lustre/lmv/lmv_obd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  */
  31
  32 #define DEBUG_SUBSYSTEM S_LMV
  33
  34 #include <linux/file.h>
  35 #include <linux/module.h>
  36 #include <linux/init.h>
  37 #include <linux/user_namespace.h>
  38 #include <linux/uidgid.h>
  39 #include <linux/slab.h>
  40 #include <linux/pagemap.h>
  41 #include <linux/mm.h>
  42 #include <linux/math64.h>
  43 #include <linux/seq_file.h>
  44 #include <linux/namei.h>
  45
  46 #include <obd_support.h>
  47 #include <lustre_lib.h>
  48 #include <lustre_net.h>
  49 #include <obd_class.h>
  50 #include <lustre_lmv.h>
  51 #include <lprocfs_status.h>
  52 #include <cl_object.h>
  53 #include <lustre_fid.h>
  54 #include <uapi/linux/lustre/lustre_ioctl.h>
  55 #include <lustre_kernelcomm.h>
  56 #include "lmv_internal.h"
  57
  58 static int lmv_check_connect(struct obd_device *obd);
  59
  60 void lmv_activate_target(struct lmv_obd *lmv, struct lmv_tgt_desc *tgt,
  61                          int activate)
  62 {
  63         if (tgt->ltd_active == activate)
  64                 return;
  65
  66         tgt->ltd_active = activate;
  67         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count +=
  68                 (activate ? 1 : -1);
  69
  70         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
  71 }
  72
  73 /**
  74  * Error codes:
  75  *
  76  *  -EINVAL  : UUID can't be found in the LMV's target list
  77  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  78  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  79  */
  80 static int lmv_set_mdc_active(struct lmv_obd *lmv,
  81                               const struct obd_uuid *uuid,
  82                               int activate)
  83 {
  84         struct lu_tgt_desc *tgt = NULL;
  85         struct obd_device *obd;
  86         int rc = 0;
  87
  88         ENTRY;
  89
  90         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  91                         lmv, uuid->uuid, activate);
  92
  93         spin_lock(&lmv->lmv_lock);
  94         lmv_foreach_connected_tgt(lmv, tgt) {
  95                 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n",
  96                        tgt->ltd_index, tgt->ltd_uuid.uuid,
  97                        tgt->ltd_exp->exp_handle.h_cookie);
  98
  99                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 100                         break;
 101         }
 102
 103         if (!tgt)
 104                 GOTO(out_lmv_lock, rc = -EINVAL);
 105
 106         obd = class_exp2obd(tgt->ltd_exp);
 107         if (obd == NULL)
 108                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 109
 110         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 111                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 112                obd->obd_type->typ_name, tgt->ltd_index);
 113         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 114
 115         if (tgt->ltd_active == activate) {
 116                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 117                        activate ? "" : "in");
 118                 GOTO(out_lmv_lock, rc);
 119         }
 120
 121         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 122                activate ? "" : "in");
 123         lmv_activate_target(lmv, tgt, activate);
 124         EXIT;
 125
 126  out_lmv_lock:
 127         spin_unlock(&lmv->lmv_lock);
 128         return rc;
 129 }
 130
 131 static struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
 132 {
 133         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
 134         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
 135
 136         return tgt ? obd_get_uuid(tgt->ltd_exp) : NULL;
 137 }
 138
 139 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 140                       enum obd_notify_event ev)
 141 {
 142         struct obd_connect_data *conn_data;
 143         struct lmv_obd          *lmv = &obd->u.lmv;
 144         struct obd_uuid         *uuid;
 145         int                      rc = 0;
 146         ENTRY;
 147
 148         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 149                 CERROR("unexpected notification of %s %s!\n",
 150                        watched->obd_type->typ_name,
 151                        watched->obd_name);
 152                 RETURN(-EINVAL);
 153         }
 154
 155         uuid = &watched->u.cli.cl_target_uuid;
 156         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 157                 /*
 158                  * Set MDC as active before notifying the observer, so the
 159                  * observer can use the MDC normally.
 160                  */
 161                 rc = lmv_set_mdc_active(lmv, uuid,
 162                                         ev == OBD_NOTIFY_ACTIVE);
 163                 if (rc) {
 164                         CERROR("%sactivation of %s failed: %d\n",
 165                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 166                                uuid->uuid, rc);
 167                         RETURN(rc);
 168                 }
 169         } else if (ev == OBD_NOTIFY_OCD) {
 170                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 171                 /*
 172                  * XXX: Make sure that ocd_connect_flags from all targets are
 173                  * the same. Otherwise one of MDTs runs wrong version or
 174                  * something like this.  --umka
 175                  */
 176                 obd->obd_self_export->exp_connect_data = *conn_data;
 177         }
 178
 179         /*
 180          * Pass the notification up the chain.
 181          */
 182         if (obd->obd_observer)
 183                 rc = obd_notify(obd->obd_observer, watched, ev);
 184
 185         RETURN(rc);
 186 }
 187
 188 static int lmv_connect(const struct lu_env *env,
 189                        struct obd_export **pexp, struct obd_device *obd,
 190                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 191                        void *localdata)
 192 {
 193         struct lmv_obd *lmv = &obd->u.lmv;
 194         struct lustre_handle conn = { 0 };
 195         struct obd_export *exp;
 196         int rc;
 197         ENTRY;
 198
 199         rc = class_connect(&conn, obd, cluuid);
 200         if (rc) {
 201                 CERROR("class_connection() returned %d\n", rc);
 202                 RETURN(rc);
 203         }
 204
 205         exp = class_conn2export(&conn);
 206
 207         lmv->connected = 0;
 208         lmv->conn_data = *data;
 209         lmv->lmv_cache = localdata;
 210
 211         lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
 212                                                     &obd->obd_kset.kobj);
 213         if (!lmv->lmv_tgts_kobj) {
 214                 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
 215                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
 216         }
 217
 218         rc = lmv_check_connect(obd);
 219         if (rc != 0)
 220                 GOTO(out_sysfs, rc);
 221
 222         *pexp = exp;
 223
 224         RETURN(rc);
 225
 226 out_sysfs:
 227         if (lmv->lmv_tgts_kobj)
 228                 kobject_put(lmv->lmv_tgts_kobj);
 229
 230         class_disconnect(exp);
 231
 232         return rc;
 233 }
 234
 235 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
 236                             __u32 def_easize)
 237 {
 238         struct obd_device *obd = exp->exp_obd;
 239         struct lmv_obd *lmv = &obd->u.lmv;
 240         struct lmv_tgt_desc *tgt;
 241         int change = 0;
 242         int rc = 0;
 243
 244         ENTRY;
 245
 246         if (lmv->max_easize < easize) {
 247                 lmv->max_easize = easize;
 248                 change = 1;
 249         }
 250         if (lmv->max_def_easize < def_easize) {
 251                 lmv->max_def_easize = def_easize;
 252                 change = 1;
 253         }
 254
 255         if (change == 0)
 256                 RETURN(0);
 257
 258         if (lmv->connected == 0)
 259                 RETURN(0);
 260
 261         lmv_foreach_connected_tgt(lmv, tgt) {
 262                 if (!tgt->ltd_active)
 263                         continue;
 264
 265                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
 266                 if (rc) {
 267                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
 268                                " rc = %d\n", obd->obd_name, tgt->ltd_index, rc);
 269                         break;
 270                 }
 271         }
 272         RETURN(rc);
 273 }
 274
 275 #define MAX_STRING_SIZE 128
 276
 277 static int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 278 {
 279         struct lmv_obd *lmv = &obd->u.lmv;
 280         struct obd_device *mdc_obd;
 281         struct obd_export *mdc_exp;
 282         struct lu_fld_target target;
 283         int  rc;
 284         ENTRY;
 285
 286         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 287                                         &obd->obd_uuid);
 288         if (!mdc_obd) {
 289                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 290                 RETURN(-EINVAL);
 291         }
 292
 293         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
 294                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 295                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
 296
 297         if (!mdc_obd->obd_set_up) {
 298                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 299                 RETURN(-EINVAL);
 300         }
 301
 302         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
 303                          &lmv->conn_data, lmv->lmv_cache);
 304         if (rc) {
 305                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 306                 RETURN(rc);
 307         }
 308
 309         /*
 310          * Init fid sequence client for this mdc and add new fld target.
 311          */
 312         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
 313         if (rc)
 314                 RETURN(rc);
 315
 316         target.ft_srv = NULL;
 317         target.ft_exp = mdc_exp;
 318         target.ft_idx = tgt->ltd_index;
 319
 320         fld_client_add_target(&lmv->lmv_fld, &target);
 321
 322         rc = obd_register_observer(mdc_obd, obd);
 323         if (rc) {
 324                 obd_disconnect(mdc_exp);
 325                 CERROR("target %s register_observer error %d\n",
 326                        tgt->ltd_uuid.uuid, rc);
 327                 RETURN(rc);
 328         }
 329
 330         if (obd->obd_observer) {
 331                 /*
 332                  * Tell the observer about the new target.
 333                  */
 334                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 335                                 OBD_NOTIFY_ACTIVE);
 336                 if (rc) {
 337                         obd_disconnect(mdc_exp);
 338                         RETURN(rc);
 339                 }
 340         }
 341
 342         tgt->ltd_active = 1;
 343         tgt->ltd_exp = mdc_exp;
 344         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count++;
 345
 346         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
 347
 348         rc = lu_qos_add_tgt(&lmv->lmv_qos, tgt);
 349         if (rc) {
 350                 obd_disconnect(mdc_exp);
 351                 RETURN(rc);
 352         }
 353
 354         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 355                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 356                atomic_read(&obd->obd_refcount));
 357
 358         lmv_statfs_check_update(obd, tgt);
 359
 360         if (lmv->lmv_tgts_kobj)
 361                 /* Even if we failed to create the link, that's fine */
 362                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
 363                                        &mdc_obd->obd_kset.kobj,
 364                                        mdc_obd->obd_name);
 365         RETURN(0);
 366 }
 367
 368 static void lmv_del_target(struct lmv_obd *lmv, struct lu_tgt_desc *tgt)
 369 {
 370         LASSERT(tgt);
 371         ltd_del_tgt(&lmv->lmv_mdt_descs, tgt);
 372         OBD_FREE_PTR(tgt);
 373 }
 374
 375 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 376                            __u32 index, int gen)
 377 {
 378         struct obd_device *mdc_obd;
 379         struct lmv_obd *lmv = &obd->u.lmv;
 380         struct lmv_tgt_desc *tgt;
 381         struct lu_tgt_descs *ltd = &lmv->lmv_mdt_descs;
 382         int rc = 0;
 383
 384         ENTRY;
 385
 386         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
 387         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
 388                                         &obd->obd_uuid);
 389         if (!mdc_obd) {
 390                 CERROR("%s: Target %s not attached: rc = %d\n",
 391                        obd->obd_name, uuidp->uuid, -EINVAL);
 392                 RETURN(-EINVAL);
 393         }
 394
 395         OBD_ALLOC_PTR(tgt);
 396         if (!tgt)
 397                 RETURN(-ENOMEM);
 398
 399         mutex_init(&tgt->ltd_fid_mutex);
 400         tgt->ltd_index = index;
 401         tgt->ltd_uuid = *uuidp;
 402         tgt->ltd_active = 0;
 403
 404         mutex_lock(&ltd->ltd_mutex);
 405         rc = ltd_add_tgt(ltd, tgt);
 406         mutex_unlock(&ltd->ltd_mutex);
 407
 408         if (rc)
 409                 GOTO(out_tgt, rc);
 410
 411         if (!lmv->connected)
 412                 /* lmv_check_connect() will connect this target. */
 413                 RETURN(0);
 414
 415         rc = lmv_connect_mdc(obd, tgt);
 416         if (!rc) {
 417                 int easize = sizeof(struct lmv_stripe_md) +
 418                         lmv->lmv_mdt_count * sizeof(struct lu_fid);
 419
 420                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
 421         }
 422
 423         RETURN(rc);
 424
 425 out_tgt:
 426         OBD_FREE_PTR(tgt);
 427         return rc;
 428 }
 429
 430 static int lmv_check_connect(struct obd_device *obd)
 431 {
 432         struct lmv_obd *lmv = &obd->u.lmv;
 433         struct lmv_tgt_desc *tgt;
 434         int easize;
 435         int rc;
 436
 437         ENTRY;
 438
 439         if (lmv->connected)
 440                 RETURN(0);
 441
 442         mutex_lock(&lmv->lmv_mdt_descs.ltd_mutex);
 443         if (lmv->connected)
 444                 GOTO(unlock, rc = 0);
 445
 446         if (!lmv->lmv_mdt_count) {
 447                 CERROR("%s: no targets configured: rc = -EINVAL\n",
 448                        obd->obd_name);
 449                 GOTO(unlock, rc = -EINVAL);
 450         }
 451
 452         if (!lmv_mdt0_inited(lmv)) {
 453                 CERROR("%s: no target configured for index 0: rc = -EINVAL.\n",
 454                        obd->obd_name);
 455                 GOTO(unlock, rc = -EINVAL);
 456         }
 457
 458         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 459                obd->obd_uuid.uuid, obd->obd_name);
 460
 461         lmv_foreach_tgt(lmv, tgt) {
 462                 rc = lmv_connect_mdc(obd, tgt);
 463                 if (rc)
 464                         GOTO(out_disc, rc);
 465         }
 466
 467         lmv->connected = 1;
 468         easize = lmv_mds_md_size(lmv->lmv_mdt_count, LMV_MAGIC);
 469         lmv_init_ea_size(obd->obd_self_export, easize, 0);
 470         EXIT;
 471 unlock:
 472         mutex_unlock(&lmv->lmv_mdt_descs.ltd_mutex);
 473
 474         return rc;
 475
 476 out_disc:
 477         lmv_foreach_tgt(lmv, tgt) {
 478                 tgt->ltd_active = 0;
 479                 if (!tgt->ltd_exp)
 480                         continue;
 481
 482                 --lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count;
 483                 obd_disconnect(tgt->ltd_exp);
 484         }
 485
 486         goto unlock;
 487 }
 488
 489 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 490 {
 491         struct lmv_obd *lmv = &obd->u.lmv;
 492         struct obd_device *mdc_obd;
 493         int rc;
 494         ENTRY;
 495
 496         LASSERT(tgt != NULL);
 497         LASSERT(obd != NULL);
 498
 499         mdc_obd = class_exp2obd(tgt->ltd_exp);
 500
 501         if (mdc_obd) {
 502                 mdc_obd->obd_force = obd->obd_force;
 503                 mdc_obd->obd_fail = obd->obd_fail;
 504                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 505
 506                 if (lmv->lmv_tgts_kobj)
 507                         sysfs_remove_link(lmv->lmv_tgts_kobj,
 508                                           mdc_obd->obd_name);
 509         }
 510
 511         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
 512         if (rc)
 513                 CERROR("Can't finalize fids factory\n");
 514
 515         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 516                tgt->ltd_exp->exp_obd->obd_name,
 517                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 518
 519         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 520         rc = obd_disconnect(tgt->ltd_exp);
 521         if (rc) {
 522                 if (tgt->ltd_active) {
 523                         CERROR("Target %s disconnect error %d\n",
 524                                tgt->ltd_uuid.uuid, rc);
 525                 }
 526         }
 527
 528         lmv_activate_target(lmv, tgt, 0);
 529         tgt->ltd_exp = NULL;
 530         RETURN(0);
 531 }
 532
 533 static int lmv_disconnect(struct obd_export *exp)
 534 {
 535         struct obd_device *obd = class_exp2obd(exp);
 536         struct lmv_obd *lmv = &obd->u.lmv;
 537         struct lmv_tgt_desc *tgt;
 538         int rc;
 539
 540         ENTRY;
 541
 542         lmv_foreach_connected_tgt(lmv, tgt)
 543                 lmv_disconnect_mdc(obd, tgt);
 544
 545         if (lmv->lmv_tgts_kobj)
 546                 kobject_put(lmv->lmv_tgts_kobj);
 547
 548         if (!lmv->connected)
 549                 class_export_put(exp);
 550         rc = class_disconnect(exp);
 551         lmv->connected = 0;
 552
 553         RETURN(rc);
 554 }
 555
 556 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
 557                         void __user *uarg)
 558 {
 559         struct obd_device *obd = class_exp2obd(exp);
 560         struct lmv_obd *lmv = &obd->u.lmv;
 561         struct getinfo_fid2path *gf;
 562         struct lmv_tgt_desc *tgt;
 563         struct getinfo_fid2path *remote_gf = NULL;
 564         struct lu_fid root_fid;
 565         int remote_gf_size = 0;
 566         int rc;
 567
 568         gf = karg;
 569         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 570         if (IS_ERR(tgt))
 571                 RETURN(PTR_ERR(tgt));
 572
 573         root_fid = *gf->gf_u.gf_root_fid;
 574         LASSERT(fid_is_sane(&root_fid));
 575
 576 repeat_fid2path:
 577         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
 578         if (rc != 0 && rc != -EREMOTE)
 579                 GOTO(out_fid2path, rc);
 580
 581         /* If remote_gf != NULL, it means just building the
 582          * path on the remote MDT, copy this path segement to gf */
 583         if (remote_gf != NULL) {
 584                 struct getinfo_fid2path *ori_gf;
 585                 char *ptr;
 586                 int len;
 587
 588                 ori_gf = (struct getinfo_fid2path *)karg;
 589                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
 590                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
 591                         GOTO(out_fid2path, rc = -EOVERFLOW);
 592
 593                 ptr = ori_gf->gf_u.gf_path;
 594
 595                 len = strlen(gf->gf_u.gf_path);
 596                 /* move the current path to the right to release space
 597                  * for closer-to-root part */
 598                 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
 599                 memcpy(ptr, gf->gf_u.gf_path, len);
 600                 ptr[len] = '/';
 601         }
 602
 603         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
 604                tgt->ltd_exp->exp_obd->obd_name,
 605                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
 606                gf->gf_linkno);
 607
 608         if (rc == 0)
 609                 GOTO(out_fid2path, rc);
 610
 611         /* sigh, has to go to another MDT to do path building further */
 612         if (remote_gf == NULL) {
 613                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
 614                 OBD_ALLOC(remote_gf, remote_gf_size);
 615                 if (remote_gf == NULL)
 616                         GOTO(out_fid2path, rc = -ENOMEM);
 617                 remote_gf->gf_pathlen = PATH_MAX;
 618         }
 619
 620         if (!fid_is_sane(&gf->gf_fid)) {
 621                 CERROR("%s: invalid FID "DFID": rc = %d\n",
 622                        tgt->ltd_exp->exp_obd->obd_name,
 623                        PFID(&gf->gf_fid), -EINVAL);
 624                 GOTO(out_fid2path, rc = -EINVAL);
 625         }
 626
 627         tgt = lmv_fid2tgt(lmv, &gf->gf_fid);
 628         if (IS_ERR(tgt))
 629                 GOTO(out_fid2path, rc = -EINVAL);
 630
 631         remote_gf->gf_fid = gf->gf_fid;
 632         remote_gf->gf_recno = -1;
 633         remote_gf->gf_linkno = -1;
 634         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
 635         *remote_gf->gf_u.gf_root_fid = root_fid;
 636         gf = remote_gf;
 637         goto repeat_fid2path;
 638
 639 out_fid2path:
 640         if (remote_gf != NULL)
 641                 OBD_FREE(remote_gf, remote_gf_size);
 642         RETURN(rc);
 643 }
 644
 645 static int lmv_hsm_req_count(struct lmv_obd *lmv,
 646                              const struct hsm_user_request *hur,
 647                              const struct lmv_tgt_desc *tgt_mds)
 648 {
 649         struct lmv_tgt_desc *curr_tgt;
 650         __u32 i;
 651         int nr = 0;
 652
 653         /* count how many requests must be sent to the given target */
 654         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
 655                 curr_tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[i].hui_fid);
 656                 if (IS_ERR(curr_tgt))
 657                         RETURN(PTR_ERR(curr_tgt));
 658                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
 659                         nr++;
 660         }
 661         return nr;
 662 }
 663
 664 static int lmv_hsm_req_build(struct lmv_obd *lmv,
 665                               struct hsm_user_request *hur_in,
 666                               const struct lmv_tgt_desc *tgt_mds,
 667                               struct hsm_user_request *hur_out)
 668 {
 669         __u32 i, nr_out;
 670         struct lmv_tgt_desc *curr_tgt;
 671
 672         /* build the hsm_user_request for the given target */
 673         hur_out->hur_request = hur_in->hur_request;
 674         nr_out = 0;
 675         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
 676                 curr_tgt = lmv_fid2tgt(lmv, &hur_in->hur_user_item[i].hui_fid);
 677                 if (IS_ERR(curr_tgt))
 678                         RETURN(PTR_ERR(curr_tgt));
 679                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
 680                         hur_out->hur_user_item[nr_out] =
 681                                                 hur_in->hur_user_item[i];
 682                         nr_out++;
 683                 }
 684         }
 685         hur_out->hur_request.hr_itemcount = nr_out;
 686         memcpy(hur_data(hur_out), hur_data(hur_in),
 687                hur_in->hur_request.hr_data_len);
 688
 689         RETURN(0);
 690 }
 691
 692 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
 693                                  int len, struct lustre_kernelcomm *lk,
 694                                  void __user *uarg)
 695 {
 696         struct lmv_obd *lmv = &obd->u.lmv;
 697         struct lu_tgt_desc *tgt;
 698         int rc;
 699
 700         ENTRY;
 701
 702         /* unregister request (call from llapi_hsm_copytool_fini) */
 703         lmv_foreach_connected_tgt(lmv, tgt)
 704                 /* best effort: try to clean as much as possible
 705                  * (continue on error) */
 706                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 707
 708         /* Whatever the result, remove copytool from kuc groups.
 709          * Unreached coordinators will get EPIPE on next requests
 710          * and will unregister automatically.
 711          */
 712         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 713
 714         RETURN(rc);
 715 }
 716
 717 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
 718                                int len, struct lustre_kernelcomm *lk,
 719                                void __user *uarg)
 720 {
 721         struct lmv_obd *lmv = &obd->u.lmv;
 722         struct file *filp;
 723         bool any_set = false;
 724         struct kkuc_ct_data *kcd;
 725         size_t kcd_size;
 726         struct lu_tgt_desc *tgt;
 727         __u32 i;
 728         int err;
 729         int rc = 0;
 730
 731         ENTRY;
 732
 733         filp = fget(lk->lk_wfd);
 734         if (!filp)
 735                 RETURN(-EBADF);
 736
 737         if (lk->lk_flags & LK_FLG_DATANR)
 738                 kcd_size = offsetof(struct kkuc_ct_data,
 739                                     kcd_archives[lk->lk_data_count]);
 740         else
 741                 kcd_size = sizeof(*kcd);
 742
 743         OBD_ALLOC(kcd, kcd_size);
 744         if (kcd == NULL)
 745                 GOTO(err_fput, rc = -ENOMEM);
 746
 747         kcd->kcd_nr_archives = lk->lk_data_count;
 748         if (lk->lk_flags & LK_FLG_DATANR) {
 749                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
 750                 if (lk->lk_data_count > 0)
 751                         memcpy(kcd->kcd_archives, lk->lk_data,
 752                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
 753         } else {
 754                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
 755         }
 756
 757         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
 758                                    lk->lk_group, kcd, kcd_size);
 759         OBD_FREE(kcd, kcd_size);
 760         if (rc)
 761                 GOTO(err_fput, rc);
 762
 763         /* All or nothing: try to register to all MDS.
 764          * In case of failure, unregister from previous MDS,
 765          * except if it because of inactive target. */
 766         lmv_foreach_connected_tgt(lmv, tgt) {
 767                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 768                 if (err) {
 769                         if (tgt->ltd_active) {
 770                                 /* permanent error */
 771                                 CERROR("%s: iocontrol MDC %s on MDT"
 772                                        " idx %d cmd %x: err = %d\n",
 773                                        lmv2obd_dev(lmv)->obd_name,
 774                                        tgt->ltd_uuid.uuid, tgt->ltd_index, cmd,
 775                                        err);
 776                                 rc = err;
 777                                 lk->lk_flags |= LK_FLG_STOP;
 778                                 i = tgt->ltd_index;
 779                                 /* unregister from previous MDS */
 780                                 lmv_foreach_connected_tgt(lmv, tgt) {
 781                                         if (tgt->ltd_index >= i)
 782                                                 break;
 783
 784                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
 785                                                       lk, uarg);
 786                                 }
 787                                 GOTO(err_kkuc_rem, rc);
 788                         }
 789                         /* else: transient error.
 790                          * kuc will register to the missing MDT
 791                          * when it is back */
 792                 } else {
 793                         any_set = true;
 794                 }
 795         }
 796
 797         if (!any_set)
 798                 /* no registration done: return error */
 799                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
 800
 801         RETURN(0);
 802
 803 err_kkuc_rem:
 804         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 805
 806 err_fput:
 807         fput(filp);
 808         return rc;
 809 }
 810
 811 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 812                          int len, void *karg, void __user *uarg)
 813 {
 814         struct obd_device *obd = class_exp2obd(exp);
 815         struct lmv_obd *lmv = &obd->u.lmv;
 816         struct lu_tgt_desc *tgt = NULL;
 817         int set = 0;
 818         __u32 count = lmv->lmv_mdt_count;
 819         int rc = 0;
 820
 821         ENTRY;
 822
 823         if (count == 0)
 824                 RETURN(-ENOTTY);
 825
 826         switch (cmd) {
 827         case IOC_OBD_STATFS: {
 828                 struct obd_ioctl_data *data = karg;
 829                 struct obd_device *mdc_obd;
 830                 struct obd_statfs stat_buf = {0};
 831                 __u32 index;
 832
 833                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 834
 835                 if (index >= lmv->lmv_mdt_descs.ltd_tgts_size)
 836                         RETURN(-ENODEV);
 837
 838                 tgt = lmv_tgt(lmv, index);
 839                 if (!tgt)
 840                         RETURN(-EAGAIN);
 841
 842                 if (!tgt->ltd_active)
 843                         RETURN(-ENODATA);
 844
 845                 mdc_obd = class_exp2obd(tgt->ltd_exp);
 846                 if (!mdc_obd)
 847                         RETURN(-EINVAL);
 848
 849                 /* copy UUID */
 850                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 851                                  min((int) data->ioc_plen2,
 852                                      (int) sizeof(struct obd_uuid))))
 853                         RETURN(-EFAULT);
 854
 855                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
 856                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
 857                                 0);
 858                 if (rc)
 859                         RETURN(rc);
 860                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
 861                                  min((int) data->ioc_plen1,
 862                                      (int) sizeof(stat_buf))))
 863                         RETURN(-EFAULT);
 864                 break;
 865         }
 866         case OBD_IOC_QUOTACTL: {
 867                 struct if_quotactl *qctl = karg;
 868                 struct obd_quotactl *oqctl;
 869
 870                 if (qctl->qc_valid == QC_MDTIDX) {
 871                         tgt = lmv_tgt(lmv, qctl->qc_idx);
 872                 } else if (qctl->qc_valid == QC_UUID) {
 873                         lmv_foreach_tgt(lmv, tgt) {
 874                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 875                                                      &qctl->obd_uuid))
 876                                         continue;
 877
 878                                 if (!tgt->ltd_exp)
 879                                         RETURN(-EINVAL);
 880
 881                                 break;
 882                         }
 883                 } else {
 884                         RETURN(-EINVAL);
 885                 }
 886
 887                 if (!tgt || !tgt->ltd_exp)
 888                         RETURN(-EINVAL);
 889
 890                 OBD_ALLOC_PTR(oqctl);
 891                 if (!oqctl)
 892                         RETURN(-ENOMEM);
 893
 894                 QCTL_COPY(oqctl, qctl);
 895                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 896                 if (rc == 0) {
 897                         QCTL_COPY(qctl, oqctl);
 898                         qctl->qc_valid = QC_MDTIDX;
 899                         qctl->obd_uuid = tgt->ltd_uuid;
 900                 }
 901                 OBD_FREE_PTR(oqctl);
 902                 break;
 903         }
 904         case LL_IOC_GET_CONNECT_FLAGS: {
 905                 tgt = lmv_tgt(lmv, 0);
 906                 rc = -ENODATA;
 907                 if (tgt && tgt->ltd_exp)
 908                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 909                 break;
 910         }
 911         case LL_IOC_FID2MDTIDX: {
 912                 struct lu_fid *fid = karg;
 913                 int             mdt_index;
 914
 915                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
 916                 if (rc != 0)
 917                         RETURN(rc);
 918
 919                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
 920                  * point to user space memory for FID2MDTIDX. */
 921                 *(__u32 *)uarg = mdt_index;
 922                 break;
 923         }
 924         case OBD_IOC_FID2PATH: {
 925                 rc = lmv_fid2path(exp, len, karg, uarg);
 926                 break;
 927         }
 928         case LL_IOC_HSM_STATE_GET:
 929         case LL_IOC_HSM_STATE_SET:
 930         case LL_IOC_HSM_ACTION: {
 931                 struct md_op_data *op_data = karg;
 932
 933                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
 934                 if (IS_ERR(tgt))
 935                         RETURN(PTR_ERR(tgt));
 936
 937                 if (tgt->ltd_exp == NULL)
 938                         RETURN(-EINVAL);
 939
 940                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 941                 break;
 942         }
 943         case LL_IOC_HSM_PROGRESS: {
 944                 const struct hsm_progress_kernel *hpk = karg;
 945
 946                 tgt = lmv_fid2tgt(lmv, &hpk->hpk_fid);
 947                 if (IS_ERR(tgt))
 948                         RETURN(PTR_ERR(tgt));
 949                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 950                 break;
 951         }
 952         case LL_IOC_HSM_REQUEST: {
 953                 struct hsm_user_request *hur = karg;
 954                 unsigned int reqcount = hur->hur_request.hr_itemcount;
 955
 956                 if (reqcount == 0)
 957                         RETURN(0);
 958
 959                 /* if the request is about a single fid
 960                  * or if there is a single MDS, no need to split
 961                  * the request. */
 962                 if (reqcount == 1 || count == 1) {
 963                         tgt = lmv_fid2tgt(lmv, &hur->hur_user_item[0].hui_fid);
 964                         if (IS_ERR(tgt))
 965                                 RETURN(PTR_ERR(tgt));
 966                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 967                 } else {
 968                         /* split fid list to their respective MDS */
 969                         lmv_foreach_connected_tgt(lmv, tgt) {
 970                                 int nr, rc1;
 971                                 size_t reqlen;
 972                                 struct hsm_user_request *req;
 973
 974                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
 975                                 if (nr < 0)
 976                                         RETURN(nr);
 977                                 if (nr == 0) /* nothing for this MDS */
 978                                         continue;
 979
 980                                 /* build a request with fids for this MDS */
 981                                 reqlen = offsetof(typeof(*hur),
 982                                                   hur_user_item[nr])
 983                                                 + hur->hur_request.hr_data_len;
 984                                 OBD_ALLOC_LARGE(req, reqlen);
 985                                 if (req == NULL)
 986                                         RETURN(-ENOMEM);
 987                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
 988                                 if (rc1 < 0)
 989                                         GOTO(hsm_req_err, rc1);
 990                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
 991                                                     req, uarg);
 992 hsm_req_err:
 993                                 if (rc1 != 0 && rc == 0)
 994                                         rc = rc1;
 995                                 OBD_FREE_LARGE(req, reqlen);
 996                         }
 997                 }
 998                 break;
 999         }
1000         case LL_IOC_LOV_SWAP_LAYOUTS: {
1001                 struct md_op_data *op_data = karg;
1002                 struct lmv_tgt_desc *tgt1, *tgt2;
1003
1004                 tgt1 = lmv_fid2tgt(lmv, &op_data->op_fid1);
1005                 if (IS_ERR(tgt1))
1006                         RETURN(PTR_ERR(tgt1));
1007
1008                 tgt2 = lmv_fid2tgt(lmv, &op_data->op_fid2);
1009                 if (IS_ERR(tgt2))
1010                         RETURN(PTR_ERR(tgt2));
1011
1012                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1013                         RETURN(-EINVAL);
1014
1015                 /* only files on same MDT can have their layouts swapped */
1016                 if (tgt1->ltd_index != tgt2->ltd_index)
1017                         RETURN(-EPERM);
1018
1019                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1020                 break;
1021         }
1022         case LL_IOC_HSM_CT_START: {
1023                 struct lustre_kernelcomm *lk = karg;
1024                 if (lk->lk_flags & LK_FLG_STOP)
1025                         rc = lmv_hsm_ct_unregister(obd, cmd, len, lk, uarg);
1026                 else
1027                         rc = lmv_hsm_ct_register(obd, cmd, len, lk, uarg);
1028                 break;
1029         }
1030         default:
1031                 lmv_foreach_connected_tgt(lmv, tgt) {
1032                         struct obd_device *mdc_obd;
1033                         int err;
1034
1035                         /* ll_umount_begin() sets force flag but for lmv, not
1036                          * mdc. Let's pass it through */
1037                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1038                         mdc_obd->obd_force = obd->obd_force;
1039                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1040                         if (err) {
1041                                 if (tgt->ltd_active) {
1042                                         CERROR("error: iocontrol MDC %s on MDT"
1043                                                " idx %d cmd %x: err = %d\n",
1044                                                tgt->ltd_uuid.uuid,
1045                                                tgt->ltd_index, cmd, err);
1046                                         if (!rc)
1047                                                 rc = err;
1048                                 }
1049                         } else
1050                                 set = 1;
1051                 }
1052                 if (!set && !rc)
1053                         rc = -EIO;
1054         }
1055         RETURN(rc);
1056 }
1057
1058 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1059                   struct lu_fid *fid, struct md_op_data *op_data)
1060 {
1061         struct obd_device *obd = class_exp2obd(exp);
1062         struct lmv_obd *lmv = &obd->u.lmv;
1063         struct lmv_tgt_desc *tgt;
1064         int rc;
1065
1066         ENTRY;
1067
1068         LASSERT(op_data);
1069         LASSERT(fid);
1070
1071         tgt = lmv_tgt(lmv, op_data->op_mds);
1072         if (!tgt)
1073                 RETURN(-ENODEV);
1074
1075         if (!tgt->ltd_active || !tgt->ltd_exp)
1076                 RETURN(-ENODEV);
1077
1078         /*
1079          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1080          * on server that seq in new allocated fid is not yet known.
1081          */
1082         mutex_lock(&tgt->ltd_fid_mutex);
1083         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1084         mutex_unlock(&tgt->ltd_fid_mutex);
1085         if (rc > 0) {
1086                 LASSERT(fid_is_sane(fid));
1087                 rc = 0;
1088         }
1089
1090         RETURN(rc);
1091 }
1092
1093 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1094 {
1095         struct lmv_obd *lmv = &obd->u.lmv;
1096         struct lmv_desc *desc;
1097         struct lnet_process_id lnet_id;
1098         int i = 0;
1099         int rc;
1100
1101         ENTRY;
1102
1103         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1104                 CERROR("LMV setup requires a descriptor\n");
1105                 RETURN(-EINVAL);
1106         }
1107
1108         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1109         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1110                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1111                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1112                 RETURN(-EINVAL);
1113         }
1114
1115         obd_str2uuid(&lmv->lmv_mdt_descs.ltd_lmv_desc.ld_uuid,
1116                      desc->ld_uuid.uuid);
1117         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count = 0;
1118         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count = 0;
1119         lmv->lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage =
1120                 LMV_DESC_QOS_MAXAGE_DEFAULT;
1121         lmv->max_def_easize = 0;
1122         lmv->max_easize = 0;
1123
1124         spin_lock_init(&lmv->lmv_lock);
1125
1126         /*
1127          * initialize rr_index to lower 32bit of netid, so that client
1128          * can distribute subdirs evenly from the beginning.
1129          */
1130         while (LNetGetId(i++, &lnet_id) != -ENOENT) {
1131                 if (lnet_id.nid != LNET_NID_LO_0) {
1132                         lmv->lmv_qos_rr_index = (u32)lnet_id.nid;
1133                         break;
1134                 }
1135         }
1136
1137         rc = lmv_tunables_init(obd);
1138         if (rc)
1139                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1140                       obd->obd_name, rc);
1141
1142         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1143                              LUSTRE_CLI_FLD_HASH_DHT);
1144         if (rc)
1145                 CERROR("Can't init FLD, err %d\n", rc);
1146
1147         rc = lu_tgt_descs_init(&lmv->lmv_mdt_descs, true);
1148         if (rc)
1149                 CWARN("%s: error initialize target table: rc = %d\n",
1150                       obd->obd_name, rc);
1151
1152         RETURN(rc);
1153 }
1154
1155 static int lmv_cleanup(struct obd_device *obd)
1156 {
1157         struct lmv_obd *lmv = &obd->u.lmv;
1158         struct lu_tgt_desc *tgt;
1159         struct lu_tgt_desc *tmp;
1160
1161         ENTRY;
1162
1163         fld_client_fini(&lmv->lmv_fld);
1164         lmv_foreach_tgt_safe(lmv, tgt, tmp)
1165                 lmv_del_target(lmv, tgt);
1166         lu_tgt_descs_fini(&lmv->lmv_mdt_descs);
1167
1168         RETURN(0);
1169 }
1170
1171 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1172 {
1173         struct lustre_cfg       *lcfg = buf;
1174         struct obd_uuid         obd_uuid;
1175         int                     gen;
1176         __u32                   index;
1177         int                     rc;
1178         ENTRY;
1179
1180         switch (lcfg->lcfg_command) {
1181         case LCFG_ADD_MDC:
1182                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1183                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1184                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1185                         GOTO(out, rc = -EINVAL);
1186
1187                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1188
1189                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1190                         GOTO(out, rc = -EINVAL);
1191                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1192                         GOTO(out, rc = -EINVAL);
1193                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1194                 GOTO(out, rc);
1195         default:
1196                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1197                 GOTO(out, rc = -EINVAL);
1198         }
1199 out:
1200         RETURN(rc);
1201 }
1202
1203 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1204 {
1205         int i;
1206
1207         if (flags & OBD_STATFS_FOR_MDT0)
1208                 return 0;
1209
1210         if (lmv->lmv_statfs_start || lmv->lmv_mdt_count == 1)
1211                 return lmv->lmv_statfs_start;
1212
1213         /* choose initial MDT for this client */
1214         for (i = 0;; i++) {
1215                 struct lnet_process_id lnet_id;
1216                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1217                         break;
1218
1219                 if (lnet_id.nid != LNET_NID_LO_0) {
1220                         /* We dont need a full 64-bit modulus, just enough
1221                          * to distribute the requests across MDTs evenly.
1222                          */
1223                         lmv->lmv_statfs_start = (u32)lnet_id.nid %
1224                                                 lmv->lmv_mdt_count;
1225                         break;
1226                 }
1227         }
1228
1229         return lmv->lmv_statfs_start;
1230 }
1231
1232 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1233                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1234 {
1235         struct obd_device *obd = class_exp2obd(exp);
1236         struct lmv_obd *lmv = &obd->u.lmv;
1237         struct obd_statfs *temp;
1238         struct lu_tgt_desc *tgt;
1239         __u32 i;
1240         __u32 idx;
1241         int rc = 0;
1242
1243         ENTRY;
1244
1245         OBD_ALLOC(temp, sizeof(*temp));
1246         if (temp == NULL)
1247                 RETURN(-ENOMEM);
1248
1249         /* distribute statfs among MDTs */
1250         idx = lmv_select_statfs_mdt(lmv, flags);
1251
1252         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++, idx++) {
1253                 idx = idx % lmv->lmv_mdt_descs.ltd_tgts_size;
1254                 tgt = lmv_tgt(lmv, idx);
1255                 if (!tgt || !tgt->ltd_exp)
1256                         continue;
1257
1258                 rc = obd_statfs(env, tgt->ltd_exp, temp, max_age,
1259                                 flags | OBD_STATFS_NESTED);
1260                 if (rc) {
1261                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1262                                tgt->ltd_exp->exp_obd->obd_name, i, rc);
1263                         GOTO(out_free_temp, rc);
1264                 }
1265
1266                 if (temp->os_state & OS_STATFS_SUM ||
1267                     flags == OBD_STATFS_FOR_MDT0) {
1268                         /* reset to the last aggregated values
1269                          * and don't sum with non-aggrated data */
1270                         /* If the statfs is from mount, it needs to retrieve
1271                          * necessary information from MDT0. i.e. mount does
1272                          * not need the merged osfs from all of MDT. Also
1273                          * clients can be mounted as long as MDT0 is in
1274                          * service */
1275                         *osfs = *temp;
1276                         break;
1277                 }
1278
1279                 if (i == 0) {
1280                         *osfs = *temp;
1281                 } else {
1282                         osfs->os_bavail += temp->os_bavail;
1283                         osfs->os_blocks += temp->os_blocks;
1284                         osfs->os_ffree += temp->os_ffree;
1285                         osfs->os_files += temp->os_files;
1286                         osfs->os_granted += temp->os_granted;
1287                 }
1288         }
1289
1290         EXIT;
1291 out_free_temp:
1292         OBD_FREE(temp, sizeof(*temp));
1293         return rc;
1294 }
1295
1296 static int lmv_statfs_update(void *cookie, int rc)
1297 {
1298         struct obd_info *oinfo = cookie;
1299         struct obd_device *obd = oinfo->oi_obd;
1300         struct lmv_obd *lmv = &obd->u.lmv;
1301         struct lmv_tgt_desc *tgt = oinfo->oi_tgt;
1302         struct obd_statfs *osfs = oinfo->oi_osfs;
1303
1304         /*
1305          * NB: don't deactivate TGT upon error, because we may not trigger async
1306          * statfs any longer, then there is no chance to activate TGT.
1307          */
1308         if (!rc) {
1309                 spin_lock(&lmv->lmv_lock);
1310                 tgt->ltd_statfs = *osfs;
1311                 tgt->ltd_statfs_age = ktime_get_seconds();
1312                 spin_unlock(&lmv->lmv_lock);
1313                 set_bit(LQ_DIRTY, &lmv->lmv_qos.lq_flags);
1314         }
1315
1316         return rc;
1317 }
1318
1319 /* update tgt statfs async if it's ld_qos_maxage old */
1320 int lmv_statfs_check_update(struct obd_device *obd, struct lmv_tgt_desc *tgt)
1321 {
1322         struct obd_info oinfo = {
1323                 .oi_obd = obd,
1324                 .oi_tgt = tgt,
1325                 .oi_cb_up = lmv_statfs_update,
1326         };
1327         int rc;
1328
1329         if (ktime_get_seconds() - tgt->ltd_statfs_age <
1330             obd->u.lmv.lmv_mdt_descs.ltd_lmv_desc.ld_qos_maxage)
1331                 return 0;
1332
1333         rc = obd_statfs_async(tgt->ltd_exp, &oinfo, 0, NULL);
1334
1335         return rc;
1336 }
1337
1338 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1339                         struct lu_fid *fid)
1340 {
1341         struct obd_device *obd = exp->exp_obd;
1342         struct lmv_obd *lmv = &obd->u.lmv;
1343         struct lu_tgt_desc *tgt = lmv_tgt(lmv, 0);
1344         int rc;
1345
1346         ENTRY;
1347
1348         if (!tgt)
1349                 RETURN(-ENODEV);
1350
1351         rc = md_get_root(tgt->ltd_exp, fileset, fid);
1352         RETURN(rc);
1353 }
1354
1355 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1356                         u64 obd_md_valid, const char *name, size_t buf_size,
1357                         struct ptlrpc_request **req)
1358 {
1359         struct obd_device *obd = exp->exp_obd;
1360         struct lmv_obd *lmv = &obd->u.lmv;
1361         struct lmv_tgt_desc *tgt;
1362         int rc;
1363
1364         ENTRY;
1365
1366         tgt = lmv_fid2tgt(lmv, fid);
1367         if (IS_ERR(tgt))
1368                 RETURN(PTR_ERR(tgt));
1369
1370         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1371
1372         RETURN(rc);
1373 }
1374
1375 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1376                         u64 obd_md_valid, const char *name,
1377                         const void *value, size_t value_size,
1378                         unsigned int xattr_flags, u32 suppgid,
1379                         struct ptlrpc_request **req)
1380 {
1381         struct obd_device *obd = exp->exp_obd;
1382         struct lmv_obd *lmv = &obd->u.lmv;
1383         struct lmv_tgt_desc *tgt;
1384         int rc;
1385
1386         ENTRY;
1387
1388         tgt = lmv_fid2tgt(lmv, fid);
1389         if (IS_ERR(tgt))
1390                 RETURN(PTR_ERR(tgt));
1391
1392         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1393                          value, value_size, xattr_flags, suppgid, req);
1394
1395         RETURN(rc);
1396 }
1397
1398 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1399                        struct ptlrpc_request **request)
1400 {
1401         struct obd_device *obd = exp->exp_obd;
1402         struct lmv_obd *lmv = &obd->u.lmv;
1403         struct lmv_tgt_desc *tgt;
1404         int rc;
1405
1406         ENTRY;
1407
1408         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1409         if (IS_ERR(tgt))
1410                 RETURN(PTR_ERR(tgt));
1411
1412         if (op_data->op_flags & MF_GET_MDT_IDX) {
1413                 op_data->op_mds = tgt->ltd_index;
1414                 RETURN(0);
1415         }
1416
1417         rc = md_getattr(tgt->ltd_exp, op_data, request);
1418
1419         RETURN(rc);
1420 }
1421
1422 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1423 {
1424         struct obd_device *obd = exp->exp_obd;
1425         struct lmv_obd *lmv = &obd->u.lmv;
1426         struct lu_tgt_desc *tgt;
1427
1428         ENTRY;
1429
1430         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1431
1432         /*
1433          * With DNE every object can have two locks in different namespaces:
1434          * lookup lock in space of MDT storing direntry and update/open lock in
1435          * space of MDT storing inode.
1436          */
1437         lmv_foreach_connected_tgt(lmv, tgt)
1438                 md_null_inode(tgt->ltd_exp, fid);
1439
1440         RETURN(0);
1441 }
1442
1443 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1444                      struct md_open_data *mod, struct ptlrpc_request **request)
1445 {
1446         struct obd_device *obd = exp->exp_obd;
1447         struct lmv_obd *lmv = &obd->u.lmv;
1448         struct lmv_tgt_desc *tgt;
1449         int rc;
1450
1451         ENTRY;
1452
1453         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
1454         if (IS_ERR(tgt))
1455                 RETURN(PTR_ERR(tgt));
1456
1457         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1458         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1459         RETURN(rc);
1460 }
1461
1462 static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 mdt,
1463                                               unsigned short dir_depth)
1464 {
1465         struct lu_tgt_desc *tgt, *cur = NULL;
1466         __u64 total_avail = 0;
1467         __u64 total_weight = 0;
1468         __u64 cur_weight = 0;
1469         int total_usable = 0;
1470         __u64 rand;
1471         int rc;
1472
1473         ENTRY;
1474
1475         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1476                 RETURN(ERR_PTR(-EAGAIN));
1477
1478         down_write(&lmv->lmv_qos.lq_rw_sem);
1479
1480         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1481                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1482
1483         rc = ltd_qos_penalties_calc(&lmv->lmv_mdt_descs);
1484         if (rc)
1485                 GOTO(unlock, tgt = ERR_PTR(rc));
1486
1487         lmv_foreach_tgt(lmv, tgt) {
1488                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1489                         tgt->ltd_qos.ltq_usable = 0;
1490                         continue;
1491                 }
1492
1493                 tgt->ltd_qos.ltq_usable = 1;
1494                 lu_tgt_qos_weight_calc(tgt);
1495                 if (tgt->ltd_index == mdt)
1496                         cur = tgt;
1497                 total_avail += tgt->ltd_qos.ltq_avail;
1498                 total_weight += tgt->ltd_qos.ltq_weight;
1499                 total_usable++;
1500         }
1501
1502         /* if current MDT has above-average space, within range of the QOS
1503          * threshold, stay on the same MDT to avoid creating needless remote
1504          * MDT directories. It's more likely for low level directories.
1505          */
1506         rand = total_avail * (256 - lmv->lmv_qos.lq_threshold_rr) /
1507                (total_usable * 256 * (1 + dir_depth / 4));
1508         if (cur && cur->ltd_qos.ltq_avail >= rand) {
1509                 tgt = cur;
1510                 GOTO(unlock, tgt);
1511         }
1512
1513         rand = lu_prandom_u64_max(total_weight);
1514
1515         lmv_foreach_connected_tgt(lmv, tgt) {
1516                 if (!tgt->ltd_qos.ltq_usable)
1517                         continue;
1518
1519                 cur_weight += tgt->ltd_qos.ltq_weight;
1520                 if (cur_weight < rand)
1521                         continue;
1522
1523                 ltd_qos_update(&lmv->lmv_mdt_descs, tgt, &total_weight);
1524                 GOTO(unlock, tgt);
1525         }
1526
1527         /* no proper target found */
1528         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1529 unlock:
1530         up_write(&lmv->lmv_qos.lq_rw_sem);
1531
1532         return tgt;
1533 }
1534
1535 static struct lu_tgt_desc *lmv_locate_tgt_rr(struct lmv_obd *lmv)
1536 {
1537         struct lu_tgt_desc *tgt;
1538         int i;
1539         int index;
1540
1541         ENTRY;
1542
1543         spin_lock(&lmv->lmv_lock);
1544         for (i = 0; i < lmv->lmv_mdt_descs.ltd_tgts_size; i++) {
1545                 index = (i + lmv->lmv_qos_rr_index) %
1546                         lmv->lmv_mdt_descs.ltd_tgts_size;
1547                 tgt = lmv_tgt(lmv, index);
1548                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
1549                         continue;
1550
1551                 lmv->lmv_qos_rr_index = (tgt->ltd_index + 1) %
1552                                         lmv->lmv_mdt_descs.ltd_tgts_size;
1553                 spin_unlock(&lmv->lmv_lock);
1554
1555                 RETURN(tgt);
1556         }
1557         spin_unlock(&lmv->lmv_lock);
1558
1559         RETURN(ERR_PTR(-ENODEV));
1560 }
1561
1562 /* locate MDT which is less full (avoid the most full MDT) */
1563 static struct lu_tgt_desc *lmv_locate_tgt_lf(struct lmv_obd *lmv)
1564 {
1565         struct lu_tgt_desc *min = NULL;
1566         struct lu_tgt_desc *tgt;
1567         __u64 avail = 0;
1568         __u64 rand;
1569
1570         ENTRY;
1571
1572         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1573                 RETURN(ERR_PTR(-EAGAIN));
1574
1575         down_write(&lmv->lmv_qos.lq_rw_sem);
1576
1577         if (!ltd_qos_is_usable(&lmv->lmv_mdt_descs))
1578                 GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1579
1580         lmv_foreach_tgt(lmv, tgt) {
1581                 if (!tgt->ltd_exp || !tgt->ltd_active) {
1582                         tgt->ltd_qos.ltq_usable = 0;
1583                         continue;
1584                 }
1585
1586                 tgt->ltd_qos.ltq_usable = 1;
1587                 lu_tgt_qos_weight_calc(tgt);
1588                 avail += tgt->ltd_qos.ltq_avail;
1589                 if (!min || min->ltd_qos.ltq_avail > tgt->ltd_qos.ltq_avail)
1590                         min = tgt;
1591         }
1592
1593         /* avoid the most full MDT */
1594         if (min)
1595                 avail -= min->ltd_qos.ltq_avail;
1596
1597         rand = lu_prandom_u64_max(avail);
1598         avail = 0;
1599         lmv_foreach_connected_tgt(lmv, tgt) {
1600                 if (!tgt->ltd_qos.ltq_usable)
1601                         continue;
1602
1603                 if (tgt == min)
1604                         continue;
1605
1606                 avail += tgt->ltd_qos.ltq_avail;
1607                 if (avail < rand)
1608                         continue;
1609
1610                 GOTO(unlock, tgt);
1611         }
1612
1613         /* no proper target found */
1614         GOTO(unlock, tgt = ERR_PTR(-EAGAIN));
1615 unlock:
1616         up_write(&lmv->lmv_qos.lq_rw_sem);
1617
1618         RETURN(tgt);
1619 }
1620
1621 /* locate MDT by file name, for striped directory, the file name hash decides
1622  * which stripe its dirent is stored.
1623  */
1624 static struct lmv_tgt_desc *
1625 lmv_locate_tgt_by_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1626                        const char *name, int namelen, struct lu_fid *fid,
1627                        __u32 *mds, bool new_layout)
1628 {
1629         struct lmv_tgt_desc *tgt;
1630         const struct lmv_oinfo *oinfo;
1631
1632         if (!lmv_dir_striped(lsm) || !namelen) {
1633                 tgt = lmv_fid2tgt(lmv, fid);
1634                 if (IS_ERR(tgt))
1635                         return tgt;
1636
1637                 *mds = tgt->ltd_index;
1638                 return tgt;
1639         }
1640
1641         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1642                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1643                         return ERR_PTR(-EBADF);
1644                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1645         } else {
1646                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, new_layout);
1647                 if (IS_ERR(oinfo))
1648                         return ERR_CAST(oinfo);
1649         }
1650
1651         /* check stripe FID is sane */
1652         if (!fid_is_sane(&oinfo->lmo_fid))
1653                 return ERR_PTR(-ENODEV);
1654
1655         *fid = oinfo->lmo_fid;
1656         *mds = oinfo->lmo_mds;
1657         tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1658
1659         CDEBUG(D_INODE, "locate MDT %u parent "DFID"\n", *mds, PFID(fid));
1660
1661         return tgt ? tgt : ERR_PTR(-ENODEV);
1662 }
1663
1664 /**
1665  * Locate MDT of op_data->op_fid1
1666  *
1667  * For striped directory, it will locate the stripe by name hash, if hash_type
1668  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1669  * which is set outside, and if dir is migrating, 'op_data->op_new_layout'
1670  * indicates whether old or new layout is used to locate.
1671  *
1672  * For plain direcotry, it just locate the MDT of op_data->op_fid1.
1673  *
1674  * \param[in] lmv               LMV device
1675  * \param[in/out] op_data       client MD stack parameters, name, namelen etc,
1676  *                              op_mds and op_fid1 will be updated if op_mea1
1677  *                              indicates fid1 represents a striped directory.
1678  *
1679  * retval               pointer to the lmv_tgt_desc if succeed.
1680  *                      ERR_PTR(errno) if failed.
1681  */
1682 struct lmv_tgt_desc *
1683 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data)
1684 {
1685         struct lmv_stripe_md *lsm = op_data->op_mea1;
1686         struct lmv_oinfo *oinfo;
1687         struct lmv_tgt_desc *tgt;
1688
1689         if (lmv_dir_foreign(lsm))
1690                 return ERR_PTR(-ENODATA);
1691
1692         /* During creating VOLATILE file, it should honor the mdt
1693          * index if the file under striped dir is being restored, see
1694          * ct_restore(). */
1695         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1696             op_data->op_mds != LMV_OFFSET_DEFAULT) {
1697                 tgt = lmv_tgt(lmv, op_data->op_mds);
1698                 if (!tgt)
1699                         return ERR_PTR(-ENODEV);
1700
1701                 if (lmv_dir_striped(lsm)) {
1702                         int i;
1703
1704                         /* refill the right parent fid */
1705                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1706                                 oinfo = &lsm->lsm_md_oinfo[i];
1707                                 if (oinfo->lmo_mds == op_data->op_mds) {
1708                                         op_data->op_fid1 = oinfo->lmo_fid;
1709                                         break;
1710                                 }
1711                         }
1712
1713                         if (i == lsm->lsm_md_stripe_count)
1714                                 op_data->op_fid1 = lsm->lsm_md_oinfo[0].lmo_fid;
1715                 }
1716         } else if (lmv_dir_bad_hash(lsm)) {
1717                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1718                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1719
1720                 op_data->op_fid1 = oinfo->lmo_fid;
1721                 op_data->op_mds = oinfo->lmo_mds;
1722                 tgt = lmv_tgt(lmv, oinfo->lmo_mds);
1723                 if (!tgt)
1724                         return ERR_PTR(-ENODEV);
1725         } else {
1726                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea1,
1727                                 op_data->op_name, op_data->op_namelen,
1728                                 &op_data->op_fid1, &op_data->op_mds,
1729                                 op_data->op_new_layout);
1730         }
1731
1732         return tgt;
1733 }
1734
1735 /* Locate MDT of op_data->op_fid2 for link/rename */
1736 static struct lmv_tgt_desc *
1737 lmv_locate_tgt2(struct lmv_obd *lmv, struct md_op_data *op_data)
1738 {
1739         struct lmv_tgt_desc *tgt;
1740         int rc;
1741
1742         LASSERT(op_data->op_name);
1743         if (lmv_dir_layout_changing(op_data->op_mea2)) {
1744                 struct lu_fid fid1 = op_data->op_fid1;
1745                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1746                 struct ptlrpc_request *request = NULL;
1747
1748                 /*
1749                  * avoid creating new file under old layout of migrating
1750                  * directory, check it here.
1751                  */
1752                 tgt = lmv_locate_tgt_by_name(lmv, op_data->op_mea2,
1753                                 op_data->op_name, op_data->op_namelen,
1754                                 &op_data->op_fid2, &op_data->op_mds, false);
1755                 if (IS_ERR(tgt))
1756                         RETURN(tgt);
1757
1758                 op_data->op_fid1 = op_data->op_fid2;
1759                 op_data->op_mea1 = op_data->op_mea2;
1760                 rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1761                 op_data->op_fid1 = fid1;
1762                 op_data->op_mea1 = lsm1;
1763                 if (!rc) {
1764                         ptlrpc_req_finished(request);
1765                         RETURN(ERR_PTR(-EEXIST));
1766                 }
1767
1768                 if (rc != -ENOENT)
1769                         RETURN(ERR_PTR(rc));
1770         }
1771
1772         return lmv_locate_tgt_by_name(lmv, op_data->op_mea2, op_data->op_name,
1773                                 op_data->op_namelen, &op_data->op_fid2,
1774                                 &op_data->op_mds, true);
1775 }
1776
1777 int lmv_old_layout_lookup(struct lmv_obd *lmv, struct md_op_data *op_data)
1778 {
1779         struct lu_tgt_desc *tgt;
1780         struct ptlrpc_request *request;
1781         int rc;
1782
1783         LASSERT(lmv_dir_layout_changing(op_data->op_mea1));
1784         LASSERT(!op_data->op_new_layout);
1785
1786         tgt = lmv_locate_tgt(lmv, op_data);
1787         if (IS_ERR(tgt))
1788                 return PTR_ERR(tgt);
1789
1790         rc = md_getattr_name(tgt->ltd_exp, op_data, &request);
1791         if (!rc) {
1792                 ptlrpc_req_finished(request);
1793                 return -EEXIST;
1794         }
1795
1796         return rc;
1797 }
1798
1799 /* mkdir by QoS upon 'lfs mkdir -i -1'.
1800  *
1801  * NB, mkdir by QoS only if parent is not striped, this is to avoid remote
1802  * directories under striped directory.
1803  */
1804 static inline bool lmv_op_user_qos_mkdir(const struct md_op_data *op_data)
1805 {
1806         const struct lmv_user_md *lum = op_data->op_data;
1807
1808         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1809                 return false;
1810
1811         if (lmv_dir_striped(op_data->op_mea1))
1812                 return false;
1813
1814         return (op_data->op_cli_flags & CLI_SET_MEA) && lum &&
1815                le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC &&
1816                le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
1817 }
1818
1819 /* mkdir by QoS if either ROOT or parent default LMV is space balanced. */
1820 static inline bool lmv_op_default_qos_mkdir(const struct md_op_data *op_data)
1821 {
1822         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1823
1824         if (op_data->op_code != LUSTRE_OPC_MKDIR)
1825                 return false;
1826
1827         if (lmv_dir_striped(op_data->op_mea1))
1828                 return false;
1829
1830         return (op_data->op_flags & MF_QOS_MKDIR) ||
1831                (lsm && lsm->lsm_md_master_mdt_index == LMV_OFFSET_DEFAULT);
1832 }
1833
1834 /* if parent default LMV is space balanced, and
1835  * 1. max_inherit_rr is set
1836  * 2. or parent is ROOT
1837  * mkdir roundrobin. Or if parent doesn't have default LMV, while ROOT default
1838  * LMV requests roundrobin mkdir, do the same.
1839  * NB, this needs to check server is balanced, which is done by caller.
1840  */
1841 static inline bool lmv_op_default_rr_mkdir(const struct md_op_data *op_data)
1842 {
1843         const struct lmv_stripe_md *lsm = op_data->op_default_mea1;
1844
1845         if (!lmv_op_default_qos_mkdir(op_data))
1846                 return false;
1847
1848         return (op_data->op_flags & MF_RR_MKDIR) ||
1849                (lsm && lsm->lsm_md_max_inherit_rr != LMV_INHERIT_RR_NONE) ||
1850                fid_is_root(&op_data->op_fid1);
1851 }
1852
1853 /* 'lfs mkdir -i <specific_MDT>' */
1854 static inline bool lmv_op_user_specific_mkdir(const struct md_op_data *op_data)
1855 {
1856         const struct lmv_user_md *lum = op_data->op_data;
1857
1858         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1859                op_data->op_cli_flags & CLI_SET_MEA && lum &&
1860                (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC ||
1861                 le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) &&
1862                le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
1863 }
1864
1865 /* parent default LMV master_mdt_index is not -1. */
1866 static inline bool
1867 lmv_op_default_specific_mkdir(const struct md_op_data *op_data)
1868 {
1869         return op_data->op_code == LUSTRE_OPC_MKDIR &&
1870                op_data->op_default_mea1 &&
1871                op_data->op_default_mea1->lsm_md_master_mdt_index !=
1872                         LMV_OFFSET_DEFAULT;
1873 }
1874
1875 /* locate MDT by space usage */
1876 static struct lu_tgt_desc *lmv_locate_tgt_by_space(struct lmv_obd *lmv,
1877                                                    struct md_op_data *op_data,
1878                                                    struct lmv_tgt_desc *tgt)
1879 {
1880         struct lmv_tgt_desc *tmp = tgt;
1881
1882         tgt = lmv_locate_tgt_qos(lmv, op_data->op_mds, op_data->op_dir_depth);
1883         if (tgt == ERR_PTR(-EAGAIN)) {
1884                 if (ltd_qos_is_balanced(&lmv->lmv_mdt_descs) &&
1885                     !lmv_op_default_rr_mkdir(op_data) &&
1886                     !lmv_op_user_qos_mkdir(op_data))
1887                         /* if not necessary, don't create remote directory. */
1888                         tgt = tmp;
1889                 else
1890                         tgt = lmv_locate_tgt_rr(lmv);
1891         }
1892
1893         /*
1894          * only update statfs after QoS mkdir, this means the cached statfs may
1895          * be stale, and current mkdir may not follow QoS accurately, but it's
1896          * not serious, and avoids periodic statfs when client doesn't mkdir by
1897          * QoS.
1898          */
1899         if (!IS_ERR(tgt)) {
1900                 op_data->op_mds = tgt->ltd_index;
1901                 lmv_statfs_check_update(lmv2obd_dev(lmv), tgt);
1902         }
1903
1904         return tgt;
1905 }
1906
1907 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1908                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1909                 gid_t gid, kernel_cap_t cap_effective, __u64 rdev,
1910                 struct ptlrpc_request **request)
1911 {
1912         struct obd_device *obd = exp->exp_obd;
1913         struct lmv_obd *lmv = &obd->u.lmv;
1914         struct lmv_tgt_desc *tgt;
1915         struct mdt_body *repbody;
1916         int rc;
1917
1918         ENTRY;
1919
1920         if (!lmv->lmv_mdt_descs.ltd_lmv_desc.ld_active_tgt_count)
1921                 RETURN(-EIO);
1922
1923         if (lmv_dir_bad_hash(op_data->op_mea1))
1924                 RETURN(-EBADF);
1925
1926         if (lmv_dir_layout_changing(op_data->op_mea1)) {
1927                 /*
1928                  * if parent is migrating, create() needs to lookup existing
1929                  * name in both old and new layout, check old layout on client.
1930                  */
1931                 rc = lmv_old_layout_lookup(lmv, op_data);
1932                 if (rc != -ENOENT)
1933                         RETURN(rc);
1934
1935                 op_data->op_new_layout = true;
1936         }
1937
1938         tgt = lmv_locate_tgt(lmv, op_data);
1939         if (IS_ERR(tgt))
1940                 RETURN(PTR_ERR(tgt));
1941
1942         /* the order to apply policy in mkdir:
1943          * 1. is "lfs mkdir -i N"? mkdir on MDT N.
1944          * 2. is "lfs mkdir -i -1"? mkdir by space usage.
1945          * 3. is starting MDT specified in default LMV? mkdir on MDT N.
1946          * 4. is default LMV space balanced? mkdir by space usage.
1947          */
1948         if (lmv_op_user_specific_mkdir(op_data)) {
1949                 struct lmv_user_md *lum = op_data->op_data;
1950
1951                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
1952                 tgt = lmv_tgt(lmv, op_data->op_mds);
1953                 if (!tgt)
1954                         RETURN(-ENODEV);
1955         } else if (lmv_op_user_qos_mkdir(op_data)) {
1956                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
1957                 if (IS_ERR(tgt))
1958                         RETURN(PTR_ERR(tgt));
1959         } else if (lmv_op_default_specific_mkdir(op_data)) {
1960                 op_data->op_mds =
1961                         op_data->op_default_mea1->lsm_md_master_mdt_index;
1962                 tgt = lmv_tgt(lmv, op_data->op_mds);
1963                 if (!tgt)
1964                         RETURN(-ENODEV);
1965         } else if (lmv_op_default_qos_mkdir(op_data)) {
1966                 tgt = lmv_locate_tgt_by_space(lmv, op_data, tgt);
1967                 if (IS_ERR(tgt))
1968                         RETURN(PTR_ERR(tgt));
1969         }
1970
1971 retry:
1972         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1973         if (rc)
1974                 RETURN(rc);
1975
1976         CDEBUG(D_INODE, "CREATE name '%.*s' "DFID" on "DFID" -> mds #%x\n",
1977                 (int)op_data->op_namelen, op_data->op_name,
1978                 PFID(&op_data->op_fid2), PFID(&op_data->op_fid1),
1979                 op_data->op_mds);
1980
1981         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1982         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1983                        cap_effective, rdev, request);
1984         if (rc == 0) {
1985                 if (*request == NULL)
1986                         RETURN(rc);
1987                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1988         }
1989
1990         /* dir restripe needs to send to MDT where dir is located */
1991         if (rc != -EREMOTE ||
1992             !(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH))
1993                 RETURN(rc);
1994
1995         repbody = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
1996         if (repbody == NULL)
1997                 RETURN(-EPROTO);
1998
1999         /* Not cross-ref case, just get out of here. */
2000         if (likely(!(repbody->mbo_valid & OBD_MD_MDS)))
2001                 RETURN(rc);
2002
2003         op_data->op_fid2 = repbody->mbo_fid1;
2004         ptlrpc_req_finished(*request);
2005         *request = NULL;
2006
2007         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
2008         if (IS_ERR(tgt))
2009                 RETURN(PTR_ERR(tgt));
2010
2011         op_data->op_mds = tgt->ltd_index;
2012         goto retry;
2013 }
2014
2015 static int
2016 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
2017             const union ldlm_policy_data *policy, struct md_op_data *op_data,
2018             struct lustre_handle *lockh, __u64 extra_lock_flags)
2019 {
2020         struct obd_device *obd = exp->exp_obd;
2021         struct lmv_obd *lmv = &obd->u.lmv;
2022         struct lmv_tgt_desc *tgt;
2023         int rc;
2024
2025         ENTRY;
2026
2027         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
2028
2029         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2030         if (IS_ERR(tgt))
2031                 RETURN(PTR_ERR(tgt));
2032
2033         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
2034                PFID(&op_data->op_fid1), tgt->ltd_index);
2035
2036         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
2037                         extra_lock_flags);
2038
2039         RETURN(rc);
2040 }
2041
2042 int
2043 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
2044                  struct ptlrpc_request **preq)
2045 {
2046         struct obd_device *obd = exp->exp_obd;
2047         struct lmv_obd *lmv = &obd->u.lmv;
2048         struct lmv_tgt_desc *tgt;
2049         struct mdt_body *body;
2050         int rc;
2051
2052         ENTRY;
2053
2054 retry:
2055         if (op_data->op_namelen == 2 &&
2056             op_data->op_name[0] == '.' && op_data->op_name[1] == '.')
2057                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2058         else
2059                 tgt = lmv_locate_tgt(lmv, op_data);
2060         if (IS_ERR(tgt))
2061                 RETURN(PTR_ERR(tgt));
2062
2063         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
2064                 (int)op_data->op_namelen, op_data->op_name,
2065                 PFID(&op_data->op_fid1), tgt->ltd_index);
2066
2067         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
2068         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2069                 ptlrpc_req_finished(*preq);
2070                 *preq = NULL;
2071                 goto retry;
2072         }
2073
2074         if (rc)
2075                 RETURN(rc);
2076
2077         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
2078         LASSERT(body != NULL);
2079
2080         if (body->mbo_valid & OBD_MD_MDS) {
2081                 op_data->op_fid1 = body->mbo_fid1;
2082                 op_data->op_valid |= OBD_MD_FLCROSSREF;
2083                 op_data->op_namelen = 0;
2084                 op_data->op_name = NULL;
2085
2086                 ptlrpc_req_finished(*preq);
2087                 *preq = NULL;
2088
2089                 goto retry;
2090         }
2091
2092         RETURN(rc);
2093 }
2094
2095 #define md_op_data_fid(op_data, fl)                     \
2096         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
2097          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
2098          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
2099          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
2100          NULL)
2101
2102 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
2103                             struct md_op_data *op_data, __u32 op_tgt,
2104                             enum ldlm_mode mode, int bits, int flag)
2105 {
2106         struct lu_fid *fid = md_op_data_fid(op_data, flag);
2107         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
2108         union ldlm_policy_data policy = { { 0 } };
2109         int rc = 0;
2110         ENTRY;
2111
2112         if (!fid_is_sane(fid))
2113                 RETURN(0);
2114
2115         if (tgt == NULL) {
2116                 tgt = lmv_fid2tgt(lmv, fid);
2117                 if (IS_ERR(tgt))
2118                         RETURN(PTR_ERR(tgt));
2119         }
2120
2121         if (tgt->ltd_index != op_tgt) {
2122                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
2123                 policy.l_inodebits.bits = bits;
2124                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
2125                                       mode, LCF_ASYNC, NULL);
2126         } else {
2127                 CDEBUG(D_INODE,
2128                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
2129                        op_tgt, PFID(fid));
2130                 op_data->op_flags |= flag;
2131                 rc = 0;
2132         }
2133
2134         RETURN(rc);
2135 }
2136
2137 /*
2138  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
2139  * op_data->op_fid2
2140  */
2141 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
2142                     struct ptlrpc_request **request)
2143 {
2144         struct obd_device       *obd = exp->exp_obd;
2145         struct lmv_obd          *lmv = &obd->u.lmv;
2146         struct lmv_tgt_desc     *tgt;
2147         int                      rc;
2148         ENTRY;
2149
2150         LASSERT(op_data->op_namelen != 0);
2151
2152         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
2153                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
2154                op_data->op_name, PFID(&op_data->op_fid1));
2155
2156         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2157         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2158         op_data->op_cap = current_cap();
2159
2160         tgt = lmv_locate_tgt2(lmv, op_data);
2161         if (IS_ERR(tgt))
2162                 RETURN(PTR_ERR(tgt));
2163
2164         /*
2165          * Cancel UPDATE lock on child (fid1).
2166          */
2167         op_data->op_flags |= MF_MDC_CANCEL_FID2;
2168         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
2169                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2170         if (rc != 0)
2171                 RETURN(rc);
2172
2173         rc = md_link(tgt->ltd_exp, op_data, request);
2174
2175         RETURN(rc);
2176 }
2177
2178 /* migrate the top directory */
2179 static inline bool lmv_op_topdir_migrate(const struct md_op_data *op_data)
2180 {
2181         if (!S_ISDIR(op_data->op_mode))
2182                 return false;
2183
2184         if (lmv_dir_layout_changing(op_data->op_mea1))
2185                 return false;
2186
2187         return true;
2188 }
2189
2190 /* migrate top dir to specific MDTs */
2191 static inline bool lmv_topdir_specific_migrate(const struct md_op_data *op_data)
2192 {
2193         const struct lmv_user_md *lum = op_data->op_data;
2194
2195         if (!lmv_op_topdir_migrate(op_data))
2196                 return false;
2197
2198         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2199 }
2200
2201 /* migrate top dir in QoS mode if user issued "lfs migrate -m -1..." */
2202 static inline bool lmv_topdir_qos_migrate(const struct md_op_data *op_data)
2203 {
2204         const struct lmv_user_md *lum = op_data->op_data;
2205
2206         if (!lmv_op_topdir_migrate(op_data))
2207                 return false;
2208
2209         return le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT;
2210 }
2211
2212 static inline bool lmv_subdir_specific_migrate(const struct md_op_data *op_data)
2213 {
2214         const struct lmv_user_md *lum = op_data->op_data;
2215
2216         if (!S_ISDIR(op_data->op_mode))
2217                 return false;
2218
2219         if (!lmv_dir_layout_changing(op_data->op_mea1))
2220                 return false;
2221
2222         return le32_to_cpu(lum->lum_stripe_offset) != LMV_OFFSET_DEFAULT;
2223 }
2224
2225 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
2226                         const char *name, size_t namelen,
2227                         struct ptlrpc_request **request)
2228 {
2229         struct obd_device *obd = exp->exp_obd;
2230         struct lmv_obd *lmv = &obd->u.lmv;
2231         struct lmv_stripe_md *lsm = op_data->op_mea1;
2232         struct lmv_tgt_desc *parent_tgt;
2233         struct lmv_tgt_desc *sp_tgt;
2234         struct lmv_tgt_desc *tp_tgt = NULL;
2235         struct lmv_tgt_desc *child_tgt;
2236         struct lmv_tgt_desc *tgt;
2237         struct lu_fid target_fid = { 0 };
2238         int rc;
2239
2240         ENTRY;
2241
2242         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
2243
2244         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
2245                PFID(&op_data->op_fid1), (int)namelen, name);
2246
2247         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2248         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2249         op_data->op_cap = current_cap();
2250
2251         parent_tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2252         if (IS_ERR(parent_tgt))
2253                 RETURN(PTR_ERR(parent_tgt));
2254
2255         if (lmv_dir_striped(lsm)) {
2256                 const struct lmv_oinfo *oinfo;
2257
2258                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
2259                 if (IS_ERR(oinfo))
2260                         RETURN(PTR_ERR(oinfo));
2261
2262                 /* save source stripe FID in fid4 temporarily for ELC */
2263                 op_data->op_fid4 = oinfo->lmo_fid;
2264                 sp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2265                 if (!sp_tgt)
2266                         RETURN(-ENODEV);
2267
2268                 /*
2269                  * if parent is being migrated too, fill op_fid2 with target
2270                  * stripe fid, otherwise the target stripe is not created yet.
2271                  */
2272                 if (lmv_dir_layout_changing(lsm)) {
2273                         oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
2274                                                         true);
2275                         if (IS_ERR(oinfo))
2276                                 RETURN(PTR_ERR(oinfo));
2277
2278                         op_data->op_fid2 = oinfo->lmo_fid;
2279                         tp_tgt = lmv_tgt(lmv, oinfo->lmo_mds);
2280                         if (!tp_tgt)
2281                                 RETURN(-ENODEV);
2282
2283                         /* parent unchanged and update namespace only */
2284                         if (lu_fid_eq(&op_data->op_fid4, &op_data->op_fid2) &&
2285                             op_data->op_bias & MDS_MIGRATE_NSONLY)
2286                                 RETURN(-EALREADY);
2287                 }
2288         } else {
2289                 sp_tgt = parent_tgt;
2290         }
2291
2292         child_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2293         if (IS_ERR(child_tgt))
2294                 RETURN(PTR_ERR(child_tgt));
2295
2296         if (lmv_topdir_specific_migrate(op_data)) {
2297                 struct lmv_user_md *lum = op_data->op_data;
2298
2299                 op_data->op_mds = le32_to_cpu(lum->lum_stripe_offset);
2300         } else if (lmv_topdir_qos_migrate(op_data)) {
2301                 tgt = lmv_locate_tgt_lf(lmv);
2302                 if (tgt == ERR_PTR(-EAGAIN))
2303                         tgt = lmv_locate_tgt_rr(lmv);
2304                 if (IS_ERR(tgt))
2305                         RETURN(PTR_ERR(tgt));
2306
2307                 op_data->op_mds = tgt->ltd_index;
2308         } else if (lmv_subdir_specific_migrate(op_data)) {
2309                 struct lmv_user_md *lum = op_data->op_data;
2310                 __u32 i;
2311
2312                 LASSERT(tp_tgt);
2313                 if (le32_to_cpu(lum->lum_magic) == LMV_USER_MAGIC_SPECIFIC) {
2314                         /* adjust MDTs in lum, since subdir is located on where
2315                          * its parent stripe is, not the first specified MDT.
2316                          */
2317                         for (i = 0; i < le32_to_cpu(lum->lum_stripe_count);
2318                              i++) {
2319                                 if (le32_to_cpu(lum->lum_objects[i].lum_mds) ==
2320                                     tp_tgt->ltd_index)
2321                                         break;
2322                         }
2323
2324                         if (i == le32_to_cpu(lum->lum_stripe_count))
2325                                 RETURN(-ENODEV);
2326
2327                         lum->lum_objects[i].lum_mds =
2328                                 lum->lum_objects[0].lum_mds;
2329                         lum->lum_objects[0].lum_mds =
2330                                 cpu_to_le32(tp_tgt->ltd_index);
2331                 }
2332                 /* NB, the above adjusts subdir migration for command like
2333                  * "lfs migrate -m 0,1,2 ...", but for migration like
2334                  * "lfs migrate -m 0 -c 2 ...", the top dir is migrated to MDT0
2335                  * and MDT1, however its subdir may be migrated to MDT1 and MDT2
2336                  */
2337
2338                 lum->lum_stripe_offset = cpu_to_le32(tp_tgt->ltd_index);
2339                 op_data->op_mds = tp_tgt->ltd_index;
2340         } else if (tp_tgt) {
2341                 op_data->op_mds = tp_tgt->ltd_index;
2342         } else {
2343                 op_data->op_mds = sp_tgt->ltd_index;
2344         }
2345
2346         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2347         if (rc)
2348                 RETURN(rc);
2349
2350         /*
2351          * for directory, send migrate request to the MDT where the object will
2352          * be migrated to, because we can't create a striped directory remotely.
2353          *
2354          * otherwise, send to the MDT where source is located because regular
2355          * file may open lease.
2356          *
2357          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2358          * backward compatibility.
2359          */
2360         if (S_ISDIR(op_data->op_mode) &&
2361             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2362                 tgt = lmv_fid2tgt(lmv, &target_fid);
2363                 if (IS_ERR(tgt))
2364                         RETURN(PTR_ERR(tgt));
2365         } else {
2366                 tgt = child_tgt;
2367         }
2368
2369         /* cancel UPDATE lock of parent master object */
2370         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index, LCK_EX,
2371                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2372         if (rc)
2373                 RETURN(rc);
2374
2375         /* cancel UPDATE lock of source parent */
2376         if (sp_tgt != parent_tgt) {
2377                 /*
2378                  * migrate RPC packs master object FID, because we can only pack
2379                  * two FIDs in reint RPC, but MDS needs to know both source
2380                  * parent and target parent, and it will obtain them from master
2381                  * FID and LMV, the other FID in RPC is kept for target.
2382                  *
2383                  * since this FID is not passed to MDC, cancel it anyway.
2384                  */
2385                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2386                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2387                 if (rc)
2388                         RETURN(rc);
2389
2390                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2391         }
2392         op_data->op_fid4 = target_fid;
2393
2394         /* cancel UPDATE locks of target parent */
2395         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2396                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2397         if (rc)
2398                 RETURN(rc);
2399
2400         /* cancel LOOKUP lock of source if source is remote object */
2401         if (child_tgt != sp_tgt) {
2402                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index,
2403                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2404                                       MF_MDC_CANCEL_FID3);
2405                 if (rc)
2406                         RETURN(rc);
2407         }
2408
2409         /* cancel ELC locks of source */
2410         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_index, LCK_EX,
2411                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2412         if (rc)
2413                 RETURN(rc);
2414
2415         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2416
2417         RETURN(rc);
2418 }
2419
2420 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2421                       const char *old, size_t oldlen,
2422                       const char *new, size_t newlen,
2423                       struct ptlrpc_request **request)
2424 {
2425         struct obd_device *obd = exp->exp_obd;
2426         struct lmv_obd *lmv = &obd->u.lmv;
2427         struct lmv_tgt_desc *sp_tgt;
2428         struct lmv_tgt_desc *tp_tgt = NULL;
2429         struct lmv_tgt_desc *src_tgt = NULL;
2430         struct lmv_tgt_desc *tgt;
2431         struct mdt_body *body;
2432         int rc;
2433
2434         ENTRY;
2435
2436         LASSERT(oldlen != 0);
2437
2438         if (op_data->op_cli_flags & CLI_MIGRATE) {
2439                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2440                 RETURN(rc);
2441         }
2442
2443         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2444         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2445         op_data->op_cap = current_cap();
2446
2447         op_data->op_name = new;
2448         op_data->op_namelen = newlen;
2449
2450         tp_tgt = lmv_locate_tgt2(lmv, op_data);
2451         if (IS_ERR(tp_tgt))
2452                 RETURN(PTR_ERR(tp_tgt));
2453
2454         /* Since the target child might be destroyed, and it might become
2455          * orphan, and we can only check orphan on the local MDT right now, so
2456          * we send rename request to the MDT where target child is located. If
2457          * target child does not exist, then it will send the request to the
2458          * target parent */
2459         if (fid_is_sane(&op_data->op_fid4)) {
2460                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2461                 if (IS_ERR(tgt))
2462                         RETURN(PTR_ERR(tgt));
2463         } else {
2464                 tgt = tp_tgt;
2465         }
2466
2467         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2468
2469         /* cancel UPDATE locks of target parent */
2470         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_index, LCK_EX,
2471                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2472         if (rc != 0)
2473                 RETURN(rc);
2474
2475         if (fid_is_sane(&op_data->op_fid4)) {
2476                 /* cancel LOOKUP lock of target on target parent */
2477                 if (tgt != tp_tgt) {
2478                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2479                                               tgt->ltd_index, LCK_EX,
2480                                               MDS_INODELOCK_LOOKUP,
2481                                               MF_MDC_CANCEL_FID4);
2482                         if (rc != 0)
2483                                 RETURN(rc);
2484                 }
2485         }
2486
2487         if (fid_is_sane(&op_data->op_fid3)) {
2488                 src_tgt = lmv_fid2tgt(lmv, &op_data->op_fid3);
2489                 if (IS_ERR(src_tgt))
2490                         RETURN(PTR_ERR(src_tgt));
2491
2492                 /* cancel ELC locks of source */
2493                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_index,
2494                                       LCK_EX, MDS_INODELOCK_ELC,
2495                                       MF_MDC_CANCEL_FID3);
2496                 if (rc != 0)
2497                         RETURN(rc);
2498         }
2499
2500         op_data->op_name = old;
2501         op_data->op_namelen = oldlen;
2502 retry:
2503         sp_tgt = lmv_locate_tgt(lmv, op_data);
2504         if (IS_ERR(sp_tgt))
2505                 RETURN(PTR_ERR(sp_tgt));
2506
2507         /* cancel UPDATE locks of source parent */
2508         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_index, LCK_EX,
2509                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2510         if (rc != 0)
2511                 RETURN(rc);
2512
2513         if (fid_is_sane(&op_data->op_fid3)) {
2514                 /* cancel LOOKUP lock of source on source parent */
2515                 if (src_tgt != sp_tgt) {
2516                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2517                                               tgt->ltd_index, LCK_EX,
2518                                               MDS_INODELOCK_LOOKUP,
2519                                               MF_MDC_CANCEL_FID3);
2520                         if (rc != 0)
2521                                 RETURN(rc);
2522                 }
2523         }
2524
2525 rename:
2526         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2527                 PFID(&op_data->op_fid1), (int)oldlen, old,
2528                 PFID(&op_data->op_fid2), (int)newlen, new);
2529
2530         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2531                         request);
2532         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2533                 ptlrpc_req_finished(*request);
2534                 *request = NULL;
2535                 goto retry;
2536         }
2537
2538         if (rc && rc != -EXDEV)
2539                 RETURN(rc);
2540
2541         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2542         if (body == NULL)
2543                 RETURN(-EPROTO);
2544
2545         /* Not cross-ref case, just get out of here. */
2546         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2547                 RETURN(rc);
2548
2549         op_data->op_fid4 = body->mbo_fid1;
2550
2551         ptlrpc_req_finished(*request);
2552         *request = NULL;
2553
2554         tgt = lmv_fid2tgt(lmv, &op_data->op_fid4);
2555         if (IS_ERR(tgt))
2556                 RETURN(PTR_ERR(tgt));
2557
2558         if (fid_is_sane(&op_data->op_fid4)) {
2559                 /* cancel LOOKUP lock of target on target parent */
2560                 if (tgt != tp_tgt) {
2561                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2562                                               tgt->ltd_index, LCK_EX,
2563                                               MDS_INODELOCK_LOOKUP,
2564                                               MF_MDC_CANCEL_FID4);
2565                         if (rc != 0)
2566                                 RETURN(rc);
2567                 }
2568         }
2569
2570         goto rename;
2571 }
2572
2573 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2574                        void *ea, size_t ealen, struct ptlrpc_request **request)
2575 {
2576         struct obd_device *obd = exp->exp_obd;
2577         struct lmv_obd *lmv = &obd->u.lmv;
2578         struct lmv_tgt_desc *tgt;
2579         int rc = 0;
2580
2581         ENTRY;
2582
2583         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2584                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2585                op_data->op_xvalid);
2586
2587         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2588         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2589         if (IS_ERR(tgt))
2590                 RETURN(PTR_ERR(tgt));
2591
2592         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2593
2594         RETURN(rc);
2595 }
2596
2597 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2598                      struct ptlrpc_request **request)
2599 {
2600         struct obd_device *obd = exp->exp_obd;
2601         struct lmv_obd *lmv = &obd->u.lmv;
2602         struct lmv_tgt_desc *tgt;
2603         int rc;
2604
2605         ENTRY;
2606
2607         tgt = lmv_fid2tgt(lmv, fid);
2608         if (IS_ERR(tgt))
2609                 RETURN(PTR_ERR(tgt));
2610
2611         rc = md_fsync(tgt->ltd_exp, fid, request);
2612         RETURN(rc);
2613 }
2614
2615 struct stripe_dirent {
2616         struct page             *sd_page;
2617         struct lu_dirpage       *sd_dp;
2618         struct lu_dirent        *sd_ent;
2619         bool                     sd_eof;
2620 };
2621
2622 struct lmv_dir_ctxt {
2623         struct lmv_obd          *ldc_lmv;
2624         struct md_op_data       *ldc_op_data;
2625         struct md_callback      *ldc_cb_op;
2626         __u64                    ldc_hash;
2627         int                      ldc_count;
2628         struct stripe_dirent     ldc_stripes[0];
2629 };
2630
2631 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2632 {
2633         if (stripe->sd_page) {
2634                 kunmap(stripe->sd_page);
2635                 put_page(stripe->sd_page);
2636                 stripe->sd_page = NULL;
2637                 stripe->sd_ent = NULL;
2638         }
2639 }
2640
2641 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2642 {
2643         int i;
2644
2645         for (i = 0; i < ctxt->ldc_count; i++)
2646                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2647 }
2648
2649 /* if @ent is dummy, or . .., get next */
2650 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2651                                            struct lu_dirent *ent,
2652                                            int stripe_index)
2653 {
2654         for (; ent; ent = lu_dirent_next(ent)) {
2655                 /* Skip dummy entry */
2656                 if (le16_to_cpu(ent->lde_namelen) == 0)
2657                         continue;
2658
2659                 /* skip . and .. for other stripes */
2660                 if (stripe_index &&
2661                     (strncmp(ent->lde_name, ".",
2662                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2663                      strncmp(ent->lde_name, "..",
2664                              le16_to_cpu(ent->lde_namelen)) == 0))
2665                         continue;
2666
2667                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2668                         break;
2669         }
2670
2671         return ent;
2672 }
2673
2674 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2675                                             struct stripe_dirent *stripe,
2676                                             int stripe_index)
2677 {
2678         struct md_op_data *op_data = ctxt->ldc_op_data;
2679         struct lmv_oinfo *oinfo;
2680         struct lu_fid fid = op_data->op_fid1;
2681         struct inode *inode = op_data->op_data;
2682         struct lmv_tgt_desc *tgt;
2683         struct lu_dirent *ent = stripe->sd_ent;
2684         __u64 hash = ctxt->ldc_hash;
2685         int rc = 0;
2686
2687         ENTRY;
2688
2689         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2690         LASSERT(!ent);
2691
2692         do {
2693                 if (stripe->sd_page) {
2694                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2695
2696                         /* @hash should be the last dirent hash */
2697                         LASSERTF(hash <= end,
2698                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2699                                  ctxt, stripe, hash, end);
2700                         /* unload last page */
2701                         stripe_dirent_unload(stripe);
2702                         /* eof */
2703                         if (end == MDS_DIR_END_OFF) {
2704                                 stripe->sd_eof = true;
2705                                 break;
2706                         }
2707                         hash = end;
2708                 }
2709
2710                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2711                 if (!oinfo->lmo_root) {
2712                         rc = -ENOENT;
2713                         break;
2714                 }
2715
2716                 tgt = lmv_tgt(ctxt->ldc_lmv, oinfo->lmo_mds);
2717                 if (!tgt) {
2718                         rc = -ENODEV;
2719                         break;
2720                 }
2721
2722                 /* op_data is shared by stripes, reset after use */
2723                 op_data->op_fid1 = oinfo->lmo_fid;
2724                 op_data->op_fid2 = oinfo->lmo_fid;
2725                 op_data->op_data = oinfo->lmo_root;
2726
2727                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2728                                   &stripe->sd_page);
2729
2730                 op_data->op_fid1 = fid;
2731                 op_data->op_fid2 = fid;
2732                 op_data->op_data = inode;
2733
2734                 if (rc)
2735                         break;
2736
2737                 stripe->sd_dp = page_address(stripe->sd_page);
2738                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2739                                         stripe_index);
2740                 /* in case a page filled with ., .. and dummy, read next */
2741         } while (!ent);
2742
2743         stripe->sd_ent = ent;
2744         if (rc) {
2745                 LASSERT(!ent);
2746                 /* treat error as eof, so dir can be partially accessed */
2747                 stripe->sd_eof = true;
2748                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2749                               "directory is partially accessed!\n",
2750                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2751                               rc);
2752         }
2753
2754         RETURN(ent);
2755 }
2756
2757 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2758 {
2759         struct obd_device *obd = exp->exp_obd;
2760         struct lmv_obd *lmv = &obd->u.lmv;
2761         struct lmv_tgt_desc *tgt;
2762         int rc;
2763
2764         ENTRY;
2765
2766         rc = lmv_check_connect(obd);
2767         if (rc != 0)
2768                 RETURN(rc);
2769
2770         tgt = lmv_fid2tgt(lmv, &data->op_fid1);
2771         if (IS_ERR(tgt))
2772                 RETURN(PTR_ERR(tgt));
2773
2774         data->op_flags |= MF_MDC_CANCEL_FID1;
2775         rc = md_file_resync(tgt->ltd_exp, data);
2776         RETURN(rc);
2777 }
2778
2779 /**
2780  * Get dirent with the closest hash for striped directory
2781  *
2782  * This function will search the dir entry, whose hash value is the
2783  * closest(>=) to hash from all of sub-stripes, and it is only being called
2784  * for striped directory.
2785  *
2786  * \param[in] ctxt              dir read context
2787  *
2788  * \retval                      dirent get the entry successfully
2789  *                              NULL does not get the entry, normally it means
2790  *                              it reaches the end of the directory, while read
2791  *                              stripe dirent error is ignored to allow partial
2792  *                              access.
2793  */
2794 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2795 {
2796         struct stripe_dirent *stripe;
2797         struct lu_dirent *ent = NULL;
2798         int i;
2799         int min = -1;
2800
2801         /* TODO: optimize with k-way merge sort */
2802         for (i = 0; i < ctxt->ldc_count; i++) {
2803                 stripe = &ctxt->ldc_stripes[i];
2804                 if (stripe->sd_eof)
2805                         continue;
2806
2807                 if (!stripe->sd_ent) {
2808                         stripe_dirent_load(ctxt, stripe, i);
2809                         if (!stripe->sd_ent) {
2810                                 LASSERT(stripe->sd_eof);
2811                                 continue;
2812                         }
2813                 }
2814
2815                 if (min == -1 ||
2816                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2817                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2818                         min = i;
2819                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2820                             ctxt->ldc_hash)
2821                                 break;
2822                 }
2823         }
2824
2825         if (min != -1) {
2826                 stripe = &ctxt->ldc_stripes[min];
2827                 ent = stripe->sd_ent;
2828                 /* pop found dirent */
2829                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2830                                                    min);
2831         }
2832
2833         return ent;
2834 }
2835
2836 /**
2837  * Build dir entry page for striped directory
2838  *
2839  * This function gets one entry by @offset from a striped directory. It will
2840  * read entries from all of stripes, and choose one closest to the required
2841  * offset(&offset). A few notes
2842  * 1. skip . and .. for non-zero stripes, because there can only have one .
2843  * and .. in a directory.
2844  * 2. op_data will be shared by all of stripes, instead of allocating new
2845  * one, so need to restore before reusing.
2846  *
2847  * \param[in] exp       obd export refer to LMV
2848  * \param[in] op_data   hold those MD parameters of read_entry
2849  * \param[in] cb_op     ldlm callback being used in enqueue in mdc_read_entry
2850  * \param[in] offset    starting hash offset
2851  * \param[out] ppage    the page holding the entry. Note: because the entry
2852  *                      will be accessed in upper layer, so we need hold the
2853  *                      page until the usages of entry is finished, see
2854  *                      ll_dir_entry_next.
2855  *
2856  * retval               =0 if get entry successfully
2857  *                      <0 cannot get entry
2858  */
2859 static int lmv_striped_read_page(struct obd_export *exp,
2860                                  struct md_op_data *op_data,
2861                                  struct md_callback *cb_op,
2862                                  __u64 offset, struct page **ppage)
2863 {
2864         struct page *page = NULL;
2865         struct lu_dirpage *dp;
2866         void *start;
2867         struct lu_dirent *ent;
2868         struct lu_dirent *last_ent;
2869         int stripe_count;
2870         struct lmv_dir_ctxt *ctxt;
2871         struct lu_dirent *next = NULL;
2872         __u16 ent_size;
2873         size_t left_bytes;
2874         int rc = 0;
2875         ENTRY;
2876
2877         /* Allocate a page and read entries from all of stripes and fill
2878          * the page by hash order */
2879         page = alloc_page(GFP_KERNEL);
2880         if (!page)
2881                 RETURN(-ENOMEM);
2882
2883         /* Initialize the entry page */
2884         dp = kmap(page);
2885         memset(dp, 0, sizeof(*dp));
2886         dp->ldp_hash_start = cpu_to_le64(offset);
2887
2888         start = dp + 1;
2889         left_bytes = PAGE_SIZE - sizeof(*dp);
2890         ent = start;
2891         last_ent = ent;
2892
2893         /* initalize dir read context */
2894         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2895         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2896         if (!ctxt)
2897                 GOTO(free_page, rc = -ENOMEM);
2898         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2899         ctxt->ldc_op_data = op_data;
2900         ctxt->ldc_cb_op = cb_op;
2901         ctxt->ldc_hash = offset;
2902         ctxt->ldc_count = stripe_count;
2903
2904         while (1) {
2905                 next = lmv_dirent_next(ctxt);
2906
2907                 /* end of directory */
2908                 if (!next) {
2909                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2910                         break;
2911                 }
2912                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2913
2914                 ent_size = le16_to_cpu(next->lde_reclen);
2915
2916                 /* the last entry lde_reclen is 0, but it might not be the last
2917                  * one of this temporay dir page */
2918                 if (!ent_size)
2919                         ent_size = lu_dirent_calc_size(
2920                                         le16_to_cpu(next->lde_namelen),
2921                                         le32_to_cpu(next->lde_attrs));
2922                 /* page full */
2923                 if (ent_size > left_bytes)
2924                         break;
2925
2926                 memcpy(ent, next, ent_size);
2927
2928                 /* Replace . with master FID and Replace .. with the parent FID
2929                  * of master object */
2930                 if (strncmp(ent->lde_name, ".",
2931                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2932                     le16_to_cpu(ent->lde_namelen) == 1)
2933                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2934                 else if (strncmp(ent->lde_name, "..",
2935                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2936                            le16_to_cpu(ent->lde_namelen) == 2)
2937                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2938
2939                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2940                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2941                        le64_to_cpu(ent->lde_hash));
2942
2943                 left_bytes -= ent_size;
2944                 ent->lde_reclen = cpu_to_le16(ent_size);
2945                 last_ent = ent;
2946                 ent = (void *)ent + ent_size;
2947         };
2948
2949         last_ent->lde_reclen = 0;
2950
2951         if (ent == start)
2952                 dp->ldp_flags |= LDF_EMPTY;
2953         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2954                 dp->ldp_flags |= LDF_COLLIDE;
2955         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2956         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2957
2958         put_lmv_dir_ctxt(ctxt);
2959         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2960
2961         *ppage = page;
2962
2963         RETURN(0);
2964
2965 free_page:
2966         kunmap(page);
2967         __free_page(page);
2968
2969         return rc;
2970 }
2971
2972 static int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2973                          struct md_callback *cb_op, __u64 offset,
2974                          struct page **ppage)
2975 {
2976         struct obd_device *obd = exp->exp_obd;
2977         struct lmv_obd *lmv = &obd->u.lmv;
2978         struct lmv_tgt_desc *tgt;
2979         int rc;
2980
2981         ENTRY;
2982
2983         if (unlikely(lmv_dir_foreign(op_data->op_mea1)))
2984                 RETURN(-ENODATA);
2985
2986         if (unlikely(lmv_dir_striped(op_data->op_mea1))) {
2987                 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2988                 RETURN(rc);
2989         }
2990
2991         tgt = lmv_fid2tgt(lmv, &op_data->op_fid1);
2992         if (IS_ERR(tgt))
2993                 RETURN(PTR_ERR(tgt));
2994
2995         rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2996
2997         RETURN(rc);
2998 }
2999
3000 /**
3001  * Unlink a file/directory
3002  *
3003  * Unlink a file or directory under the parent dir. The unlink request
3004  * usually will be sent to the MDT where the child is located, but if
3005  * the client does not have the child FID then request will be sent to the
3006  * MDT where the parent is located.
3007  *
3008  * If the parent is a striped directory then it also needs to locate which
3009  * stripe the name of the child is located, and replace the parent FID
3010  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
3011  * it will walk through all of sub-stripes until the child is being
3012  * unlinked finally.
3013  *
3014  * \param[in] exp       export refer to LMV
3015  * \param[in] op_data   different parameters transferred beween client
3016  *                      MD stacks, name, namelen, FIDs etc.
3017  *                      op_fid1 is the parent FID, op_fid2 is the child
3018  *                      FID.
3019  * \param[out] request  point to the request of unlink.
3020  *
3021  * retval               0 if succeed
3022  *                      negative errno if failed.
3023  */
3024 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
3025                       struct ptlrpc_request **request)
3026 {
3027         struct obd_device *obd = exp->exp_obd;
3028         struct lmv_obd *lmv = &obd->u.lmv;
3029         struct lmv_tgt_desc *tgt;
3030         struct lmv_tgt_desc *parent_tgt;
3031         struct mdt_body *body;
3032         int rc;
3033
3034         ENTRY;
3035
3036         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
3037         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
3038         op_data->op_cap = current_cap();
3039
3040 retry:
3041         parent_tgt = lmv_locate_tgt(lmv, op_data);
3042         if (IS_ERR(parent_tgt))
3043                 RETURN(PTR_ERR(parent_tgt));
3044
3045         if (likely(!fid_is_zero(&op_data->op_fid2))) {
3046                 tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3047                 if (IS_ERR(tgt))
3048                         RETURN(PTR_ERR(tgt));
3049         } else {
3050                 tgt = parent_tgt;
3051         }
3052
3053         /*
3054          * If child's fid is given, cancel unused locks for it if it is from
3055          * another export than parent.
3056          *
3057          * LOOKUP lock for child (fid3) should also be cancelled on parent
3058          * tgt_tgt in mdc_unlink().
3059          */
3060         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
3061
3062         if (parent_tgt != tgt)
3063                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_index,
3064                                       LCK_EX, MDS_INODELOCK_LOOKUP,
3065                                       MF_MDC_CANCEL_FID3);
3066
3067         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_index, LCK_EX,
3068                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
3069         if (rc)
3070                 RETURN(rc);
3071
3072         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
3073                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2),
3074                tgt->ltd_index);
3075
3076         rc = md_unlink(tgt->ltd_exp, op_data, request);
3077         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
3078                 ptlrpc_req_finished(*request);
3079                 *request = NULL;
3080                 goto retry;
3081         }
3082
3083         if (rc != -EREMOTE)
3084                 RETURN(rc);
3085
3086         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
3087         if (body == NULL)
3088                 RETURN(-EPROTO);
3089
3090         /* Not cross-ref case, just get out of here. */
3091         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
3092                 RETURN(rc);
3093
3094         /* This is a remote object, try remote MDT. */
3095         op_data->op_fid2 = body->mbo_fid1;
3096         ptlrpc_req_finished(*request);
3097         *request = NULL;
3098
3099         tgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3100         if (IS_ERR(tgt))
3101                 RETURN(PTR_ERR(tgt));
3102
3103         goto retry;
3104 }
3105
3106 static int lmv_precleanup(struct obd_device *obd)
3107 {
3108         ENTRY;
3109         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
3110         fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
3111         lprocfs_obd_cleanup(obd);
3112         lprocfs_free_md_stats(obd);
3113         RETURN(0);
3114 }
3115
3116 /**
3117  * Get by key a value associated with a LMV device.
3118  *
3119  * Dispatch request to lower-layer devices as needed.
3120  *
3121  * \param[in] env               execution environment for this thread
3122  * \param[in] exp               export for the LMV device
3123  * \param[in] keylen            length of key identifier
3124  * \param[in] key               identifier of key to get value for
3125  * \param[in] vallen            size of \a val
3126  * \param[out] val              pointer to storage location for value
3127  * \param[in] lsm               optional striping metadata of object
3128  *
3129  * \retval 0            on success
3130  * \retval negative     negated errno on failure
3131  */
3132 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
3133                         __u32 keylen, void *key, __u32 *vallen, void *val)
3134 {
3135         struct obd_device *obd;
3136         struct lmv_obd *lmv;
3137         struct lu_tgt_desc *tgt;
3138         int rc = 0;
3139
3140         ENTRY;
3141
3142         obd = class_exp2obd(exp);
3143         if (obd == NULL) {
3144                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3145                        exp->exp_handle.h_cookie);
3146                 RETURN(-EINVAL);
3147         }
3148
3149         lmv = &obd->u.lmv;
3150         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
3151                 LASSERT(*vallen == sizeof(__u32));
3152                 lmv_foreach_connected_tgt(lmv, tgt) {
3153                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
3154                                           vallen, val))
3155                                 RETURN(0);
3156                 }
3157                 RETURN(-EINVAL);
3158         } else if (KEY_IS(KEY_MAX_EASIZE) ||
3159                    KEY_IS(KEY_DEFAULT_EASIZE) ||
3160                    KEY_IS(KEY_CONN_DATA)) {
3161                 /*
3162                  * Forwarding this request to first MDS, it should know LOV
3163                  * desc.
3164                  */
3165                 tgt = lmv_tgt(lmv, 0);
3166                 if (!tgt)
3167                         RETURN(-ENODEV);
3168
3169                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key, vallen, val);
3170                 if (!rc && KEY_IS(KEY_CONN_DATA))
3171                         exp->exp_connect_data = *(struct obd_connect_data *)val;
3172                 RETURN(rc);
3173         } else if (KEY_IS(KEY_TGT_COUNT)) {
3174                 *((int *)val) = lmv->lmv_mdt_descs.ltd_lmv_desc.ld_tgt_count;
3175                 RETURN(0);
3176         }
3177
3178         CDEBUG(D_IOCTL, "Invalid key\n");
3179         RETURN(-EINVAL);
3180 }
3181
3182 static int lmv_rmfid(struct obd_export *exp, struct fid_array *fa,
3183                      int *__rcs, struct ptlrpc_request_set *_set)
3184 {
3185         struct obd_device *obd = class_exp2obd(exp);
3186         struct ptlrpc_request_set *set = _set;
3187         struct lmv_obd *lmv = &obd->u.lmv;
3188         int tgt_count = lmv->lmv_mdt_count;
3189         struct lu_tgt_desc *tgt;
3190         struct fid_array *fat, **fas = NULL;
3191         int i, rc, **rcs = NULL;
3192
3193         if (!set) {
3194                 set = ptlrpc_prep_set();
3195                 if (!set)
3196                         RETURN(-ENOMEM);
3197         }
3198
3199         /* split FIDs by targets */
3200         OBD_ALLOC_PTR_ARRAY(fas, tgt_count);
3201         if (fas == NULL)
3202                 GOTO(out, rc = -ENOMEM);
3203         OBD_ALLOC_PTR_ARRAY(rcs, tgt_count);
3204         if (rcs == NULL)
3205                 GOTO(out_fas, rc = -ENOMEM);
3206
3207         for (i = 0; i < fa->fa_nr; i++) {
3208                 unsigned int idx;
3209
3210                 rc = lmv_fld_lookup(lmv, &fa->fa_fids[i], &idx);
3211                 if (rc) {
3212                         CDEBUG(D_OTHER, "can't lookup "DFID": rc = %d\n",
3213                                PFID(&fa->fa_fids[i]), rc);
3214                         continue;
3215                 }
3216                 LASSERT(idx < tgt_count);
3217                 if (!fas[idx])
3218                         OBD_ALLOC(fas[idx], offsetof(struct fid_array,
3219                                   fa_fids[fa->fa_nr]));
3220                 if (!fas[idx])
3221                         GOTO(out, rc = -ENOMEM);
3222                 if (!rcs[idx])
3223                         OBD_ALLOC_PTR_ARRAY(rcs[idx], fa->fa_nr);
3224                 if (!rcs[idx])
3225                         GOTO(out, rc = -ENOMEM);
3226
3227                 fat = fas[idx];
3228                 fat->fa_fids[fat->fa_nr++] = fa->fa_fids[i];
3229         }
3230
3231         lmv_foreach_connected_tgt(lmv, tgt) {
3232                 fat = fas[tgt->ltd_index];
3233                 if (!fat || fat->fa_nr == 0)
3234                         continue;
3235                 rc = md_rmfid(tgt->ltd_exp, fat, rcs[tgt->ltd_index], set);
3236         }
3237
3238         rc = ptlrpc_set_wait(NULL, set);
3239         if (rc == 0) {
3240                 int j = 0;
3241                 for (i = 0; i < tgt_count; i++) {
3242                         fat = fas[i];
3243                         if (!fat || fat->fa_nr == 0)
3244                                 continue;
3245                         /* copy FIDs back */
3246                         memcpy(fa->fa_fids + j, fat->fa_fids,
3247                                fat->fa_nr * sizeof(struct lu_fid));
3248                         /* copy rcs back */
3249                         memcpy(__rcs + j, rcs[i], fat->fa_nr * sizeof(**rcs));
3250                         j += fat->fa_nr;
3251                 }
3252         }
3253         if (set != _set)
3254                 ptlrpc_set_destroy(set);
3255
3256 out:
3257         for (i = 0; i < tgt_count; i++) {
3258                 if (fas && fas[i])
3259                         OBD_FREE(fas[i], offsetof(struct fid_array,
3260                                                 fa_fids[fa->fa_nr]));
3261                 if (rcs && rcs[i])
3262                         OBD_FREE_PTR_ARRAY(rcs[i], fa->fa_nr);
3263         }
3264         if (rcs)
3265                 OBD_FREE_PTR_ARRAY(rcs, tgt_count);
3266 out_fas:
3267         if (fas)
3268                 OBD_FREE_PTR_ARRAY(fas, tgt_count);
3269
3270         RETURN(rc);
3271 }
3272
3273 /**
3274  * Asynchronously set by key a value associated with a LMV device.
3275  *
3276  * Dispatch request to lower-layer devices as needed.
3277  *
3278  * \param[in] env       execution environment for this thread
3279  * \param[in] exp       export for the LMV device
3280  * \param[in] keylen    length of key identifier
3281  * \param[in] key       identifier of key to store value for
3282  * \param[in] vallen    size of value to store
3283  * \param[in] val       pointer to data to be stored
3284  * \param[in] set       optional list of related ptlrpc requests
3285  *
3286  * \retval 0            on success
3287  * \retval negative     negated errno on failure
3288  */
3289 static int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
3290                               __u32 keylen, void *key, __u32 vallen, void *val,
3291                               struct ptlrpc_request_set *set)
3292 {
3293         struct lmv_tgt_desc *tgt;
3294         struct obd_device *obd;
3295         struct lmv_obd *lmv;
3296         int rc = 0;
3297         ENTRY;
3298
3299         obd = class_exp2obd(exp);
3300         if (obd == NULL) {
3301                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
3302                        exp->exp_handle.h_cookie);
3303                 RETURN(-EINVAL);
3304         }
3305         lmv = &obd->u.lmv;
3306
3307         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
3308             KEY_IS(KEY_DEFAULT_EASIZE)) {
3309                 int err = 0;
3310
3311                 lmv_foreach_connected_tgt(lmv, tgt) {
3312                         err = obd_set_info_async(env, tgt->ltd_exp,
3313                                                  keylen, key, vallen, val, set);
3314                         if (err && rc == 0)
3315                                 rc = err;
3316                 }
3317
3318                 RETURN(rc);
3319         }
3320
3321         RETURN(-EINVAL);
3322 }
3323
3324 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
3325                             const struct lmv_mds_md_v1 *lmm1)
3326 {
3327         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
3328         int             stripe_count;
3329         int             cplen;
3330         int             i;
3331         int             rc = 0;
3332         ENTRY;
3333
3334         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
3335         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3336         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
3337         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
3338                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
3339         else
3340                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
3341         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
3342         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
3343         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
3344         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
3345                         sizeof(lsm->lsm_md_pool_name));
3346
3347         if (cplen >= sizeof(lsm->lsm_md_pool_name))
3348                 RETURN(-E2BIG);
3349
3350         CDEBUG(D_INFO, "unpack lsm count %d/%d, master %d hash_type %#x/%#x "
3351                "layout_version %d\n", lsm->lsm_md_stripe_count,
3352                lsm->lsm_md_migrate_offset, lsm->lsm_md_master_mdt_index,
3353                lsm->lsm_md_hash_type, lsm->lsm_md_migrate_hash,
3354                lsm->lsm_md_layout_version);
3355
3356         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3357         for (i = 0; i < stripe_count; i++) {
3358                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3359                               &lmm1->lmv_stripe_fids[i]);
3360                 /*
3361                  * set default value -1, so lmv_locate_tgt() knows this stripe
3362                  * target is not initialized.
3363                  */
3364                 lsm->lsm_md_oinfo[i].lmo_mds = LMV_OFFSET_DEFAULT;
3365                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3366                         continue;
3367
3368                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3369                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3370                 if (rc == -ENOENT)
3371                         continue;
3372
3373                 if (rc)
3374                         RETURN(rc);
3375
3376                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3377                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3378         }
3379
3380         RETURN(rc);
3381 }
3382
3383 static inline int lmv_unpack_user_md(struct obd_export *exp,
3384                                      struct lmv_stripe_md *lsm,
3385                                      const struct lmv_user_md *lmu)
3386 {
3387         lsm->lsm_md_magic = le32_to_cpu(lmu->lum_magic);
3388         lsm->lsm_md_stripe_count = le32_to_cpu(lmu->lum_stripe_count);
3389         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmu->lum_stripe_offset);
3390         lsm->lsm_md_hash_type = le32_to_cpu(lmu->lum_hash_type);
3391         lsm->lsm_md_max_inherit = lmu->lum_max_inherit;
3392         lsm->lsm_md_max_inherit_rr = lmu->lum_max_inherit_rr;
3393         lsm->lsm_md_pool_name[LOV_MAXPOOLNAME] = 0;
3394
3395         return 0;
3396 }
3397
3398 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3399                         const union lmv_mds_md *lmm, size_t lmm_size)
3400 {
3401         struct lmv_stripe_md     *lsm;
3402         int                      lsm_size;
3403         int                      rc;
3404         bool                     allocated = false;
3405         ENTRY;
3406
3407         LASSERT(lsmp != NULL);
3408
3409         lsm = *lsmp;
3410         /* Free memmd */
3411         if (lsm != NULL && lmm == NULL) {
3412                 int i;
3413                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3414
3415                 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3416                         size_t lfm_size;
3417
3418                         lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3419                                                               lfm_value[0]);
3420                         OBD_FREE_LARGE(lfm, lfm_size);
3421                         RETURN(0);
3422                 }
3423
3424                 if (lmv_dir_striped(lsm)) {
3425                         for (i = 0; i < lsm->lsm_md_stripe_count; i++)
3426                                 iput(lsm->lsm_md_oinfo[i].lmo_root);
3427                         lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3428                 } else {
3429                         lsm_size = lmv_stripe_md_size(0);
3430                 }
3431                 OBD_FREE(lsm, lsm_size);
3432                 *lsmp = NULL;
3433                 RETURN(0);
3434         }
3435
3436         /* foreign lmv case */
3437         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3438                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3439
3440                 if (lfm == NULL) {
3441                         OBD_ALLOC_LARGE(lfm, lmm_size);
3442                         if (lfm == NULL)
3443                                 RETURN(-ENOMEM);
3444                         *lsmp = (struct lmv_stripe_md *)lfm;
3445                 }
3446                 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3447                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3448                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3449                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3450                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3451                        lfm->lfm_length);
3452                 RETURN(lmm_size);
3453         }
3454
3455         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3456                 RETURN(-EPERM);
3457
3458         /* Unpack memmd */
3459         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3460             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3461                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3462                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3463                        -EIO);
3464                 RETURN(-EIO);
3465         }
3466
3467         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3468                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3469         else
3470                 /**
3471                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3472                  * stripecount should be 0 then.
3473                  */
3474                 lsm_size = lmv_stripe_md_size(0);
3475
3476         if (lsm == NULL) {
3477                 OBD_ALLOC(lsm, lsm_size);
3478                 if (lsm == NULL)
3479                         RETURN(-ENOMEM);
3480                 allocated = true;
3481                 *lsmp = lsm;
3482         }
3483
3484         switch (le32_to_cpu(lmm->lmv_magic)) {
3485         case LMV_MAGIC_V1:
3486                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3487                 break;
3488         case LMV_USER_MAGIC:
3489                 rc = lmv_unpack_user_md(exp, lsm, &lmm->lmv_user_md);
3490                 break;
3491         default:
3492                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3493                        le32_to_cpu(lmm->lmv_magic));
3494                 rc = -EINVAL;
3495                 break;
3496         }
3497
3498         if (rc != 0 && allocated) {
3499                 OBD_FREE(lsm, lsm_size);
3500                 *lsmp = NULL;
3501                 lsm_size = rc;
3502         }
3503         RETURN(lsm_size);
3504 }
3505
3506 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3507 {
3508         lmv_unpackmd(NULL, &lsm, NULL, 0);
3509 }
3510 EXPORT_SYMBOL(lmv_free_memmd);
3511
3512 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3513                              union ldlm_policy_data *policy,
3514                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3515                              void *opaque)
3516 {
3517         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3518         struct lu_tgt_desc *tgt;
3519         int err;
3520         int rc = 0;
3521
3522         ENTRY;
3523
3524         LASSERT(fid != NULL);
3525
3526         lmv_foreach_connected_tgt(lmv, tgt) {
3527                 if (!tgt->ltd_active)
3528                         continue;
3529
3530                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3531                                        opaque);
3532                 if (!rc)
3533                         rc = err;
3534         }
3535         RETURN(rc);
3536 }
3537
3538 static int lmv_set_lock_data(struct obd_export *exp,
3539                              const struct lustre_handle *lockh,
3540                              void *data, __u64 *bits)
3541 {
3542         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3543         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3544         int rc;
3545
3546         ENTRY;
3547
3548         if (tgt == NULL || tgt->ltd_exp == NULL)
3549                 RETURN(-EINVAL);
3550         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3551         RETURN(rc);
3552 }
3553
3554 static enum ldlm_mode
3555 lmv_lock_match(struct obd_export *exp, __u64 flags,
3556                const struct lu_fid *fid, enum ldlm_type type,
3557                union ldlm_policy_data *policy,
3558                enum ldlm_mode mode, struct lustre_handle *lockh)
3559 {
3560         struct obd_device *obd = exp->exp_obd;
3561         struct lmv_obd *lmv = &obd->u.lmv;
3562         enum ldlm_mode rc;
3563         struct lu_tgt_desc *tgt;
3564         int i;
3565         int index;
3566
3567         ENTRY;
3568
3569         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
3570
3571         /*
3572          * With DNE every object can have two locks in different namespaces:
3573          * lookup lock in space of MDT storing direntry and update/open lock in
3574          * space of MDT storing inode.  Try the MDT that the FID maps to first,
3575          * since this can be easily found, and only try others if that fails.
3576          */
3577         for (i = 0, index = lmv_fid2tgt_index(lmv, fid);
3578              i < lmv->lmv_mdt_descs.ltd_tgts_size;
3579              i++, index = (index + 1) % lmv->lmv_mdt_descs.ltd_tgts_size) {
3580                 if (index < 0) {
3581                         CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
3582                                obd->obd_name, PFID(fid), index);
3583                         index = 0;
3584                 }
3585
3586                 tgt = lmv_tgt(lmv, index);
3587                 if (!tgt || !tgt->ltd_exp || !tgt->ltd_active)
3588                         continue;
3589
3590                 rc = md_lock_match(tgt->ltd_exp, flags, fid, type, policy, mode,
3591                                    lockh);
3592                 if (rc)
3593                         RETURN(rc);
3594         }
3595
3596         RETURN(0);
3597 }
3598
3599 static int
3600 lmv_get_lustre_md(struct obd_export *exp, struct req_capsule *pill,
3601                   struct obd_export *dt_exp, struct obd_export *md_exp,
3602                   struct lustre_md *md)
3603 {
3604         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3605         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3606
3607         if (!tgt || !tgt->ltd_exp)
3608                 return -EINVAL;
3609
3610         return md_get_lustre_md(tgt->ltd_exp, pill, dt_exp, md_exp, md);
3611 }
3612
3613 static int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
3614 {
3615         struct obd_device *obd = exp->exp_obd;
3616         struct lmv_obd *lmv = &obd->u.lmv;
3617         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3618
3619         ENTRY;
3620
3621         if (md->default_lmv) {
3622                 lmv_free_memmd(md->default_lmv);
3623                 md->default_lmv = NULL;
3624         }
3625         if (md->lmv != NULL) {
3626                 lmv_free_memmd(md->lmv);
3627                 md->lmv = NULL;
3628         }
3629         if (!tgt || !tgt->ltd_exp)
3630                 RETURN(-EINVAL);
3631         RETURN(md_free_lustre_md(tgt->ltd_exp, md));
3632 }
3633
3634 static int lmv_set_open_replay_data(struct obd_export *exp,
3635                                     struct obd_client_handle *och,
3636                                     struct lookup_intent *it)
3637 {
3638         struct obd_device *obd = exp->exp_obd;
3639         struct lmv_obd *lmv = &obd->u.lmv;
3640         struct lmv_tgt_desc *tgt;
3641
3642         ENTRY;
3643
3644         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3645         if (IS_ERR(tgt))
3646                 RETURN(PTR_ERR(tgt));
3647
3648         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3649 }
3650
3651 static int lmv_clear_open_replay_data(struct obd_export *exp,
3652                                       struct obd_client_handle *och)
3653 {
3654         struct obd_device *obd = exp->exp_obd;
3655         struct lmv_obd *lmv = &obd->u.lmv;
3656         struct lmv_tgt_desc *tgt;
3657
3658         ENTRY;
3659
3660         tgt = lmv_fid2tgt(lmv, &och->och_fid);
3661         if (IS_ERR(tgt))
3662                 RETURN(PTR_ERR(tgt));
3663
3664         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3665 }
3666
3667 static int lmv_intent_getattr_async(struct obd_export *exp,
3668                                     struct md_enqueue_info *minfo)
3669 {
3670         struct md_op_data *op_data = &minfo->mi_data;
3671         struct obd_device *obd = exp->exp_obd;
3672         struct lmv_obd *lmv = &obd->u.lmv;
3673         struct lmv_tgt_desc *ptgt;
3674         struct lmv_tgt_desc *ctgt;
3675         int rc;
3676
3677         ENTRY;
3678
3679         if (!fid_is_sane(&op_data->op_fid2))
3680                 RETURN(-EINVAL);
3681
3682         ptgt = lmv_locate_tgt(lmv, op_data);
3683         if (IS_ERR(ptgt))
3684                 RETURN(PTR_ERR(ptgt));
3685
3686         ctgt = lmv_fid2tgt(lmv, &op_data->op_fid2);
3687         if (IS_ERR(ctgt))
3688                 RETURN(PTR_ERR(ctgt));
3689
3690         /*
3691          * remote object needs two RPCs to lookup and getattr, considering the
3692          * complexity don't support statahead for now.
3693          */
3694         if (ctgt != ptgt)
3695                 RETURN(-EREMOTE);
3696
3697         rc = md_intent_getattr_async(ptgt->ltd_exp, minfo);
3698
3699         RETURN(rc);
3700 }
3701
3702 static int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3703                                struct lu_fid *fid, __u64 *bits)
3704 {
3705         struct obd_device *obd = exp->exp_obd;
3706         struct lmv_obd *lmv = &obd->u.lmv;
3707         struct lmv_tgt_desc *tgt;
3708         int rc;
3709
3710         ENTRY;
3711
3712         tgt = lmv_fid2tgt(lmv, fid);
3713         if (IS_ERR(tgt))
3714                 RETURN(PTR_ERR(tgt));
3715
3716         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3717         RETURN(rc);
3718 }
3719
3720 static int lmv_get_fid_from_lsm(struct obd_export *exp,
3721                                 const struct lmv_stripe_md *lsm,
3722                                 const char *name, int namelen,
3723                                 struct lu_fid *fid)
3724 {
3725         const struct lmv_oinfo *oinfo;
3726
3727         LASSERT(lmv_dir_striped(lsm));
3728
3729         oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
3730         if (IS_ERR(oinfo))
3731                 return PTR_ERR(oinfo);
3732
3733         *fid = oinfo->lmo_fid;
3734
3735         RETURN(0);
3736 }
3737
3738 /**
3739  * For lmv, only need to send request to master MDT, and the master MDT will
3740  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3741  * we directly fetch data from the slave MDTs.
3742  */
3743 static int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3744                         struct obd_quotactl *oqctl)
3745 {
3746         struct obd_device *obd = class_exp2obd(exp);
3747         struct lmv_obd *lmv = &obd->u.lmv;
3748         struct lmv_tgt_desc *tgt = lmv_tgt(lmv, 0);
3749         __u64 curspace, curinodes;
3750         int rc = 0;
3751
3752         ENTRY;
3753
3754         if (!tgt || !tgt->ltd_exp || !tgt->ltd_active) {
3755                 CERROR("master lmv inactive\n");
3756                 RETURN(-EIO);
3757         }
3758
3759         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3760                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3761                 RETURN(rc);
3762         }
3763
3764         curspace = curinodes = 0;
3765         lmv_foreach_connected_tgt(lmv, tgt) {
3766                 int err;
3767
3768                 if (!tgt->ltd_active)
3769                         continue;
3770
3771                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3772                 if (err) {
3773                         CERROR("getquota on mdt %d failed. %d\n",
3774                                tgt->ltd_index, err);
3775                         if (!rc)
3776                                 rc = err;
3777                 } else {
3778                         curspace += oqctl->qc_dqblk.dqb_curspace;
3779                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3780                 }
3781         }
3782         oqctl->qc_dqblk.dqb_curspace = curspace;
3783         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3784
3785         RETURN(rc);
3786 }
3787
3788 static int lmv_merge_attr(struct obd_export *exp,
3789                           const struct lmv_stripe_md *lsm,
3790                           struct cl_attr *attr,
3791                           ldlm_blocking_callback cb_blocking)
3792 {
3793         int rc;
3794         int i;
3795
3796         if (!lmv_dir_striped(lsm))
3797                 return 0;
3798
3799         rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3800         if (rc < 0)
3801                 return rc;
3802
3803         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3804                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3805
3806                 if (!inode)
3807                         continue;
3808
3809                 CDEBUG(D_INFO,
3810                        "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
3811                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3812                        i_size_read(inode), (unsigned long long)inode->i_blocks,
3813                        inode->i_nlink, (s64)inode->i_atime.tv_sec,
3814                        (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
3815
3816                 /* for slave stripe, it needs to subtract nlink for . and .. */
3817                 if (i != 0)
3818                         attr->cat_nlink += inode->i_nlink - 2;
3819                 else
3820                         attr->cat_nlink = inode->i_nlink;
3821
3822                 attr->cat_size += i_size_read(inode);
3823                 attr->cat_blocks += inode->i_blocks;
3824
3825                 if (attr->cat_atime < inode->i_atime.tv_sec)
3826                         attr->cat_atime = inode->i_atime.tv_sec;
3827
3828                 if (attr->cat_ctime < inode->i_ctime.tv_sec)
3829                         attr->cat_ctime = inode->i_ctime.tv_sec;
3830
3831                 if (attr->cat_mtime < inode->i_mtime.tv_sec)
3832                         attr->cat_mtime = inode->i_mtime.tv_sec;
3833         }
3834         return 0;
3835 }
3836
3837 static const struct obd_ops lmv_obd_ops = {
3838         .o_owner                = THIS_MODULE,
3839         .o_setup                = lmv_setup,
3840         .o_cleanup              = lmv_cleanup,
3841         .o_precleanup           = lmv_precleanup,
3842         .o_process_config       = lmv_process_config,
3843         .o_connect              = lmv_connect,
3844         .o_disconnect           = lmv_disconnect,
3845         .o_statfs               = lmv_statfs,
3846         .o_get_info             = lmv_get_info,
3847         .o_set_info_async       = lmv_set_info_async,
3848         .o_notify               = lmv_notify,
3849         .o_get_uuid             = lmv_get_uuid,
3850         .o_fid_alloc            = lmv_fid_alloc,
3851         .o_iocontrol            = lmv_iocontrol,
3852         .o_quotactl             = lmv_quotactl
3853 };
3854
3855 static const struct md_ops lmv_md_ops = {
3856         .m_get_root             = lmv_get_root,
3857         .m_null_inode           = lmv_null_inode,
3858         .m_close                = lmv_close,
3859         .m_create               = lmv_create,
3860         .m_enqueue              = lmv_enqueue,
3861         .m_getattr              = lmv_getattr,
3862         .m_getxattr             = lmv_getxattr,
3863         .m_getattr_name         = lmv_getattr_name,
3864         .m_intent_lock          = lmv_intent_lock,
3865         .m_link                 = lmv_link,
3866         .m_rename               = lmv_rename,
3867         .m_setattr              = lmv_setattr,
3868         .m_setxattr             = lmv_setxattr,
3869         .m_fsync                = lmv_fsync,
3870         .m_file_resync          = lmv_file_resync,
3871         .m_read_page            = lmv_read_page,
3872         .m_unlink               = lmv_unlink,
3873         .m_init_ea_size         = lmv_init_ea_size,
3874         .m_cancel_unused        = lmv_cancel_unused,
3875         .m_set_lock_data        = lmv_set_lock_data,
3876         .m_lock_match           = lmv_lock_match,
3877         .m_get_lustre_md        = lmv_get_lustre_md,
3878         .m_free_lustre_md       = lmv_free_lustre_md,
3879         .m_merge_attr           = lmv_merge_attr,
3880         .m_set_open_replay_data = lmv_set_open_replay_data,
3881         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3882         .m_intent_getattr_async = lmv_intent_getattr_async,
3883         .m_revalidate_lock      = lmv_revalidate_lock,
3884         .m_get_fid_from_lsm     = lmv_get_fid_from_lsm,
3885         .m_unpackmd             = lmv_unpackmd,
3886         .m_rmfid                = lmv_rmfid,
3887 };
3888
3889 static int __init lmv_init(void)
3890 {
3891         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true,
3892                                    LUSTRE_LMV_NAME, NULL);
3893 }
3894
3895 static void __exit lmv_exit(void)
3896 {
3897         class_unregister_type(LUSTRE_LMV_NAME);
3898 }
3899
3900 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3901 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
3902 MODULE_VERSION(LUSTRE_VERSION_STRING);
3903 MODULE_LICENSE("GPL");
3904
3905 module_init(lmv_init);
3906 module_exit(lmv_exit);