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