx < tgt_count);
3048                 if (!fas[idx])
3049                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
3050                                   fa_fids[fa->fa_nr]));
3051                 if (!fas[idx])
3052                         GOTO(out, rc = -ENOMEM);
3053                 if (!rcs[idx])
3054                         OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
3055                 if (!rcs[idx])
3056                         GOTO(out, rc = -ENOMEM);
3057
3058                 fat = fas[idx];
3059                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
3060         }
3061
3062         lmv_foreach_connected_tgt(lmv, tgt) {
3063                 fat = fas[tgt->ltd_index];
3064                 if (!fat || fat->fa_nr == 0)
3065                         continue;
3066                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3067         }
3068
3069         rc = ptlrpc_set_wait(NULL, set);
3070         if (rc == 0) {
3071                 int j = 0;
3072                 for (i = 0; i < tgt_count; i++) {
3073                         fat = fas[i];
3074                         if (!fat || fat->fa_nr == 0)
3075                                 continue;
3076                         /* copy FIDs back */
3077                         memcpy(fa->fa_fids + j, fat->fa_fids,
3078                                fat->fa_nr * sizeof(struct lu_fid));
3079                         /* copy rcs back */
3080                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3081                         j += fat->fa_nr;
3082                 }
3083         }
3084         if (set != _set)
3085                 ptlrpc_set_destroy(set);
3086
3087 out:
3088         for (i = 0; i < tgt_count; i++) {
3089                 if (fas && fas[i])
3090                         OBD_FREE(fas[i], offsetof(struct fid_array,
3091                                                 fa_fids[fa->fa_nr]));
3092                 if (rcs && rcs[i])
3093                         OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3094         }
3095         if (rcs)
3096                 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3097 out_fas:
3098         if (fas)
3099                 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3100
3101         RETURN(rc);
3102 }
3103
3104 /**
3105  * Asynchronously set by key a value associated with a LMV device.
3106  *
3107  * Dispatch request to lower-layer devices as needed.
3108  *
3109  * \param[in] env       execution environment for this thread
3110  * \param[in] exp       export for the LMV device
3111  * \param[in] keylen    length of key identifier
3112  * \param[in] key       identifier of key to store value for
3113  * \param[in] vallen    size of value to store
3114  * \param[in] val       pointer to data to be stored
3115  * \param[in] set       optional list of related ptlrpc requests
3116  *
3117  * \retval 0            on success
3118  * \retval negative     negated errno on failure
3119  */
3120 static int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3121                               __u32 keylen, void *key, __u32 vallen, void *val,
3122                               struct ptlrpc_request_set *set)
3123 {
3124         struct lmv_tgt_desc *tgt;
3125         struct obd_device *obd;
3126         struct lmv_obd *lmv;
3127         int rc = 0;
3128         ENTRY;
3129
3130         obd = class_exp2obd(exp);
3131         if (obd == NULL) {
3132                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3133                        exp->exp_handle.h_cookie);
3134                 RETURN(-EINVAL);
3135         }
3136         lmv = &obd->u.lmv;
3137
3138         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3139             KEY_IS(KEY_DEFAULT_EASIZE)) {
3140                 int err = 0;
3141
3142                 lmv_foreach_connected_tgt(lmv, tgt) {
3143                         err = obd_set_info_async(env, tgt->ltd_exp,
3144                                                  keylen, key, vallen, val, set);
3145                         if (err && rc == 0)
3146                                 rc = err;
3147                 }
3148
3149                 RETURN(rc);
3150         }
3151
3152         RETURN(-EINVAL);
3153 }
3154
3155 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3156                             const struct lmv_mds_md_v1 *lmm1)
3157 {
3158         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3159         int             stripe_count;
3160         int             cplen;
3161         int             i;
3162         int             rc = 0;
3163         ENTRY;
3164
3165         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3166         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3167         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3168         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3169                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3170         else
3171                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3172         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3173         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3174         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3175         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3176                         sizeof(lsm->lsm_md_pool_name));
3177
3178         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3179                 RETURN(-E2BIG);
3180
3181         CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3182                "layout_version %d\n", lsm->lsm_md_stripe_count,
3183                lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3184                lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3185                lsm->lsm_md_layout_version);
3186
3187         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3188         for (i = 0; i < stripe_count; i++) {
3189                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3190                               &lmm1->lmv_stripe_fids[i]);
3191                 /*
3192                  * set default value -1, so lmv_locate_tgt() knows this stripe
3193                  * target is not initialized.
3194                  */
3195                 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3196                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3197                         continue;
3198
3199                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3200                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3201                 if (rc == -ENOENT)
3202                         continue;
3203
3204                 if (rc)
3205                         RETURN(rc);
3206
3207                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3208                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3209         }
3210
3211         RETURN(rc);
3212 }
3213
3214 static inline int lmv_unpack_user_md(struct obd_export *exp,
3215                                      struct lmv_stripe_md *lsm,
3216                                      const struct lmv_user_md *lmu)
3217 {
3218         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3219         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3220         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3221         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3222         lsm->lsm_md_max_inherit = lmu->lum_max_inherit;
3223         lsm->lsm_md_max_inherit_rr = lmu->lum_max_inherit_rr;
3224         lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3225
3226         return 0;
3227 }
3228
3229 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3230                         const union lmv_mds_md *lmm, size_t lmm_size)
3231 {
3232         struct lmv_stripe_md     *lsm;
3233         int                      lsm_size;
3234         int                      rc;
3235         bool                     allocated = false;
3236         ENTRY;
3237
3238         LASSERT(lsmp != NULL);
3239
3240         lsm = *lsmp;
3241         /* Free memmd */
3242         if (lsm != NULL && lmm == NULL) {
3243                 int i;
3244                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3245
3246                 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3247                         size_t lfm_size;
3248
3249                         lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3250                                                               lfm_value[0]);
3251                         OBD_FREE_LARGE(lfm, lfm_size);
3252                         RETURN(0);
3253                 }
3254
3255                 if (lmv_dir_striped(lsm)) {
3256                         for (i = 0; i < lsm->lsm_md_stripe_count; i++)
3257                                 iput(lsm->lsm_md_oinfo[i].lmo_root);
3258                         lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3259                 } else {
3260                         lsm_size = lmv_stripe_md_size(0);
3261                 }
3262                 OBD_FREE(lsm, lsm_size);
3263                 *lsmp = NULL;
3264                 RETURN(0);
3265         }
3266
3267         /* foreign lmv case */
3268         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3269                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3270
3271                 if (lfm == NULL) {
3272                         OBD_ALLOC_LARGE(lfm, lmm_size);
3273                         if (lfm == NULL)
3274                                 RETURN(-ENOMEM);
3275                         *lsmp = (struct lmv_stripe_md *)lfm;
3276                 }
3277                 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3278                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3279                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3280                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3281                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3282                        lfm->lfm_length);
3283                 RETURN(lmm_size);
3284         }
3285
3286         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3287                 RETURN(-EPERM);
3288
3289         /* Unpack memmd */
3290         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3291             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3292                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3293                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3294                        -EIO);
3295                 RETURN(-EIO);
3296         }
3297
3298         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3299                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3300         else
3301                 /**
3302                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3303                  * stripecount should be 0 then.
3304                  */
3305                 lsm_size = lmv_stripe_md_size(0);
3306
3307         if (lsm == NULL) {
3308                 OBD_ALLOC(lsm, lsm_size);
3309                 if (lsm == NULL)
3310                         RETURN(-ENOMEM);
3311                 allocated = true;
3312                 *lsmp = lsm;
3313         }
3314
3315         switch (le32_to_cpu(lmm->lmv_magic)) {
3316         case LMV_MAGIC_V1:
3317                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3318                 break;
3319         case LMV_USER_MAGIC:
3320                 rc = lmv_unpack_user_md(exp, lsm, &lmm->lmv_user_md);
3321                 break;
3322         default:
3323                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3324                        le32_to_cpu(lmm->lmv_magic));
3325                 rc = -EINVAL;
3326                 break;
3327         }
3328
3329         if (rc != 0 && allocated) {
3330                 OBD_FREE(lsm, lsm_size);
3331                 *lsmp = NULL;
3332                 lsm_size = rc;
3333         }
3334         RETURN(lsm_size);
3335 }
3336
3337 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3338 {
3339         lmv_unpackmd(NULL, &lsm, NULL, 0);
3340 }
3341 EXPORT_SYMBOL(lmv_free_memmd);
3342
3343 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3344                              union ldlm_policy_data *policy,
3345                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3346                              void *opaque)
3347 {
3348         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3349         struct lu_tgt_desc *tgt;
3350         int err;
3351         int rc = 0;
3352
3353         ENTRY;
3354
3355         LASSERT(fid != NULL);
3356
3357         lmv_foreach_connected_tgt(lmv, tgt) {
3358                 if (!tgt->ltd_active)
3359                         continue;
3360
3361                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3362                                        opaque);
3363                 if (!rc)
3364                         rc = err;
3365         }
3366         RETURN(rc);
3367 }
3368
3369 static int lmv_set_lock_data(struct obd_export *exp,
3370                              const struct lustre_handle *lockh,
3371                              void *data, __u64 *bits)
3372 {
3373         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3374         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3375         int rc;
3376
3377         ENTRY;
3378
3379         if (tgt == NULL || tgt->ltd_exp == NULL)
3380                 RETURN(-EINVAL);
3381         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3382         RETURN(rc);
3383 }
3384
3385 static enum ldlm_mode
3386 lmv_lock_match(struct obd_export *exp, __u64 flags,
3387                const struct lu_fid *fid, enum ldlm_type type,
3388                union ldlm_policy_data *policy,
3389                enum ldlm_mode mode, struct lustre_handle *lockh)
3390 {
3391         struct obd_device *obd = exp->exp_obd;
3392         struct lmv_obd *lmv = &obd->u.lmv;
3393         enum ldlm_mode rc;
3394         struct lu_tgt_desc *tgt;
3395         int i;
3396         int index;
3397
3398         ENTRY;
3399
3400         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
3401
3402         /*
3403          * With DNE every object can have two locks in different namespaces:
3404          * lookup lock in space of MDT storing direntry and update/open lock in
3405          * space of MDT storing inode.  Try the MDT that the FID maps to first,
3406          * since this can be easily found, and only try others if that fails.
3407          */
3408         for (i = 0, index = lmv_fid2tgt_index(lmv, fid);
3409              i < lmv->lmv_mdt_descs.ltd_tgts_size;
3410              i++, index = (index + 1) % lmv->lmv_mdt_descs.ltd_tgts_size) {
3411                 if (index < 0) {
3412                         CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
3413                                obd->obd_name, PFID(fid), index);
3414                         index = 0;
3415                 }
3416
3417                 tgt = lmv_tgt(lmv, index);
3418                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
3419                         continue;
3420
3421                 rc = md_lock_match(tgt->ltd_exp, flags, fid, type, policy, mode,
3422                                    lockh);
3423                 if (rc)
3424                         RETURN(rc);
3425         }
3426
3427         RETURN(0);
3428 }
3429
3430 static int
3431 lmv_get_lustre_md(struct obd_export *exp, struct req_capsule *pill,
3432                   struct obd_export *dt_exp, struct obd_export *md_exp,
3433                   struct lustre_md *md)
3434 {
3435         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3436         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3437
3438         if (!tgt || !tgt->ltd_exp)
3439                 return -EINVAL;
3440
3441         return md_get_lustre_md(tgt->ltd_exp, pill, dt_exp, md_exp, md);
3442 }
3443
3444 static int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
3445 {
3446         struct obd_device *obd = exp->exp_obd;
3447         struct lmv_obd *lmv = &obd->u.lmv;
3448         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3449
3450         ENTRY;
3451
3452         if (md->default_lmv) {
3453                 lmv_free_memmd(md->default_lmv);
3454                 md->default_lmv = NULL;
3455         }
3456         if (md->lmv != NULL) {
3457                 lmv_free_memmd(md->lmv);
3458                 md->lmv = NULL;
3459         }
3460         if (!tgt || !tgt->ltd_exp)
3461                 RETURN(-EINVAL);
3462         RETURN(md_free_lustre_md(tgt->ltd_exp, md));
3463 }
3464
3465 static int lmv_set_open_replay_data(struct obd_export *exp,
3466                                     struct obd_client_handle *och,
3467                                     struct lookup_intent *it)
3468 {
3469         struct obd_device *obd = exp->exp_obd;
3470         struct lmv_obd *lmv = &obd->u.lmv;
3471         struct lmv_tgt_desc *tgt;
3472
3473         ENTRY;
3474
3475         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3476         if (IS_ERR(tgt))
3477                 RETURN(PTR_ERR(tgt));
3478
3479         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3480 }
3481
3482 static int lmv_clear_open_replay_data(struct obd_export *exp,
3483                                       struct obd_client_handle *och)
3484 {
3485         struct obd_device *obd = exp->exp_obd;
3486         struct lmv_obd *lmv = &obd->u.lmv;
3487         struct lmv_tgt_desc *tgt;
3488
3489         ENTRY;
3490
3491         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3492         if (IS_ERR(tgt))
3493                 RETURN(PTR_ERR(tgt));
3494
3495         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3496 }
3497
3498 static int lmv_intent_getattr_async(struct obd_export *exp,
3499                                     struct md_op_item *item)
3500 {
3501         struct md_op_data *op_data = &item->mop_data;
3502         struct obd_device *obd = exp->exp_obd;
3503         struct lmv_obd *lmv = &obd->u.lmv;
3504         struct lmv_tgt_desc *ptgt;
3505         struct lmv_tgt_desc *ctgt;
3506         int rc;
3507
3508         ENTRY;
3509
3510         if (!fid_is_sane(&op_data->op_fid2))
3511                 RETURN(-EINVAL);
3512
3513         ptgt = lmv_locate_tgt(lmv, op_data);
3514         if (IS_ERR(ptgt))
3515                 RETURN(PTR_ERR(ptgt));
3516
3517         ctgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3518         if (IS_ERR(ctgt))
3519                 RETURN(PTR_ERR(ctgt));
3520
3521         /*
3522          * remote object needs two RPCs to lookup and getattr, considering the
3523          * complexity don't support statahead for now.
3524          */
3525         if (ctgt != ptgt)
3526                 RETURN(-EREMOTE);
3527
3528         rc = md_intent_getattr_async(ptgt->ltd_exp, item);
3529
3530         RETURN(rc);
3531 }
3532
3533 static int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3534                                struct lu_fid *fid, __u64 *bits)
3535 {
3536         struct obd_device *obd = exp->exp_obd;
3537         struct lmv_obd *lmv = &obd->u.lmv;
3538         struct lmv_tgt_desc *tgt;
3539         int rc;
3540
3541         ENTRY;
3542
3543         tgt = lmv_fid2tgt(lmv, fid);
3544         if (IS_ERR(tgt))
3545                 RETURN(PTR_ERR(tgt));
3546
3547         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3548         RETURN(rc);
3549 }
3550
3551 static int lmv_get_fid_from_lsm(struct obd_export *exp,
3552                                 const struct lmv_stripe_md *lsm,
3553                                 const char *name, int namelen,
3554                                 struct lu_fid *fid)
3555 {
3556         const struct lmv_oinfo *oinfo;
3557
3558         LASSERT(lmv_dir_striped(lsm));
3559
3560         oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
3561         if (IS_ERR(oinfo))
3562                 return PTR_ERR(oinfo);
3563
3564         *fid = oinfo->lmo_fid;
3565
3566         RETURN(0);
3567 }
3568
3569 /**
3570  * For lmv, only need to send request to master MDT, and the master MDT will
3571  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3572  * we directly fetch data from the slave MDTs.
3573  */
3574 static int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3575                         struct obd_quotactl *oqctl)
3576 {
3577         struct obd_device *obd = class_exp2obd(exp);
3578         struct lmv_obd *lmv = &obd->u.lmv;
3579         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3580         __u64 curspace, curinodes;
3581         int rc = 0;
3582
3583         ENTRY;
3584
3585         if (!tgt || !tgt->ltd_exp || !tgt->ltd_active) {
3586                 CERROR("master lmv inactive\n");
3587                 RETURN(-EIO);
3588         }
3589
3590         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3591                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3592                 RETURN(rc);
3593         }
3594
3595         curspace = curinodes = 0;
3596         lmv_foreach_connected_tgt(lmv, tgt) {
3597                 int err;
3598
3599                 if (!tgt->ltd_active)
3600                         continue;
3601
3602                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3603                 if (err) {
3604                         CERROR("getquota on mdt %d failed. %d\n",
3605                                tgt->ltd_index, err);
3606                         if (!rc)
3607                                 rc = err;
3608                 } else {
3609                         curspace += oqctl->qc_dqblk.dqb_curspace;
3610                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3611                 }
3612         }
3613         oqctl->qc_dqblk.dqb_curspace = curspace;
3614         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3615
3616         RETURN(rc);
3617 }
3618
3619 static int lmv_merge_attr(struct obd_export *exp,
3620                           const struct lmv_stripe_md *lsm,
3621                           struct cl_attr *attr,
3622                           ldlm_blocking_callback cb_blocking)
3623 {
3624         int rc;
3625         int i;
3626
3627         if (!lmv_dir_striped(lsm))
3628                 return 0;
3629
3630         rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3631         if (rc < 0)
3632                 return rc;
3633
3634         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3635                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3636
3637                 if (!inode)
3638                         continue;
3639
3640                 CDEBUG(D_INFO,
3641                        "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
3642                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3643                        i_size_read(inode), (unsigned long long)inode->i_blocks,
3644                        inode->i_nlink, (s64)inode->i_atime.tv_sec,
3645                        (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
3646
3647                 /* for slave stripe, it needs to subtract nlink for . and .. */
3648                 if (i != 0)
3649                         attr->cat_nlink += inode->i_nlink - 2;
3650                 else
3651                         attr->cat_nlink = inode->i_nlink;
3652
3653                 attr->cat_size += i_size_read(inode);
3654                 attr->cat_blocks += inode->i_blocks;
3655
3656                 if (attr->cat_atime < inode->i_atime.tv_sec)
3657                         attr->cat_atime = inode->i_atime.tv_sec;
3658
3659                 if (attr->cat_ctime < inode->i_ctime.tv_sec)
3660                         attr->cat_ctime = inode->i_ctime.tv_sec;
3661
3662                 if (attr->cat_mtime < inode->i_mtime.tv_sec)
3663                         attr->cat_mtime = inode->i_mtime.tv_sec;
3664         }
3665         return 0;
3666 }
3667
3668 static const struct obd_ops lmv_obd_ops = {
3669         .o_owner                = THIS_MODULE,
3670         .o_setup                = lmv_setup,
3671         .o_cleanup              = lmv_cleanup,
3672         .o_precleanup           = lmv_precleanup,
3673         .o_process_config       = lmv_process_config,
3674         .o_connect              = lmv_connect,
3675         .o_disconnect           = lmv_disconnect,
3676         .o_statfs               = lmv_statfs,
3677         .o_get_info             = lmv_get_info,
3678         .o_set_info_async       = lmv_set_info_async,
3679         .o_notify               = lmv_notify,
3680         .o_get_uuid             = lmv_get_uuid,
3681         .o_fid_alloc            = lmv_fid_alloc,
3682         .o_iocontrol            = lmv_iocontrol,
3683         .o_quotactl             = lmv_quotactl
3684 };
3685
3686 static const struct md_ops lmv_md_ops = {
3687         .m_get_root             = lmv_get_root,
3688         .m_null_inode           = lmv_null_inode,
3689         .m_close                = lmv_close,
3690         .m_create               = lmv_create,
3691         .m_enqueue              = lmv_enqueue,
3692         .m_getattr              = lmv_getattr,
3693         .m_getxattr             = lmv_getxattr,
3694         .m_getattr_name         = lmv_getattr_name,
3695         .m_intent_lock          = lmv_intent_lock,
3696         .m_link                 = lmv_link,
3697         .m_rename               = lmv_rename,
3698         .m_setattr              = lmv_setattr,
3699         .m_setxattr             = lmv_setxattr,
3700         .m_fsync                = lmv_fsync,
3701         .m_file_resync          = lmv_file_resync,
3702         .m_read_page            = lmv_read_page,
3703         .m_unlink               = lmv_unlink,
3704         .m_init_ea_size         = lmv_init_ea_size,
3705         .m_cancel_unused        = lmv_cancel_unused,
3706         .m_set_lock_data        = lmv_set_lock_data,
3707         .m_lock_match           = lmv_lock_match,
3708         .m_get_lustre_md        = lmv_get_lustre_md,
3709         .m_free_lustre_md       = lmv_free_lustre_md,
3710         .m_merge_attr           = lmv_merge_attr,
3711         .m_set_open_replay_data = lmv_set_open_replay_data,
3712         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3713         .m_intent_getattr_async = lmv_intent_getattr_async,
3714         .m_revalidate_lock      = lmv_revalidate_lock,
3715         .m_get_fid_from_lsm     = lmv_get_fid_from_lsm,
3716         .m_unpackmd             = lmv_unpackmd,
3717         .m_rmfid                = lmv_rmfid,
3718 };
3719
3720 static int __init lmv_init(void)
3721 {
3722         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true,
3723                                    LUSTRE_LMV_NAME, NULL);
3724 }
3725
3726 static void __exit lmv_exit(void)
3727 {
3728         class_unregister_type(LUSTRE_LMV_NAME);
3729 }
3730
3731 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3732 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
3733 MODULE_VERSION(LUSTRE_VERSION_STRING);
3734 MODULE_LICENSE("GPL");
3735
3736 module_init(lmv_init);
3737 module_exit(lmv_exit);