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_processid 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 (!nid_is_lo0(&lnet_id.nid)) {
1132                         lmv->lmv_qos_rr_index = ntohl(lnet_id.nid.nid_addr[0]);
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_processid lnet_id;
1216                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1217                         break;
1218
1219                 if (!nid_is_lo0(&lnet_id.nid)) {
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 = nidhash(&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         int err = 0;
1243
1244         ENTRY;
1245
1246         OBD_ALLOC(temp, sizeof(*temp));
1247         if (temp == NULL)
1248                 RETURN(-ENOMEM);
1249
1250         /* distribute statfs among MDTs */
1251         idx = lmv_select_statfs_mdt(lmv, flags);
1252
1253         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1254                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1255                 tgt = lmv_tgt(lmv, idx);
1256                 if (!tgt || !tgt->ltd_exp)
1257                         continue;
1258
1259                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age,
1260                                 flags | OBD_STATFS_NESTED);
1261                 if (rc) {
1262                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1263                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1264                         err = rc;
1265                         /* Try another MDT */
1266                         if (flags & OBD_STATFS_SUM)
1267                                 continue;
1268                         GOTO(out_free_temp, rc);
1269                 }
1270
1271                 if (temp->os_state & OS_STATFS_SUM ||
1272                     flags == OBD_STATFS_FOR_MDT0) {
1273                         /* reset to the last aggregated values
1274                          * and don't sum with non-aggrated data */
1275                         /* If the statfs is from mount, it needs to retrieve
1276                          * necessary information from MDT0. i.e. mount does
1277                          * not need the merged osfs from all of MDT. Also
1278                          * clients can be mounted as long as MDT0 is in
1279                          * service */
1280                         *osfs = *temp;
1281                         GOTO(out_free_temp, rc);
1282                 }
1283
1284                 if (i == 0) {
1285                         *osfs = *temp;
1286                 } else {
1287                         osfs->os_bavail += temp->os_bavail;
1288                         osfs->os_blocks += temp->os_blocks;
1289                         osfs->os_ffree += temp->os_ffree;
1290                         osfs->os_files += temp->os_files;
1291                         osfs->os_granted += temp->os_granted;
1292                 }
1293         }
1294         /* There is no stats from some MDTs, data incomplete */
1295         if (err)
1296                 rc = err;
1297 out_free_temp:
1298         OBD_FREE(temp, sizeof(*temp));
1299         RETURN(rc);
1300 }
1301
1302 static int lmv_statfs_update(void *cookie, int rc)
1303 {
1304         struct obd_info *oinfo = cookie;
1305         struct obd_device *obd = oinfo->oi_obd;
1306         struct lmv_obd *lmv = &obd->u.lmv;
1307         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1308         struct obd_statfs *osfs = oinfo->oi_osfs;
1309
1310         /*
1311          * NB: don't deactivate TGT upon error, because we may not trigger async
1312          * statfs any longer, then there is no chance to activate TGT.
1313          */
1314         if (!rc) {
1315                 spin_lock(&lmv->lmv_lock);
1316                 tgt->ltd_statfs = *osfs;
1317                 tgt->ltd_statfs_age = ktime_get_seconds();
1318                 spin_unlock(&lmv->lmv_lock);
1319                 set_bit(LQ_DIRTY, &lmv->lmv_qos.lq_flags);
1320         }
1321
1322         return rc;
1323 }
1324
1325 /* update tgt statfs async if it's ld_qos_maxage old */
1326 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1327 {
1328         struct obd_info oinfo = {
1329                 .oi_obd = obd,
1330                 .oi_tgt = tgt,
1331                 .oi_cb_up = lmv_statfs_update,
1332         };
1333         int rc;
1334
1335         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1336             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1337                 return 0;
1338
1339         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1340
1341         return rc;
1342 }
1343
1344 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1345                         struct lu_fid *fid)
1346 {
1347         struct obd_device *obd = exp->exp_obd;
1348         struct lmv_obd *lmv = &obd->u.lmv;
1349         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1350         int rc;
1351
1352         ENTRY;
1353
1354         if (!tgt)
1355                 RETURN(-ENODEV);
1356
1357         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1358         RETURN(rc);
1359 }
1360
1361 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1362                         u64 obd_md_valid, const char *name, size_t buf_size,
1363                         struct ptlrpc_request **req)
1364 {
1365         struct obd_device *obd = exp->exp_obd;
1366         struct lmv_obd *lmv = &obd->u.lmv;
1367         struct lmv_tgt_desc *tgt;
1368         int rc;
1369
1370         ENTRY;
1371
1372         tgt = lmv_fid2tgt(lmv, fid);
1373         if (IS_ERR(tgt))
1374                 RETURN(PTR_ERR(tgt));
1375
1376         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1377
1378         RETURN(rc);
1379 }
1380
1381 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1382                         u64 obd_md_valid, const char *name,
1383                         const void *value, size_t value_size,
1384                         unsigned int xattr_flags, u32 suppgid,
1385                         struct ptlrpc_request **req)
1386 {
1387         struct obd_device *obd = exp->exp_obd;
1388         struct lmv_obd *lmv = &obd->u.lmv;
1389         struct lmv_tgt_desc *tgt;
1390         int rc;
1391
1392         ENTRY;
1393
1394         tgt = lmv_fid2tgt(lmv, fid);
1395         if (IS_ERR(tgt))
1396                 RETURN(PTR_ERR(tgt));
1397
1398         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1399                          value, value_size, xattr_flags, suppgid, req);
1400
1401         RETURN(rc);
1402 }
1403
1404 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1405                        struct ptlrpc_request **request)
1406 {
1407         struct obd_device *obd = exp->exp_obd;
1408         struct lmv_obd *lmv = &obd->u.lmv;
1409         struct lmv_tgt_desc *tgt;
1410         int rc;
1411
1412         ENTRY;
1413
1414         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1415         if (IS_ERR(tgt))
1416                 RETURN(PTR_ERR(tgt));
1417
1418         if (op_data->op_flags & MF_GET_MDT_IDX) {
1419                 op_data->op_mds = tgt->ltd_index;
1420                 RETURN(0);
1421         }
1422
1423         rc = md_getattr(tgt->ltd_exp, op_data, request);
1424
1425         RETURN(rc);
1426 }
1427
1428 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1429 {
1430         struct obd_device *obd = exp->exp_obd;
1431         struct lmv_obd *lmv = &obd->u.lmv;
1432         struct lu_tgt_desc *tgt;
1433
1434         ENTRY;
1435
1436         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1437
1438         /*
1439          * With DNE every object can have two locks in different namespaces:
1440          * lookup lock in space of MDT storing direntry and update/open lock in
1441          * space of MDT storing inode.
1442          */
1443         lmv_foreach_connected_tgt(lmv, tgt)
1444                 md_null_inode(tgt->ltd_exp, fid);
1445
1446         RETURN(0);
1447 }
1448
1449 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1450                      struct md_open_data *mod, struct ptlrpc_request **request)
1451 {
1452         struct obd_device *obd = exp->exp_obd;
1453         struct lmv_obd *lmv = &obd->u.lmv;
1454         struct lmv_tgt_desc *tgt;
1455         int rc;
1456
1457         ENTRY;
1458
1459         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1460         if (IS_ERR(tgt))
1461                 RETURN(PTR_ERR(tgt));
1462
1463         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1464         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1465         RETURN(rc);
1466 }
1467
1468 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 mdt,
1469                                               unsigned short dir_depth)
1470 {
1471         struct lu_tgt_desc *tgt, *cur = NULL;
1472         __u64 total_avail = 0;
1473         __u64 total_weight = 0;
1474         __u64 cur_weight = 0;
1475         int total_usable = 0;
1476         __u64 rand;
1477         int rc;
1478
1479         ENTRY;
1480
1481         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1482                 RETURN(ERR_PTR(-EAGAIN));
1483
1484         down_write(&lmv->lmv_qos.lq_rw_sem);
1485
1486         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1487                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1488
1489         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1490         if (rc)
1491                 GOTO(unlock, tgt = ERR_PTR(rc));
1492
1493         lmv_foreach_tgt(lmv, tgt) {
1494                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1495                         tgt->ltd_qos.ltq_usable = 0;
1496                         continue;
1497                 }
1498
1499                 tgt->ltd_qos.ltq_usable = 1;
1500                 lu_tgt_qos_weight_calc(tgt);
1501                 if (tgt->ltd_index == mdt)
1502                         cur = tgt;
1503                 total_avail += tgt->ltd_qos.ltq_avail;
1504                 total_weight += tgt->ltd_qos.ltq_weight;
1505                 total_usable++;
1506         }
1507
1508         /* if current MDT has above-average space, within range of the QOS
1509          * threshold, stay on the same MDT to avoid creating needless remote
1510          * MDT directories. It's more likely for low level directories
1511          * "16 / (dir_depth + 10)" is the factor to make it more unlikely for
1512          * top level directories, while more likely for low levels.
1513          */
1514         rand = total_avail * 16 / (total_usable * (dir_depth + 10));
1515         if (cur && cur->ltd_qos.ltq_avail >= rand) {
1516                 tgt = cur;
1517                 GOTO(unlock, tgt);
1518         }
1519
1520         rand = lu_prandom_u64_max(total_weight);
1521
1522         lmv_foreach_connected_tgt(lmv, tgt) {
1523                 if (!tgt->ltd_qos.ltq_usable)
1524                         continue;
1525
1526                 cur_weight += tgt->ltd_qos.ltq_weight;
1527                 if (cur_weight < rand)
1528                         continue;
1529
1530                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1531                 GOTO(unlock, tgt);
1532         }
1533
1534         /* no proper target found */
1535         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1536 unlock:
1537         up_write(&lmv->lmv_qos.lq_rw_sem);
1538
1539         return tgt;
1540 }
1541
1542 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv)
1543 {
1544         struct lu_tgt_desc *tgt;
1545         int i;
1546         int index;
1547
1548         ENTRY;
1549
1550         spin_lock(&lmv->lmv_lock);
1551         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1552                 index = (i + lmv->lmv_qos_rr_index) %
1553                         lmv->lmv_mdt_descs.ltd_tgts_size;
1554                 tgt = lmv_tgt(lmv, index);
1555                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1556                         continue;
1557
1558                 lmv->lmv_qos_rr_index = (tgt->ltd_index + 1) %
1559                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1560                 spin_unlock(&lmv->lmv_lock);
1561
1562                 RETURN(tgt);
1563         }
1564         spin_unlock(&lmv->lmv_lock);
1565
1566         RETURN(ERR_PTR(-ENODEV));
1567 }
1568
1569 /* locate MDT which is less full (avoid the most full MDT) */
1570 static struct lu_tgt_desc *lmv_locate_tgt_lf(struct lmv_obd *lmv)
1571 {
1572         struct lu_tgt_desc *min = NULL;
1573         struct lu_tgt_desc *tgt;
1574         __u64 avail = 0;
1575         __u64 rand;
1576
1577         ENTRY;
1578
1579         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1580                 RETURN(ERR_PTR(-EAGAIN));
1581
1582         down_write(&lmv->lmv_qos.lq_rw_sem);
1583
1584         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1585                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1586
1587         lmv_foreach_tgt(lmv, tgt) {
1588                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1589                         tgt->ltd_qos.ltq_usable = 0;
1590                         continue;
1591                 }
1592
1593                 tgt->ltd_qos.ltq_usable = 1;
1594                 lu_tgt_qos_weight_calc(tgt);
1595                 avail += tgt->ltd_qos.ltq_avail;
1596                 if (!min || min->ltd_qos.ltq_avail > tgt->ltd_qos.ltq_avail)
1597                         min = tgt;
1598         }
1599
1600         /* avoid the most full MDT */
1601         if (min)
1602                 avail -= min->ltd_qos.ltq_avail;
1603
1604         rand = lu_prandom_u64_max(avail);
1605         avail = 0;
1606         lmv_foreach_connected_tgt(lmv, tgt) {
1607                 if (!tgt->ltd_qos.ltq_usable)
1608                         continue;
1609
1610                 if (tgt == min)
1611                         continue;
1612
1613                 avail += tgt->ltd_qos.ltq_avail;
1614                 if (avail < rand)
1615                         continue;
1616
1617                 GOTO(unlock, tgt);
1618         }
1619
1620         /* no proper target found */
1621         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1622 unlock:
1623         up_write(&lmv->lmv_qos.lq_rw_sem);
1624
1625         RETURN(tgt);
1626 }
1627
1628 /* locate MDT by file name, for striped directory, the file name hash decides
1629  * which stripe its dirent is stored.
1630  */
1631 static struct lmv_tgt_desc *
1632 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1633                        const char *name, int namelen, struct lu_fid *fid,
1634                        __u32 *mds, bool new_layout)
1635 {
1636         struct lmv_tgt_desc *tgt;
1637         const struct lmv_oinfo *oinfo;
1638
1639         if (!lmv_dir_striped(lsm) || !namelen) {
1640                 tgt = lmv_fid2tgt(lmv, fid);
1641                 if (IS_ERR(tgt))
1642                         return tgt;
1643
1644                 *mds = tgt->ltd_index;
1645                 return tgt;
1646         }
1647
1648         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1649                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1650                         return ERR_PTR(-EBADF);
1651                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1652         } else {
1653                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, new_layout);
1654                 if (IS_ERR(oinfo))
1655                         return ERR_CAST(oinfo);
1656         }
1657
1658         /* check stripe FID is sane */
1659         if (!fid_is_sane(&oinfo->lmo_fid))
1660                 return ERR_PTR(-ENODEV);
1661
1662         *fid = oinfo->lmo_fid;
1663         *mds = oinfo->lmo_mds;
1664         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1665
1666         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1667
1668         return tgt ? tgt : ERR_PTR(-ENODEV);
1669 }
1670
1671 /**
1672  * Locate MDT of op_data->op_fid1
1673  *
1674  * For striped directory, it will locate the stripe by name hash, if hash_type
1675  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1676  * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1677  * indicates whether old or new layout is used to locate.
1678  *
1679  * For plain direcotry, it just locate the MDT of op_data->op_fid1.
1680  *
1681  * \param[in] lmv               LMV device
1682  * \param[in/out] op_data       client MD stack parameters, name, namelen etc,
1683  *                              op_mds and op_fid1 will be updated if op_mea1
1684  *                              indicates fid1 represents a striped directory.
1685  *
1686  * retval               pointer to the lmv_tgt_desc if succeed.
1687  *                      ERR_PTR(errno) if failed.
1688  */
1689 struct lmv_tgt_desc *
1690 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1691 {
1692         struct lmv_stripe_md *lsm = op_data->op_mea1;
1693         struct lmv_oinfo *oinfo;
1694         struct lmv_tgt_desc *tgt;
1695
1696         if (lmv_dir_foreign(lsm))
1697                 return ERR_PTR(-ENODATA);
1698
1699         /* During creating VOLATILE file, it should honor the mdt
1700          * index if the file under striped dir is being restored, see
1701          * ct_restore(). */
1702         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1703             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1704                 tgt = lmv_tgt(lmv, op_data->op_mds);
1705                 if (!tgt)
1706                         return ERR_PTR(-ENODEV);
1707
1708                 if (lmv_dir_striped(lsm)) {
1709                         int i;
1710
1711                         /* refill the right parent fid */
1712                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1713                                 oinfo = &lsm->lsm_md_oinfo[i];
1714                                 if (oinfo->lmo_mds == op_data->op_mds) {
1715                                         op_data->op_fid1 = oinfo->lmo_fid;
1716                                         break;
1717                                 }
1718                         }
1719
1720                         if (i == lsm->lsm_md_stripe_count)
1721                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1722                 }
1723         } else if (lmv_dir_bad_hash(lsm)) {
1724                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1725                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1726
1727                 op_data->op_fid1 = oinfo->lmo_fid;
1728                 op_data->op_mds = oinfo->lmo_mds;
1729                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1730                 if (!tgt)
1731                         return ERR_PTR(-ENODEV);
1732         } else {
1733                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1734                                 op_data->op_name, op_data->op_namelen,
1735                                 &op_data->op_fid1, &op_data->op_mds,
1736                                 op_data->op_new_layout);
1737         }
1738
1739         return tgt;
1740 }
1741
1742 /* Locate MDT of op_data->op_fid2 for link/rename */
1743 static struct lmv_tgt_desc *
1744 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1745 {
1746         struct lmv_tgt_desc *tgt;
1747         int rc;
1748
1749         LASSERT(op_data->op_name);
1750         if (lmv_dir_layout_changing(op_data->op_mea2)) {
1751                 struct lu_fid fid1 = op_data->op_fid1;
1752                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1753                 struct ptlrpc_request *request = NULL;
1754
1755                 /*
1756                  * avoid creating new file under old layout of migrating
1757                  * directory, check it here.
1758                  */
1759                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1760                                 op_data->op_name, op_data->op_namelen,
1761                                 &op_data->op_fid2, &op_data->op_mds, false);
1762                 if (IS_ERR(tgt))
1763                         RETURN(tgt);
1764
1765                 op_data->op_fid1 = op_data->op_fid2;
1766                 op_data->op_mea1 = op_data->op_mea2;
1767                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1768                 op_data->op_fid1 = fid1;
1769                 op_data->op_mea1 = lsm1;
1770                 if (!rc) {
1771                         ptlrpc_req_finished(request);
1772                         RETURN(ERR_PTR(-EEXIST));
1773                 }
1774
1775                 if (rc != -ENOENT)
1776                         RETURN(ERR_PTR(rc));
1777         }
1778
1779         return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1780                                 op_data->op_namelen, &op_data->op_fid2,
1781                                 &op_data->op_mds, true);
1782 }
1783
1784 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1785 {
1786         struct lu_tgt_desc *tgt;
1787         struct ptlrpc_request *request;
1788         int rc;
1789
1790         LASSERT(lmv_dir_layout_changing(op_data->op_mea1));
1791         LASSERT(!op_data->op_new_layout);
1792
1793         tgt = lmv_locate_tgt(lmv, op_data);
1794         if (IS_ERR(tgt))
1795                 return PTR_ERR(tgt);
1796
1797         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1798         if (!rc) {
1799                 ptlrpc_req_finished(request);
1800                 return -EEXIST;
1801         }
1802
1803         return rc;
1804 }
1805
1806 /* mkdir by QoS upon 'lfs mkdir -i -1'.
1807  *
1808  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1809  * directories under striped directory.
1810  */
1811 static inline bool lmv_op_user_qos_mkdir(const struct md_op_data *op_data)
1812 {
1813         const struct lmv_user_md *lum = op_data->op_data;
1814
1815         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1816                 return false;
1817
1818         if (lmv_dir_striped(op_data->op_mea1))
1819                 return false;
1820
1821         return (op_data->op_cli_flags & CLI_SET_MEA) && lum &&
1822                le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC &&
1823                le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
1824 }
1825
1826 /* mkdir by QoS if either ROOT or parent default LMV is space balanced. */
1827 static inline bool lmv_op_default_qos_mkdir(const struct md_op_data *op_data)
1828 {
1829         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1830
1831         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1832                 return false;
1833
1834         if (lmv_dir_striped(op_data->op_mea1))
1835                 return false;
1836
1837         return (op_data->op_flags & MF_QOS_MKDIR) ||
1838                (lsm && lsm->lsm_md_master_mdt_index == LMV_OFFSET_DEFAULT);
1839 }
1840
1841 /* if parent default LMV is space balanced, and
1842  * 1. max_inherit_rr is set
1843  * 2. or parent is ROOT
1844  * mkdir roundrobin. Or if parent doesn't have default LMV, while ROOT default
1845  * LMV requests roundrobin mkdir, do the same.
1846  * NB, this needs to check server is balanced, which is done by caller.
1847  */
1848 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data)
1849 {
1850         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1851
1852         if (!lmv_op_default_qos_mkdir(op_data))
1853                 return false;
1854
1855         return (op_data->op_flags & MF_RR_MKDIR) ||
1856                (lsm && lsm->lsm_md_max_inherit_rr != LMV_INHERIT_RR_NONE) ||
1857                fid_is_root(&op_data->op_fid1);
1858 }
1859
1860 /* 'lfs mkdir -i <specific_MDT>' */
1861 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1862 {
1863         const struct lmv_user_md *lum = op_data->op_data;
1864
1865         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1866                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1867                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1868                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1869                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1870 }
1871
1872 /* parent default LMV master_mdt_index is not -1. */
1873 static inline bool
1874 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1875 {
1876         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1877                op_data->op_default_mea1 &&
1878                op_data->op_default_mea1->lsm_md_master_mdt_index !=
1879                         LMV_OFFSET_DEFAULT;
1880 }
1881
1882 /* locate MDT by space usage */
1883 static struct lu_tgt_desc *lmv_locate_tgt_by_space(struct lmv_obd *lmv,
1884                                                    struct md_op_data *op_data,
1885                                                    struct lmv_tgt_desc *tgt)
1886 {
1887         struct lmv_tgt_desc *tmp = tgt;
1888
1889         tgt = lmv_locate_tgt_qos(lmv, op_data->op_mds, op_data->op_dir_depth);
1890         if (tgt == ERR_PTR(-EAGAIN)) {
1891                 if (ltd_qos_is_balanced(&lmv->lmv_mdt_descs) &&
1892                     !lmv_op_default_rr_mkdir(op_data) &&
1893                     !lmv_op_user_qos_mkdir(op_data))
1894                         /* if not necessary, don't create remote directory. */
1895                         tgt = tmp;
1896                 else
1897                         tgt = lmv_locate_tgt_rr(lmv);
1898         }
1899
1900         /*
1901          * only update statfs after QoS mkdir, this means the cached statfs may
1902          * be stale, and current mkdir may not follow QoS accurately, but it's
1903          * not serious, and avoids periodic statfs when client doesn't mkdir by
1904          * QoS.
1905          */
1906         if (!IS_ERR(tgt)) {
1907                 op_data->op_mds = tgt->ltd_index;
1908                 lmv_statfs_check_update(lmv2obd_dev(lmv), tgt);
1909         }
1910
1911         return tgt;
1912 }
1913
1914 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1915                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1916                 gid_t gid, kernel_cap_t cap_effective, __u64 rdev,
1917                 struct ptlrpc_request **request)
1918 {
1919         struct obd_device *obd = exp->exp_obd;
1920         struct lmv_obd *lmv = &obd->u.lmv;
1921         struct lmv_tgt_desc *tgt;
1922         struct mdt_body *repbody;
1923         int rc;
1924
1925         ENTRY;
1926
1927         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1928                 RETURN(-EIO);
1929
1930         if (lmv_dir_bad_hash(op_data->op_mea1))
1931                 RETURN(-EBADF);
1932
1933         if (lmv_dir_layout_changing(op_data->op_mea1)) {
1934                 /*
1935                  * if parent is migrating, create() needs to lookup existing
1936                  * name in both old and new layout, check old layout on client.
1937                  */
1938                 rc = lmv_old_layout_lookup(lmv, op_data);
1939                 if (rc != -ENOENT)
1940                         RETURN(rc);
1941
1942                 op_data->op_new_layout = true;
1943         }
1944
1945         tgt = lmv_locate_tgt(lmv, op_data);
1946         if (IS_ERR(tgt))
1947                 RETURN(PTR_ERR(tgt));
1948
1949         /* the order to apply policy in mkdir:
1950          * 1. is "lfs mkdir -i N"? mkdir on MDT N.
1951          * 2. is "lfs mkdir -i -1"? mkdir by space usage.
1952          * 3. is starting MDT specified in default LMV? mkdir on MDT N.
1953          * 4. is default LMV space balanced? mkdir by space usage.
1954          */
1955         if (lmv_op_user_specific_mkdir(op_data)) {
1956                 struct lmv_user_md *lum = op_data->op_data;
1957
1958                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
1959                 tgt = lmv_tgt(lmv, op_data->op_mds);
1960                 if (!tgt)
1961                         RETURN(-ENODEV);
1962         } else if (lmv_op_user_qos_mkdir(op_data)) {
1963                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
1964                 if (IS_ERR(tgt))
1965                         RETURN(PTR_ERR(tgt));
1966         } else if (lmv_op_default_specific_mkdir(op_data)) {
1967                 op_data->op_mds =
1968                         op_data->op_default_mea1->lsm_md_master_mdt_index;
1969                 tgt = lmv_tgt(lmv, op_data->op_mds);
1970                 if (!tgt)
1971                         RETURN(-ENODEV);
1972         } else if (lmv_op_default_qos_mkdir(op_data)) {
1973                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
1974                 if (IS_ERR(tgt))
1975                         RETURN(PTR_ERR(tgt));
1976         }
1977
1978 retry:
1979         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1980         if (rc)
1981                 RETURN(rc);
1982
1983         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
1984                 (int)op_data->op_namelen, op_data->op_name,
1985                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
1986                 op_data->op_mds);
1987
1988         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1989         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1990                        cap_effective, rdev, request);
1991         if (rc == 0) {
1992                 if (*request == NULL)
1993                         RETURN(rc);
1994                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1995         }
1996
1997         /* dir restripe needs to send to MDT where dir is located */
1998         if (rc != -EREMOTE ||
1999             !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
2000                 RETURN(rc);
2001
2002         repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2003         if (repbody == NULL)
2004                 RETURN(-EPROTO);
2005
2006         /* Not cross-ref case, just get out of here. */
2007         if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
2008                 RETURN(rc);
2009
2010         op_data->op_fid2 = repbody->mbo_fid1;
2011         ptlrpc_req_finished(*request);
2012         *request = NULL;
2013
2014         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2015         if (IS_ERR(tgt))
2016                 RETURN(PTR_ERR(tgt));
2017
2018         op_data->op_mds = tgt->ltd_index;
2019         goto retry;
2020 }
2021
2022 static int
2023 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
2024             const union ldlm_policy_data *policy, struct md_op_data *op_data,
2025             struct lustre_handle *lockh, __u64 extra_lock_flags)
2026 {
2027         struct obd_device *obd = exp->exp_obd;
2028         struct lmv_obd *lmv = &obd->u.lmv;
2029         struct lmv_tgt_desc *tgt;
2030         int rc;
2031
2032         ENTRY;
2033
2034         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
2035
2036         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2037         if (IS_ERR(tgt))
2038                 RETURN(PTR_ERR(tgt));
2039
2040         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
2041                PFID(&op_data->op_fid1), tgt->ltd_index);
2042
2043         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
2044                         extra_lock_flags);
2045
2046         RETURN(rc);
2047 }
2048
2049 int
2050 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
2051                  struct ptlrpc_request **preq)
2052 {
2053         struct obd_device *obd = exp->exp_obd;
2054         struct lmv_obd *lmv = &obd->u.lmv;
2055         struct lmv_tgt_desc *tgt;
2056         struct mdt_body *body;
2057         int rc;
2058
2059         ENTRY;
2060
2061 retry:
2062         if (op_data->op_namelen == 2 &&
2063             op_data->op_name[0] == '.' && op_data->op_name[1] == '.')
2064                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2065         else
2066                 tgt = lmv_locate_tgt(lmv, op_data);
2067         if (IS_ERR(tgt))
2068                 RETURN(PTR_ERR(tgt));
2069
2070         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
2071                 (int)op_data->op_namelen, op_data->op_name,
2072                 PFID(&op_data->op_fid1), tgt->ltd_index);
2073
2074         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
2075         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2076                 ptlrpc_req_finished(*preq);
2077                 *preq = NULL;
2078                 goto retry;
2079         }
2080
2081         if (rc)
2082                 RETURN(rc);
2083
2084         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
2085         LASSERT(body != NULL);
2086
2087         if (body->mbo_valid & OBD_MD_MDS) {
2088                 op_data->op_fid1 = body->mbo_fid1;
2089                 op_data->op_valid |= OBD_MD_FLCROSSREF;
2090                 op_data->op_namelen = 0;
2091                 op_data->op_name = NULL;
2092
2093                 ptlrpc_req_finished(*preq);
2094                 *preq = NULL;
2095
2096                 goto retry;
2097         }
2098
2099         RETURN(rc);
2100 }
2101
2102 #define md_op_data_fid(op_data, fl)                     \
2103         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2104          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2105          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2106          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2107          NULL)
2108
2109 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2110                             struct md_op_data *op_data, __u32 op_tgt,
2111                             enum ldlm_mode mode, int bits, int flag)
2112 {
2113         struct lu_fid *fid = md_op_data_fid(op_data, flag);
2114         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
2115         union ldlm_policy_data policy = { { 0 } };
2116         int rc = 0;
2117         ENTRY;
2118
2119         if (!fid_is_sane(fid))
2120                 RETURN(0);
2121
2122         if (tgt == NULL) {
2123                 tgt = lmv_fid2tgt(lmv, fid);
2124                 if (IS_ERR(tgt))
2125                         RETURN(PTR_ERR(tgt));
2126         }
2127
2128         if (tgt->ltd_index != op_tgt) {
2129                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2130                 policy.l_inodebits.bits = bits;
2131                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2132                                       mode, LCF_ASYNC, NULL);
2133         } else {
2134                 CDEBUG(D_INODE,
2135                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2136                        op_tgt, PFID(fid));
2137                 op_data->op_flags |= flag;
2138                 rc = 0;
2139         }
2140
2141         RETURN(rc);
2142 }
2143
2144 /*
2145  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2146  * op_data->op_fid2
2147  */
2148 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2149                     struct ptlrpc_request **request)
2150 {
2151         struct obd_device       *obd = exp->exp_obd;
2152         struct lmv_obd          *lmv = &obd->u.lmv;
2153         struct lmv_tgt_desc     *tgt;
2154         int                      rc;
2155         ENTRY;
2156
2157         LASSERT(op_data->op_namelen != 0);
2158
2159         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2160                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2161                op_data->op_name, PFID(&op_data->op_fid1));
2162
2163         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2164         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2165         op_data->op_cap = current_cap();
2166
2167         tgt = lmv_locate_tgt2(lmv, op_data);
2168         if (IS_ERR(tgt))
2169                 RETURN(PTR_ERR(tgt));
2170
2171         /*
2172          * Cancel UPDATE lock on child (fid1).
2173          */
2174         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2175         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2176                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2177         if (rc != 0)
2178                 RETURN(rc);
2179
2180         rc = md_link(tgt->ltd_exp, op_data, request);
2181
2182         RETURN(rc);
2183 }
2184
2185 /* migrate the top directory */
2186 static inline bool lmv_op_topdir_migrate(const struct md_op_data *op_data)
2187 {
2188         if (!S_ISDIR(op_data->op_mode))
2189                 return false;
2190
2191         if (lmv_dir_layout_changing(op_data->op_mea1))
2192                 return false;
2193
2194         return true;
2195 }
2196
2197 /* migrate top dir to specific MDTs */
2198 static inline bool lmv_topdir_specific_migrate(const struct md_op_data *op_data)
2199 {
2200         const struct lmv_user_md *lum = op_data->op_data;
2201
2202         if (!lmv_op_topdir_migrate(op_data))
2203                 return false;
2204
2205         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2206 }
2207
2208 /* migrate top dir in QoS mode if user issued "lfs migrate -m -1..." */
2209 static inline bool lmv_topdir_qos_migrate(const struct md_op_data *op_data)
2210 {
2211         const struct lmv_user_md *lum = op_data->op_data;
2212
2213         if (!lmv_op_topdir_migrate(op_data))
2214                 return false;
2215
2216         return le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
2217 }
2218
2219 static inline bool lmv_subdir_specific_migrate(const struct md_op_data *op_data)
2220 {
2221         const struct lmv_user_md *lum = op_data->op_data;
2222
2223         if (!S_ISDIR(op_data->op_mode))
2224                 return false;
2225
2226         if (!lmv_dir_layout_changing(op_data->op_mea1))
2227                 return false;
2228
2229         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2230 }
2231
2232 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2233                         const char *name, size_t namelen,
2234                         struct ptlrpc_request **request)
2235 {
2236         struct obd_device *obd = exp->exp_obd;
2237         struct lmv_obd *lmv = &obd->u.lmv;
2238         struct lmv_stripe_md *lsm = op_data->op_mea1;
2239         struct lmv_tgt_desc *parent_tgt;
2240         struct lmv_tgt_desc *sp_tgt;
2241         struct lmv_tgt_desc *tp_tgt = NULL;
2242         struct lmv_tgt_desc *child_tgt;
2243         struct lmv_tgt_desc *tgt;
2244         struct lu_fid target_fid = { 0 };
2245         int rc;
2246
2247         ENTRY;
2248
2249         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2250
2251         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2252                PFID(&op_data->op_fid1), (int)namelen, name);
2253
2254         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2255         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2256         op_data->op_cap = current_cap();
2257
2258         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2259         if (IS_ERR(parent_tgt))
2260                 RETURN(PTR_ERR(parent_tgt));
2261
2262         if (lmv_dir_striped(lsm)) {
2263                 const struct lmv_oinfo *oinfo;
2264
2265                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
2266                 if (IS_ERR(oinfo))
2267                         RETURN(PTR_ERR(oinfo));
2268
2269                 /* save source stripe FID in fid4 temporarily for ELC */
2270                 op_data->op_fid4 = oinfo->lmo_fid;
2271                 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2272                 if (!sp_tgt)
2273                         RETURN(-ENODEV);
2274
2275                 /*
2276                  * if parent is being migrated too, fill op_fid2 with target
2277                  * stripe fid, otherwise the target stripe is not created yet.
2278                  */
2279                 if (lmv_dir_layout_changing(lsm)) {
2280                         oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
2281                                                         true);
2282                         if (IS_ERR(oinfo))
2283                                 RETURN(PTR_ERR(oinfo));
2284
2285                         op_data->op_fid2 = oinfo->lmo_fid;
2286                         tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2287                         if (!tp_tgt)
2288                                 RETURN(-ENODEV);
2289
2290                         /* parent unchanged and update namespace only */
2291                         if (lu_fid_eq(&op_data->op_fid4, &op_data->op_fid2) &&
2292                             op_data->op_bias & MDS_MIGRATE_NSONLY)
2293                                 RETURN(-EALREADY);
2294                 }
2295         } else {
2296                 sp_tgt = parent_tgt;
2297         }
2298
2299         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2300         if (IS_ERR(child_tgt))
2301                 RETURN(PTR_ERR(child_tgt));
2302
2303         if (lmv_topdir_specific_migrate(op_data)) {
2304                 struct lmv_user_md *lum = op_data->op_data;
2305
2306                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2307         } else if (lmv_topdir_qos_migrate(op_data)) {
2308                 tgt = lmv_locate_tgt_lf(lmv);
2309                 if (tgt == ERR_PTR(-EAGAIN))
2310                         tgt = lmv_locate_tgt_rr(lmv);
2311                 if (IS_ERR(tgt))
2312                         RETURN(PTR_ERR(tgt));
2313
2314                 op_data->op_mds = tgt->ltd_index;
2315         } else if (lmv_subdir_specific_migrate(op_data)) {
2316                 struct lmv_user_md *lum = op_data->op_data;
2317                 __u32 i;
2318
2319                 LASSERT(tp_tgt);
2320                 if (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) {
2321                         /* adjust MDTs in lum, since subdir is located on where
2322                          * its parent stripe is, not the first specified MDT.
2323                          */
2324                         for (i = 0; i < le32_to_cpu(lum->lum_stripe_count);
2325                              i++) {
2326                                 if (le32_to_cpu(lum->lum_objects[i].lum_mds) ==
2327                                     tp_tgt->ltd_index)
2328                                         break;
2329                         }
2330
2331                         if (i == le32_to_cpu(lum->lum_stripe_count))
2332                                 RETURN(-ENODEV);
2333
2334                         lum->lum_objects[i].lum_mds =
2335                                 lum->lum_objects[0].lum_mds;
2336                         lum->lum_objects[0].lum_mds =
2337                                 cpu_to_le32(tp_tgt->ltd_index);
2338                 }
2339                 /* NB, the above adjusts subdir migration for command like
2340                  * "lfs migrate -m 0,1,2 ...", but for migration like
2341                  * "lfs migrate -m 0 -c 2 ...", the top dir is migrated to MDT0
2342                  * and MDT1, however its subdir may be migrated to MDT1 and MDT2
2343                  */
2344
2345                 lum->lum_stripe_offset = cpu_to_le32(tp_tgt->ltd_index);
2346                 op_data->op_mds = tp_tgt->ltd_index;
2347         } else if (tp_tgt) {
2348                 op_data->op_mds = tp_tgt->ltd_index;
2349         } else {
2350                 op_data->op_mds = sp_tgt->ltd_index;
2351         }
2352
2353         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2354         if (rc)
2355                 RETURN(rc);
2356
2357         /*
2358          * for directory, send migrate request to the MDT where the object will
2359          * be migrated to, because we can't create a striped directory remotely.
2360          *
2361          * otherwise, send to the MDT where source is located because regular
2362          * file may open lease.
2363          *
2364          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2365          * backward compatibility.
2366          */
2367         if (S_ISDIR(op_data->op_mode) &&
2368             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2369                 tgt = lmv_fid2tgt(lmv, &target_fid);
2370                 if (IS_ERR(tgt))
2371                         RETURN(PTR_ERR(tgt));
2372         } else {
2373                 tgt = child_tgt;
2374         }
2375
2376         /* cancel UPDATE lock of parent master object */
2377         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2378                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2379         if (rc)
2380                 RETURN(rc);
2381
2382         /* cancel UPDATE lock of source parent */
2383         if (sp_tgt != parent_tgt) {
2384                 /*
2385                  * migrate RPC packs master object FID, because we can only pack
2386                  * two FIDs in reint RPC, but MDS needs to know both source
2387                  * parent and target parent, and it will obtain them from master
2388                  * FID and LMV, the other FID in RPC is kept for target.
2389                  *
2390                  * since this FID is not passed to MDC, cancel it anyway.
2391                  */
2392                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2393                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2394                 if (rc)
2395                         RETURN(rc);
2396
2397                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2398         }
2399         op_data->op_fid4 = target_fid;
2400
2401         /* cancel UPDATE locks of target parent */
2402         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2403                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2404         if (rc)
2405                 RETURN(rc);
2406
2407         /* cancel LOOKUP lock of source if source is remote object */
2408         if (child_tgt != sp_tgt) {
2409                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2410                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2411                                       MF_MDC_CANCEL_FID3);
2412                 if (rc)
2413                         RETURN(rc);
2414         }
2415
2416         /* cancel ELC locks of source */
2417         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2418                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2419         if (rc)
2420                 RETURN(rc);
2421
2422         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2423
2424         RETURN(rc);
2425 }
2426
2427 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2428                       const char *old, size_t oldlen,
2429                       const char *new, size_t newlen,
2430                       struct ptlrpc_request **request)
2431 {
2432         struct obd_device *obd = exp->exp_obd;
2433         struct lmv_obd *lmv = &obd->u.lmv;
2434         struct lmv_tgt_desc *sp_tgt;
2435         struct lmv_tgt_desc *tp_tgt = NULL;
2436         struct lmv_tgt_desc *src_tgt = NULL;
2437         struct lmv_tgt_desc *tgt;
2438         struct mdt_body *body;
2439         int rc;
2440
2441         ENTRY;
2442
2443         LASSERT(oldlen != 0);
2444
2445         if (op_data->op_cli_flags & CLI_MIGRATE) {
2446                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2447                 RETURN(rc);
2448         }
2449
2450         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2451         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2452         op_data->op_cap = current_cap();
2453
2454         op_data->op_name = new;
2455         op_data->op_namelen = newlen;
2456
2457         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2458         if (IS_ERR(tp_tgt))
2459                 RETURN(PTR_ERR(tp_tgt));
2460
2461         /* Since the target child might be destroyed, and it might become
2462          * orphan, and we can only check orphan on the local MDT right now, so
2463          * we send rename request to the MDT where target child is located. If
2464          * target child does not exist, then it will send the request to the
2465          * target parent */
2466         if (fid_is_sane(&op_data->op_fid4)) {
2467                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2468                 if (IS_ERR(tgt))
2469                         RETURN(PTR_ERR(tgt));
2470         } else {
2471                 tgt = tp_tgt;
2472         }
2473
2474         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2475
2476         /* cancel UPDATE locks of target parent */
2477         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2478                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2479         if (rc != 0)
2480                 RETURN(rc);
2481
2482         if (fid_is_sane(&op_data->op_fid4)) {
2483                 /* cancel LOOKUP lock of target on target parent */
2484                 if (tgt != tp_tgt) {
2485                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2486                                               tgt->ltd_index, LCK_EX,
2487                                               MDS_INODELOCK_LOOKUP,
2488                                               MF_MDC_CANCEL_FID4);
2489                         if (rc != 0)
2490                                 RETURN(rc);
2491                 }
2492         }
2493
2494         if (fid_is_sane(&op_data->op_fid3)) {
2495                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2496                 if (IS_ERR(src_tgt))
2497                         RETURN(PTR_ERR(src_tgt));
2498
2499                 /* cancel ELC locks of source */
2500                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2501                                       LCK_EX, MDS_INODELOCK_ELC,
2502                                       MF_MDC_CANCEL_FID3);
2503                 if (rc != 0)
2504                         RETURN(rc);
2505         }
2506
2507         op_data->op_name = old;
2508         op_data->op_namelen = oldlen;
2509 retry:
2510         sp_tgt = lmv_locate_tgt(lmv, op_data);
2511         if (IS_ERR(sp_tgt))
2512                 RETURN(PTR_ERR(sp_tgt));
2513
2514         /* cancel UPDATE locks of source parent */
2515         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2516                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2517         if (rc != 0)
2518                 RETURN(rc);
2519
2520         if (fid_is_sane(&op_data->op_fid3)) {
2521                 /* cancel LOOKUP lock of source on source parent */
2522                 if (src_tgt != sp_tgt) {
2523                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2524                                               tgt->ltd_index, LCK_EX,
2525                                               MDS_INODELOCK_LOOKUP,
2526                                               MF_MDC_CANCEL_FID3);
2527                         if (rc != 0)
2528                                 RETURN(rc);
2529                 }
2530         }
2531
2532 rename:
2533         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2534                 PFID(&op_data->op_fid1), (int)oldlen, old,
2535                 PFID(&op_data->op_fid2), (int)newlen, new);
2536
2537         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2538                         request);
2539         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2540                 ptlrpc_req_finished(*request);
2541                 *request = NULL;
2542                 goto retry;
2543         }
2544
2545         if (rc && rc != -EXDEV)
2546                 RETURN(rc);
2547
2548         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2549         if (body == NULL)
2550                 RETURN(-EPROTO);
2551
2552         /* Not cross-ref case, just get out of here. */
2553         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2554                 RETURN(rc);
2555
2556         op_data->op_fid4 = body->mbo_fid1;
2557
2558         ptlrpc_req_finished(*request);
2559         *request = NULL;
2560
2561         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2562         if (IS_ERR(tgt))
2563                 RETURN(PTR_ERR(tgt));
2564
2565         if (fid_is_sane(&op_data->op_fid4)) {
2566                 /* cancel LOOKUP lock of target on target parent */
2567                 if (tgt != tp_tgt) {
2568                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2569                                               tgt->ltd_index, LCK_EX,
2570                                               MDS_INODELOCK_LOOKUP,
2571                                               MF_MDC_CANCEL_FID4);
2572                         if (rc != 0)
2573                                 RETURN(rc);
2574                 }
2575         }
2576
2577         goto rename;
2578 }
2579
2580 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2581                        void *ea, size_t ealen, struct ptlrpc_request **request)
2582 {
2583         struct obd_device *obd = exp->exp_obd;
2584         struct lmv_obd *lmv = &obd->u.lmv;
2585         struct lmv_tgt_desc *tgt;
2586         int rc = 0;
2587
2588         ENTRY;
2589
2590         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2591                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2592                op_data->op_xvalid);
2593
2594         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2595         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2596         if (IS_ERR(tgt))
2597                 RETURN(PTR_ERR(tgt));
2598
2599         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2600
2601         RETURN(rc);
2602 }
2603
2604 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2605                      struct ptlrpc_request **request)
2606 {
2607         struct obd_device *obd = exp->exp_obd;
2608         struct lmv_obd *lmv = &obd->u.lmv;
2609         struct lmv_tgt_desc *tgt;
2610         int rc;
2611
2612         ENTRY;
2613
2614         tgt = lmv_fid2tgt(lmv, fid);
2615         if (IS_ERR(tgt))
2616                 RETURN(PTR_ERR(tgt));
2617
2618         rc = md_fsync(tgt->ltd_exp, fid, request);
2619         RETURN(rc);
2620 }
2621
2622 struct stripe_dirent {
2623         struct page             *sd_page;
2624         struct lu_dirpage       *sd_dp;
2625         struct lu_dirent        *sd_ent;
2626         bool                     sd_eof;
2627 };
2628
2629 struct lmv_dir_ctxt {
2630         struct lmv_obd          *ldc_lmv;
2631         struct md_op_data       *ldc_op_data;
2632         struct md_readdir_info  *ldc_mrinfo;
2633         __u64                    ldc_hash;
2634         int                      ldc_count;
2635         struct stripe_dirent     ldc_stripes[0];
2636 };
2637
2638 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2639 {
2640         if (stripe->sd_page) {
2641                 kunmap(stripe->sd_page);
2642                 put_page(stripe->sd_page);
2643                 stripe->sd_page = NULL;
2644                 stripe->sd_ent = NULL;
2645         }
2646 }
2647
2648 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2649 {
2650         int i;
2651
2652         for (i = 0; i < ctxt->ldc_count; i++)
2653                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2654 }
2655
2656 /* if @ent is dummy, or . .., get next */
2657 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2658                                            struct lu_dirent *ent,
2659                                            int stripe_index)
2660 {
2661         for (; ent; ent = lu_dirent_next(ent)) {
2662                 /* Skip dummy entry */
2663                 if (le16_to_cpu(ent->lde_namelen) == 0)
2664                         continue;
2665
2666                 /* skip . and .. for other stripes */
2667                 if (stripe_index &&
2668                     (strncmp(ent->lde_name, ".",
2669                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2670                      strncmp(ent->lde_name, "..",
2671                              le16_to_cpu(ent->lde_namelen)) == 0))
2672                         continue;
2673
2674                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2675                         break;
2676         }
2677
2678         return ent;
2679 }
2680
2681 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2682                                             struct stripe_dirent *stripe,
2683                                             int stripe_index)
2684 {
2685         struct md_op_data *op_data = ctxt->ldc_op_data;
2686         struct lmv_oinfo *oinfo;
2687         struct lu_fid fid = op_data->op_fid1;
2688         struct inode *inode = op_data->op_data;
2689         struct lmv_tgt_desc *tgt;
2690         struct lu_dirent *ent = stripe->sd_ent;
2691         __u64 hash = ctxt->ldc_hash;
2692         int rc = 0;
2693
2694         ENTRY;
2695
2696         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2697         LASSERT(!ent);
2698
2699         do {
2700                 if (stripe->sd_page) {
2701                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2702
2703                         /* @hash should be the last dirent hash */
2704                         LASSERTF(hash <= end,
2705                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2706                                  ctxt, stripe, hash, end);
2707                         /* unload last page */
2708                         stripe_dirent_unload(stripe);
2709                         /* eof */
2710                         if (end == MDS_DIR_END_OFF) {
2711                                 stripe->sd_eof = true;
2712                                 break;
2713                         }
2714                         hash = end;
2715                 }
2716
2717                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2718                 if (!oinfo->lmo_root) {
2719                         rc = -ENOENT;
2720                         break;
2721                 }
2722
2723                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2724                 if (!tgt) {
2725                         rc = -ENODEV;
2726                         break;
2727                 }
2728
2729                 /* op_data is shared by stripes, reset after use */
2730                 op_data->op_fid1 = oinfo->lmo_fid;
2731                 op_data->op_fid2 = oinfo->lmo_fid;
2732                 op_data->op_data = oinfo->lmo_root;
2733
2734                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_mrinfo, hash,
2735                                   &stripe->sd_page);
2736
2737                 op_data->op_fid1 = fid;
2738                 op_data->op_fid2 = fid;
2739                 op_data->op_data = inode;
2740
2741                 if (rc)
2742                         break;
2743
2744                 stripe->sd_dp = page_address(stripe->sd_page);
2745                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2746                                         stripe_index);
2747                 /* in case a page filled with ., .. and dummy, read next */
2748         } while (!ent);
2749
2750         stripe->sd_ent = ent;
2751         if (rc) {
2752                 LASSERT(!ent);
2753                 /* treat error as eof, so dir can be partially accessed */
2754                 stripe->sd_eof = true;
2755                 ctxt->ldc_mrinfo->mr_partial_readdir_rc = rc;
2756                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2757                               "directory is partially accessed!\n",
2758                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2759                               rc);
2760         }
2761
2762         RETURN(ent);
2763 }
2764
2765 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2766 {
2767         struct obd_device *obd = exp->exp_obd;
2768         struct lmv_obd *lmv = &obd->u.lmv;
2769         struct lmv_tgt_desc *tgt;
2770         int rc;
2771
2772         ENTRY;
2773
2774         rc = lmv_check_connect(obd);
2775         if (rc != 0)
2776                 RETURN(rc);
2777
2778         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2779         if (IS_ERR(tgt))
2780                 RETURN(PTR_ERR(tgt));
2781
2782         data->op_flags |= MF_MDC_CANCEL_FID1;
2783         rc = md_file_resync(tgt->ltd_exp, data);
2784         RETURN(rc);
2785 }
2786
2787 /**
2788  * Get dirent with the closest hash for striped directory
2789  *
2790  * This function will search the dir entry, whose hash value is the
2791  * closest(>=) to hash from all of sub-stripes, and it is only being called
2792  * for striped directory.
2793  *
2794  * \param[in] ctxt              dir read context
2795  *
2796  * \retval                      dirent get the entry successfully
2797  *                              NULL does not get the entry, normally it means
2798  *                              it reaches the end of the directory, while read
2799  *                              stripe dirent error is ignored to allow partial
2800  *                              access.
2801  */
2802 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2803 {
2804         struct stripe_dirent *stripe;
2805         struct lu_dirent *ent = NULL;
2806         int i;
2807         int min = -1;
2808
2809         /* TODO: optimize with k-way merge sort */
2810         for (i = 0; i < ctxt->ldc_count; i++) {
2811                 stripe = &ctxt->ldc_stripes[i];
2812                 if (stripe->sd_eof)
2813                         continue;
2814
2815                 if (!stripe->sd_ent) {
2816                         stripe_dirent_load(ctxt, stripe, i);
2817                         if (!stripe->sd_ent) {
2818                                 LASSERT(stripe->sd_eof);
2819                                 continue;
2820                         }
2821                 }
2822
2823                 if (min == -1 ||
2824                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2825                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2826                         min = i;
2827                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2828                             ctxt->ldc_hash)
2829                                 break;
2830                 }
2831         }
2832
2833         if (min != -1) {
2834                 stripe = &ctxt->ldc_stripes[min];
2835                 ent = stripe->sd_ent;
2836                 /* pop found dirent */
2837                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2838                                                    min);
2839         }
2840
2841         return ent;
2842 }
2843
2844 /**
2845  * Build dir entry page for striped directory
2846  *
2847  * This function gets one entry by @offset from a striped directory. It will
2848  * read entries from all of stripes, and choose one closest to the required
2849  * offset(&offset). A few notes
2850  * 1. skip . and .. for non-zero stripes, because there can only have one .
2851  * and .. in a directory.
2852  * 2. op_data will be shared by all of stripes, instead of allocating new
2853  * one, so need to restore before reusing.
2854  *
2855  * \param[in] exp       obd export refer to LMV
2856  * \param[in] op_data   hold those MD parameters of read_entry
2857  * \param[in] mrinfo    ldlm callback being used in enqueue in mdc_read_entry,
2858  *                      and partial readdir result will be stored in it.
2859  * \param[in] offset    starting hash offset
2860  * \param[out] ppage    the page holding the entry. Note: because the entry
2861  *                      will be accessed in upper layer, so we need hold the
2862  *                      page until the usages of entry is finished, see
2863  *                      ll_dir_entry_next.
2864  *
2865  * retval               =0 if get entry successfully
2866  *                      <0 cannot get entry
2867  */
2868 static int lmv_striped_read_page(struct obd_export *exp,
2869                                  struct md_op_data *op_data,
2870                                  struct md_readdir_info *mrinfo, __u64 offset,
2871                                  struct page **ppage)
2872 {
2873         struct page *page = NULL;
2874         struct lu_dirpage *dp;
2875         void *start;
2876         struct lu_dirent *ent;
2877         struct lu_dirent *last_ent;
2878         int stripe_count;
2879         struct lmv_dir_ctxt *ctxt;
2880         struct lu_dirent *next = NULL;
2881         __u16 ent_size;
2882         size_t left_bytes;
2883         int rc = 0;
2884         ENTRY;
2885
2886         /* Allocate a page and read entries from all of stripes and fill
2887          * the page by hash order */
2888         page = alloc_page(GFP_KERNEL);
2889         if (!page)
2890                 RETURN(-ENOMEM);
2891
2892         /* Initialize the entry page */
2893         dp = kmap(page);
2894         memset(dp, 0, sizeof(*dp));
2895         dp->ldp_hash_start = cpu_to_le64(offset);
2896
2897         start = dp + 1;
2898         left_bytes = PAGE_SIZE - sizeof(*dp);
2899         ent = start;
2900         last_ent = ent;
2901
2902         /* initalize dir read context */
2903         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2904         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2905         if (!ctxt)
2906                 GOTO(free_page, rc = -ENOMEM);
2907         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2908         ctxt->ldc_op_data = op_data;
2909         ctxt->ldc_mrinfo = mrinfo;
2910         ctxt->ldc_hash = offset;
2911         ctxt->ldc_count = stripe_count;
2912
2913         while (1) {
2914                 next = lmv_dirent_next(ctxt);
2915
2916                 /* end of directory */
2917                 if (!next) {
2918                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2919                         break;
2920                 }
2921                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2922
2923                 ent_size = le16_to_cpu(next->lde_reclen);
2924
2925                 /* the last entry lde_reclen is 0, but it might not be the last
2926                  * one of this temporay dir page */
2927                 if (!ent_size)
2928                         ent_size = lu_dirent_calc_size(
2929                                         le16_to_cpu(next->lde_namelen),
2930                                         le32_to_cpu(next->lde_attrs));
2931                 /* page full */
2932                 if (ent_size > left_bytes)
2933                         break;
2934
2935                 memcpy(ent, next, ent_size);
2936
2937                 /* Replace . with master FID and Replace .. with the parent FID
2938                  * of master object */
2939                 if (strncmp(ent->lde_name, ".",
2940                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2941                     le16_to_cpu(ent->lde_namelen) == 1)
2942                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2943                 else if (strncmp(ent->lde_name, "..",
2944                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2945                            le16_to_cpu(ent->lde_namelen) == 2)
2946                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2947
2948                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2949                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2950                        le64_to_cpu(ent->lde_hash));
2951
2952                 left_bytes -= ent_size;
2953                 ent->lde_reclen = cpu_to_le16(ent_size);
2954                 last_ent = ent;
2955                 ent = (void *)ent + ent_size;
2956         };
2957
2958         last_ent->lde_reclen = 0;
2959
2960         if (ent == start)
2961                 dp->ldp_flags |= LDF_EMPTY;
2962         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2963                 dp->ldp_flags |= LDF_COLLIDE;
2964         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2965         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2966
2967         put_lmv_dir_ctxt(ctxt);
2968         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2969
2970         *ppage = page;
2971
2972         RETURN(0);
2973
2974 free_page:
2975         kunmap(page);
2976         __free_page(page);
2977
2978         return rc;
2979 }
2980
2981 static int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2982                          struct md_readdir_info *mrinfo, __u64 offset,
2983                          struct page **ppage)
2984 {
2985         struct obd_device *obd = exp->exp_obd;
2986         struct lmv_obd *lmv = &obd->u.lmv;
2987         struct lmv_tgt_desc *tgt;
2988         int rc;
2989
2990         ENTRY;
2991
2992         if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2993                 RETURN(-ENODATA);
2994
2995         if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2996                 rc = lmv_striped_read_page(exp, op_data, mrinfo, offset, ppage);
2997                 RETURN(rc);
2998         }
2999
3000         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
3001         if (IS_ERR(tgt))
3002                 RETURN(PTR_ERR(tgt));
3003
3004         rc = md_read_page(tgt->ltd_exp, op_data, mrinfo, offset, ppage);
3005
3006         RETURN(rc);
3007 }
3008
3009 /**
3010  * Unlink a file/directory
3011  *
3012  * Unlink a file or directory under the parent dir. The unlink request
3013  * usually will be sent to the MDT where the child is located, but if
3014  * the client does not have the child FID then request will be sent to the
3015  * MDT where the parent is located.
3016  *
3017  * If the parent is a striped directory then it also needs to locate which
3018  * stripe the name of the child is located, and replace the parent FID
3019  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
3020  * it will walk through all of sub-stripes until the child is being
3021  * unlinked finally.
3022  *
3023  * \param[in] exp       export refer to LMV
3024  * \param[in] op_data   different parameters transferred beween client
3025  *                      MD stacks, name, namelen, FIDs etc.
3026  *                      op_fid1 is the parent FID, op_fid2 is the child
3027  *                      FID.
3028  * \param[out] request  point to the request of unlink.
3029  *
3030  * retval               0 if succeed
3031  *                      negative errno if failed.
3032  */
3033 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
3034                       struct ptlrpc_request **request)
3035 {
3036         struct obd_device *obd = exp->exp_obd;
3037         struct lmv_obd *lmv = &obd->u.lmv;
3038         struct lmv_tgt_desc *tgt;
3039         struct lmv_tgt_desc *parent_tgt;
3040         struct mdt_body *body;
3041         int rc;
3042
3043         ENTRY;
3044
3045         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
3046         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
3047         op_data->op_cap = current_cap();
3048
3049 retry:
3050         parent_tgt = lmv_locate_tgt(lmv, op_data);
3051         if (IS_ERR(parent_tgt))
3052                 RETURN(PTR_ERR(parent_tgt));
3053
3054         if (likely(!fid_is_zero(&op_data->op_fid2))) {
3055                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3056                 if (IS_ERR(tgt))
3057                         RETURN(PTR_ERR(tgt));
3058         } else {
3059                 tgt = parent_tgt;
3060         }
3061
3062         /*
3063          * If child's fid is given, cancel unused locks for it if it is from
3064          * another export than parent.
3065          *
3066          * LOOKUP lock for child (fid3) should also be cancelled on parent
3067          * tgt_tgt in mdc_unlink().
3068          */
3069         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
3070
3071         if (parent_tgt != tgt)
3072                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
3073                                       LCK_EX, MDS_INODELOCK_LOOKUP,
3074                                       MF_MDC_CANCEL_FID3);
3075
3076         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
3077                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
3078         if (rc)
3079                 RETURN(rc);
3080
3081         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
3082                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
3083                tgt->ltd_index);
3084
3085         rc = md_unlink(tgt->ltd_exp, op_data, request);
3086         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
3087                 ptlrpc_req_finished(*request);
3088                 *request = NULL;
3089                 goto retry;
3090         }
3091
3092         if (rc != -EREMOTE)
3093                 RETURN(rc);
3094
3095         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
3096         if (body == NULL)
3097                 RETURN(-EPROTO);
3098
3099         /* Not cross-ref case, just get out of here. */
3100         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
3101                 RETURN(rc);
3102
3103         /* This is a remote object, try remote MDT. */
3104         op_data->op_fid2 = body->mbo_fid1;
3105         ptlrpc_req_finished(*request);
3106         *request = NULL;
3107
3108         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3109         if (IS_ERR(tgt))
3110                 RETURN(PTR_ERR(tgt));
3111
3112         goto retry;
3113 }
3114
3115 static int lmv_precleanup(struct obd_device *obd)
3116 {
3117         ENTRY;
3118         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
3119         fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
3120         lprocfs_obd_cleanup(obd);
3121         lprocfs_free_md_stats(obd);
3122         RETURN(0);
3123 }
3124
3125 /**
3126  * Get by key a value associated with a LMV device.
3127  *
3128  * Dispatch request to lower-layer devices as needed.
3129  *
3130  * \param[in] env               execution environment for this thread
3131  * \param[in] exp               export for the LMV device
3132  * \param[in] keylen            length of key identifier
3133  * \param[in] key               identifier of key to get value for
3134  * \param[in] vallen            size of \a val
3135  * \param[out] val              pointer to storage location for value
3136  * \param[in] lsm               optional striping metadata of object
3137  *
3138  * \retval 0            on success
3139  * \retval negative     negated errno on failure
3140  */
3141 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
3142                         __u32 keylen, void *key, __u32 *vallen, void *val)
3143 {
3144         struct obd_device *obd;
3145         struct lmv_obd *lmv;
3146         struct lu_tgt_desc *tgt;
3147         int rc = 0;
3148
3149         ENTRY;
3150
3151         obd = class_exp2obd(exp);
3152         if (obd == NULL) {
3153                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3154                        exp->exp_handle.h_cookie);
3155                 RETURN(-EINVAL);
3156         }
3157
3158         lmv = &obd->u.lmv;
3159         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
3160                 LASSERT(*vallen == sizeof(__u32));
3161                 lmv_foreach_connected_tgt(lmv, tgt) {
3162                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
3163                                           vallen, val))
3164                                 RETURN(0);
3165                 }
3166                 RETURN(-EINVAL);
3167         } else if (KEY_IS(KEY_MAX_EASIZE) ||
3168                    KEY_IS(KEY_DEFAULT_EASIZE) ||
3169                    KEY_IS(KEY_CONN_DATA)) {
3170                 /*
3171                  * Forwarding this request to first MDS, it should know LOV
3172                  * desc.
3173                  */
3174                 tgt = lmv_tgt(lmv, 0);
3175                 if (!tgt)
3176                         RETURN(-ENODEV);
3177
3178                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
3179                 if (!rc && KEY_IS(KEY_CONN_DATA))
3180                         exp->exp_connect_data = *(struct obd_connect_data *)val;
3181                 RETURN(rc);
3182         } else if (KEY_IS(KEY_TGT_COUNT)) {
3183                 *((int *)val) = lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count;
3184                 RETURN(0);
3185         }
3186
3187         CDEBUG(D_IOCTL, "Invalid key\n");
3188         RETURN(-EINVAL);
3189 }
3190
3191 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
3192                      int *__rcs, struct ptlrpc_request_set *_set)
3193 {
3194         struct obd_device *obd = class_exp2obd(exp);
3195         struct ptlrpc_request_set *set = _set;
3196         struct lmv_obd *lmv = &obd->u.lmv;
3197         int tgt_count = lmv->lmv_mdt_count;
3198         struct lu_tgt_desc *tgt;
3199         struct fid_array *fat, **fas = NULL;
3200         int i, rc, **rcs = NULL;
3201
3202         if (!set) {
3203                 set = ptlrpc_prep_set();
3204                 if (!set)
3205                         RETURN(-ENOMEM);
3206         }
3207
3208         /* split FIDs by targets */
3209         OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
3210         if (fas == NULL)
3211                 GOTO(out, rc = -ENOMEM);
3212         OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
3213         if (rcs == NULL)
3214                 GOTO(out_fas, rc = -ENOMEM);
3215
3216         for (i = 0; i < fa->fa_nr; i++) {
3217                 unsigned int idx;
3218
3219                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
3220                 if (rc) {
3221                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
3222                                PFID(&fa->fa_fids[i]), rc);
3223                         continue;
3224                 }
3225                 LASSERT(idx < tgt_count);
3226                 if (!fas[idx])
3227                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
3228                                   fa_fids[fa->fa_nr]));
3229                 if (!fas[idx])
3230                         GOTO(out, rc = -ENOMEM);
3231                 if (!rcs[idx])
3232                         OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
3233                 if (!rcs[idx])
3234                         GOTO(out, rc = -ENOMEM);
3235
3236                 fat = fas[idx];
3237                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
3238         }
3239
3240         lmv_foreach_connected_tgt(lmv, tgt) {
3241                 fat = fas[tgt->ltd_index];
3242                 if (!fat || fat->fa_nr == 0)
3243                         continue;
3244                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3245         }
3246
3247         rc = ptlrpc_set_wait(NULL, set);
3248         if (rc == 0) {
3249                 int j = 0;
3250                 for (i = 0; i < tgt_count; i++) {
3251                         fat = fas[i];
3252                         if (!fat || fat->fa_nr == 0)
3253                                 continue;
3254                         /* copy FIDs back */
3255                         memcpy(fa->fa_fids + j, fat->fa_fids,
3256                                fat->fa_nr * sizeof(struct lu_fid));
3257                         /* copy rcs back */
3258                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3259                         j += fat->fa_nr;
3260                 }
3261         }
3262         if (set != _set)
3263                 ptlrpc_set_destroy(set);
3264
3265 out:
3266         for (i = 0; i < tgt_count; i++) {
3267                 if (fas && fas[i])
3268                         OBD_FREE(fas[i], offsetof(struct fid_array,
3269                                                 fa_fids[fa->fa_nr]));
3270                 if (rcs && rcs[i])
3271                         OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3272         }
3273         if (rcs)
3274                 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3275 out_fas:
3276         if (fas)
3277                 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3278
3279         RETURN(rc);
3280 }
3281
3282 /**
3283  * Asynchronously set by key a value associated with a LMV device.
3284  *
3285  * Dispatch request to lower-layer devices as needed.
3286  *
3287  * \param[in] env       execution environment for this thread
3288  * \param[in] exp       export for the LMV device
3289  * \param[in] keylen    length of key identifier
3290  * \param[in] key       identifier of key to store value for
3291  * \param[in] vallen    size of value to store
3292  * \param[in] val       pointer to data to be stored
3293  * \param[in] set       optional list of related ptlrpc requests
3294  *
3295  * \retval 0            on success
3296  * \retval negative     negated errno on failure
3297  */
3298 static int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3299                               __u32 keylen, void *key, __u32 vallen, void *val,
3300                               struct ptlrpc_request_set *set)
3301 {
3302         struct lmv_tgt_desc *tgt;
3303         struct obd_device *obd;
3304         struct lmv_obd *lmv;
3305         int rc = 0;
3306         ENTRY;
3307
3308         obd = class_exp2obd(exp);
3309         if (obd == NULL) {
3310                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3311                        exp->exp_handle.h_cookie);
3312                 RETURN(-EINVAL);
3313         }
3314         lmv = &obd->u.lmv;
3315
3316         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3317             KEY_IS(KEY_DEFAULT_EASIZE)) {
3318                 int err = 0;
3319
3320                 lmv_foreach_connected_tgt(lmv, tgt) {
3321                         err = obd_set_info_async(env, tgt->ltd_exp,
3322                                                  keylen, key, vallen, val, set);
3323                         if (err && rc == 0)
3324                                 rc = err;
3325                 }
3326
3327                 RETURN(rc);
3328         }
3329
3330         RETURN(-EINVAL);
3331 }
3332
3333 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3334                             const struct lmv_mds_md_v1 *lmm1)
3335 {
3336         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3337         int             stripe_count;
3338         int             cplen;
3339         int             i;
3340         int             rc = 0;
3341         ENTRY;
3342
3343         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3344         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3345         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3346         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3347                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3348         else
3349                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3350         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3351         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3352         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3353         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3354                         sizeof(lsm->lsm_md_pool_name));
3355
3356         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3357                 RETURN(-E2BIG);
3358
3359         CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3360                "layout_version %d\n", lsm->lsm_md_stripe_count,
3361                lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3362                lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3363                lsm->lsm_md_layout_version);
3364
3365         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3366         for (i = 0; i < stripe_count; i++) {
3367                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3368                               &lmm1->lmv_stripe_fids[i]);
3369                 /*
3370                  * set default value -1, so lmv_locate_tgt() knows this stripe
3371                  * target is not initialized.
3372                  */
3373                 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3374                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3375                         continue;
3376
3377                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3378                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3379                 if (rc == -ENOENT)
3380                         continue;
3381
3382                 if (rc)
3383                         RETURN(rc);
3384
3385                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3386                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3387         }
3388
3389         RETURN(rc);
3390 }
3391
3392 static inline int lmv_unpack_user_md(struct obd_export *exp,
3393                                      struct lmv_stripe_md *lsm,
3394                                      const struct lmv_user_md *lmu)
3395 {
3396         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3397         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3398         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3399         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3400         lsm->lsm_md_max_inherit = lmu->lum_max_inherit;
3401         lsm->lsm_md_max_inherit_rr = lmu->lum_max_inherit_rr;
3402         lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3403
3404         return 0;
3405 }
3406
3407 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3408                         const union lmv_mds_md *lmm, size_t lmm_size)
3409 {
3410         struct lmv_stripe_md     *lsm;
3411         int                      lsm_size;
3412         int                      rc;
3413         bool                     allocated = false;
3414         ENTRY;
3415
3416         LASSERT(lsmp != NULL);
3417
3418         lsm = *lsmp;
3419         /* Free memmd */
3420         if (lsm != NULL && lmm == NULL) {
3421                 int i;
3422                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3423
3424                 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3425                         size_t lfm_size;
3426
3427                         lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3428                                                               lfm_value[0]);
3429                         OBD_FREE_LARGE(lfm, lfm_size);
3430                         RETURN(0);
3431                 }
3432
3433                 if (lmv_dir_striped(lsm)) {
3434                         for (i = 0; i < lsm->lsm_md_stripe_count; i++)
3435                                 iput(lsm->lsm_md_oinfo[i].lmo_root);
3436                         lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3437                 } else {
3438                         lsm_size = lmv_stripe_md_size(0);
3439                 }
3440                 OBD_FREE(lsm, lsm_size);
3441                 *lsmp = NULL;
3442                 RETURN(0);
3443         }
3444
3445         /* foreign lmv case */
3446         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3447                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3448
3449                 if (lfm == NULL) {
3450                         OBD_ALLOC_LARGE(lfm, lmm_size);
3451                         if (lfm == NULL)
3452                                 RETURN(-ENOMEM);
3453                         *lsmp = (struct lmv_stripe_md *)lfm;
3454                 }
3455                 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3456                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3457                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3458                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3459                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3460                        lfm->lfm_length);
3461                 RETURN(lmm_size);
3462         }
3463
3464         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3465                 RETURN(-EPERM);
3466
3467         /* Unpack memmd */
3468         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3469             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3470                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3471                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3472                        -EIO);
3473                 RETURN(-EIO);
3474         }
3475
3476         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3477                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3478         else
3479                 /**
3480                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3481                  * stripecount should be 0 then.
3482                  */
3483                 lsm_size = lmv_stripe_md_size(0);
3484
3485         if (lsm == NULL) {
3486                 OBD_ALLOC(lsm, lsm_size);
3487                 if (lsm == NULL)
3488                         RETURN(-ENOMEM);
3489                 allocated = true;
3490                 *lsmp = lsm;
3491         }
3492
3493         switch (le32_to_cpu(lmm->lmv_magic)) {
3494         case LMV_MAGIC_V1:
3495                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3496                 break;
3497         case LMV_USER_MAGIC:
3498                 rc = lmv_unpack_user_md(exp, lsm, &lmm->lmv_user_md);
3499                 break;
3500         default:
3501                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3502                        le32_to_cpu(lmm->lmv_magic));
3503                 rc = -EINVAL;
3504                 break;
3505         }
3506
3507         if (rc != 0 && allocated) {
3508                 OBD_FREE(lsm, lsm_size);
3509                 *lsmp = NULL;
3510                 lsm_size = rc;
3511         }
3512         RETURN(lsm_size);
3513 }
3514
3515 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3516 {
3517         lmv_unpackmd(NULL, &lsm, NULL, 0);
3518 }
3519 EXPORT_SYMBOL(lmv_free_memmd);
3520
3521 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3522                              union ldlm_policy_data *policy,
3523                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3524                              void *opaque)
3525 {
3526         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3527         struct lu_tgt_desc *tgt;
3528         int err;
3529         int rc = 0;
3530
3531         ENTRY;
3532
3533         LASSERT(fid != NULL);
3534
3535         lmv_foreach_connected_tgt(lmv, tgt) {
3536                 if (!tgt->ltd_active)
3537                         continue;
3538
3539                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3540                                        opaque);
3541                 if (!rc)
3542                         rc = err;
3543         }
3544         RETURN(rc);
3545 }
3546
3547 static int lmv_set_lock_data(struct obd_export *exp,
3548                              const struct lustre_handle *lockh,
3549                              void *data, __u64 *bits)
3550 {
3551         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3552         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3553         int rc;
3554
3555         ENTRY;
3556
3557         if (tgt == NULL || tgt->ltd_exp == NULL)
3558                 RETURN(-EINVAL);
3559         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3560         RETURN(rc);
3561 }
3562
3563 static enum ldlm_mode
3564 lmv_lock_match(struct obd_export *exp, __u64 flags,
3565                const struct lu_fid *fid, enum ldlm_type type,
3566                union ldlm_policy_data *policy,
3567                enum ldlm_mode mode, struct lustre_handle *lockh)
3568 {
3569         struct obd_device *obd = exp->exp_obd;
3570         struct lmv_obd *lmv = &obd->u.lmv;
3571         enum ldlm_mode rc;
3572         struct lu_tgt_desc *tgt;
3573         int i;
3574         int index;
3575
3576         ENTRY;
3577
3578         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
3579
3580         /*
3581          * With DNE every object can have two locks in different namespaces:
3582          * lookup lock in space of MDT storing direntry and update/open lock in
3583          * space of MDT storing inode.  Try the MDT that the FID maps to first,
3584          * since this can be easily found, and only try others if that fails.
3585          */
3586         for (i = 0, index = lmv_fid2tgt_index(lmv, fid);
3587              i < lmv->lmv_mdt_descs.ltd_tgts_size;
3588              i++, index = (index + 1) % lmv->lmv_mdt_descs.ltd_tgts_size) {
3589                 if (index < 0) {
3590                         CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
3591                                obd->obd_name, PFID(fid), index);
3592                         index = 0;
3593                 }
3594
3595                 tgt = lmv_tgt(lmv, index);
3596                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
3597                         continue;
3598
3599                 rc = md_lock_match(tgt->ltd_exp, flags, fid, type, policy, mode,
3600                                    lockh);
3601                 if (rc)
3602                         RETURN(rc);
3603         }
3604
3605         RETURN(0);
3606 }
3607
3608 static int
3609 lmv_get_lustre_md(struct obd_export *exp, struct req_capsule *pill,
3610                   struct obd_export *dt_exp, struct obd_export *md_exp,
3611                   struct lustre_md *md)
3612 {
3613         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3614         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3615
3616         if (!tgt || !tgt->ltd_exp)
3617                 return -EINVAL;
3618
3619         return md_get_lustre_md(tgt->ltd_exp, pill, dt_exp, md_exp, md);
3620 }
3621
3622 static int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
3623 {
3624         struct obd_device *obd = exp->exp_obd;
3625         struct lmv_obd *lmv = &obd->u.lmv;
3626         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3627
3628         ENTRY;
3629
3630         if (md->default_lmv) {
3631                 lmv_free_memmd(md->default_lmv);
3632                 md->default_lmv = NULL;
3633         }
3634         if (md->lmv != NULL) {
3635                 lmv_free_memmd(md->lmv);
3636                 md->lmv = NULL;
3637         }
3638         if (!tgt || !tgt->ltd_exp)
3639                 RETURN(-EINVAL);
3640         RETURN(md_free_lustre_md(tgt->ltd_exp, md));
3641 }
3642
3643 static int lmv_set_open_replay_data(struct obd_export *exp,
3644                                     struct obd_client_handle *och,
3645                                     struct lookup_intent *it)
3646 {
3647         struct obd_device *obd = exp->exp_obd;
3648         struct lmv_obd *lmv = &obd->u.lmv;
3649         struct lmv_tgt_desc *tgt;
3650
3651         ENTRY;
3652
3653         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3654         if (IS_ERR(tgt))
3655                 RETURN(PTR_ERR(tgt));
3656
3657         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3658 }
3659
3660 static int lmv_clear_open_replay_data(struct obd_export *exp,
3661                                       struct obd_client_handle *och)
3662 {
3663         struct obd_device *obd = exp->exp_obd;
3664         struct lmv_obd *lmv = &obd->u.lmv;
3665         struct lmv_tgt_desc *tgt;
3666
3667         ENTRY;
3668
3669         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3670         if (IS_ERR(tgt))
3671                 RETURN(PTR_ERR(tgt));
3672
3673         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3674 }
3675
3676 static int lmv_intent_getattr_async(struct obd_export *exp,
3677                                     struct md_enqueue_info *minfo)
3678 {
3679         struct md_op_data *op_data = &minfo->mi_data;
3680         struct obd_device *obd = exp->exp_obd;
3681         struct lmv_obd *lmv = &obd->u.lmv;
3682         struct lmv_tgt_desc *ptgt;
3683         struct lmv_tgt_desc *ctgt;
3684         int rc;
3685
3686         ENTRY;
3687
3688         if (!fid_is_sane(&op_data->op_fid2))
3689                 RETURN(-EINVAL);
3690
3691         ptgt = lmv_locate_tgt(lmv, op_data);
3692         if (IS_ERR(ptgt))
3693                 RETURN(PTR_ERR(ptgt));
3694
3695         ctgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3696         if (IS_ERR(ctgt))
3697                 RETURN(PTR_ERR(ctgt));
3698
3699         /*
3700          * remote object needs two RPCs to lookup and getattr, considering the
3701          * complexity don't support statahead for now.
3702          */
3703         if (ctgt != ptgt)
3704                 RETURN(-EREMOTE);
3705
3706         rc = md_intent_getattr_async(ptgt->ltd_exp, minfo);
3707
3708         RETURN(rc);
3709 }
3710
3711 static int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3712                                struct lu_fid *fid, __u64 *bits)
3713 {
3714         struct obd_device *obd = exp->exp_obd;
3715         struct lmv_obd *lmv = &obd->u.lmv;
3716         struct lmv_tgt_desc *tgt;
3717         int rc;
3718
3719         ENTRY;
3720
3721         tgt = lmv_fid2tgt(lmv, fid);
3722         if (IS_ERR(tgt))
3723                 RETURN(PTR_ERR(tgt));
3724
3725         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3726         RETURN(rc);
3727 }
3728
3729 static int lmv_get_fid_from_lsm(struct obd_export *exp,
3730                                 const struct lmv_stripe_md *lsm,
3731                                 const char *name, int namelen,
3732                                 struct lu_fid *fid)
3733 {
3734         const struct lmv_oinfo *oinfo;
3735
3736         if (!lmv_dir_striped(lsm))
3737                 RETURN(-ESTALE);
3738
3739         oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
3740         if (IS_ERR(oinfo))
3741                 return PTR_ERR(oinfo);
3742
3743         *fid = oinfo->lmo_fid;
3744
3745         RETURN(0);
3746 }
3747
3748 /**
3749  * For lmv, only need to send request to master MDT, and the master MDT will
3750  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3751  * we directly fetch data from the slave MDTs.
3752  */
3753 static int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3754                         struct obd_quotactl *oqctl)
3755 {
3756         struct obd_device *obd = class_exp2obd(exp);
3757         struct lmv_obd *lmv = &obd->u.lmv;
3758         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3759         __u64 curspace, curinodes;
3760         int rc = 0;
3761
3762         ENTRY;
3763
3764         if (!tgt || !tgt->ltd_exp || !tgt->ltd_active) {
3765                 CERROR("master lmv inactive\n");
3766                 RETURN(-EIO);
3767         }
3768
3769         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3770                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3771                 RETURN(rc);
3772         }
3773
3774         curspace = curinodes = 0;
3775         lmv_foreach_connected_tgt(lmv, tgt) {
3776                 int err;
3777
3778                 if (!tgt->ltd_active)
3779                         continue;
3780
3781                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3782                 if (err) {
3783                         CERROR("getquota on mdt %d failed. %d\n",
3784                                tgt->ltd_index, err);
3785                         if (!rc)
3786                                 rc = err;
3787                 } else {
3788                         curspace += oqctl->qc_dqblk.dqb_curspace;
3789                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3790                 }
3791         }
3792         oqctl->qc_dqblk.dqb_curspace = curspace;
3793         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3794
3795         RETURN(rc);
3796 }
3797
3798 static int lmv_merge_attr(struct obd_export *exp,
3799                           const struct lmv_stripe_md *lsm,
3800                           struct cl_attr *attr,
3801                           ldlm_blocking_callback cb_blocking)
3802 {
3803         int rc;
3804         int i;
3805
3806         if (!lmv_dir_striped(lsm))
3807                 return 0;
3808
3809         rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3810         if (rc < 0)
3811                 return rc;
3812
3813         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3814                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3815
3816                 if (!inode)
3817                         continue;
3818
3819                 CDEBUG(D_INFO,
3820                        "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
3821                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3822                        i_size_read(inode), (unsigned long long)inode->i_blocks,
3823                        inode->i_nlink, (s64)inode->i_atime.tv_sec,
3824                        (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
3825
3826                 /* for slave stripe, it needs to subtract nlink for . and .. */
3827                 if (i != 0)
3828                         attr->cat_nlink += inode->i_nlink - 2;
3829                 else
3830                         attr->cat_nlink = inode->i_nlink;
3831
3832                 attr->cat_size += i_size_read(inode);
3833                 attr->cat_blocks += inode->i_blocks;
3834
3835                 if (attr->cat_atime < inode->i_atime.tv_sec)
3836                         attr->cat_atime = inode->i_atime.tv_sec;
3837
3838                 if (attr->cat_ctime < inode->i_ctime.tv_sec)
3839                         attr->cat_ctime = inode->i_ctime.tv_sec;
3840
3841                 if (attr->cat_mtime < inode->i_mtime.tv_sec)
3842                         attr->cat_mtime = inode->i_mtime.tv_sec;
3843         }
3844         return 0;
3845 }
3846
3847 static const struct obd_ops lmv_obd_ops = {
3848         .o_owner                = THIS_MODULE,
3849         .o_setup                = lmv_setup,
3850         .o_cleanup              = lmv_cleanup,
3851         .o_precleanup           = lmv_precleanup,
3852         .o_process_config       = lmv_process_config,
3853         .o_connect              = lmv_connect,
3854         .o_disconnect           = lmv_disconnect,
3855         .o_statfs               = lmv_statfs,
3856         .o_get_info             = lmv_get_info,
3857         .o_set_info_async       = lmv_set_info_async,
3858         .o_notify               = lmv_notify,
3859         .o_get_uuid             = lmv_get_uuid,
3860         .o_fid_alloc            = lmv_fid_alloc,
3861         .o_iocontrol            = lmv_iocontrol,
3862         .o_quotactl             = lmv_quotactl
3863 };
3864
3865 static const struct md_ops lmv_md_ops = {
3866         .m_get_root             = lmv_get_root,
3867         .m_null_inode           = lmv_null_inode,
3868         .m_close                = lmv_close,
3869         .m_create               = lmv_create,
3870         .m_enqueue              = lmv_enqueue,
3871         .m_getattr              = lmv_getattr,
3872         .m_getxattr             = lmv_getxattr,
3873         .m_getattr_name         = lmv_getattr_name,
3874         .m_intent_lock          = lmv_intent_lock,
3875         .m_link                 = lmv_link,
3876         .m_rename               = lmv_rename,
3877         .m_setattr              = lmv_setattr,
3878         .m_setxattr             = lmv_setxattr,
3879         .m_fsync                = lmv_fsync,
3880         .m_file_resync          = lmv_file_resync,
3881         .m_read_page            = lmv_read_page,
3882         .m_unlink               = lmv_unlink,
3883         .m_init_ea_size         = lmv_init_ea_size,
3884         .m_cancel_unused        = lmv_cancel_unused,
3885         .m_set_lock_data        = lmv_set_lock_data,
3886         .m_lock_match           = lmv_lock_match,
3887         .m_get_lustre_md        = lmv_get_lustre_md,
3888         .m_free_lustre_md       = lmv_free_lustre_md,
3889         .m_merge_attr           = lmv_merge_attr,
3890         .m_set_open_replay_data = lmv_set_open_replay_data,
3891         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3892         .m_intent_getattr_async = lmv_intent_getattr_async,
3893         .m_revalidate_lock      = lmv_revalidate_lock,
3894         .m_get_fid_from_lsm     = lmv_get_fid_from_lsm,
3895         .m_unpackmd             = lmv_unpackmd,
3896         .m_rmfid                = lmv_rmfid,
3897 };
3898
3899 static int __init lmv_init(void)
3900 {
3901         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true,
3902                                    LUSTRE_LMV_NAME, NULL);
3903 }
3904
3905 static void __exit lmv_exit(void)
3906 {
3907         class_unregister_type(LUSTRE_LMV_NAME);
3908 }
3909
3910 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3911 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
3912 MODULE_VERSION(LUSTRE_VERSION_STRING);
3913 MODULE_LICENSE("GPL");
3914
3915 module_init(lmv_init);
3916 module_exit(lmv_exit);