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  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_LMV
  34
  35 #include <linux/file.h>
  36 #include <linux/module.h>
  37 #include <linux/init.h>
  38 #include <linux/user_namespace.h>
  39 #ifdef HAVE_UIDGID_HEADER
  40 # include <linux/uidgid.h>
  41 #endif
  42 #include <linux/slab.h>
  43 #include <linux/pagemap.h>
  44 #include <linux/mm.h>
  45 #include <linux/math64.h>
  46 #include <linux/seq_file.h>
  47 #include <linux/namei.h>
  48
  49 #include <obd_support.h>
  50 #include <lustre_lib.h>
  51 #include <lustre_net.h>
  52 #include <obd_class.h>
  53 #include <lustre_lmv.h>
  54 #include <lprocfs_status.h>
  55 #include <cl_object.h>
  56 #include <lustre_fid.h>
  57 #include <uapi/linux/lustre/lustre_ioctl.h>
  58 #include <lustre_kernelcomm.h>
  59 #include "lmv_internal.h"
  60
  61 static int lmv_check_connect(struct obd_device *obd);
  62
  63 static void lmv_activate_target(struct lmv_obd *lmv,
  64                                 struct lmv_tgt_desc *tgt,
  65                                 int activate)
  66 {
  67         if (tgt->ltd_active == activate)
  68                 return;
  69
  70         tgt->ltd_active = activate;
  71         lmv->desc.ld_active_tgt_count += (activate ? 1 : -1);
  72
  73         tgt->ltd_exp->exp_obd->obd_inactive = !activate;
  74 }
  75
  76 /**
  77  * Error codes:
  78  *
  79  *  -EINVAL  : UUID can't be found in the LMV's target list
  80  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  81  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  82  */
  83 static int lmv_set_mdc_active(struct lmv_obd *lmv,
  84                               const struct obd_uuid *uuid,
  85                               int activate)
  86 {
  87         struct lmv_tgt_desc     *tgt = NULL;
  88         struct obd_device       *obd;
  89         __u32                    i;
  90         int                      rc = 0;
  91         ENTRY;
  92
  93         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  94                         lmv, uuid->uuid, activate);
  95
  96         spin_lock(&lmv->lmv_lock);
  97         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
  98                 tgt = lmv->tgts[i];
  99                 if (tgt == NULL || tgt->ltd_exp == NULL)
 100                         continue;
 101
 102                 CDEBUG(D_INFO, "Target idx %d is %s conn %#llx\n", i,
 103                        tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
 104
 105                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 106                         break;
 107         }
 108
 109         if (i == lmv->desc.ld_tgt_count)
 110                 GOTO(out_lmv_lock, rc = -EINVAL);
 111
 112         obd = class_exp2obd(tgt->ltd_exp);
 113         if (obd == NULL)
 114                 GOTO(out_lmv_lock, rc = -ENOTCONN);
 115
 116         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
 117                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
 118                obd->obd_type->typ_name, i);
 119         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
 120
 121         if (tgt->ltd_active == activate) {
 122                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
 123                        activate ? "" : "in");
 124                 GOTO(out_lmv_lock, rc);
 125         }
 126
 127         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd,
 128                activate ? "" : "in");
 129         lmv_activate_target(lmv, tgt, activate);
 130         EXIT;
 131
 132  out_lmv_lock:
 133         spin_unlock(&lmv->lmv_lock);
 134         return rc;
 135 }
 136
 137 struct obd_uuid *lmv_get_uuid(struct obd_export *exp)
 138 {
 139         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
 140         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
 141
 142         return (tgt == NULL) ? NULL : obd_get_uuid(tgt->ltd_exp);
 143 }
 144
 145 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 146                       enum obd_notify_event ev)
 147 {
 148         struct obd_connect_data *conn_data;
 149         struct lmv_obd          *lmv = &obd->u.lmv;
 150         struct obd_uuid         *uuid;
 151         int                      rc = 0;
 152         ENTRY;
 153
 154         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 155                 CERROR("unexpected notification of %s %s!\n",
 156                        watched->obd_type->typ_name,
 157                        watched->obd_name);
 158                 RETURN(-EINVAL);
 159         }
 160
 161         uuid = &watched->u.cli.cl_target_uuid;
 162         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 163                 /*
 164                  * Set MDC as active before notifying the observer, so the
 165                  * observer can use the MDC normally.
 166                  */
 167                 rc = lmv_set_mdc_active(lmv, uuid,
 168                                         ev == OBD_NOTIFY_ACTIVE);
 169                 if (rc) {
 170                         CERROR("%sactivation of %s failed: %d\n",
 171                                ev == OBD_NOTIFY_ACTIVE ? "" : "de",
 172                                uuid->uuid, rc);
 173                         RETURN(rc);
 174                 }
 175         } else if (ev == OBD_NOTIFY_OCD) {
 176                 conn_data = &watched->u.cli.cl_import->imp_connect_data;
 177                 /*
 178                  * XXX: Make sure that ocd_connect_flags from all targets are
 179                  * the same. Otherwise one of MDTs runs wrong version or
 180                  * something like this.  --umka
 181                  */
 182                 obd->obd_self_export->exp_connect_data = *conn_data;
 183         }
 184
 185         /*
 186          * Pass the notification up the chain.
 187          */
 188         if (obd->obd_observer)
 189                 rc = obd_notify(obd->obd_observer, watched, ev);
 190
 191         RETURN(rc);
 192 }
 193
 194 static int lmv_connect(const struct lu_env *env,
 195                        struct obd_export **pexp, struct obd_device *obd,
 196                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 197                        void *localdata)
 198 {
 199         struct lmv_obd *lmv = &obd->u.lmv;
 200         struct lustre_handle conn = { 0 };
 201         struct obd_export *exp;
 202         int rc;
 203         ENTRY;
 204
 205         rc = class_connect(&conn, obd, cluuid);
 206         if (rc) {
 207                 CERROR("class_connection() returned %d\n", rc);
 208                 RETURN(rc);
 209         }
 210
 211         exp = class_conn2export(&conn);
 212
 213         lmv->connected = 0;
 214         lmv->conn_data = *data;
 215         lmv->lmv_cache = localdata;
 216
 217         lmv->lmv_tgts_kobj = kobject_create_and_add("target_obds",
 218                                                     &obd->obd_kset.kobj);
 219         if (!lmv->lmv_tgts_kobj) {
 220                 CERROR("%s: cannot create /sys/fs/lustre/%s/%s/target_obds\n",
 221                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
 222         }
 223
 224         rc = lmv_check_connect(obd);
 225         if (rc != 0)
 226                 GOTO(out_sysfs, rc);
 227
 228         *pexp = exp;
 229
 230         RETURN(rc);
 231
 232 out_sysfs:
 233         if (lmv->lmv_tgts_kobj)
 234                 kobject_put(lmv->lmv_tgts_kobj);
 235
 236         class_disconnect(exp);
 237
 238         return rc;
 239 }
 240
 241 static int lmv_init_ea_size(struct obd_export *exp, __u32 easize,
 242                             __u32 def_easize)
 243 {
 244         struct obd_device       *obd = exp->exp_obd;
 245         struct lmv_obd          *lmv = &obd->u.lmv;
 246         __u32                    i;
 247         int                      rc = 0;
 248         int                      change = 0;
 249         ENTRY;
 250
 251         if (lmv->max_easize < easize) {
 252                 lmv->max_easize = easize;
 253                 change = 1;
 254         }
 255         if (lmv->max_def_easize < def_easize) {
 256                 lmv->max_def_easize = def_easize;
 257                 change = 1;
 258         }
 259
 260         if (change == 0)
 261                 RETURN(0);
 262
 263         if (lmv->connected == 0)
 264                 RETURN(0);
 265
 266         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 267                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 268
 269                 if (tgt == NULL || tgt->ltd_exp == NULL) {
 270                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
 271                         continue;
 272                 }
 273                 if (!tgt->ltd_active)
 274                         continue;
 275
 276                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
 277                 if (rc) {
 278                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
 279                                " rc = %d\n", obd->obd_name, i, rc);
 280                         break;
 281                 }
 282         }
 283         RETURN(rc);
 284 }
 285
 286 #define MAX_STRING_SIZE 128
 287
 288 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 289 {
 290         struct lmv_obd          *lmv = &obd->u.lmv;
 291         struct obd_device       *mdc_obd;
 292         struct obd_export       *mdc_exp;
 293         struct lu_fld_target     target;
 294         int                      rc;
 295         ENTRY;
 296
 297         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 298                                         &obd->obd_uuid);
 299         if (!mdc_obd) {
 300                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 301                 RETURN(-EINVAL);
 302         }
 303
 304         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
 305                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 306                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
 307
 308         if (!mdc_obd->obd_set_up) {
 309                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 310                 RETURN(-EINVAL);
 311         }
 312
 313         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
 314                          &lmv->conn_data, lmv->lmv_cache);
 315         if (rc) {
 316                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 317                 RETURN(rc);
 318         }
 319
 320         /*
 321          * Init fid sequence client for this mdc and add new fld target.
 322          */
 323         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
 324         if (rc)
 325                 RETURN(rc);
 326
 327         target.ft_srv = NULL;
 328         target.ft_exp = mdc_exp;
 329         target.ft_idx = tgt->ltd_idx;
 330
 331         fld_client_add_target(&lmv->lmv_fld, &target);
 332
 333         rc = obd_register_observer(mdc_obd, obd);
 334         if (rc) {
 335                 obd_disconnect(mdc_exp);
 336                 CERROR("target %s register_observer error %d\n",
 337                        tgt->ltd_uuid.uuid, rc);
 338                 RETURN(rc);
 339         }
 340
 341         if (obd->obd_observer) {
 342                 /*
 343                  * Tell the observer about the new target.
 344                  */
 345                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 346                                 OBD_NOTIFY_ACTIVE);
 347                 if (rc) {
 348                         obd_disconnect(mdc_exp);
 349                         RETURN(rc);
 350                 }
 351         }
 352
 353         tgt->ltd_active = 1;
 354         tgt->ltd_exp = mdc_exp;
 355         lmv->desc.ld_active_tgt_count++;
 356
 357         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
 358
 359         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 360                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 361                 atomic_read(&obd->obd_refcount));
 362
 363         if (lmv->lmv_tgts_kobj)
 364                 /* Even if we failed to create the link, that's fine */
 365                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
 366                                        &mdc_obd->obd_kset.kobj,
 367                                        mdc_obd->obd_name);
 368         RETURN(0);
 369 }
 370
 371 static void lmv_del_target(struct lmv_obd *lmv, int index)
 372 {
 373         if (lmv->tgts[index] == NULL)
 374                 return;
 375
 376         OBD_FREE_PTR(lmv->tgts[index]);
 377         lmv->tgts[index] = NULL;
 378         return;
 379 }
 380
 381 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 382                            __u32 index, int gen)
 383 {
 384         struct obd_device *mdc_obd;
 385         struct lmv_obd      *lmv = &obd->u.lmv;
 386         struct lmv_tgt_desc *tgt;
 387         int                  orig_tgt_count = 0;
 388         int                  rc = 0;
 389         ENTRY;
 390
 391         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
 392         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
 393                                         &obd->obd_uuid);
 394         if (!mdc_obd) {
 395                 CERROR("%s: Target %s not attached: rc = %d\n",
 396                        obd->obd_name, uuidp->uuid, -EINVAL);
 397                 RETURN(-EINVAL);
 398         }
 399
 400         mutex_lock(&lmv->lmv_init_mutex);
 401         if ((index < lmv->tgts_size) && (lmv->tgts[index] != NULL)) {
 402                 tgt = lmv->tgts[index];
 403                 CERROR("%s: UUID %s already assigned at LMV target index %d:"
 404                        " rc = %d\n", obd->obd_name,
 405                        obd_uuid2str(&tgt->ltd_uuid), index, -EEXIST);
 406                 mutex_unlock(&lmv->lmv_init_mutex);
 407                 RETURN(-EEXIST);
 408         }
 409
 410         if (index >= lmv->tgts_size) {
 411                 /* We need to reallocate the lmv target array. */
 412                 struct lmv_tgt_desc **newtgts, **old = NULL;
 413                 __u32 newsize = 1;
 414                 __u32 oldsize = 0;
 415
 416                 while (newsize < index + 1)
 417                         newsize = newsize << 1;
 418                 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
 419                 if (newtgts == NULL) {
 420                         mutex_unlock(&lmv->lmv_init_mutex);
 421                         RETURN(-ENOMEM);
 422                 }
 423
 424                 if (lmv->tgts_size) {
 425                         memcpy(newtgts, lmv->tgts,
 426                                sizeof(*newtgts) * lmv->tgts_size);
 427                         old = lmv->tgts;
 428                         oldsize = lmv->tgts_size;
 429                 }
 430
 431                 lmv->tgts = newtgts;
 432                 lmv->tgts_size = newsize;
 433                 smp_rmb();
 434                 if (old)
 435                         OBD_FREE(old, sizeof(*old) * oldsize);
 436
 437                 CDEBUG(D_CONFIG, "tgts: %p size: %d\n", lmv->tgts,
 438                        lmv->tgts_size);
 439         }
 440
 441         OBD_ALLOC_PTR(tgt);
 442         if (!tgt) {
 443                 mutex_unlock(&lmv->lmv_init_mutex);
 444                 RETURN(-ENOMEM);
 445         }
 446
 447         mutex_init(&tgt->ltd_fid_mutex);
 448         tgt->ltd_idx = index;
 449         tgt->ltd_uuid = *uuidp;
 450         tgt->ltd_active = 0;
 451         lmv->tgts[index] = tgt;
 452         if (index >= lmv->desc.ld_tgt_count) {
 453                 orig_tgt_count = lmv->desc.ld_tgt_count;
 454                 lmv->desc.ld_tgt_count = index + 1;
 455         }
 456
 457         if (lmv->connected == 0) {
 458                 /* lmv_check_connect() will connect this target. */
 459                 mutex_unlock(&lmv->lmv_init_mutex);
 460                 RETURN(0);
 461         }
 462
 463         /* Otherwise let's connect it ourselves */
 464         mutex_unlock(&lmv->lmv_init_mutex);
 465         rc = lmv_connect_mdc(obd, tgt);
 466         if (rc != 0) {
 467                 spin_lock(&lmv->lmv_lock);
 468                 if (lmv->desc.ld_tgt_count == index + 1)
 469                         lmv->desc.ld_tgt_count = orig_tgt_count;
 470                 memset(tgt, 0, sizeof(*tgt));
 471                 spin_unlock(&lmv->lmv_lock);
 472         } else {
 473                 int easize = sizeof(struct lmv_stripe_md) +
 474                         lmv->desc.ld_tgt_count * sizeof(struct lu_fid);
 475                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
 476         }
 477
 478         RETURN(rc);
 479 }
 480
 481 static int lmv_check_connect(struct obd_device *obd)
 482 {
 483         struct lmv_obd          *lmv = &obd->u.lmv;
 484         struct lmv_tgt_desc     *tgt;
 485         __u32                    i;
 486         int                      rc;
 487         int                      easize;
 488         ENTRY;
 489
 490         if (lmv->connected)
 491                 RETURN(0);
 492
 493         mutex_lock(&lmv->lmv_init_mutex);
 494         if (lmv->connected) {
 495                 mutex_unlock(&lmv->lmv_init_mutex);
 496                 RETURN(0);
 497         }
 498
 499         if (lmv->desc.ld_tgt_count == 0) {
 500                 mutex_unlock(&lmv->lmv_init_mutex);
 501                 CERROR("%s: no targets configured.\n", obd->obd_name);
 502                 RETURN(-EINVAL);
 503         }
 504
 505         LASSERT(lmv->tgts != NULL);
 506
 507         if (lmv->tgts[0] == NULL) {
 508                 mutex_unlock(&lmv->lmv_init_mutex);
 509                 CERROR("%s: no target configured for index 0.\n",
 510                        obd->obd_name);
 511                 RETURN(-EINVAL);
 512         }
 513
 514         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 515                obd->obd_uuid.uuid, obd->obd_name);
 516
 517         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 518                 tgt = lmv->tgts[i];
 519                 if (tgt == NULL)
 520                         continue;
 521                 rc = lmv_connect_mdc(obd, tgt);
 522                 if (rc)
 523                         GOTO(out_disc, rc);
 524         }
 525
 526         lmv->connected = 1;
 527         easize = lmv_mds_md_size(lmv->desc.ld_tgt_count, LMV_MAGIC);
 528         lmv_init_ea_size(obd->obd_self_export, easize, 0);
 529         mutex_unlock(&lmv->lmv_init_mutex);
 530         RETURN(0);
 531
 532  out_disc:
 533         while (i-- > 0) {
 534                 int rc2;
 535                 tgt = lmv->tgts[i];
 536                 if (tgt == NULL)
 537                         continue;
 538                 tgt->ltd_active = 0;
 539                 if (tgt->ltd_exp) {
 540                         --lmv->desc.ld_active_tgt_count;
 541                         rc2 = obd_disconnect(tgt->ltd_exp);
 542                         if (rc2) {
 543                                 CERROR("LMV target %s disconnect on "
 544                                        "MDC idx %d: error %d\n",
 545                                        tgt->ltd_uuid.uuid, i, rc2);
 546                         }
 547                 }
 548         }
 549
 550         mutex_unlock(&lmv->lmv_init_mutex);
 551
 552         RETURN(rc);
 553 }
 554
 555 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 556 {
 557         struct lmv_obd         *lmv = &obd->u.lmv;
 558         struct obd_device      *mdc_obd;
 559         int                     rc;
 560         ENTRY;
 561
 562         LASSERT(tgt != NULL);
 563         LASSERT(obd != NULL);
 564
 565         mdc_obd = class_exp2obd(tgt->ltd_exp);
 566
 567         if (mdc_obd) {
 568                 mdc_obd->obd_force = obd->obd_force;
 569                 mdc_obd->obd_fail = obd->obd_fail;
 570                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 571
 572                 if (lmv->lmv_tgts_kobj)
 573                         sysfs_remove_link(lmv->lmv_tgts_kobj,
 574                                           mdc_obd->obd_name);
 575         }
 576
 577         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
 578         if (rc)
 579                 CERROR("Can't finanize fids factory\n");
 580
 581         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 582                tgt->ltd_exp->exp_obd->obd_name,
 583                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 584
 585         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 586         rc = obd_disconnect(tgt->ltd_exp);
 587         if (rc) {
 588                 if (tgt->ltd_active) {
 589                         CERROR("Target %s disconnect error %d\n",
 590                                tgt->ltd_uuid.uuid, rc);
 591                 }
 592         }
 593
 594         lmv_activate_target(lmv, tgt, 0);
 595         tgt->ltd_exp = NULL;
 596         RETURN(0);
 597 }
 598
 599 static int lmv_disconnect(struct obd_export *exp)
 600 {
 601         struct obd_device       *obd = class_exp2obd(exp);
 602         struct lmv_obd          *lmv = &obd->u.lmv;
 603         int                      rc;
 604         __u32                    i;
 605         ENTRY;
 606
 607         if (!lmv->tgts)
 608                 goto out_local;
 609
 610         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 611                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
 612                         continue;
 613
 614                 lmv_disconnect_mdc(obd, lmv->tgts[i]);
 615         }
 616
 617         if (lmv->lmv_tgts_kobj)
 618                 kobject_put(lmv->lmv_tgts_kobj);
 619
 620 out_local:
 621         /*
 622          * This is the case when no real connection is established by
 623          * lmv_check_connect().
 624          */
 625         if (!lmv->connected)
 626                 class_export_put(exp);
 627         rc = class_disconnect(exp);
 628         lmv->connected = 0;
 629
 630         RETURN(rc);
 631 }
 632
 633 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
 634                         void __user *uarg)
 635 {
 636         struct obd_device       *obddev = class_exp2obd(exp);
 637         struct lmv_obd          *lmv = &obddev->u.lmv;
 638         struct getinfo_fid2path *gf;
 639         struct lmv_tgt_desc     *tgt;
 640         struct getinfo_fid2path *remote_gf = NULL;
 641         struct lu_fid           root_fid;
 642         int                     remote_gf_size = 0;
 643         int                     rc;
 644
 645         gf = karg;
 646         tgt = lmv_find_target(lmv, &gf->gf_fid);
 647         if (IS_ERR(tgt))
 648                 RETURN(PTR_ERR(tgt));
 649
 650         root_fid = *gf->gf_u.gf_root_fid;
 651         LASSERT(fid_is_sane(&root_fid));
 652
 653 repeat_fid2path:
 654         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
 655         if (rc != 0 && rc != -EREMOTE)
 656                 GOTO(out_fid2path, rc);
 657
 658         /* If remote_gf != NULL, it means just building the
 659          * path on the remote MDT, copy this path segement to gf */
 660         if (remote_gf != NULL) {
 661                 struct getinfo_fid2path *ori_gf;
 662                 char *ptr;
 663                 int len;
 664
 665                 ori_gf = (struct getinfo_fid2path *)karg;
 666                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
 667                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
 668                         GOTO(out_fid2path, rc = -EOVERFLOW);
 669
 670                 ptr = ori_gf->gf_u.gf_path;
 671
 672                 len = strlen(gf->gf_u.gf_path);
 673                 /* move the current path to the right to release space
 674                  * for closer-to-root part */
 675                 memmove(ptr + len + 1, ptr, strlen(ori_gf->gf_u.gf_path));
 676                 memcpy(ptr, gf->gf_u.gf_path, len);
 677                 ptr[len] = '/';
 678         }
 679
 680         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
 681                tgt->ltd_exp->exp_obd->obd_name,
 682                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
 683                gf->gf_linkno);
 684
 685         if (rc == 0)
 686                 GOTO(out_fid2path, rc);
 687
 688         /* sigh, has to go to another MDT to do path building further */
 689         if (remote_gf == NULL) {
 690                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
 691                 OBD_ALLOC(remote_gf, remote_gf_size);
 692                 if (remote_gf == NULL)
 693                         GOTO(out_fid2path, rc = -ENOMEM);
 694                 remote_gf->gf_pathlen = PATH_MAX;
 695         }
 696
 697         if (!fid_is_sane(&gf->gf_fid)) {
 698                 CERROR("%s: invalid FID "DFID": rc = %d\n",
 699                        tgt->ltd_exp->exp_obd->obd_name,
 700                        PFID(&gf->gf_fid), -EINVAL);
 701                 GOTO(out_fid2path, rc = -EINVAL);
 702         }
 703
 704         tgt = lmv_find_target(lmv, &gf->gf_fid);
 705         if (IS_ERR(tgt))
 706                 GOTO(out_fid2path, rc = -EINVAL);
 707
 708         remote_gf->gf_fid = gf->gf_fid;
 709         remote_gf->gf_recno = -1;
 710         remote_gf->gf_linkno = -1;
 711         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
 712         *remote_gf->gf_u.gf_root_fid = root_fid;
 713         gf = remote_gf;
 714         goto repeat_fid2path;
 715
 716 out_fid2path:
 717         if (remote_gf != NULL)
 718                 OBD_FREE(remote_gf, remote_gf_size);
 719         RETURN(rc);
 720 }
 721
 722 static int lmv_hsm_req_count(struct lmv_obd *lmv,
 723                              const struct hsm_user_request *hur,
 724                              const struct lmv_tgt_desc *tgt_mds)
 725 {
 726         __u32                    i;
 727         int                      nr = 0;
 728         struct lmv_tgt_desc     *curr_tgt;
 729
 730         /* count how many requests must be sent to the given target */
 731         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
 732                 curr_tgt = lmv_find_target(lmv, &hur->hur_user_item[i].hui_fid);
 733                 if (IS_ERR(curr_tgt))
 734                         RETURN(PTR_ERR(curr_tgt));
 735                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
 736                         nr++;
 737         }
 738         return nr;
 739 }
 740
 741 static int lmv_hsm_req_build(struct lmv_obd *lmv,
 742                               struct hsm_user_request *hur_in,
 743                               const struct lmv_tgt_desc *tgt_mds,
 744                               struct hsm_user_request *hur_out)
 745 {
 746         __u32                    i, nr_out;
 747         struct lmv_tgt_desc     *curr_tgt;
 748
 749         /* build the hsm_user_request for the given target */
 750         hur_out->hur_request = hur_in->hur_request;
 751         nr_out = 0;
 752         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
 753                 curr_tgt = lmv_find_target(lmv,
 754                                            &hur_in->hur_user_item[i].hui_fid);
 755                 if (IS_ERR(curr_tgt))
 756                         RETURN(PTR_ERR(curr_tgt));
 757                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
 758                         hur_out->hur_user_item[nr_out] =
 759                                                 hur_in->hur_user_item[i];
 760                         nr_out++;
 761                 }
 762         }
 763         hur_out->hur_request.hr_itemcount = nr_out;
 764         memcpy(hur_data(hur_out), hur_data(hur_in),
 765                hur_in->hur_request.hr_data_len);
 766
 767         RETURN(0);
 768 }
 769
 770 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
 771                                  int len, struct lustre_kernelcomm *lk,
 772                                  void __user *uarg)
 773 {
 774         struct lmv_obd *lmv = &obd->u.lmv;
 775         __u32   i;
 776         int     rc;
 777         ENTRY;
 778
 779         /* unregister request (call from llapi_hsm_copytool_fini) */
 780         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 781                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 782
 783                 if (tgt == NULL || tgt->ltd_exp == NULL)
 784                         continue;
 785                 /* best effort: try to clean as much as possible
 786                  * (continue on error) */
 787                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 788         }
 789
 790         /* Whatever the result, remove copytool from kuc groups.
 791          * Unreached coordinators will get EPIPE on next requests
 792          * and will unregister automatically.
 793          */
 794         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 795
 796         RETURN(rc);
 797 }
 798
 799 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
 800                                int len, struct lustre_kernelcomm *lk,
 801                                void __user *uarg)
 802 {
 803         struct lmv_obd *lmv = &obd->u.lmv;
 804         struct file *filp;
 805         __u32 i, j;
 806         int err;
 807         bool any_set = false;
 808         struct kkuc_ct_data *kcd;
 809         size_t kcd_size;
 810         int rc = 0;
 811         ENTRY;
 812
 813         filp = fget(lk->lk_wfd);
 814         if (!filp)
 815                 RETURN(-EBADF);
 816
 817         if (lk->lk_flags & LK_FLG_DATANR)
 818                 kcd_size = offsetof(struct kkuc_ct_data,
 819                                     kcd_archives[lk->lk_data_count]);
 820         else
 821                 kcd_size = sizeof(*kcd);
 822
 823         OBD_ALLOC(kcd, kcd_size);
 824         if (kcd == NULL)
 825                 GOTO(err_fput, rc = -ENOMEM);
 826
 827         kcd->kcd_nr_archives = lk->lk_data_count;
 828         if (lk->lk_flags & LK_FLG_DATANR) {
 829                 kcd->kcd_magic = KKUC_CT_DATA_ARRAY_MAGIC;
 830                 if (lk->lk_data_count > 0)
 831                         memcpy(kcd->kcd_archives, lk->lk_data,
 832                                sizeof(*kcd->kcd_archives) * lk->lk_data_count);
 833         } else {
 834                 kcd->kcd_magic = KKUC_CT_DATA_BITMAP_MAGIC;
 835         }
 836
 837         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
 838                                    lk->lk_group, kcd, kcd_size);
 839         OBD_FREE(kcd, kcd_size);
 840         if (rc)
 841                 GOTO(err_fput, rc);
 842
 843         /* All or nothing: try to register to all MDS.
 844          * In case of failure, unregister from previous MDS,
 845          * except if it because of inactive target. */
 846         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 847                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 848
 849                 if (tgt == NULL || tgt->ltd_exp == NULL)
 850                         continue;
 851
 852                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 853                 if (err) {
 854                         if (tgt->ltd_active) {
 855                                 /* permanent error */
 856                                 CERROR("%s: iocontrol MDC %s on MDT"
 857                                        " idx %d cmd %x: err = %d\n",
 858                                        lmv2obd_dev(lmv)->obd_name,
 859                                        tgt->ltd_uuid.uuid, i, cmd, err);
 860                                 rc = err;
 861                                 lk->lk_flags |= LK_FLG_STOP;
 862                                 /* unregister from previous MDS */
 863                                 for (j = 0; j < i; j++) {
 864                                         tgt = lmv->tgts[j];
 865                                         if (tgt == NULL || tgt->ltd_exp == NULL)
 866                                                 continue;
 867                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
 868                                                       lk, uarg);
 869                                 }
 870                                 GOTO(err_kkuc_rem, rc);
 871                         }
 872                         /* else: transient error.
 873                          * kuc will register to the missing MDT
 874                          * when it is back */
 875                 } else {
 876                         any_set = true;
 877                 }
 878         }
 879
 880         if (!any_set)
 881                 /* no registration done: return error */
 882                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
 883
 884         RETURN(0);
 885
 886 err_kkuc_rem:
 887         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 888
 889 err_fput:
 890         fput(filp);
 891         return rc;
 892 }
 893
 894
 895
 896
 897 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 898                          int len, void *karg, void __user *uarg)
 899 {
 900         struct obd_device       *obddev = class_exp2obd(exp);
 901         struct lmv_obd          *lmv = &obddev->u.lmv;
 902         struct lmv_tgt_desc     *tgt = NULL;
 903         __u32                    i = 0;
 904         int                      rc = 0;
 905         int                      set = 0;
 906         __u32                    count = lmv->desc.ld_tgt_count;
 907         ENTRY;
 908
 909         if (count == 0)
 910                 RETURN(-ENOTTY);
 911
 912         switch (cmd) {
 913         case IOC_OBD_STATFS: {
 914                 struct obd_ioctl_data *data = karg;
 915                 struct obd_device *mdc_obd;
 916                 struct obd_statfs stat_buf = {0};
 917                 __u32 index;
 918
 919                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 920                 if ((index >= count))
 921                         RETURN(-ENODEV);
 922
 923                 tgt = lmv->tgts[index];
 924                 if (tgt == NULL || !tgt->ltd_active)
 925                         RETURN(-ENODATA);
 926
 927                 mdc_obd = class_exp2obd(tgt->ltd_exp);
 928                 if (!mdc_obd)
 929                         RETURN(-EINVAL);
 930
 931                 /* copy UUID */
 932                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 933                                  min((int) data->ioc_plen2,
 934                                      (int) sizeof(struct obd_uuid))))
 935                         RETURN(-EFAULT);
 936
 937                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
 938                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
 939                                 0);
 940                 if (rc)
 941                         RETURN(rc);
 942                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
 943                                  min((int) data->ioc_plen1,
 944                                      (int) sizeof(stat_buf))))
 945                         RETURN(-EFAULT);
 946                 break;
 947         }
 948         case OBD_IOC_QUOTACTL: {
 949                 struct if_quotactl *qctl = karg;
 950                 struct obd_quotactl *oqctl;
 951
 952                 if (qctl->qc_valid == QC_MDTIDX) {
 953                         if (count <= qctl->qc_idx)
 954                                 RETURN(-EINVAL);
 955
 956                         tgt = lmv->tgts[qctl->qc_idx];
 957                         if (tgt == NULL || tgt->ltd_exp == NULL)
 958                                 RETURN(-EINVAL);
 959                 } else if (qctl->qc_valid == QC_UUID) {
 960                         for (i = 0; i < count; i++) {
 961                                 tgt = lmv->tgts[i];
 962                                 if (tgt == NULL)
 963                                         continue;
 964                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 965                                                      &qctl->obd_uuid))
 966                                         continue;
 967
 968                                 if (tgt->ltd_exp == NULL)
 969                                         RETURN(-EINVAL);
 970
 971                                 break;
 972                         }
 973                 } else {
 974                         RETURN(-EINVAL);
 975                 }
 976
 977                 if (i >= count)
 978                         RETURN(-EAGAIN);
 979
 980                 LASSERT(tgt != NULL && tgt->ltd_exp != NULL);
 981                 OBD_ALLOC_PTR(oqctl);
 982                 if (!oqctl)
 983                         RETURN(-ENOMEM);
 984
 985                 QCTL_COPY(oqctl, qctl);
 986                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 987                 if (rc == 0) {
 988                         QCTL_COPY(qctl, oqctl);
 989                         qctl->qc_valid = QC_MDTIDX;
 990                         qctl->obd_uuid = tgt->ltd_uuid;
 991                 }
 992                 OBD_FREE_PTR(oqctl);
 993                 break;
 994         }
 995         case LL_IOC_GET_CONNECT_FLAGS: {
 996                 tgt = lmv->tgts[0];
 997                 if (tgt == NULL || tgt->ltd_exp == NULL)
 998                         RETURN(-ENODATA);
 999                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1000                 break;
1001         }
1002         case LL_IOC_FID2MDTIDX: {
1003                 struct lu_fid *fid = karg;
1004                 int             mdt_index;
1005
1006                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
1007                 if (rc != 0)
1008                         RETURN(rc);
1009
1010                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
1011                  * point to user space memory for FID2MDTIDX. */
1012                 *(__u32 *)uarg = mdt_index;
1013                 break;
1014         }
1015         case OBD_IOC_FID2PATH: {
1016                 rc = lmv_fid2path(exp, len, karg, uarg);
1017                 break;
1018         }
1019         case LL_IOC_HSM_STATE_GET:
1020         case LL_IOC_HSM_STATE_SET:
1021         case LL_IOC_HSM_ACTION: {
1022                 struct md_op_data       *op_data = karg;
1023
1024                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1025                 if (IS_ERR(tgt))
1026                         RETURN(PTR_ERR(tgt));
1027
1028                 if (tgt->ltd_exp == NULL)
1029                         RETURN(-EINVAL);
1030
1031                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1032                 break;
1033         }
1034         case LL_IOC_HSM_PROGRESS: {
1035                 const struct hsm_progress_kernel *hpk = karg;
1036
1037                 tgt = lmv_find_target(lmv, &hpk->hpk_fid);
1038                 if (IS_ERR(tgt))
1039                         RETURN(PTR_ERR(tgt));
1040                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1041                 break;
1042         }
1043         case LL_IOC_HSM_REQUEST: {
1044                 struct hsm_user_request *hur = karg;
1045                 unsigned int reqcount = hur->hur_request.hr_itemcount;
1046
1047                 if (reqcount == 0)
1048                         RETURN(0);
1049
1050                 /* if the request is about a single fid
1051                  * or if there is a single MDS, no need to split
1052                  * the request. */
1053                 if (reqcount == 1 || count == 1) {
1054                         tgt = lmv_find_target(lmv,
1055                                               &hur->hur_user_item[0].hui_fid);
1056                         if (IS_ERR(tgt))
1057                                 RETURN(PTR_ERR(tgt));
1058                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1059                 } else {
1060                         /* split fid list to their respective MDS */
1061                         for (i = 0; i < count; i++) {
1062                                 int nr, rc1;
1063                                 size_t reqlen;
1064                                 struct hsm_user_request *req;
1065
1066                                 tgt = lmv->tgts[i];
1067                                 if (tgt == NULL || tgt->ltd_exp == NULL)
1068                                         continue;
1069
1070                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
1071                                 if (nr < 0)
1072                                         RETURN(nr);
1073                                 if (nr == 0) /* nothing for this MDS */
1074                                         continue;
1075
1076                                 /* build a request with fids for this MDS */
1077                                 reqlen = offsetof(typeof(*hur),
1078                                                   hur_user_item[nr])
1079                                                 + hur->hur_request.hr_data_len;
1080                                 OBD_ALLOC_LARGE(req, reqlen);
1081                                 if (req == NULL)
1082                                         RETURN(-ENOMEM);
1083                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
1084                                 if (rc1 < 0)
1085                                         GOTO(hsm_req_err, rc1);
1086                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
1087                                                     req, uarg);
1088 hsm_req_err:
1089                                 if (rc1 != 0 && rc == 0)
1090                                         rc = rc1;
1091                                 OBD_FREE_LARGE(req, reqlen);
1092                         }
1093                 }
1094                 break;
1095         }
1096         case LL_IOC_LOV_SWAP_LAYOUTS: {
1097                 struct md_op_data       *op_data = karg;
1098                 struct lmv_tgt_desc     *tgt1, *tgt2;
1099
1100                 tgt1 = lmv_find_target(lmv, &op_data->op_fid1);
1101                 if (IS_ERR(tgt1))
1102                         RETURN(PTR_ERR(tgt1));
1103
1104                 tgt2 = lmv_find_target(lmv, &op_data->op_fid2);
1105                 if (IS_ERR(tgt2))
1106                         RETURN(PTR_ERR(tgt2));
1107
1108                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1109                         RETURN(-EINVAL);
1110
1111                 /* only files on same MDT can have their layouts swapped */
1112                 if (tgt1->ltd_idx != tgt2->ltd_idx)
1113                         RETURN(-EPERM);
1114
1115                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1116                 break;
1117         }
1118         case LL_IOC_HSM_CT_START: {
1119                 struct lustre_kernelcomm *lk = karg;
1120                 if (lk->lk_flags & LK_FLG_STOP)
1121                         rc = lmv_hsm_ct_unregister(obddev, cmd, len, lk, uarg);
1122                 else
1123                         rc = lmv_hsm_ct_register(obddev, cmd, len, lk, uarg);
1124                 break;
1125         }
1126         default:
1127                 for (i = 0; i < count; i++) {
1128                         struct obd_device *mdc_obd;
1129                         int err;
1130
1131                         tgt = lmv->tgts[i];
1132                         if (tgt == NULL || tgt->ltd_exp == NULL)
1133                                 continue;
1134                         /* ll_umount_begin() sets force flag but for lmv, not
1135                          * mdc. Let's pass it through */
1136                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1137                         mdc_obd->obd_force = obddev->obd_force;
1138                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1139                         if (err) {
1140                                 if (tgt->ltd_active) {
1141                                         CERROR("error: iocontrol MDC %s on MDT"
1142                                                " idx %d cmd %x: err = %d\n",
1143                                                tgt->ltd_uuid.uuid, i, cmd, err);
1144                                         if (!rc)
1145                                                 rc = err;
1146                                 }
1147                         } else
1148                                 set = 1;
1149                 }
1150                 if (!set && !rc)
1151                         rc = -EIO;
1152         }
1153         RETURN(rc);
1154 }
1155
1156 /**
1157  * This is _inode_ placement policy function (not name).
1158  */
1159 static int lmv_placement_policy(struct obd_device *obd,
1160                                 struct md_op_data *op_data, u32 *mds)
1161 {
1162         struct lmv_obd     *lmv = &obd->u.lmv;
1163         struct lmv_user_md *lum;
1164
1165         ENTRY;
1166
1167         LASSERT(mds != NULL);
1168
1169         if (lmv->desc.ld_tgt_count == 1) {
1170                 *mds = 0;
1171                 RETURN(0);
1172         }
1173
1174         lum = op_data->op_data;
1175         /* Choose MDS by
1176          * 1. See if the stripe offset is specified by lum.
1177          * 2. Then check if there is default stripe offset.
1178          * 3. Finally choose MDS by name hash if the parent
1179          *    is striped directory. (see lmv_locate_tgt()).
1180          *
1181          * presently explicit MDT location is not supported
1182          * for foreign dirs (as it can't be embedded into free
1183          * format LMV, like with lum_stripe_offset), so we only
1184          * rely on default stripe offset or then name hashing.
1185          */
1186         if (op_data->op_cli_flags & CLI_SET_MEA && lum != NULL &&
1187             le32_to_cpu(lum->lum_magic != LMV_MAGIC_FOREIGN) &&
1188             le32_to_cpu(lum->lum_stripe_offset) != (__u32)-1) {
1189                 *mds = le32_to_cpu(lum->lum_stripe_offset);
1190         } else if (op_data->op_default_stripe_offset != (__u32)-1) {
1191                 *mds = op_data->op_default_stripe_offset;
1192                 op_data->op_mds = *mds;
1193                 /* Correct the stripe offset in lum */
1194                 if (lum != NULL &&
1195                     le32_to_cpu(lum->lum_magic != LMV_MAGIC_FOREIGN))
1196                         lum->lum_stripe_offset = cpu_to_le32(*mds);
1197         } else {
1198                 *mds = op_data->op_mds;
1199         }
1200
1201         RETURN(0);
1202 }
1203
1204 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid, u32 mds)
1205 {
1206         struct lmv_tgt_desc     *tgt;
1207         int                      rc;
1208         ENTRY;
1209
1210         tgt = lmv_get_target(lmv, mds, NULL);
1211         if (IS_ERR(tgt))
1212                 RETURN(PTR_ERR(tgt));
1213
1214         /*
1215          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1216          * on server that seq in new allocated fid is not yet known.
1217          */
1218         mutex_lock(&tgt->ltd_fid_mutex);
1219
1220         if (tgt->ltd_active == 0 || tgt->ltd_exp == NULL)
1221                 GOTO(out, rc = -ENODEV);
1222
1223         /*
1224          * Asking underlying tgt layer to allocate new fid.
1225          */
1226         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1227         if (rc > 0) {
1228                 LASSERT(fid_is_sane(fid));
1229                 rc = 0;
1230         }
1231
1232         EXIT;
1233 out:
1234         mutex_unlock(&tgt->ltd_fid_mutex);
1235         return rc;
1236 }
1237
1238 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1239                   struct lu_fid *fid, struct md_op_data *op_data)
1240 {
1241         struct obd_device     *obd = class_exp2obd(exp);
1242         struct lmv_obd        *lmv = &obd->u.lmv;
1243         u32                    mds = 0;
1244         int                    rc;
1245         ENTRY;
1246
1247         LASSERT(op_data != NULL);
1248         LASSERT(fid != NULL);
1249
1250         rc = lmv_placement_policy(obd, op_data, &mds);
1251         if (rc) {
1252                 CERROR("Can't get target for allocating fid, "
1253                        "rc %d\n", rc);
1254                 RETURN(rc);
1255         }
1256
1257         rc = __lmv_fid_alloc(lmv, fid, mds);
1258         if (rc) {
1259                 CERROR("Can't alloc new fid, rc %d\n", rc);
1260                 RETURN(rc);
1261         }
1262
1263         RETURN(rc);
1264 }
1265
1266 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1267 {
1268         struct lmv_obd  *lmv = &obd->u.lmv;
1269         struct lmv_desc *desc;
1270         int             rc;
1271         ENTRY;
1272
1273         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1274                 CERROR("LMV setup requires a descriptor\n");
1275                 RETURN(-EINVAL);
1276         }
1277
1278         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1279         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1280                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1281                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1282                 RETURN(-EINVAL);
1283         }
1284
1285         lmv->tgts_size = 32U;
1286         OBD_ALLOC(lmv->tgts, sizeof(*lmv->tgts) * lmv->tgts_size);
1287         if (lmv->tgts == NULL)
1288                 RETURN(-ENOMEM);
1289
1290         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1291         lmv->desc.ld_tgt_count = 0;
1292         lmv->desc.ld_active_tgt_count = 0;
1293         lmv->max_def_easize = 0;
1294         lmv->max_easize = 0;
1295
1296         spin_lock_init(&lmv->lmv_lock);
1297         mutex_init(&lmv->lmv_init_mutex);
1298
1299         rc = lmv_tunables_init(obd);
1300         if (rc)
1301                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1302                       obd->obd_name, rc);
1303
1304         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1305                              LUSTRE_CLI_FLD_HASH_DHT);
1306         if (rc) {
1307                 CERROR("Can't init FLD, err %d\n", rc);
1308                 GOTO(out, rc);
1309         }
1310
1311         RETURN(0);
1312
1313 out:
1314         return rc;
1315 }
1316
1317 static int lmv_cleanup(struct obd_device *obd)
1318 {
1319         struct lmv_obd   *lmv = &obd->u.lmv;
1320         ENTRY;
1321
1322         fld_client_fini(&lmv->lmv_fld);
1323         if (lmv->tgts != NULL) {
1324                 int i;
1325                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1326                         if (lmv->tgts[i] == NULL)
1327                                 continue;
1328                         lmv_del_target(lmv, i);
1329                 }
1330                 OBD_FREE(lmv->tgts, sizeof(*lmv->tgts) * lmv->tgts_size);
1331                 lmv->tgts_size = 0;
1332         }
1333         RETURN(0);
1334 }
1335
1336 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1337 {
1338         struct lustre_cfg       *lcfg = buf;
1339         struct obd_uuid         obd_uuid;
1340         int                     gen;
1341         __u32                   index;
1342         int                     rc;
1343         ENTRY;
1344
1345         switch (lcfg->lcfg_command) {
1346         case LCFG_ADD_MDC:
1347                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1348                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1349                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1350                         GOTO(out, rc = -EINVAL);
1351
1352                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1353
1354                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1355                         GOTO(out, rc = -EINVAL);
1356                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1357                         GOTO(out, rc = -EINVAL);
1358                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1359                 GOTO(out, rc);
1360         default:
1361                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1362                 GOTO(out, rc = -EINVAL);
1363         }
1364 out:
1365         RETURN(rc);
1366 }
1367
1368 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1369 {
1370         int i;
1371
1372         if (flags & OBD_STATFS_FOR_MDT0)
1373                 return 0;
1374
1375         if (lmv->lmv_statfs_start || lmv->desc.ld_tgt_count == 1)
1376                 return lmv->lmv_statfs_start;
1377
1378         /* choose initial MDT for this client */
1379         for (i = 0;; i++) {
1380                 struct lnet_process_id lnet_id;
1381                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1382                         break;
1383
1384                 if (LNET_NETTYP(LNET_NIDNET(lnet_id.nid)) != LOLND) {
1385                         /* We dont need a full 64-bit modulus, just enough
1386                          * to distribute the requests across MDTs evenly.
1387                          */
1388                         lmv->lmv_statfs_start =
1389                                 (u32)lnet_id.nid % lmv->desc.ld_tgt_count;
1390                         break;
1391                 }
1392         }
1393
1394         return lmv->lmv_statfs_start;
1395 }
1396
1397 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1398                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1399 {
1400         struct obd_device       *obd = class_exp2obd(exp);
1401         struct lmv_obd          *lmv = &obd->u.lmv;
1402         struct obd_statfs       *temp;
1403         int                      rc = 0;
1404         __u32                    i, idx;
1405         ENTRY;
1406
1407         OBD_ALLOC(temp, sizeof(*temp));
1408         if (temp == NULL)
1409                 RETURN(-ENOMEM);
1410
1411         /* distribute statfs among MDTs */
1412         idx = lmv_select_statfs_mdt(lmv, flags);
1413
1414         for (i = 0; i < lmv->desc.ld_tgt_count; i++, idx++) {
1415                 idx = idx % lmv->desc.ld_tgt_count;
1416                 if (lmv->tgts[idx] == NULL || lmv->tgts[idx]->ltd_exp == NULL)
1417                         continue;
1418
1419                 rc = obd_statfs(env, lmv->tgts[idx]->ltd_exp, temp,
1420                                 max_age, flags);
1421                 if (rc) {
1422                         CERROR("%s: can't stat MDS #%d: rc = %d\n",
1423                                lmv->tgts[idx]->ltd_exp->exp_obd->obd_name, i,
1424                                rc);
1425                         GOTO(out_free_temp, rc);
1426                 }
1427
1428                 if (temp->os_state & OS_STATE_SUM ||
1429                     flags == OBD_STATFS_FOR_MDT0) {
1430                         /* reset to the last aggregated values
1431                          * and don't sum with non-aggrated data */
1432                         /* If the statfs is from mount, it needs to retrieve
1433                          * necessary information from MDT0. i.e. mount does
1434                          * not need the merged osfs from all of MDT. Also
1435                          * clients can be mounted as long as MDT0 is in
1436                          * service */
1437                         *osfs = *temp;
1438                         break;
1439                 }
1440
1441                 if (i == 0) {
1442                         *osfs = *temp;
1443                 } else {
1444                         osfs->os_bavail += temp->os_bavail;
1445                         osfs->os_blocks += temp->os_blocks;
1446                         osfs->os_ffree += temp->os_ffree;
1447                         osfs->os_files += temp->os_files;
1448                         osfs->os_granted += temp->os_granted;
1449                 }
1450         }
1451
1452         EXIT;
1453 out_free_temp:
1454         OBD_FREE(temp, sizeof(*temp));
1455         return rc;
1456 }
1457
1458 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1459                         struct lu_fid *fid)
1460 {
1461         struct obd_device    *obd = exp->exp_obd;
1462         struct lmv_obd       *lmv = &obd->u.lmv;
1463         int                   rc;
1464         ENTRY;
1465
1466         rc = md_get_root(lmv->tgts[0]->ltd_exp, fileset, fid);
1467         RETURN(rc);
1468 }
1469
1470 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1471                         u64 obd_md_valid, const char *name, size_t buf_size,
1472                         struct ptlrpc_request **req)
1473 {
1474         struct obd_device      *obd = exp->exp_obd;
1475         struct lmv_obd         *lmv = &obd->u.lmv;
1476         struct lmv_tgt_desc    *tgt;
1477         int                     rc;
1478         ENTRY;
1479
1480         tgt = lmv_find_target(lmv, fid);
1481         if (IS_ERR(tgt))
1482                 RETURN(PTR_ERR(tgt));
1483
1484         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1485
1486         RETURN(rc);
1487 }
1488
1489 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1490                         u64 obd_md_valid, const char *name,
1491                         const void *value, size_t value_size,
1492                         unsigned int xattr_flags, u32 suppgid,
1493                         struct ptlrpc_request **req)
1494 {
1495         struct obd_device      *obd = exp->exp_obd;
1496         struct lmv_obd         *lmv = &obd->u.lmv;
1497         struct lmv_tgt_desc    *tgt;
1498         int                     rc;
1499         ENTRY;
1500
1501         tgt = lmv_find_target(lmv, fid);
1502         if (IS_ERR(tgt))
1503                 RETURN(PTR_ERR(tgt));
1504
1505         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1506                          value, value_size, xattr_flags, suppgid, req);
1507
1508         RETURN(rc);
1509 }
1510
1511 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1512                        struct ptlrpc_request **request)
1513 {
1514         struct obd_device       *obd = exp->exp_obd;
1515         struct lmv_obd          *lmv = &obd->u.lmv;
1516         struct lmv_tgt_desc     *tgt;
1517         int                      rc;
1518         ENTRY;
1519
1520         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1521         if (IS_ERR(tgt))
1522                 RETURN(PTR_ERR(tgt));
1523
1524         if (op_data->op_flags & MF_GET_MDT_IDX) {
1525                 op_data->op_mds = tgt->ltd_idx;
1526                 RETURN(0);
1527         }
1528
1529         rc = md_getattr(tgt->ltd_exp, op_data, request);
1530
1531         RETURN(rc);
1532 }
1533
1534 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1535 {
1536         struct obd_device   *obd = exp->exp_obd;
1537         struct lmv_obd      *lmv = &obd->u.lmv;
1538         __u32                i;
1539         ENTRY;
1540
1541         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1542
1543         /*
1544          * With DNE every object can have two locks in different namespaces:
1545          * lookup lock in space of MDT storing direntry and update/open lock in
1546          * space of MDT storing inode.
1547          */
1548         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1549                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
1550                         continue;
1551                 md_null_inode(lmv->tgts[i]->ltd_exp, fid);
1552         }
1553
1554         RETURN(0);
1555 }
1556
1557 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1558                      struct md_open_data *mod, struct ptlrpc_request **request)
1559 {
1560         struct obd_device     *obd = exp->exp_obd;
1561         struct lmv_obd        *lmv = &obd->u.lmv;
1562         struct lmv_tgt_desc   *tgt;
1563         int                    rc;
1564         ENTRY;
1565
1566         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1567         if (IS_ERR(tgt))
1568                 RETURN(PTR_ERR(tgt));
1569
1570         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1571         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1572         RETURN(rc);
1573 }
1574
1575 struct lmv_tgt_desc*
1576 __lmv_locate_tgt(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1577                  const char *name, int namelen, struct lu_fid *fid, u32 *mds,
1578                  bool post_migrate)
1579 {
1580         struct lmv_tgt_desc *tgt;
1581         const struct lmv_oinfo *oinfo;
1582
1583         if (lsm == NULL || namelen == 0) {
1584                 tgt = lmv_find_target(lmv, fid);
1585                 if (IS_ERR(tgt))
1586                         return tgt;
1587
1588                 LASSERT(mds);
1589                 *mds = tgt->ltd_idx;
1590                 return tgt;
1591         }
1592
1593         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1594                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1595                         return ERR_PTR(-EBADF);
1596                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1597         } else {
1598                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen,
1599                                                 post_migrate);
1600                 if (IS_ERR(oinfo))
1601                         return ERR_CAST(oinfo);
1602         }
1603
1604         if (fid != NULL)
1605                 *fid = oinfo->lmo_fid;
1606         if (mds != NULL)
1607                 *mds = oinfo->lmo_mds;
1608
1609         tgt = lmv_get_target(lmv, oinfo->lmo_mds, NULL);
1610
1611         CDEBUG(D_INFO, "locate on mds %u "DFID"\n", oinfo->lmo_mds,
1612                PFID(&oinfo->lmo_fid));
1613
1614         return tgt;
1615 }
1616
1617
1618 /**
1619  * Locate mdt by fid or name
1620  *
1621  * For striped directory, it will locate the stripe by name hash, if hash_type
1622  * is unknown, it will return the stripe specified by 'op_data->op_stripe_index'
1623  * which is set outside, and if dir is migrating, 'op_data->op_post_migrate'
1624  * indicates whether old or new layout is used to locate.
1625  *
1626  * For normal direcotry, it will locate MDS by FID directly.
1627  *
1628  * \param[in] lmv       LMV device
1629  * \param[in] op_data   client MD stack parameters, name, namelen
1630  *                      mds_num etc.
1631  * \param[in] fid       object FID used to locate MDS.
1632  *
1633  * retval               pointer to the lmv_tgt_desc if succeed.
1634  *                      ERR_PTR(errno) if failed.
1635  */
1636 struct lmv_tgt_desc*
1637 lmv_locate_tgt(struct lmv_obd *lmv, struct md_op_data *op_data,
1638                struct lu_fid *fid)
1639 {
1640         struct lmv_stripe_md *lsm = op_data->op_mea1;
1641         struct lmv_oinfo *oinfo;
1642         struct lmv_tgt_desc *tgt;
1643
1644         /* foreign dir is not striped dir */
1645         if (lsm && lsm->lsm_md_magic == LMV_MAGIC_FOREIGN)
1646                 return ERR_PTR(-ENODATA);
1647
1648         /* During creating VOLATILE file, it should honor the mdt
1649          * index if the file under striped dir is being restored, see
1650          * ct_restore(). */
1651         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1652             (int)op_data->op_mds != -1) {
1653                 tgt = lmv_get_target(lmv, op_data->op_mds, NULL);
1654                 if (IS_ERR(tgt))
1655                         return tgt;
1656
1657                 if (lsm) {
1658                         int i;
1659
1660                         /* refill the right parent fid */
1661                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1662                                 oinfo = &lsm->lsm_md_oinfo[i];
1663                                 if (oinfo->lmo_mds == op_data->op_mds) {
1664                                         *fid = oinfo->lmo_fid;
1665                                         break;
1666                                 }
1667                         }
1668
1669                         if (i == lsm->lsm_md_stripe_count)
1670                                 *fid = lsm->lsm_md_oinfo[0].lmo_fid;
1671                 }
1672         } else if (lmv_is_dir_bad_hash(lsm)) {
1673                 LASSERT(op_data->op_stripe_index < lsm->lsm_md_stripe_count);
1674                 oinfo = &lsm->lsm_md_oinfo[op_data->op_stripe_index];
1675
1676                 *fid = oinfo->lmo_fid;
1677                 op_data->op_mds = oinfo->lmo_mds;
1678                 tgt = lmv_get_target(lmv, oinfo->lmo_mds, NULL);
1679         } else {
1680                 tgt = __lmv_locate_tgt(lmv, lsm, op_data->op_name,
1681                                        op_data->op_namelen, fid,
1682                                        &op_data->op_mds,
1683                                        op_data->op_post_migrate);
1684         }
1685
1686         return tgt;
1687 }
1688
1689 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1690                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1691                 gid_t gid, cfs_cap_t cap_effective, __u64 rdev,
1692                 struct ptlrpc_request **request)
1693 {
1694         struct obd_device       *obd = exp->exp_obd;
1695         struct lmv_obd          *lmv = &obd->u.lmv;
1696         struct lmv_tgt_desc     *tgt;
1697         int                      rc;
1698         ENTRY;
1699
1700         if (!lmv->desc.ld_active_tgt_count)
1701                 RETURN(-EIO);
1702
1703         if (lmv_is_dir_bad_hash(op_data->op_mea1))
1704                 RETURN(-EBADF);
1705
1706         if (lmv_is_dir_migrating(op_data->op_mea1)) {
1707                 /*
1708                  * if parent is migrating, create() needs to lookup existing
1709                  * name, to avoid creating new file under old layout of
1710                  * migrating directory, check old layout here.
1711                  */
1712                 tgt = lmv_locate_tgt(lmv, op_data, &op_data->op_fid1);
1713                 if (IS_ERR(tgt))
1714                         RETURN(PTR_ERR(tgt));
1715
1716                 rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1717                 if (!rc) {
1718                         ptlrpc_req_finished(*request);
1719                         *request = NULL;
1720                         RETURN(-EEXIST);
1721                 }
1722
1723                 if (rc != -ENOENT)
1724                         RETURN(rc);
1725
1726                 op_data->op_post_migrate = true;
1727         }
1728
1729         tgt = lmv_locate_tgt(lmv, op_data, &op_data->op_fid1);
1730         if (IS_ERR(tgt))
1731                 RETURN(PTR_ERR(tgt));
1732
1733         CDEBUG(D_INODE, "CREATE name '%.*s' on "DFID" -> mds #%x\n",
1734                 (int)op_data->op_namelen, op_data->op_name,
1735                 PFID(&op_data->op_fid1), op_data->op_mds);
1736
1737         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1738         if (rc)
1739                 RETURN(rc);
1740
1741         if (exp_connect_flags(exp) & OBD_CONNECT_DIR_STRIPE) {
1742                 /* Send the create request to the MDT where the object
1743                  * will be located */
1744                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
1745                 if (IS_ERR(tgt))
1746                         RETURN(PTR_ERR(tgt));
1747
1748                 op_data->op_mds = tgt->ltd_idx;
1749         } else {
1750                 CDEBUG(D_CONFIG, "Server doesn't support striped dirs\n");
1751         }
1752
1753         CDEBUG(D_INODE, "CREATE obj "DFID" -> mds #%x\n",
1754                PFID(&op_data->op_fid2), op_data->op_mds);
1755
1756         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1757         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1758                        cap_effective, rdev, request);
1759         if (rc == 0) {
1760                 if (*request == NULL)
1761                         RETURN(rc);
1762                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1763         }
1764         RETURN(rc);
1765 }
1766
1767 static int
1768 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1769             const union ldlm_policy_data *policy, struct md_op_data *op_data,
1770             struct lustre_handle *lockh, __u64 extra_lock_flags)
1771 {
1772         struct obd_device        *obd = exp->exp_obd;
1773         struct lmv_obd           *lmv = &obd->u.lmv;
1774         struct lmv_tgt_desc      *tgt;
1775         int                       rc;
1776         ENTRY;
1777
1778         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
1779
1780         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1781         if (IS_ERR(tgt))
1782                 RETURN(PTR_ERR(tgt));
1783
1784         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
1785                PFID(&op_data->op_fid1), tgt->ltd_idx);
1786
1787         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
1788                         extra_lock_flags);
1789
1790         RETURN(rc);
1791 }
1792
1793 int
1794 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1795                  struct ptlrpc_request **preq)
1796 {
1797         struct obd_device *obd = exp->exp_obd;
1798         struct lmv_obd *lmv = &obd->u.lmv;
1799         struct lmv_tgt_desc *tgt;
1800         struct mdt_body *body;
1801         int rc;
1802
1803         ENTRY;
1804
1805 retry:
1806         tgt = lmv_locate_tgt(lmv, op_data, &op_data->op_fid1);
1807         if (IS_ERR(tgt))
1808                 RETURN(PTR_ERR(tgt));
1809
1810         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1811                 (int)op_data->op_namelen, op_data->op_name,
1812                 PFID(&op_data->op_fid1), tgt->ltd_idx);
1813
1814         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
1815         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
1816                 ptlrpc_req_finished(*preq);
1817                 *preq = NULL;
1818                 goto retry;
1819         }
1820
1821         if (rc)
1822                 RETURN(rc);
1823
1824         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
1825         LASSERT(body != NULL);
1826
1827         if (body->mbo_valid & OBD_MD_MDS) {
1828                 op_data->op_fid1 = body->mbo_fid1;
1829                 op_data->op_valid |= OBD_MD_FLCROSSREF;
1830                 op_data->op_namelen = 0;
1831                 op_data->op_name = NULL;
1832
1833                 ptlrpc_req_finished(*preq);
1834                 *preq = NULL;
1835
1836                 goto retry;
1837         }
1838
1839         RETURN(rc);
1840 }
1841
1842 #define md_op_data_fid(op_data, fl)                     \
1843         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1844          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1845          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1846          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1847          NULL)
1848
1849 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
1850                             struct md_op_data *op_data, __u32 op_tgt,
1851                             enum ldlm_mode mode, int bits, int flag)
1852 {
1853         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1854         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
1855         union ldlm_policy_data policy = { { 0 } };
1856         int rc = 0;
1857         ENTRY;
1858
1859         if (!fid_is_sane(fid))
1860                 RETURN(0);
1861
1862         if (tgt == NULL) {
1863                 tgt = lmv_find_target(lmv, fid);
1864                 if (IS_ERR(tgt))
1865                         RETURN(PTR_ERR(tgt));
1866         }
1867
1868         if (tgt->ltd_idx != op_tgt) {
1869                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1870                 policy.l_inodebits.bits = bits;
1871                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1872                                       mode, LCF_ASYNC, NULL);
1873         } else {
1874                 CDEBUG(D_INODE,
1875                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
1876                        op_tgt, PFID(fid));
1877                 op_data->op_flags |= flag;
1878                 rc = 0;
1879         }
1880
1881         RETURN(rc);
1882 }
1883
1884 /*
1885  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1886  * op_data->op_fid2
1887  */
1888 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1889                     struct ptlrpc_request **request)
1890 {
1891         struct obd_device       *obd = exp->exp_obd;
1892         struct lmv_obd          *lmv = &obd->u.lmv;
1893         struct lmv_tgt_desc     *tgt;
1894         int                      rc;
1895         ENTRY;
1896
1897         LASSERT(op_data->op_namelen != 0);
1898
1899         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
1900                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
1901                op_data->op_name, PFID(&op_data->op_fid1));
1902
1903         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
1904         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
1905         op_data->op_cap = cfs_curproc_cap_pack();
1906
1907         if (lmv_is_dir_migrating(op_data->op_mea2)) {
1908                 struct lu_fid fid1 = op_data->op_fid1;
1909                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
1910
1911                 /*
1912                  * avoid creating new file under old layout of migrating
1913                  * directory, check it here.
1914                  */
1915                 tgt = __lmv_locate_tgt(lmv, op_data->op_mea2, op_data->op_name,
1916                                        op_data->op_namelen, &op_data->op_fid2,
1917                                        &op_data->op_mds, false);
1918                 tgt = lmv_locate_tgt(lmv, op_data, &op_data->op_fid1);
1919                 if (IS_ERR(tgt))
1920                         RETURN(PTR_ERR(tgt));
1921
1922                 op_data->op_fid1 = op_data->op_fid2;
1923                 op_data->op_mea1 = op_data->op_mea2;
1924                 rc = md_getattr_name(tgt->ltd_exp, op_data, request);
1925                 op_data->op_fid1 = fid1;
1926                 op_data->op_mea1 = lsm1;
1927                 if (!rc) {
1928                         ptlrpc_req_finished(*request);
1929                         *request = NULL;
1930                         RETURN(-EEXIST);
1931                 }
1932
1933                 if (rc != -ENOENT)
1934                         RETURN(rc);
1935         }
1936
1937         tgt = __lmv_locate_tgt(lmv, op_data->op_mea2, op_data->op_name,
1938                                op_data->op_namelen, &op_data->op_fid2,
1939                                &op_data->op_mds, true);
1940         if (IS_ERR(tgt))
1941                 RETURN(PTR_ERR(tgt));
1942
1943         /*
1944          * Cancel UPDATE lock on child (fid1).
1945          */
1946         op_data->op_flags |= MF_MDC_CANCEL_FID2;
1947         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_idx, LCK_EX,
1948                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1949         if (rc != 0)
1950                 RETURN(rc);
1951
1952         rc = md_link(tgt->ltd_exp, op_data, request);
1953
1954         RETURN(rc);
1955 }
1956
1957 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
1958                         const char *name, size_t namelen,
1959                         struct ptlrpc_request **request)
1960 {
1961         struct obd_device *obd = exp->exp_obd;
1962         struct lmv_obd *lmv = &obd->u.lmv;
1963         struct lmv_stripe_md *lsm = op_data->op_mea1;
1964         struct lmv_tgt_desc *parent_tgt;
1965         struct lmv_tgt_desc *sp_tgt;
1966         struct lmv_tgt_desc *tp_tgt = NULL;
1967         struct lmv_tgt_desc *child_tgt;
1968         struct lmv_tgt_desc *tgt;
1969         struct lu_fid target_fid;
1970         int rc;
1971
1972         ENTRY;
1973
1974         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
1975
1976         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
1977                PFID(&op_data->op_fid1), (int)namelen, name);
1978
1979         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
1980         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
1981         op_data->op_cap = cfs_curproc_cap_pack();
1982
1983         parent_tgt = lmv_find_target(lmv, &op_data->op_fid1);
1984         if (IS_ERR(parent_tgt))
1985                 RETURN(PTR_ERR(parent_tgt));
1986
1987         if (lsm) {
1988                 __u32 hash_type = lsm->lsm_md_hash_type;
1989                 __u32 stripe_count = lsm->lsm_md_stripe_count;
1990
1991                 /*
1992                  * old stripes are appended after new stripes for migrating
1993                  * directory.
1994                  */
1995                 if (lsm->lsm_md_hash_type & LMV_HASH_FLAG_MIGRATION) {
1996                         hash_type = lsm->lsm_md_migrate_hash;
1997                         stripe_count -= lsm->lsm_md_migrate_offset;
1998                 }
1999
2000                 rc = lmv_name_to_stripe_index(hash_type, stripe_count, name,
2001                                               namelen);
2002                 if (rc < 0)
2003                         RETURN(rc);
2004
2005                 if (lsm->lsm_md_hash_type & LMV_HASH_FLAG_MIGRATION)
2006                         rc += lsm->lsm_md_migrate_offset;
2007
2008                 /* save it in fid4 temporarily for early cancel */
2009                 op_data->op_fid4 = lsm->lsm_md_oinfo[rc].lmo_fid;
2010                 sp_tgt = lmv_get_target(lmv, lsm->lsm_md_oinfo[rc].lmo_mds,
2011                                         NULL);
2012                 if (IS_ERR(sp_tgt))
2013                         RETURN(PTR_ERR(sp_tgt));
2014
2015                 /*
2016                  * if parent is being migrated too, fill op_fid2 with target
2017                  * stripe fid, otherwise the target stripe is not created yet.
2018                  */
2019                 if (lsm->lsm_md_hash_type & LMV_HASH_FLAG_MIGRATION) {
2020                         hash_type = lsm->lsm_md_hash_type &
2021                                     ~LMV_HASH_FLAG_MIGRATION;
2022                         stripe_count = lsm->lsm_md_migrate_offset;
2023
2024                         rc = lmv_name_to_stripe_index(hash_type, stripe_count,
2025                                                       name, namelen);
2026                         if (rc < 0)
2027                                 RETURN(rc);
2028
2029                         op_data->op_fid2 = lsm->lsm_md_oinfo[rc].lmo_fid;
2030                         tp_tgt = lmv_get_target(lmv,
2031                                                 lsm->lsm_md_oinfo[rc].lmo_mds,
2032                                                 NULL);
2033                         if (IS_ERR(tp_tgt))
2034                                 RETURN(PTR_ERR(tp_tgt));
2035                 }
2036         } else {
2037                 sp_tgt = parent_tgt;
2038         }
2039
2040         child_tgt = lmv_find_target(lmv, &op_data->op_fid3);
2041         if (IS_ERR(child_tgt))
2042                 RETURN(PTR_ERR(child_tgt));
2043
2044         if (!S_ISDIR(op_data->op_mode) && tp_tgt)
2045                 rc = __lmv_fid_alloc(lmv, &target_fid, tp_tgt->ltd_idx);
2046         else
2047                 rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
2048         if (rc)
2049                 RETURN(rc);
2050
2051         /*
2052          * for directory, send migrate request to the MDT where the object will
2053          * be migrated to, because we can't create a striped directory remotely.
2054          *
2055          * otherwise, send to the MDT where source is located because regular
2056          * file may open lease.
2057          *
2058          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
2059          * backward compatibility.
2060          */
2061         if (S_ISDIR(op_data->op_mode) &&
2062             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
2063                 tgt = lmv_find_target(lmv, &target_fid);
2064                 if (IS_ERR(tgt))
2065                         RETURN(PTR_ERR(tgt));
2066         } else {
2067                 tgt = child_tgt;
2068         }
2069
2070         /* cancel UPDATE lock of parent master object */
2071         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_idx, LCK_EX,
2072                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2073         if (rc)
2074                 RETURN(rc);
2075
2076         /* cancel UPDATE lock of source parent */
2077         if (sp_tgt != parent_tgt) {
2078                 /*
2079                  * migrate RPC packs master object FID, because we can only pack
2080                  * two FIDs in reint RPC, but MDS needs to know both source
2081                  * parent and target parent, and it will obtain them from master
2082                  * FID and LMV, the other FID in RPC is kept for target.
2083                  *
2084                  * since this FID is not passed to MDC, cancel it anyway.
2085                  */
2086                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
2087                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
2088                 if (rc)
2089                         RETURN(rc);
2090
2091                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2092         }
2093         op_data->op_fid4 = target_fid;
2094
2095         /* cancel UPDATE locks of target parent */
2096         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_idx, LCK_EX,
2097                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2098         if (rc)
2099                 RETURN(rc);
2100
2101         /* cancel LOOKUP lock of source if source is remote object */
2102         if (child_tgt != sp_tgt) {
2103                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_idx,
2104                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2105                                       MF_MDC_CANCEL_FID3);
2106                 if (rc)
2107                         RETURN(rc);
2108         }
2109
2110         /* cancel ELC locks of source */
2111         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_idx, LCK_EX,
2112                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2113         if (rc)
2114                 RETURN(rc);
2115
2116         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2117
2118         RETURN(rc);
2119 }
2120
2121 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2122                       const char *old, size_t oldlen,
2123                       const char *new, size_t newlen,
2124                       struct ptlrpc_request **request)
2125 {
2126         struct obd_device *obd = exp->exp_obd;
2127         struct lmv_obd *lmv = &obd->u.lmv;
2128         struct lmv_tgt_desc *sp_tgt;
2129         struct lmv_tgt_desc *tp_tgt = NULL;
2130         struct lmv_tgt_desc *src_tgt = NULL;
2131         struct lmv_tgt_desc *tgt;
2132         struct mdt_body *body;
2133         int rc;
2134
2135         ENTRY;
2136
2137         LASSERT(oldlen != 0);
2138
2139         if (op_data->op_cli_flags & CLI_MIGRATE) {
2140                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2141                 RETURN(rc);
2142         }
2143
2144         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2145         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2146         op_data->op_cap = cfs_curproc_cap_pack();
2147
2148         if (lmv_is_dir_migrating(op_data->op_mea2)) {
2149                 struct lu_fid fid1 = op_data->op_fid1;
2150                 struct lmv_stripe_md *lsm1 = op_data->op_mea1;
2151
2152                 /*
2153                  * we avoid creating new file under old layout of migrating
2154                  * directory, if there is an existing file with new name under
2155                  * old layout, we can't unlink file in old layout and rename to
2156                  * new layout in one transaction, so return -EBUSY here.`
2157                  */
2158                 tgt = __lmv_locate_tgt(lmv, op_data->op_mea2, new, newlen,
2159                                        &op_data->op_fid2, &op_data->op_mds,
2160                                        false);
2161                 if (IS_ERR(tgt))
2162                         RETURN(PTR_ERR(tgt));
2163
2164                 op_data->op_fid1 = op_data->op_fid2;
2165                 op_data->op_mea1 = op_data->op_mea2;
2166                 op_data->op_name = new;
2167                 op_data->op_namelen = newlen;
2168                 rc = md_getattr_name(tgt->ltd_exp, op_data, request);
2169                 op_data->op_fid1 = fid1;
2170                 op_data->op_mea1 = lsm1;
2171                 op_data->op_name = NULL;
2172                 op_data->op_namelen = 0;
2173                 if (!rc) {
2174                         ptlrpc_req_finished(*request);
2175                         *request = NULL;
2176                         RETURN(-EBUSY);
2177                 }
2178
2179                 if (rc != -ENOENT)
2180                         RETURN(rc);
2181         }
2182
2183         /* rename to new layout for migrating directory */
2184         tp_tgt = __lmv_locate_tgt(lmv, op_data->op_mea2, new, newlen,
2185                                   &op_data->op_fid2, &op_data->op_mds, true);
2186         if (IS_ERR(tp_tgt))
2187                 RETURN(PTR_ERR(tp_tgt));
2188
2189         /* Since the target child might be destroyed, and it might become
2190          * orphan, and we can only check orphan on the local MDT right now, so
2191          * we send rename request to the MDT where target child is located. If
2192          * target child does not exist, then it will send the request to the
2193          * target parent */
2194         if (fid_is_sane(&op_data->op_fid4)) {
2195                 tgt = lmv_find_target(lmv, &op_data->op_fid4);
2196                 if (IS_ERR(tgt))
2197                         RETURN(PTR_ERR(tgt));
2198         } else {
2199                 tgt = tp_tgt;
2200         }
2201
2202         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2203
2204         /* cancel UPDATE locks of target parent */
2205         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_idx, LCK_EX,
2206                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2207         if (rc != 0)
2208                 RETURN(rc);
2209
2210         if (fid_is_sane(&op_data->op_fid4)) {
2211                 /* cancel LOOKUP lock of target on target parent */
2212                 if (tgt != tp_tgt) {
2213                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2214                                               tgt->ltd_idx, LCK_EX,
2215                                               MDS_INODELOCK_LOOKUP,
2216                                               MF_MDC_CANCEL_FID4);
2217                         if (rc != 0)
2218                                 RETURN(rc);
2219                 }
2220         }
2221
2222         if (fid_is_sane(&op_data->op_fid3)) {
2223                 src_tgt = lmv_find_target(lmv, &op_data->op_fid3);
2224                 if (IS_ERR(src_tgt))
2225                         RETURN(PTR_ERR(src_tgt));
2226
2227                 /* cancel ELC locks of source */
2228                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_idx,
2229                                       LCK_EX, MDS_INODELOCK_ELC,
2230                                       MF_MDC_CANCEL_FID3);
2231                 if (rc != 0)
2232                         RETURN(rc);
2233         }
2234
2235 retry:
2236         sp_tgt = __lmv_locate_tgt(lmv, op_data->op_mea1, old, oldlen,
2237                                   &op_data->op_fid1, &op_data->op_mds,
2238                                   op_data->op_post_migrate);
2239         if (IS_ERR(sp_tgt))
2240                 RETURN(PTR_ERR(sp_tgt));
2241
2242         /* cancel UPDATE locks of source parent */
2243         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_idx, LCK_EX,
2244                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2245         if (rc != 0)
2246                 RETURN(rc);
2247
2248         if (fid_is_sane(&op_data->op_fid3)) {
2249                 /* cancel LOOKUP lock of source on source parent */
2250                 if (src_tgt != sp_tgt) {
2251                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2252                                               tgt->ltd_idx, LCK_EX,
2253                                               MDS_INODELOCK_LOOKUP,
2254                                               MF_MDC_CANCEL_FID3);
2255                         if (rc != 0)
2256                                 RETURN(rc);
2257                 }
2258         }
2259
2260 rename:
2261         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2262                 PFID(&op_data->op_fid1), (int)oldlen, old,
2263                 PFID(&op_data->op_fid2), (int)newlen, new);
2264
2265         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2266                         request);
2267         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2268                 ptlrpc_req_finished(*request);
2269                 *request = NULL;
2270                 goto retry;
2271         }
2272
2273         if (rc && rc != -EXDEV)
2274                 RETURN(rc);
2275
2276         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2277         if (body == NULL)
2278                 RETURN(-EPROTO);
2279
2280         /* Not cross-ref case, just get out of here. */
2281         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2282                 RETURN(rc);
2283
2284         op_data->op_fid4 = body->mbo_fid1;
2285
2286         ptlrpc_req_finished(*request);
2287         *request = NULL;
2288
2289         tgt = lmv_find_target(lmv, &op_data->op_fid4);
2290         if (IS_ERR(tgt))
2291                 RETURN(PTR_ERR(tgt));
2292
2293         if (fid_is_sane(&op_data->op_fid4)) {
2294                 /* cancel LOOKUP lock of target on target parent */
2295                 if (tgt != tp_tgt) {
2296                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2297                                               tgt->ltd_idx, LCK_EX,
2298                                               MDS_INODELOCK_LOOKUP,
2299                                               MF_MDC_CANCEL_FID4);
2300                         if (rc != 0)
2301                                 RETURN(rc);
2302                 }
2303         }
2304
2305         goto rename;
2306 }
2307
2308 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2309                        void *ea, size_t ealen, struct ptlrpc_request **request)
2310 {
2311         struct obd_device       *obd = exp->exp_obd;
2312         struct lmv_obd          *lmv = &obd->u.lmv;
2313         struct lmv_tgt_desc     *tgt;
2314         int                      rc = 0;
2315         ENTRY;
2316
2317         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x/0x%x\n",
2318                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid,
2319                op_data->op_xvalid);
2320
2321         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2322         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2323         if (IS_ERR(tgt))
2324                 RETURN(PTR_ERR(tgt));
2325
2326         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2327
2328         RETURN(rc);
2329 }
2330
2331 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2332                      struct ptlrpc_request **request)
2333 {
2334         struct obd_device       *obd = exp->exp_obd;
2335         struct lmv_obd          *lmv = &obd->u.lmv;
2336         struct lmv_tgt_desc     *tgt;
2337         int                      rc;
2338         ENTRY;
2339
2340         tgt = lmv_find_target(lmv, fid);
2341         if (IS_ERR(tgt))
2342                 RETURN(PTR_ERR(tgt));
2343
2344         rc = md_fsync(tgt->ltd_exp, fid, request);
2345         RETURN(rc);
2346 }
2347
2348 struct stripe_dirent {
2349         struct page             *sd_page;
2350         struct lu_dirpage       *sd_dp;
2351         struct lu_dirent        *sd_ent;
2352         bool                     sd_eof;
2353 };
2354
2355 struct lmv_dir_ctxt {
2356         struct lmv_obd          *ldc_lmv;
2357         struct md_op_data       *ldc_op_data;
2358         struct md_callback      *ldc_cb_op;
2359         __u64                    ldc_hash;
2360         int                      ldc_count;
2361         struct stripe_dirent     ldc_stripes[0];
2362 };
2363
2364 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2365 {
2366         if (stripe->sd_page) {
2367                 kunmap(stripe->sd_page);
2368                 put_page(stripe->sd_page);
2369                 stripe->sd_page = NULL;
2370                 stripe->sd_ent = NULL;
2371         }
2372 }
2373
2374 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2375 {
2376         int i;
2377
2378         for (i = 0; i < ctxt->ldc_count; i++)
2379                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2380 }
2381
2382 /* if @ent is dummy, or . .., get next */
2383 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2384                                            struct lu_dirent *ent,
2385                                            int stripe_index)
2386 {
2387         for (; ent; ent = lu_dirent_next(ent)) {
2388                 /* Skip dummy entry */
2389                 if (le16_to_cpu(ent->lde_namelen) == 0)
2390                         continue;
2391
2392                 /* skip . and .. for other stripes */
2393                 if (stripe_index &&
2394                     (strncmp(ent->lde_name, ".",
2395                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2396                      strncmp(ent->lde_name, "..",
2397                              le16_to_cpu(ent->lde_namelen)) == 0))
2398                         continue;
2399
2400                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2401                         break;
2402         }
2403
2404         return ent;
2405 }
2406
2407 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2408                                             struct stripe_dirent *stripe,
2409                                             int stripe_index)
2410 {
2411         struct md_op_data *op_data = ctxt->ldc_op_data;
2412         struct lmv_oinfo *oinfo;
2413         struct lu_fid fid = op_data->op_fid1;
2414         struct inode *inode = op_data->op_data;
2415         struct lmv_tgt_desc *tgt;
2416         struct lu_dirent *ent = stripe->sd_ent;
2417         __u64 hash = ctxt->ldc_hash;
2418         int rc = 0;
2419
2420         ENTRY;
2421
2422         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2423         LASSERT(!ent);
2424
2425         do {
2426                 if (stripe->sd_page) {
2427                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2428
2429                         /* @hash should be the last dirent hash */
2430                         LASSERTF(hash <= end,
2431                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2432                                  ctxt, stripe, hash, end);
2433                         /* unload last page */
2434                         stripe_dirent_unload(stripe);
2435                         /* eof */
2436                         if (end == MDS_DIR_END_OFF) {
2437                                 stripe->sd_eof = true;
2438                                 break;
2439                         }
2440                         hash = end;
2441                 }
2442
2443                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2444                 if (!oinfo->lmo_root) {
2445                         rc = -ENOENT;
2446                         break;
2447                 }
2448
2449                 tgt = lmv_get_target(ctxt->ldc_lmv, oinfo->lmo_mds, NULL);
2450                 if (IS_ERR(tgt)) {
2451                         rc = PTR_ERR(tgt);
2452                         break;
2453                 }
2454
2455                 /* op_data is shared by stripes, reset after use */
2456                 op_data->op_fid1 = oinfo->lmo_fid;
2457                 op_data->op_fid2 = oinfo->lmo_fid;
2458                 op_data->op_data = oinfo->lmo_root;
2459
2460                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2461                                   &stripe->sd_page);
2462
2463                 op_data->op_fid1 = fid;
2464                 op_data->op_fid2 = fid;
2465                 op_data->op_data = inode;
2466
2467                 if (rc)
2468                         break;
2469
2470                 stripe->sd_dp = page_address(stripe->sd_page);
2471                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2472                                         stripe_index);
2473                 /* in case a page filled with ., .. and dummy, read next */
2474         } while (!ent);
2475
2476         stripe->sd_ent = ent;
2477         if (rc) {
2478                 LASSERT(!ent);
2479                 /* treat error as eof, so dir can be partially accessed */
2480                 stripe->sd_eof = true;
2481                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2482                               "directory is partially accessed!\n",
2483                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2484                               rc);
2485         }
2486
2487         RETURN(ent);
2488 }
2489
2490 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2491 {
2492         struct obd_device       *obd = exp->exp_obd;
2493         struct lmv_obd          *lmv = &obd->u.lmv;
2494         struct lmv_tgt_desc     *tgt;
2495         int                      rc;
2496         ENTRY;
2497
2498         rc = lmv_check_connect(obd);
2499         if (rc != 0)
2500                 RETURN(rc);
2501
2502         tgt = lmv_find_target(lmv, &data->op_fid1);
2503         if (IS_ERR(tgt))
2504                 RETURN(PTR_ERR(tgt));
2505
2506         data->op_flags |= MF_MDC_CANCEL_FID1;
2507         rc = md_file_resync(tgt->ltd_exp, data);
2508         RETURN(rc);
2509 }
2510
2511 /**
2512  * Get dirent with the closest hash for striped directory
2513  *
2514  * This function will search the dir entry, whose hash value is the
2515  * closest(>=) to hash from all of sub-stripes, and it is only being called
2516  * for striped directory.
2517  *
2518  * \param[in] ctxt              dir read context
2519  *
2520  * \retval                      dirent get the entry successfully
2521  *                              NULL does not get the entry, normally it means
2522  *                              it reaches the end of the directory, while read
2523  *                              stripe dirent error is ignored to allow partial
2524  *                              access.
2525  */
2526 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2527 {
2528         struct stripe_dirent *stripe;
2529         struct lu_dirent *ent = NULL;
2530         int i;
2531         int min = -1;
2532
2533         /* TODO: optimize with k-way merge sort */
2534         for (i = 0; i < ctxt->ldc_count; i++) {
2535                 stripe = &ctxt->ldc_stripes[i];
2536                 if (stripe->sd_eof)
2537                         continue;
2538
2539                 if (!stripe->sd_ent) {
2540                         stripe_dirent_load(ctxt, stripe, i);
2541                         if (!stripe->sd_ent) {
2542                                 LASSERT(stripe->sd_eof);
2543                                 continue;
2544                         }
2545                 }
2546
2547                 if (min == -1 ||
2548                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2549                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2550                         min = i;
2551                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2552                             ctxt->ldc_hash)
2553                                 break;
2554                 }
2555         }
2556
2557         if (min != -1) {
2558                 stripe = &ctxt->ldc_stripes[min];
2559                 ent = stripe->sd_ent;
2560                 /* pop found dirent */
2561                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2562                                                    min);
2563         }
2564
2565         return ent;
2566 }
2567
2568 /**
2569  * Build dir entry page for striped directory
2570  *
2571  * This function gets one entry by @offset from a striped directory. It will
2572  * read entries from all of stripes, and choose one closest to the required
2573  * offset(&offset). A few notes
2574  * 1. skip . and .. for non-zero stripes, because there can only have one .
2575  * and .. in a directory.
2576  * 2. op_data will be shared by all of stripes, instead of allocating new
2577  * one, so need to restore before reusing.
2578  *
2579  * \param[in] exp       obd export refer to LMV
2580  * \param[in] op_data   hold those MD parameters of read_entry
2581  * \param[in] cb_op     ldlm callback being used in enqueue in mdc_read_entry
2582  * \param[in] offset    starting hash offset
2583  * \param[out] ppage    the page holding the entry. Note: because the entry
2584  *                      will be accessed in upper layer, so we need hold the
2585  *                      page until the usages of entry is finished, see
2586  *                      ll_dir_entry_next.
2587  *
2588  * retval               =0 if get entry successfully
2589  *                      <0 cannot get entry
2590  */
2591 static int lmv_striped_read_page(struct obd_export *exp,
2592                                  struct md_op_data *op_data,
2593                                  struct md_callback *cb_op,
2594                                  __u64 offset, struct page **ppage)
2595 {
2596         struct page *page = NULL;
2597         struct lu_dirpage *dp;
2598         void *start;
2599         struct lu_dirent *ent;
2600         struct lu_dirent *last_ent;
2601         int stripe_count;
2602         struct lmv_dir_ctxt *ctxt;
2603         struct lu_dirent *next = NULL;
2604         __u16 ent_size;
2605         size_t left_bytes;
2606         int rc = 0;
2607         ENTRY;
2608
2609         /* Allocate a page and read entries from all of stripes and fill
2610          * the page by hash order */
2611         page = alloc_page(GFP_KERNEL);
2612         if (!page)
2613                 RETURN(-ENOMEM);
2614
2615         /* Initialize the entry page */
2616         dp = kmap(page);
2617         memset(dp, 0, sizeof(*dp));
2618         dp->ldp_hash_start = cpu_to_le64(offset);
2619
2620         start = dp + 1;
2621         left_bytes = PAGE_SIZE - sizeof(*dp);
2622         ent = start;
2623         last_ent = ent;
2624
2625         /* initalize dir read context */
2626         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2627         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2628         if (!ctxt)
2629                 GOTO(free_page, rc = -ENOMEM);
2630         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2631         ctxt->ldc_op_data = op_data;
2632         ctxt->ldc_cb_op = cb_op;
2633         ctxt->ldc_hash = offset;
2634         ctxt->ldc_count = stripe_count;
2635
2636         while (1) {
2637                 next = lmv_dirent_next(ctxt);
2638
2639                 /* end of directory */
2640                 if (!next) {
2641                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2642                         break;
2643                 }
2644                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2645
2646                 ent_size = le16_to_cpu(next->lde_reclen);
2647
2648                 /* the last entry lde_reclen is 0, but it might not be the last
2649                  * one of this temporay dir page */
2650                 if (!ent_size)
2651                         ent_size = lu_dirent_calc_size(
2652                                         le16_to_cpu(next->lde_namelen),
2653                                         le32_to_cpu(next->lde_attrs));
2654                 /* page full */
2655                 if (ent_size > left_bytes)
2656                         break;
2657
2658                 memcpy(ent, next, ent_size);
2659
2660                 /* Replace . with master FID and Replace .. with the parent FID
2661                  * of master object */
2662                 if (strncmp(ent->lde_name, ".",
2663                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2664                     le16_to_cpu(ent->lde_namelen) == 1)
2665                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2666                 else if (strncmp(ent->lde_name, "..",
2667                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2668                            le16_to_cpu(ent->lde_namelen) == 2)
2669                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2670
2671                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2672                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2673                        le64_to_cpu(ent->lde_hash));
2674
2675                 left_bytes -= ent_size;
2676                 ent->lde_reclen = cpu_to_le16(ent_size);
2677                 last_ent = ent;
2678                 ent = (void *)ent + ent_size;
2679         };
2680
2681         last_ent->lde_reclen = 0;
2682
2683         if (ent == start)
2684                 dp->ldp_flags |= LDF_EMPTY;
2685         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2686                 dp->ldp_flags |= LDF_COLLIDE;
2687         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2688         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2689
2690         put_lmv_dir_ctxt(ctxt);
2691         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2692
2693         *ppage = page;
2694
2695         RETURN(0);
2696
2697 free_page:
2698         kunmap(page);
2699         __free_page(page);
2700
2701         return rc;
2702 }
2703
2704 int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2705                   struct md_callback *cb_op, __u64 offset,
2706                   struct page **ppage)
2707 {
2708         struct obd_device       *obd = exp->exp_obd;
2709         struct lmv_obd          *lmv = &obd->u.lmv;
2710         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2711         struct lmv_tgt_desc     *tgt;
2712         int                     rc;
2713         ENTRY;
2714
2715         if (unlikely(lsm != NULL)) {
2716                 /* foreign dir is not striped dir */
2717                 if (lsm->lsm_md_magic == LMV_MAGIC_FOREIGN)
2718                         return -ENODATA;
2719
2720                 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2721                 RETURN(rc);
2722         }
2723
2724         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2725         if (IS_ERR(tgt))
2726                 RETURN(PTR_ERR(tgt));
2727
2728         rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2729
2730         RETURN(rc);
2731 }
2732
2733 /**
2734  * Unlink a file/directory
2735  *
2736  * Unlink a file or directory under the parent dir. The unlink request
2737  * usually will be sent to the MDT where the child is located, but if
2738  * the client does not have the child FID then request will be sent to the
2739  * MDT where the parent is located.
2740  *
2741  * If the parent is a striped directory then it also needs to locate which
2742  * stripe the name of the child is located, and replace the parent FID
2743  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2744  * it will walk through all of sub-stripes until the child is being
2745  * unlinked finally.
2746  *
2747  * \param[in] exp       export refer to LMV
2748  * \param[in] op_data   different parameters transferred beween client
2749  *                      MD stacks, name, namelen, FIDs etc.
2750  *                      op_fid1 is the parent FID, op_fid2 is the child
2751  *                      FID.
2752  * \param[out] request  point to the request of unlink.
2753  *
2754  * retval               0 if succeed
2755  *                      negative errno if failed.
2756  */
2757 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2758                       struct ptlrpc_request **request)
2759 {
2760         struct obd_device *obd = exp->exp_obd;
2761         struct lmv_obd *lmv = &obd->u.lmv;
2762         struct lmv_tgt_desc *tgt;
2763         struct lmv_tgt_desc *parent_tgt;
2764         struct mdt_body *body;
2765         int rc;
2766
2767         ENTRY;
2768
2769         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2770         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2771         op_data->op_cap = cfs_curproc_cap_pack();
2772
2773 retry:
2774         parent_tgt = lmv_locate_tgt(lmv, op_data, &op_data->op_fid1);
2775         if (IS_ERR(parent_tgt))
2776                 RETURN(PTR_ERR(parent_tgt));
2777
2778         if (likely(!fid_is_zero(&op_data->op_fid2))) {
2779                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2780                 if (IS_ERR(tgt))
2781                         RETURN(PTR_ERR(tgt));
2782         } else {
2783                 tgt = parent_tgt;
2784         }
2785
2786         /*
2787          * If child's fid is given, cancel unused locks for it if it is from
2788          * another export than parent.
2789          *
2790          * LOOKUP lock for child (fid3) should also be cancelled on parent
2791          * tgt_tgt in mdc_unlink().
2792          */
2793         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2794
2795         if (parent_tgt != tgt)
2796                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_idx,
2797                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2798                                       MF_MDC_CANCEL_FID3);
2799
2800         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_idx, LCK_EX,
2801                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2802         if (rc)
2803                 RETURN(rc);
2804
2805         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
2806                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2), tgt->ltd_idx);
2807
2808         rc = md_unlink(tgt->ltd_exp, op_data, request);
2809         if (rc == -ENOENT && lmv_dir_retry_check_update(op_data)) {
2810                 ptlrpc_req_finished(*request);
2811                 *request = NULL;
2812                 goto retry;
2813         }
2814
2815         if (rc != -EREMOTE)
2816                 RETURN(rc);
2817
2818         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2819         if (body == NULL)
2820                 RETURN(-EPROTO);
2821
2822         /* Not cross-ref case, just get out of here. */
2823         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2824                 RETURN(rc);
2825
2826         /* This is a remote object, try remote MDT. */
2827         op_data->op_fid2 = body->mbo_fid1;
2828         ptlrpc_req_finished(*request);
2829         *request = NULL;
2830
2831         tgt = lmv_find_target(lmv, &op_data->op_fid2);
2832         if (IS_ERR(tgt))
2833                 RETURN(PTR_ERR(tgt));
2834
2835         goto retry;
2836 }
2837
2838 static int lmv_precleanup(struct obd_device *obd)
2839 {
2840         ENTRY;
2841         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
2842         fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
2843         lprocfs_obd_cleanup(obd);
2844         lprocfs_free_md_stats(obd);
2845         RETURN(0);
2846 }
2847
2848 /**
2849  * Get by key a value associated with a LMV device.
2850  *
2851  * Dispatch request to lower-layer devices as needed.
2852  *
2853  * \param[in] env               execution environment for this thread
2854  * \param[in] exp               export for the LMV device
2855  * \param[in] keylen            length of key identifier
2856  * \param[in] key               identifier of key to get value for
2857  * \param[in] vallen            size of \a val
2858  * \param[out] val              pointer to storage location for value
2859  * \param[in] lsm               optional striping metadata of object
2860  *
2861  * \retval 0            on success
2862  * \retval negative     negated errno on failure
2863  */
2864 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2865                         __u32 keylen, void *key, __u32 *vallen, void *val)
2866 {
2867         struct obd_device       *obd;
2868         struct lmv_obd          *lmv;
2869         int                      rc = 0;
2870         ENTRY;
2871
2872         obd = class_exp2obd(exp);
2873         if (obd == NULL) {
2874                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2875                        exp->exp_handle.h_cookie);
2876                 RETURN(-EINVAL);
2877         }
2878
2879         lmv = &obd->u.lmv;
2880         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2881                 int i;
2882
2883                 LASSERT(*vallen == sizeof(__u32));
2884                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2885                         struct lmv_tgt_desc *tgt = lmv->tgts[i];
2886                         /*
2887                          * All tgts should be connected when this gets called.
2888                          */
2889                         if (tgt == NULL || tgt->ltd_exp == NULL)
2890                                 continue;
2891
2892                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2893                                           vallen, val))
2894                                 RETURN(0);
2895                 }
2896                 RETURN(-EINVAL);
2897         } else if (KEY_IS(KEY_MAX_EASIZE) ||
2898                    KEY_IS(KEY_DEFAULT_EASIZE) ||
2899                    KEY_IS(KEY_CONN_DATA)) {
2900                 /*
2901                  * Forwarding this request to first MDS, it should know LOV
2902                  * desc.
2903                  */
2904                 rc = obd_get_info(env, lmv->tgts[0]->ltd_exp, keylen, key,
2905                                   vallen, val);
2906                 if (!rc && KEY_IS(KEY_CONN_DATA))
2907                         exp->exp_connect_data = *(struct obd_connect_data *)val;
2908                 RETURN(rc);
2909         } else if (KEY_IS(KEY_TGT_COUNT)) {
2910                 *((int *)val) = lmv->desc.ld_tgt_count;
2911                 RETURN(0);
2912         }
2913
2914         CDEBUG(D_IOCTL, "Invalid key\n");
2915         RETURN(-EINVAL);
2916 }
2917
2918 /**
2919  * Asynchronously set by key a value associated with a LMV device.
2920  *
2921  * Dispatch request to lower-layer devices as needed.
2922  *
2923  * \param[in] env       execution environment for this thread
2924  * \param[in] exp       export for the LMV device
2925  * \param[in] keylen    length of key identifier
2926  * \param[in] key       identifier of key to store value for
2927  * \param[in] vallen    size of value to store
2928  * \param[in] val       pointer to data to be stored
2929  * \param[in] set       optional list of related ptlrpc requests
2930  *
2931  * \retval 0            on success
2932  * \retval negative     negated errno on failure
2933  */
2934 int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
2935                         __u32 keylen, void *key, __u32 vallen, void *val,
2936                         struct ptlrpc_request_set *set)
2937 {
2938         struct lmv_tgt_desc     *tgt = NULL;
2939         struct obd_device       *obd;
2940         struct lmv_obd          *lmv;
2941         int rc = 0;
2942         ENTRY;
2943
2944         obd = class_exp2obd(exp);
2945         if (obd == NULL) {
2946                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2947                        exp->exp_handle.h_cookie);
2948                 RETURN(-EINVAL);
2949         }
2950         lmv = &obd->u.lmv;
2951
2952         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2953             KEY_IS(KEY_DEFAULT_EASIZE)) {
2954                 int i, err = 0;
2955
2956                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2957                         tgt = lmv->tgts[i];
2958
2959                         if (tgt == NULL || tgt->ltd_exp == NULL)
2960                                 continue;
2961
2962                         err = obd_set_info_async(env, tgt->ltd_exp,
2963                                                  keylen, key, vallen, val, set);
2964                         if (err && rc == 0)
2965                                 rc = err;
2966                 }
2967
2968                 RETURN(rc);
2969         }
2970
2971         RETURN(-EINVAL);
2972 }
2973
2974 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
2975                             const struct lmv_mds_md_v1 *lmm1)
2976 {
2977         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
2978         int             stripe_count;
2979         int             cplen;
2980         int             i;
2981         int             rc = 0;
2982         ENTRY;
2983
2984         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
2985         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
2986         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
2987         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
2988                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
2989         else
2990                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
2991         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
2992         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
2993         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
2994         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
2995                         sizeof(lsm->lsm_md_pool_name));
2996
2997         if (cplen >= sizeof(lsm->lsm_md_pool_name))
2998                 RETURN(-E2BIG);
2999
3000         CDEBUG(D_INFO, "unpack lsm count %d, master %d hash_type %#x "
3001                "layout_version %d\n", lsm->lsm_md_stripe_count,
3002                lsm->lsm_md_master_mdt_index, lsm->lsm_md_hash_type,
3003                lsm->lsm_md_layout_version);
3004
3005         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
3006         for (i = 0; i < stripe_count; i++) {
3007                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
3008                               &lmm1->lmv_stripe_fids[i]);
3009                 /*
3010                  * set default value -1, so lmv_locate_tgt() knows this stripe
3011                  * target is not initialized.
3012                  */
3013                 lsm->lsm_md_oinfo[i].lmo_mds = (u32)-1;
3014                 if (!fid_is_sane(&lsm->lsm_md_oinfo[i].lmo_fid))
3015                         continue;
3016
3017                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
3018                                     &lsm->lsm_md_oinfo[i].lmo_mds);
3019                 if (rc == -ENOENT)
3020                         continue;
3021
3022                 if (rc)
3023                         RETURN(rc);
3024
3025                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
3026                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
3027         }
3028
3029         RETURN(rc);
3030 }
3031
3032 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
3033                         const union lmv_mds_md *lmm, size_t lmm_size)
3034 {
3035         struct lmv_stripe_md     *lsm;
3036         int                      lsm_size;
3037         int                      rc;
3038         bool                     allocated = false;
3039         ENTRY;
3040
3041         LASSERT(lsmp != NULL);
3042
3043         lsm = *lsmp;
3044         /* Free memmd */
3045         if (lsm != NULL && lmm == NULL) {
3046                 int i;
3047                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3048
3049                 if (lfm->lfm_magic == LMV_MAGIC_FOREIGN) {
3050                         size_t lfm_size;
3051
3052                         lfm_size = lfm->lfm_length + offsetof(typeof(*lfm),
3053                                                               lfm_value[0]);
3054                         OBD_FREE_LARGE(lfm, lfm_size);
3055                         RETURN(0);
3056                 }
3057
3058                 for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3059                         if (lsm->lsm_md_oinfo[i].lmo_root)
3060                                 iput(lsm->lsm_md_oinfo[i].lmo_root);
3061                 }
3062                 lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
3063                 OBD_FREE(lsm, lsm_size);
3064                 *lsmp = NULL;
3065                 RETURN(0);
3066         }
3067
3068         /* foreign lmv case */
3069         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_FOREIGN) {
3070                 struct lmv_foreign_md *lfm = (struct lmv_foreign_md *)lsm;
3071
3072                 if (lfm == NULL) {
3073                         OBD_ALLOC_LARGE(lfm, lmm_size);
3074                         if (lfm == NULL)
3075                                 RETURN(-ENOMEM);
3076                         *lsmp = (struct lmv_stripe_md *)lfm;
3077                 }
3078                 lfm->lfm_magic = le32_to_cpu(lmm->lmv_foreign_md.lfm_magic);
3079                 lfm->lfm_length = le32_to_cpu(lmm->lmv_foreign_md.lfm_length);
3080                 lfm->lfm_type = le32_to_cpu(lmm->lmv_foreign_md.lfm_type);
3081                 lfm->lfm_flags = le32_to_cpu(lmm->lmv_foreign_md.lfm_flags);
3082                 memcpy(&lfm->lfm_value, &lmm->lmv_foreign_md.lfm_value,
3083                        lfm->lfm_length);
3084                 RETURN(lmm_size);
3085         }
3086
3087         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
3088                 RETURN(-EPERM);
3089
3090         /* Unpack memmd */
3091         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
3092             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
3093                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
3094                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
3095                        -EIO);
3096                 RETURN(-EIO);
3097         }
3098
3099         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
3100                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
3101         else
3102                 /**
3103                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
3104                  * stripecount should be 0 then.
3105                  */
3106                 lsm_size = lmv_stripe_md_size(0);
3107
3108         if (lsm == NULL) {
3109                 OBD_ALLOC(lsm, lsm_size);
3110                 if (lsm == NULL)
3111                         RETURN(-ENOMEM);
3112                 allocated = true;
3113                 *lsmp = lsm;
3114         }
3115
3116         switch (le32_to_cpu(lmm->lmv_magic)) {
3117         case LMV_MAGIC_V1:
3118                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3119                 break;
3120         default:
3121                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3122                        le32_to_cpu(lmm->lmv_magic));
3123                 rc = -EINVAL;
3124                 break;
3125         }
3126
3127         if (rc != 0 && allocated) {
3128                 OBD_FREE(lsm, lsm_size);
3129                 *lsmp = NULL;
3130                 lsm_size = rc;
3131         }
3132         RETURN(lsm_size);
3133 }
3134
3135 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3136 {
3137         lmv_unpackmd(NULL, &lsm, NULL, 0);
3138 }
3139 EXPORT_SYMBOL(lmv_free_memmd);
3140
3141 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3142                              union ldlm_policy_data *policy,
3143                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3144                              void *opaque)
3145 {
3146         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3147         int rc = 0;
3148         __u32 i;
3149         ENTRY;
3150
3151         LASSERT(fid != NULL);
3152
3153         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3154                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
3155                 int err;
3156
3157                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
3158                         continue;
3159
3160                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3161                                        opaque);
3162                 if (!rc)
3163                         rc = err;
3164         }
3165         RETURN(rc);
3166 }
3167
3168 static int lmv_set_lock_data(struct obd_export *exp,
3169                              const struct lustre_handle *lockh,
3170                              void *data, __u64 *bits)
3171 {
3172         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
3173         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3174         int                      rc;
3175         ENTRY;
3176
3177         if (tgt == NULL || tgt->ltd_exp == NULL)
3178                 RETURN(-EINVAL);
3179         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3180         RETURN(rc);
3181 }
3182
3183 enum ldlm_mode lmv_lock_match(struct obd_export *exp, __u64 flags,
3184                               const struct lu_fid *fid, enum ldlm_type type,
3185                               union ldlm_policy_data *policy,
3186                               enum ldlm_mode mode, struct lustre_handle *lockh)
3187 {
3188         struct obd_device       *obd = exp->exp_obd;
3189         struct lmv_obd          *lmv = &obd->u.lmv;
3190         enum ldlm_mode          rc;
3191         int                     tgt;
3192         int                     i;
3193         ENTRY;
3194
3195         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
3196
3197         /*
3198          * With DNE every object can have two locks in different namespaces:
3199          * lookup lock in space of MDT storing direntry and update/open lock in
3200          * space of MDT storing inode.  Try the MDT that the FID maps to first,
3201          * since this can be easily found, and only try others if that fails.
3202          */
3203         for (i = 0, tgt = lmv_find_target_index(lmv, fid);
3204              i < lmv->desc.ld_tgt_count;
3205              i++, tgt = (tgt + 1) % lmv->desc.ld_tgt_count) {
3206                 if (tgt < 0) {
3207                         CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
3208                                obd->obd_name, PFID(fid), tgt);
3209                         tgt = 0;
3210                 }
3211
3212                 if (lmv->tgts[tgt] == NULL ||
3213                     lmv->tgts[tgt]->ltd_exp == NULL ||
3214                     lmv->tgts[tgt]->ltd_active == 0)
3215                         continue;
3216
3217                 rc = md_lock_match(lmv->tgts[tgt]->ltd_exp, flags, fid,
3218                                    type, policy, mode, lockh);
3219                 if (rc)
3220                         RETURN(rc);
3221         }
3222
3223         RETURN(0);
3224 }
3225
3226 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
3227                       struct obd_export *dt_exp, struct obd_export *md_exp,
3228                       struct lustre_md *md)
3229 {
3230         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
3231         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3232
3233         if (tgt == NULL || tgt->ltd_exp == NULL)
3234                 RETURN(-EINVAL);
3235
3236         return md_get_lustre_md(lmv->tgts[0]->ltd_exp, req, dt_exp, md_exp, md);
3237 }
3238
3239 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
3240 {
3241         struct obd_device       *obd = exp->exp_obd;
3242         struct lmv_obd          *lmv = &obd->u.lmv;
3243         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3244         ENTRY;
3245
3246         if (md->lmv != NULL) {
3247                 lmv_free_memmd(md->lmv);
3248                 md->lmv = NULL;
3249         }
3250         if (tgt == NULL || tgt->ltd_exp == NULL)
3251                 RETURN(-EINVAL);
3252         RETURN(md_free_lustre_md(lmv->tgts[0]->ltd_exp, md));
3253 }
3254
3255 int lmv_set_open_replay_data(struct obd_export *exp,
3256                              struct obd_client_handle *och,
3257                              struct lookup_intent *it)
3258 {
3259         struct obd_device       *obd = exp->exp_obd;
3260         struct lmv_obd          *lmv = &obd->u.lmv;
3261         struct lmv_tgt_desc     *tgt;
3262         ENTRY;
3263
3264         tgt = lmv_find_target(lmv, &och->och_fid);
3265         if (IS_ERR(tgt))
3266                 RETURN(PTR_ERR(tgt));
3267
3268         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3269 }
3270
3271 int lmv_clear_open_replay_data(struct obd_export *exp,
3272                                struct obd_client_handle *och)
3273 {
3274         struct obd_device       *obd = exp->exp_obd;
3275         struct lmv_obd          *lmv = &obd->u.lmv;
3276         struct lmv_tgt_desc     *tgt;
3277         ENTRY;
3278
3279         tgt = lmv_find_target(lmv, &och->och_fid);
3280         if (IS_ERR(tgt))
3281                 RETURN(PTR_ERR(tgt));
3282
3283         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3284 }
3285
3286 int lmv_intent_getattr_async(struct obd_export *exp,
3287                              struct md_enqueue_info *minfo)
3288 {
3289         struct md_op_data *op_data = &minfo->mi_data;
3290         struct obd_device *obd = exp->exp_obd;
3291         struct lmv_obd *lmv = &obd->u.lmv;
3292         struct lmv_tgt_desc *tgt = NULL;
3293         int rc;
3294         ENTRY;
3295
3296         if (!fid_is_sane(&op_data->op_fid2))
3297                 RETURN(-EINVAL);
3298
3299         tgt = lmv_find_target(lmv, &op_data->op_fid1);
3300         if (IS_ERR(tgt))
3301                 RETURN(PTR_ERR(tgt));
3302
3303         /*
3304          * no special handle for remote dir, which needs to fetch both LOOKUP
3305          * lock on parent, and then UPDATE lock on child MDT, which makes all
3306          * complicated because this is done async. So only LOOKUP lock is
3307          * fetched for remote dir, but considering remote dir is rare case,
3308          * and not supporting it in statahead won't cause any issue, just leave
3309          * it as is.
3310          */
3311
3312         rc = md_intent_getattr_async(tgt->ltd_exp, minfo);
3313         RETURN(rc);
3314 }
3315
3316 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3317                         struct lu_fid *fid, __u64 *bits)
3318 {
3319         struct obd_device       *obd = exp->exp_obd;
3320         struct lmv_obd          *lmv = &obd->u.lmv;
3321         struct lmv_tgt_desc     *tgt;
3322         int                      rc;
3323         ENTRY;
3324
3325         tgt = lmv_find_target(lmv, fid);
3326         if (IS_ERR(tgt))
3327                 RETURN(PTR_ERR(tgt));
3328
3329         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3330         RETURN(rc);
3331 }
3332
3333 int lmv_get_fid_from_lsm(struct obd_export *exp,
3334                          const struct lmv_stripe_md *lsm,
3335                          const char *name, int namelen, struct lu_fid *fid)
3336 {
3337         const struct lmv_oinfo *oinfo;
3338
3339         LASSERT(lsm != NULL);
3340         oinfo = lsm_name_to_stripe_info(lsm, name, namelen, false);
3341         if (IS_ERR(oinfo))
3342                 return PTR_ERR(oinfo);
3343
3344         *fid = oinfo->lmo_fid;
3345
3346         RETURN(0);
3347 }
3348
3349 /**
3350  * For lmv, only need to send request to master MDT, and the master MDT will
3351  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3352  * we directly fetch data from the slave MDTs.
3353  */
3354 int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3355                  struct obd_quotactl *oqctl)
3356 {
3357         struct obd_device   *obd = class_exp2obd(exp);
3358         struct lmv_obd      *lmv = &obd->u.lmv;
3359         struct lmv_tgt_desc *tgt = lmv->tgts[0];
3360         int                  rc = 0;
3361         __u32                i;
3362         __u64                curspace, curinodes;
3363         ENTRY;
3364
3365         if (tgt == NULL ||
3366             tgt->ltd_exp == NULL ||
3367             !tgt->ltd_active ||
3368             lmv->desc.ld_tgt_count == 0) {
3369                 CERROR("master lmv inactive\n");
3370                 RETURN(-EIO);
3371         }
3372
3373         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3374                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3375                 RETURN(rc);
3376         }
3377
3378         curspace = curinodes = 0;
3379         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3380                 int err;
3381                 tgt = lmv->tgts[i];
3382
3383                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
3384                         continue;
3385
3386                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3387                 if (err) {
3388                         CERROR("getquota on mdt %d failed. %d\n", i, err);
3389                         if (!rc)
3390                                 rc = err;
3391                 } else {
3392                         curspace += oqctl->qc_dqblk.dqb_curspace;
3393                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3394                 }
3395         }
3396         oqctl->qc_dqblk.dqb_curspace = curspace;
3397         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3398
3399         RETURN(rc);
3400 }
3401
3402 static int lmv_merge_attr(struct obd_export *exp,
3403                           const struct lmv_stripe_md *lsm,
3404                           struct cl_attr *attr,
3405                           ldlm_blocking_callback cb_blocking)
3406 {
3407         int rc;
3408         int i;
3409
3410         /* foreign dir is not striped dir */
3411         if (lsm->lsm_md_magic == LMV_MAGIC_FOREIGN)
3412                 return 0;
3413
3414         rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3415         if (rc < 0)
3416                 return rc;
3417
3418         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3419                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3420
3421                 if (!inode)
3422                         continue;
3423
3424                 CDEBUG(D_INFO,
3425                        "" DFID " size %llu, blocks %llu nlink %u, atime %lld ctime %lld, mtime %lld.\n",
3426                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3427                        i_size_read(inode), (unsigned long long)inode->i_blocks,
3428                        inode->i_nlink, (s64)inode->i_atime.tv_sec,
3429                        (s64)inode->i_ctime.tv_sec, (s64)inode->i_mtime.tv_sec);
3430
3431                 /* for slave stripe, it needs to subtract nlink for . and .. */
3432                 if (i != 0)
3433                         attr->cat_nlink += inode->i_nlink - 2;
3434                 else
3435                         attr->cat_nlink = inode->i_nlink;
3436
3437                 attr->cat_size += i_size_read(inode);
3438                 attr->cat_blocks += inode->i_blocks;
3439
3440                 if (attr->cat_atime < inode->i_atime.tv_sec)
3441                         attr->cat_atime = inode->i_atime.tv_sec;
3442
3443                 if (attr->cat_ctime < inode->i_ctime.tv_sec)
3444                         attr->cat_ctime = inode->i_ctime.tv_sec;
3445
3446                 if (attr->cat_mtime < inode->i_mtime.tv_sec)
3447                         attr->cat_mtime = inode->i_mtime.tv_sec;
3448         }
3449         return 0;
3450 }
3451
3452 struct obd_ops lmv_obd_ops = {
3453         .o_owner                = THIS_MODULE,
3454         .o_setup                = lmv_setup,
3455         .o_cleanup              = lmv_cleanup,
3456         .o_precleanup           = lmv_precleanup,
3457         .o_process_config       = lmv_process_config,
3458         .o_connect              = lmv_connect,
3459         .o_disconnect           = lmv_disconnect,
3460         .o_statfs               = lmv_statfs,
3461         .o_get_info             = lmv_get_info,
3462         .o_set_info_async       = lmv_set_info_async,
3463         .o_notify               = lmv_notify,
3464         .o_get_uuid             = lmv_get_uuid,
3465         .o_iocontrol            = lmv_iocontrol,
3466         .o_quotactl             = lmv_quotactl
3467 };
3468
3469 struct md_ops lmv_md_ops = {
3470         .m_get_root             = lmv_get_root,
3471         .m_null_inode           = lmv_null_inode,
3472         .m_close                = lmv_close,
3473         .m_create               = lmv_create,
3474         .m_enqueue              = lmv_enqueue,
3475         .m_getattr              = lmv_getattr,
3476         .m_getxattr             = lmv_getxattr,
3477         .m_getattr_name         = lmv_getattr_name,
3478         .m_intent_lock          = lmv_intent_lock,
3479         .m_link                 = lmv_link,
3480         .m_rename               = lmv_rename,
3481         .m_setattr              = lmv_setattr,
3482         .m_setxattr             = lmv_setxattr,
3483         .m_fsync                = lmv_fsync,
3484         .m_file_resync          = lmv_file_resync,
3485         .m_read_page            = lmv_read_page,
3486         .m_unlink               = lmv_unlink,
3487         .m_init_ea_size         = lmv_init_ea_size,
3488         .m_cancel_unused        = lmv_cancel_unused,
3489         .m_set_lock_data        = lmv_set_lock_data,
3490         .m_lock_match           = lmv_lock_match,
3491         .m_get_lustre_md        = lmv_get_lustre_md,
3492         .m_free_lustre_md       = lmv_free_lustre_md,
3493         .m_merge_attr           = lmv_merge_attr,
3494         .m_set_open_replay_data = lmv_set_open_replay_data,
3495         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3496         .m_intent_getattr_async = lmv_intent_getattr_async,
3497         .m_revalidate_lock      = lmv_revalidate_lock,
3498         .m_get_fid_from_lsm     = lmv_get_fid_from_lsm,
3499         .m_unpackmd             = lmv_unpackmd,
3500 };
3501
3502 static int __init lmv_init(void)
3503 {
3504         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true, NULL,
3505                                    LUSTRE_LMV_NAME, NULL);
3506 }
3507
3508 static void __exit lmv_exit(void)
3509 {
3510         class_unregister_type(LUSTRE_LMV_NAME);
3511 }
3512
3513 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3514 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
3515 MODULE_VERSION(LUSTRE_VERSION_STRING);
3516 MODULE_LICENSE("GPL");
3517
3518 module_init(lmv_init);
3519 module_exit(lmv_exit);