lustre/lmv/lmv_obd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  */
  31
  32 #define DEBUG_SUBSYSTEM S_LMV
  33
  34 #include <linux/file.h>
  35 #include <linux/module.h>
  36 #include <linux/init.h>
  37 #include <linux/user_namespace.h>
  38 #include <linux/uidgid.h>
  39 #include <linux/slab.h>
  40 #include <linux/pagemap.h>
  41 #include <linux/mm.h>
  42 #include <linux/math64.h>
  43 #include <linux/seq_file.h>
  44 #include <linux/namei.h>
  45
  46 #include <obd_support.h>
  47 #include <lustre_lib.h>
  48 #include <lustre_net.h>
  49 #include <obd_class.h>
  50 #include <lustre_lmv.h>
  51 #include <lprocfs_status.h>
  52 #include <cl_object.h>
  53 #include <lustre_fid.h>
  54 #include <uapi/linux/lustre/lustre_ioctl.h>
  55 #include <lustre_kernelcomm.h>
  56 #include "lmv_internal.h"
  57
  58 static int lmv_check_connect(struct obd_device *obd);
  59
  60 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
  61                          int activate)
  62 {
  63         if (tgt->ltd_active == activate)
  64                 return;
  65
  66         tgt->ltd_active = activate;
  67         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
  68                 (activate ? 1 : -1);
  69
  70         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
  71 }
  72
  73 /**
  74  * Error codes:
  75  *
  76  *  -EINVAL  : UUID can't be found in the LMV's target list
  77  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  78  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  79  */
  80 static int lmv_set_mdc_active(struct lmv_obd *lmv,
  81                               const struct obd_uuid *uuid,
  82                               int activate)
  83 {
  84         struct lu_tgt_desc *tgt = NULL;
  85         struct obd_device *obd;
  86         int rc = 0;
  87
  88         ENTRY;
  89
  90         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  91                         lmv, uuid->uuid, activate);
  92
  93         spin_lock(&lmv->lmv_lock);
  94         lmv_foreach_connected_tgt(lmv, tgt) {
  95                 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
  96                        tgt->ltd_index, tgt->ltd_uuid.uuid,
  97                        tgt->ltd_exp->exp_handle.h_cookie);
  98
  99                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 100                         break;
 101         }
 102
 103         if (!tgt)
 104                 GOTO(out_lmv_lock, rc = -EINVAL);
 105
 106         obd = class_exp2obd(tgt->ltd_exp);
 107         if (obd == NULL)
 108                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 109
 110         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 111                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 112                obd->obd_type->typ_name, tgt->ltd_index);
 113         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 114
 115         if (tgt->ltd_active == activate) {
 116                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 117                        activate ? "" : "in");
 118                 GOTO(out_lmv_lock, rc);
 119         }
 120
 121         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 122                activate ? "" : "in");
 123         lmv_activate_target(lmv, tgt, activate);
 124         EXIT;
 125
 126  out_lmv_lock:
 127         spin_unlock(&lmv->lmv_lock);
 128         return rc;
 129 }
 130
 131 static struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
 132 {
 133         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
 134         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
 135
 136         return tgt ? obd_get_uuid(tgt->ltd_exp) : NULL;
 137 }
 138
 139 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 140                       enum obd_notify_event ev)
 141 {
 142         struct obd_connect_data *conn_data;
 143         struct lmv_obd          *lmv = &obd->u.lmv;
 144         struct obd_uuid         *uuid;
 145         int                      rc = 0;
 146         ENTRY;
 147
 148         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 149                 CERROR("unexpected notification of %s %s!\n",
 150                        watched->obd_type->typ_name,
 151                        watched->obd_name);
 152                 RETURN(-EINVAL);
 153         }
 154
 155         uuid = &watched->u.cli.cl_target_uuid;
 156         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 157                 /*
 158                  * Set MDC as active before notifying the observer, so the
 159                  * observer can use the MDC normally.
 160                  */
 161                 rc = lmv_set_mdc_active(lmv, uuid,
 162                                         ev == OBD_NOTIFY_ACTIVE);
 163                 if (rc) {
 164                         CERROR("%sactivation of %s failed: %d\n",
 165                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 166                                uuid->uuid, rc);
 167                         RETURN(rc);
 168                 }
 169         } else if (ev == OBD_NOTIFY_OCD) {
 170                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 171                 /*
 172                  * XXX: Make sure that ocd_connect_flags from all targets are
 173                  * the same. Otherwise one of MDTs runs wrong version or
 174                  * something like this.  --umka
 175                  */
 176                 obd->obd_self_export->exp_connect_data = *conn_data;
 177         }
 178
 179         /*
 180          * Pass the notification up the chain.
 181          */
 182         if (obd->obd_observer)
 183                 rc = obd_notify(obd->obd_observer, watched, ev);
 184
 185         RETURN(rc);
 186 }
 187
 188 static int lmv_connect(const struct lu_env *env,
 189                        struct obd_export **pexp, struct obd_device *obd,
 190                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 191                        void *localdata)
 192 {
 193         struct lmv_obd *lmv = &obd->u.lmv;
 194         struct lustre_handle conn = { 0 };
 195         struct obd_export *exp;
 196         int rc;
 197         ENTRY;
 198
 199         rc = class_connect(&conn, obd, cluuid);
 200         if (rc) {
 201                 CERROR("class_connection() returned %d\n", rc);
 202                 RETURN(rc);
 203         }
 204
 205         exp = class_conn2export(&conn);
 206
 207         lmv->connected = 0;
 208         lmv->conn_data = *data;
 209         lmv->lmv_cache = localdata;
 210
 211         lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
 212                                                     &obd->obd_kset.kobj);
 213         if (!lmv->lmv_tgts_kobj) {
 214                 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
 215                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
 216         }
 217
 218         rc = lmv_check_connect(obd);
 219         if (rc != 0)
 220                 GOTO(out_sysfs, rc);
 221
 222         *pexp = exp;
 223
 224         RETURN(rc);
 225
 226 out_sysfs:
 227         if (lmv->lmv_tgts_kobj)
 228                 kobject_put(lmv->lmv_tgts_kobj);
 229
 230         class_disconnect(exp);
 231
 232         return rc;
 233 }
 234
 235 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
 236                             __u32 def_easize)
 237 {
 238         struct obd_device *obd = exp->exp_obd;
 239         struct lmv_obd *lmv = &obd->u.lmv;
 240         struct lmv_tgt_desc *tgt;
 241         int change = 0;
 242         int rc = 0;
 243
 244         ENTRY;
 245
 246         if (lmv->max_easize < easize) {
 247                 lmv->max_easize = easize;
 248                 change = 1;
 249         }
 250         if (lmv->max_def_easize < def_easize) {
 251                 lmv->max_def_easize = def_easize;
 252                 change = 1;
 253         }
 254
 255         if (change == 0)
 256                 RETURN(0);
 257
 258         if (lmv->connected == 0)
 259                 RETURN(0);
 260
 261         lmv_foreach_connected_tgt(lmv, tgt) {
 262                 if (!tgt->ltd_active)
 263                         continue;
 264
 265                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
 266                 if (rc) {
 267                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
 268                                " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
 269                         break;
 270                 }
 271         }
 272         RETURN(rc);
 273 }
 274
 275 #define MAX_STRING_SIZE 128
 276
 277 static int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 278 {
 279         struct lmv_obd *lmv = &obd->u.lmv;
 280         struct obd_device *mdc_obd;
 281         struct obd_export *mdc_exp;
 282         struct lu_fld_target target;
 283         int  rc;
 284         ENTRY;
 285
 286         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 287                                         &obd->obd_uuid);
 288         if (!mdc_obd) {
 289                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 290                 RETURN(-EINVAL);
 291         }
 292
 293         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
 294                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 295                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
 296
 297         if (!mdc_obd->obd_set_up) {
 298                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 299                 RETURN(-EINVAL);
 300         }
 301
 302         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
 303                          &lmv->conn_data, lmv->lmv_cache);
 304         if (rc) {
 305                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 306                 RETURN(rc);
 307         }
 308
 309         /*
 310          * Init fid sequence client for this mdc and add new fld target.
 311          */
 312         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
 313         if (rc)
 314                 RETURN(rc);
 315
 316         target.ft_srv = NULL;
 317         target.ft_exp = mdc_exp;
 318         target.ft_idx = tgt->ltd_index;
 319
 320         fld_client_add_target(&lmv->lmv_fld, &target);
 321
 322         rc = obd_register_observer(mdc_obd, obd);
 323         if (rc) {
 324                 obd_disconnect(mdc_exp);
 325                 CERROR("target %s register_observer error %d\n",
 326                        tgt->ltd_uuid.uuid, rc);
 327                 RETURN(rc);
 328         }
 329
 330         if (obd->obd_observer) {
 331                 /*
 332                  * Tell the observer about the new target.
 333                  */
 334                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 335                                 OBD_NOTIFY_ACTIVE);
 336                 if (rc) {
 337                         obd_disconnect(mdc_exp);
 338                         RETURN(rc);
 339                 }
 340         }
 341
 342         tgt->ltd_active = 1;
 343         tgt->ltd_exp = mdc_exp;
 344         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
 345
 346         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
 347
 348         rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
 349         if (rc) {
 350                 obd_disconnect(mdc_exp);
 351                 RETURN(rc);
 352         }
 353
 354         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 355                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 356                atomic_read(&obd->obd_refcount));
 357
 358         lmv_statfs_check_update(obd, tgt);
 359
 360         if (lmv->lmv_tgts_kobj)
 361                 /* Even if we failed to create the link, that's fine */
 362                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
 363                                        &mdc_obd->obd_kset.kobj,
 364                                        mdc_obd->obd_name);
 365         RETURN(0);
 366 }
 367
 368 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
 369 {
 370         LASSERT(tgt);
 371         ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
 372         OBD_FREE_PTR(tgt);
 373 }
 374
 375 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 376                            __u32 index, int gen)
 377 {
 378         struct obd_device *mdc_obd;
 379         struct lmv_obd *lmv = &obd->u.lmv;
 380         struct lmv_tgt_desc *tgt;
 381         struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
 382         int rc = 0;
 383
 384         ENTRY;
 385
 386         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
 387         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
 388                                         &obd->obd_uuid);
 389         if (!mdc_obd) {
 390                 CERROR("%s: Target %s not attached: rc = %d\n",
 391                        obd->obd_name, uuidp->uuid, -EINVAL);
 392                 RETURN(-EINVAL);
 393         }
 394
 395         OBD_ALLOC_PTR(tgt);
 396         if (!tgt)
 397                 RETURN(-ENOMEM);
 398
 399         mutex_init(&tgt->ltd_fid_mutex);
 400         tgt->ltd_index = index;
 401         tgt->ltd_uuid = *uuidp;
 402         tgt->ltd_active = 0;
 403
 404         mutex_lock(&ltd->ltd_mutex);
 405         rc = ltd_add_tgt(ltd, tgt);
 406         mutex_unlock(&ltd->ltd_mutex);
 407
 408         if (rc)
 409                 GOTO(out_tgt, rc);
 410
 411         if (!lmv->connected)
 412                 /* lmv_check_connect() will connect this target. */
 413                 RETURN(0);
 414
 415         rc = lmv_connect_mdc(obd, tgt);
 416         if (!rc) {
 417                 int easize = sizeof(struct lmv_stripe_md) +
 418                         lmv->lmv_mdt_count * sizeof(struct lu_fid);
 419
 420                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
 421         }
 422
 423         RETURN(rc);
 424
 425 out_tgt:
 426         OBD_FREE_PTR(tgt);
 427         return rc;
 428 }
 429
 430 static int lmv_check_connect(struct obd_device *obd)
 431 {
 432         struct lmv_obd *lmv = &obd->u.lmv;
 433         struct lmv_tgt_desc *tgt;
 434         int easize;
 435         int rc;
 436
 437         ENTRY;
 438
 439         if (lmv->connected)
 440                 RETURN(0);
 441
 442         mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
 443         if (lmv->connected)
 444                 GOTO(unlock, rc = 0);
 445
 446         if (!lmv->lmv_mdt_count) {
 447                 CERROR("%s: no targets configured: rc = -EINVAL\n",
 448                        obd->obd_name);
 449                 GOTO(unlock, rc = -EINVAL);
 450         }
 451
 452         if (!lmv_mdt0_inited(lmv)) {
 453                 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
 454                        obd->obd_name);
 455                 GOTO(unlock, rc = -EINVAL);
 456         }
 457
 458         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 459                obd->obd_uuid.uuid, obd->obd_name);
 460
 461         lmv_foreach_tgt(lmv, tgt) {
 462                 rc = lmv_connect_mdc(obd, tgt);
 463                 if (rc)
 464                         GOTO(out_disc, rc);
 465         }
 466
 467         lmv->connected = 1;
 468         easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
 469         lmv_init_ea_size(obd->obd_self_export, easize, 0);
 470         EXIT;
 471 unlock:
 472         mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
 473
 474         return rc;
 475
 476 out_disc:
 477         lmv_foreach_tgt(lmv, tgt) {
 478                 tgt->ltd_active = 0;
 479                 if (!tgt->ltd_exp)
 480                         continue;
 481
 482                 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
 483                 obd_disconnect(tgt->ltd_exp);
 484         }
 485
 486         goto unlock;
 487 }
 488
 489 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 490 {
 491         struct lmv_obd *lmv = &obd->u.lmv;
 492         struct obd_device *mdc_obd;
 493         int rc;
 494         ENTRY;
 495
 496         LASSERT(tgt != NULL);
 497         LASSERT(obd != NULL);
 498
 499         mdc_obd = class_exp2obd(tgt->ltd_exp);
 500
 501         if (mdc_obd) {
 502                 mdc_obd->obd_force = obd->obd_force;
 503                 mdc_obd->obd_fail = obd->obd_fail;
 504                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 505
 506                 if (lmv->lmv_tgts_kobj)
 507                         sysfs_remove_link(lmv->lmv_tgts_kobj,
 508                                           mdc_obd->obd_name);
 509         }
 510
 511         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
 512         if (rc)
 513                 CERROR("Can't finalize fids factory\n");
 514
 515         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 516                tgt->ltd_exp->exp_obd->obd_name,
 517                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 518
 519         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 520         rc = obd_disconnect(tgt->ltd_exp);
 521         if (rc) {
 522                 if (tgt->ltd_active) {
 523                         CERROR("Target %s disconnect error %d\n",
 524                                tgt->ltd_uuid.uuid, rc);
 525                 }
 526         }
 527
 528         lmv_activate_target(lmv, tgt, 0);
 529         tgt->ltd_exp = NULL;
 530         RETURN(0);
 531 }
 532
 533 static int lmv_disconnect(struct obd_export *exp)
 534 {
 535         struct obd_device *obd = class_exp2obd(exp);
 536         struct lmv_obd *lmv = &obd->u.lmv;
 537         struct lmv_tgt_desc *tgt;
 538         int rc;
 539
 540         ENTRY;
 541
 542         lmv_foreach_connected_tgt(lmv, tgt)
 543                 lmv_disconnect_mdc(obd, tgt);
 544
 545         if (lmv->lmv_tgts_kobj)
 546                 kobject_put(lmv->lmv_tgts_kobj);
 547
 548         if (!lmv->connected)
 549                 class_export_put(exp);
 550         rc = class_disconnect(exp);
 551         lmv->connected = 0;
 552
 553         RETURN(rc);
 554 }
 555
 556 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
 557                         void __user *uarg)
 558 {
 559         struct obd_device *obd = class_exp2obd(exp);
 560         struct lmv_obd *lmv = &obd->u.lmv;
 561         struct getinfo_fid2path *gf;
 562         struct lmv_tgt_desc *tgt;
 563         struct getinfo_fid2path *remote_gf = NULL;
 564         struct lu_fid root_fid;
 565         int remote_gf_size = 0;
 566         int rc;
 567
 568         gf = karg;
 569         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 570         if (IS_ERR(tgt))
 571                 RETURN(PTR_ERR(tgt));
 572
 573         root_fid = *gf->gf_u.gf_root_fid;
 574         LASSERT(fid_is_sane(&root_fid));
 575
 576 repeat_fid2path:
 577         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
 578         if (rc != 0 && rc != -EREMOTE)
 579                 GOTO(out_fid2path, rc);
 580
 581         /* If remote_gf != NULL, it means just building the
 582          * path on the remote MDT, copy this path segement to gf */
 583         if (remote_gf != NULL) {
 584                 struct getinfo_fid2path *ori_gf;
 585                 char *ptr;
 586                 int len;
 587
 588                 ori_gf = (struct getinfo_fid2path *)karg;
 589                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
 590                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
 591                         GOTO(out_fid2path, rc = -EOVERFLOW);
 592
 593                 ptr = ori_gf->gf_u.gf_path;
 594
 595                 len = strlen(gf->gf_u.gf_path);
 596                 /* move the current path to the right to release space
 597                  * for closer-to-root part */
 598                 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
 599                 memcpy(ptr, gf->gf_u.gf_path, len);
 600                 ptr[len] = '/';
 601         }
 602
 603         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
 604                tgt->ltd_exp->exp_obd->obd_name,
 605                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
 606                gf->gf_linkno);
 607
 608         if (rc == 0)
 609                 GOTO(out_fid2path, rc);
 610
 611         /* sigh, has to go to another MDT to do path building further */
 612         if (remote_gf == NULL) {
 613                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
 614                 OBD_ALLOC(remote_gf, remote_gf_size);
 615                 if (remote_gf == NULL)
 616                         GOTO(out_fid2path, rc = -ENOMEM);
 617                 remote_gf->gf_pathlen = PATH_MAX;
 618         }
 619
 620         if (!fid_is_sane(&gf->gf_fid)) {
 621                 CERROR("%s: invalid FID "DFID": rc = %d\n",
 622                        tgt->ltd_exp->exp_obd->obd_name,
 623                        PFID(&gf->gf_fid), -EINVAL);
 624                 GOTO(out_fid2path, rc = -EINVAL);
 625         }
 626
 627         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 628         if (IS_ERR(tgt))
 629                 GOTO(out_fid2path, rc = -EINVAL);
 630
 631         remote_gf->gf_fid = gf->gf_fid;
 632         remote_gf->gf_recno = -1;
 633         remote_gf->gf_linkno = -1;
 634         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
 635         *remote_gf->gf_u.gf_root_fid = root_fid;
 636         gf = remote_gf;
 637         goto repeat_fid2path;
 638
 639 out_fid2path:
 640         if (remote_gf != NULL)
 641                 OBD_FREE(remote_gf, remote_gf_size);
 642         RETURN(rc);
 643 }
 644
 645 static int lmv_hsm_req_count(struct lmv_obd *lmv,
 646                              const struct hsm_user_request *hur,
 647                              const struct lmv_tgt_desc *tgt_mds)
 648 {
 649         struct lmv_tgt_desc *curr_tgt;
 650         __u32 i;
 651         int nr = 0;
 652
 653         /* count how many requests must be sent to the given target */
 654         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
 655                 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
 656                 if (IS_ERR(curr_tgt))
 657                         RETURN(PTR_ERR(curr_tgt));
 658                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
 659                         nr++;
 660         }
 661         return nr;
 662 }
 663
 664 static int lmv_hsm_req_build(struct lmv_obd *lmv,
 665                               struct hsm_user_request *hur_in,
 666                               const struct lmv_tgt_desc *tgt_mds,
 667                               struct hsm_user_request *hur_out)
 668 {
 669         __u32 i, nr_out;
 670         struct lmv_tgt_desc *curr_tgt;
 671
 672         /* build the hsm_user_request for the given target */
 673         hur_out->hur_request = hur_in->hur_request;
 674         nr_out = 0;
 675         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
 676                 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
 677                 if (IS_ERR(curr_tgt))
 678                         RETURN(PTR_ERR(curr_tgt));
 679                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
 680                         hur_out->hur_user_item[nr_out] =
 681                                                 hur_in->hur_user_item[i];
 682                         nr_out++;
 683                 }
 684         }
 685         hur_out->hur_request.hr_itemcount = nr_out;
 686         memcpy(hur_data(hur_out), hur_data(hur_in),
 687                hur_in->hur_request.hr_data_len);
 688
 689         RETURN(0);
 690 }
 691
 692 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
 693                                  int len, struct lustre_kernelcomm *lk,
 694                                  void __user *uarg)
 695 {
 696         struct lmv_obd *lmv = &obd->u.lmv;
 697         struct lu_tgt_desc *tgt;
 698         int rc;
 699
 700         ENTRY;
 701
 702         /* unregister request (call from llapi_hsm_copytool_fini) */
 703         lmv_foreach_connected_tgt(lmv, tgt)
 704                 /* best effort: try to clean as much as possible
 705                  * (continue on error) */
 706                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 707
 708         /* Whatever the result, remove copytool from kuc groups.
 709          * Unreached coordinators will get EPIPE on next requests
 710          * and will unregister automatically.
 711          */
 712         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 713
 714         RETURN(rc);
 715 }
 716
 717 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
 718                                int len, struct lustre_kernelcomm *lk,
 719                                void __user *uarg)
 720 {
 721         struct lmv_obd *lmv = &obd->u.lmv;
 722         struct file *filp;
 723         bool any_set = false;
 724         struct kkuc_ct_data *kcd;
 725         size_t kcd_size;
 726         struct lu_tgt_desc *tgt;
 727         __u32 i;
 728         int err;
 729         int rc = 0;
 730
 731         ENTRY;
 732
 733         filp = fget(lk->lk_wfd);
 734         if (!filp)
 735                 RETURN(-EBADF);
 736
 737         if (lk->lk_flags & LK_FLG_DATANR)
 738                 kcd_size = offsetof(struct kkuc_ct_data,
 739                                     kcd_archives[lk->lk_data_count]);
 740         else
 741                 kcd_size = sizeof(*kcd);
 742
 743         OBD_ALLOC(kcd, kcd_size);
 744         if (kcd == NULL)
 745                 GOTO(err_fput, rc = -ENOMEM);
 746
 747         kcd->kcd_nr_archives = lk->lk_data_count;
 748         if (lk->lk_flags & LK_FLG_DATANR) {
 749                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
 750                 if (lk->lk_data_count > 0)
 751                         memcpy(kcd->kcd_archives, lk->lk_data,
 752                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
 753         } else {
 754                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
 755         }
 756
 757         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
 758                                    lk->lk_group, kcd, kcd_size);
 759         OBD_FREE(kcd, kcd_size);
 760         if (rc)
 761                 GOTO(err_fput, rc);
 762
 763         /* All or nothing: try to register to all MDS.
 764          * In case of failure, unregister from previous MDS,
 765          * except if it because of inactive target. */
 766         lmv_foreach_connected_tgt(lmv, tgt) {
 767                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 768                 if (err) {
 769                         if (tgt->ltd_active) {
 770                                 /* permanent error */
 771                                 CERROR("%s: iocontrol MDC %s on MDT"
 772                                        " idx %d cmd %x: err = %d\n",
 773                                        lmv2obd_dev(lmv)->obd_name,
 774                                        tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
 775                                        err);
 776                                 rc = err;
 777                                 lk->lk_flags |= LK_FLG_STOP;
 778                                 i = tgt->ltd_index;
 779                                 /* unregister from previous MDS */
 780                                 lmv_foreach_connected_tgt(lmv, tgt) {
 781                                         if (tgt->ltd_index >= i)
 782                                                 break;
 783
 784                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
 785                                                       lk, uarg);
 786                                 }
 787                                 GOTO(err_kkuc_rem, rc);
 788                         }
 789                         /* else: transient error.
 790                          * kuc will register to the missing MDT
 791                          * when it is back */
 792                 } else {
 793                         any_set = true;
 794                 }
 795         }
 796
 797         if (!any_set)
 798                 /* no registration done: return error */
 799                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
 800
 801         RETURN(0);
 802
 803 err_kkuc_rem:
 804         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 805
 806 err_fput:
 807         fput(filp);
 808         return rc;
 809 }
 810
 811 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 812                          int len, void *karg, void __user *uarg)
 813 {
 814         struct obd_device *obd = class_exp2obd(exp);
 815         struct lmv_obd *lmv = &obd->u.lmv;
 816         struct lu_tgt_desc *tgt = NULL;
 817         int set = 0;
 818         __u32 count = lmv->lmv_mdt_count;
 819         int rc = 0;
 820
 821         ENTRY;
 822
 823         if (count == 0)
 824                 RETURN(-ENOTTY);
 825
 826         switch (cmd) {
 827         case IOC_OBD_STATFS: {
 828                 struct obd_ioctl_data *data = karg;
 829                 struct obd_device *mdc_obd;
 830                 struct obd_statfs stat_buf = {0};
 831                 __u32 index;
 832
 833                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 834
 835                 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
 836                         RETURN(-ENODEV);
 837
 838                 tgt = lmv_tgt(lmv, index);
 839                 if (!tgt)
 840                         RETURN(-EAGAIN);
 841
 842                 if (!tgt->ltd_active)
 843                         RETURN(-ENODATA);
 844
 845                 mdc_obd = class_exp2obd(tgt->ltd_exp);
 846                 if (!mdc_obd)
 847                         RETURN(-EINVAL);
 848
 849                 /* copy UUID */
 850                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 851                                  min((int) data->ioc_plen2,
 852                                      (int) sizeof(struct obd_uuid))))
 853                         RETURN(-EFAULT);
 854
 855                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
 856                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
 857                                 0);
 858                 if (rc)
 859                         RETURN(rc);
 860                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
 861                                  min((int) data->ioc_plen1,
 862                                      (int) sizeof(stat_buf))))
 863                         RETURN(-EFAULT);
 864                 break;
 865         }
 866         case OBD_IOC_QUOTACTL: {
 867                 struct if_quotactl *qctl = karg;
 868                 struct obd_quotactl *oqctl;
 869
 870                 if (qctl->qc_valid == QC_MDTIDX) {
 871                         tgt = lmv_tgt(lmv, qctl->qc_idx);
 872                 } else if (qctl->qc_valid == QC_UUID) {
 873                         lmv_foreach_tgt(lmv, tgt) {
 874                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 875                                                      &qctl->obd_uuid))
 876                                         continue;
 877
 878                                 if (!tgt->ltd_exp)
 879                                         RETURN(-EINVAL);
 880
 881                                 break;
 882                         }
 883                 } else {
 884                         RETURN(-EINVAL);
 885                 }
 886
 887                 if (!tgt || !tgt->ltd_exp)
 888                         RETURN(-EINVAL);
 889
 890                 OBD_ALLOC_PTR(oqctl);
 891                 if (!oqctl)
 892                         RETURN(-ENOMEM);
 893
 894                 QCTL_COPY(oqctl, qctl);
 895                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 896                 if (rc == 0) {
 897                         QCTL_COPY(qctl, oqctl);
 898                         qctl->qc_valid = QC_MDTIDX;
 899                         qctl->obd_uuid = tgt->ltd_uuid;
 900                 }
 901                 OBD_FREE_PTR(oqctl);
 902                 break;
 903         }
 904         case LL_IOC_GET_CONNECT_FLAGS: {
 905                 tgt = lmv_tgt(lmv, 0);
 906                 rc = -ENODATA;
 907                 if (tgt && tgt->ltd_exp)
 908                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 909                 break;
 910         }
 911         case LL_IOC_FID2MDTIDX: {
 912                 struct lu_fid *fid = karg;
 913                 int             mdt_index;
 914
 915                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
 916                 if (rc != 0)
 917                         RETURN(rc);
 918
 919                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
 920                  * point to user space memory for FID2MDTIDX. */
 921                 *(__u32 *)uarg = mdt_index;
 922                 break;
 923         }
 924         case OBD_IOC_FID2PATH: {
 925                 rc = lmv_fid2path(exp, len, karg, uarg);
 926                 break;
 927         }
 928         case LL_IOC_HSM_STATE_GET:
 929         case LL_IOC_HSM_STATE_SET:
 930         case LL_IOC_HSM_ACTION: {
 931                 struct md_op_data *op_data = karg;
 932
 933                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
 934                 if (IS_ERR(tgt))
 935                         RETURN(PTR_ERR(tgt));
 936
 937                 if (tgt->ltd_exp == NULL)
 938                         RETURN(-EINVAL);
 939
 940                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 941                 break;
 942         }
 943         case LL_IOC_HSM_PROGRESS: {
 944                 const struct hsm_progress_kernel *hpk = karg;
 945
 946                 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
 947                 if (IS_ERR(tgt))
 948                         RETURN(PTR_ERR(tgt));
 949                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 950                 break;
 951         }
 952         case LL_IOC_HSM_REQUEST: {
 953                 struct hsm_user_request *hur = karg;
 954                 unsigned int reqcount = hur->hur_request.hr_itemcount;
 955
 956                 if (reqcount == 0)
 957                         RETURN(0);
 958
 959                 /* if the request is about a single fid
 960                  * or if there is a single MDS, no need to split
 961                  * the request. */
 962                 if (reqcount == 1 || count == 1) {
 963                         tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
 964                         if (IS_ERR(tgt))
 965                                 RETURN(PTR_ERR(tgt));
 966                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 967                 } else {
 968                         /* split fid list to their respective MDS */
 969                         lmv_foreach_connected_tgt(lmv, tgt) {
 970                                 int nr, rc1;
 971                                 size_t reqlen;
 972                                 struct hsm_user_request *req;
 973
 974                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
 975                                 if (nr < 0)
 976                                         RETURN(nr);
 977                                 if (nr == 0) /* nothing for this MDS */
 978                                         continue;
 979
 980                                 /* build a request with fids for this MDS */
 981                                 reqlen = offsetof(typeof(*hur),
 982                                                   hur_user_item[nr])
 983                                                 + hur->hur_request.hr_data_len;
 984                                 OBD_ALLOC_LARGE(req, reqlen);
 985                                 if (req == NULL)
 986                                         RETURN(-ENOMEM);
 987                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
 988                                 if (rc1 < 0)
 989                                         GOTO(hsm_req_err, rc1);
 990                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
 991                                                     req, uarg);
 992 hsm_req_err:
 993                                 if (rc1 != 0 && rc == 0)
 994                                         rc = rc1;
 995                                 OBD_FREE_LARGE(req, reqlen);
 996                         }
 997                 }
 998                 break;
 999         }
1000         case LL_IOC_LOV_SWAP_LAYOUTS: {
1001                 struct md_op_data *op_data = karg;
1002                 struct lmv_tgt_desc *tgt1, *tgt2;
1003
1004                 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1005                 if (IS_ERR(tgt1))
1006                         RETURN(PTR_ERR(tgt1));
1007
1008                 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1009                 if (IS_ERR(tgt2))
1010                         RETURN(PTR_ERR(tgt2));
1011
1012                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1013                         RETURN(-EINVAL);
1014
1015                 /* only files on same MDT can have their layouts swapped */
1016                 if (tgt1->ltd_index != tgt2->ltd_index)
1017                         RETURN(-EPERM);
1018
1019                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1020                 break;
1021         }
1022         case LL_IOC_HSM_CT_START: {
1023                 struct lustre_kernelcomm *lk = karg;
1024                 if (lk->lk_flags & LK_FLG_STOP)
1025                         rc = lmv_hsm_ct_unregister(obd, cmd, len, lk, uarg);
1026                 else
1027                         rc = lmv_hsm_ct_register(obd, cmd, len, lk, uarg);
1028                 break;
1029         }
1030         default:
1031                 lmv_foreach_connected_tgt(lmv, tgt) {
1032                         struct obd_device *mdc_obd;
1033                         int err;
1034
1035                         /* ll_umount_begin() sets force flag but for lmv, not
1036                          * mdc. Let's pass it through */
1037                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1038                         mdc_obd->obd_force = obd->obd_force;
1039                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1040                         if (err) {
1041                                 if (tgt->ltd_active) {
1042                                         CERROR("error: iocontrol MDC %s on MDT"
1043                                                " idx %d cmd %x: err = %d\n",
1044                                                tgt->ltd_uuid.uuid,
1045                                                tgt->ltd_index, cmd, err);
1046                                         if (!rc)
1047                                                 rc = err;
1048                                 }
1049                         } else
1050                                 set = 1;
1051                 }
1052                 if (!set && !rc)
1053                         rc = -EIO;
1054         }
1055         RETURN(rc);
1056 }
1057
1058 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1059                   struct lu_fid *fid, struct md_op_data *op_data)
1060 {
1061         struct obd_device *obd = class_exp2obd(exp);
1062         struct lmv_obd *lmv = &obd->u.lmv;
1063         struct lmv_tgt_desc *tgt;
1064         int rc;
1065
1066         ENTRY;
1067
1068         LASSERT(op_data);
1069         LASSERT(fid);
1070
1071         tgt = lmv_tgt(lmv, op_data->op_mds);
1072         if (!tgt)
1073                 RETURN(-ENODEV);
1074
1075         if (!tgt->ltd_active || !tgt->ltd_exp)
1076                 RETURN(-ENODEV);
1077
1078         /*
1079          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1080          * on server that seq in new allocated fid is not yet known.
1081          */
1082         mutex_lock(&tgt->ltd_fid_mutex);
1083         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1084         mutex_unlock(&tgt->ltd_fid_mutex);
1085         if (rc > 0) {
1086                 LASSERT(fid_is_sane(fid));
1087                 rc = 0;
1088         }
1089
1090         RETURN(rc);
1091 }
1092
1093 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1094 {
1095         struct lmv_obd *lmv = &obd->u.lmv;
1096         struct lmv_desc *desc;
1097         struct lnet_process_id lnet_id;
1098         int i = 0;
1099         int rc;
1100
1101         ENTRY;
1102
1103         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1104                 CERROR("LMV setup requires a descriptor\n");
1105                 RETURN(-EINVAL);
1106         }
1107
1108         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1109         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1110                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1111                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1112                 RETURN(-EINVAL);
1113         }
1114
1115         obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1116                      desc->ld_uuid.uuid);
1117         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1118         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1119         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1120                 LMV_DESC_QOS_MAXAGE_DEFAULT;
1121         lmv->max_def_easize = 0;
1122         lmv->max_easize = 0;
1123
1124         spin_lock_init(&lmv->lmv_lock);
1125
1126         /*
1127          * initialize rr_index to lower 32bit of netid, so that client
1128          * can distribute subdirs evenly from the beginning.
1129          */
1130         while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1131                 if (lnet_id.nid != LNET_NID_LO_0) {
1132                         lmv->lmv_qos_rr_index = (u32)lnet_id.nid;
1133                         break;
1134                 }
1135         }
1136
1137         rc = lmv_tunables_init(obd);
1138         if (rc)
1139                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1140                       obd->obd_name, rc);
1141
1142         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1143                              LUSTRE_CLI_FLD_HASH_DHT);
1144         if (rc)
1145                 CERROR("Can't init FLD, err %d\n", rc);
1146
1147         rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1148         if (rc)
1149                 CWARN("%s: error initialize target table: rc = %d\n",
1150                       obd->obd_name, rc);
1151
1152         RETURN(rc);
1153 }
1154
1155 static int lmv_cleanup(struct obd_device *obd)
1156 {
1157         struct lmv_obd *lmv = &obd->u.lmv;
1158         struct lu_tgt_desc *tgt;
1159         struct lu_tgt_desc *tmp;
1160
1161         ENTRY;
1162
1163         fld_client_fini(&lmv->lmv_fld);
1164         lmv_foreach_tgt_safe(lmv, tgt, tmp)
1165                 lmv_del_target(lmv, tgt);
1166         lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1167
1168         RETURN(0);
1169 }
1170
1171 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1172 {
1173         struct lustre_cfg       *lcfg = buf;
1174         struct obd_uuid         obd_uuid;
1175         int                     gen;
1176         __u32                   index;
1177         int                     rc;
1178         ENTRY;
1179
1180         switch (lcfg->lcfg_command) {
1181         case LCFG_ADD_MDC:
1182                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1183                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1184                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1185                         GOTO(out, rc = -EINVAL);
1186
1187                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1188
1189                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1190                         GOTO(out, rc = -EINVAL);
1191                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1192                         GOTO(out, rc = -EINVAL);
1193                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1194                 GOTO(out, rc);
1195         default:
1196                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1197                 GOTO(out, rc = -EINVAL);
1198         }
1199 out:
1200         RETURN(rc);
1201 }
1202
1203 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1204 {
1205         int i;
1206
1207         if (flags & OBD_STATFS_FOR_MDT0)
1208                 return 0;
1209
1210         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1211                 return lmv->lmv_statfs_start;
1212
1213         /* choose initial MDT for this client */
1214         for (i = 0;; i++) {
1215                 struct lnet_process_id lnet_id;
1216                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1217                         break;
1218
1219                 if (lnet_id.nid != LNET_NID_LO_0) {
1220                         /* We dont need a full 64-bit modulus, just enough
1221                          * to distribute the requests across MDTs evenly.
1222                          */
1223                         lmv->lmv_statfs_start = (u32)lnet_id.nid %
1224                                                 lmv->lmv_mdt_count;
1225                         break;
1226                 }
1227         }
1228
1229         return lmv->lmv_statfs_start;
1230 }
1231
1232 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1233                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1234 {
1235         struct obd_device *obd = class_exp2obd(exp);
1236         struct lmv_obd *lmv = &obd->u.lmv;
1237         struct obd_statfs *temp;
1238         struct lu_tgt_desc *tgt;
1239         __u32 i;
1240         __u32 idx;
1241         int rc = 0;
1242
1243         ENTRY;
1244
1245         OBD_ALLOC(temp, sizeof(*temp));
1246         if (temp == NULL)
1247                 RETURN(-ENOMEM);
1248
1249         /* distribute statfs among MDTs */
1250         idx = lmv_select_statfs_mdt(lmv, flags);
1251
1252         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1253                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1254                 tgt = lmv_tgt(lmv, idx);
1255                 if (!tgt || !tgt->ltd_exp)
1256                         continue;
1257
1258                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age,
1259                                 flags | OBD_STATFS_NESTED);
1260                 if (rc) {
1261                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1262                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1263                         GOTO(out_free_temp, rc);
1264                 }
1265
1266                 if (temp->os_state & OS_STATFS_SUM ||
1267                     flags == OBD_STATFS_FOR_MDT0) {
1268                         /* reset to the last aggregated values
1269                          * and don't sum with non-aggrated data */
1270                         /* If the statfs is from mount, it needs to retrieve
1271                          * necessary information from MDT0. i.e. mount does
1272                          * not need the merged osfs from all of MDT. Also
1273                          * clients can be mounted as long as MDT0 is in
1274                          * service */
1275                         *osfs = *temp;
1276                         break;
1277                 }
1278
1279                 if (i == 0) {
1280                         *osfs = *temp;
1281                 } else {
1282                         osfs->os_bavail += temp->os_bavail;
1283                         osfs->os_blocks += temp->os_blocks;
1284                         osfs->os_ffree += temp->os_ffree;
1285                         osfs->os_files += temp->os_files;
1286                         osfs->os_granted += temp->os_granted;
1287                 }
1288         }
1289
1290         EXIT;
1291 out_free_temp:
1292         OBD_FREE(temp, sizeof(*temp));
1293         return rc;
1294 }
1295
1296 static int lmv_statfs_update(void *cookie, int rc)
1297 {
1298         struct obd_info *oinfo = cookie;
1299         struct obd_device *obd = oinfo->oi_obd;
1300         struct lmv_obd *lmv = &obd->u.lmv;
1301         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1302         struct obd_statfs *osfs = oinfo->oi_osfs;
1303
1304         /*
1305          * NB: don't deactivate TGT upon error, because we may not trigger async
1306          * statfs any longer, then there is no chance to activate TGT.
1307          */
1308         if (!rc) {
1309                 spin_lock(&lmv->lmv_lock);
1310                 tgt->ltd_statfs = *osfs;
1311                 tgt->ltd_statfs_age = ktime_get_seconds();
1312                 spin_unlock(&lmv->lmv_lock);
1313                 set_bit(LQ_DIRTY, &lmv->lmv_qos.lq_flags);
1314         }
1315
1316         return rc;
1317 }
1318
1319 /* update tgt statfs async if it's ld_qos_maxage old */
1320 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1321 {
1322         struct obd_info oinfo = {
1323                 .oi_obd = obd,
1324                 .oi_tgt = tgt,
1325                 .oi_cb_up = lmv_statfs_update,
1326         };
1327         int rc;
1328
1329         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1330             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1331                 return 0;
1332
1333         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1334
1335         return rc;
1336 }
1337
1338 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1339                         struct lu_fid *fid)
1340 {
1341         struct obd_device *obd = exp->exp_obd;
1342         struct lmv_obd *lmv = &obd->u.lmv;
1343         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1344         int rc;
1345
1346         ENTRY;
1347
1348         if (!tgt)
1349                 RETURN(-ENODEV);
1350
1351         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1352         RETURN(rc);
1353 }
1354
1355 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1356                         u64 obd_md_valid, const char *name, size_t buf_size,
1357                         struct ptlrpc_request **req)
1358 {
1359         struct obd_device *obd = exp->exp_obd;
1360         struct lmv_obd *lmv = &obd->u.lmv;
1361         struct lmv_tgt_desc *tgt;
1362         int rc;
1363
1364         ENTRY;
1365
1366         tgt = lmv_fid2tgt(lmv, fid);
1367         if (IS_ERR(tgt))
1368                 RETURN(PTR_ERR(tgt));
1369
1370         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1371
1372         RETURN(rc);
1373 }
1374
1375 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1376                         u64 obd_md_valid, const char *name,
1377                         const void *value, size_t value_size,
1378                         unsigned int xattr_flags, u32 suppgid,
1379                         struct ptlrpc_request **req)
1380 {
1381         struct obd_device *obd = exp->exp_obd;
1382         struct lmv_obd *lmv = &obd->u.lmv;
1383         struct lmv_tgt_desc *tgt;
1384         int rc;
1385
1386         ENTRY;
1387
1388         tgt = lmv_fid2tgt(lmv, fid);
1389         if (IS_ERR(tgt))
1390                 RETURN(PTR_ERR(tgt));
1391
1392         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1393                          value, value_size, xattr_flags, suppgid, req);
1394
1395         RETURN(rc);
1396 }
1397
1398 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1399                        struct ptlrpc_request **request)
1400 {
1401         struct obd_device *obd = exp->exp_obd;
1402         struct lmv_obd *lmv = &obd->u.lmv;
1403         struct lmv_tgt_desc *tgt;
1404         int rc;
1405
1406         ENTRY;
1407
1408         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1409         if (IS_ERR(tgt))
1410                 RETURN(PTR_ERR(tgt));
1411
1412         if (op_data->op_flags & MF_GET_MDT_IDX) {
1413                 op_data->op_mds = tgt->ltd_index;
1414                 RETURN(0);
1415         }
1416
1417         rc = md_getattr(tgt->ltd_exp, op_data, request);
1418
1419         RETURN(rc);
1420 }
1421
1422 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1423 {
1424         struct obd_device *obd = exp->exp_obd;
1425         struct lmv_obd *lmv = &obd->u.lmv;
1426         struct lu_tgt_desc *tgt;
1427
1428         ENTRY;
1429
1430         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1431
1432         /*
1433          * With DNE every object can have two locks in different namespaces:
1434          * lookup lock in space of MDT storing direntry and update/open lock in
1435          * space of MDT storing inode.
1436          */
1437         lmv_foreach_connected_tgt(lmv, tgt)
1438                 md_null_inode(tgt->ltd_exp, fid);
1439
1440         RETURN(0);
1441 }
1442
1443 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1444                      struct md_open_data *mod, struct ptlrpc_request **request)
1445 {
1446         struct obd_device *obd = exp->exp_obd;
1447         struct lmv_obd *lmv = &obd->u.lmv;
1448         struct lmv_tgt_desc *tgt;
1449         int rc;
1450
1451         ENTRY;
1452
1453         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1454         if (IS_ERR(tgt))
1455                 RETURN(PTR_ERR(tgt));
1456
1457         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1458         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1459         RETURN(rc);
1460 }
1461
1462 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 mdt,
1463                                               unsigned short dir_depth)
1464 {
1465         struct lu_tgt_desc *tgt, *cur = NULL;
1466         __u64 total_avail = 0;
1467         __u64 total_weight = 0;
1468         __u64 cur_weight = 0;
1469         int total_usable = 0;
1470         __u64 rand;
1471         int rc;
1472
1473         ENTRY;
1474
1475         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1476                 RETURN(ERR_PTR(-EAGAIN));
1477
1478         down_write(&lmv->lmv_qos.lq_rw_sem);
1479
1480         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1481                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1482
1483         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1484         if (rc)
1485                 GOTO(unlock, tgt = ERR_PTR(rc));
1486
1487         lmv_foreach_tgt(lmv, tgt) {
1488                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1489                         tgt->ltd_qos.ltq_usable = 0;
1490                         continue;
1491                 }
1492
1493                 tgt->ltd_qos.ltq_usable = 1;
1494                 lu_tgt_qos_weight_calc(tgt);
1495                 if (tgt->ltd_index == mdt)
1496                         cur = tgt;
1497                 total_avail += tgt->ltd_qos.ltq_avail;
1498                 total_weight += tgt->ltd_qos.ltq_weight;
1499                 total_usable++;
1500         }
1501
1502         /* if current MDT has above-average space, within range of the QOS
1503          * threshold, stay on the same MDT to avoid creating needless remote
1504          * MDT directories. It's more likely for low level directories.
1505          */
1506         rand = total_avail * (256 - lmv->lmv_qos.lq_threshold_rr) /
1507                (total_usable * 256 * (1 + dir_depth / 4));
1508         if (cur && cur->ltd_qos.ltq_avail >= rand) {
1509                 tgt = cur;
1510                 GOTO(unlock, tgt);
1511         }
1512
1513         rand = lu_prandom_u64_max(total_weight);
1514
1515         lmv_foreach_connected_tgt(lmv, tgt) {
1516                 if (!tgt->ltd_qos.ltq_usable)
1517                         continue;
1518
1519                 cur_weight += tgt->ltd_qos.ltq_weight;
1520                 if (cur_weight < rand)
1521                         continue;
1522
1523                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1524                 GOTO(unlock, tgt);
1525         }
1526
1527         /* no proper target found */
1528         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1529 unlock:
1530         up_write(&lmv->lmv_qos.lq_rw_sem);
1531
1532         return tgt;
1533 }
1534
1535 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv)
1536 {
1537         struct lu_tgt_desc *tgt;
1538         int i;
1539         int index;
1540
1541         ENTRY;
1542
1543         spin_lock(&lmv->lmv_lock);
1544         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1545                 index = (i + lmv->lmv_qos_rr_index) %
1546                         lmv->lmv_mdt_descs.ltd_tgts_size;
1547                 tgt = lmv_tgt(lmv, index);
1548                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1549                         continue;
1550
1551                 lmv->lmv_qos_rr_index = (tgt->ltd_index + 1) %
1552                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1553                 spin_unlock(&lmv->lmv_lock);
1554
1555                 RETURN(tgt);
1556         }
1557         spin_unlock(&lmv->lmv_lock);
1558
1559         RETURN(ERR_PTR(-ENODEV));
1560 }
1561
1562 /* locate MDT which is less full (avoid the most full MDT) */
1563 static struct lu_tgt_desc *lmv_locate_tgt_lf(struct lmv_obd *lmv)
1564 {
1565         struct lu_tgt_desc *min = NULL;
1566         struct lu_tgt_desc *tgt;
1567         __u64 avail = 0;
1568         __u64 rand;
1569
1570         ENTRY;
1571
1572         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1573                 RETURN(ERR_PTR(-EAGAIN));
1574
1575         down_write(&lmv->lmv_qos.lq_rw_sem);
1576
1577         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1578                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1579
1580         lmv_foreach_tgt(lmv, tgt) {
1581                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1582                         tgt->ltd_qos.ltq_usable = 0;
1583                         continue;
1584                 }
1585
1586                 tgt->ltd_qos.ltq_usable = 1;
1587                 lu_tgt_qos_weight_calc(tgt);
1588                 avail += tgt->ltd_qos.ltq_avail;
1589                 if (!min || min->ltd_qos.ltq_avail > tgt->ltd_qos.ltq_avail)
1590                         min = tgt;
1591         }
1592
1593         /* avoid the most full MDT */
1594         if (min)
1595                 avail -= min->ltd_qos.ltq_avail;
1596
1597         rand = lu_prandom_u64_max(avail);
1598         avail = 0;
1599         lmv_foreach_connected_tgt(lmv, tgt) {
1600                 if (!tgt->ltd_qos.ltq_usable)
1601                         continue;
1602
1603                 if (tgt == min)
1604                         continue;
1605
1606                 avail += tgt->ltd_qos.ltq_avail;
1607                 if (avail < rand)
1608                         continue;
1609
1610                 GOTO(unlock, tgt);
1611         }
1612
1613         /* no proper target found */
1614         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1615 unlock:
1616         up_write(&lmv->lmv_qos.lq_rw_sem);
1617
1618         RETURN(tgt);
1619 }
1620
1621 /* locate MDT by file name, for striped directory, the file name hash decides
1622  * which stripe its dirent is stored.
1623  */
1624 static struct lmv_tgt_desc *
1625 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1626                        const char *name, int namelen, struct lu_fid *fid,
1627                        __u32 *mds, bool new_layout)
1628 {
1629         struct lmv_tgt_desc *tgt;
1630         const struct lmv_oinfo *oinfo;
1631
1632         if (!lmv_dir_striped(lsm) || !namelen) {
1633                 tgt = lmv_fid2tgt(lmv, fid);
1634                 if (IS_ERR(tgt))
1635                         return tgt;
1636
1637                 *mds = tgt->ltd_index;
1638                 return tgt;
1639         }
1640
1641         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1642                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1643                         return ERR_PTR(-EBADF);
1644                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1645         } else {
1646                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, new_layout);
1647                 if (IS_ERR(oinfo))
1648                         return ERR_CAST(oinfo);
1649         }
1650
1651         /* check stripe FID is sane */
1652         if (!fid_is_sane(&oinfo->lmo_fid))
1653                 return ERR_PTR(-ENODEV);
1654
1655         *fid = oinfo->lmo_fid;
1656         *mds = oinfo->lmo_mds;
1657         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1658
1659         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1660
1661         return tgt ? tgt : ERR_PTR(-ENODEV);
1662 }
1663
1664 /**
1665  * Locate MDT of op_data->op_fid1
1666  *
1667  * For striped directory, it will locate the stripe by name hash, if hash_type
1668  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1669  * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1670  * indicates whether old or new layout is used to locate.
1671  *
1672  * For plain direcotry, it just locate the MDT of op_data->op_fid1.
1673  *
1674  * \param[in] lmv               LMV device
1675  * \param[in/out] op_data       client MD stack parameters, name, namelen etc,
1676  *                              op_mds and op_fid1 will be updated if op_mea1
1677  *                              indicates fid1 represents a striped directory.
1678  *
1679  * retval               pointer to the lmv_tgt_desc if succeed.
1680  *                      ERR_PTR(errno) if failed.
1681  */
1682 struct lmv_tgt_desc *
1683 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1684 {
1685         struct lmv_stripe_md *lsm = op_data->op_mea1;
1686         struct lmv_oinfo *oinfo;
1687         struct lmv_tgt_desc *tgt;
1688
1689         if (lmv_dir_foreign(lsm))
1690                 return ERR_PTR(-ENODATA);
1691
1692         /* During creating VOLATILE file, it should honor the mdt
1693          * index if the file under striped dir is being restored, see
1694          * ct_restore(). */
1695         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1696             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1697                 tgt = lmv_tgt(lmv, op_data->op_mds);
1698                 if (!tgt)
1699                         return ERR_PTR(-ENODEV);
1700
1701                 if (lmv_dir_striped(lsm)) {
1702                         int i;
1703
1704                         /* refill the right parent fid */
1705                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1706                                 oinfo = &lsm->lsm_md_oinfo[i];
1707                                 if (oinfo->lmo_mds == op_data->op_mds) {
1708                                         op_data->op_fid1 = oinfo->lmo_fid;
1709                                         break;
1710                                 }
1711                         }
1712
1713                         if (i == lsm->lsm_md_stripe_count)
1714                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1715                 }
1716         } else if (lmv_dir_bad_hash(lsm)) {
1717                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1718                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1719
1720                 op_data->op_fid1 = oinfo->lmo_fid;
1721                 op_data->op_mds = oinfo->lmo_mds;
1722                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1723                 if (!tgt)
1724                         return ERR_PTR(-ENODEV);
1725         } else {
1726                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1727                                 op_data->op_name, op_data->op_namelen,
1728                                 &op_data->op_fid1, &op_data->op_mds,
1729                                 op_data->op_new_layout);
1730         }
1731
1732         return tgt;
1733 }
1734
1735 /* Locate MDT of op_data->op_fid2 for link/rename */
1736 static struct lmv_tgt_desc *
1737 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1738 {
1739         struct lmv_tgt_desc *tgt;
1740         int rc;
1741
1742         LASSERT(op_data->op_name);
1743         if (lmv_dir_layout_changing(op_data->op_mea2)) {
1744                 struct lu_fid fid1 = op_data->op_fid1;
1745                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1746                 struct ptlrpc_request *request = NULL;
1747
1748                 /*
1749                  * avoid creating new file under old layout of migrating
1750                  * directory, check it here.
1751                  */
1752                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1753                                 op_data->op_name, op_data->op_namelen,
1754                                 &op_data->op_fid2, &op_data->op_mds, false);
1755                 if (IS_ERR(tgt))
1756                         RETURN(tgt);
1757
1758                 op_data->op_fid1 = op_data->op_fid2;
1759                 op_data->op_mea1 = op_data->op_mea2;
1760                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1761                 op_data->op_fid1 = fid1;
1762                 op_data->op_mea1 = lsm1;
1763                 if (!rc) {
1764                         ptlrpc_req_finished(request);
1765                         RETURN(ERR_PTR(-EEXIST));
1766                 }
1767
1768                 if (rc != -ENOENT)
1769                         RETURN(ERR_PTR(rc));
1770         }
1771
1772         return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1773                                 op_data->op_namelen, &op_data->op_fid2,
1774                                 &op_data->op_mds, true);
1775 }
1776
1777 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1778 {
1779         struct lu_tgt_desc *tgt;
1780         struct ptlrpc_request *request;
1781         int rc;
1782
1783         LASSERT(lmv_dir_layout_changing(op_data->op_mea1));
1784         LASSERT(!op_data->op_new_layout);
1785
1786         tgt = lmv_locate_tgt(lmv, op_data);
1787         if (IS_ERR(tgt))
1788                 return PTR_ERR(tgt);
1789
1790         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1791         if (!rc) {
1792                 ptlrpc_req_finished(request);
1793                 return -EEXIST;
1794         }
1795
1796         return rc;
1797 }
1798
1799 static inline bool lmv_op_user_qos_mkdir(const struct md_op_data *op_data)
1800 {
1801         const struct lmv_user_md *lum = op_data->op_data;
1802
1803         return (op_data->op_cli_flags & CLI_SET_MEA) && lum &&
1804                le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC &&
1805                le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
1806 }
1807
1808 static inline bool lmv_op_default_qos_mkdir(const struct md_op_data *op_data)
1809 {
1810         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1811
1812         return (op_data->op_flags & MF_QOS_MKDIR) ||
1813                (lsm && lsm->lsm_md_master_mdt_index == LMV_OFFSET_DEFAULT);
1814 }
1815
1816 /* mkdir by QoS in three cases:
1817  * 1. ROOT default LMV is space balanced.
1818  * 2. 'lfs mkdir -i -1'
1819  * 3. parent default LMV master_mdt_index is -1
1820  *
1821  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1822  * directories under striped directory.
1823  */
1824 static inline bool lmv_op_qos_mkdir(const struct md_op_data *op_data)
1825 {
1826         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1827                 return false;
1828
1829         if (lmv_dir_striped(op_data->op_mea1))
1830                 return false;
1831
1832         if (lmv_op_user_qos_mkdir(op_data))
1833                 return true;
1834
1835         if (lmv_op_default_qos_mkdir(op_data))
1836                 return true;
1837
1838         return false;
1839 }
1840
1841 /* if parent default LMV is space balanced, and
1842  * 1. max_inherit_rr is set
1843  * 2. or parent is ROOT
1844  * mkdir roundrobin. Or if parent doesn't have default LMV, while ROOT default
1845  * LMV requests roundrobin mkdir, do the same.
1846  * NB, this needs to check server is balanced, which is done by caller.
1847  */
1848 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data)
1849 {
1850         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1851
1852         if (!lmv_op_default_qos_mkdir(op_data))
1853                 return false;
1854
1855         return (op_data->op_flags & MF_RR_MKDIR) ||
1856                (lsm && lsm->lsm_md_max_inherit_rr != LMV_INHERIT_RR_NONE) ||
1857                fid_is_root(&op_data->op_fid1);
1858 }
1859
1860 /* 'lfs mkdir -i <specific_MDT>' */
1861 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1862 {
1863         const struct lmv_user_md *lum = op_data->op_data;
1864
1865         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1866                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1867                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1868                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1869                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1870 }
1871
1872 /* parent default LMV master_mdt_index is not -1. */
1873 static inline bool
1874 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1875 {
1876         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1877                op_data->op_default_mea1 &&
1878                op_data->op_default_mea1->lsm_md_master_mdt_index !=
1879                         LMV_OFFSET_DEFAULT;
1880 }
1881
1882 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1883                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1884                 gid_t gid, kernel_cap_t cap_effective, __u64 rdev,
1885                 struct ptlrpc_request **request)
1886 {
1887         struct obd_device *obd = exp->exp_obd;
1888         struct lmv_obd *lmv = &obd->u.lmv;
1889         struct lmv_tgt_desc *tgt;
1890         struct mdt_body *repbody;
1891         int rc;
1892
1893         ENTRY;
1894
1895         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1896                 RETURN(-EIO);
1897
1898         if (lmv_dir_bad_hash(op_data->op_mea1))
1899                 RETURN(-EBADF);
1900
1901         if (lmv_dir_layout_changing(op_data->op_mea1)) {
1902                 /*
1903                  * if parent is migrating, create() needs to lookup existing
1904                  * name in both old and new layout, check old layout on client.
1905                  */
1906                 rc = lmv_old_layout_lookup(lmv, op_data);
1907                 if (rc != -ENOENT)
1908                         RETURN(rc);
1909
1910                 op_data->op_new_layout = true;
1911         }
1912
1913         tgt = lmv_locate_tgt(lmv, op_data);
1914         if (IS_ERR(tgt))
1915                 RETURN(PTR_ERR(tgt));
1916
1917         if (lmv_op_user_specific_mkdir(op_data)) {
1918                 struct lmv_user_md *lum = op_data->op_data;
1919
1920                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
1921                 tgt = lmv_tgt(lmv, op_data->op_mds);
1922                 if (!tgt)
1923                         RETURN(-ENODEV);
1924         } else if (lmv_op_default_specific_mkdir(op_data)) {
1925                 op_data->op_mds =
1926                         op_data->op_default_mea1->lsm_md_master_mdt_index;
1927                 tgt = lmv_tgt(lmv, op_data->op_mds);
1928                 if (!tgt)
1929                         RETURN(-ENODEV);
1930         } else if (lmv_op_qos_mkdir(op_data)) {
1931                 struct lmv_tgt_desc *tmp = tgt;
1932
1933                 tgt = lmv_locate_tgt_qos(lmv, op_data->op_mds,
1934                                          op_data->op_dir_depth);
1935                 if (tgt == ERR_PTR(-EAGAIN)) {
1936                         if (ltd_qos_is_balanced(&lmv->lmv_mdt_descs) &&
1937                             !lmv_op_default_rr_mkdir(op_data) &&
1938                             !lmv_op_user_qos_mkdir(op_data))
1939                                 /* if it's not necessary, don't create remote
1940                                  * directory.
1941                                  */
1942                                 tgt = tmp;
1943                         else
1944                                 tgt = lmv_locate_tgt_rr(lmv);
1945                 }
1946                 if (IS_ERR(tgt))
1947                         RETURN(PTR_ERR(tgt));
1948
1949                 op_data->op_mds = tgt->ltd_index;
1950                 /*
1951                  * only update statfs after QoS mkdir, this means the cached
1952                  * statfs may be stale, and current mkdir may not follow QoS
1953                  * accurately, but it's not serious, and avoids periodic statfs
1954                  * when client doesn't mkdir by QoS.
1955                  */
1956                 lmv_statfs_check_update(obd, tgt);
1957         }
1958
1959 retry:
1960         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1961         if (rc)
1962                 RETURN(rc);
1963
1964         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
1965                 (int)op_data->op_namelen, op_data->op_name,
1966                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
1967                 op_data->op_mds);
1968
1969         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1970         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1971                        cap_effective, rdev, request);
1972         if (rc == 0) {
1973                 if (*request == NULL)
1974                         RETURN(rc);
1975                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1976         }
1977
1978         /* dir restripe needs to send to MDT where dir is located */
1979         if (rc != -EREMOTE ||
1980             !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
1981                 RETURN(rc);
1982
1983         repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
1984         if (repbody == NULL)
1985                 RETURN(-EPROTO);
1986
1987         /* Not cross-ref case, just get out of here. */
1988         if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
1989                 RETURN(rc);
1990
1991         op_data->op_fid2 = repbody->mbo_fid1;
1992         ptlrpc_req_finished(*request);
1993         *request = NULL;
1994
1995         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
1996         if (IS_ERR(tgt))
1997                 RETURN(PTR_ERR(tgt));
1998
1999         op_data->op_mds = tgt->ltd_index;
2000         goto retry;
2001 }
2002
2003 static int
2004 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
2005             const union ldlm_policy_data *policy, struct md_op_data *op_data,
2006             struct lustre_handle *lockh, __u64 extra_lock_flags)
2007 {
2008         struct obd_device *obd = exp->exp_obd;
2009         struct lmv_obd *lmv = &obd->u.lmv;
2010         struct lmv_tgt_desc *tgt;
2011         int rc;
2012
2013         ENTRY;
2014
2015         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
2016
2017         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2018         if (IS_ERR(tgt))
2019                 RETURN(PTR_ERR(tgt));
2020
2021         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
2022                PFID(&op_data->op_fid1), tgt->ltd_index);
2023
2024         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
2025                         extra_lock_flags);
2026
2027         RETURN(rc);
2028 }
2029
2030 int
2031 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
2032                  struct ptlrpc_request **preq)
2033 {
2034         struct obd_device *obd = exp->exp_obd;
2035         struct lmv_obd *lmv = &obd->u.lmv;
2036         struct lmv_tgt_desc *tgt;
2037         struct mdt_body *body;
2038         int rc;
2039
2040         ENTRY;
2041
2042 retry:
2043         if (op_data->op_namelen == 2 &&
2044             op_data->op_name[0] == '.' && op_data->op_name[1] == '.')
2045                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2046         else
2047                 tgt = lmv_locate_tgt(lmv, op_data);
2048         if (IS_ERR(tgt))
2049                 RETURN(PTR_ERR(tgt));
2050
2051         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
2052                 (int)op_data->op_namelen, op_data->op_name,
2053                 PFID(&op_data->op_fid1), tgt->ltd_index);
2054
2055         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
2056         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2057                 ptlrpc_req_finished(*preq);
2058                 *preq = NULL;
2059                 goto retry;
2060         }
2061
2062         if (rc)
2063                 RETURN(rc);
2064
2065         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
2066         LASSERT(body != NULL);
2067
2068         if (body->mbo_valid & OBD_MD_MDS) {
2069                 op_data->op_fid1 = body->mbo_fid1;
2070                 op_data->op_valid |= OBD_MD_FLCROSSREF;
2071                 op_data->op_namelen = 0;
2072                 op_data->op_name = NULL;
2073
2074                 ptlrpc_req_finished(*preq);
2075                 *preq = NULL;
2076
2077                 goto retry;
2078         }
2079
2080         RETURN(rc);
2081 }
2082
2083 #define md_op_data_fid(op_data, fl)                     \
2084         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2085          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2086          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2087          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2088          NULL)
2089
2090 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2091                             struct md_op_data *op_data, __u32 op_tgt,
2092                             enum ldlm_mode mode, int bits, int flag)
2093 {
2094         struct lu_fid *fid = md_op_data_fid(op_data, flag);
2095         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
2096         union ldlm_policy_data policy = { { 0 } };
2097         int rc = 0;
2098         ENTRY;
2099
2100         if (!fid_is_sane(fid))
2101                 RETURN(0);
2102
2103         if (tgt == NULL) {
2104                 tgt = lmv_fid2tgt(lmv, fid);
2105                 if (IS_ERR(tgt))
2106                         RETURN(PTR_ERR(tgt));
2107         }
2108
2109         if (tgt->ltd_index != op_tgt) {
2110                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2111                 policy.l_inodebits.bits = bits;
2112                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2113                                       mode, LCF_ASYNC, NULL);
2114         } else {
2115                 CDEBUG(D_INODE,
2116                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2117                        op_tgt, PFID(fid));
2118                 op_data->op_flags |= flag;
2119                 rc = 0;
2120         }
2121
2122         RETURN(rc);
2123 }
2124
2125 /*
2126  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2127  * op_data->op_fid2
2128  */
2129 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2130                     struct ptlrpc_request **request)
2131 {
2132         struct obd_device       *obd = exp->exp_obd;
2133         struct lmv_obd          *lmv = &obd->u.lmv;
2134         struct lmv_tgt_desc     *tgt;
2135         int                      rc;
2136         ENTRY;
2137
2138         LASSERT(op_data->op_namelen != 0);
2139
2140         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2141                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2142                op_data->op_name, PFID(&op_data->op_fid1));
2143
2144         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2145         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2146         op_data->op_cap = current_cap();
2147
2148         tgt = lmv_locate_tgt2(lmv, op_data);
2149         if (IS_ERR(tgt))
2150                 RETURN(PTR_ERR(tgt));
2151
2152         /*
2153          * Cancel UPDATE lock on child (fid1).
2154          */
2155         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2156         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2157                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2158         if (rc != 0)
2159                 RETURN(rc);
2160
2161         rc = md_link(tgt->ltd_exp, op_data, request);
2162
2163         RETURN(rc);
2164 }
2165
2166 /* migrate the top directory */
2167 static inline bool lmv_op_topdir_migrate(const struct md_op_data *op_data)
2168 {
2169         if (!S_ISDIR(op_data->op_mode))
2170                 return false;
2171
2172         if (lmv_dir_layout_changing(op_data->op_mea1))
2173                 return false;
2174
2175         return true;
2176 }
2177
2178 /* migrate top dir to specific MDTs */
2179 static inline bool lmv_topdir_specific_migrate(const struct md_op_data *op_data)
2180 {
2181         const struct lmv_user_md *lum = op_data->op_data;
2182
2183         if (!lmv_op_topdir_migrate(op_data))
2184                 return false;
2185
2186         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2187 }
2188
2189 /* migrate top dir in QoS mode if user issued "lfs migrate -m -1..." */
2190 static inline bool lmv_topdir_qos_migrate(const struct md_op_data *op_data)
2191 {
2192         const struct lmv_user_md *lum = op_data->op_data;
2193
2194         if (!lmv_op_topdir_migrate(op_data))
2195                 return false;
2196
2197         return le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
2198 }
2199
2200 static inline bool lmv_subdir_specific_migrate(const struct md_op_data *op_data)
2201 {
2202         const struct lmv_user_md *lum = op_data->op_data;
2203
2204         if (!S_ISDIR(op_data->op_mode))
2205                 return false;
2206
2207         if (!lmv_dir_layout_changing(op_data->op_mea1))
2208                 return false;
2209
2210         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2211 }
2212
2213 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2214                         const char *name, size_t namelen,
2215                         struct ptlrpc_request **request)
2216 {
2217         struct obd_device *obd = exp->exp_obd;
2218         struct lmv_obd *lmv = &obd->u.lmv;
2219         struct lmv_stripe_md *lsm = op_data->op_mea1;
2220         struct lmv_tgt_desc *parent_tgt;
2221         struct lmv_tgt_desc *sp_tgt;
2222         struct lmv_tgt_desc *tp_tgt = NULL;
2223         struct lmv_tgt_desc *child_tgt;
2224         struct lmv_tgt_desc *tgt;
2225         struct lu_fid target_fid = { 0 };
2226         int rc;
2227
2228         ENTRY;
2229
2230         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2231
2232         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2233                PFID(&op_data->op_fid1), (int)namelen, name);
2234
2235         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2236         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2237         op_data->op_cap = current_cap();
2238
2239         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2240         if (IS_ERR(parent_tgt))
2241                 RETURN(PTR_ERR(parent_tgt));
2242
2243         if (lmv_dir_striped(lsm)) {
2244                 const struct lmv_oinfo *oinfo;
2245
2246                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
2247                 if (IS_ERR(oinfo))
2248                         RETURN(PTR_ERR(oinfo));
2249
2250                 /* save source stripe FID in fid4 temporarily for ELC */
2251                 op_data->op_fid4 = oinfo->lmo_fid;
2252                 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2253                 if (!sp_tgt)
2254                         RETURN(-ENODEV);
2255
2256                 /*
2257                  * if parent is being migrated too, fill op_fid2 with target
2258                  * stripe fid, otherwise the target stripe is not created yet.
2259                  */
2260                 if (lmv_dir_layout_changing(lsm)) {
2261                         oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
2262                                                         true);
2263                         if (IS_ERR(oinfo))
2264                                 RETURN(PTR_ERR(oinfo));
2265
2266                         op_data->op_fid2 = oinfo->lmo_fid;
2267                         tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2268                         if (!tp_tgt)
2269                                 RETURN(-ENODEV);
2270
2271                         /* parent unchanged and update namespace only */
2272                         if (lu_fid_eq(&op_data->op_fid4, &op_data->op_fid2) &&
2273                             op_data->op_bias & MDS_MIGRATE_NSONLY)
2274                                 RETURN(-EALREADY);
2275                 }
2276         } else {
2277                 sp_tgt = parent_tgt;
2278         }
2279
2280         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2281         if (IS_ERR(child_tgt))
2282                 RETURN(PTR_ERR(child_tgt));
2283
2284         if (lmv_topdir_specific_migrate(op_data)) {
2285                 struct lmv_user_md *lum = op_data->op_data;
2286
2287                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2288         } else if (lmv_topdir_qos_migrate(op_data)) {
2289                 tgt = lmv_locate_tgt_lf(lmv);
2290                 if (tgt == ERR_PTR(-EAGAIN))
2291                         tgt = lmv_locate_tgt_rr(lmv);
2292                 if (IS_ERR(tgt))
2293                         RETURN(PTR_ERR(tgt));
2294
2295                 op_data->op_mds = tgt->ltd_index;
2296         } else if (lmv_subdir_specific_migrate(op_data)) {
2297                 struct lmv_user_md *lum = op_data->op_data;
2298                 __u32 i;
2299
2300                 LASSERT(tp_tgt);
2301                 if (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) {
2302                         /* adjust MDTs in lum, since subdir is located on where
2303                          * its parent stripe is, not the first specified MDT.
2304                          */
2305                         for (i = 0; i < le32_to_cpu(lum->lum_stripe_count);
2306                              i++) {
2307                                 if (le32_to_cpu(lum->lum_objects[i].lum_mds) ==
2308                                     tp_tgt->ltd_index)
2309                                         break;
2310                         }
2311
2312                         if (i == le32_to_cpu(lum->lum_stripe_count))
2313                                 RETURN(-ENODEV);
2314
2315                         lum->lum_objects[i].lum_mds =
2316                                 lum->lum_objects[0].lum_mds;
2317                         lum->lum_objects[0].lum_mds =
2318                                 cpu_to_le32(tp_tgt->ltd_index);
2319                 }
2320                 /* NB, the above adjusts subdir migration for command like
2321                  * "lfs migrate -m 0,1,2 ...", but for migration like
2322                  * "lfs migrate -m 0 -c 2 ...", the top dir is migrated to MDT0
2323                  * and MDT1, however its subdir may be migrated to MDT1 and MDT2
2324                  */
2325
2326                 lum->lum_stripe_offset = cpu_to_le32(tp_tgt->ltd_index);
2327                 op_data->op_mds = tp_tgt->ltd_index;
2328         } else if (tp_tgt) {
2329                 op_data->op_mds = tp_tgt->ltd_index;
2330         } else {
2331                 op_data->op_mds = sp_tgt->ltd_index;
2332         }
2333
2334         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2335         if (rc)
2336                 RETURN(rc);
2337
2338         /*
2339          * for directory, send migrate request to the MDT where the object will
2340          * be migrated to, because we can't create a striped directory remotely.
2341          *
2342          * otherwise, send to the MDT where source is located because regular
2343          * file may open lease.
2344          *
2345          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2346          * backward compatibility.
2347          */
2348         if (S_ISDIR(op_data->op_mode) &&
2349             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2350                 tgt = lmv_fid2tgt(lmv, &target_fid);
2351                 if (IS_ERR(tgt))
2352                         RETURN(PTR_ERR(tgt));
2353         } else {
2354                 tgt = child_tgt;
2355         }
2356
2357         /* cancel UPDATE lock of parent master object */
2358         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2359                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2360         if (rc)
2361                 RETURN(rc);
2362
2363         /* cancel UPDATE lock of source parent */
2364         if (sp_tgt != parent_tgt) {
2365                 /*
2366                  * migrate RPC packs master object FID, because we can only pack
2367                  * two FIDs in reint RPC, but MDS needs to know both source
2368                  * parent and target parent, and it will obtain them from master
2369                  * FID and LMV, the other FID in RPC is kept for target.
2370                  *
2371                  * since this FID is not passed to MDC, cancel it anyway.
2372                  */
2373                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2374                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2375                 if (rc)
2376                         RETURN(rc);
2377
2378                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2379         }
2380         op_data->op_fid4 = target_fid;
2381
2382         /* cancel UPDATE locks of target parent */
2383         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2384                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2385         if (rc)
2386                 RETURN(rc);
2387
2388         /* cancel LOOKUP lock of source if source is remote object */
2389         if (child_tgt != sp_tgt) {
2390                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2391                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2392                                       MF_MDC_CANCEL_FID3);
2393                 if (rc)
2394                         RETURN(rc);
2395         }
2396
2397         /* cancel ELC locks of source */
2398         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2399                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2400         if (rc)
2401                 RETURN(rc);
2402
2403         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2404
2405         RETURN(rc);
2406 }
2407
2408 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2409                       const char *old, size_t oldlen,
2410                       const char *new, size_t newlen,
2411                       struct ptlrpc_request **request)
2412 {
2413         struct obd_device *obd = exp->exp_obd;
2414         struct lmv_obd *lmv = &obd->u.lmv;
2415         struct lmv_tgt_desc *sp_tgt;
2416         struct lmv_tgt_desc *tp_tgt = NULL;
2417         struct lmv_tgt_desc *src_tgt = NULL;
2418         struct lmv_tgt_desc *tgt;
2419         struct mdt_body *body;
2420         int rc;
2421
2422         ENTRY;
2423
2424         LASSERT(oldlen != 0);
2425
2426         if (op_data->op_cli_flags & CLI_MIGRATE) {
2427                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2428                 RETURN(rc);
2429         }
2430
2431         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2432         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2433         op_data->op_cap = current_cap();
2434
2435         op_data->op_name = new;
2436         op_data->op_namelen = newlen;
2437
2438         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2439         if (IS_ERR(tp_tgt))
2440                 RETURN(PTR_ERR(tp_tgt));
2441
2442         /* Since the target child might be destroyed, and it might become
2443          * orphan, and we can only check orphan on the local MDT right now, so
2444          * we send rename request to the MDT where target child is located. If
2445          * target child does not exist, then it will send the request to the
2446          * target parent */
2447         if (fid_is_sane(&op_data->op_fid4)) {
2448                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2449                 if (IS_ERR(tgt))
2450                         RETURN(PTR_ERR(tgt));
2451         } else {
2452                 tgt = tp_tgt;
2453         }
2454
2455         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2456
2457         /* cancel UPDATE locks of target parent */
2458         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2459                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2460         if (rc != 0)
2461                 RETURN(rc);
2462
2463         if (fid_is_sane(&op_data->op_fid4)) {
2464                 /* cancel LOOKUP lock of target on target parent */
2465                 if (tgt != tp_tgt) {
2466                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2467                                               tgt->ltd_index, LCK_EX,
2468                                               MDS_INODELOCK_LOOKUP,
2469                                               MF_MDC_CANCEL_FID4);
2470                         if (rc != 0)
2471                                 RETURN(rc);
2472                 }
2473         }
2474
2475         if (fid_is_sane(&op_data->op_fid3)) {
2476                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2477                 if (IS_ERR(src_tgt))
2478                         RETURN(PTR_ERR(src_tgt));
2479
2480                 /* cancel ELC locks of source */
2481                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2482                                       LCK_EX, MDS_INODELOCK_ELC,
2483                                       MF_MDC_CANCEL_FID3);
2484                 if (rc != 0)
2485                         RETURN(rc);
2486         }
2487
2488         op_data->op_name = old;
2489         op_data->op_namelen = oldlen;
2490 retry:
2491         sp_tgt = lmv_locate_tgt(lmv, op_data);
2492         if (IS_ERR(sp_tgt))
2493                 RETURN(PTR_ERR(sp_tgt));
2494
2495         /* cancel UPDATE locks of source parent */
2496         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2497                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2498         if (rc != 0)
2499                 RETURN(rc);
2500
2501         if (fid_is_sane(&op_data->op_fid3)) {
2502                 /* cancel LOOKUP lock of source on source parent */
2503                 if (src_tgt != sp_tgt) {
2504                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2505                                               tgt->ltd_index, LCK_EX,
2506                                               MDS_INODELOCK_LOOKUP,
2507                                               MF_MDC_CANCEL_FID3);
2508                         if (rc != 0)
2509                                 RETURN(rc);
2510                 }
2511         }
2512
2513 rename:
2514         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2515                 PFID(&op_data->op_fid1), (int)oldlen, old,
2516                 PFID(&op_data->op_fid2), (int)newlen, new);
2517
2518         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2519                         request);
2520         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2521                 ptlrpc_req_finished(*request);
2522                 *request = NULL;
2523                 goto retry;
2524         }
2525
2526         if (rc && rc != -EXDEV)
2527                 RETURN(rc);
2528
2529         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2530         if (body == NULL)
2531                 RETURN(-EPROTO);
2532
2533         /* Not cross-ref case, just get out of here. */
2534         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2535                 RETURN(rc);
2536
2537         op_data->op_fid4 = body->mbo_fid1;
2538
2539         ptlrpc_req_finished(*request);
2540         *request = NULL;
2541
2542         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2543         if (IS_ERR(tgt))
2544                 RETURN(PTR_ERR(tgt));
2545
2546         if (fid_is_sane(&op_data->op_fid4)) {
2547                 /* cancel LOOKUP lock of target on target parent */
2548                 if (tgt != tp_tgt) {
2549                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2550                                               tgt->ltd_index, LCK_EX,
2551                                               MDS_INODELOCK_LOOKUP,
2552                                               MF_MDC_CANCEL_FID4);
2553                         if (rc != 0)
2554                                 RETURN(rc);
2555                 }
2556         }
2557
2558         goto rename;
2559 }
2560
2561 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2562                        void *ea, size_t ealen, struct ptlrpc_request **request)
2563 {
2564         struct obd_device *obd = exp->exp_obd;
2565         struct lmv_obd *lmv = &obd->u.lmv;
2566         struct lmv_tgt_desc *tgt;
2567         int rc = 0;
2568
2569         ENTRY;
2570
2571         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2572                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2573                op_data->op_xvalid);
2574
2575         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2576         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2577         if (IS_ERR(tgt))
2578                 RETURN(PTR_ERR(tgt));
2579
2580         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2581
2582         RETURN(rc);
2583 }
2584
2585 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2586                      struct ptlrpc_request **request)
2587 {
2588         struct obd_device *obd = exp->exp_obd;
2589         struct lmv_obd *lmv = &obd->u.lmv;
2590         struct lmv_tgt_desc *tgt;
2591         int rc;
2592
2593         ENTRY;
2594
2595         tgt = lmv_fid2tgt(lmv, fid);
2596         if (IS_ERR(tgt))
2597                 RETURN(PTR_ERR(tgt));
2598
2599         rc = md_fsync(tgt->ltd_exp, fid, request);
2600         RETURN(rc);
2601 }
2602
2603 struct stripe_dirent {
2604         struct page             *sd_page;
2605         struct lu_dirpage       *sd_dp;
2606         struct lu_dirent        *sd_ent;
2607         bool                     sd_eof;
2608 };
2609
2610 struct lmv_dir_ctxt {
2611         struct lmv_obd          *ldc_lmv;
2612         struct md_op_data       *ldc_op_data;
2613         struct md_callback      *ldc_cb_op;
2614         __u64                    ldc_hash;
2615         int                      ldc_count;
2616         struct stripe_dirent     ldc_stripes[0];
2617 };
2618
2619 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2620 {
2621         if (stripe->sd_page) {
2622                 kunmap(stripe->sd_page);
2623                 put_page(stripe->sd_page);
2624                 stripe->sd_page = NULL;
2625                 stripe->sd_ent = NULL;
2626         }
2627 }
2628
2629 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2630 {
2631         int i;
2632
2633         for (i = 0; i < ctxt->ldc_count; i++)
2634                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2635 }
2636
2637 /* if @ent is dummy, or . .., get next */
2638 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2639                                            struct lu_dirent *ent,
2640                                            int stripe_index)
2641 {
2642         for (; ent; ent = lu_dirent_next(ent)) {
2643                 /* Skip dummy entry */
2644                 if (le16_to_cpu(ent->lde_namelen) == 0)
2645                         continue;
2646
2647                 /* skip . and .. for other stripes */
2648                 if (stripe_index &&
2649                     (strncmp(ent->lde_name, ".",
2650                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2651                      strncmp(ent->lde_name, "..",
2652                              le16_to_cpu(ent->lde_namelen)) == 0))
2653                         continue;
2654
2655                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2656                         break;
2657         }
2658
2659         return ent;
2660 }
2661
2662 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2663                                             struct stripe_dirent *stripe,
2664                                             int stripe_index)
2665 {
2666         struct md_op_data *op_data = ctxt->ldc_op_data;
2667         struct lmv_oinfo *oinfo;
2668         struct lu_fid fid = op_data->op_fid1;
2669         struct inode *inode = op_data->op_data;
2670         struct lmv_tgt_desc *tgt;
2671         struct lu_dirent *ent = stripe->sd_ent;
2672         __u64 hash = ctxt->ldc_hash;
2673         int rc = 0;
2674
2675         ENTRY;
2676
2677         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2678         LASSERT(!ent);
2679
2680         do {
2681                 if (stripe->sd_page) {
2682                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2683
2684                         /* @hash should be the last dirent hash */
2685                         LASSERTF(hash <= end,
2686                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2687                                  ctxt, stripe, hash, end);
2688                         /* unload last page */
2689                         stripe_dirent_unload(stripe);
2690                         /* eof */
2691                         if (end == MDS_DIR_END_OFF) {
2692                                 stripe->sd_eof = true;
2693                                 break;
2694                         }
2695                         hash = end;
2696                 }
2697
2698                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2699                 if (!oinfo->lmo_root) {
2700                         rc = -ENOENT;
2701                         break;
2702                 }
2703
2704                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2705                 if (!tgt) {
2706                         rc = -ENODEV;
2707                         break;
2708                 }
2709
2710                 /* op_data is shared by stripes, reset after use */
2711                 op_data->op_fid1 = oinfo->lmo_fid;
2712                 op_data->op_fid2 = oinfo->lmo_fid;
2713                 op_data->op_data = oinfo->lmo_root;
2714
2715                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2716                                   &stripe->sd_page);
2717
2718                 op_data->op_fid1 = fid;
2719                 op_data->op_fid2 = fid;
2720                 op_data->op_data = inode;
2721
2722                 if (rc)
2723                         break;
2724
2725                 stripe->sd_dp = page_address(stripe->sd_page);
2726                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2727                                         stripe_index);
2728                 /* in case a page filled with ., .. and dummy, read next */
2729         } while (!ent);
2730
2731         stripe->sd_ent = ent;
2732         if (rc) {
2733                 LASSERT(!ent);
2734                 /* treat error as eof, so dir can be partially accessed */
2735                 stripe->sd_eof = true;
2736                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2737                               "directory is partially accessed!\n",
2738                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2739                               rc);
2740         }
2741
2742         RETURN(ent);
2743 }
2744
2745 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2746 {
2747         struct obd_device *obd = exp->exp_obd;
2748         struct lmv_obd *lmv = &obd->u.lmv;
2749         struct lmv_tgt_desc *tgt;
2750         int rc;
2751
2752         ENTRY;
2753
2754         rc = lmv_check_connect(obd);
2755         if (rc != 0)
2756                 RETURN(rc);
2757
2758         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2759         if (IS_ERR(tgt))
2760                 RETURN(PTR_ERR(tgt));
2761
2762         data->op_flags |= MF_MDC_CANCEL_FID1;
2763         rc = md_file_resync(tgt->ltd_exp, data);
2764         RETURN(rc);
2765 }
2766
2767 /**
2768  * Get dirent with the closest hash for striped directory
2769  *
2770  * This function will search the dir entry, whose hash value is the
2771  * closest(>=) to hash from all of sub-stripes, and it is only being called
2772  * for striped directory.
2773  *
2774  * \param[in] ctxt              dir read context
2775  *
2776  * \retval                      dirent get the entry successfully
2777  *                              NULL does not get the entry, normally it means
2778  *                              it reaches the end of the directory, while read
2779  *                              stripe dirent error is ignored to allow partial
2780  *                              access.
2781  */
2782 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2783 {
2784         struct stripe_dirent *stripe;
2785         struct lu_dirent *ent = NULL;
2786         int i;
2787         int min = -1;
2788
2789         /* TODO: optimize with k-way merge sort */
2790         for (i = 0; i < ctxt->ldc_count; i++) {
2791                 stripe = &ctxt->ldc_stripes[i];
2792                 if (stripe->sd_eof)
2793                         continue;
2794
2795                 if (!stripe->sd_ent) {
2796                         stripe_dirent_load(ctxt, stripe, i);
2797                         if (!stripe->sd_ent) {
2798                                 LASSERT(stripe->sd_eof);
2799                                 continue;
2800                         }
2801                 }
2802
2803                 if (min == -1 ||
2804                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2805                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2806                         min = i;
2807                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2808                             ctxt->ldc_hash)
2809                                 break;
2810                 }
2811         }
2812
2813         if (min != -1) {
2814                 stripe = &ctxt->ldc_stripes[min];
2815                 ent = stripe->sd_ent;
2816                 /* pop found dirent */
2817                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2818                                                    min);
2819         }
2820
2821         return ent;
2822 }
2823
2824 /**
2825  * Build dir entry page for striped directory
2826  *
2827  * This function gets one entry by @offset from a striped directory. It will
2828  * read entries from all of stripes, and choose one closest to the required
2829  * offset(&offset). A few notes
2830  * 1. skip . and .. for non-zero stripes, because there can only have one .
2831  * and .. in a directory.
2832  * 2. op_data will be shared by all of stripes, instead of allocating new
2833  * one, so need to restore before reusing.
2834  *
2835  * \param[in] exp       obd export refer to LMV
2836  * \param[in] op_data   hold those MD parameters of read_entry
2837  * \param[in] cb_op     ldlm callback being used in enqueue in mdc_read_entry
2838  * \param[in] offset    starting hash offset
2839  * \param[out] ppage    the page holding the entry. Note: because the entry
2840  *                      will be accessed in upper layer, so we need hold the
2841  *                      page until the usages of entry is finished, see
2842  *                      ll_dir_entry_next.
2843  *
2844  * retval               =0 if get entry successfully
2845  *                      <0 cannot get entry
2846  */
2847 static int lmv_striped_read_page(struct obd_export *exp,
2848                                  struct md_op_data *op_data,
2849                                  struct md_callback *cb_op,
2850                                  __u64 offset, struct page **ppage)
2851 {
2852         struct page *page = NULL;
2853         struct lu_dirpage *dp;
2854         void *start;
2855         struct lu_dirent *ent;
2856         struct lu_dirent *last_ent;
2857         int stripe_count;
2858         struct lmv_dir_ctxt *ctxt;
2859         struct lu_dirent *next = NULL;
2860         __u16 ent_size;
2861         size_t left_bytes;
2862         int rc = 0;
2863         ENTRY;
2864
2865         /* Allocate a page and read entries from all of stripes and fill
2866          * the page by hash order */
2867         page = alloc_page(GFP_KERNEL);
2868         if (!page)
2869                 RETURN(-ENOMEM);
2870
2871         /* Initialize the entry page */
2872         dp = kmap(page);
2873         memset(dp, 0, sizeof(*dp));
2874         dp->ldp_hash_start = cpu_to_le64(offset);
2875
2876         start = dp + 1;
2877         left_bytes = PAGE_SIZE - sizeof(*dp);
2878         ent = start;
2879         last_ent = ent;
2880
2881         /* initalize dir read context */
2882         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2883         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2884         if (!ctxt)
2885                 GOTO(free_page, rc = -ENOMEM);
2886         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2887         ctxt->ldc_op_data = op_data;
2888         ctxt->ldc_cb_op = cb_op;
2889         ctxt->ldc_hash = offset;
2890         ctxt->ldc_count = stripe_count;
2891
2892         while (1) {
2893                 next = lmv_dirent_next(ctxt);
2894
2895                 /* end of directory */
2896                 if (!next) {
2897                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2898                         break;
2899                 }
2900                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2901
2902                 ent_size = le16_to_cpu(next->lde_reclen);
2903
2904                 /* the last entry lde_reclen is 0, but it might not be the last
2905                  * one of this temporay dir page */
2906                 if (!ent_size)
2907                         ent_size = lu_dirent_calc_size(
2908                                         le16_to_cpu(next->lde_namelen),
2909                                         le32_to_cpu(next->lde_attrs));
2910                 /* page full */
2911                 if (ent_size > left_bytes)
2912                         break;
2913
2914                 memcpy(ent, next, ent_size);
2915
2916                 /* Replace . with master FID and Replace .. with the parent FID
2917                  * of master object */
2918                 if (strncmp(ent->lde_name, ".",
2919                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2920                     le16_to_cpu(ent->lde_namelen) == 1)
2921                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2922                 else if (strncmp(ent->lde_name, "..",
2923                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2924                            le16_to_cpu(ent->lde_namelen) == 2)
2925                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2926
2927                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2928                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2929                        le64_to_cpu(ent->lde_hash));
2930
2931                 left_bytes -= ent_size;
2932                 ent->lde_reclen = cpu_to_le16(ent_size);
2933                 last_ent = ent;
2934                 ent = (void *)ent + ent_size;
2935         };
2936
2937         last_ent->lde_reclen = 0;
2938
2939         if (ent == start)
2940                 dp->ldp_flags |= LDF_EMPTY;
2941         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2942                 dp->ldp_flags |= LDF_COLLIDE;
2943         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2944         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2945
2946         put_lmv_dir_ctxt(ctxt);
2947         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2948
2949         *ppage = page;
2950
2951         RETURN(0);
2952
2953 free_page:
2954         kunmap(page);
2955         __free_page(page);
2956
2957         return rc;
2958 }
2959
2960 static int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2961                          struct md_callback *cb_op, __u64 offset,
2962                          struct page **ppage)
2963 {
2964         struct obd_device *obd = exp->exp_obd;
2965         struct lmv_obd *lmv = &obd->u.lmv;
2966         struct lmv_tgt_desc *tgt;
2967         int rc;
2968
2969         ENTRY;
2970
2971         if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2972                 RETURN(-ENODATA);
2973
2974         if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2975                 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2976                 RETURN(rc);
2977         }
2978
2979         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2980         if (IS_ERR(tgt))
2981                 RETURN(PTR_ERR(tgt));
2982
2983         rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2984
2985         RETURN(rc);
2986 }
2987
2988 /**
2989  * Unlink a file/directory
2990  *
2991  * Unlink a file or directory under the parent dir. The unlink request
2992  * usually will be sent to the MDT where the child is located, but if
2993  * the client does not have the child FID then request will be sent to the
2994  * MDT where the parent is located.
2995  *
2996  * If the parent is a striped directory then it also needs to locate which
2997  * stripe the name of the child is located, and replace the parent FID
2998  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2999  * it will walk through all of sub-stripes until the child is being
3000  * unlinked finally.
3001  *
3002  * \param[in] exp       export refer to LMV
3003  * \param[in] op_data   different parameters transferred beween client
3004  *                      MD stacks, name, namelen, FIDs etc.
3005  *                      op_fid1 is the parent FID, op_fid2 is the child
3006  *                      FID.
3007  * \param[out] request  point to the request of unlink.
3008  *
3009  * retval               0 if succeed
3010  *                      negative errno if failed.
3011  */
3012 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
3013                       struct ptlrpc_request **request)
3014 {
3015         struct obd_device *obd = exp->exp_obd;
3016         struct lmv_obd *lmv = &obd->u.lmv;
3017         struct lmv_tgt_desc *tgt;
3018         struct lmv_tgt_desc *parent_tgt;
3019         struct mdt_body *body;
3020         int rc;
3021
3022         ENTRY;
3023
3024         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
3025         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
3026         op_data->op_cap = current_cap();
3027
3028 retry:
3029         parent_tgt = lmv_locate_tgt(lmv, op_data);
3030         if (IS_ERR(parent_tgt))
3031                 RETURN(PTR_ERR(parent_tgt));
3032
3033         if (likely(!fid_is_zero(&op_data->op_fid2))) {
3034                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3035                 if (IS_ERR(tgt))
3036                         RETURN(PTR_ERR(tgt));
3037         } else {
3038                 tgt = parent_tgt;
3039         }
3040
3041         /*
3042          * If child's fid is given, cancel unused locks for it if it is from
3043          * another export than parent.
3044          *
3045          * LOOKUP lock for child (fid3) should also be cancelled on parent
3046          * tgt_tgt in mdc_unlink().
3047          */
3048         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
3049
3050         if (parent_tgt != tgt)
3051                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
3052                                       LCK_EX, MDS_INODELOCK_LOOKUP,
3053                                       MF_MDC_CANCEL_FID3);
3054
3055         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
3056                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
3057         if (rc)
3058                 RETURN(rc);
3059
3060         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
3061                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
3062                tgt->ltd_index);
3063
3064         rc = md_unlink(tgt->ltd_exp, op_data, request);
3065         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
3066                 ptlrpc_req_finished(*request);
3067                 *request = NULL;
3068                 goto retry;
3069         }
3070
3071         if (rc != -EREMOTE)
3072                 RETURN(rc);
3073
3074         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
3075         if (body == NULL)
3076                 RETURN(-EPROTO);
3077
3078         /* Not cross-ref case, just get out of here. */
3079         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
3080                 RETURN(rc);
3081
3082         /* This is a remote object, try remote MDT. */
3083         op_data->op_fid2 = body->mbo_fid1;
3084         ptlrpc_req_finished(*request);
3085         *request = NULL;
3086
3087         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3088         if (IS_ERR(tgt))
3089                 RETURN(PTR_ERR(tgt));
3090
3091         goto retry;
3092 }
3093
3094 static int lmv_precleanup(struct obd_device *obd)
3095 {
3096         ENTRY;
3097         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
3098         fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
3099         lprocfs_obd_cleanup(obd);
3100         lprocfs_free_md_stats(obd);
3101         RETURN(0);
3102 }
3103
3104 /**
3105  * Get by key a value associated with a LMV device.
3106  *
3107  * Dispatch request to lower-layer devices as needed.
3108  *
3109  * \param[in] env               execution environment for this thread
3110  * \param[in] exp               export for the LMV device
3111  * \param[in] keylen            length of key identifier
3112  * \param[in] key               identifier of key to get value for
3113  * \param[in] vallen            size of \a val
3114  * \param[out] val              pointer to storage location for value
3115  * \param[in] lsm               optional striping metadata of object
3116  *
3117  * \retval 0            on success
3118  * \retval negative     negated errno on failure
3119  */
3120 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
3121                         __u32 keylen, void *key, __u32 *vallen, void *val)
3122 {
3123         struct obd_device *obd;
3124         struct lmv_obd *lmv;
3125         struct lu_tgt_desc *tgt;
3126         int rc = 0;
3127
3128         ENTRY;
3129
3130         obd = class_exp2obd(exp);
3131         if (obd == NULL) {
3132                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3133                        exp->exp_handle.h_cookie);
3134                 RETURN(-EINVAL);
3135         }
3136
3137         lmv = &obd->u.lmv;
3138         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
3139                 LASSERT(*vallen == sizeof(__u32));
3140                 lmv_foreach_connected_tgt(lmv, tgt) {
3141                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
3142                                           vallen, val))
3143                                 RETURN(0);
3144                 }
3145                 RETURN(-EINVAL);
3146         } else if (KEY_IS(KEY_MAX_EASIZE) ||
3147                    KEY_IS(KEY_DEFAULT_EASIZE) ||
3148                    KEY_IS(KEY_CONN_DATA)) {
3149                 /*
3150                  * Forwarding this request to first MDS, it should know LOV
3151                  * desc.
3152                  */
3153                 tgt = lmv_tgt(lmv, 0);
3154                 if (!tgt)
3155                         RETURN(-ENODEV);
3156
3157                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
3158                 if (!rc && KEY_IS(KEY_CONN_DATA))
3159                         exp->exp_connect_data = *(struct obd_connect_data *)val;
3160                 RETURN(rc);
3161         } else if (KEY_IS(KEY_TGT_COUNT)) {
3162                 *((int *)val) = lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count;
3163                 RETURN(0);
3164         }
3165
3166         CDEBUG(D_IOCTL, "Invalid key\n");
3167         RETURN(-EINVAL);
3168 }
3169
3170 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
3171                      int *__rcs, struct ptlrpc_request_set *_set)
3172 {
3173         struct obd_device *obd = class_exp2obd(exp);
3174         struct ptlrpc_request_set *set = _set;
3175         struct lmv_obd *lmv = &obd->u.lmv;
3176         int tgt_count = lmv->lmv_mdt_count;
3177         struct lu_tgt_desc *tgt;
3178         struct fid_array *fat, **fas = NULL;
3179         int i, rc, **rcs = NULL;
3180
3181         if (!set) {
3182                 set = ptlrpc_prep_set();
3183                 if (!set)
3184                         RETURN(-ENOMEM);
3185         }
3186
3187         /* split FIDs by targets */
3188         OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
3189         if (fas == NULL)
3190                 GOTO(out, rc = -ENOMEM);
3191         OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
3192         if (rcs == NULL)
3193                 GOTO(out_fas, rc = -ENOMEM);
3194
3195         for (i = 0; i < fa->fa_nr; i++) {
3196                 unsigned int idx;
3197
3198                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
3199                 if (rc) {
3200                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
3201                                PFID(&fa->fa_fids[i]), rc);
3202                         continue;
3203                 }
3204                 LASSERT(idx < tgt_count);
3205                 if (!fas[idx])
3206                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
3207                                   fa_fids[fa->fa_nr]));
3208                 if (!fas[idx])
3209                         GOTO(out, rc = -ENOMEM);
3210                 if (!rcs[idx])
3211                         OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
3212                 if (!rcs[idx])
3213                         GOTO(out, rc = -ENOMEM);
3214
3215                 fat = fas[idx];
3216                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
3217         }
3218
3219         lmv_foreach_connected_tgt(lmv, tgt) {
3220                 fat = fas[tgt->ltd_index];
3221                 if (!fat || fat->fa_nr == 0)
3222                         continue;
3223                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3224         }
3225
3226         rc = ptlrpc_set_wait(NULL, set);
3227         if (rc == 0) {
3228                 int j = 0;
3229                 for (i = 0; i < tgt_count; i++) {
3230                         fat = fas[i];
3231                         if (!fat || fat->fa_nr == 0)
3232                                 continue;
3233                         /* copy FIDs back */
3234                         memcpy(fa->fa_fids + j, fat->fa_fids,
3235                                fat->fa_nr * sizeof(struct lu_fid));
3236                         /* copy rcs back */
3237                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3238                         j += fat->fa_nr;
3239                 }
3240         }
3241         if (set != _set)
3242                 ptlrpc_set_destroy(set);
3243
3244 out:
3245         for (i = 0; i < tgt_count; i++) {
3246                 if (fas && fas[i])
3247                         OBD_FREE(fas[i], offsetof(struct fid_array,
3248                                                 fa_fids[fa->fa_nr]));
3249                 if (rcs && rcs[i])
3250                         OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3251         }
3252         if (rcs)
3253                 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3254 out_fas:
3255         if (fas)
3256                 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3257
3258         RETURN(rc);
3259 }
3260
3261 /**
3262  * Asynchronously set by key a value associated with a LMV device.
3263  *
3264  * Dispatch request to lower-layer devices as needed.
3265  *
3266  * \param[in] env       execution environment for this thread
3267  * \param[in] exp       export for the LMV device
3268  * \param[in] keylen    length of key identifier
3269  * \param[in] key       identifier of key to store value for
3270  * \param[in] vallen    size of value to store
3271  * \param[in] val       pointer to data to be stored
3272  * \param[in] set       optional list of related ptlrpc requests
3273  *
3274  * \retval 0            on success
3275  * \retval negative     negated errno on failure
3276  */
3277 static int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3278                               __u32 keylen, void *key, __u32 vallen, void *val,
3279                               struct ptlrpc_request_set *set)
3280 {
3281         struct lmv_tgt_desc *tgt;
3282         struct obd_device *obd;
3283         struct lmv_obd *lmv;
3284         int rc = 0;
3285         ENTRY;
3286
3287         obd = class_exp2obd(exp);
3288         if (obd == NULL) {
3289                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3290                        exp->exp_handle.h_cookie);
3291                 RETURN(-EINVAL);
3292         }
3293         lmv = &obd->u.lmv;
3294
3295         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3296             KEY_IS(KEY_DEFAULT_EASIZE)) {
3297                 int err = 0;
3298
3299                 lmv_foreach_connected_tgt(lmv, tgt) {
3300                         err = obd_set_info_async(env, tgt->ltd_exp,
3301                                                  keylen, key, vallen, val, set);
3302                         if (err && rc == 0)
3303                                 rc = err;
3304                 }
3305
3306                 RETURN(rc);
3307         }
3308
3309         RETURN(-EINVAL);
3310 }
3311
3312 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3313                             const struct lmv_mds_md_v1 *lmm1)
3314 {
3315         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3316         int             stripe_count;
3317         int             cplen;
3318         int             i;
3319         int             rc = 0;
3320         ENTRY;
3321
3322         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3323         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3324         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3325         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3326                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3327         else
3328                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3329         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3330         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3331         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3332         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3333                         sizeof(lsm->lsm_md_pool_name));
3334
3335         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3336                 RETURN(-E2BIG);
3337
3338         CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3339                "layout_version %d\n", lsm->lsm_md_stripe_count,
3340                lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3341                lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3342                lsm->lsm_md_layout_version);
3343
3344         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3345         for (i = 0; i < stripe_count; i++) {
3346                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3347                               &lmm1->lmv_stripe_fids[i]);
3348                 /*
3349                  * set default value -1, so lmv_locate_tgt() knows this stripe
3350                  * target is not initialized.
3351                  */
3352                 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3353                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3354                         continue;
3355
3356                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3357                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3358                 if (rc == -ENOENT)
3359                         continue;
3360
3361                 if (rc)
3362                         RETURN(rc);
3363
3364                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3365                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3366         }
3367
3368         RETURN(rc);
3369 }
3370
3371 static inline int lmv_unpack_user_md(struct obd_export *exp,
3372                                      struct lmv_stripe_md *lsm,
3373                                      const struct lmv_user_md *lmu)
3374 {
3375         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3376         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3377         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3378         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3379         lsm->lsm_md_max_inherit = lmu->lum_max_inherit;
3380         lsm->lsm_md_max_inherit_rr = lmu->lum_max_inherit_rr;
3381         lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3382
3383         return 0;
3384 }
3385
3386 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3387                         const union lmv_mds_md *lmm, size_t lmm_size)
3388 {
3389         struct lmv_stripe_md     *lsm;
3390         int                      lsm_size;
3391         int                      rc;
3392         bool                     allocated = false;
3393         ENTRY;
3394
3395         LASSERT(lsmp != NULL);
3396
3397         lsm = *lsmp;
3398         /* Free memmd */
3399         if (lsm != NULL && lmm == NULL) {
3400                 int i;
3401                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3402
3403                 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3404                         size_t lfm_size;
3405
3406                         lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3407                                                               lfm_value[0]);
3408                         OBD_FREE_LARGE(lfm, lfm_size);
3409                         RETURN(0);
3410                 }
3411
3412                 if (lmv_dir_striped(lsm)) {
3413                         for (i = 0; i < lsm->lsm_md_stripe_count; i++)
3414                                 iput(lsm->lsm_md_oinfo[i].lmo_root);
3415                         lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3416                 } else {
3417                         lsm_size = lmv_stripe_md_size(0);
3418                 }
3419                 OBD_FREE(lsm, lsm_size);
3420                 *lsmp = NULL;
3421                 RETURN(0);
3422         }
3423
3424         /* foreign lmv case */
3425         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3426                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3427
3428                 if (lfm == NULL) {
3429                         OBD_ALLOC_LARGE(lfm, lmm_size);
3430                         if (lfm == NULL)
3431                                 RETURN(-ENOMEM);
3432                         *lsmp = (struct lmv_stripe_md *)lfm;
3433                 }
3434                 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3435                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3436                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3437                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3438                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3439                        lfm->lfm_length);
3440                 RETURN(lmm_size);
3441         }
3442
3443         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3444                 RETURN(-EPERM);
3445
3446         /* Unpack memmd */
3447         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3448             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3449                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3450                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3451                        -EIO);
3452                 RETURN(-EIO);
3453         }
3454
3455         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3456                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3457         else
3458                 /**
3459                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3460                  * stripecount should be 0 then.
3461                  */
3462                 lsm_size = lmv_stripe_md_size(0);
3463
3464         if (lsm == NULL) {
3465                 OBD_ALLOC(lsm, lsm_size);
3466                 if (lsm == NULL)
3467                         RETURN(-ENOMEM);
3468                 allocated = true;
3469                 *lsmp = lsm;
3470         }
3471
3472         switch (le32_to_cpu(lmm->lmv_magic)) {
3473         case LMV_MAGIC_V1:
3474                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3475                 break;
3476         case LMV_USER_MAGIC:
3477                 rc = lmv_unpack_user_md(exp, lsm, &lmm->lmv_user_md);
3478                 break;
3479         default:
3480                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3481                        le32_to_cpu(lmm->lmv_magic));
3482                 rc = -EINVAL;
3483                 break;
3484         }
3485
3486         if (rc != 0 && allocated) {
3487                 OBD_FREE(lsm, lsm_size);
3488                 *lsmp = NULL;
3489                 lsm_size = rc;
3490         }
3491         RETURN(lsm_size);
3492 }
3493
3494 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3495 {
3496         lmv_unpackmd(NULL, &lsm, NULL, 0);
3497 }
3498 EXPORT_SYMBOL(lmv_free_memmd);
3499
3500 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3501                              union ldlm_policy_data *policy,
3502                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3503                              void *opaque)
3504 {
3505         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3506         struct lu_tgt_desc *tgt;
3507         int err;
3508         int rc = 0;
3509
3510         ENTRY;
3511
3512         LASSERT(fid != NULL);
3513
3514         lmv_foreach_connected_tgt(lmv, tgt) {
3515                 if (!tgt->ltd_active)
3516                         continue;
3517
3518                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3519                                        opaque);
3520                 if (!rc)
3521                         rc = err;
3522         }
3523         RETURN(rc);
3524 }
3525
3526 static int lmv_set_lock_data(struct obd_export *exp,
3527                              const struct lustre_handle *lockh,
3528                              void *data, __u64 *bits)
3529 {
3530         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3531         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3532         int rc;
3533
3534         ENTRY;
3535
3536         if (tgt == NULL || tgt->ltd_exp == NULL)
3537                 RETURN(-EINVAL);
3538         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3539         RETURN(rc);
3540 }
3541
3542 static enum ldlm_mode
3543 lmv_lock_match(struct obd_export *exp, __u64 flags,
3544                const struct lu_fid *fid, enum ldlm_type type,
3545                union ldlm_policy_data *policy,
3546                enum ldlm_mode mode, struct lustre_handle *lockh)
3547 {
3548         struct obd_device *obd = exp->exp_obd;
3549         struct lmv_obd *lmv = &obd->u.lmv;
3550         enum ldlm_mode rc;
3551         struct lu_tgt_desc *tgt;
3552         int i;
3553         int index;
3554
3555         ENTRY;
3556
3557         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
3558
3559         /*
3560          * With DNE every object can have two locks in different namespaces:
3561          * lookup lock in space of MDT storing direntry and update/open lock in
3562          * space of MDT storing inode.  Try the MDT that the FID maps to first,
3563          * since this can be easily found, and only try others if that fails.
3564          */
3565         for (i = 0, index = lmv_fid2tgt_index(lmv, fid);
3566              i < lmv->lmv_mdt_descs.ltd_tgts_size;
3567              i++, index = (index + 1) % lmv->lmv_mdt_descs.ltd_tgts_size) {
3568                 if (index < 0) {
3569                         CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
3570                                obd->obd_name, PFID(fid), index);
3571                         index = 0;
3572                 }
3573
3574                 tgt = lmv_tgt(lmv, index);
3575                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
3576                         continue;
3577
3578                 rc = md_lock_match(tgt->ltd_exp, flags, fid, type, policy, mode,
3579                                    lockh);
3580                 if (rc)
3581                         RETURN(rc);
3582         }
3583
3584         RETURN(0);
3585 }
3586
3587 static int
3588 lmv_get_lustre_md(struct obd_export *exp, struct req_capsule *pill,
3589                   struct obd_export *dt_exp, struct obd_export *md_exp,
3590                   struct lustre_md *md)
3591 {
3592         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3593         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3594
3595         if (!tgt || !tgt->ltd_exp)
3596                 return -EINVAL;
3597
3598         return md_get_lustre_md(tgt->ltd_exp, pill, dt_exp, md_exp, md);
3599 }
3600
3601 static int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
3602 {
3603         struct obd_device *obd = exp->exp_obd;
3604         struct lmv_obd *lmv = &obd->u.lmv;
3605         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3606
3607         ENTRY;
3608
3609         if (md->default_lmv) {
3610                 lmv_free_memmd(md->default_lmv);
3611                 md->default_lmv = NULL;
3612         }
3613         if (md->lmv != NULL) {
3614                 lmv_free_memmd(md->lmv);
3615                 md->lmv = NULL;
3616         }
3617         if (!tgt || !tgt->ltd_exp)
3618                 RETURN(-EINVAL);
3619         RETURN(md_free_lustre_md(tgt->ltd_exp, md));
3620 }
3621
3622 static int lmv_set_open_replay_data(struct obd_export *exp,
3623                                     struct obd_client_handle *och,
3624                                     struct lookup_intent *it)
3625 {
3626         struct obd_device *obd = exp->exp_obd;
3627         struct lmv_obd *lmv = &obd->u.lmv;
3628         struct lmv_tgt_desc *tgt;
3629
3630         ENTRY;
3631
3632         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3633         if (IS_ERR(tgt))
3634                 RETURN(PTR_ERR(tgt));
3635
3636         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3637 }
3638
3639 static int lmv_clear_open_replay_data(struct obd_export *exp,
3640                                       struct obd_client_handle *och)
3641 {
3642         struct obd_device *obd = exp->exp_obd;
3643         struct lmv_obd *lmv = &obd->u.lmv;
3644         struct lmv_tgt_desc *tgt;
3645
3646         ENTRY;
3647
3648         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3649         if (IS_ERR(tgt))
3650                 RETURN(PTR_ERR(tgt));
3651
3652         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3653 }
3654
3655 static int lmv_intent_getattr_async(struct obd_export *exp,
3656                                     struct md_enqueue_info *minfo)
3657 {
3658         struct md_op_data *op_data = &minfo->mi_data;
3659         struct obd_device *obd = exp->exp_obd;
3660         struct lmv_obd *lmv = &obd->u.lmv;
3661         struct lmv_tgt_desc *ptgt;
3662         struct lmv_tgt_desc *ctgt;
3663         int rc;
3664
3665         ENTRY;
3666
3667         if (!fid_is_sane(&op_data->op_fid2))
3668                 RETURN(-EINVAL);
3669
3670         ptgt = lmv_locate_tgt(lmv, op_data);
3671         if (IS_ERR(ptgt))
3672                 RETURN(PTR_ERR(ptgt));
3673
3674         ctgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3675         if (IS_ERR(ctgt))
3676                 RETURN(PTR_ERR(ctgt));
3677
3678         /*
3679          * remote object needs two RPCs to lookup and getattr, considering the
3680          * complexity don't support statahead for now.
3681          */
3682         if (ctgt != ptgt)
3683                 RETURN(-EREMOTE);
3684
3685         rc = md_intent_getattr_async(ptgt->ltd_exp, minfo);
3686
3687         RETURN(rc);
3688 }
3689
3690 static int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3691                                struct lu_fid *fid, __u64 *bits)
3692 {
3693         struct obd_device *obd = exp->exp_obd;
3694         struct lmv_obd *lmv = &obd->u.lmv;
3695         struct lmv_tgt_desc *tgt;
3696         int rc;
3697
3698         ENTRY;
3699
3700         tgt = lmv_fid2tgt(lmv, fid);
3701         if (IS_ERR(tgt))
3702                 RETURN(PTR_ERR(tgt));
3703
3704         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3705         RETURN(rc);
3706 }
3707
3708 static int lmv_get_fid_from_lsm(struct obd_export *exp,
3709                                 const struct lmv_stripe_md *lsm,
3710                                 const char *name, int namelen,
3711                                 struct lu_fid *fid)
3712 {
3713         const struct lmv_oinfo *oinfo;
3714
3715         LASSERT(lmv_dir_striped(lsm));
3716
3717         oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
3718         if (IS_ERR(oinfo))
3719                 return PTR_ERR(oinfo);
3720
3721         *fid = oinfo->lmo_fid;
3722
3723         RETURN(0);
3724 }
3725
3726 /**
3727  * For lmv, only need to send request to master MDT, and the master MDT will
3728  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3729  * we directly fetch data from the slave MDTs.
3730  */
3731 static int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3732                         struct obd_quotactl *oqctl)
3733 {
3734         struct obd_device *obd = class_exp2obd(exp);
3735         struct lmv_obd *lmv = &obd->u.lmv;
3736         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3737         __u64 curspace, curinodes;
3738         int rc = 0;
3739
3740         ENTRY;
3741
3742         if (!tgt || !tgt->ltd_exp || !tgt->ltd_active) {
3743                 CERROR("master lmv inactive\n");
3744                 RETURN(-EIO);
3745         }
3746
3747         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3748                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3749                 RETURN(rc);
3750         }
3751
3752         curspace = curinodes = 0;
3753         lmv_foreach_connected_tgt(lmv, tgt) {
3754                 int err;
3755
3756                 if (!tgt->ltd_active)
3757                         continue;
3758
3759                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3760                 if (err) {
3761                         CERROR("getquota on mdt %d failed. %d\n",
3762                                tgt->ltd_index, err);
3763                         if (!rc)
3764                                 rc = err;
3765                 } else {
3766                         curspace += oqctl->qc_dqblk.dqb_curspace;
3767                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3768                 }
3769         }
3770         oqctl->qc_dqblk.dqb_curspace = curspace;
3771         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3772
3773         RETURN(rc);
3774 }
3775
3776 static int lmv_merge_attr(struct obd_export *exp,
3777                           const struct lmv_stripe_md *lsm,
3778                           struct cl_attr *attr,
3779                           ldlm_blocking_callback cb_blocking)
3780 {
3781         int rc;
3782         int i;
3783
3784         if (!lmv_dir_striped(lsm))
3785                 return 0;
3786
3787         rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3788         if (rc < 0)
3789                 return rc;
3790
3791         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3792                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3793
3794                 if (!inode)
3795                         continue;
3796
3797                 CDEBUG(D_INFO,
3798                        "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
3799                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3800                        i_size_read(inode), (unsigned long long)inode->i_blocks,
3801                        inode->i_nlink, (s64)inode->i_atime.tv_sec,
3802                        (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
3803
3804                 /* for slave stripe, it needs to subtract nlink for . and .. */
3805                 if (i != 0)
3806                         attr->cat_nlink += inode->i_nlink - 2;
3807                 else
3808                         attr->cat_nlink = inode->i_nlink;
3809
3810                 attr->cat_size += i_size_read(inode);
3811                 attr->cat_blocks += inode->i_blocks;
3812
3813                 if (attr->cat_atime < inode->i_atime.tv_sec)
3814                         attr->cat_atime = inode->i_atime.tv_sec;
3815
3816                 if (attr->cat_ctime < inode->i_ctime.tv_sec)
3817                         attr->cat_ctime = inode->i_ctime.tv_sec;
3818
3819                 if (attr->cat_mtime < inode->i_mtime.tv_sec)
3820                         attr->cat_mtime = inode->i_mtime.tv_sec;
3821         }
3822         return 0;
3823 }
3824
3825 static const struct obd_ops lmv_obd_ops = {
3826         .o_owner                = THIS_MODULE,
3827         .o_setup                = lmv_setup,
3828         .o_cleanup              = lmv_cleanup,
3829         .o_precleanup           = lmv_precleanup,
3830         .o_process_config       = lmv_process_config,
3831         .o_connect              = lmv_connect,
3832         .o_disconnect           = lmv_disconnect,
3833         .o_statfs               = lmv_statfs,
3834         .o_get_info             = lmv_get_info,
3835         .o_set_info_async       = lmv_set_info_async,
3836         .o_notify               = lmv_notify,
3837         .o_get_uuid             = lmv_get_uuid,
3838         .o_fid_alloc            = lmv_fid_alloc,
3839         .o_iocontrol            = lmv_iocontrol,
3840         .o_quotactl             = lmv_quotactl
3841 };
3842
3843 static const struct md_ops lmv_md_ops = {
3844         .m_get_root             = lmv_get_root,
3845         .m_null_inode           = lmv_null_inode,
3846         .m_close                = lmv_close,
3847         .m_create               = lmv_create,
3848         .m_enqueue              = lmv_enqueue,
3849         .m_getattr              = lmv_getattr,
3850         .m_getxattr             = lmv_getxattr,
3851         .m_getattr_name         = lmv_getattr_name,
3852         .m_intent_lock          = lmv_intent_lock,
3853         .m_link                 = lmv_link,
3854         .m_rename               = lmv_rename,
3855         .m_setattr              = lmv_setattr,
3856         .m_setxattr             = lmv_setxattr,
3857         .m_fsync                = lmv_fsync,
3858         .m_file_resync          = lmv_file_resync,
3859         .m_read_page            = lmv_read_page,
3860         .m_unlink               = lmv_unlink,
3861         .m_init_ea_size         = lmv_init_ea_size,
3862         .m_cancel_unused        = lmv_cancel_unused,
3863         .m_set_lock_data        = lmv_set_lock_data,
3864         .m_lock_match           = lmv_lock_match,
3865         .m_get_lustre_md        = lmv_get_lustre_md,
3866         .m_free_lustre_md       = lmv_free_lustre_md,
3867         .m_merge_attr           = lmv_merge_attr,
3868         .m_set_open_replay_data = lmv_set_open_replay_data,
3869         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3870         .m_intent_getattr_async = lmv_intent_getattr_async,
3871         .m_revalidate_lock      = lmv_revalidate_lock,
3872         .m_get_fid_from_lsm     = lmv_get_fid_from_lsm,
3873         .m_unpackmd             = lmv_unpackmd,
3874         .m_rmfid                = lmv_rmfid,
3875 };
3876
3877 static int __init lmv_init(void)
3878 {
3879         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true,
3880                                    LUSTRE_LMV_NAME, NULL);
3881 }
3882
3883 static void __exit lmv_exit(void)
3884 {
3885         class_unregister_type(LUSTRE_LMV_NAME);
3886 }
3887
3888 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3889 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
3890 MODULE_VERSION(LUSTRE_VERSION_STRING);
3891 MODULE_LICENSE("GPL");
3892
3893 module_init(lmv_init);
3894 module_exit(lmv_exit);