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 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data);
  60
  61 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
  62                          int activate)
  63 {
  64         if (tgt->ltd_active == activate)
  65                 return;
  66
  67         tgt->ltd_active = activate;
  68         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
  69                 (activate ? 1 : -1);
  70
  71         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
  72 }
  73
  74 /**
  75  * Error codes:
  76  *
  77  *  -EINVAL  : UUID can't be found in the LMV's target list
  78  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  79  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  80  */
  81 static int lmv_set_mdc_active(struct lmv_obd *lmv,
  82                               const struct obd_uuid *uuid,
  83                               int activate)
  84 {
  85         struct lu_tgt_desc *tgt = NULL;
  86         struct obd_device *obd;
  87         int rc = 0;
  88
  89         ENTRY;
  90
  91         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  92                         lmv, uuid->uuid, activate);
  93
  94         spin_lock(&lmv->lmv_lock);
  95         lmv_foreach_connected_tgt(lmv, tgt) {
  96                 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
  97                        tgt->ltd_index, tgt->ltd_uuid.uuid,
  98                        tgt->ltd_exp->exp_handle.h_cookie);
  99
 100                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 101                         break;
 102         }
 103
 104         if (!tgt)
 105                 GOTO(out_lmv_lock, rc = -EINVAL);
 106
 107         obd = class_exp2obd(tgt->ltd_exp);
 108         if (obd == NULL)
 109                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 110
 111         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 112                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 113                obd->obd_type->typ_name, tgt->ltd_index);
 114         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 115
 116         if (tgt->ltd_active == activate) {
 117                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 118                        activate ? "" : "in");
 119                 GOTO(out_lmv_lock, rc);
 120         }
 121
 122         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 123                activate ? "" : "in");
 124         lmv_activate_target(lmv, tgt, activate);
 125         EXIT;
 126
 127  out_lmv_lock:
 128         spin_unlock(&lmv->lmv_lock);
 129         return rc;
 130 }
 131
 132 static struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
 133 {
 134         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
 135         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
 136
 137         return tgt ? obd_get_uuid(tgt->ltd_exp) : NULL;
 138 }
 139
 140 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 141                       enum obd_notify_event ev)
 142 {
 143         struct obd_connect_data *conn_data;
 144         struct lmv_obd          *lmv = &obd->u.lmv;
 145         struct obd_uuid         *uuid;
 146         int                      rc = 0;
 147         ENTRY;
 148
 149         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 150                 CERROR("unexpected notification of %s %s!\n",
 151                        watched->obd_type->typ_name,
 152                        watched->obd_name);
 153                 RETURN(-EINVAL);
 154         }
 155
 156         uuid = &watched->u.cli.cl_target_uuid;
 157         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 158                 /*
 159                  * Set MDC as active before notifying the observer, so the
 160                  * observer can use the MDC normally.
 161                  */
 162                 rc = lmv_set_mdc_active(lmv, uuid,
 163                                         ev == OBD_NOTIFY_ACTIVE);
 164                 if (rc) {
 165                         CERROR("%sactivation of %s failed: %d\n",
 166                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 167                                uuid->uuid, rc);
 168                         RETURN(rc);
 169                 }
 170         } else if (ev == OBD_NOTIFY_OCD) {
 171                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 172                 /*
 173                  * XXX: Make sure that ocd_connect_flags from all targets are
 174                  * the same. Otherwise one of MDTs runs wrong version or
 175                  * something like this.  --umka
 176                  */
 177                 obd->obd_self_export->exp_connect_data = *conn_data;
 178         }
 179
 180         /*
 181          * Pass the notification up the chain.
 182          */
 183         if (obd->obd_observer)
 184                 rc = obd_notify(obd->obd_observer, watched, ev);
 185
 186         RETURN(rc);
 187 }
 188
 189 static int lmv_connect(const struct lu_env *env,
 190                        struct obd_export **pexp, struct obd_device *obd,
 191                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 192                        void *localdata)
 193 {
 194         struct lmv_obd *lmv = &obd->u.lmv;
 195         struct lustre_handle conn = { 0 };
 196         struct obd_export *exp;
 197         int rc;
 198         ENTRY;
 199
 200         rc = class_connect(&conn, obd, cluuid);
 201         if (rc) {
 202                 CERROR("class_connection() returned %d\n", rc);
 203                 RETURN(rc);
 204         }
 205
 206         exp = class_conn2export(&conn);
 207
 208         lmv->connected = 0;
 209         lmv->conn_data = *data;
 210         lmv->lmv_cache = localdata;
 211
 212         lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
 213                                                     &obd->obd_kset.kobj);
 214         if (!lmv->lmv_tgts_kobj) {
 215                 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
 216                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
 217         }
 218
 219         rc = lmv_check_connect(obd);
 220         if (rc != 0)
 221                 GOTO(out_sysfs, rc);
 222
 223         *pexp = exp;
 224
 225         RETURN(rc);
 226
 227 out_sysfs:
 228         if (lmv->lmv_tgts_kobj)
 229                 kobject_put(lmv->lmv_tgts_kobj);
 230
 231         class_disconnect(exp);
 232
 233         return rc;
 234 }
 235
 236 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
 237                             __u32 def_easize)
 238 {
 239         struct obd_device *obd = exp->exp_obd;
 240         struct lmv_obd *lmv = &obd->u.lmv;
 241         struct lmv_tgt_desc *tgt;
 242         int change = 0;
 243         int rc = 0;
 244
 245         ENTRY;
 246
 247         if (lmv->max_easize < easize) {
 248                 lmv->max_easize = easize;
 249                 change = 1;
 250         }
 251         if (lmv->max_def_easize < def_easize) {
 252                 lmv->max_def_easize = def_easize;
 253                 change = 1;
 254         }
 255
 256         if (change == 0)
 257                 RETURN(0);
 258
 259         if (lmv->connected == 0)
 260                 RETURN(0);
 261
 262         lmv_foreach_connected_tgt(lmv, tgt) {
 263                 if (!tgt->ltd_active)
 264                         continue;
 265
 266                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
 267                 if (rc) {
 268                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
 269                                " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
 270                         break;
 271                 }
 272         }
 273         RETURN(rc);
 274 }
 275
 276 #define MAX_STRING_SIZE 128
 277
 278 static int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 279 {
 280         struct lmv_obd *lmv = &obd->u.lmv;
 281         struct obd_device *mdc_obd;
 282         struct obd_export *mdc_exp;
 283         struct lu_fld_target target;
 284         int  rc;
 285         ENTRY;
 286
 287         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 288                                         &obd->obd_uuid);
 289         if (!mdc_obd) {
 290                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 291                 RETURN(-EINVAL);
 292         }
 293
 294         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
 295                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 296                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
 297
 298         if (!mdc_obd->obd_set_up) {
 299                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 300                 RETURN(-EINVAL);
 301         }
 302
 303         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
 304                          &lmv->conn_data, lmv->lmv_cache);
 305         if (rc) {
 306                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 307                 RETURN(rc);
 308         }
 309
 310         /*
 311          * Init fid sequence client for this mdc and add new fld target.
 312          */
 313         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
 314         if (rc)
 315                 RETURN(rc);
 316
 317         target.ft_srv = NULL;
 318         target.ft_exp = mdc_exp;
 319         target.ft_idx = tgt->ltd_index;
 320
 321         fld_client_add_target(&lmv->lmv_fld, &target);
 322
 323         rc = obd_register_observer(mdc_obd, obd);
 324         if (rc) {
 325                 obd_disconnect(mdc_exp);
 326                 CERROR("target %s register_observer error %d\n",
 327                        tgt->ltd_uuid.uuid, rc);
 328                 RETURN(rc);
 329         }
 330
 331         if (obd->obd_observer) {
 332                 /*
 333                  * Tell the observer about the new target.
 334                  */
 335                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 336                                 OBD_NOTIFY_ACTIVE);
 337                 if (rc) {
 338                         obd_disconnect(mdc_exp);
 339                         RETURN(rc);
 340                 }
 341         }
 342
 343         tgt->ltd_active = 1;
 344         tgt->ltd_exp = mdc_exp;
 345         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
 346
 347         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
 348
 349         rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
 350         if (rc) {
 351                 obd_disconnect(mdc_exp);
 352                 RETURN(rc);
 353         }
 354
 355         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 356                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 357                atomic_read(&obd->obd_refcount));
 358
 359         lmv_statfs_check_update(obd, tgt);
 360
 361         if (lmv->lmv_tgts_kobj)
 362                 /* Even if we failed to create the link, that's fine */
 363                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
 364                                        &mdc_obd->obd_kset.kobj,
 365                                        mdc_obd->obd_name);
 366         RETURN(0);
 367 }
 368
 369 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
 370 {
 371         LASSERT(tgt);
 372         ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
 373         OBD_FREE_PTR(tgt);
 374 }
 375
 376 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 377                            __u32 index, int gen)
 378 {
 379         struct obd_device *mdc_obd;
 380         struct lmv_obd *lmv = &obd->u.lmv;
 381         struct lmv_tgt_desc *tgt;
 382         struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
 383         int rc = 0;
 384
 385         ENTRY;
 386
 387         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
 388         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
 389                                         &obd->obd_uuid);
 390         if (!mdc_obd) {
 391                 CERROR("%s: Target %s not attached: rc = %d\n",
 392                        obd->obd_name, uuidp->uuid, -EINVAL);
 393                 RETURN(-EINVAL);
 394         }
 395
 396         OBD_ALLOC_PTR(tgt);
 397         if (!tgt)
 398                 RETURN(-ENOMEM);
 399
 400         mutex_init(&tgt->ltd_fid_mutex);
 401         tgt->ltd_index = index;
 402         tgt->ltd_uuid = *uuidp;
 403         tgt->ltd_active = 0;
 404
 405         mutex_lock(&ltd->ltd_mutex);
 406         rc = ltd_add_tgt(ltd, tgt);
 407         mutex_unlock(&ltd->ltd_mutex);
 408
 409         if (rc)
 410                 GOTO(out_tgt, rc);
 411
 412         if (!lmv->connected)
 413                 /* lmv_check_connect() will connect this target. */
 414                 RETURN(0);
 415
 416         rc = lmv_connect_mdc(obd, tgt);
 417         if (!rc) {
 418                 int easize = sizeof(struct lmv_stripe_md) +
 419                         lmv->lmv_mdt_count * sizeof(struct lu_fid);
 420
 421                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
 422         }
 423
 424         RETURN(rc);
 425
 426 out_tgt:
 427         OBD_FREE_PTR(tgt);
 428         return rc;
 429 }
 430
 431 static int lmv_check_connect(struct obd_device *obd)
 432 {
 433         struct lmv_obd *lmv = &obd->u.lmv;
 434         struct lmv_tgt_desc *tgt;
 435         int easize;
 436         int rc;
 437
 438         ENTRY;
 439
 440         if (lmv->connected)
 441                 RETURN(0);
 442
 443         mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
 444         if (lmv->connected)
 445                 GOTO(unlock, rc = 0);
 446
 447         if (!lmv->lmv_mdt_count) {
 448                 CERROR("%s: no targets configured: rc = -EINVAL\n",
 449                        obd->obd_name);
 450                 GOTO(unlock, rc = -EINVAL);
 451         }
 452
 453         if (!lmv_mdt0_inited(lmv)) {
 454                 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
 455                        obd->obd_name);
 456                 GOTO(unlock, rc = -EINVAL);
 457         }
 458
 459         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 460                obd->obd_uuid.uuid, obd->obd_name);
 461
 462         lmv_foreach_tgt(lmv, tgt) {
 463                 rc = lmv_connect_mdc(obd, tgt);
 464                 if (rc)
 465                         GOTO(out_disc, rc);
 466         }
 467
 468         lmv->connected = 1;
 469         easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
 470         lmv_init_ea_size(obd->obd_self_export, easize, 0);
 471         EXIT;
 472 unlock:
 473         mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
 474
 475         return rc;
 476
 477 out_disc:
 478         lmv_foreach_tgt(lmv, tgt) {
 479                 tgt->ltd_active = 0;
 480                 if (!tgt->ltd_exp)
 481                         continue;
 482
 483                 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
 484                 obd_disconnect(tgt->ltd_exp);
 485         }
 486
 487         goto unlock;
 488 }
 489
 490 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 491 {
 492         struct lmv_obd *lmv = &obd->u.lmv;
 493         struct obd_device *mdc_obd;
 494         int rc;
 495         ENTRY;
 496
 497         LASSERT(tgt != NULL);
 498         LASSERT(obd != NULL);
 499
 500         mdc_obd = class_exp2obd(tgt->ltd_exp);
 501
 502         if (mdc_obd) {
 503                 mdc_obd->obd_force = obd->obd_force;
 504                 mdc_obd->obd_fail = obd->obd_fail;
 505                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 506
 507                 if (lmv->lmv_tgts_kobj)
 508                         sysfs_remove_link(lmv->lmv_tgts_kobj,
 509                                           mdc_obd->obd_name);
 510         }
 511
 512         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
 513         if (rc)
 514                 CERROR("Can't finalize fids factory\n");
 515
 516         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 517                tgt->ltd_exp->exp_obd->obd_name,
 518                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 519
 520         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 521         rc = obd_disconnect(tgt->ltd_exp);
 522         if (rc) {
 523                 if (tgt->ltd_active) {
 524                         CERROR("Target %s disconnect error %d\n",
 525                                tgt->ltd_uuid.uuid, rc);
 526                 }
 527         }
 528
 529         lmv_activate_target(lmv, tgt, 0);
 530         tgt->ltd_exp = NULL;
 531         RETURN(0);
 532 }
 533
 534 static int lmv_disconnect(struct obd_export *exp)
 535 {
 536         struct obd_device *obd = class_exp2obd(exp);
 537         struct lmv_obd *lmv = &obd->u.lmv;
 538         struct lmv_tgt_desc *tgt;
 539         int rc;
 540
 541         ENTRY;
 542
 543         lmv_foreach_connected_tgt(lmv, tgt)
 544                 lmv_disconnect_mdc(obd, tgt);
 545
 546         if (lmv->lmv_tgts_kobj)
 547                 kobject_put(lmv->lmv_tgts_kobj);
 548
 549         lmv->connected = 0;
 550         rc = class_disconnect(exp);
 551
 552         RETURN(rc);
 553 }
 554
 555 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
 556                         void __user *uarg)
 557 {
 558         struct obd_device *obd = class_exp2obd(exp);
 559         struct lmv_obd *lmv = &obd->u.lmv;
 560         struct getinfo_fid2path *gf;
 561         struct lmv_tgt_desc *tgt;
 562         struct getinfo_fid2path *remote_gf = NULL;
 563         struct lu_fid root_fid;
 564         int remote_gf_size = 0;
 565         int rc;
 566
 567         gf = karg;
 568         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 569         if (IS_ERR(tgt))
 570                 RETURN(PTR_ERR(tgt));
 571
 572         root_fid = *gf->gf_u.gf_root_fid;
 573         LASSERT(fid_is_sane(&root_fid));
 574
 575 repeat_fid2path:
 576         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
 577         if (rc != 0 && rc != -EREMOTE)
 578                 GOTO(out_fid2path, rc);
 579
 580         /* If remote_gf != NULL, it means just building the
 581          * path on the remote MDT, copy this path segement to gf */
 582         if (remote_gf != NULL) {
 583                 struct getinfo_fid2path *ori_gf;
 584                 char *ptr;
 585                 int len;
 586
 587                 ori_gf = (struct getinfo_fid2path *)karg;
 588                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
 589                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
 590                         GOTO(out_fid2path, rc = -EOVERFLOW);
 591
 592                 ptr = ori_gf->gf_u.gf_path;
 593
 594                 len = strlen(gf->gf_u.gf_path);
 595                 /* move the current path to the right to release space
 596                  * for closer-to-root part */
 597                 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
 598                 memcpy(ptr, gf->gf_u.gf_path, len);
 599                 ptr[len] = '/';
 600         }
 601
 602         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
 603                tgt->ltd_exp->exp_obd->obd_name,
 604                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
 605                gf->gf_linkno);
 606
 607         if (rc == 0)
 608                 GOTO(out_fid2path, rc);
 609
 610         /* sigh, has to go to another MDT to do path building further */
 611         if (remote_gf == NULL) {
 612                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
 613                 OBD_ALLOC(remote_gf, remote_gf_size);
 614                 if (remote_gf == NULL)
 615                         GOTO(out_fid2path, rc = -ENOMEM);
 616                 remote_gf->gf_pathlen = PATH_MAX;
 617         }
 618
 619         if (!fid_is_sane(&gf->gf_fid)) {
 620                 CERROR("%s: invalid FID "DFID": rc = %d\n",
 621                        tgt->ltd_exp->exp_obd->obd_name,
 622                        PFID(&gf->gf_fid), -EINVAL);
 623                 GOTO(out_fid2path, rc = -EINVAL);
 624         }
 625
 626         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 627         if (IS_ERR(tgt))
 628                 GOTO(out_fid2path, rc = -EINVAL);
 629
 630         remote_gf->gf_fid = gf->gf_fid;
 631         remote_gf->gf_recno = -1;
 632         remote_gf->gf_linkno = -1;
 633         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
 634         *remote_gf->gf_u.gf_root_fid = root_fid;
 635         gf = remote_gf;
 636         goto repeat_fid2path;
 637
 638 out_fid2path:
 639         if (remote_gf != NULL)
 640                 OBD_FREE(remote_gf, remote_gf_size);
 641         RETURN(rc);
 642 }
 643
 644 static int lmv_hsm_req_count(struct lmv_obd *lmv,
 645                              const struct hsm_user_request *hur,
 646                              const struct lmv_tgt_desc *tgt_mds)
 647 {
 648         struct lmv_tgt_desc *curr_tgt;
 649         __u32 i;
 650         int nr = 0;
 651
 652         /* count how many requests must be sent to the given target */
 653         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
 654                 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
 655                 if (IS_ERR(curr_tgt))
 656                         RETURN(PTR_ERR(curr_tgt));
 657                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
 658                         nr++;
 659         }
 660         return nr;
 661 }
 662
 663 static int lmv_hsm_req_build(struct lmv_obd *lmv,
 664                               struct hsm_user_request *hur_in,
 665                               const struct lmv_tgt_desc *tgt_mds,
 666                               struct hsm_user_request *hur_out)
 667 {
 668         __u32 i, nr_out;
 669         struct lmv_tgt_desc *curr_tgt;
 670
 671         /* build the hsm_user_request for the given target */
 672         hur_out->hur_request = hur_in->hur_request;
 673         nr_out = 0;
 674         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
 675                 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
 676                 if (IS_ERR(curr_tgt))
 677                         RETURN(PTR_ERR(curr_tgt));
 678                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
 679                         hur_out->hur_user_item[nr_out] =
 680                                                 hur_in->hur_user_item[i];
 681                         nr_out++;
 682                 }
 683         }
 684         hur_out->hur_request.hr_itemcount = nr_out;
 685         memcpy(hur_data(hur_out), hur_data(hur_in),
 686                hur_in->hur_request.hr_data_len);
 687
 688         RETURN(0);
 689 }
 690
 691 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
 692                                  int len, struct lustre_kernelcomm *lk,
 693                                  void __user *uarg)
 694 {
 695         struct lmv_obd *lmv = &obd->u.lmv;
 696         struct lu_tgt_desc *tgt;
 697         int rc;
 698
 699         ENTRY;
 700
 701         /* unregister request (call from llapi_hsm_copytool_fini) */
 702         lmv_foreach_connected_tgt(lmv, tgt)
 703                 /* best effort: try to clean as much as possible
 704                  * (continue on error) */
 705                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 706
 707         /* Whatever the result, remove copytool from kuc groups.
 708          * Unreached coordinators will get EPIPE on next requests
 709          * and will unregister automatically.
 710          */
 711         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 712
 713         RETURN(rc);
 714 }
 715
 716 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
 717                                int len, struct lustre_kernelcomm *lk,
 718                                void __user *uarg)
 719 {
 720         struct lmv_obd *lmv = &obd->u.lmv;
 721         struct file *filp;
 722         bool any_set = false;
 723         struct kkuc_ct_data *kcd;
 724         size_t kcd_size;
 725         struct lu_tgt_desc *tgt;
 726         __u32 i;
 727         int err;
 728         int rc = 0;
 729
 730         ENTRY;
 731
 732         filp = fget(lk->lk_wfd);
 733         if (!filp)
 734                 RETURN(-EBADF);
 735
 736         if (lk->lk_flags & LK_FLG_DATANR)
 737                 kcd_size = offsetof(struct kkuc_ct_data,
 738                                     kcd_archives[lk->lk_data_count]);
 739         else
 740                 kcd_size = sizeof(*kcd);
 741
 742         OBD_ALLOC(kcd, kcd_size);
 743         if (kcd == NULL)
 744                 GOTO(err_fput, rc = -ENOMEM);
 745
 746         kcd->kcd_nr_archives = lk->lk_data_count;
 747         if (lk->lk_flags & LK_FLG_DATANR) {
 748                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
 749                 if (lk->lk_data_count > 0)
 750                         memcpy(kcd->kcd_archives, lk->lk_data,
 751                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
 752         } else {
 753                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
 754         }
 755
 756         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
 757                                    lk->lk_group, kcd, kcd_size);
 758         OBD_FREE(kcd, kcd_size);
 759         if (rc)
 760                 GOTO(err_fput, rc);
 761
 762         /* All or nothing: try to register to all MDS.
 763          * In case of failure, unregister from previous MDS,
 764          * except if it because of inactive target. */
 765         lmv_foreach_connected_tgt(lmv, tgt) {
 766                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 767                 if (err) {
 768                         if (tgt->ltd_active) {
 769                                 /* permanent error */
 770                                 CERROR("%s: iocontrol MDC %s on MDT"
 771                                        " idx %d cmd %x: err = %d\n",
 772                                        lmv2obd_dev(lmv)->obd_name,
 773                                        tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
 774                                        err);
 775                                 rc = err;
 776                                 lk->lk_flags |= LK_FLG_STOP;
 777                                 i = tgt->ltd_index;
 778                                 /* unregister from previous MDS */
 779                                 lmv_foreach_connected_tgt(lmv, tgt) {
 780                                         if (tgt->ltd_index >= i)
 781                                                 break;
 782
 783                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
 784                                                       lk, uarg);
 785                                 }
 786                                 GOTO(err_kkuc_rem, rc);
 787                         }
 788                         /* else: transient error.
 789                          * kuc will register to the missing MDT
 790                          * when it is back */
 791                 } else {
 792                         any_set = true;
 793                 }
 794         }
 795
 796         if (!any_set)
 797                 /* no registration done: return error */
 798                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
 799
 800         RETURN(0);
 801
 802 err_kkuc_rem:
 803         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 804
 805 err_fput:
 806         fput(filp);
 807         return rc;
 808 }
 809
 810 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 811                          int len, void *karg, void __user *uarg)
 812 {
 813         struct obd_device *obd = class_exp2obd(exp);
 814         struct lmv_obd *lmv = &obd->u.lmv;
 815         struct lu_tgt_desc *tgt = NULL;
 816         int set = 0;
 817         __u32 count = lmv->lmv_mdt_count;
 818         int rc = 0;
 819
 820         ENTRY;
 821
 822         if (count == 0)
 823                 RETURN(-ENOTTY);
 824
 825         switch (cmd) {
 826         case IOC_OBD_STATFS: {
 827                 struct obd_ioctl_data *data = karg;
 828                 struct obd_device *mdc_obd;
 829                 struct obd_statfs stat_buf = {0};
 830                 __u32 index;
 831
 832                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 833
 834                 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
 835                         RETURN(-ENODEV);
 836
 837                 tgt = lmv_tgt(lmv, index);
 838                 if (!tgt)
 839                         RETURN(-EAGAIN);
 840
 841                 if (!tgt->ltd_active)
 842                         RETURN(-ENODATA);
 843
 844                 mdc_obd = class_exp2obd(tgt->ltd_exp);
 845                 if (!mdc_obd)
 846                         RETURN(-EINVAL);
 847
 848                 /* copy UUID */
 849                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 850                                  min((int) data->ioc_plen2,
 851                                      (int) sizeof(struct obd_uuid))))
 852                         RETURN(-EFAULT);
 853
 854                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
 855                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
 856                                 0);
 857                 if (rc)
 858                         RETURN(rc);
 859                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
 860                                  min((int) data->ioc_plen1,
 861                                      (int) sizeof(stat_buf))))
 862                         RETURN(-EFAULT);
 863                 break;
 864         }
 865         case OBD_IOC_QUOTACTL: {
 866                 struct if_quotactl *qctl = karg;
 867                 struct obd_quotactl *oqctl;
 868                 struct obd_import *imp;
 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)
 888                         RETURN(-ENODEV);
 889
 890                 if (!tgt->ltd_exp)
 891                         RETURN(-EINVAL);
 892
 893                 imp = class_exp2cliimp(tgt->ltd_exp);
 894                 if (!tgt->ltd_active && imp->imp_state != LUSTRE_IMP_IDLE) {
 895                         qctl->qc_valid = QC_MDTIDX;
 896                         qctl->obd_uuid = tgt->ltd_uuid;
 897                         RETURN(-ENODATA);
 898                 }
 899
 900                 OBD_ALLOC_PTR(oqctl);
 901                 if (!oqctl)
 902                         RETURN(-ENOMEM);
 903
 904                 QCTL_COPY(oqctl, qctl);
 905                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 906                 if (rc == 0) {
 907                         QCTL_COPY(qctl, oqctl);
 908                         qctl->qc_valid = QC_MDTIDX;
 909                         qctl->obd_uuid = tgt->ltd_uuid;
 910                 }
 911                 OBD_FREE_PTR(oqctl);
 912                 break;
 913         }
 914         case LL_IOC_GET_CONNECT_FLAGS: {
 915                 tgt = lmv_tgt(lmv, 0);
 916                 rc = -ENODATA;
 917                 if (tgt && tgt->ltd_exp)
 918                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 919                 break;
 920         }
 921         case LL_IOC_FID2MDTIDX: {
 922                 struct lu_fid *fid = karg;
 923                 int             mdt_index;
 924
 925                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
 926                 if (rc != 0)
 927                         RETURN(rc);
 928
 929                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
 930                  * point to user space memory for FID2MDTIDX. */
 931                 *(__u32 *)uarg = mdt_index;
 932                 break;
 933         }
 934         case OBD_IOC_FID2PATH: {
 935                 rc = lmv_fid2path(exp, len, karg, uarg);
 936                 break;
 937         }
 938         case LL_IOC_HSM_STATE_GET:
 939         case LL_IOC_HSM_STATE_SET:
 940         case LL_IOC_HSM_ACTION: {
 941                 struct md_op_data *op_data = karg;
 942
 943                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
 944                 if (IS_ERR(tgt))
 945                         RETURN(PTR_ERR(tgt));
 946
 947                 if (tgt->ltd_exp == NULL)
 948                         RETURN(-EINVAL);
 949
 950                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 951                 break;
 952         }
 953         case LL_IOC_HSM_PROGRESS: {
 954                 const struct hsm_progress_kernel *hpk = karg;
 955
 956                 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
 957                 if (IS_ERR(tgt))
 958                         RETURN(PTR_ERR(tgt));
 959                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 960                 break;
 961         }
 962         case LL_IOC_HSM_REQUEST: {
 963                 struct hsm_user_request *hur = karg;
 964                 unsigned int reqcount = hur->hur_request.hr_itemcount;
 965
 966                 if (reqcount == 0)
 967                         RETURN(0);
 968
 969                 /* if the request is about a single fid
 970                  * or if there is a single MDS, no need to split
 971                  * the request. */
 972                 if (reqcount == 1 || count == 1) {
 973                         tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
 974                         if (IS_ERR(tgt))
 975                                 RETURN(PTR_ERR(tgt));
 976                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 977                 } else {
 978                         /* split fid list to their respective MDS */
 979                         lmv_foreach_connected_tgt(lmv, tgt) {
 980                                 int nr, rc1;
 981                                 size_t reqlen;
 982                                 struct hsm_user_request *req;
 983
 984                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
 985                                 if (nr < 0)
 986                                         RETURN(nr);
 987                                 if (nr == 0) /* nothing for this MDS */
 988                                         continue;
 989
 990                                 /* build a request with fids for this MDS */
 991                                 reqlen = offsetof(typeof(*hur),
 992                                                   hur_user_item[nr])
 993                                                 + hur->hur_request.hr_data_len;
 994                                 OBD_ALLOC_LARGE(req, reqlen);
 995                                 if (req == NULL)
 996                                         RETURN(-ENOMEM);
 997                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
 998                                 if (rc1 < 0)
 999                                         GOTO(hsm_req_err, rc1);
1000                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
1001                                                     req, uarg);
1002 hsm_req_err:
1003                                 if (rc1 != 0 && rc == 0)
1004                                         rc = rc1;
1005                                 OBD_FREE_LARGE(req, reqlen);
1006                         }
1007                 }
1008                 break;
1009         }
1010         case LL_IOC_LOV_SWAP_LAYOUTS: {
1011                 struct md_op_data *op_data = karg;
1012                 struct lmv_tgt_desc *tgt1, *tgt2;
1013
1014                 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1015                 if (IS_ERR(tgt1))
1016                         RETURN(PTR_ERR(tgt1));
1017
1018                 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1019                 if (IS_ERR(tgt2))
1020                         RETURN(PTR_ERR(tgt2));
1021
1022                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1023                         RETURN(-EINVAL);
1024
1025                 /* only files on same MDT can have their layouts swapped */
1026                 if (tgt1->ltd_index != tgt2->ltd_index)
1027                         RETURN(-EPERM);
1028
1029                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1030                 break;
1031         }
1032         case LL_IOC_HSM_CT_START: {
1033                 struct lustre_kernelcomm *lk = karg;
1034                 if (lk->lk_flags & LK_FLG_STOP)
1035                         rc = lmv_hsm_ct_unregister(obd, cmd, len, lk, uarg);
1036                 else
1037                         rc = lmv_hsm_ct_register(obd, cmd, len, lk, uarg);
1038                 break;
1039         }
1040         default:
1041                 lmv_foreach_connected_tgt(lmv, tgt) {
1042                         struct obd_device *mdc_obd;
1043                         int err;
1044
1045                         /* ll_umount_begin() sets force flag but for lmv, not
1046                          * mdc. Let's pass it through */
1047                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1048                         mdc_obd->obd_force = obd->obd_force;
1049                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1050                         if (err) {
1051                                 if (tgt->ltd_active) {
1052                                         CERROR("error: iocontrol MDC %s on MDT"
1053                                                " idx %d cmd %x: err = %d\n",
1054                                                tgt->ltd_uuid.uuid,
1055                                                tgt->ltd_index, cmd, err);
1056                                         if (!rc)
1057                                                 rc = err;
1058                                 }
1059                         } else
1060                                 set = 1;
1061                 }
1062                 if (!set && !rc)
1063                         rc = -EIO;
1064         }
1065         RETURN(rc);
1066 }
1067
1068 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1069                   struct lu_fid *fid, struct md_op_data *op_data)
1070 {
1071         struct obd_device *obd = class_exp2obd(exp);
1072         struct lmv_obd *lmv = &obd->u.lmv;
1073         struct lmv_tgt_desc *tgt;
1074         int rc;
1075
1076         ENTRY;
1077
1078         LASSERT(op_data);
1079         LASSERT(fid);
1080
1081         tgt = lmv_tgt(lmv, op_data->op_mds);
1082         if (!tgt)
1083                 RETURN(-ENODEV);
1084
1085         if (!tgt->ltd_active || !tgt->ltd_exp)
1086                 RETURN(-ENODEV);
1087
1088         /*
1089          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1090          * on server that seq in new allocated fid is not yet known.
1091          */
1092         mutex_lock(&tgt->ltd_fid_mutex);
1093         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1094         mutex_unlock(&tgt->ltd_fid_mutex);
1095         if (rc > 0) {
1096                 LASSERT(fid_is_sane(fid));
1097                 rc = 0;
1098         }
1099
1100         RETURN(rc);
1101 }
1102
1103 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1104 {
1105         struct lmv_obd *lmv = &obd->u.lmv;
1106         struct lmv_desc *desc;
1107         struct lnet_processid lnet_id;
1108         int i = 0;
1109         int rc;
1110
1111         ENTRY;
1112
1113         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1114                 CERROR("LMV setup requires a descriptor\n");
1115                 RETURN(-EINVAL);
1116         }
1117
1118         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1119         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1120                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1121                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1122                 RETURN(-EINVAL);
1123         }
1124
1125         obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1126                      desc->ld_uuid.uuid);
1127         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1128         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1129         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1130                 LMV_DESC_QOS_MAXAGE_DEFAULT;
1131         lmv->max_def_easize = 0;
1132         lmv->max_easize = 0;
1133
1134         spin_lock_init(&lmv->lmv_lock);
1135
1136         /*
1137          * initialize rr_index to lower 32bit of netid, so that client
1138          * can distribute subdirs evenly from the beginning.
1139          */
1140         while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1141                 if (!nid_is_lo0(&lnet_id.nid)) {
1142                         lmv->lmv_qos_rr_index = ntohl(lnet_id.nid.nid_addr[0]);
1143                         break;
1144                 }
1145         }
1146
1147         rc = lmv_tunables_init(obd);
1148         if (rc)
1149                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1150                       obd->obd_name, rc);
1151
1152         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1153                              LUSTRE_CLI_FLD_HASH_DHT);
1154         if (rc)
1155                 CERROR("Can't init FLD, err %d\n", rc);
1156
1157         rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1158         if (rc)
1159                 CWARN("%s: error initialize target table: rc = %d\n",
1160                       obd->obd_name, rc);
1161
1162         RETURN(rc);
1163 }
1164
1165 static int lmv_cleanup(struct obd_device *obd)
1166 {
1167         struct lmv_obd *lmv = &obd->u.lmv;
1168         struct lu_tgt_desc *tgt;
1169         struct lu_tgt_desc *tmp;
1170
1171         ENTRY;
1172
1173         fld_client_fini(&lmv->lmv_fld);
1174         fld_client_debugfs_fini(&lmv->lmv_fld);
1175
1176         lprocfs_obd_cleanup(obd);
1177         lprocfs_free_md_stats(obd);
1178
1179         lmv_foreach_tgt_safe(lmv, tgt, tmp)
1180                 lmv_del_target(lmv, tgt);
1181         lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1182
1183         RETURN(0);
1184 }
1185
1186 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1187 {
1188         struct lustre_cfg       *lcfg = buf;
1189         struct obd_uuid         obd_uuid;
1190         int                     gen;
1191         __u32                   index;
1192         int                     rc;
1193         ENTRY;
1194
1195         switch (lcfg->lcfg_command) {
1196         case LCFG_ADD_MDC:
1197                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1198                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1199                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1200                         GOTO(out, rc = -EINVAL);
1201
1202                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1203
1204                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1205                         GOTO(out, rc = -EINVAL);
1206                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1207                         GOTO(out, rc = -EINVAL);
1208                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1209                 GOTO(out, rc);
1210         default:
1211                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1212                 GOTO(out, rc = -EINVAL);
1213         }
1214 out:
1215         RETURN(rc);
1216 }
1217
1218 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1219 {
1220         int i;
1221
1222         if (flags & OBD_STATFS_FOR_MDT0)
1223                 return 0;
1224
1225         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1226                 return lmv->lmv_statfs_start;
1227
1228         /* choose initial MDT for this client */
1229         for (i = 0;; i++) {
1230                 struct lnet_processid lnet_id;
1231                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1232                         break;
1233
1234                 if (!nid_is_lo0(&lnet_id.nid)) {
1235                         /* We dont need a full 64-bit modulus, just enough
1236                          * to distribute the requests across MDTs evenly.
1237                          */
1238                         lmv->lmv_statfs_start = nidhash(&lnet_id.nid) %
1239                                                 lmv->lmv_mdt_count;
1240                         break;
1241                 }
1242         }
1243
1244         return lmv->lmv_statfs_start;
1245 }
1246
1247 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1248                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1249 {
1250         struct obd_device *obd = class_exp2obd(exp);
1251         struct lmv_obd *lmv = &obd->u.lmv;
1252         struct obd_statfs *temp;
1253         struct lu_tgt_desc *tgt;
1254         __u32 i;
1255         __u32 idx;
1256         int rc = 0;
1257         int err = 0;
1258
1259         ENTRY;
1260
1261         OBD_ALLOC(temp, sizeof(*temp));
1262         if (temp == NULL)
1263                 RETURN(-ENOMEM);
1264
1265         /* distribute statfs among MDTs */
1266         idx = lmv_select_statfs_mdt(lmv, flags);
1267
1268         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1269                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1270                 tgt = lmv_tgt(lmv, idx);
1271                 if (!tgt || !tgt->ltd_exp)
1272                         continue;
1273
1274                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age,
1275                                 flags | OBD_STATFS_NESTED);
1276                 if (rc) {
1277                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1278                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1279                         err = rc;
1280                         /* Try another MDT */
1281                         if (flags & OBD_STATFS_SUM)
1282                                 continue;
1283                         GOTO(out_free_temp, rc);
1284                 }
1285
1286                 if (temp->os_state & OS_STATFS_SUM ||
1287                     flags == OBD_STATFS_FOR_MDT0) {
1288                         /* reset to the last aggregated values
1289                          * and don't sum with non-aggrated data */
1290                         /* If the statfs is from mount, it needs to retrieve
1291                          * necessary information from MDT0. i.e. mount does
1292                          * not need the merged osfs from all of MDT. Also
1293                          * clients can be mounted as long as MDT0 is in
1294                          * service */
1295                         *osfs = *temp;
1296                         GOTO(out_free_temp, rc);
1297                 }
1298
1299                 if (i == 0) {
1300                         *osfs = *temp;
1301                 } else {
1302                         osfs->os_bavail += temp->os_bavail;
1303                         osfs->os_blocks += temp->os_blocks;
1304                         osfs->os_ffree += temp->os_ffree;
1305                         osfs->os_files += temp->os_files;
1306                         osfs->os_granted += temp->os_granted;
1307                 }
1308         }
1309         /* There is no stats from some MDTs, data incomplete */
1310         if (err)
1311                 rc = err;
1312 out_free_temp:
1313         OBD_FREE(temp, sizeof(*temp));
1314         RETURN(rc);
1315 }
1316
1317 static int lmv_statfs_update(void *cookie, int rc)
1318 {
1319         struct obd_info *oinfo = cookie;
1320         struct obd_device *obd = oinfo->oi_obd;
1321         struct lmv_obd *lmv = &obd->u.lmv;
1322         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1323         struct obd_statfs *osfs = oinfo->oi_osfs;
1324
1325         /*
1326          * NB: don't deactivate TGT upon error, because we may not trigger async
1327          * statfs any longer, then there is no chance to activate TGT.
1328          */
1329         if (!rc) {
1330                 spin_lock(&lmv->lmv_lock);
1331                 tgt->ltd_statfs = *osfs;
1332                 tgt->ltd_statfs_age = ktime_get_seconds();
1333                 spin_unlock(&lmv->lmv_lock);
1334                 set_bit(LQ_DIRTY, &lmv->lmv_qos.lq_flags);
1335         }
1336
1337         return rc;
1338 }
1339
1340 /* update tgt statfs async if it's ld_qos_maxage old */
1341 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1342 {
1343         struct obd_info oinfo = {
1344                 .oi_obd = obd,
1345                 .oi_tgt = tgt,
1346                 .oi_cb_up = lmv_statfs_update,
1347         };
1348         int rc;
1349
1350         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1351             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1352                 return 0;
1353
1354         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1355
1356         return rc;
1357 }
1358
1359 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1360                         struct lu_fid *fid)
1361 {
1362         struct obd_device *obd = exp->exp_obd;
1363         struct lmv_obd *lmv = &obd->u.lmv;
1364         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1365         int rc;
1366
1367         ENTRY;
1368
1369         if (!tgt)
1370                 RETURN(-ENODEV);
1371
1372         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1373         RETURN(rc);
1374 }
1375
1376 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1377                         u64 obd_md_valid, const char *name, size_t buf_size,
1378                         struct ptlrpc_request **req)
1379 {
1380         struct obd_device *obd = exp->exp_obd;
1381         struct lmv_obd *lmv = &obd->u.lmv;
1382         struct lmv_tgt_desc *tgt;
1383         int rc;
1384
1385         ENTRY;
1386
1387         tgt = lmv_fid2tgt(lmv, fid);
1388         if (IS_ERR(tgt))
1389                 RETURN(PTR_ERR(tgt));
1390
1391         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1392
1393         RETURN(rc);
1394 }
1395
1396 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1397                         u64 obd_md_valid, const char *name,
1398                         const void *value, size_t value_size,
1399                         unsigned int xattr_flags, u32 suppgid,
1400                         struct ptlrpc_request **req)
1401 {
1402         struct obd_device *obd = exp->exp_obd;
1403         struct lmv_obd *lmv = &obd->u.lmv;
1404         struct lmv_tgt_desc *tgt;
1405         int rc;
1406
1407         ENTRY;
1408
1409         tgt = lmv_fid2tgt(lmv, fid);
1410         if (IS_ERR(tgt))
1411                 RETURN(PTR_ERR(tgt));
1412
1413         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1414                          value, value_size, xattr_flags, suppgid, req);
1415
1416         RETURN(rc);
1417 }
1418
1419 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1420                        struct ptlrpc_request **request)
1421 {
1422         struct obd_device *obd = exp->exp_obd;
1423         struct lmv_obd *lmv = &obd->u.lmv;
1424         struct lmv_tgt_desc *tgt;
1425         int rc;
1426
1427         ENTRY;
1428
1429         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1430         if (IS_ERR(tgt))
1431                 RETURN(PTR_ERR(tgt));
1432
1433         if (op_data->op_flags & MF_GET_MDT_IDX) {
1434                 op_data->op_mds = tgt->ltd_index;
1435                 RETURN(0);
1436         }
1437
1438         rc = md_getattr(tgt->ltd_exp, op_data, request);
1439
1440         RETURN(rc);
1441 }
1442
1443 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1444 {
1445         struct obd_device *obd = exp->exp_obd;
1446         struct lmv_obd *lmv = &obd->u.lmv;
1447         struct lu_tgt_desc *tgt;
1448
1449         ENTRY;
1450
1451         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1452
1453         /*
1454          * With DNE every object can have two locks in different namespaces:
1455          * lookup lock in space of MDT storing direntry and update/open lock in
1456          * space of MDT storing inode.
1457          */
1458         lmv_foreach_connected_tgt(lmv, tgt)
1459                 md_null_inode(tgt->ltd_exp, fid);
1460
1461         RETURN(0);
1462 }
1463
1464 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1465                      struct md_open_data *mod, struct ptlrpc_request **request)
1466 {
1467         struct obd_device *obd = exp->exp_obd;
1468         struct lmv_obd *lmv = &obd->u.lmv;
1469         struct lmv_tgt_desc *tgt;
1470         int rc;
1471
1472         ENTRY;
1473
1474         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1475         if (IS_ERR(tgt))
1476                 RETURN(PTR_ERR(tgt));
1477
1478         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1479         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1480         RETURN(rc);
1481 }
1482
1483 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv,
1484                                               struct md_op_data *op_data)
1485 {
1486         struct lu_tgt_desc *tgt, *cur = NULL;
1487         __u64 total_avail = 0;
1488         __u64 total_weight = 0;
1489         __u64 cur_weight = 0;
1490         int total_usable = 0;
1491         __u64 rand;
1492         int rc;
1493
1494         ENTRY;
1495
1496         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1497                 RETURN(ERR_PTR(-EAGAIN));
1498
1499         down_write(&lmv->lmv_qos.lq_rw_sem);
1500
1501         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1502                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1503
1504         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1505         if (rc)
1506                 GOTO(unlock, tgt = ERR_PTR(rc));
1507
1508         lmv_foreach_tgt(lmv, tgt) {
1509                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1510                         tgt->ltd_qos.ltq_usable = 0;
1511                         continue;
1512                 }
1513
1514                 tgt->ltd_qos.ltq_usable = 1;
1515                 lu_tgt_qos_weight_calc(tgt);
1516                 if (tgt->ltd_index == op_data->op_mds)
1517                         cur = tgt;
1518                 total_avail += tgt->ltd_qos.ltq_avail;
1519                 total_weight += tgt->ltd_qos.ltq_weight;
1520                 total_usable++;
1521         }
1522
1523         /* If current MDT has above-average space and dir is not aleady using
1524          * round-robin to spread across more MDTs, stay on the parent MDT
1525          * to avoid creating needless remote MDT directories.  Remote dirs
1526          * close to the root balance space more effectively than bottom dirs,
1527          * so prefer to create remote dirs at top level of directory tree.
1528          * "16 / (dir_depth + 10)" is the factor to make it less likely
1529          * for top-level directories to stay local unless they have more than
1530          * average free space, while deep dirs prefer local until more full.
1531          *    depth=0 -> 160%, depth=3 -> 123%, depth=6 -> 100%,
1532          *    depth=9 -> 84%, depth=12 -> 73%, depth=15 -> 64%
1533          */
1534         if (!lmv_op_default_rr_mkdir(op_data)) {
1535                 rand = total_avail * 16 /
1536                         (total_usable * (op_data->op_dir_depth + 10));
1537                 if (cur && cur->ltd_qos.ltq_avail >= rand) {
1538                         tgt = cur;
1539                         GOTO(unlock, tgt);
1540                 }
1541         }
1542
1543         rand = lu_prandom_u64_max(total_weight);
1544
1545         lmv_foreach_connected_tgt(lmv, tgt) {
1546                 if (!tgt->ltd_qos.ltq_usable)
1547                         continue;
1548
1549                 cur_weight += tgt->ltd_qos.ltq_weight;
1550                 if (cur_weight < rand)
1551                         continue;
1552
1553                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1554                 GOTO(unlock, tgt);
1555         }
1556
1557         /* no proper target found */
1558         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1559 unlock:
1560         up_write(&lmv->lmv_qos.lq_rw_sem);
1561
1562         return tgt;
1563 }
1564
1565 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv)
1566 {
1567         struct lu_tgt_desc *tgt;
1568         int i;
1569         int index;
1570
1571         ENTRY;
1572
1573         spin_lock(&lmv->lmv_lock);
1574         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1575                 index = (i + lmv->lmv_qos_rr_index) %
1576                         lmv->lmv_mdt_descs.ltd_tgts_size;
1577                 tgt = lmv_tgt(lmv, index);
1578                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1579                         continue;
1580
1581                 lmv->lmv_qos_rr_index = (tgt->ltd_index + 1) %
1582                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1583                 spin_unlock(&lmv->lmv_lock);
1584
1585                 RETURN(tgt);
1586         }
1587         spin_unlock(&lmv->lmv_lock);
1588
1589         RETURN(ERR_PTR(-ENODEV));
1590 }
1591
1592 /* locate MDT which is less full (avoid the most full MDT) */
1593 static struct lu_tgt_desc *lmv_locate_tgt_lf(struct lmv_obd *lmv)
1594 {
1595         struct lu_tgt_desc *min = NULL;
1596         struct lu_tgt_desc *tgt;
1597         __u64 avail = 0;
1598         __u64 rand;
1599
1600         ENTRY;
1601
1602         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1603                 RETURN(ERR_PTR(-EAGAIN));
1604
1605         down_write(&lmv->lmv_qos.lq_rw_sem);
1606
1607         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1608                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1609
1610         lmv_foreach_tgt(lmv, tgt) {
1611                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1612                         tgt->ltd_qos.ltq_usable = 0;
1613                         continue;
1614                 }
1615
1616                 tgt->ltd_qos.ltq_usable = 1;
1617                 lu_tgt_qos_weight_calc(tgt);
1618                 avail += tgt->ltd_qos.ltq_avail;
1619                 if (!min || min->ltd_qos.ltq_avail > tgt->ltd_qos.ltq_avail)
1620                         min = tgt;
1621         }
1622
1623         /* avoid the most full MDT */
1624         if (min)
1625                 avail -= min->ltd_qos.ltq_avail;
1626
1627         rand = lu_prandom_u64_max(avail);
1628         avail = 0;
1629         lmv_foreach_connected_tgt(lmv, tgt) {
1630                 if (!tgt->ltd_qos.ltq_usable)
1631                         continue;
1632
1633                 if (tgt == min)
1634                         continue;
1635
1636                 avail += tgt->ltd_qos.ltq_avail;
1637                 if (avail < rand)
1638                         continue;
1639
1640                 GOTO(unlock, tgt);
1641         }
1642
1643         /* no proper target found */
1644         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1645 unlock:
1646         up_write(&lmv->lmv_qos.lq_rw_sem);
1647
1648         RETURN(tgt);
1649 }
1650
1651 /* locate MDT by file name, for striped directory, the file name hash decides
1652  * which stripe its dirent is stored.
1653  */
1654 static struct lmv_tgt_desc *
1655 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1656                        const char *name, int namelen, struct lu_fid *fid,
1657                        __u32 *mds, bool new_layout)
1658 {
1659         struct lmv_tgt_desc *tgt;
1660         const struct lmv_oinfo *oinfo;
1661
1662         if (!lmv_dir_striped(lsm) || !namelen) {
1663                 tgt = lmv_fid2tgt(lmv, fid);
1664                 if (IS_ERR(tgt))
1665                         return tgt;
1666
1667                 *mds = tgt->ltd_index;
1668                 return tgt;
1669         }
1670
1671         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1672                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1673                         return ERR_PTR(-EBADF);
1674                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1675         } else {
1676                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, new_layout);
1677                 if (IS_ERR(oinfo))
1678                         return ERR_CAST(oinfo);
1679         }
1680
1681         /* check stripe FID is sane */
1682         if (!fid_is_sane(&oinfo->lmo_fid))
1683                 return ERR_PTR(-ENODEV);
1684
1685         *fid = oinfo->lmo_fid;
1686         *mds = oinfo->lmo_mds;
1687         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1688
1689         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1690
1691         return tgt ? tgt : ERR_PTR(-ENODEV);
1692 }
1693
1694 /**
1695  * Locate MDT of op_data->op_fid1
1696  *
1697  * For striped directory, it will locate the stripe by name hash, if hash_type
1698  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1699  * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1700  * indicates whether old or new layout is used to locate.
1701  *
1702  * For plain direcotry, it just locate the MDT of op_data->op_fid1.
1703  *
1704  * \param[in] lmv               LMV device
1705  * \param[in/out] op_data       client MD stack parameters, name, namelen etc,
1706  *                              op_mds and op_fid1 will be updated if op_mea1
1707  *                              indicates fid1 represents a striped directory.
1708  *
1709  * retval               pointer to the lmv_tgt_desc if succeed.
1710  *                      ERR_PTR(errno) if failed.
1711  */
1712 struct lmv_tgt_desc *
1713 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1714 {
1715         struct lmv_stripe_md *lsm = op_data->op_mea1;
1716         struct lmv_oinfo *oinfo;
1717         struct lmv_tgt_desc *tgt;
1718
1719         if (lmv_dir_foreign(lsm))
1720                 return ERR_PTR(-ENODATA);
1721
1722         /* During creating VOLATILE file, it should honor the mdt
1723          * index if the file under striped dir is being restored, see
1724          * ct_restore(). */
1725         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1726             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1727                 tgt = lmv_tgt(lmv, op_data->op_mds);
1728                 if (!tgt)
1729                         return ERR_PTR(-ENODEV);
1730
1731                 if (lmv_dir_striped(lsm)) {
1732                         int i;
1733
1734                         /* refill the right parent fid */
1735                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1736                                 oinfo = &lsm->lsm_md_oinfo[i];
1737                                 if (oinfo->lmo_mds == op_data->op_mds) {
1738                                         op_data->op_fid1 = oinfo->lmo_fid;
1739                                         break;
1740                                 }
1741                         }
1742
1743                         if (i == lsm->lsm_md_stripe_count)
1744                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1745                 }
1746         } else if (lmv_dir_bad_hash(lsm)) {
1747                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1748                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1749
1750                 op_data->op_fid1 = oinfo->lmo_fid;
1751                 op_data->op_mds = oinfo->lmo_mds;
1752                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1753                 if (!tgt)
1754                         return ERR_PTR(-ENODEV);
1755         } else {
1756                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1757                                 op_data->op_name, op_data->op_namelen,
1758                                 &op_data->op_fid1, &op_data->op_mds,
1759                                 op_data->op_new_layout);
1760         }
1761
1762         return tgt;
1763 }
1764
1765 /* Locate MDT of op_data->op_fid2 for link/rename */
1766 static struct lmv_tgt_desc *
1767 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1768 {
1769         struct lmv_tgt_desc *tgt;
1770         int rc;
1771
1772         LASSERT(op_data->op_name);
1773         if (lmv_dir_layout_changing(op_data->op_mea2)) {
1774                 struct lu_fid fid1 = op_data->op_fid1;
1775                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1776                 struct ptlrpc_request *request = NULL;
1777
1778                 /*
1779                  * avoid creating new file under old layout of migrating
1780                  * directory, check it here.
1781                  */
1782                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1783                                 op_data->op_name, op_data->op_namelen,
1784                                 &op_data->op_fid2, &op_data->op_mds, false);
1785                 if (IS_ERR(tgt))
1786                         RETURN(tgt);
1787
1788                 op_data->op_fid1 = op_data->op_fid2;
1789                 op_data->op_mea1 = op_data->op_mea2;
1790                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1791                 op_data->op_fid1 = fid1;
1792                 op_data->op_mea1 = lsm1;
1793                 if (!rc) {
1794                         ptlrpc_req_finished(request);
1795                         RETURN(ERR_PTR(-EEXIST));
1796                 }
1797
1798                 if (rc != -ENOENT)
1799                         RETURN(ERR_PTR(rc));
1800         }
1801
1802         return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1803                                 op_data->op_namelen, &op_data->op_fid2,
1804                                 &op_data->op_mds, true);
1805 }
1806
1807 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1808 {
1809         struct lu_tgt_desc *tgt;
1810         struct ptlrpc_request *request;
1811         int rc;
1812
1813         LASSERT(lmv_dir_layout_changing(op_data->op_mea1));
1814         LASSERT(!op_data->op_new_layout);
1815
1816         tgt = lmv_locate_tgt(lmv, op_data);
1817         if (IS_ERR(tgt))
1818                 return PTR_ERR(tgt);
1819
1820         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1821         if (!rc) {
1822                 ptlrpc_req_finished(request);
1823                 return -EEXIST;
1824         }
1825
1826         return rc;
1827 }
1828
1829 /* mkdir by QoS upon 'lfs mkdir -i -1'.
1830  *
1831  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1832  * directories under striped directory.
1833  */
1834 static inline bool lmv_op_user_qos_mkdir(const struct md_op_data *op_data)
1835 {
1836         const struct lmv_user_md *lum = op_data->op_data;
1837
1838         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1839                 return false;
1840
1841         if (lmv_dir_striped(op_data->op_mea1))
1842                 return false;
1843
1844         return (op_data->op_cli_flags & CLI_SET_MEA) && lum &&
1845                le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC &&
1846                le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
1847 }
1848
1849 /* mkdir by QoS if either ROOT or parent default LMV is space balanced. */
1850 static inline bool lmv_op_default_qos_mkdir(const struct md_op_data *op_data)
1851 {
1852         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1853
1854         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1855                 return false;
1856
1857         if (lmv_dir_striped(op_data->op_mea1))
1858                 return false;
1859
1860         return (op_data->op_flags & MF_QOS_MKDIR) ||
1861                (lsm && lsm->lsm_md_master_mdt_index == LMV_OFFSET_DEFAULT);
1862 }
1863
1864 /* if parent default LMV is space balanced, and
1865  * 1. max_inherit_rr is set
1866  * 2. or parent is ROOT
1867  * mkdir roundrobin. Or if parent doesn't have default LMV, while ROOT default
1868  * LMV requests roundrobin mkdir, do the same.
1869  * NB, this needs to check server is balanced, which is done by caller.
1870  */
1871 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data)
1872 {
1873         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1874
1875         return (op_data->op_flags & MF_RR_MKDIR) ||
1876                (lsm && lsm->lsm_md_max_inherit_rr != LMV_INHERIT_RR_NONE) ||
1877                fid_is_root(&op_data->op_fid1);
1878 }
1879
1880 /* 'lfs mkdir -i <specific_MDT>' */
1881 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1882 {
1883         const struct lmv_user_md *lum = op_data->op_data;
1884
1885         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1886                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1887                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1888                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1889                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1890 }
1891
1892 /* parent default LMV master_mdt_index is not -1. */
1893 static inline bool
1894 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1895 {
1896         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1897                op_data->op_default_mea1 &&
1898                op_data->op_default_mea1->lsm_md_master_mdt_index !=
1899                         LMV_OFFSET_DEFAULT;
1900 }
1901
1902 /* locate MDT by space usage */
1903 static struct lu_tgt_desc *lmv_locate_tgt_by_space(struct lmv_obd *lmv,
1904                                                    struct md_op_data *op_data,
1905                                                    struct lmv_tgt_desc *tgt)
1906 {
1907         struct lmv_tgt_desc *tmp = tgt;
1908
1909         tgt = lmv_locate_tgt_qos(lmv, op_data);
1910         if (tgt == ERR_PTR(-EAGAIN)) {
1911                 if (ltd_qos_is_balanced(&lmv->lmv_mdt_descs) &&
1912                     !lmv_op_default_rr_mkdir(op_data) &&
1913                     !lmv_op_user_qos_mkdir(op_data))
1914                         /* if not necessary, don't create remote directory. */
1915                         tgt = tmp;
1916                 else
1917                         tgt = lmv_locate_tgt_rr(lmv);
1918         }
1919
1920         /*
1921          * only update statfs after QoS mkdir, this means the cached statfs may
1922          * be stale, and current mkdir may not follow QoS accurately, but it's
1923          * not serious, and avoids periodic statfs when client doesn't mkdir by
1924          * QoS.
1925          */
1926         if (!IS_ERR(tgt)) {
1927                 op_data->op_mds = tgt->ltd_index;
1928                 lmv_statfs_check_update(lmv2obd_dev(lmv), tgt);
1929         }
1930
1931         return tgt;
1932 }
1933
1934 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1935                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1936                 gid_t gid, kernel_cap_t cap_effective, __u64 rdev,
1937                 struct ptlrpc_request **request)
1938 {
1939         struct obd_device *obd = exp->exp_obd;
1940         struct lmv_obd *lmv = &obd->u.lmv;
1941         struct lmv_tgt_desc *tgt;
1942         struct mdt_body *repbody;
1943         int rc;
1944
1945         ENTRY;
1946
1947         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1948                 RETURN(-EIO);
1949
1950         if (lmv_dir_bad_hash(op_data->op_mea1))
1951                 RETURN(-EBADF);
1952
1953         if (lmv_dir_layout_changing(op_data->op_mea1)) {
1954                 /*
1955                  * if parent is migrating, create() needs to lookup existing
1956                  * name in both old and new layout, check old layout on client.
1957                  */
1958                 rc = lmv_old_layout_lookup(lmv, op_data);
1959                 if (rc != -ENOENT)
1960                         RETURN(rc);
1961
1962                 op_data->op_new_layout = true;
1963         }
1964
1965         tgt = lmv_locate_tgt(lmv, op_data);
1966         if (IS_ERR(tgt))
1967                 RETURN(PTR_ERR(tgt));
1968
1969         /* the order to apply policy in mkdir:
1970          * 1. is "lfs mkdir -i N"? mkdir on MDT N.
1971          * 2. is "lfs mkdir -i -1"? mkdir by space usage.
1972          * 3. is starting MDT specified in default LMV? mkdir on MDT N.
1973          * 4. is default LMV space balanced? mkdir by space usage.
1974          */
1975         if (lmv_op_user_specific_mkdir(op_data)) {
1976                 struct lmv_user_md *lum = op_data->op_data;
1977
1978                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
1979                 tgt = lmv_tgt(lmv, op_data->op_mds);
1980                 if (!tgt)
1981                         RETURN(-ENODEV);
1982         } else if (lmv_op_user_qos_mkdir(op_data)) {
1983                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
1984                 if (IS_ERR(tgt))
1985                         RETURN(PTR_ERR(tgt));
1986         } else if (lmv_op_default_specific_mkdir(op_data)) {
1987                 op_data->op_mds =
1988                         op_data->op_default_mea1->lsm_md_master_mdt_index;
1989                 tgt = lmv_tgt(lmv, op_data->op_mds);
1990                 if (!tgt)
1991                         RETURN(-ENODEV);
1992         } else if (lmv_op_default_qos_mkdir(op_data)) {
1993                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
1994                 if (IS_ERR(tgt))
1995                         RETURN(PTR_ERR(tgt));
1996         }
1997
1998 retry:
1999         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
2000         if (rc)
2001                 RETURN(rc);
2002
2003         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
2004                 (int)op_data->op_namelen, op_data->op_name,
2005                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
2006                 op_data->op_mds);
2007
2008         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2009         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
2010                        cap_effective, rdev, request);
2011         if (rc == 0) {
2012                 if (*request == NULL)
2013                         RETURN(rc);
2014                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
2015         }
2016
2017         /* dir restripe needs to send to MDT where dir is located */
2018         if (rc != -EREMOTE ||
2019             !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
2020                 RETURN(rc);
2021
2022         repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2023         if (repbody == NULL)
2024                 RETURN(-EPROTO);
2025
2026         /* Not cross-ref case, just get out of here. */
2027         if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
2028                 RETURN(rc);
2029
2030         op_data->op_fid2 = repbody->mbo_fid1;
2031         ptlrpc_req_finished(*request);
2032         *request = NULL;
2033
2034         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2035         if (IS_ERR(tgt))
2036                 RETURN(PTR_ERR(tgt));
2037
2038         op_data->op_mds = tgt->ltd_index;
2039         goto retry;
2040 }
2041
2042 static int
2043 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
2044             const union ldlm_policy_data *policy, struct md_op_data *op_data,
2045             struct lustre_handle *lockh, __u64 extra_lock_flags)
2046 {
2047         struct obd_device *obd = exp->exp_obd;
2048         struct lmv_obd *lmv = &obd->u.lmv;
2049         struct lmv_tgt_desc *tgt;
2050         int rc;
2051
2052         ENTRY;
2053
2054         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
2055
2056         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2057         if (IS_ERR(tgt))
2058                 RETURN(PTR_ERR(tgt));
2059
2060         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
2061                PFID(&op_data->op_fid1), tgt->ltd_index);
2062
2063         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
2064                         extra_lock_flags);
2065
2066         RETURN(rc);
2067 }
2068
2069 int
2070 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
2071                  struct ptlrpc_request **preq)
2072 {
2073         struct obd_device *obd = exp->exp_obd;
2074         struct lmv_obd *lmv = &obd->u.lmv;
2075         struct lmv_tgt_desc *tgt;
2076         struct mdt_body *body;
2077         int rc;
2078
2079         ENTRY;
2080
2081 retry:
2082         if (op_data->op_namelen == 2 &&
2083             op_data->op_name[0] == '.' && op_data->op_name[1] == '.')
2084                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2085         else
2086                 tgt = lmv_locate_tgt(lmv, op_data);
2087         if (IS_ERR(tgt))
2088                 RETURN(PTR_ERR(tgt));
2089
2090         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
2091                 (int)op_data->op_namelen, op_data->op_name,
2092                 PFID(&op_data->op_fid1), tgt->ltd_index);
2093
2094         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
2095         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2096                 ptlrpc_req_finished(*preq);
2097                 *preq = NULL;
2098                 goto retry;
2099         }
2100
2101         if (rc)
2102                 RETURN(rc);
2103
2104         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
2105         LASSERT(body != NULL);
2106
2107         if (body->mbo_valid & OBD_MD_MDS) {
2108                 op_data->op_fid1 = body->mbo_fid1;
2109                 op_data->op_valid |= OBD_MD_FLCROSSREF;
2110                 op_data->op_namelen = 0;
2111                 op_data->op_name = NULL;
2112
2113                 ptlrpc_req_finished(*preq);
2114                 *preq = NULL;
2115
2116                 goto retry;
2117         }
2118
2119         RETURN(rc);
2120 }
2121
2122 #define md_op_data_fid(op_data, fl)                     \
2123         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2124          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2125          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2126          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2127          NULL)
2128
2129 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2130                             struct md_op_data *op_data, __u32 op_tgt,
2131                             enum ldlm_mode mode, int bits, int flag)
2132 {
2133         struct lu_fid *fid = md_op_data_fid(op_data, flag);
2134         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
2135         union ldlm_policy_data policy = { { 0 } };
2136         int rc = 0;
2137         ENTRY;
2138
2139         if (!fid_is_sane(fid))
2140                 RETURN(0);
2141
2142         if (tgt == NULL) {
2143                 tgt = lmv_fid2tgt(lmv, fid);
2144                 if (IS_ERR(tgt))
2145                         RETURN(PTR_ERR(tgt));
2146         }
2147
2148         if (tgt->ltd_index != op_tgt) {
2149                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2150                 policy.l_inodebits.bits = bits;
2151                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2152                                       mode, LCF_ASYNC, NULL);
2153         } else {
2154                 CDEBUG(D_INODE,
2155                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2156                        op_tgt, PFID(fid));
2157                 op_data->op_flags |= flag;
2158                 rc = 0;
2159         }
2160
2161         RETURN(rc);
2162 }
2163
2164 /*
2165  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2166  * op_data->op_fid2
2167  */
2168 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2169                     struct ptlrpc_request **request)
2170 {
2171         struct obd_device       *obd = exp->exp_obd;
2172         struct lmv_obd          *lmv = &obd->u.lmv;
2173         struct lmv_tgt_desc     *tgt;
2174         int                      rc;
2175         ENTRY;
2176
2177         LASSERT(op_data->op_namelen != 0);
2178
2179         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2180                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2181                op_data->op_name, PFID(&op_data->op_fid1));
2182
2183         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2184         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2185         op_data->op_cap = current_cap();
2186
2187         tgt = lmv_locate_tgt2(lmv, op_data);
2188         if (IS_ERR(tgt))
2189                 RETURN(PTR_ERR(tgt));
2190
2191         /*
2192          * Cancel UPDATE lock on child (fid1).
2193          */
2194         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2195         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2196                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2197         if (rc != 0)
2198                 RETURN(rc);
2199
2200         rc = md_link(tgt->ltd_exp, op_data, request);
2201
2202         RETURN(rc);
2203 }
2204
2205 /* migrate the top directory */
2206 static inline bool lmv_op_topdir_migrate(const struct md_op_data *op_data)
2207 {
2208         if (!S_ISDIR(op_data->op_mode))
2209                 return false;
2210
2211         if (lmv_dir_layout_changing(op_data->op_mea1))
2212                 return false;
2213
2214         return true;
2215 }
2216
2217 /* migrate top dir to specific MDTs */
2218 static inline bool lmv_topdir_specific_migrate(const struct md_op_data *op_data)
2219 {
2220         const struct lmv_user_md *lum = op_data->op_data;
2221
2222         if (!lmv_op_topdir_migrate(op_data))
2223                 return false;
2224
2225         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2226 }
2227
2228 /* migrate top dir in QoS mode if user issued "lfs migrate -m -1..." */
2229 static inline bool lmv_topdir_qos_migrate(const struct md_op_data *op_data)
2230 {
2231         const struct lmv_user_md *lum = op_data->op_data;
2232
2233         if (!lmv_op_topdir_migrate(op_data))
2234                 return false;
2235
2236         return le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
2237 }
2238
2239 static inline bool lmv_subdir_specific_migrate(const struct md_op_data *op_data)
2240 {
2241         const struct lmv_user_md *lum = op_data->op_data;
2242
2243         if (!S_ISDIR(op_data->op_mode))
2244                 return false;
2245
2246         if (!lmv_dir_layout_changing(op_data->op_mea1))
2247                 return false;
2248
2249         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2250 }
2251
2252 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2253                         const char *name, size_t namelen,
2254                         struct ptlrpc_request **request)
2255 {
2256         struct obd_device *obd = exp->exp_obd;
2257         struct lmv_obd *lmv = &obd->u.lmv;
2258         struct lmv_stripe_md *lsm = op_data->op_mea1;
2259         struct lmv_tgt_desc *parent_tgt;
2260         struct lmv_tgt_desc *sp_tgt;
2261         struct lmv_tgt_desc *tp_tgt = NULL;
2262         struct lmv_tgt_desc *child_tgt;
2263         struct lmv_tgt_desc *tgt;
2264         struct lu_fid target_fid = { 0 };
2265         int rc;
2266
2267         ENTRY;
2268
2269         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2270
2271         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2272                PFID(&op_data->op_fid1), (int)namelen, name);
2273
2274         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2275         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2276         op_data->op_cap = current_cap();
2277
2278         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2279         if (IS_ERR(parent_tgt))
2280                 RETURN(PTR_ERR(parent_tgt));
2281
2282         if (lmv_dir_striped(lsm)) {
2283                 const struct lmv_oinfo *oinfo;
2284
2285                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
2286                 if (IS_ERR(oinfo))
2287                         RETURN(PTR_ERR(oinfo));
2288
2289                 /* save source stripe FID in fid4 temporarily for ELC */
2290                 op_data->op_fid4 = oinfo->lmo_fid;
2291                 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2292                 if (!sp_tgt)
2293                         RETURN(-ENODEV);
2294
2295                 /*
2296                  * if parent is being migrated too, fill op_fid2 with target
2297                  * stripe fid, otherwise the target stripe is not created yet.
2298                  */
2299                 if (lmv_dir_layout_changing(lsm)) {
2300                         oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
2301                                                         true);
2302                         if (IS_ERR(oinfo))
2303                                 RETURN(PTR_ERR(oinfo));
2304
2305                         op_data->op_fid2 = oinfo->lmo_fid;
2306                         tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2307                         if (!tp_tgt)
2308                                 RETURN(-ENODEV);
2309
2310                         /* parent unchanged and update namespace only */
2311                         if (lu_fid_eq(&op_data->op_fid4, &op_data->op_fid2) &&
2312                             op_data->op_bias & MDS_MIGRATE_NSONLY)
2313                                 RETURN(-EALREADY);
2314                 }
2315         } else {
2316                 sp_tgt = parent_tgt;
2317         }
2318
2319         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2320         if (IS_ERR(child_tgt))
2321                 RETURN(PTR_ERR(child_tgt));
2322
2323         if (lmv_topdir_specific_migrate(op_data)) {
2324                 struct lmv_user_md *lum = op_data->op_data;
2325
2326                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2327         } else if (lmv_topdir_qos_migrate(op_data)) {
2328                 tgt = lmv_locate_tgt_lf(lmv);
2329                 if (tgt == ERR_PTR(-EAGAIN))
2330                         tgt = lmv_locate_tgt_rr(lmv);
2331                 if (IS_ERR(tgt))
2332                         RETURN(PTR_ERR(tgt));
2333
2334                 op_data->op_mds = tgt->ltd_index;
2335         } else if (lmv_subdir_specific_migrate(op_data)) {
2336                 struct lmv_user_md *lum = op_data->op_data;
2337                 __u32 i;
2338
2339                 LASSERT(tp_tgt);
2340                 if (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) {
2341                         /* adjust MDTs in lum, since subdir is located on where
2342                          * its parent stripe is, not the first specified MDT.
2343                          */
2344                         for (i = 0; i < le32_to_cpu(lum->lum_stripe_count);
2345                              i++) {
2346                                 if (le32_to_cpu(lum->lum_objects[i].lum_mds) ==
2347                                     tp_tgt->ltd_index)
2348                                         break;
2349                         }
2350
2351                         if (i == le32_to_cpu(lum->lum_stripe_count))
2352                                 RETURN(-ENODEV);
2353
2354                         lum->lum_objects[i].lum_mds =
2355                                 lum->lum_objects[0].lum_mds;
2356                         lum->lum_objects[0].lum_mds =
2357                                 cpu_to_le32(tp_tgt->ltd_index);
2358                 }
2359                 /* NB, the above adjusts subdir migration for command like
2360                  * "lfs migrate -m 0,1,2 ...", but for migration like
2361                  * "lfs migrate -m 0 -c 2 ...", the top dir is migrated to MDT0
2362                  * and MDT1, however its subdir may be migrated to MDT1 and MDT2
2363                  */
2364
2365                 lum->lum_stripe_offset = cpu_to_le32(tp_tgt->ltd_index);
2366                 op_data->op_mds = tp_tgt->ltd_index;
2367         } else if (tp_tgt) {
2368                 op_data->op_mds = tp_tgt->ltd_index;
2369         } else {
2370                 op_data->op_mds = sp_tgt->ltd_index;
2371         }
2372
2373         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2374         if (rc)
2375                 RETURN(rc);
2376
2377         /*
2378          * for directory, send migrate request to the MDT where the object will
2379          * be migrated to, because we can't create a striped directory remotely.
2380          *
2381          * otherwise, send to the MDT where source is located because regular
2382          * file may open lease.
2383          *
2384          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2385          * backward compatibility.
2386          */
2387         if (S_ISDIR(op_data->op_mode) &&
2388             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2389                 tgt = lmv_fid2tgt(lmv, &target_fid);
2390                 if (IS_ERR(tgt))
2391                         RETURN(PTR_ERR(tgt));
2392         } else {
2393                 tgt = child_tgt;
2394         }
2395
2396         /* cancel UPDATE lock of parent master object */
2397         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2398                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2399         if (rc)
2400                 RETURN(rc);
2401
2402         /* cancel UPDATE lock of source parent */
2403         if (sp_tgt != parent_tgt) {
2404                 /*
2405                  * migrate RPC packs master object FID, because we can only pack
2406                  * two FIDs in reint RPC, but MDS needs to know both source
2407                  * parent and target parent, and it will obtain them from master
2408                  * FID and LMV, the other FID in RPC is kept for target.
2409                  *
2410                  * since this FID is not passed to MDC, cancel it anyway.
2411                  */
2412                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2413                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2414                 if (rc)
2415                         RETURN(rc);
2416
2417                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2418         }
2419         op_data->op_fid4 = target_fid;
2420
2421         /* cancel UPDATE locks of target parent */
2422         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2423                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2424         if (rc)
2425                 RETURN(rc);
2426
2427         /* cancel LOOKUP lock of source if source is remote object */
2428         if (child_tgt != sp_tgt) {
2429                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2430                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2431                                       MF_MDC_CANCEL_FID3);
2432                 if (rc)
2433                         RETURN(rc);
2434         }
2435
2436         /* cancel ELC locks of source */
2437         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2438                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2439         if (rc)
2440                 RETURN(rc);
2441
2442         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2443
2444         RETURN(rc);
2445 }
2446
2447 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2448                       const char *old, size_t oldlen,
2449                       const char *new, size_t newlen,
2450                       struct ptlrpc_request **request)
2451 {
2452         struct obd_device *obd = exp->exp_obd;
2453         struct lmv_obd *lmv = &obd->u.lmv;
2454         struct lmv_tgt_desc *sp_tgt;
2455         struct lmv_tgt_desc *tp_tgt = NULL;
2456         struct lmv_tgt_desc *src_tgt = NULL;
2457         struct lmv_tgt_desc *tgt;
2458         struct mdt_body *body;
2459         int rc;
2460
2461         ENTRY;
2462
2463         LASSERT(oldlen != 0);
2464
2465         if (op_data->op_cli_flags & CLI_MIGRATE) {
2466                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2467                 RETURN(rc);
2468         }
2469
2470         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2471         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2472         op_data->op_cap = current_cap();
2473
2474         op_data->op_name = new;
2475         op_data->op_namelen = newlen;
2476
2477         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2478         if (IS_ERR(tp_tgt))
2479                 RETURN(PTR_ERR(tp_tgt));
2480
2481         /* Since the target child might be destroyed, and it might become
2482          * orphan, and we can only check orphan on the local MDT right now, so
2483          * we send rename request to the MDT where target child is located. If
2484          * target child does not exist, then it will send the request to the
2485          * target parent */
2486         if (fid_is_sane(&op_data->op_fid4)) {
2487                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2488                 if (IS_ERR(tgt))
2489                         RETURN(PTR_ERR(tgt));
2490         } else {
2491                 tgt = tp_tgt;
2492         }
2493
2494         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2495
2496         /* cancel UPDATE locks of target parent */
2497         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2498                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2499         if (rc != 0)
2500                 RETURN(rc);
2501
2502         if (fid_is_sane(&op_data->op_fid4)) {
2503                 /* cancel LOOKUP lock of target on target parent */
2504                 if (tgt != tp_tgt) {
2505                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2506                                               tgt->ltd_index, LCK_EX,
2507                                               MDS_INODELOCK_LOOKUP,
2508                                               MF_MDC_CANCEL_FID4);
2509                         if (rc != 0)
2510                                 RETURN(rc);
2511                 }
2512         }
2513
2514         if (fid_is_sane(&op_data->op_fid3)) {
2515                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2516                 if (IS_ERR(src_tgt))
2517                         RETURN(PTR_ERR(src_tgt));
2518
2519                 /* cancel ELC locks of source */
2520                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2521                                       LCK_EX, MDS_INODELOCK_ELC,
2522                                       MF_MDC_CANCEL_FID3);
2523                 if (rc != 0)
2524                         RETURN(rc);
2525         }
2526
2527         op_data->op_name = old;
2528         op_data->op_namelen = oldlen;
2529 retry:
2530         sp_tgt = lmv_locate_tgt(lmv, op_data);
2531         if (IS_ERR(sp_tgt))
2532                 RETURN(PTR_ERR(sp_tgt));
2533
2534         /* cancel UPDATE locks of source parent */
2535         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2536                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2537         if (rc != 0)
2538                 RETURN(rc);
2539
2540         if (fid_is_sane(&op_data->op_fid3)) {
2541                 /* cancel LOOKUP lock of source on source parent */
2542                 if (src_tgt != sp_tgt) {
2543                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2544                                               tgt->ltd_index, LCK_EX,
2545                                               MDS_INODELOCK_LOOKUP,
2546                                               MF_MDC_CANCEL_FID3);
2547                         if (rc != 0)
2548                                 RETURN(rc);
2549                 }
2550         }
2551
2552 rename:
2553         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2554                 PFID(&op_data->op_fid1), (int)oldlen, old,
2555                 PFID(&op_data->op_fid2), (int)newlen, new);
2556
2557         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2558                         request);
2559         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2560                 ptlrpc_req_finished(*request);
2561                 *request = NULL;
2562                 goto retry;
2563         }
2564
2565         if (rc && rc != -EXDEV)
2566                 RETURN(rc);
2567
2568         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2569         if (body == NULL)
2570                 RETURN(-EPROTO);
2571
2572         /* Not cross-ref case, just get out of here. */
2573         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2574                 RETURN(rc);
2575
2576         op_data->op_fid4 = body->mbo_fid1;
2577
2578         ptlrpc_req_finished(*request);
2579         *request = NULL;
2580
2581         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2582         if (IS_ERR(tgt))
2583                 RETURN(PTR_ERR(tgt));
2584
2585         if (fid_is_sane(&op_data->op_fid4)) {
2586                 /* cancel LOOKUP lock of target on target parent */
2587                 if (tgt != tp_tgt) {
2588                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2589                                               tgt->ltd_index, LCK_EX,
2590                                               MDS_INODELOCK_LOOKUP,
2591                                               MF_MDC_CANCEL_FID4);
2592                         if (rc != 0)
2593                                 RETURN(rc);
2594                 }
2595         }
2596
2597         goto rename;
2598 }
2599
2600 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2601                        void *ea, size_t ealen, struct ptlrpc_request **request)
2602 {
2603         struct obd_device *obd = exp->exp_obd;
2604         struct lmv_obd *lmv = &obd->u.lmv;
2605         struct lmv_tgt_desc *tgt;
2606         int rc = 0;
2607
2608         ENTRY;
2609
2610         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2611                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2612                op_data->op_xvalid);
2613
2614         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2615         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2616         if (IS_ERR(tgt))
2617                 RETURN(PTR_ERR(tgt));
2618
2619         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2620
2621         RETURN(rc);
2622 }
2623
2624 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2625                      struct ptlrpc_request **request)
2626 {
2627         struct obd_device *obd = exp->exp_obd;
2628         struct lmv_obd *lmv = &obd->u.lmv;
2629         struct lmv_tgt_desc *tgt;
2630         int rc;
2631
2632         ENTRY;
2633
2634         tgt = lmv_fid2tgt(lmv, fid);
2635         if (IS_ERR(tgt))
2636                 RETURN(PTR_ERR(tgt));
2637
2638         rc = md_fsync(tgt->ltd_exp, fid, request);
2639         RETURN(rc);
2640 }
2641
2642 struct stripe_dirent {
2643         struct page             *sd_page;
2644         struct lu_dirpage       *sd_dp;
2645         struct lu_dirent        *sd_ent;
2646         bool                     sd_eof;
2647 };
2648
2649 struct lmv_dir_ctxt {
2650         struct lmv_obd          *ldc_lmv;
2651         struct md_op_data       *ldc_op_data;
2652         struct md_readdir_info  *ldc_mrinfo;
2653         __u64                    ldc_hash;
2654         int                      ldc_count;
2655         struct stripe_dirent     ldc_stripes[0];
2656 };
2657
2658 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2659 {
2660         if (stripe->sd_page) {
2661                 kunmap(stripe->sd_page);
2662                 put_page(stripe->sd_page);
2663                 stripe->sd_page = NULL;
2664                 stripe->sd_ent = NULL;
2665         }
2666 }
2667
2668 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2669 {
2670         int i;
2671
2672         for (i = 0; i < ctxt->ldc_count; i++)
2673                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2674 }
2675
2676 /* if @ent is dummy, or . .., get next */
2677 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2678                                            struct lu_dirent *ent,
2679                                            int stripe_index)
2680 {
2681         for (; ent; ent = lu_dirent_next(ent)) {
2682                 /* Skip dummy entry */
2683                 if (le16_to_cpu(ent->lde_namelen) == 0)
2684                         continue;
2685
2686                 /* skip . and .. for other stripes */
2687                 if (stripe_index &&
2688                     (strncmp(ent->lde_name, ".",
2689                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2690                      strncmp(ent->lde_name, "..",
2691                              le16_to_cpu(ent->lde_namelen)) == 0))
2692                         continue;
2693
2694                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2695                         break;
2696         }
2697
2698         return ent;
2699 }
2700
2701 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2702                                             struct stripe_dirent *stripe,
2703                                             int stripe_index)
2704 {
2705         struct md_op_data *op_data = ctxt->ldc_op_data;
2706         struct lmv_oinfo *oinfo;
2707         struct lu_fid fid = op_data->op_fid1;
2708         struct inode *inode = op_data->op_data;
2709         struct lmv_tgt_desc *tgt;
2710         struct lu_dirent *ent = stripe->sd_ent;
2711         __u64 hash = ctxt->ldc_hash;
2712         int rc = 0;
2713
2714         ENTRY;
2715
2716         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2717         LASSERT(!ent);
2718
2719         do {
2720                 if (stripe->sd_page) {
2721                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2722
2723                         /* @hash should be the last dirent hash */
2724                         LASSERTF(hash <= end,
2725                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2726                                  ctxt, stripe, hash, end);
2727                         /* unload last page */
2728                         stripe_dirent_unload(stripe);
2729                         /* eof */
2730                         if (end == MDS_DIR_END_OFF) {
2731                                 stripe->sd_eof = true;
2732                                 break;
2733                         }
2734                         hash = end;
2735                 }
2736
2737                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2738                 if (!oinfo->lmo_root) {
2739                         rc = -ENOENT;
2740                         break;
2741                 }
2742
2743                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2744                 if (!tgt) {
2745                         rc = -ENODEV;
2746                         break;
2747                 }
2748
2749                 /* op_data is shared by stripes, reset after use */
2750                 op_data->op_fid1 = oinfo->lmo_fid;
2751                 op_data->op_fid2 = oinfo->lmo_fid;
2752                 op_data->op_data = oinfo->lmo_root;
2753
2754                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_mrinfo, hash,
2755                                   &stripe->sd_page);
2756
2757                 op_data->op_fid1 = fid;
2758                 op_data->op_fid2 = fid;
2759                 op_data->op_data = inode;
2760
2761                 if (rc)
2762                         break;
2763
2764                 stripe->sd_dp = page_address(stripe->sd_page);
2765                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2766                                         stripe_index);
2767                 /* in case a page filled with ., .. and dummy, read next */
2768         } while (!ent);
2769
2770         stripe->sd_ent = ent;
2771         if (rc) {
2772                 LASSERT(!ent);
2773                 /* treat error as eof, so dir can be partially accessed */
2774                 stripe->sd_eof = true;
2775                 ctxt->ldc_mrinfo->mr_partial_readdir_rc = rc;
2776                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2777                               "directory is partially accessed!\n",
2778                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2779                               rc);
2780         }
2781
2782         RETURN(ent);
2783 }
2784
2785 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2786 {
2787         struct obd_device *obd = exp->exp_obd;
2788         struct lmv_obd *lmv = &obd->u.lmv;
2789         struct lmv_tgt_desc *tgt;
2790         int rc;
2791
2792         ENTRY;
2793
2794         rc = lmv_check_connect(obd);
2795         if (rc != 0)
2796                 RETURN(rc);
2797
2798         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2799         if (IS_ERR(tgt))
2800                 RETURN(PTR_ERR(tgt));
2801
2802         data->op_flags |= MF_MDC_CANCEL_FID1;
2803         rc = md_file_resync(tgt->ltd_exp, data);
2804         RETURN(rc);
2805 }
2806
2807 /**
2808  * Get dirent with the closest hash for striped directory
2809  *
2810  * This function will search the dir entry, whose hash value is the
2811  * closest(>=) to hash from all of sub-stripes, and it is only being called
2812  * for striped directory.
2813  *
2814  * \param[in] ctxt              dir read context
2815  *
2816  * \retval                      dirent get the entry successfully
2817  *                              NULL does not get the entry, normally it means
2818  *                              it reaches the end of the directory, while read
2819  *                              stripe dirent error is ignored to allow partial
2820  *                              access.
2821  */
2822 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2823 {
2824         struct stripe_dirent *stripe;
2825         struct lu_dirent *ent = NULL;
2826         int i;
2827         int min = -1;
2828
2829         /* TODO: optimize with k-way merge sort */
2830         for (i = 0; i < ctxt->ldc_count; i++) {
2831                 stripe = &ctxt->ldc_stripes[i];
2832                 if (stripe->sd_eof)
2833                         continue;
2834
2835                 if (!stripe->sd_ent) {
2836                         stripe_dirent_load(ctxt, stripe, i);
2837                         if (!stripe->sd_ent) {
2838                                 LASSERT(stripe->sd_eof);
2839                                 continue;
2840                         }
2841                 }
2842
2843                 if (min == -1 ||
2844                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2845                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2846                         min = i;
2847                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2848                             ctxt->ldc_hash)
2849                                 break;
2850                 }
2851         }
2852
2853         if (min != -1) {
2854                 stripe = &ctxt->ldc_stripes[min];
2855                 ent = stripe->sd_ent;
2856                 /* pop found dirent */
2857                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2858                                                    min);
2859         }
2860
2861         return ent;
2862 }
2863
2864 /**
2865  * Build dir entry page for striped directory
2866  *
2867  * This function gets one entry by @offset from a striped directory. It will
2868  * read entries from all of stripes, and choose one closest to the required
2869  * offset(&offset). A few notes
2870  * 1. skip . and .. for non-zero stripes, because there can only have one .
2871  * and .. in a directory.
2872  * 2. op_data will be shared by all of stripes, instead of allocating new
2873  * one, so need to restore before reusing.
2874  *
2875  * \param[in] exp       obd export refer to LMV
2876  * \param[in] op_data   hold those MD parameters of read_entry
2877  * \param[in] mrinfo    ldlm callback being used in enqueue in mdc_read_entry,
2878  *                      and partial readdir result will be stored in it.
2879  * \param[in] offset    starting hash offset
2880  * \param[out] ppage    the page holding the entry. Note: because the entry
2881  *                      will be accessed in upper layer, so we need hold the
2882  *                      page until the usages of entry is finished, see
2883  *                      ll_dir_entry_next.
2884  *
2885  * retval               =0 if get entry successfully
2886  *                      <0 cannot get entry
2887  */
2888 static int lmv_striped_read_page(struct obd_export *exp,
2889                                  struct md_op_data *op_data,
2890                                  struct md_readdir_info *mrinfo, __u64 offset,
2891                                  struct page **ppage)
2892 {
2893         struct page *page = NULL;
2894         struct lu_dirpage *dp;
2895         void *start;
2896         struct lu_dirent *ent;
2897         struct lu_dirent *last_ent;
2898         int stripe_count;
2899         struct lmv_dir_ctxt *ctxt;
2900         struct lu_dirent *next = NULL;
2901         __u16 ent_size;
2902         size_t left_bytes;
2903         int rc = 0;
2904         ENTRY;
2905
2906         /* Allocate a page and read entries from all of stripes and fill
2907          * the page by hash order */
2908         page = alloc_page(GFP_KERNEL);
2909         if (!page)
2910                 RETURN(-ENOMEM);
2911
2912         /* Initialize the entry page */
2913         dp = kmap(page);
2914         memset(dp, 0, sizeof(*dp));
2915         dp->ldp_hash_start = cpu_to_le64(offset);
2916
2917         start = dp + 1;
2918         left_bytes = PAGE_SIZE - sizeof(*dp);
2919         ent = start;
2920         last_ent = ent;
2921
2922         /* initalize dir read context */
2923         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2924         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2925         if (!ctxt)
2926                 GOTO(free_page, rc = -ENOMEM);
2927         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2928         ctxt->ldc_op_data = op_data;
2929         ctxt->ldc_mrinfo = mrinfo;
2930         ctxt->ldc_hash = offset;
2931         ctxt->ldc_count = stripe_count;
2932
2933         while (1) {
2934                 next = lmv_dirent_next(ctxt);
2935
2936                 /* end of directory */
2937                 if (!next) {
2938                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2939                         break;
2940                 }
2941                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2942
2943                 ent_size = le16_to_cpu(next->lde_reclen);
2944
2945                 /* the last entry lde_reclen is 0, but it might not be the last
2946                  * one of this temporay dir page */
2947                 if (!ent_size)
2948                         ent_size = lu_dirent_calc_size(
2949                                         le16_to_cpu(next->lde_namelen),
2950                                         le32_to_cpu(next->lde_attrs));
2951                 /* page full */
2952                 if (ent_size > left_bytes)
2953                         break;
2954
2955                 memcpy(ent, next, ent_size);
2956
2957                 /* Replace . with master FID and Replace .. with the parent FID
2958                  * of master object */
2959                 if (strncmp(ent->lde_name, ".",
2960                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2961                     le16_to_cpu(ent->lde_namelen) == 1)
2962                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2963                 else if (strncmp(ent->lde_name, "..",
2964                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2965                            le16_to_cpu(ent->lde_namelen) == 2)
2966                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2967
2968                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2969                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2970                        le64_to_cpu(ent->lde_hash));
2971
2972                 left_bytes -= ent_size;
2973                 ent->lde_reclen = cpu_to_le16(ent_size);
2974                 last_ent = ent;
2975                 ent = (void *)ent + ent_size;
2976         };
2977
2978         last_ent->lde_reclen = 0;
2979
2980         if (ent == start)
2981                 dp->ldp_flags |= LDF_EMPTY;
2982         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2983                 dp->ldp_flags |= LDF_COLLIDE;
2984         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2985         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2986
2987         put_lmv_dir_ctxt(ctxt);
2988         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2989
2990         *ppage = page;
2991
2992         RETURN(0);
2993
2994 free_page:
2995         kunmap(page);
2996         __free_page(page);
2997
2998         return rc;
2999 }
3000
3001 static int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
3002                          struct md_readdir_info *mrinfo, __u64 offset,
3003                          struct page **ppage)
3004 {
3005         struct obd_device *obd = exp->exp_obd;
3006         struct lmv_obd *lmv = &obd->u.lmv;
3007         struct lmv_tgt_desc *tgt;
3008         int rc;
3009
3010         ENTRY;
3011
3012         if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
3013                 RETURN(-ENODATA);
3014
3015         if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
3016                 rc = lmv_striped_read_page(exp, op_data, mrinfo, offset, ppage);
3017                 RETURN(rc);
3018         }
3019
3020         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
3021         if (IS_ERR(tgt))
3022                 RETURN(PTR_ERR(tgt));
3023
3024         rc = md_read_page(tgt->ltd_exp, op_data, mrinfo, offset, ppage);
3025
3026         RETURN(rc);
3027 }
3028
3029 /**
3030  * Unlink a file/directory
3031  *
3032  * Unlink a file or directory under the parent dir. The unlink request
3033  * usually will be sent to the MDT where the child is located, but if
3034  * the client does not have the child FID then request will be sent to the
3035  * MDT where the parent is located.
3036  *
3037  * If the parent is a striped directory then it also needs to locate which
3038  * stripe the name of the child is located, and replace the parent FID
3039  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
3040  * it will walk through all of sub-stripes until the child is being
3041  * unlinked finally.
3042  *
3043  * \param[in] exp       export refer to LMV
3044  * \param[in] op_data   different parameters transferred beween client
3045  *                      MD stacks, name, namelen, FIDs etc.
3046  *                      op_fid1 is the parent FID, op_fid2 is the child
3047  *                      FID.
3048  * \param[out] request  point to the request of unlink.
3049  *
3050  * retval               0 if succeed
3051  *                      negative errno if failed.
3052  */
3053 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
3054                       struct ptlrpc_request **request)
3055 {
3056         struct obd_device *obd = exp->exp_obd;
3057         struct lmv_obd *lmv = &obd->u.lmv;
3058         struct lmv_tgt_desc *tgt;
3059         struct lmv_tgt_desc *parent_tgt;
3060         struct mdt_body *body;
3061         int rc;
3062
3063         ENTRY;
3064
3065         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
3066         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
3067         op_data->op_cap = current_cap();
3068
3069 retry:
3070         parent_tgt = lmv_locate_tgt(lmv, op_data);
3071         if (IS_ERR(parent_tgt))
3072                 RETURN(PTR_ERR(parent_tgt));
3073
3074         if (likely(!fid_is_zero(&op_data->op_fid2))) {
3075                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3076                 if (IS_ERR(tgt))
3077                         RETURN(PTR_ERR(tgt));
3078         } else {
3079                 tgt = parent_tgt;
3080         }
3081
3082         /*
3083          * If child's fid is given, cancel unused locks for it if it is from
3084          * another export than parent.
3085          *
3086          * LOOKUP lock for child (fid3) should also be cancelled on parent
3087          * tgt_tgt in mdc_unlink().
3088          */
3089         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
3090
3091         if (parent_tgt != tgt)
3092                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
3093                                       LCK_EX, MDS_INODELOCK_LOOKUP,
3094                                       MF_MDC_CANCEL_FID3);
3095
3096         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
3097                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
3098         if (rc)
3099                 RETURN(rc);
3100
3101         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
3102                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
3103                tgt->ltd_index);
3104
3105         rc = md_unlink(tgt->ltd_exp, op_data, request);
3106         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
3107                 ptlrpc_req_finished(*request);
3108                 *request = NULL;
3109                 goto retry;
3110         }
3111
3112         if (rc != -EREMOTE)
3113                 RETURN(rc);
3114
3115         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
3116         if (body == NULL)
3117                 RETURN(-EPROTO);
3118
3119         /* Not cross-ref case, just get out of here. */
3120         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
3121                 RETURN(rc);
3122
3123         /* This is a remote object, try remote MDT. */
3124         op_data->op_fid2 = body->mbo_fid1;
3125         ptlrpc_req_finished(*request);
3126         *request = NULL;
3127
3128         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3129         if (IS_ERR(tgt))
3130                 RETURN(PTR_ERR(tgt));
3131
3132         goto retry;
3133 }
3134
3135 static int lmv_precleanup(struct obd_device *obd)
3136 {
3137         ENTRY;
3138         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
3139         RETURN(0);
3140 }
3141
3142 /**
3143  * Get by key a value associated with a LMV device.
3144  *
3145  * Dispatch request to lower-layer devices as needed.
3146  *
3147  * \param[in] env               execution environment for this thread
3148  * \param[in] exp               export for the LMV device
3149  * \param[in] keylen            length of key identifier
3150  * \param[in] key               identifier of key to get value for
3151  * \param[in] vallen            size of \a val
3152  * \param[out] val              pointer to storage location for value
3153  * \param[in] lsm               optional striping metadata of object
3154  *
3155  * \retval 0            on success
3156  * \retval negative     negated errno on failure
3157  */
3158 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
3159                         __u32 keylen, void *key, __u32 *vallen, void *val)
3160 {
3161         struct obd_device *obd;
3162         struct lmv_obd *lmv;
3163         struct lu_tgt_desc *tgt;
3164         int rc = 0;
3165
3166         ENTRY;
3167
3168         obd = class_exp2obd(exp);
3169         if (obd == NULL) {
3170                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3171                        exp->exp_handle.h_cookie);
3172                 RETURN(-EINVAL);
3173         }
3174
3175         lmv = &obd->u.lmv;
3176         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
3177                 LASSERT(*vallen == sizeof(__u32));
3178                 lmv_foreach_connected_tgt(lmv, tgt) {
3179                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
3180                                           vallen, val))
3181                                 RETURN(0);
3182                 }
3183                 RETURN(-EINVAL);
3184         } else if (KEY_IS(KEY_MAX_EASIZE) ||
3185                    KEY_IS(KEY_DEFAULT_EASIZE) ||
3186                    KEY_IS(KEY_CONN_DATA)) {
3187                 /*
3188                  * Forwarding this request to first MDS, it should know LOV
3189                  * desc.
3190                  */
3191                 tgt = lmv_tgt(lmv, 0);
3192                 if (!tgt)
3193                         RETURN(-ENODEV);
3194
3195                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
3196                 if (!rc && KEY_IS(KEY_CONN_DATA))
3197                         exp->exp_connect_data = *(struct obd_connect_data *)val;
3198                 RETURN(rc);
3199         } else if (KEY_IS(KEY_TGT_COUNT)) {
3200                 *((int *)val) = lmv->lmv_mdt_descs.ltd_tgts_size;
3201                 RETURN(0);
3202         }
3203
3204         CDEBUG(D_IOCTL, "Invalid key\n");
3205         RETURN(-EINVAL);
3206 }
3207
3208 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
3209                      int *__rcs, struct ptlrpc_request_set *_set)
3210 {
3211         struct obd_device *obd = class_exp2obd(exp);
3212         struct ptlrpc_request_set *set = _set;
3213         struct lmv_obd *lmv = &obd->u.lmv;
3214         int tgt_count = lmv->lmv_mdt_count;
3215         struct lu_tgt_desc *tgt;
3216         struct fid_array *fat, **fas = NULL;
3217         int i, rc, **rcs = NULL;
3218
3219         if (!set) {
3220                 set = ptlrpc_prep_set();
3221                 if (!set)
3222                         RETURN(-ENOMEM);
3223         }
3224
3225         /* split FIDs by targets */
3226         OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
3227         if (fas == NULL)
3228                 GOTO(out, rc = -ENOMEM);
3229         OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
3230         if (rcs == NULL)
3231                 GOTO(out_fas, rc = -ENOMEM);
3232
3233         for (i = 0; i < fa->fa_nr; i++) {
3234                 unsigned int idx;
3235
3236                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
3237                 if (rc) {
3238                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
3239                                PFID(&fa->fa_fids[i]), rc);
3240                         continue;
3241                 }
3242                 LASSERT(idx < tgt_count);
3243                 if (!fas[idx])
3244                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
3245                                   fa_fids[fa->fa_nr]));
3246                 if (!fas[idx])
3247                         GOTO(out, rc = -ENOMEM);
3248                 if (!rcs[idx])
3249                         OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
3250                 if (!rcs[idx])
3251                         GOTO(out, rc = -ENOMEM);
3252
3253                 fat = fas[idx];
3254                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
3255         }
3256
3257         lmv_foreach_connected_tgt(lmv, tgt) {
3258                 fat = fas[tgt->ltd_index];
3259                 if (!fat || fat->fa_nr == 0)
3260                         continue;
3261                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3262         }
3263
3264         rc = ptlrpc_set_wait(NULL, set);
3265         if (rc == 0) {
3266                 int j = 0;
3267                 for (i = 0; i < tgt_count; i++) {
3268                         fat = fas[i];
3269                         if (!fat || fat->fa_nr == 0)
3270                                 continue;
3271                         /* copy FIDs back */
3272                         memcpy(fa->fa_fids + j, fat->fa_fids,
3273                                fat->fa_nr * sizeof(struct lu_fid));
3274                         /* copy rcs back */
3275                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3276                         j += fat->fa_nr;
3277                 }
3278         }
3279         if (set != _set)
3280                 ptlrpc_set_destroy(set);
3281
3282 out:
3283         for (i = 0; i < tgt_count; i++) {
3284                 if (fas && fas[i])
3285                         OBD_FREE(fas[i], offsetof(struct fid_array,
3286                                                 fa_fids[fa->fa_nr]));
3287                 if (rcs && rcs[i])
3288                         OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3289         }
3290         if (rcs)
3291                 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3292 out_fas:
3293         if (fas)
3294                 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3295
3296         RETURN(rc);
3297 }
3298
3299 /**
3300  * Asynchronously set by key a value associated with a LMV device.
3301  *
3302  * Dispatch request to lower-layer devices as needed.
3303  *
3304  * \param[in] env       execution environment for this thread
3305  * \param[in] exp       export for the LMV device
3306  * \param[in] keylen    length of key identifier
3307  * \param[in] key       identifier of key to store value for
3308  * \param[in] vallen    size of value to store
3309  * \param[in] val       pointer to data to be stored
3310  * \param[in] set       optional list of related ptlrpc requests
3311  *
3312  * \retval 0            on success
3313  * \retval negative     negated errno on failure
3314  */
3315 static int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3316                               __u32 keylen, void *key, __u32 vallen, void *val,
3317                               struct ptlrpc_request_set *set)
3318 {
3319         struct lmv_tgt_desc *tgt;
3320         struct obd_device *obd;
3321         struct lmv_obd *lmv;
3322         int rc = 0;
3323         ENTRY;
3324
3325         obd = class_exp2obd(exp);
3326         if (obd == NULL) {
3327                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3328                        exp->exp_handle.h_cookie);
3329                 RETURN(-EINVAL);
3330         }
3331         lmv = &obd->u.lmv;
3332
3333         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3334             KEY_IS(KEY_DEFAULT_EASIZE)) {
3335                 int err = 0;
3336
3337                 lmv_foreach_connected_tgt(lmv, tgt) {
3338                         err = obd_set_info_async(env, tgt->ltd_exp,
3339                                                  keylen, key, vallen, val, set);
3340                         if (err && rc == 0)
3341                                 rc = err;
3342                 }
3343
3344                 RETURN(rc);
3345         }
3346
3347         RETURN(-EINVAL);
3348 }
3349
3350 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3351                             const struct lmv_mds_md_v1 *lmm1)
3352 {
3353         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3354         int             stripe_count;
3355         int             cplen;
3356         int             i;
3357         int             rc = 0;
3358         ENTRY;
3359
3360         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3361         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3362         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3363         if (CFS_FAIL_CHECK(OBD_FAIL_LMV_UNKNOWN_STRIPE))
3364                 lsm->lsm_md_hash_type = cfs_fail_val ?: LMV_HASH_TYPE_UNKNOWN;
3365         else
3366                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3367         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3368         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3369         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3370         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3371                         sizeof(lsm->lsm_md_pool_name));
3372
3373         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3374                 RETURN(-E2BIG);
3375
3376         CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3377                "layout_version %d\n", lsm->lsm_md_stripe_count,
3378                lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3379                lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3380                lsm->lsm_md_layout_version);
3381
3382         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3383         for (i = 0; i < stripe_count; i++) {
3384                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3385                               &lmm1->lmv_stripe_fids[i]);
3386                 /*
3387                  * set default value -1, so lmv_locate_tgt() knows this stripe
3388                  * target is not initialized.
3389                  */
3390                 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3391                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3392                         continue;
3393
3394                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3395                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3396                 if (rc == -ENOENT)
3397                         continue;
3398
3399                 if (rc)
3400                         RETURN(rc);
3401
3402                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3403                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3404         }
3405
3406         RETURN(rc);
3407 }
3408
3409 static inline int lmv_unpack_user_md(struct obd_export *exp,
3410                                      struct lmv_stripe_md *lsm,
3411                                      const struct lmv_user_md *lmu)
3412 {
3413         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3414         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3415         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3416         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3417         lsm->lsm_md_max_inherit = lmu->lum_max_inherit;
3418         lsm->lsm_md_max_inherit_rr = lmu->lum_max_inherit_rr;
3419         lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3420
3421         return 0;
3422 }
3423
3424 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3425                         const union lmv_mds_md *lmm, size_t lmm_size)
3426 {
3427         struct lmv_stripe_md     *lsm;
3428         int                      lsm_size;
3429         int                      rc;
3430         bool                     allocated = false;
3431         ENTRY;
3432
3433         LASSERT(lsmp != NULL);
3434
3435         lsm = *lsmp;
3436         /* Free memmd */
3437         if (lsm != NULL && lmm == NULL) {
3438                 int i;
3439                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3440
3441                 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3442                         size_t lfm_size;
3443
3444                         lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3445                                                               lfm_value[0]);
3446                         OBD_FREE_LARGE(lfm, lfm_size);
3447                         RETURN(0);
3448                 }
3449
3450                 if (lmv_dir_striped(lsm)) {
3451                         for (i = 0; i < lsm->lsm_md_stripe_count; i++)
3452                                 iput(lsm->lsm_md_oinfo[i].lmo_root);
3453                         lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3454                 } else {
3455                         lsm_size = lmv_stripe_md_size(0);
3456                 }
3457                 OBD_FREE(lsm, lsm_size);
3458                 *lsmp = NULL;
3459                 RETURN(0);
3460         }
3461
3462         /* foreign lmv case */
3463         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3464                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3465
3466                 if (lfm == NULL) {
3467                         OBD_ALLOC_LARGE(lfm, lmm_size);
3468                         if (lfm == NULL)
3469                                 RETURN(-ENOMEM);
3470                         *lsmp = (struct lmv_stripe_md *)lfm;
3471                 }
3472                 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3473                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3474                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3475                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3476                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3477                        lfm->lfm_length);
3478                 RETURN(lmm_size);
3479         }
3480
3481         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3482                 RETURN(-EPERM);
3483
3484         /* Unpack memmd */
3485         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3486             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3487                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3488                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3489                        -EIO);
3490                 RETURN(-EIO);
3491         }
3492
3493         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3494                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3495         else
3496                 /**
3497                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3498                  * stripecount should be 0 then.
3499                  */
3500                 lsm_size = lmv_stripe_md_size(0);
3501
3502         if (lsm == NULL) {
3503                 OBD_ALLOC(lsm, lsm_size);
3504                 if (lsm == NULL)
3505                         RETURN(-ENOMEM);
3506                 allocated = true;
3507                 *lsmp = lsm;
3508         }
3509
3510         switch (le32_to_cpu(lmm->lmv_magic)) {
3511         case LMV_MAGIC_V1:
3512                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3513                 break;
3514         case LMV_USER_MAGIC:
3515                 rc = lmv_unpack_user_md(exp, lsm, &lmm->lmv_user_md);
3516                 break;
3517         default:
3518                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3519                        le32_to_cpu(lmm->lmv_magic));
3520                 rc = -EINVAL;
3521                 break;
3522         }
3523
3524         if (rc != 0 && allocated) {
3525                 OBD_FREE(lsm, lsm_size);
3526                 *lsmp = NULL;
3527                 lsm_size = rc;
3528         }
3529         RETURN(lsm_size);
3530 }
3531
3532 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3533 {
3534         lmv_unpackmd(NULL, &lsm, NULL, 0);
3535 }
3536 EXPORT_SYMBOL(lmv_free_memmd);
3537
3538 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3539                              union ldlm_policy_data *policy,
3540                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3541                              void *opaque)
3542 {
3543         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3544         struct lu_tgt_desc *tgt;
3545         int err;
3546         int rc = 0;
3547
3548         ENTRY;
3549
3550         LASSERT(fid != NULL);
3551
3552         lmv_foreach_connected_tgt(lmv, tgt) {
3553                 if (!tgt->ltd_active)
3554                         continue;
3555
3556                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3557                                        opaque);
3558                 if (!rc)
3559                         rc = err;
3560         }
3561         RETURN(rc);
3562 }
3563
3564 static int lmv_set_lock_data(struct obd_export *exp,
3565                              const struct lustre_handle *lockh,
3566                              void *data, __u64 *bits)
3567 {
3568         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3569         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3570         int rc;
3571
3572         ENTRY;
3573
3574         if (tgt == NULL || tgt->ltd_exp == NULL)
3575                 RETURN(-EINVAL);
3576         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3577         RETURN(rc);
3578 }
3579
3580 static enum ldlm_mode
3581 lmv_lock_match(struct obd_export *exp, __u64 flags,
3582                const struct lu_fid *fid, enum ldlm_type type,
3583                union ldlm_policy_data *policy,
3584                enum ldlm_mode mode, struct lustre_handle *lockh)
3585 {
3586         struct obd_device *obd = exp->exp_obd;
3587         struct lmv_obd *lmv = &obd->u.lmv;
3588         enum ldlm_mode rc;
3589         struct lu_tgt_desc *tgt;
3590         int i;
3591         int index;
3592
3593         ENTRY;
3594
3595         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
3596
3597         /*
3598          * With DNE every object can have two locks in different namespaces:
3599          * lookup lock in space of MDT storing direntry and update/open lock in
3600          * space of MDT storing inode.  Try the MDT that the FID maps to first,
3601          * since this can be easily found, and only try others if that fails.
3602          */
3603         for (i = 0, index = lmv_fid2tgt_index(lmv, fid);
3604              i < lmv->lmv_mdt_descs.ltd_tgts_size;
3605              i++, index = (index + 1) % lmv->lmv_mdt_descs.ltd_tgts_size) {
3606                 if (index < 0) {
3607                         CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
3608                                obd->obd_name, PFID(fid), index);
3609                         index = 0;
3610                 }
3611
3612                 tgt = lmv_tgt(lmv, index);
3613                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
3614                         continue;
3615
3616                 rc = md_lock_match(tgt->ltd_exp, flags, fid, type, policy, mode,
3617                                    lockh);
3618                 if (rc)
3619                         RETURN(rc);
3620         }
3621
3622         RETURN(0);
3623 }
3624
3625 static int
3626 lmv_get_lustre_md(struct obd_export *exp, struct req_capsule *pill,
3627                   struct obd_export *dt_exp, struct obd_export *md_exp,
3628                   struct lustre_md *md)
3629 {
3630         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3631         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3632
3633         if (!tgt || !tgt->ltd_exp)
3634                 return -EINVAL;
3635
3636         return md_get_lustre_md(tgt->ltd_exp, pill, dt_exp, md_exp, md);
3637 }
3638
3639 static int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
3640 {
3641         struct obd_device *obd = exp->exp_obd;
3642         struct lmv_obd *lmv = &obd->u.lmv;
3643         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3644
3645         ENTRY;
3646
3647         if (md->default_lmv) {
3648                 lmv_free_memmd(md->default_lmv);
3649                 md->default_lmv = NULL;
3650         }
3651         if (md->lmv != NULL) {
3652                 lmv_free_memmd(md->lmv);
3653                 md->lmv = NULL;
3654         }
3655         if (!tgt || !tgt->ltd_exp)
3656                 RETURN(-EINVAL);
3657         RETURN(md_free_lustre_md(tgt->ltd_exp, md));
3658 }
3659
3660 static int lmv_set_open_replay_data(struct obd_export *exp,
3661                                     struct obd_client_handle *och,
3662                                     struct lookup_intent *it)
3663 {
3664         struct obd_device *obd = exp->exp_obd;
3665         struct lmv_obd *lmv = &obd->u.lmv;
3666         struct lmv_tgt_desc *tgt;
3667
3668         ENTRY;
3669
3670         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3671         if (IS_ERR(tgt))
3672                 RETURN(PTR_ERR(tgt));
3673
3674         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3675 }
3676
3677 static int lmv_clear_open_replay_data(struct obd_export *exp,
3678                                       struct obd_client_handle *och)
3679 {
3680         struct obd_device *obd = exp->exp_obd;
3681         struct lmv_obd *lmv = &obd->u.lmv;
3682         struct lmv_tgt_desc *tgt;
3683
3684         ENTRY;
3685
3686         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3687         if (IS_ERR(tgt))
3688                 RETURN(PTR_ERR(tgt));
3689
3690         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3691 }
3692
3693 static int lmv_intent_getattr_async(struct obd_export *exp,
3694                                     struct md_enqueue_info *minfo)
3695 {
3696         struct md_op_data *op_data = &minfo->mi_data;
3697         struct obd_device *obd = exp->exp_obd;
3698         struct lmv_obd *lmv = &obd->u.lmv;
3699         struct lmv_tgt_desc *ptgt;
3700         struct lmv_tgt_desc *ctgt;
3701         int rc;
3702
3703         ENTRY;
3704
3705         if (!fid_is_sane(&op_data->op_fid2))
3706                 RETURN(-EINVAL);
3707
3708         ptgt = lmv_locate_tgt(lmv, op_data);
3709         if (IS_ERR(ptgt))
3710                 RETURN(PTR_ERR(ptgt));
3711
3712         ctgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3713         if (IS_ERR(ctgt))
3714                 RETURN(PTR_ERR(ctgt));
3715
3716         /*
3717          * remote object needs two RPCs to lookup and getattr, considering the
3718          * complexity don't support statahead for now.
3719          */
3720         if (ctgt != ptgt)
3721                 RETURN(-EREMOTE);
3722
3723         rc = md_intent_getattr_async(ptgt->ltd_exp, minfo);
3724
3725         RETURN(rc);
3726 }
3727
3728 static int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3729                                struct lu_fid *fid, __u64 *bits)
3730 {
3731         struct obd_device *obd = exp->exp_obd;
3732         struct lmv_obd *lmv = &obd->u.lmv;
3733         struct lmv_tgt_desc *tgt;
3734         int rc;
3735
3736         ENTRY;
3737
3738         tgt = lmv_fid2tgt(lmv, fid);
3739         if (IS_ERR(tgt))
3740                 RETURN(PTR_ERR(tgt));
3741
3742         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3743         RETURN(rc);
3744 }
3745
3746 static int lmv_get_fid_from_lsm(struct obd_export *exp,
3747                                 const struct lmv_stripe_md *lsm,
3748                                 const char *name, int namelen,
3749                                 struct lu_fid *fid)
3750 {
3751         const struct lmv_oinfo *oinfo;
3752
3753         if (!lmv_dir_striped(lsm))
3754                 RETURN(-ESTALE);
3755
3756         oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
3757         if (IS_ERR(oinfo))
3758                 return PTR_ERR(oinfo);
3759
3760         *fid = oinfo->lmo_fid;
3761
3762         RETURN(0);
3763 }
3764
3765 /**
3766  * For lmv, only need to send request to master MDT, and the master MDT will
3767  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3768  * we directly fetch data from the slave MDTs.
3769  */
3770 static int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3771                         struct obd_quotactl *oqctl)
3772 {
3773         struct obd_device *obd = class_exp2obd(exp);
3774         struct lmv_obd *lmv = &obd->u.lmv;
3775         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3776         __u64 curspace, curinodes;
3777         int rc = 0;
3778
3779         ENTRY;
3780
3781         if (!tgt || !tgt->ltd_exp || !tgt->ltd_active) {
3782                 CERROR("master lmv inactive\n");
3783                 RETURN(-EIO);
3784         }
3785
3786         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3787                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3788                 RETURN(rc);
3789         }
3790
3791         curspace = curinodes = 0;
3792         lmv_foreach_connected_tgt(lmv, tgt) {
3793                 int err;
3794
3795                 if (!tgt->ltd_active)
3796                         continue;
3797
3798                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3799                 if (err) {
3800                         CERROR("getquota on mdt %d failed. %d\n",
3801                                tgt->ltd_index, err);
3802                         if (!rc)
3803                                 rc = err;
3804                 } else {
3805                         curspace += oqctl->qc_dqblk.dqb_curspace;
3806                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3807                 }
3808         }
3809         oqctl->qc_dqblk.dqb_curspace = curspace;
3810         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3811
3812         RETURN(rc);
3813 }
3814
3815 static int lmv_merge_attr(struct obd_export *exp,
3816                           const struct lmv_stripe_md *lsm,
3817                           struct cl_attr *attr,
3818                           ldlm_blocking_callback cb_blocking)
3819 {
3820         int rc;
3821         int i;
3822
3823         if (!lmv_dir_striped(lsm))
3824                 return 0;
3825
3826         rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3827         if (rc < 0)
3828                 return rc;
3829
3830         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3831                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3832
3833                 if (!inode)
3834                         continue;
3835
3836                 CDEBUG(D_INFO,
3837                        "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
3838                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3839                        i_size_read(inode), (unsigned long long)inode->i_blocks,
3840                        inode->i_nlink, (s64)inode->i_atime.tv_sec,
3841                        (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
3842
3843                 /* for slave stripe, it needs to subtract nlink for . and .. */
3844                 if (i != 0)
3845                         attr->cat_nlink += inode->i_nlink - 2;
3846                 else
3847                         attr->cat_nlink = inode->i_nlink;
3848
3849                 attr->cat_size += i_size_read(inode);
3850                 attr->cat_blocks += inode->i_blocks;
3851
3852                 if (attr->cat_atime < inode->i_atime.tv_sec)
3853                         attr->cat_atime = inode->i_atime.tv_sec;
3854
3855                 if (attr->cat_ctime < inode->i_ctime.tv_sec)
3856                         attr->cat_ctime = inode->i_ctime.tv_sec;
3857
3858                 if (attr->cat_mtime < inode->i_mtime.tv_sec)
3859                         attr->cat_mtime = inode->i_mtime.tv_sec;
3860         }
3861         return 0;
3862 }
3863
3864 static const struct obd_ops lmv_obd_ops = {
3865         .o_owner                = THIS_MODULE,
3866         .o_setup                = lmv_setup,
3867         .o_cleanup              = lmv_cleanup,
3868         .o_precleanup           = lmv_precleanup,
3869         .o_process_config       = lmv_process_config,
3870         .o_connect              = lmv_connect,
3871         .o_disconnect           = lmv_disconnect,
3872         .o_statfs               = lmv_statfs,
3873         .o_get_info             = lmv_get_info,
3874         .o_set_info_async       = lmv_set_info_async,
3875         .o_notify               = lmv_notify,
3876         .o_get_uuid             = lmv_get_uuid,
3877         .o_fid_alloc            = lmv_fid_alloc,
3878         .o_iocontrol            = lmv_iocontrol,
3879         .o_quotactl             = lmv_quotactl
3880 };
3881
3882 static const struct md_ops lmv_md_ops = {
3883         .m_get_root             = lmv_get_root,
3884         .m_null_inode           = lmv_null_inode,
3885         .m_close                = lmv_close,
3886         .m_create               = lmv_create,
3887         .m_enqueue              = lmv_enqueue,
3888         .m_getattr              = lmv_getattr,
3889         .m_getxattr             = lmv_getxattr,
3890         .m_getattr_name         = lmv_getattr_name,
3891         .m_intent_lock          = lmv_intent_lock,
3892         .m_link                 = lmv_link,
3893         .m_rename               = lmv_rename,
3894         .m_setattr              = lmv_setattr,
3895         .m_setxattr             = lmv_setxattr,
3896         .m_fsync                = lmv_fsync,
3897         .m_file_resync          = lmv_file_resync,
3898         .m_read_page            = lmv_read_page,
3899         .m_unlink               = lmv_unlink,
3900         .m_init_ea_size         = lmv_init_ea_size,
3901         .m_cancel_unused        = lmv_cancel_unused,
3902         .m_set_lock_data        = lmv_set_lock_data,
3903         .m_lock_match           = lmv_lock_match,
3904         .m_get_lustre_md        = lmv_get_lustre_md,
3905         .m_free_lustre_md       = lmv_free_lustre_md,
3906         .m_merge_attr           = lmv_merge_attr,
3907         .m_set_open_replay_data = lmv_set_open_replay_data,
3908         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3909         .m_intent_getattr_async = lmv_intent_getattr_async,
3910         .m_revalidate_lock      = lmv_revalidate_lock,
3911         .m_get_fid_from_lsm     = lmv_get_fid_from_lsm,
3912         .m_unpackmd             = lmv_unpackmd,
3913         .m_rmfid                = lmv_rmfid,
3914 };
3915
3916 static int __init lmv_init(void)
3917 {
3918         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true,
3919                                    LUSTRE_LMV_NAME, NULL);
3920 }
3921
3922 static void __exit lmv_exit(void)
3923 {
3924         class_unregister_type(LUSTRE_LMV_NAME);
3925 }
3926
3927 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3928 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
3929 MODULE_VERSION(LUSTRE_VERSION_STRING);
3930 MODULE_LICENSE("GPL");
3931
3932 module_init(lmv_init);
3933 module_exit(lmv_exit);