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