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 || !tgt->ltd_active) {
 269                         CWARN("%s: NULL export for %d\n", obd->obd_name, i);
 270                         continue;
 271                 }
 272
 273                 rc = md_init_ea_size(tgt->ltd_exp, easize, def_easize);
 274                 if (rc) {
 275                         CERROR("%s: obd_init_ea_size() failed on MDT target %d:"
 276                                " rc = %d\n", obd->obd_name, i, rc);
 277                         break;
 278                 }
 279         }
 280         RETURN(rc);
 281 }
 282
 283 #define MAX_STRING_SIZE 128
 284
 285 int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 286 {
 287         struct lmv_obd          *lmv = &obd->u.lmv;
 288         struct obd_device       *mdc_obd;
 289         struct obd_export       *mdc_exp;
 290         struct lu_fld_target     target;
 291         int                      rc;
 292         ENTRY;
 293
 294         mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME,
 295                                         &obd->obd_uuid);
 296         if (!mdc_obd) {
 297                 CERROR("target %s not attached\n", tgt->ltd_uuid.uuid);
 298                 RETURN(-EINVAL);
 299         }
 300
 301         CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s\n",
 302                mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 303                tgt->ltd_uuid.uuid, obd->obd_uuid.uuid);
 304
 305         if (!mdc_obd->obd_set_up) {
 306                 CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid);
 307                 RETURN(-EINVAL);
 308         }
 309
 310         rc = obd_connect(NULL, &mdc_exp, mdc_obd, &obd->obd_uuid,
 311                          &lmv->conn_data, NULL);
 312         if (rc) {
 313                 CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc);
 314                 RETURN(rc);
 315         }
 316
 317         /*
 318          * Init fid sequence client for this mdc and add new fld target.
 319          */
 320         rc = obd_fid_init(mdc_obd, mdc_exp, LUSTRE_SEQ_METADATA);
 321         if (rc)
 322                 RETURN(rc);
 323
 324         target.ft_srv = NULL;
 325         target.ft_exp = mdc_exp;
 326         target.ft_idx = tgt->ltd_idx;
 327
 328         fld_client_add_target(&lmv->lmv_fld, &target);
 329
 330         rc = obd_register_observer(mdc_obd, obd);
 331         if (rc) {
 332                 obd_disconnect(mdc_exp);
 333                 CERROR("target %s register_observer error %d\n",
 334                        tgt->ltd_uuid.uuid, rc);
 335                 RETURN(rc);
 336         }
 337
 338         if (obd->obd_observer) {
 339                 /*
 340                  * Tell the observer about the new target.
 341                  */
 342                 rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd,
 343                                 OBD_NOTIFY_ACTIVE);
 344                 if (rc) {
 345                         obd_disconnect(mdc_exp);
 346                         RETURN(rc);
 347                 }
 348         }
 349
 350         tgt->ltd_active = 1;
 351         tgt->ltd_exp = mdc_exp;
 352         lmv->desc.ld_active_tgt_count++;
 353
 354         md_init_ea_size(tgt->ltd_exp, lmv->max_easize, lmv->max_def_easize);
 355
 356         CDEBUG(D_CONFIG, "Connected to %s(%s) successfully (%d)\n",
 357                 mdc_obd->obd_name, mdc_obd->obd_uuid.uuid,
 358                 atomic_read(&obd->obd_refcount));
 359
 360         if (lmv->lmv_tgts_kobj)
 361                 /* Even if we failed to create the link, that's fine */
 362                 rc = sysfs_create_link(lmv->lmv_tgts_kobj,
 363                                        &mdc_obd->obd_kset.kobj,
 364                                        mdc_obd->obd_name);
 365         RETURN(0);
 366 }
 367
 368 static void lmv_del_target(struct lmv_obd *lmv, int index)
 369 {
 370         if (lmv->tgts[index] == NULL)
 371                 return;
 372
 373         OBD_FREE_PTR(lmv->tgts[index]);
 374         lmv->tgts[index] = NULL;
 375         return;
 376 }
 377
 378 static int lmv_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 379                            __u32 index, int gen)
 380 {
 381         struct obd_device *mdc_obd;
 382         struct lmv_obd      *lmv = &obd->u.lmv;
 383         struct lmv_tgt_desc *tgt;
 384         int                  orig_tgt_count = 0;
 385         int                  rc = 0;
 386         ENTRY;
 387
 388         CDEBUG(D_CONFIG, "Target uuid: %s. index %d\n", uuidp->uuid, index);
 389         mdc_obd = class_find_client_obd(uuidp, LUSTRE_MDC_NAME,
 390                                         &obd->obd_uuid);
 391         if (!mdc_obd) {
 392                 CERROR("%s: Target %s not attached: rc = %d\n",
 393                        obd->obd_name, uuidp->uuid, -EINVAL);
 394                 RETURN(-EINVAL);
 395         }
 396
 397         mutex_lock(&lmv->lmv_init_mutex);
 398         if ((index < lmv->tgts_size) && (lmv->tgts[index] != NULL)) {
 399                 tgt = lmv->tgts[index];
 400                 CERROR("%s: UUID %s already assigned at LMV target index %d:"
 401                        " rc = %d\n", obd->obd_name,
 402                        obd_uuid2str(&tgt->ltd_uuid), index, -EEXIST);
 403                 mutex_unlock(&lmv->lmv_init_mutex);
 404                 RETURN(-EEXIST);
 405         }
 406
 407         if (index >= lmv->tgts_size) {
 408                 /* We need to reallocate the lmv target array. */
 409                 struct lmv_tgt_desc **newtgts, **old = NULL;
 410                 __u32 newsize = 1;
 411                 __u32 oldsize = 0;
 412
 413                 while (newsize < index + 1)
 414                         newsize = newsize << 1;
 415                 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
 416                 if (newtgts == NULL) {
 417                         mutex_unlock(&lmv->lmv_init_mutex);
 418                         RETURN(-ENOMEM);
 419                 }
 420
 421                 if (lmv->tgts_size) {
 422                         memcpy(newtgts, lmv->tgts,
 423                                sizeof(*newtgts) * lmv->tgts_size);
 424                         old = lmv->tgts;
 425                         oldsize = lmv->tgts_size;
 426                 }
 427
 428                 lmv->tgts = newtgts;
 429                 lmv->tgts_size = newsize;
 430                 smp_rmb();
 431                 if (old)
 432                         OBD_FREE(old, sizeof(*old) * oldsize);
 433
 434                 CDEBUG(D_CONFIG, "tgts: %p size: %d\n", lmv->tgts,
 435                        lmv->tgts_size);
 436         }
 437
 438         OBD_ALLOC_PTR(tgt);
 439         if (!tgt) {
 440                 mutex_unlock(&lmv->lmv_init_mutex);
 441                 RETURN(-ENOMEM);
 442         }
 443
 444         mutex_init(&tgt->ltd_fid_mutex);
 445         tgt->ltd_idx = index;
 446         tgt->ltd_uuid = *uuidp;
 447         tgt->ltd_active = 0;
 448         lmv->tgts[index] = tgt;
 449         if (index >= lmv->desc.ld_tgt_count) {
 450                 orig_tgt_count = lmv->desc.ld_tgt_count;
 451                 lmv->desc.ld_tgt_count = index + 1;
 452         }
 453
 454         if (lmv->connected == 0) {
 455                 /* lmv_check_connect() will connect this target. */
 456                 mutex_unlock(&lmv->lmv_init_mutex);
 457                 RETURN(0);
 458         }
 459
 460         /* Otherwise let's connect it ourselves */
 461         mutex_unlock(&lmv->lmv_init_mutex);
 462         rc = lmv_connect_mdc(obd, tgt);
 463         if (rc != 0) {
 464                 spin_lock(&lmv->lmv_lock);
 465                 if (lmv->desc.ld_tgt_count == index + 1)
 466                         lmv->desc.ld_tgt_count = orig_tgt_count;
 467                 memset(tgt, 0, sizeof(*tgt));
 468                 spin_unlock(&lmv->lmv_lock);
 469         } else {
 470                 int easize = sizeof(struct lmv_stripe_md) +
 471                         lmv->desc.ld_tgt_count * sizeof(struct lu_fid);
 472                 lmv_init_ea_size(obd->obd_self_export, easize, 0);
 473         }
 474
 475         RETURN(rc);
 476 }
 477
 478 static int lmv_check_connect(struct obd_device *obd)
 479 {
 480         struct lmv_obd          *lmv = &obd->u.lmv;
 481         struct lmv_tgt_desc     *tgt;
 482         __u32                    i;
 483         int                      rc;
 484         int                      easize;
 485         ENTRY;
 486
 487         if (lmv->connected)
 488                 RETURN(0);
 489
 490         mutex_lock(&lmv->lmv_init_mutex);
 491         if (lmv->connected) {
 492                 mutex_unlock(&lmv->lmv_init_mutex);
 493                 RETURN(0);
 494         }
 495
 496         if (lmv->desc.ld_tgt_count == 0) {
 497                 mutex_unlock(&lmv->lmv_init_mutex);
 498                 CERROR("%s: no targets configured.\n", obd->obd_name);
 499                 RETURN(-EINVAL);
 500         }
 501
 502         LASSERT(lmv->tgts != NULL);
 503
 504         if (lmv->tgts[0] == NULL) {
 505                 mutex_unlock(&lmv->lmv_init_mutex);
 506                 CERROR("%s: no target configured for index 0.\n",
 507                        obd->obd_name);
 508                 RETURN(-EINVAL);
 509         }
 510
 511         CDEBUG(D_CONFIG, "Time to connect %s to %s\n",
 512                obd->obd_uuid.uuid, obd->obd_name);
 513
 514         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 515                 tgt = lmv->tgts[i];
 516                 if (tgt == NULL)
 517                         continue;
 518                 rc = lmv_connect_mdc(obd, tgt);
 519                 if (rc)
 520                         GOTO(out_disc, rc);
 521         }
 522
 523         lmv->connected = 1;
 524         easize = lmv_mds_md_size(lmv->desc.ld_tgt_count, LMV_MAGIC);
 525         lmv_init_ea_size(obd->obd_self_export, easize, 0);
 526         mutex_unlock(&lmv->lmv_init_mutex);
 527         RETURN(0);
 528
 529  out_disc:
 530         while (i-- > 0) {
 531                 int rc2;
 532                 tgt = lmv->tgts[i];
 533                 if (tgt == NULL)
 534                         continue;
 535                 tgt->ltd_active = 0;
 536                 if (tgt->ltd_exp) {
 537                         --lmv->desc.ld_active_tgt_count;
 538                         rc2 = obd_disconnect(tgt->ltd_exp);
 539                         if (rc2) {
 540                                 CERROR("LMV target %s disconnect on "
 541                                        "MDC idx %d: error %d\n",
 542                                        tgt->ltd_uuid.uuid, i, rc2);
 543                         }
 544                 }
 545         }
 546
 547         mutex_unlock(&lmv->lmv_init_mutex);
 548
 549         RETURN(rc);
 550 }
 551
 552 static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt)
 553 {
 554         struct lmv_obd         *lmv = &obd->u.lmv;
 555         struct obd_device      *mdc_obd;
 556         int                     rc;
 557         ENTRY;
 558
 559         LASSERT(tgt != NULL);
 560         LASSERT(obd != NULL);
 561
 562         mdc_obd = class_exp2obd(tgt->ltd_exp);
 563
 564         if (mdc_obd) {
 565                 mdc_obd->obd_force = obd->obd_force;
 566                 mdc_obd->obd_fail = obd->obd_fail;
 567                 mdc_obd->obd_no_recov = obd->obd_no_recov;
 568
 569                 if (lmv->lmv_tgts_kobj)
 570                         sysfs_remove_link(lmv->lmv_tgts_kobj,
 571                                           mdc_obd->obd_name);
 572         }
 573
 574         rc = obd_fid_fini(tgt->ltd_exp->exp_obd);
 575         if (rc)
 576                 CERROR("Can't finanize fids factory\n");
 577
 578         CDEBUG(D_INFO, "Disconnected from %s(%s) successfully\n",
 579                tgt->ltd_exp->exp_obd->obd_name,
 580                tgt->ltd_exp->exp_obd->obd_uuid.uuid);
 581
 582         obd_register_observer(tgt->ltd_exp->exp_obd, NULL);
 583         rc = obd_disconnect(tgt->ltd_exp);
 584         if (rc) {
 585                 if (tgt->ltd_active) {
 586                         CERROR("Target %s disconnect error %d\n",
 587                                tgt->ltd_uuid.uuid, rc);
 588                 }
 589         }
 590
 591         lmv_activate_target(lmv, tgt, 0);
 592         tgt->ltd_exp = NULL;
 593         RETURN(0);
 594 }
 595
 596 static int lmv_disconnect(struct obd_export *exp)
 597 {
 598         struct obd_device       *obd = class_exp2obd(exp);
 599         struct lmv_obd          *lmv = &obd->u.lmv;
 600         int                      rc;
 601         __u32                    i;
 602         ENTRY;
 603
 604         if (!lmv->tgts)
 605                 goto out_local;
 606
 607         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 608                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
 609                         continue;
 610
 611                 lmv_disconnect_mdc(obd, lmv->tgts[i]);
 612         }
 613
 614         if (lmv->lmv_tgts_kobj)
 615                 kobject_put(lmv->lmv_tgts_kobj);
 616
 617 out_local:
 618         /*
 619          * This is the case when no real connection is established by
 620          * lmv_check_connect().
 621          */
 622         if (!lmv->connected)
 623                 class_export_put(exp);
 624         rc = class_disconnect(exp);
 625         lmv->connected = 0;
 626
 627         RETURN(rc);
 628 }
 629
 630 static int lmv_fid2path(struct obd_export *exp, int len, void *karg,
 631                         void __user *uarg)
 632 {
 633         struct obd_device       *obddev = class_exp2obd(exp);
 634         struct lmv_obd          *lmv = &obddev->u.lmv;
 635         struct getinfo_fid2path *gf;
 636         struct lmv_tgt_desc     *tgt;
 637         struct getinfo_fid2path *remote_gf = NULL;
 638         struct lu_fid           root_fid;
 639         int                     remote_gf_size = 0;
 640         int                     rc;
 641
 642         gf = karg;
 643         tgt = lmv_find_target(lmv, &gf->gf_fid);
 644         if (IS_ERR(tgt))
 645                 RETURN(PTR_ERR(tgt));
 646
 647         root_fid = *gf->gf_u.gf_root_fid;
 648         LASSERT(fid_is_sane(&root_fid));
 649
 650 repeat_fid2path:
 651         rc = obd_iocontrol(OBD_IOC_FID2PATH, tgt->ltd_exp, len, gf, uarg);
 652         if (rc != 0 && rc != -EREMOTE)
 653                 GOTO(out_fid2path, rc);
 654
 655         /* If remote_gf != NULL, it means just building the
 656          * path on the remote MDT, copy this path segement to gf */
 657         if (remote_gf != NULL) {
 658                 struct getinfo_fid2path *ori_gf;
 659                 char *ptr;
 660
 661                 ori_gf = (struct getinfo_fid2path *)karg;
 662                 if (strlen(ori_gf->gf_u.gf_path) + 1 +
 663                     strlen(gf->gf_u.gf_path) + 1 > ori_gf->gf_pathlen)
 664                         GOTO(out_fid2path, rc = -EOVERFLOW);
 665
 666                 ptr = ori_gf->gf_u.gf_path;
 667
 668                 memmove(ptr + strlen(gf->gf_u.gf_path) + 1, ptr,
 669                         strlen(ori_gf->gf_u.gf_path));
 670
 671                 strncpy(ptr, gf->gf_u.gf_path,
 672                         strlen(gf->gf_u.gf_path));
 673                 ptr += strlen(gf->gf_u.gf_path);
 674                 *ptr = '/';
 675         }
 676
 677         CDEBUG(D_INFO, "%s: get path %s "DFID" rec: %llu ln: %u\n",
 678                tgt->ltd_exp->exp_obd->obd_name,
 679                gf->gf_u.gf_path, PFID(&gf->gf_fid), gf->gf_recno,
 680                gf->gf_linkno);
 681
 682         if (rc == 0)
 683                 GOTO(out_fid2path, rc);
 684
 685         /* sigh, has to go to another MDT to do path building further */
 686         if (remote_gf == NULL) {
 687                 remote_gf_size = sizeof(*remote_gf) + PATH_MAX;
 688                 OBD_ALLOC(remote_gf, remote_gf_size);
 689                 if (remote_gf == NULL)
 690                         GOTO(out_fid2path, rc = -ENOMEM);
 691                 remote_gf->gf_pathlen = PATH_MAX;
 692         }
 693
 694         if (!fid_is_sane(&gf->gf_fid)) {
 695                 CERROR("%s: invalid FID "DFID": rc = %d\n",
 696                        tgt->ltd_exp->exp_obd->obd_name,
 697                        PFID(&gf->gf_fid), -EINVAL);
 698                 GOTO(out_fid2path, rc = -EINVAL);
 699         }
 700
 701         tgt = lmv_find_target(lmv, &gf->gf_fid);
 702         if (IS_ERR(tgt))
 703                 GOTO(out_fid2path, rc = -EINVAL);
 704
 705         remote_gf->gf_fid = gf->gf_fid;
 706         remote_gf->gf_recno = -1;
 707         remote_gf->gf_linkno = -1;
 708         memset(remote_gf->gf_u.gf_path, 0, remote_gf->gf_pathlen);
 709         *remote_gf->gf_u.gf_root_fid = root_fid;
 710         gf = remote_gf;
 711         goto repeat_fid2path;
 712
 713 out_fid2path:
 714         if (remote_gf != NULL)
 715                 OBD_FREE(remote_gf, remote_gf_size);
 716         RETURN(rc);
 717 }
 718
 719 static int lmv_hsm_req_count(struct lmv_obd *lmv,
 720                              const struct hsm_user_request *hur,
 721                              const struct lmv_tgt_desc *tgt_mds)
 722 {
 723         __u32                    i;
 724         int                      nr = 0;
 725         struct lmv_tgt_desc     *curr_tgt;
 726
 727         /* count how many requests must be sent to the given target */
 728         for (i = 0; i < hur->hur_request.hr_itemcount; i++) {
 729                 curr_tgt = lmv_find_target(lmv, &hur->hur_user_item[i].hui_fid);
 730                 if (IS_ERR(curr_tgt))
 731                         RETURN(PTR_ERR(curr_tgt));
 732                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid))
 733                         nr++;
 734         }
 735         return nr;
 736 }
 737
 738 static int lmv_hsm_req_build(struct lmv_obd *lmv,
 739                               struct hsm_user_request *hur_in,
 740                               const struct lmv_tgt_desc *tgt_mds,
 741                               struct hsm_user_request *hur_out)
 742 {
 743         __u32                    i, nr_out;
 744         struct lmv_tgt_desc     *curr_tgt;
 745
 746         /* build the hsm_user_request for the given target */
 747         hur_out->hur_request = hur_in->hur_request;
 748         nr_out = 0;
 749         for (i = 0; i < hur_in->hur_request.hr_itemcount; i++) {
 750                 curr_tgt = lmv_find_target(lmv,
 751                                            &hur_in->hur_user_item[i].hui_fid);
 752                 if (IS_ERR(curr_tgt))
 753                         RETURN(PTR_ERR(curr_tgt));
 754                 if (obd_uuid_equals(&curr_tgt->ltd_uuid, &tgt_mds->ltd_uuid)) {
 755                         hur_out->hur_user_item[nr_out] =
 756                                                 hur_in->hur_user_item[i];
 757                         nr_out++;
 758                 }
 759         }
 760         hur_out->hur_request.hr_itemcount = nr_out;
 761         memcpy(hur_data(hur_out), hur_data(hur_in),
 762                hur_in->hur_request.hr_data_len);
 763
 764         RETURN(0);
 765 }
 766
 767 static int lmv_hsm_ct_unregister(struct obd_device *obd, unsigned int cmd,
 768                                  int len, struct lustre_kernelcomm *lk,
 769                                  void __user *uarg)
 770 {
 771         struct lmv_obd *lmv = &obd->u.lmv;
 772         __u32   i;
 773         int     rc;
 774         ENTRY;
 775
 776         /* unregister request (call from llapi_hsm_copytool_fini) */
 777         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 778                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 779
 780                 if (tgt == NULL || tgt->ltd_exp == NULL)
 781                         continue;
 782                 /* best effort: try to clean as much as possible
 783                  * (continue on error) */
 784                 obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 785         }
 786
 787         /* Whatever the result, remove copytool from kuc groups.
 788          * Unreached coordinators will get EPIPE on next requests
 789          * and will unregister automatically.
 790          */
 791         rc = libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 792
 793         RETURN(rc);
 794 }
 795
 796 static int lmv_hsm_ct_register(struct obd_device *obd, unsigned int cmd,
 797                                int len, struct lustre_kernelcomm *lk,
 798                                void __user *uarg)
 799 {
 800         struct lmv_obd *lmv = &obd->u.lmv;
 801         struct file             *filp;
 802         __u32                    i, j;
 803         int                      err;
 804         bool                     any_set = false;
 805         struct kkuc_ct_data      kcd = {
 806                 .kcd_magic   = KKUC_CT_DATA_MAGIC,
 807                 .kcd_archive = lk->lk_data,
 808         };
 809         int                      rc = 0;
 810         ENTRY;
 811
 812         filp = fget(lk->lk_wfd);
 813         if (!filp)
 814                 RETURN(-EBADF);
 815
 816         rc = libcfs_kkuc_group_add(filp, &obd->obd_uuid, lk->lk_uid,
 817                                    lk->lk_group, &kcd, sizeof(kcd));
 818         if (rc)
 819                 GOTO(err_fput, rc);
 820
 821         /* All or nothing: try to register to all MDS.
 822          * In case of failure, unregister from previous MDS,
 823          * except if it because of inactive target. */
 824         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 825                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
 826
 827                 if (tgt == NULL || tgt->ltd_exp == NULL)
 828                         continue;
 829
 830                 err = obd_iocontrol(cmd, tgt->ltd_exp, len, lk, uarg);
 831                 if (err) {
 832                         if (tgt->ltd_active) {
 833                                 /* permanent error */
 834                                 CERROR("%s: iocontrol MDC %s on MDT"
 835                                        " idx %d cmd %x: err = %d\n",
 836                                        lmv2obd_dev(lmv)->obd_name,
 837                                        tgt->ltd_uuid.uuid, i, cmd, err);
 838                                 rc = err;
 839                                 lk->lk_flags |= LK_FLG_STOP;
 840                                 /* unregister from previous MDS */
 841                                 for (j = 0; j < i; j++) {
 842                                         tgt = lmv->tgts[j];
 843                                         if (tgt == NULL || tgt->ltd_exp == NULL)
 844                                                 continue;
 845                                         obd_iocontrol(cmd, tgt->ltd_exp, len,
 846                                                       lk, uarg);
 847                                 }
 848                                 GOTO(err_kkuc_rem, rc);
 849                         }
 850                         /* else: transient error.
 851                          * kuc will register to the missing MDT
 852                          * when it is back */
 853                 } else {
 854                         any_set = true;
 855                 }
 856         }
 857
 858         if (!any_set)
 859                 /* no registration done: return error */
 860                 GOTO(err_kkuc_rem, rc = -ENOTCONN);
 861
 862         RETURN(0);
 863
 864 err_kkuc_rem:
 865         libcfs_kkuc_group_rem(&obd->obd_uuid, lk->lk_uid, lk->lk_group);
 866
 867 err_fput:
 868         fput(filp);
 869         return rc;
 870 }
 871
 872
 873
 874
 875 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 876                          int len, void *karg, void __user *uarg)
 877 {
 878         struct obd_device       *obddev = class_exp2obd(exp);
 879         struct lmv_obd          *lmv = &obddev->u.lmv;
 880         struct lmv_tgt_desc     *tgt = NULL;
 881         __u32                    i = 0;
 882         int                      rc = 0;
 883         int                      set = 0;
 884         __u32                    count = lmv->desc.ld_tgt_count;
 885         ENTRY;
 886
 887         if (count == 0)
 888                 RETURN(-ENOTTY);
 889
 890         switch (cmd) {
 891         case IOC_OBD_STATFS: {
 892                 struct obd_ioctl_data *data = karg;
 893                 struct obd_device *mdc_obd;
 894                 struct obd_statfs stat_buf = {0};
 895                 __u32 index;
 896
 897                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
 898                 if ((index >= count))
 899                         RETURN(-ENODEV);
 900
 901                 tgt = lmv->tgts[index];
 902                 if (tgt == NULL || !tgt->ltd_active)
 903                         RETURN(-ENODATA);
 904
 905                 mdc_obd = class_exp2obd(tgt->ltd_exp);
 906                 if (!mdc_obd)
 907                         RETURN(-EINVAL);
 908
 909                 /* copy UUID */
 910                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(mdc_obd),
 911                                  min((int) data->ioc_plen2,
 912                                      (int) sizeof(struct obd_uuid))))
 913                         RETURN(-EFAULT);
 914
 915                 rc = obd_statfs(NULL, tgt->ltd_exp, &stat_buf,
 916                                 ktime_get_seconds() - OBD_STATFS_CACHE_SECONDS,
 917                                 0);
 918                 if (rc)
 919                         RETURN(rc);
 920                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
 921                                  min((int) data->ioc_plen1,
 922                                      (int) sizeof(stat_buf))))
 923                         RETURN(-EFAULT);
 924                 break;
 925         }
 926         case OBD_IOC_QUOTACTL: {
 927                 struct if_quotactl *qctl = karg;
 928                 struct obd_quotactl *oqctl;
 929
 930                 if (qctl->qc_valid == QC_MDTIDX) {
 931                         if (count <= qctl->qc_idx)
 932                                 RETURN(-EINVAL);
 933
 934                         tgt = lmv->tgts[qctl->qc_idx];
 935                         if (tgt == NULL || tgt->ltd_exp == NULL)
 936                                 RETURN(-EINVAL);
 937                 } else if (qctl->qc_valid == QC_UUID) {
 938                         for (i = 0; i < count; i++) {
 939                                 tgt = lmv->tgts[i];
 940                                 if (tgt == NULL)
 941                                         continue;
 942                                 if (!obd_uuid_equals(&tgt->ltd_uuid,
 943                                                      &qctl->obd_uuid))
 944                                         continue;
 945
 946                                 if (tgt->ltd_exp == NULL)
 947                                         RETURN(-EINVAL);
 948
 949                                 break;
 950                         }
 951                 } else {
 952                         RETURN(-EINVAL);
 953                 }
 954
 955                 if (i >= count)
 956                         RETURN(-EAGAIN);
 957
 958                 LASSERT(tgt != NULL && tgt->ltd_exp != NULL);
 959                 OBD_ALLOC_PTR(oqctl);
 960                 if (!oqctl)
 961                         RETURN(-ENOMEM);
 962
 963                 QCTL_COPY(oqctl, qctl);
 964                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
 965                 if (rc == 0) {
 966                         QCTL_COPY(qctl, oqctl);
 967                         qctl->qc_valid = QC_MDTIDX;
 968                         qctl->obd_uuid = tgt->ltd_uuid;
 969                 }
 970                 OBD_FREE_PTR(oqctl);
 971                 break;
 972         }
 973         case LL_IOC_GET_CONNECT_FLAGS: {
 974                 tgt = lmv->tgts[0];
 975                 if (tgt == NULL || tgt->ltd_exp == NULL)
 976                         RETURN(-ENODATA);
 977                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
 978                 break;
 979         }
 980         case LL_IOC_FID2MDTIDX: {
 981                 struct lu_fid *fid = karg;
 982                 int             mdt_index;
 983
 984                 rc = lmv_fld_lookup(lmv, fid, &mdt_index);
 985                 if (rc != 0)
 986                         RETURN(rc);
 987
 988                 /* Note: this is from llite(see ll_dir_ioctl()), @uarg does not
 989                  * point to user space memory for FID2MDTIDX. */
 990                 *(__u32 *)uarg = mdt_index;
 991                 break;
 992         }
 993         case OBD_IOC_FID2PATH: {
 994                 rc = lmv_fid2path(exp, len, karg, uarg);
 995                 break;
 996         }
 997         case LL_IOC_HSM_STATE_GET:
 998         case LL_IOC_HSM_STATE_SET:
 999         case LL_IOC_HSM_ACTION: {
1000                 struct md_op_data       *op_data = karg;
1001
1002                 tgt = lmv_find_target(lmv, &op_data->op_fid1);
1003                 if (IS_ERR(tgt))
1004                         RETURN(PTR_ERR(tgt));
1005
1006                 if (tgt->ltd_exp == NULL)
1007                         RETURN(-EINVAL);
1008
1009                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1010                 break;
1011         }
1012         case LL_IOC_HSM_PROGRESS: {
1013                 const struct hsm_progress_kernel *hpk = karg;
1014
1015                 tgt = lmv_find_target(lmv, &hpk->hpk_fid);
1016                 if (IS_ERR(tgt))
1017                         RETURN(PTR_ERR(tgt));
1018                 rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1019                 break;
1020         }
1021         case LL_IOC_HSM_REQUEST: {
1022                 struct hsm_user_request *hur = karg;
1023                 unsigned int reqcount = hur->hur_request.hr_itemcount;
1024
1025                 if (reqcount == 0)
1026                         RETURN(0);
1027
1028                 /* if the request is about a single fid
1029                  * or if there is a single MDS, no need to split
1030                  * the request. */
1031                 if (reqcount == 1 || count == 1) {
1032                         tgt = lmv_find_target(lmv,
1033                                               &hur->hur_user_item[0].hui_fid);
1034                         if (IS_ERR(tgt))
1035                                 RETURN(PTR_ERR(tgt));
1036                         rc = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1037                 } else {
1038                         /* split fid list to their respective MDS */
1039                         for (i = 0; i < count; i++) {
1040                                 int nr, rc1;
1041                                 size_t reqlen;
1042                                 struct hsm_user_request *req;
1043
1044                                 tgt = lmv->tgts[i];
1045                                 if (tgt == NULL || tgt->ltd_exp == NULL)
1046                                         continue;
1047
1048                                 nr = lmv_hsm_req_count(lmv, hur, tgt);
1049                                 if (nr < 0)
1050                                         RETURN(nr);
1051                                 if (nr == 0) /* nothing for this MDS */
1052                                         continue;
1053
1054                                 /* build a request with fids for this MDS */
1055                                 reqlen = offsetof(typeof(*hur),
1056                                                   hur_user_item[nr])
1057                                                 + hur->hur_request.hr_data_len;
1058                                 OBD_ALLOC_LARGE(req, reqlen);
1059                                 if (req == NULL)
1060                                         RETURN(-ENOMEM);
1061                                 rc1 = lmv_hsm_req_build(lmv, hur, tgt, req);
1062                                 if (rc1 < 0)
1063                                         GOTO(hsm_req_err, rc1);
1064                                 rc1 = obd_iocontrol(cmd, tgt->ltd_exp, reqlen,
1065                                                     req, uarg);
1066 hsm_req_err:
1067                                 if (rc1 != 0 && rc == 0)
1068                                         rc = rc1;
1069                                 OBD_FREE_LARGE(req, reqlen);
1070                         }
1071                 }
1072                 break;
1073         }
1074         case LL_IOC_LOV_SWAP_LAYOUTS: {
1075                 struct md_op_data       *op_data = karg;
1076                 struct lmv_tgt_desc     *tgt1, *tgt2;
1077
1078                 tgt1 = lmv_find_target(lmv, &op_data->op_fid1);
1079                 if (IS_ERR(tgt1))
1080                         RETURN(PTR_ERR(tgt1));
1081
1082                 tgt2 = lmv_find_target(lmv, &op_data->op_fid2);
1083                 if (IS_ERR(tgt2))
1084                         RETURN(PTR_ERR(tgt2));
1085
1086                 if ((tgt1->ltd_exp == NULL) || (tgt2->ltd_exp == NULL))
1087                         RETURN(-EINVAL);
1088
1089                 /* only files on same MDT can have their layouts swapped */
1090                 if (tgt1->ltd_idx != tgt2->ltd_idx)
1091                         RETURN(-EPERM);
1092
1093                 rc = obd_iocontrol(cmd, tgt1->ltd_exp, len, karg, uarg);
1094                 break;
1095         }
1096         case LL_IOC_HSM_CT_START: {
1097                 struct lustre_kernelcomm *lk = karg;
1098                 if (lk->lk_flags & LK_FLG_STOP)
1099                         rc = lmv_hsm_ct_unregister(obddev, cmd, len, lk, uarg);
1100                 else
1101                         rc = lmv_hsm_ct_register(obddev, cmd, len, lk, uarg);
1102                 break;
1103         }
1104         default:
1105                 for (i = 0; i < count; i++) {
1106                         struct obd_device *mdc_obd;
1107                         int err;
1108
1109                         tgt = lmv->tgts[i];
1110                         if (tgt == NULL || tgt->ltd_exp == NULL)
1111                                 continue;
1112                         /* ll_umount_begin() sets force flag but for lmv, not
1113                          * mdc. Let's pass it through */
1114                         mdc_obd = class_exp2obd(tgt->ltd_exp);
1115                         mdc_obd->obd_force = obddev->obd_force;
1116                         err = obd_iocontrol(cmd, tgt->ltd_exp, len, karg, uarg);
1117                         if (err) {
1118                                 if (tgt->ltd_active) {
1119                                         CERROR("error: iocontrol MDC %s on MDT"
1120                                                " idx %d cmd %x: err = %d\n",
1121                                                tgt->ltd_uuid.uuid, i, cmd, err);
1122                                         if (!rc)
1123                                                 rc = err;
1124                                 }
1125                         } else
1126                                 set = 1;
1127                 }
1128                 if (!set && !rc)
1129                         rc = -EIO;
1130         }
1131         RETURN(rc);
1132 }
1133
1134 /**
1135  * This is _inode_ placement policy function (not name).
1136  */
1137 static int lmv_placement_policy(struct obd_device *obd,
1138                                 struct md_op_data *op_data, u32 *mds)
1139 {
1140         struct lmv_obd     *lmv = &obd->u.lmv;
1141         struct lmv_user_md *lum;
1142
1143         ENTRY;
1144
1145         LASSERT(mds != NULL);
1146
1147         if (lmv->desc.ld_tgt_count == 1) {
1148                 *mds = 0;
1149                 RETURN(0);
1150         }
1151
1152         lum = op_data->op_data;
1153         /* Choose MDS by
1154          * 1. See if the stripe offset is specified by lum.
1155          * 2. Then check if there is default stripe offset.
1156          * 3. Finally choose MDS by name hash if the parent
1157          *    is striped directory. (see lmv_locate_mds()). */
1158         if (op_data->op_cli_flags & CLI_SET_MEA && lum != NULL &&
1159             le32_to_cpu(lum->lum_stripe_offset) != (__u32)-1) {
1160                 *mds = le32_to_cpu(lum->lum_stripe_offset);
1161         } else if (op_data->op_default_stripe_offset != (__u32)-1) {
1162                 *mds = op_data->op_default_stripe_offset;
1163                 op_data->op_mds = *mds;
1164                 /* Correct the stripe offset in lum */
1165                 if (lum != NULL)
1166                         lum->lum_stripe_offset = cpu_to_le32(*mds);
1167         } else {
1168                 *mds = op_data->op_mds;
1169         }
1170
1171         RETURN(0);
1172 }
1173
1174 int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid, u32 mds)
1175 {
1176         struct lmv_tgt_desc     *tgt;
1177         int                      rc;
1178         ENTRY;
1179
1180         tgt = lmv_get_target(lmv, mds, NULL);
1181         if (IS_ERR(tgt))
1182                 RETURN(PTR_ERR(tgt));
1183
1184         /*
1185          * New seq alloc and FLD setup should be atomic. Otherwise we may find
1186          * on server that seq in new allocated fid is not yet known.
1187          */
1188         mutex_lock(&tgt->ltd_fid_mutex);
1189
1190         if (tgt->ltd_active == 0 || tgt->ltd_exp == NULL)
1191                 GOTO(out, rc = -ENODEV);
1192
1193         /*
1194          * Asking underlying tgt layer to allocate new fid.
1195          */
1196         rc = obd_fid_alloc(NULL, tgt->ltd_exp, fid, NULL);
1197         if (rc > 0) {
1198                 LASSERT(fid_is_sane(fid));
1199                 rc = 0;
1200         }
1201
1202         EXIT;
1203 out:
1204         mutex_unlock(&tgt->ltd_fid_mutex);
1205         return rc;
1206 }
1207
1208 int lmv_fid_alloc(const struct lu_env *env, struct obd_export *exp,
1209                   struct lu_fid *fid, struct md_op_data *op_data)
1210 {
1211         struct obd_device     *obd = class_exp2obd(exp);
1212         struct lmv_obd        *lmv = &obd->u.lmv;
1213         u32                    mds = 0;
1214         int                    rc;
1215         ENTRY;
1216
1217         LASSERT(op_data != NULL);
1218         LASSERT(fid != NULL);
1219
1220         rc = lmv_placement_policy(obd, op_data, &mds);
1221         if (rc) {
1222                 CERROR("Can't get target for allocating fid, "
1223                        "rc %d\n", rc);
1224                 RETURN(rc);
1225         }
1226
1227         rc = __lmv_fid_alloc(lmv, fid, mds);
1228         if (rc) {
1229                 CERROR("Can't alloc new fid, rc %d\n", rc);
1230                 RETURN(rc);
1231         }
1232
1233         RETURN(rc);
1234 }
1235
1236 static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
1237 {
1238         struct lmv_obd  *lmv = &obd->u.lmv;
1239         struct lmv_desc *desc;
1240         int             rc;
1241         ENTRY;
1242
1243         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1244                 CERROR("LMV setup requires a descriptor\n");
1245                 RETURN(-EINVAL);
1246         }
1247
1248         desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1);
1249         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1250                 CERROR("Lmv descriptor size wrong: %d > %d\n",
1251                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1252                 RETURN(-EINVAL);
1253         }
1254
1255         lmv->tgts_size = 32U;
1256         OBD_ALLOC(lmv->tgts, sizeof(*lmv->tgts) * lmv->tgts_size);
1257         if (lmv->tgts == NULL)
1258                 RETURN(-ENOMEM);
1259
1260         obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid);
1261         lmv->desc.ld_tgt_count = 0;
1262         lmv->desc.ld_active_tgt_count = 0;
1263         lmv->max_def_easize = 0;
1264         lmv->max_easize = 0;
1265
1266         spin_lock_init(&lmv->lmv_lock);
1267         mutex_init(&lmv->lmv_init_mutex);
1268
1269         rc = lmv_tunables_init(obd);
1270         if (rc)
1271                 CWARN("%s: error adding LMV sysfs/debugfs files: rc = %d\n",
1272                       obd->obd_name, rc);
1273
1274         rc = fld_client_init(&lmv->lmv_fld, obd->obd_name,
1275                              LUSTRE_CLI_FLD_HASH_DHT);
1276         if (rc) {
1277                 CERROR("Can't init FLD, err %d\n", rc);
1278                 GOTO(out, rc);
1279         }
1280
1281         RETURN(0);
1282
1283 out:
1284         return rc;
1285 }
1286
1287 static int lmv_cleanup(struct obd_device *obd)
1288 {
1289         struct lmv_obd   *lmv = &obd->u.lmv;
1290         ENTRY;
1291
1292         fld_client_fini(&lmv->lmv_fld);
1293         if (lmv->tgts != NULL) {
1294                 int i;
1295                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1296                         if (lmv->tgts[i] == NULL)
1297                                 continue;
1298                         lmv_del_target(lmv, i);
1299                 }
1300                 OBD_FREE(lmv->tgts, sizeof(*lmv->tgts) * lmv->tgts_size);
1301                 lmv->tgts_size = 0;
1302         }
1303         RETURN(0);
1304 }
1305
1306 static int lmv_process_config(struct obd_device *obd, size_t len, void *buf)
1307 {
1308         struct lustre_cfg       *lcfg = buf;
1309         struct obd_uuid         obd_uuid;
1310         int                     gen;
1311         __u32                   index;
1312         int                     rc;
1313         ENTRY;
1314
1315         switch (lcfg->lcfg_command) {
1316         case LCFG_ADD_MDC:
1317                 /* modify_mdc_tgts add 0:lustre-clilmv  1:lustre-MDT0000_UUID
1318                  * 2:0  3:1  4:lustre-MDT0000-mdc_UUID */
1319                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
1320                         GOTO(out, rc = -EINVAL);
1321
1322                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
1323
1324                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", &index) != 1)
1325                         GOTO(out, rc = -EINVAL);
1326                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
1327                         GOTO(out, rc = -EINVAL);
1328                 rc = lmv_add_target(obd, &obd_uuid, index, gen);
1329                 GOTO(out, rc);
1330         default:
1331                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1332                 GOTO(out, rc = -EINVAL);
1333         }
1334 out:
1335         RETURN(rc);
1336 }
1337
1338 static int lmv_select_statfs_mdt(struct lmv_obd *lmv, __u32 flags)
1339 {
1340         int i;
1341
1342         if (flags & OBD_STATFS_FOR_MDT0)
1343                 return 0;
1344
1345         if (lmv->lmv_statfs_start || lmv->desc.ld_tgt_count == 1)
1346                 return lmv->lmv_statfs_start;
1347
1348         /* choose initial MDT for this client */
1349         for (i = 0;; i++) {
1350                 struct lnet_process_id lnet_id;
1351                 if (LNetGetId(i, &lnet_id) == -ENOENT)
1352                         break;
1353
1354                 if (LNET_NETTYP(LNET_NIDNET(lnet_id.nid)) != LOLND) {
1355                         lmv->lmv_statfs_start =
1356                                 lnet_id.nid % lmv->desc.ld_tgt_count;
1357                         break;
1358                 }
1359         }
1360
1361         return lmv->lmv_statfs_start;
1362 }
1363
1364 static int lmv_statfs(const struct lu_env *env, struct obd_export *exp,
1365                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
1366 {
1367         struct obd_device       *obd = class_exp2obd(exp);
1368         struct lmv_obd          *lmv = &obd->u.lmv;
1369         struct obd_statfs       *temp;
1370         int                      rc = 0;
1371         __u32                    i, idx;
1372         ENTRY;
1373
1374         OBD_ALLOC(temp, sizeof(*temp));
1375         if (temp == NULL)
1376                 RETURN(-ENOMEM);
1377
1378         /* distribute statfs among MDTs */
1379         idx = lmv_select_statfs_mdt(lmv, flags);
1380
1381         for (i = 0; i < lmv->desc.ld_tgt_count; i++, idx++) {
1382                 idx = idx % lmv->desc.ld_tgt_count;
1383                 if (lmv->tgts[idx] == NULL || lmv->tgts[idx]->ltd_exp == NULL)
1384                         continue;
1385
1386                 rc = obd_statfs(env, lmv->tgts[idx]->ltd_exp, temp,
1387                                 max_age, flags);
1388                 if (rc) {
1389                         CERROR("can't stat MDS #%d (%s), error %d\n", i,
1390                                lmv->tgts[idx]->ltd_exp->exp_obd->obd_name,
1391                                rc);
1392                         GOTO(out_free_temp, rc);
1393                 }
1394
1395                 if (temp->os_state & OS_STATE_SUM ||
1396                     flags == OBD_STATFS_FOR_MDT0) {
1397                         /* reset to the last aggregated values
1398                          * and don't sum with non-aggrated data */
1399                         /* If the statfs is from mount, it needs to retrieve
1400                          * necessary information from MDT0. i.e. mount does
1401                          * not need the merged osfs from all of MDT. Also
1402                          * clients can be mounted as long as MDT0 is in
1403                          * service */
1404                         *osfs = *temp;
1405                         break;
1406                 }
1407
1408                 if (i == 0) {
1409                         *osfs = *temp;
1410                 } else {
1411                         osfs->os_bavail += temp->os_bavail;
1412                         osfs->os_blocks += temp->os_blocks;
1413                         osfs->os_ffree += temp->os_ffree;
1414                         osfs->os_files += temp->os_files;
1415                         osfs->os_granted += temp->os_granted;
1416                 }
1417         }
1418
1419         EXIT;
1420 out_free_temp:
1421         OBD_FREE(temp, sizeof(*temp));
1422         return rc;
1423 }
1424
1425 static int lmv_get_root(struct obd_export *exp, const char *fileset,
1426                         struct lu_fid *fid)
1427 {
1428         struct obd_device    *obd = exp->exp_obd;
1429         struct lmv_obd       *lmv = &obd->u.lmv;
1430         int                   rc;
1431         ENTRY;
1432
1433         rc = md_get_root(lmv->tgts[0]->ltd_exp, fileset, fid);
1434         RETURN(rc);
1435 }
1436
1437 static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid,
1438                         u64 obd_md_valid, const char *name, size_t buf_size,
1439                         struct ptlrpc_request **req)
1440 {
1441         struct obd_device      *obd = exp->exp_obd;
1442         struct lmv_obd         *lmv = &obd->u.lmv;
1443         struct lmv_tgt_desc    *tgt;
1444         int                     rc;
1445         ENTRY;
1446
1447         tgt = lmv_find_target(lmv, fid);
1448         if (IS_ERR(tgt))
1449                 RETURN(PTR_ERR(tgt));
1450
1451         rc = md_getxattr(tgt->ltd_exp, fid, obd_md_valid, name, buf_size, req);
1452
1453         RETURN(rc);
1454 }
1455
1456 static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid,
1457                         u64 obd_md_valid, const char *name,
1458                         const void *value, size_t value_size,
1459                         unsigned int xattr_flags, u32 suppgid,
1460                         struct ptlrpc_request **req)
1461 {
1462         struct obd_device      *obd = exp->exp_obd;
1463         struct lmv_obd         *lmv = &obd->u.lmv;
1464         struct lmv_tgt_desc    *tgt;
1465         int                     rc;
1466         ENTRY;
1467
1468         tgt = lmv_find_target(lmv, fid);
1469         if (IS_ERR(tgt))
1470                 RETURN(PTR_ERR(tgt));
1471
1472         rc = md_setxattr(tgt->ltd_exp, fid, obd_md_valid, name,
1473                          value, value_size, xattr_flags, suppgid, req);
1474
1475         RETURN(rc);
1476 }
1477
1478 static int lmv_getattr(struct obd_export *exp, struct md_op_data *op_data,
1479                        struct ptlrpc_request **request)
1480 {
1481         struct obd_device       *obd = exp->exp_obd;
1482         struct lmv_obd          *lmv = &obd->u.lmv;
1483         struct lmv_tgt_desc     *tgt;
1484         int                      rc;
1485         ENTRY;
1486
1487         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1488         if (IS_ERR(tgt))
1489                 RETURN(PTR_ERR(tgt));
1490
1491         if (op_data->op_flags & MF_GET_MDT_IDX) {
1492                 op_data->op_mds = tgt->ltd_idx;
1493                 RETURN(0);
1494         }
1495
1496         rc = md_getattr(tgt->ltd_exp, op_data, request);
1497
1498         RETURN(rc);
1499 }
1500
1501 static int lmv_null_inode(struct obd_export *exp, const struct lu_fid *fid)
1502 {
1503         struct obd_device   *obd = exp->exp_obd;
1504         struct lmv_obd      *lmv = &obd->u.lmv;
1505         __u32                i;
1506         ENTRY;
1507
1508         CDEBUG(D_INODE, "CBDATA for "DFID"\n", PFID(fid));
1509
1510         /*
1511          * With DNE every object can have two locks in different namespaces:
1512          * lookup lock in space of MDT storing direntry and update/open lock in
1513          * space of MDT storing inode.
1514          */
1515         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1516                 if (lmv->tgts[i] == NULL || lmv->tgts[i]->ltd_exp == NULL)
1517                         continue;
1518                 md_null_inode(lmv->tgts[i]->ltd_exp, fid);
1519         }
1520
1521         RETURN(0);
1522 }
1523
1524 static int lmv_close(struct obd_export *exp, struct md_op_data *op_data,
1525                      struct md_open_data *mod, struct ptlrpc_request **request)
1526 {
1527         struct obd_device     *obd = exp->exp_obd;
1528         struct lmv_obd        *lmv = &obd->u.lmv;
1529         struct lmv_tgt_desc   *tgt;
1530         int                    rc;
1531         ENTRY;
1532
1533         tgt = lmv_find_target(lmv, &op_data->op_fid1);
1534         if (IS_ERR(tgt))
1535                 RETURN(PTR_ERR(tgt));
1536
1537         CDEBUG(D_INODE, "CLOSE "DFID"\n", PFID(&op_data->op_fid1));
1538         rc = md_close(tgt->ltd_exp, op_data, mod, request);
1539         RETURN(rc);
1540 }
1541
1542 /**
1543  * Choosing the MDT by name or FID in @op_data.
1544  * For non-striped directory, it will locate MDT by fid.
1545  * For striped-directory, it will locate MDT by name. And also
1546  * it will reset op_fid1 with the FID of the choosen stripe.
1547  **/
1548 struct lmv_tgt_desc *
1549 lmv_locate_target_for_name(struct lmv_obd *lmv, struct lmv_stripe_md *lsm,
1550                            const char *name, int namelen, struct lu_fid *fid,
1551                            u32 *mds)
1552 {
1553         struct lmv_tgt_desc     *tgt;
1554         const struct lmv_oinfo  *oinfo;
1555
1556         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_NAME_HASH)) {
1557                 if (cfs_fail_val >= lsm->lsm_md_stripe_count)
1558                         RETURN(ERR_PTR(-EBADF));
1559                 oinfo = &lsm->lsm_md_oinfo[cfs_fail_val];
1560         } else {
1561                 oinfo = lsm_name_to_stripe_info(lsm, name, namelen);
1562                 if (IS_ERR(oinfo))
1563                         RETURN(ERR_CAST(oinfo));
1564         }
1565
1566         if (fid != NULL)
1567                 *fid = oinfo->lmo_fid;
1568         if (mds != NULL)
1569                 *mds = oinfo->lmo_mds;
1570
1571         tgt = lmv_get_target(lmv, oinfo->lmo_mds, NULL);
1572
1573         CDEBUG(D_INFO, "locate on mds %u "DFID"\n", oinfo->lmo_mds,
1574                PFID(&oinfo->lmo_fid));
1575         return tgt;
1576 }
1577
1578 /**
1579  * Locate mds by fid or name
1580  *
1581  * For striped directory (lsm != NULL), it will locate the stripe
1582  * by name hash (see lsm_name_to_stripe_info()). Note: if the hash_type
1583  * is unknown, it will return -EBADFD, and lmv_intent_lookup might need
1584  * walk through all of stripes to locate the entry.
1585  *
1586  * For normal direcotry, it will locate MDS by FID directly.
1587  * \param[in] lmv       LMV device
1588  * \param[in] op_data   client MD stack parameters, name, namelen
1589  *                      mds_num etc.
1590  * \param[in] fid       object FID used to locate MDS.
1591  *
1592  * retval               pointer to the lmv_tgt_desc if succeed.
1593  *                      ERR_PTR(errno) if failed.
1594  */
1595 struct lmv_tgt_desc*
1596 lmv_locate_mds(struct lmv_obd *lmv, struct md_op_data *op_data,
1597                struct lu_fid *fid)
1598 {
1599         struct lmv_stripe_md    *lsm = op_data->op_mea1;
1600         struct lmv_tgt_desc     *tgt;
1601
1602         /* During creating VOLATILE file, it should honor the mdt
1603          * index if the file under striped dir is being restored, see
1604          * ct_restore(). */
1605         if (op_data->op_bias & MDS_CREATE_VOLATILE &&
1606             (int)op_data->op_mds != -1) {
1607                 int i;
1608                 tgt = lmv_get_target(lmv, op_data->op_mds, NULL);
1609                 if (IS_ERR(tgt))
1610                         return tgt;
1611
1612                 if (lsm != NULL) {
1613                         /* refill the right parent fid */
1614                         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
1615                                 struct lmv_oinfo *oinfo;
1616
1617                                 oinfo = &lsm->lsm_md_oinfo[i];
1618                                 if (oinfo->lmo_mds == op_data->op_mds) {
1619                                         *fid = oinfo->lmo_fid;
1620                                         break;
1621                                 }
1622                         }
1623
1624                         if (i == lsm->lsm_md_stripe_count)
1625                                 *fid = lsm->lsm_md_oinfo[0].lmo_fid;
1626                 }
1627
1628                 return tgt;
1629         }
1630
1631         if (lsm == NULL || op_data->op_namelen == 0) {
1632                 tgt = lmv_find_target(lmv, fid);
1633                 if (IS_ERR(tgt))
1634                         return tgt;
1635
1636                 op_data->op_mds = tgt->ltd_idx;
1637                 return tgt;
1638         }
1639
1640         return lmv_locate_target_for_name(lmv, lsm, op_data->op_name,
1641                                           op_data->op_namelen, fid,
1642                                           &op_data->op_mds);
1643 }
1644
1645 int lmv_create(struct obd_export *exp, struct md_op_data *op_data,
1646                 const void *data, size_t datalen, umode_t mode, uid_t uid,
1647                 gid_t gid, cfs_cap_t cap_effective, __u64 rdev,
1648                 struct ptlrpc_request **request)
1649 {
1650         struct obd_device       *obd = exp->exp_obd;
1651         struct lmv_obd          *lmv = &obd->u.lmv;
1652         struct lmv_tgt_desc     *tgt;
1653         int                      rc;
1654         ENTRY;
1655
1656         if (!lmv->desc.ld_active_tgt_count)
1657                 RETURN(-EIO);
1658
1659         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
1660         if (IS_ERR(tgt))
1661                 RETURN(PTR_ERR(tgt));
1662
1663         CDEBUG(D_INODE, "CREATE name '%.*s' on "DFID" -> mds #%x\n",
1664                 (int)op_data->op_namelen, op_data->op_name,
1665                 PFID(&op_data->op_fid1), op_data->op_mds);
1666
1667         rc = lmv_fid_alloc(NULL, exp, &op_data->op_fid2, op_data);
1668         if (rc)
1669                 RETURN(rc);
1670         if (exp_connect_flags(exp) & OBD_CONNECT_DIR_STRIPE) {
1671                 /* Send the create request to the MDT where the object
1672                  * will be located */
1673                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
1674                 if (IS_ERR(tgt))
1675                         RETURN(PTR_ERR(tgt));
1676
1677                 op_data->op_mds = tgt->ltd_idx;
1678         } else {
1679                 CDEBUG(D_CONFIG, "Server doesn't support striped dirs\n");
1680         }
1681
1682         CDEBUG(D_INODE, "CREATE obj "DFID" -> mds #%x\n",
1683                PFID(&op_data->op_fid2), op_data->op_mds);
1684
1685         op_data->op_flags |= MF_MDC_CANCEL_FID1;
1686         rc = md_create(tgt->ltd_exp, op_data, data, datalen, mode, uid, gid,
1687                        cap_effective, rdev, request);
1688         if (rc == 0) {
1689                 if (*request == NULL)
1690                         RETURN(rc);
1691                 CDEBUG(D_INODE, "Created - "DFID"\n", PFID(&op_data->op_fid2));
1692         }
1693         RETURN(rc);
1694 }
1695
1696 static int
1697 lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo,
1698             const union ldlm_policy_data *policy, struct md_op_data *op_data,
1699             struct lustre_handle *lockh, __u64 extra_lock_flags)
1700 {
1701         struct obd_device        *obd = exp->exp_obd;
1702         struct lmv_obd           *lmv = &obd->u.lmv;
1703         struct lmv_tgt_desc      *tgt;
1704         int                       rc;
1705         ENTRY;
1706
1707         CDEBUG(D_INODE, "ENQUEUE on "DFID"\n", PFID(&op_data->op_fid1));
1708
1709         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
1710         if (IS_ERR(tgt))
1711                 RETURN(PTR_ERR(tgt));
1712
1713         CDEBUG(D_INODE, "ENQUEUE on "DFID" -> mds #%u\n",
1714                PFID(&op_data->op_fid1), tgt->ltd_idx);
1715
1716         rc = md_enqueue(tgt->ltd_exp, einfo, policy, op_data, lockh,
1717                         extra_lock_flags);
1718
1719         RETURN(rc);
1720 }
1721
1722 static int
1723 lmv_getattr_name(struct obd_export *exp,struct md_op_data *op_data,
1724                  struct ptlrpc_request **preq)
1725 {
1726         struct ptlrpc_request   *req = NULL;
1727         struct obd_device       *obd = exp->exp_obd;
1728         struct lmv_obd          *lmv = &obd->u.lmv;
1729         struct lmv_tgt_desc     *tgt;
1730         struct mdt_body         *body;
1731         int                      rc;
1732         ENTRY;
1733
1734         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
1735         if (IS_ERR(tgt))
1736                 RETURN(PTR_ERR(tgt));
1737
1738         CDEBUG(D_INODE, "GETATTR_NAME for %*s on "DFID" -> mds #%d\n",
1739                 (int)op_data->op_namelen, op_data->op_name,
1740                 PFID(&op_data->op_fid1), tgt->ltd_idx);
1741
1742         rc = md_getattr_name(tgt->ltd_exp, op_data, preq);
1743         if (rc != 0)
1744                 RETURN(rc);
1745
1746         body = req_capsule_server_get(&(*preq)->rq_pill, &RMF_MDT_BODY);
1747         LASSERT(body != NULL);
1748
1749         if (body->mbo_valid & OBD_MD_MDS) {
1750                 struct lu_fid rid = body->mbo_fid1;
1751                 CDEBUG(D_INODE, "Request attrs for "DFID"\n",
1752                        PFID(&rid));
1753
1754                 tgt = lmv_find_target(lmv, &rid);
1755                 if (IS_ERR(tgt)) {
1756                         ptlrpc_req_finished(*preq);
1757                         preq = NULL;
1758                         RETURN(PTR_ERR(tgt));
1759                 }
1760
1761                 op_data->op_fid1 = rid;
1762                 op_data->op_valid |= OBD_MD_FLCROSSREF;
1763                 op_data->op_namelen = 0;
1764                 op_data->op_name = NULL;
1765                 rc = md_getattr_name(tgt->ltd_exp, op_data, &req);
1766                 ptlrpc_req_finished(*preq);
1767                 *preq = req;
1768         }
1769
1770         RETURN(rc);
1771 }
1772
1773 #define md_op_data_fid(op_data, fl)                     \
1774         (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \
1775          fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \
1776          fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \
1777          fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \
1778          NULL)
1779
1780 static int lmv_early_cancel(struct obd_export *exp, struct lmv_tgt_desc *tgt,
1781                             struct md_op_data *op_data, __u32 op_tgt,
1782                             enum ldlm_mode mode, int bits, int flag)
1783 {
1784         struct lu_fid *fid = md_op_data_fid(op_data, flag);
1785         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
1786         union ldlm_policy_data policy = { { 0 } };
1787         int rc = 0;
1788         ENTRY;
1789
1790         if (!fid_is_sane(fid))
1791                 RETURN(0);
1792
1793         if (tgt == NULL) {
1794                 tgt = lmv_find_target(lmv, fid);
1795                 if (IS_ERR(tgt))
1796                         RETURN(PTR_ERR(tgt));
1797         }
1798
1799         if (tgt->ltd_idx != op_tgt) {
1800                 CDEBUG(D_INODE, "EARLY_CANCEL on "DFID"\n", PFID(fid));
1801                 policy.l_inodebits.bits = bits;
1802                 rc = md_cancel_unused(tgt->ltd_exp, fid, &policy,
1803                                       mode, LCF_ASYNC, NULL);
1804         } else {
1805                 CDEBUG(D_INODE,
1806                        "EARLY_CANCEL skip operation target %d on "DFID"\n",
1807                        op_tgt, PFID(fid));
1808                 op_data->op_flags |= flag;
1809                 rc = 0;
1810         }
1811
1812         RETURN(rc);
1813 }
1814
1815 /*
1816  * llite passes fid of an target inode in op_data->op_fid1 and id of directory in
1817  * op_data->op_fid2
1818  */
1819 static int lmv_link(struct obd_export *exp, struct md_op_data *op_data,
1820                     struct ptlrpc_request **request)
1821 {
1822         struct obd_device       *obd = exp->exp_obd;
1823         struct lmv_obd          *lmv = &obd->u.lmv;
1824         struct lmv_tgt_desc     *tgt;
1825         int                      rc;
1826         ENTRY;
1827
1828         LASSERT(op_data->op_namelen != 0);
1829
1830         CDEBUG(D_INODE, "LINK "DFID":%*s to "DFID"\n",
1831                PFID(&op_data->op_fid2), (int)op_data->op_namelen,
1832                op_data->op_name, PFID(&op_data->op_fid1));
1833
1834         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
1835         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
1836         op_data->op_cap = cfs_curproc_cap_pack();
1837         if (op_data->op_mea2 != NULL) {
1838                 struct lmv_stripe_md    *lsm = op_data->op_mea2;
1839                 const struct lmv_oinfo  *oinfo;
1840
1841                 oinfo = lsm_name_to_stripe_info(lsm, op_data->op_name,
1842                                                 op_data->op_namelen);
1843                 if (IS_ERR(oinfo))
1844                         RETURN(PTR_ERR(oinfo));
1845
1846                 op_data->op_fid2 = oinfo->lmo_fid;
1847         }
1848
1849         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid2);
1850         if (IS_ERR(tgt))
1851                 RETURN(PTR_ERR(tgt));
1852
1853         /*
1854          * Cancel UPDATE lock on child (fid1).
1855          */
1856         op_data->op_flags |= MF_MDC_CANCEL_FID2;
1857         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_idx, LCK_EX,
1858                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1859         if (rc != 0)
1860                 RETURN(rc);
1861
1862         rc = md_link(tgt->ltd_exp, op_data, request);
1863
1864         RETURN(rc);
1865 }
1866
1867 static int lmv_migrate(struct obd_export *exp, struct md_op_data *op_data,
1868                         const char *name, size_t namelen,
1869                         struct ptlrpc_request **request)
1870 {
1871         struct obd_device *obd = exp->exp_obd;
1872         struct lmv_obd *lmv = &obd->u.lmv;
1873         struct lmv_stripe_md *lsm = op_data->op_mea1;
1874         struct lmv_tgt_desc *parent_tgt;
1875         struct lmv_tgt_desc *sp_tgt;
1876         struct lmv_tgt_desc *tp_tgt = NULL;
1877         struct lmv_tgt_desc *child_tgt;
1878         struct lmv_tgt_desc *tgt;
1879         struct lu_fid target_fid;
1880         int rc;
1881
1882         ENTRY;
1883
1884         LASSERT(op_data->op_cli_flags & CLI_MIGRATE);
1885         LASSERTF(fid_is_sane(&op_data->op_fid3), "invalid FID "DFID"\n",
1886                  PFID(&op_data->op_fid3));
1887
1888         CDEBUG(D_INODE, "MIGRATE "DFID"/%.*s\n",
1889                PFID(&op_data->op_fid1), (int)namelen, name);
1890
1891         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
1892         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
1893         op_data->op_cap = cfs_curproc_cap_pack();
1894
1895         parent_tgt = lmv_find_target(lmv, &op_data->op_fid1);
1896         if (IS_ERR(parent_tgt))
1897                 RETURN(PTR_ERR(parent_tgt));
1898
1899         if (lsm) {
1900                 __u32 hash_type = lsm->lsm_md_hash_type;
1901                 __u32 stripe_count = lsm->lsm_md_stripe_count;
1902
1903                 /*
1904                  * old stripes are appended after new stripes for migrating
1905                  * directory.
1906                  */
1907                 if (lsm->lsm_md_hash_type & LMV_HASH_FLAG_MIGRATION) {
1908                         hash_type = lsm->lsm_md_migrate_hash;
1909                         stripe_count -= lsm->lsm_md_migrate_offset;
1910                 }
1911
1912                 rc = lmv_name_to_stripe_index(hash_type, stripe_count, name,
1913                                               namelen);
1914                 if (rc < 0)
1915                         RETURN(rc);
1916
1917                 if (lsm->lsm_md_hash_type & LMV_HASH_FLAG_MIGRATION)
1918                         rc += lsm->lsm_md_migrate_offset;
1919
1920                 /* save it in fid4 temporarily for early cancel */
1921                 op_data->op_fid4 = lsm->lsm_md_oinfo[rc].lmo_fid;
1922                 sp_tgt = lmv_get_target(lmv, lsm->lsm_md_oinfo[rc].lmo_mds,
1923                                         NULL);
1924                 if (IS_ERR(sp_tgt))
1925                         RETURN(PTR_ERR(sp_tgt));
1926
1927                 /*
1928                  * if parent is being migrated too, fill op_fid2 with target
1929                  * stripe fid, otherwise the target stripe is not created yet.
1930                  */
1931                 if (lsm->lsm_md_hash_type & LMV_HASH_FLAG_MIGRATION) {
1932                         hash_type = lsm->lsm_md_hash_type &
1933                                     ~LMV_HASH_FLAG_MIGRATION;
1934                         stripe_count = lsm->lsm_md_migrate_offset;
1935
1936                         rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1937                                                       name, namelen);
1938                         if (rc < 0)
1939                                 RETURN(rc);
1940
1941                         op_data->op_fid2 = lsm->lsm_md_oinfo[rc].lmo_fid;
1942                         tp_tgt = lmv_get_target(lmv,
1943                                                 lsm->lsm_md_oinfo[rc].lmo_mds,
1944                                                 NULL);
1945                         if (IS_ERR(tp_tgt))
1946                                 RETURN(PTR_ERR(tp_tgt));
1947                 }
1948         } else {
1949                 sp_tgt = parent_tgt;
1950         }
1951
1952         child_tgt = lmv_find_target(lmv, &op_data->op_fid3);
1953         if (IS_ERR(child_tgt))
1954                 RETURN(PTR_ERR(child_tgt));
1955
1956         rc = lmv_fid_alloc(NULL, exp, &target_fid, op_data);
1957         if (rc)
1958                 RETURN(rc);
1959
1960         /*
1961          * for directory, send migrate request to the MDT where the object will
1962          * be migrated to, because we can't create a striped directory remotely.
1963          *
1964          * otherwise, send to the MDT where source is located because regular
1965          * file may open lease.
1966          *
1967          * NB. if MDT doesn't support DIR_MIGRATE, send to source MDT too for
1968          * backward compatibility.
1969          */
1970         if (S_ISDIR(op_data->op_mode) &&
1971             (exp_connect_flags2(exp) & OBD_CONNECT2_DIR_MIGRATE)) {
1972                 tgt = lmv_find_target(lmv, &target_fid);
1973                 if (IS_ERR(tgt))
1974                         RETURN(PTR_ERR(tgt));
1975         } else {
1976                 tgt = child_tgt;
1977         }
1978
1979         /* cancel UPDATE lock of parent master object */
1980         rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_idx, LCK_EX,
1981                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
1982         if (rc)
1983                 RETURN(rc);
1984
1985         /* cancel UPDATE lock of source parent */
1986         if (sp_tgt != parent_tgt) {
1987                 /*
1988                  * migrate RPC packs master object FID, because we can only pack
1989                  * two FIDs in reint RPC, but MDS needs to know both source
1990                  * parent and target parent, and it will obtain them from master
1991                  * FID and LMV, the other FID in RPC is kept for target.
1992                  *
1993                  * since this FID is not passed to MDC, cancel it anyway.
1994                  */
1995                 rc = lmv_early_cancel(exp, sp_tgt, op_data, -1, LCK_EX,
1996                                       MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID4);
1997                 if (rc)
1998                         RETURN(rc);
1999
2000                 op_data->op_flags &= ~MF_MDC_CANCEL_FID4;
2001         }
2002         op_data->op_fid4 = target_fid;
2003
2004         /* cancel UPDATE locks of target parent */
2005         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_idx, LCK_EX,
2006                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2007         if (rc)
2008                 RETURN(rc);
2009
2010         /* cancel LOOKUP lock of source if source is remote object */
2011         if (child_tgt != sp_tgt) {
2012                 rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_idx,
2013                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2014                                       MF_MDC_CANCEL_FID3);
2015                 if (rc)
2016                         RETURN(rc);
2017         }
2018
2019         /* cancel ELC locks of source */
2020         rc = lmv_early_cancel(exp, child_tgt, op_data, tgt->ltd_idx, LCK_EX,
2021                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2022         if (rc)
2023                 RETURN(rc);
2024
2025         rc = md_rename(tgt->ltd_exp, op_data, name, namelen, NULL, 0, request);
2026
2027         RETURN(rc);
2028 }
2029
2030 static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data,
2031                       const char *old, size_t oldlen,
2032                       const char *new, size_t newlen,
2033                       struct ptlrpc_request **request)
2034 {
2035         struct obd_device *obd = exp->exp_obd;
2036         struct lmv_obd *lmv = &obd->u.lmv;
2037         struct lmv_stripe_md *lsm = op_data->op_mea1;
2038         struct lmv_tgt_desc *sp_tgt;
2039         struct lmv_tgt_desc *tp_tgt = NULL;
2040         struct lmv_tgt_desc *tgt;
2041         struct mdt_body *body;
2042         int rc;
2043
2044         ENTRY;
2045
2046         LASSERT(oldlen != 0);
2047
2048         if (op_data->op_cli_flags & CLI_MIGRATE) {
2049                 rc = lmv_migrate(exp, op_data, old, oldlen, request);
2050                 RETURN(rc);
2051         }
2052
2053         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2054         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2055         op_data->op_cap = cfs_curproc_cap_pack();
2056
2057         CDEBUG(D_INODE, "RENAME "DFID"/%.*s to "DFID"/%.*s\n",
2058                 PFID(&op_data->op_fid1), (int)oldlen, old,
2059                 PFID(&op_data->op_fid2), (int)newlen, new);
2060
2061         if (lsm)
2062                 sp_tgt = lmv_locate_target_for_name(lmv, lsm, old, oldlen,
2063                                                     &op_data->op_fid1,
2064                                                     &op_data->op_mds);
2065         else
2066                 sp_tgt = lmv_find_target(lmv, &op_data->op_fid1);
2067         if (IS_ERR(sp_tgt))
2068                 RETURN(PTR_ERR(sp_tgt));
2069
2070         lsm = op_data->op_mea2;
2071         if (lsm)
2072                 tp_tgt = lmv_locate_target_for_name(lmv, lsm, new, newlen,
2073                                                     &op_data->op_fid2,
2074                                                     &op_data->op_mds);
2075         else
2076                 tp_tgt = lmv_find_target(lmv, &op_data->op_fid2);
2077         if (IS_ERR(tp_tgt))
2078                 RETURN(PTR_ERR(tp_tgt));
2079
2080         /* Since the target child might be destroyed, and it might become
2081          * orphan, and we can only check orphan on the local MDT right now, so
2082          * we send rename request to the MDT where target child is located. If
2083          * target child does not exist, then it will send the request to the
2084          * target parent */
2085         if (fid_is_sane(&op_data->op_fid4)) {
2086                 tgt = lmv_find_target(lmv, &op_data->op_fid4);
2087                 if (IS_ERR(tgt))
2088                         RETURN(PTR_ERR(tgt));
2089         } else {
2090                 tgt = tp_tgt;
2091         }
2092
2093         op_data->op_flags |= MF_MDC_CANCEL_FID4;
2094
2095         /* cancel UPDATE locks of source parent */
2096         rc = lmv_early_cancel(exp, sp_tgt, op_data, tgt->ltd_idx, LCK_EX,
2097                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1);
2098         if (rc != 0)
2099                 RETURN(rc);
2100
2101         /* cancel UPDATE locks of target parent */
2102         rc = lmv_early_cancel(exp, tp_tgt, op_data, tgt->ltd_idx, LCK_EX,
2103                               MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2);
2104         if (rc != 0)
2105                 RETURN(rc);
2106
2107         if (fid_is_sane(&op_data->op_fid3)) {
2108                 struct lmv_tgt_desc *src_tgt;
2109
2110                 src_tgt = lmv_find_target(lmv, &op_data->op_fid3);
2111                 if (IS_ERR(src_tgt))
2112                         RETURN(PTR_ERR(src_tgt));
2113
2114                 /* cancel LOOKUP lock of source on source parent */
2115                 if (src_tgt != sp_tgt) {
2116                         rc = lmv_early_cancel(exp, sp_tgt, op_data,
2117                                               tgt->ltd_idx, LCK_EX,
2118                                               MDS_INODELOCK_LOOKUP,
2119                                               MF_MDC_CANCEL_FID3);
2120                         if (rc != 0)
2121                                 RETURN(rc);
2122                 }
2123
2124                 /* cancel ELC locks of source */
2125                 rc = lmv_early_cancel(exp, src_tgt, op_data, tgt->ltd_idx,
2126                                       LCK_EX, MDS_INODELOCK_ELC,
2127                                       MF_MDC_CANCEL_FID3);
2128                 if (rc != 0)
2129                         RETURN(rc);
2130         }
2131
2132 retry_rename:
2133         if (fid_is_sane(&op_data->op_fid4)) {
2134                 /* cancel LOOKUP lock of target on target parent */
2135                 if (tgt != tp_tgt) {
2136                         rc = lmv_early_cancel(exp, tp_tgt, op_data,
2137                                               tgt->ltd_idx, LCK_EX,
2138                                               MDS_INODELOCK_LOOKUP,
2139                                               MF_MDC_CANCEL_FID4);
2140                         if (rc != 0)
2141                                 RETURN(rc);
2142                 }
2143         }
2144
2145         rc = md_rename(tgt->ltd_exp, op_data, old, oldlen, new, newlen,
2146                         request);
2147
2148         if (rc != 0 && rc != -EXDEV)
2149                 RETURN(rc);
2150
2151         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2152         if (body == NULL)
2153                 RETURN(-EPROTO);
2154
2155         /* Not cross-ref case, just get out of here. */
2156         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2157                 RETURN(rc);
2158
2159         CDEBUG(D_INODE, "%s: try rename to another MDT for "DFID"\n",
2160                exp->exp_obd->obd_name, PFID(&body->mbo_fid1));
2161
2162         op_data->op_fid4 = body->mbo_fid1;
2163         ptlrpc_req_finished(*request);
2164         *request = NULL;
2165
2166         tgt = lmv_find_target(lmv, &op_data->op_fid4);
2167         if (IS_ERR(tgt))
2168                 RETURN(PTR_ERR(tgt));
2169
2170         goto retry_rename;
2171 }
2172
2173 static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data,
2174                        void *ea, size_t ealen, struct ptlrpc_request **request)
2175 {
2176         struct obd_device       *obd = exp->exp_obd;
2177         struct lmv_obd          *lmv = &obd->u.lmv;
2178         struct lmv_tgt_desc     *tgt;
2179         int                      rc = 0;
2180         ENTRY;
2181
2182         CDEBUG(D_INODE, "SETATTR for "DFID", valid 0x%x\n",
2183                PFID(&op_data->op_fid1), op_data->op_attr.ia_valid);
2184
2185         op_data->op_flags |= MF_MDC_CANCEL_FID1;
2186         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2187         if (IS_ERR(tgt))
2188                 RETURN(PTR_ERR(tgt));
2189
2190         rc = md_setattr(tgt->ltd_exp, op_data, ea, ealen, request);
2191
2192         RETURN(rc);
2193 }
2194
2195 static int lmv_fsync(struct obd_export *exp, const struct lu_fid *fid,
2196                      struct ptlrpc_request **request)
2197 {
2198         struct obd_device       *obd = exp->exp_obd;
2199         struct lmv_obd          *lmv = &obd->u.lmv;
2200         struct lmv_tgt_desc     *tgt;
2201         int                      rc;
2202         ENTRY;
2203
2204         tgt = lmv_find_target(lmv, fid);
2205         if (IS_ERR(tgt))
2206                 RETURN(PTR_ERR(tgt));
2207
2208         rc = md_fsync(tgt->ltd_exp, fid, request);
2209         RETURN(rc);
2210 }
2211
2212 struct stripe_dirent {
2213         struct page             *sd_page;
2214         struct lu_dirpage       *sd_dp;
2215         struct lu_dirent        *sd_ent;
2216         bool                     sd_eof;
2217 };
2218
2219 struct lmv_dir_ctxt {
2220         struct lmv_obd          *ldc_lmv;
2221         struct md_op_data       *ldc_op_data;
2222         struct md_callback      *ldc_cb_op;
2223         __u64                    ldc_hash;
2224         int                      ldc_count;
2225         struct stripe_dirent     ldc_stripes[0];
2226 };
2227
2228 static inline void stripe_dirent_unload(struct stripe_dirent *stripe)
2229 {
2230         if (stripe->sd_page) {
2231                 kunmap(stripe->sd_page);
2232                 put_page(stripe->sd_page);
2233                 stripe->sd_page = NULL;
2234                 stripe->sd_ent = NULL;
2235         }
2236 }
2237
2238 static inline void put_lmv_dir_ctxt(struct lmv_dir_ctxt *ctxt)
2239 {
2240         int i;
2241
2242         for (i = 0; i < ctxt->ldc_count; i++)
2243                 stripe_dirent_unload(&ctxt->ldc_stripes[i]);
2244 }
2245
2246 /* if @ent is dummy, or . .., get next */
2247 static struct lu_dirent *stripe_dirent_get(struct lmv_dir_ctxt *ctxt,
2248                                            struct lu_dirent *ent,
2249                                            int stripe_index)
2250 {
2251         for (; ent; ent = lu_dirent_next(ent)) {
2252                 /* Skip dummy entry */
2253                 if (le16_to_cpu(ent->lde_namelen) == 0)
2254                         continue;
2255
2256                 /* skip . and .. for other stripes */
2257                 if (stripe_index &&
2258                     (strncmp(ent->lde_name, ".",
2259                              le16_to_cpu(ent->lde_namelen)) == 0 ||
2260                      strncmp(ent->lde_name, "..",
2261                              le16_to_cpu(ent->lde_namelen)) == 0))
2262                         continue;
2263
2264                 if (le64_to_cpu(ent->lde_hash) >= ctxt->ldc_hash)
2265                         break;
2266         }
2267
2268         return ent;
2269 }
2270
2271 static struct lu_dirent *stripe_dirent_load(struct lmv_dir_ctxt *ctxt,
2272                                             struct stripe_dirent *stripe,
2273                                             int stripe_index)
2274 {
2275         struct md_op_data *op_data = ctxt->ldc_op_data;
2276         struct lmv_oinfo *oinfo;
2277         struct lu_fid fid = op_data->op_fid1;
2278         struct inode *inode = op_data->op_data;
2279         struct lmv_tgt_desc *tgt;
2280         struct lu_dirent *ent = stripe->sd_ent;
2281         __u64 hash = ctxt->ldc_hash;
2282         int rc = 0;
2283
2284         ENTRY;
2285
2286         LASSERT(stripe == &ctxt->ldc_stripes[stripe_index]);
2287         LASSERT(!ent);
2288
2289         do {
2290                 if (stripe->sd_page) {
2291                         __u64 end = le64_to_cpu(stripe->sd_dp->ldp_hash_end);
2292
2293                         /* @hash should be the last dirent hash */
2294                         LASSERTF(hash <= end,
2295                                  "ctxt@%p stripe@%p hash %llx end %llx\n",
2296                                  ctxt, stripe, hash, end);
2297                         /* unload last page */
2298                         stripe_dirent_unload(stripe);
2299                         /* eof */
2300                         if (end == MDS_DIR_END_OFF) {
2301                                 stripe->sd_eof = true;
2302                                 break;
2303                         }
2304                         hash = end;
2305                 }
2306
2307                 oinfo = &op_data->op_mea1->lsm_md_oinfo[stripe_index];
2308                 tgt = lmv_get_target(ctxt->ldc_lmv, oinfo->lmo_mds, NULL);
2309                 if (IS_ERR(tgt)) {
2310                         rc = PTR_ERR(tgt);
2311                         break;
2312                 }
2313
2314                 /* op_data is shared by stripes, reset after use */
2315                 op_data->op_fid1 = oinfo->lmo_fid;
2316                 op_data->op_fid2 = oinfo->lmo_fid;
2317                 op_data->op_data = oinfo->lmo_root;
2318
2319                 rc = md_read_page(tgt->ltd_exp, op_data, ctxt->ldc_cb_op, hash,
2320                                   &stripe->sd_page);
2321
2322                 op_data->op_fid1 = fid;
2323                 op_data->op_fid2 = fid;
2324                 op_data->op_data = inode;
2325
2326                 if (rc)
2327                         break;
2328
2329                 stripe->sd_dp = page_address(stripe->sd_page);
2330                 ent = stripe_dirent_get(ctxt, lu_dirent_start(stripe->sd_dp),
2331                                         stripe_index);
2332                 /* in case a page filled with ., .. and dummy, read next */
2333         } while (!ent);
2334
2335         stripe->sd_ent = ent;
2336         if (rc) {
2337                 LASSERT(!ent);
2338                 /* treat error as eof, so dir can be partially accessed */
2339                 stripe->sd_eof = true;
2340                 LCONSOLE_WARN("dir "DFID" stripe %d readdir failed: %d, "
2341                               "directory is partially accessed!\n",
2342                               PFID(&ctxt->ldc_op_data->op_fid1), stripe_index,
2343                               rc);
2344         }
2345
2346         RETURN(ent);
2347 }
2348
2349 static int lmv_file_resync(struct obd_export *exp, struct md_op_data *data)
2350 {
2351         struct obd_device       *obd = exp->exp_obd;
2352         struct lmv_obd          *lmv = &obd->u.lmv;
2353         struct lmv_tgt_desc     *tgt;
2354         int                      rc;
2355         ENTRY;
2356
2357         rc = lmv_check_connect(obd);
2358         if (rc != 0)
2359                 RETURN(rc);
2360
2361         tgt = lmv_find_target(lmv, &data->op_fid1);
2362         if (IS_ERR(tgt))
2363                 RETURN(PTR_ERR(tgt));
2364
2365         data->op_flags |= MF_MDC_CANCEL_FID1;
2366         rc = md_file_resync(tgt->ltd_exp, data);
2367         RETURN(rc);
2368 }
2369
2370 /**
2371  * Get dirent with the closest hash for striped directory
2372  *
2373  * This function will search the dir entry, whose hash value is the
2374  * closest(>=) to hash from all of sub-stripes, and it is only being called
2375  * for striped directory.
2376  *
2377  * \param[in] ctxt              dir read context
2378  *
2379  * \retval                      dirent get the entry successfully
2380  *                              NULL does not get the entry, normally it means
2381  *                              it reaches the end of the directory, while read
2382  *                              stripe dirent error is ignored to allow partial
2383  *                              access.
2384  */
2385 static struct lu_dirent *lmv_dirent_next(struct lmv_dir_ctxt *ctxt)
2386 {
2387         struct stripe_dirent *stripe;
2388         struct lu_dirent *ent = NULL;
2389         int i;
2390         int min = -1;
2391
2392         /* TODO: optimize with k-way merge sort */
2393         for (i = 0; i < ctxt->ldc_count; i++) {
2394                 stripe = &ctxt->ldc_stripes[i];
2395                 if (stripe->sd_eof)
2396                         continue;
2397
2398                 if (!stripe->sd_ent) {
2399                         stripe_dirent_load(ctxt, stripe, i);
2400                         if (!stripe->sd_ent) {
2401                                 LASSERT(stripe->sd_eof);
2402                                 continue;
2403                         }
2404                 }
2405
2406                 if (min == -1 ||
2407                     le64_to_cpu(ctxt->ldc_stripes[min].sd_ent->lde_hash) >
2408                     le64_to_cpu(stripe->sd_ent->lde_hash)) {
2409                         min = i;
2410                         if (le64_to_cpu(stripe->sd_ent->lde_hash) ==
2411                             ctxt->ldc_hash)
2412                                 break;
2413                 }
2414         }
2415
2416         if (min != -1) {
2417                 stripe = &ctxt->ldc_stripes[min];
2418                 ent = stripe->sd_ent;
2419                 /* pop found dirent */
2420                 stripe->sd_ent = stripe_dirent_get(ctxt, lu_dirent_next(ent),
2421                                                    min);
2422         }
2423
2424         return ent;
2425 }
2426
2427 /**
2428  * Build dir entry page for striped directory
2429  *
2430  * This function gets one entry by @offset from a striped directory. It will
2431  * read entries from all of stripes, and choose one closest to the required
2432  * offset(&offset). A few notes
2433  * 1. skip . and .. for non-zero stripes, because there can only have one .
2434  * and .. in a directory.
2435  * 2. op_data will be shared by all of stripes, instead of allocating new
2436  * one, so need to restore before reusing.
2437  *
2438  * \param[in] exp       obd export refer to LMV
2439  * \param[in] op_data   hold those MD parameters of read_entry
2440  * \param[in] cb_op     ldlm callback being used in enqueue in mdc_read_entry
2441  * \param[in] offset    starting hash offset
2442  * \param[out] ppage    the page holding the entry. Note: because the entry
2443  *                      will be accessed in upper layer, so we need hold the
2444  *                      page until the usages of entry is finished, see
2445  *                      ll_dir_entry_next.
2446  *
2447  * retval               =0 if get entry successfully
2448  *                      <0 cannot get entry
2449  */
2450 static int lmv_striped_read_page(struct obd_export *exp,
2451                                  struct md_op_data *op_data,
2452                                  struct md_callback *cb_op,
2453                                  __u64 offset, struct page **ppage)
2454 {
2455         struct page *page = NULL;
2456         struct lu_dirpage *dp;
2457         void *start;
2458         struct lu_dirent *ent;
2459         struct lu_dirent *last_ent;
2460         int stripe_count;
2461         struct lmv_dir_ctxt *ctxt;
2462         struct lu_dirent *next = NULL;
2463         __u16 ent_size;
2464         size_t left_bytes;
2465         int rc = 0;
2466         ENTRY;
2467
2468         /* Allocate a page and read entries from all of stripes and fill
2469          * the page by hash order */
2470         page = alloc_page(GFP_KERNEL);
2471         if (!page)
2472                 RETURN(-ENOMEM);
2473
2474         /* Initialize the entry page */
2475         dp = kmap(page);
2476         memset(dp, 0, sizeof(*dp));
2477         dp->ldp_hash_start = cpu_to_le64(offset);
2478
2479         start = dp + 1;
2480         left_bytes = PAGE_SIZE - sizeof(*dp);
2481         ent = start;
2482         last_ent = ent;
2483
2484         /* initalize dir read context */
2485         stripe_count = op_data->op_mea1->lsm_md_stripe_count;
2486         OBD_ALLOC(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2487         if (!ctxt)
2488                 GOTO(free_page, rc = -ENOMEM);
2489         ctxt->ldc_lmv = &exp->exp_obd->u.lmv;
2490         ctxt->ldc_op_data = op_data;
2491         ctxt->ldc_cb_op = cb_op;
2492         ctxt->ldc_hash = offset;
2493         ctxt->ldc_count = stripe_count;
2494
2495         while (1) {
2496                 next = lmv_dirent_next(ctxt);
2497
2498                 /* end of directory */
2499                 if (!next) {
2500                         ctxt->ldc_hash = MDS_DIR_END_OFF;
2501                         break;
2502                 }
2503                 ctxt->ldc_hash = le64_to_cpu(next->lde_hash);
2504
2505                 ent_size = le16_to_cpu(next->lde_reclen);
2506
2507                 /* the last entry lde_reclen is 0, but it might not be the last
2508                  * one of this temporay dir page */
2509                 if (!ent_size)
2510                         ent_size = lu_dirent_calc_size(
2511                                         le16_to_cpu(next->lde_namelen),
2512                                         le32_to_cpu(next->lde_attrs));
2513                 /* page full */
2514                 if (ent_size > left_bytes)
2515                         break;
2516
2517                 memcpy(ent, next, ent_size);
2518
2519                 /* Replace . with master FID and Replace .. with the parent FID
2520                  * of master object */
2521                 if (strncmp(ent->lde_name, ".",
2522                             le16_to_cpu(ent->lde_namelen)) == 0 &&
2523                     le16_to_cpu(ent->lde_namelen) == 1)
2524                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid1);
2525                 else if (strncmp(ent->lde_name, "..",
2526                                    le16_to_cpu(ent->lde_namelen)) == 0 &&
2527                            le16_to_cpu(ent->lde_namelen) == 2)
2528                         fid_cpu_to_le(&ent->lde_fid, &op_data->op_fid3);
2529
2530                 CDEBUG(D_INODE, "entry %.*s hash %#llx\n",
2531                        le16_to_cpu(ent->lde_namelen), ent->lde_name,
2532                        le64_to_cpu(ent->lde_hash));
2533
2534                 left_bytes -= ent_size;
2535                 ent->lde_reclen = cpu_to_le16(ent_size);
2536                 last_ent = ent;
2537                 ent = (void *)ent + ent_size;
2538         };
2539
2540         last_ent->lde_reclen = 0;
2541
2542         if (ent == start)
2543                 dp->ldp_flags |= LDF_EMPTY;
2544         else if (ctxt->ldc_hash == le64_to_cpu(last_ent->lde_hash))
2545                 dp->ldp_flags |= LDF_COLLIDE;
2546         dp->ldp_flags = cpu_to_le32(dp->ldp_flags);
2547         dp->ldp_hash_end = cpu_to_le64(ctxt->ldc_hash);
2548
2549         put_lmv_dir_ctxt(ctxt);
2550         OBD_FREE(ctxt, offsetof(typeof(*ctxt), ldc_stripes[stripe_count]));
2551
2552         *ppage = page;
2553
2554         RETURN(0);
2555
2556 free_page:
2557         kunmap(page);
2558         __free_page(page);
2559
2560         return rc;
2561 }
2562
2563 int lmv_read_page(struct obd_export *exp, struct md_op_data *op_data,
2564                   struct md_callback *cb_op, __u64 offset,
2565                   struct page **ppage)
2566 {
2567         struct obd_device       *obd = exp->exp_obd;
2568         struct lmv_obd          *lmv = &obd->u.lmv;
2569         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2570         struct lmv_tgt_desc     *tgt;
2571         int                     rc;
2572         ENTRY;
2573
2574         if (unlikely(lsm != NULL)) {
2575                 rc = lmv_striped_read_page(exp, op_data, cb_op, offset, ppage);
2576                 RETURN(rc);
2577         }
2578
2579         tgt = lmv_find_target(lmv, &op_data->op_fid1);
2580         if (IS_ERR(tgt))
2581                 RETURN(PTR_ERR(tgt));
2582
2583         rc = md_read_page(tgt->ltd_exp, op_data, cb_op, offset, ppage);
2584
2585         RETURN(rc);
2586 }
2587
2588 /**
2589  * Unlink a file/directory
2590  *
2591  * Unlink a file or directory under the parent dir. The unlink request
2592  * usually will be sent to the MDT where the child is located, but if
2593  * the client does not have the child FID then request will be sent to the
2594  * MDT where the parent is located.
2595  *
2596  * If the parent is a striped directory then it also needs to locate which
2597  * stripe the name of the child is located, and replace the parent FID
2598  * (@op->op_fid1) with the stripe FID. Note: if the stripe is unknown,
2599  * it will walk through all of sub-stripes until the child is being
2600  * unlinked finally.
2601  *
2602  * \param[in] exp       export refer to LMV
2603  * \param[in] op_data   different parameters transferred beween client
2604  *                      MD stacks, name, namelen, FIDs etc.
2605  *                      op_fid1 is the parent FID, op_fid2 is the child
2606  *                      FID.
2607  * \param[out] request  point to the request of unlink.
2608  *
2609  * retval               0 if succeed
2610  *                      negative errno if failed.
2611  */
2612 static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data,
2613                       struct ptlrpc_request **request)
2614 {
2615         struct obd_device       *obd = exp->exp_obd;
2616         struct lmv_obd          *lmv = &obd->u.lmv;
2617         struct lmv_tgt_desc     *tgt = NULL;
2618         struct lmv_tgt_desc     *parent_tgt = NULL;
2619         struct mdt_body         *body;
2620         int                     rc;
2621         int                     stripe_index = 0;
2622         struct lmv_stripe_md    *lsm = op_data->op_mea1;
2623         ENTRY;
2624
2625 retry_unlink:
2626         /* For striped dir, we need to locate the parent as well */
2627         if (lsm != NULL) {
2628                 struct lmv_tgt_desc *tmp;
2629
2630                 LASSERT(op_data->op_name != NULL &&
2631                         op_data->op_namelen != 0);
2632
2633                 tmp = lmv_locate_target_for_name(lmv, lsm,
2634                                                  op_data->op_name,
2635                                                  op_data->op_namelen,
2636                                                  &op_data->op_fid1,
2637                                                  &op_data->op_mds);
2638
2639                 /* return -EBADFD means unknown hash type, might
2640                  * need try all sub-stripe here */
2641                 if (IS_ERR(tmp) && PTR_ERR(tmp) != -EBADFD)
2642                         RETURN(PTR_ERR(tmp));
2643
2644                 /* Note: both migrating dir and unknown hash dir need to
2645                  * try all of sub-stripes, so we need start search the
2646                  * name from stripe 0, but migrating dir is already handled
2647                  * inside lmv_locate_target_for_name(), so we only check
2648                  * unknown hash type directory here */
2649                 if (!lmv_is_known_hash_type(lsm->lsm_md_hash_type)) {
2650                         struct lmv_oinfo *oinfo;
2651
2652                         oinfo = &lsm->lsm_md_oinfo[stripe_index];
2653
2654                         op_data->op_fid1 = oinfo->lmo_fid;
2655                         op_data->op_mds = oinfo->lmo_mds;
2656                 }
2657         }
2658
2659 try_next_stripe:
2660         /* Send unlink requests to the MDT where the child is located */
2661         if (likely(!fid_is_zero(&op_data->op_fid2)))
2662                 tgt = lmv_find_target(lmv, &op_data->op_fid2);
2663         else if (lsm != NULL)
2664                 tgt = lmv_get_target(lmv, op_data->op_mds, NULL);
2665         else
2666                 tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
2667
2668         if (IS_ERR(tgt))
2669                 RETURN(PTR_ERR(tgt));
2670
2671         op_data->op_fsuid = from_kuid(&init_user_ns, current_fsuid());
2672         op_data->op_fsgid = from_kgid(&init_user_ns, current_fsgid());
2673         op_data->op_cap = cfs_curproc_cap_pack();
2674
2675         /*
2676          * If child's fid is given, cancel unused locks for it if it is from
2677          * another export than parent.
2678          *
2679          * LOOKUP lock for child (fid3) should also be cancelled on parent
2680          * tgt_tgt in mdc_unlink().
2681          */
2682         op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3;
2683
2684         /*
2685          * Cancel FULL locks on child (fid3).
2686          */
2687         parent_tgt = lmv_find_target(lmv, &op_data->op_fid1);
2688         if (IS_ERR(parent_tgt))
2689                 RETURN(PTR_ERR(parent_tgt));
2690
2691         if (parent_tgt != tgt) {
2692                 rc = lmv_early_cancel(exp, parent_tgt, op_data, tgt->ltd_idx,
2693                                       LCK_EX, MDS_INODELOCK_LOOKUP,
2694                                       MF_MDC_CANCEL_FID3);
2695         }
2696
2697         rc = lmv_early_cancel(exp, NULL, op_data, tgt->ltd_idx, LCK_EX,
2698                               MDS_INODELOCK_ELC, MF_MDC_CANCEL_FID3);
2699         if (rc != 0)
2700                 RETURN(rc);
2701
2702         CDEBUG(D_INODE, "unlink with fid="DFID"/"DFID" -> mds #%u\n",
2703                PFID(&op_data->op_fid1), PFID(&op_data->op_fid2), tgt->ltd_idx);
2704
2705         rc = md_unlink(tgt->ltd_exp, op_data, request);
2706         if (rc != 0 && rc != -EREMOTE && rc != -ENOENT)
2707                 RETURN(rc);
2708
2709         /* Try next stripe if it is needed. */
2710         if (rc == -ENOENT && lsm != NULL && lmv_need_try_all_stripes(lsm)) {
2711                 struct lmv_oinfo *oinfo;
2712
2713                 stripe_index++;
2714                 if (stripe_index >= lsm->lsm_md_stripe_count)
2715                         RETURN(rc);
2716
2717                 oinfo = &lsm->lsm_md_oinfo[stripe_index];
2718
2719                 op_data->op_fid1 = oinfo->lmo_fid;
2720                 op_data->op_mds = oinfo->lmo_mds;
2721
2722                 ptlrpc_req_finished(*request);
2723                 *request = NULL;
2724
2725                 goto try_next_stripe;
2726         }
2727
2728         body = req_capsule_server_get(&(*request)->rq_pill, &RMF_MDT_BODY);
2729         if (body == NULL)
2730                 RETURN(-EPROTO);
2731
2732         /* Not cross-ref case, just get out of here. */
2733         if (likely(!(body->mbo_valid & OBD_MD_MDS)))
2734                 RETURN(rc);
2735
2736         CDEBUG(D_INODE, "%s: try unlink to another MDT for "DFID"\n",
2737                exp->exp_obd->obd_name, PFID(&body->mbo_fid1));
2738
2739         /* This is a remote object, try remote MDT, Note: it may
2740          * try more than 1 time here, Considering following case
2741          * /mnt/lustre is root on MDT0, remote1 is on MDT1
2742          * 1. Initially A does not know where remote1 is, it send
2743          *    unlink RPC to MDT0, MDT0 return -EREMOTE, it will
2744          *    resend unlink RPC to MDT1 (retry 1st time).
2745          *
2746          * 2. During the unlink RPC in flight,
2747          *    client B mv /mnt/lustre/remote1 /mnt/lustre/remote2
2748          *    and create new remote1, but on MDT0
2749          *
2750          * 3. MDT1 get unlink RPC(from A), then do remote lock on
2751          *    /mnt/lustre, then lookup get fid of remote1, and find
2752          *    it is remote dir again, and replay -EREMOTE again.
2753          *
2754          * 4. Then A will resend unlink RPC to MDT0. (retry 2nd times).
2755          *
2756          * In theory, it might try unlimited time here, but it should
2757          * be very rare case.  */
2758         op_data->op_fid2 = body->mbo_fid1;
2759         ptlrpc_req_finished(*request);
2760         *request = NULL;
2761
2762         goto retry_unlink;
2763 }
2764
2765 static int lmv_precleanup(struct obd_device *obd)
2766 {
2767         ENTRY;
2768         libcfs_kkuc_group_rem(&obd->obd_uuid, 0, KUC_GRP_HSM);
2769         fld_client_debugfs_fini(&obd->u.lmv.lmv_fld);
2770         lprocfs_obd_cleanup(obd);
2771         lprocfs_free_md_stats(obd);
2772         RETURN(0);
2773 }
2774
2775 /**
2776  * Get by key a value associated with a LMV device.
2777  *
2778  * Dispatch request to lower-layer devices as needed.
2779  *
2780  * \param[in] env               execution environment for this thread
2781  * \param[in] exp               export for the LMV device
2782  * \param[in] keylen            length of key identifier
2783  * \param[in] key               identifier of key to get value for
2784  * \param[in] vallen            size of \a val
2785  * \param[out] val              pointer to storage location for value
2786  * \param[in] lsm               optional striping metadata of object
2787  *
2788  * \retval 0            on success
2789  * \retval negative     negated errno on failure
2790  */
2791 static int lmv_get_info(const struct lu_env *env, struct obd_export *exp,
2792                         __u32 keylen, void *key, __u32 *vallen, void *val)
2793 {
2794         struct obd_device       *obd;
2795         struct lmv_obd          *lmv;
2796         int                      rc = 0;
2797         ENTRY;
2798
2799         obd = class_exp2obd(exp);
2800         if (obd == NULL) {
2801                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2802                        exp->exp_handle.h_cookie);
2803                 RETURN(-EINVAL);
2804         }
2805
2806         lmv = &obd->u.lmv;
2807         if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) {
2808                 int i;
2809
2810                 LASSERT(*vallen == sizeof(__u32));
2811                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2812                         struct lmv_tgt_desc *tgt = lmv->tgts[i];
2813                         /*
2814                          * All tgts should be connected when this gets called.
2815                          */
2816                         if (tgt == NULL || tgt->ltd_exp == NULL)
2817                                 continue;
2818
2819                         if (!obd_get_info(env, tgt->ltd_exp, keylen, key,
2820                                           vallen, val))
2821                                 RETURN(0);
2822                 }
2823                 RETURN(-EINVAL);
2824         } else if (KEY_IS(KEY_MAX_EASIZE) ||
2825                    KEY_IS(KEY_DEFAULT_EASIZE) ||
2826                    KEY_IS(KEY_CONN_DATA)) {
2827                 /*
2828                  * Forwarding this request to first MDS, it should know LOV
2829                  * desc.
2830                  */
2831                 rc = obd_get_info(env, lmv->tgts[0]->ltd_exp, keylen, key,
2832                                   vallen, val);
2833                 if (!rc && KEY_IS(KEY_CONN_DATA))
2834                         exp->exp_connect_data = *(struct obd_connect_data *)val;
2835                 RETURN(rc);
2836         } else if (KEY_IS(KEY_TGT_COUNT)) {
2837                 *((int *)val) = lmv->desc.ld_tgt_count;
2838                 RETURN(0);
2839         }
2840
2841         CDEBUG(D_IOCTL, "Invalid key\n");
2842         RETURN(-EINVAL);
2843 }
2844
2845 /**
2846  * Asynchronously set by key a value associated with a LMV device.
2847  *
2848  * Dispatch request to lower-layer devices as needed.
2849  *
2850  * \param[in] env       execution environment for this thread
2851  * \param[in] exp       export for the LMV device
2852  * \param[in] keylen    length of key identifier
2853  * \param[in] key       identifier of key to store value for
2854  * \param[in] vallen    size of value to store
2855  * \param[in] val       pointer to data to be stored
2856  * \param[in] set       optional list of related ptlrpc requests
2857  *
2858  * \retval 0            on success
2859  * \retval negative     negated errno on failure
2860  */
2861 int lmv_set_info_async(const struct lu_env *env, struct obd_export *exp,
2862                         __u32 keylen, void *key, __u32 vallen, void *val,
2863                         struct ptlrpc_request_set *set)
2864 {
2865         struct lmv_tgt_desc     *tgt = NULL;
2866         struct obd_device       *obd;
2867         struct lmv_obd          *lmv;
2868         int rc = 0;
2869         ENTRY;
2870
2871         obd = class_exp2obd(exp);
2872         if (obd == NULL) {
2873                 CDEBUG(D_IOCTL, "Invalid client cookie %#llx\n",
2874                        exp->exp_handle.h_cookie);
2875                 RETURN(-EINVAL);
2876         }
2877         lmv = &obd->u.lmv;
2878
2879         if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) ||
2880             KEY_IS(KEY_DEFAULT_EASIZE)) {
2881                 int i, err = 0;
2882
2883                 for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
2884                         tgt = lmv->tgts[i];
2885
2886                         if (tgt == NULL || tgt->ltd_exp == NULL)
2887                                 continue;
2888
2889                         err = obd_set_info_async(env, tgt->ltd_exp,
2890                                                  keylen, key, vallen, val, set);
2891                         if (err && rc == 0)
2892                                 rc = err;
2893                 }
2894
2895                 RETURN(rc);
2896         }
2897
2898         RETURN(-EINVAL);
2899 }
2900
2901 static int lmv_unpack_md_v1(struct obd_export *exp, struct lmv_stripe_md *lsm,
2902                             const struct lmv_mds_md_v1 *lmm1)
2903 {
2904         struct lmv_obd  *lmv = &exp->exp_obd->u.lmv;
2905         int             stripe_count;
2906         int             cplen;
2907         int             i;
2908         int             rc = 0;
2909         ENTRY;
2910
2911         lsm->lsm_md_magic = le32_to_cpu(lmm1->lmv_magic);
2912         lsm->lsm_md_stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
2913         lsm->lsm_md_master_mdt_index = le32_to_cpu(lmm1->lmv_master_mdt_index);
2914         if (OBD_FAIL_CHECK(OBD_FAIL_UNKNOWN_LMV_STRIPE))
2915                 lsm->lsm_md_hash_type = LMV_HASH_TYPE_UNKNOWN;
2916         else
2917                 lsm->lsm_md_hash_type = le32_to_cpu(lmm1->lmv_hash_type);
2918         lsm->lsm_md_layout_version = le32_to_cpu(lmm1->lmv_layout_version);
2919         lsm->lsm_md_migrate_offset = le32_to_cpu(lmm1->lmv_migrate_offset);
2920         lsm->lsm_md_migrate_hash = le32_to_cpu(lmm1->lmv_migrate_hash);
2921         cplen = strlcpy(lsm->lsm_md_pool_name, lmm1->lmv_pool_name,
2922                         sizeof(lsm->lsm_md_pool_name));
2923
2924         if (cplen >= sizeof(lsm->lsm_md_pool_name))
2925                 RETURN(-E2BIG);
2926
2927         CDEBUG(D_INFO, "unpack lsm count %d, master %d hash_type %#x "
2928                "layout_version %d\n", lsm->lsm_md_stripe_count,
2929                lsm->lsm_md_master_mdt_index, lsm->lsm_md_hash_type,
2930                lsm->lsm_md_layout_version);
2931
2932         stripe_count = le32_to_cpu(lmm1->lmv_stripe_count);
2933         for (i = 0; i < stripe_count; i++) {
2934                 fid_le_to_cpu(&lsm->lsm_md_oinfo[i].lmo_fid,
2935                               &lmm1->lmv_stripe_fids[i]);
2936                 rc = lmv_fld_lookup(lmv, &lsm->lsm_md_oinfo[i].lmo_fid,
2937                                     &lsm->lsm_md_oinfo[i].lmo_mds);
2938                 if (rc != 0)
2939                         RETURN(rc);
2940                 CDEBUG(D_INFO, "unpack fid #%d "DFID"\n", i,
2941                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid));
2942         }
2943
2944         RETURN(rc);
2945 }
2946
2947 static int lmv_unpackmd(struct obd_export *exp, struct lmv_stripe_md **lsmp,
2948                         const union lmv_mds_md *lmm, size_t lmm_size)
2949 {
2950         struct lmv_stripe_md     *lsm;
2951         int                      lsm_size;
2952         int                      rc;
2953         bool                     allocated = false;
2954         ENTRY;
2955
2956         LASSERT(lsmp != NULL);
2957
2958         lsm = *lsmp;
2959         /* Free memmd */
2960         if (lsm != NULL && lmm == NULL) {
2961                 int i;
2962
2963                 for (i = 0; i < lsm->lsm_md_stripe_count; i++)
2964                         iput(lsm->lsm_md_oinfo[i].lmo_root);
2965                 lsm_size = lmv_stripe_md_size(lsm->lsm_md_stripe_count);
2966                 OBD_FREE(lsm, lsm_size);
2967                 *lsmp = NULL;
2968                 RETURN(0);
2969         }
2970
2971         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_STRIPE)
2972                 RETURN(-EPERM);
2973
2974         /* Unpack memmd */
2975         if (le32_to_cpu(lmm->lmv_magic) != LMV_MAGIC_V1 &&
2976             le32_to_cpu(lmm->lmv_magic) != LMV_USER_MAGIC) {
2977                 CERROR("%s: invalid lmv magic %x: rc = %d\n",
2978                        exp->exp_obd->obd_name, le32_to_cpu(lmm->lmv_magic),
2979                        -EIO);
2980                 RETURN(-EIO);
2981         }
2982
2983         if (le32_to_cpu(lmm->lmv_magic) == LMV_MAGIC_V1)
2984                 lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
2985         else
2986                 /**
2987                  * Unpack default dirstripe(lmv_user_md) to lmv_stripe_md,
2988                  * stripecount should be 0 then.
2989                  */
2990                 lsm_size = lmv_stripe_md_size(0);
2991
2992         lsm_size = lmv_stripe_md_size(lmv_mds_md_stripe_count_get(lmm));
2993         if (lsm == NULL) {
2994                 OBD_ALLOC(lsm, lsm_size);
2995                 if (lsm == NULL)
2996                         RETURN(-ENOMEM);
2997                 allocated = true;
2998                 *lsmp = lsm;
2999         }
3000
3001         switch (le32_to_cpu(lmm->lmv_magic)) {
3002         case LMV_MAGIC_V1:
3003                 rc = lmv_unpack_md_v1(exp, lsm, &lmm->lmv_md_v1);
3004                 break;
3005         default:
3006                 CERROR("%s: unrecognized magic %x\n", exp->exp_obd->obd_name,
3007                        le32_to_cpu(lmm->lmv_magic));
3008                 rc = -EINVAL;
3009                 break;
3010         }
3011
3012         if (rc != 0 && allocated) {
3013                 OBD_FREE(lsm, lsm_size);
3014                 *lsmp = NULL;
3015                 lsm_size = rc;
3016         }
3017         RETURN(lsm_size);
3018 }
3019
3020 void lmv_free_memmd(struct lmv_stripe_md *lsm)
3021 {
3022         lmv_unpackmd(NULL, &lsm, NULL, 0);
3023 }
3024 EXPORT_SYMBOL(lmv_free_memmd);
3025
3026 static int lmv_cancel_unused(struct obd_export *exp, const struct lu_fid *fid,
3027                              union ldlm_policy_data *policy,
3028                              enum ldlm_mode mode, enum ldlm_cancel_flags flags,
3029                              void *opaque)
3030 {
3031         struct lmv_obd *lmv = &exp->exp_obd->u.lmv;
3032         int rc = 0;
3033         __u32 i;
3034         ENTRY;
3035
3036         LASSERT(fid != NULL);
3037
3038         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3039                 struct lmv_tgt_desc *tgt = lmv->tgts[i];
3040                 int err;
3041
3042                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
3043                         continue;
3044
3045                 err = md_cancel_unused(tgt->ltd_exp, fid, policy, mode, flags,
3046                                        opaque);
3047                 if (!rc)
3048                         rc = err;
3049         }
3050         RETURN(rc);
3051 }
3052
3053 static int lmv_set_lock_data(struct obd_export *exp,
3054                              const struct lustre_handle *lockh,
3055                              void *data, __u64 *bits)
3056 {
3057         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
3058         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3059         int                      rc;
3060         ENTRY;
3061
3062         if (tgt == NULL || tgt->ltd_exp == NULL)
3063                 RETURN(-EINVAL);
3064         rc =  md_set_lock_data(tgt->ltd_exp, lockh, data, bits);
3065         RETURN(rc);
3066 }
3067
3068 enum ldlm_mode lmv_lock_match(struct obd_export *exp, __u64 flags,
3069                               const struct lu_fid *fid, enum ldlm_type type,
3070                               union ldlm_policy_data *policy,
3071                               enum ldlm_mode mode, struct lustre_handle *lockh)
3072 {
3073         struct obd_device       *obd = exp->exp_obd;
3074         struct lmv_obd          *lmv = &obd->u.lmv;
3075         enum ldlm_mode          rc;
3076         int                     tgt;
3077         int                     i;
3078         ENTRY;
3079
3080         CDEBUG(D_INODE, "Lock match for "DFID"\n", PFID(fid));
3081
3082         /*
3083          * With DNE every object can have two locks in different namespaces:
3084          * lookup lock in space of MDT storing direntry and update/open lock in
3085          * space of MDT storing inode.  Try the MDT that the FID maps to first,
3086          * since this can be easily found, and only try others if that fails.
3087          */
3088         for (i = 0, tgt = lmv_find_target_index(lmv, fid);
3089              i < lmv->desc.ld_tgt_count;
3090              i++, tgt = (tgt + 1) % lmv->desc.ld_tgt_count) {
3091                 if (tgt < 0) {
3092                         CDEBUG(D_HA, "%s: "DFID" is inaccessible: rc = %d\n",
3093                                obd->obd_name, PFID(fid), tgt);
3094                         tgt = 0;
3095                 }
3096
3097                 if (lmv->tgts[tgt] == NULL ||
3098                     lmv->tgts[tgt]->ltd_exp == NULL ||
3099                     lmv->tgts[tgt]->ltd_active == 0)
3100                         continue;
3101
3102                 rc = md_lock_match(lmv->tgts[tgt]->ltd_exp, flags, fid,
3103                                    type, policy, mode, lockh);
3104                 if (rc)
3105                         RETURN(rc);
3106         }
3107
3108         RETURN(0);
3109 }
3110
3111 int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req,
3112                       struct obd_export *dt_exp, struct obd_export *md_exp,
3113                       struct lustre_md *md)
3114 {
3115         struct lmv_obd          *lmv = &exp->exp_obd->u.lmv;
3116         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3117
3118         if (tgt == NULL || tgt->ltd_exp == NULL)
3119                 RETURN(-EINVAL);
3120
3121         return md_get_lustre_md(lmv->tgts[0]->ltd_exp, req, dt_exp, md_exp, md);
3122 }
3123
3124 int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md)
3125 {
3126         struct obd_device       *obd = exp->exp_obd;
3127         struct lmv_obd          *lmv = &obd->u.lmv;
3128         struct lmv_tgt_desc     *tgt = lmv->tgts[0];
3129         ENTRY;
3130
3131         if (md->lmv != NULL) {
3132                 lmv_free_memmd(md->lmv);
3133                 md->lmv = NULL;
3134         }
3135         if (tgt == NULL || tgt->ltd_exp == NULL)
3136                 RETURN(-EINVAL);
3137         RETURN(md_free_lustre_md(lmv->tgts[0]->ltd_exp, md));
3138 }
3139
3140 int lmv_set_open_replay_data(struct obd_export *exp,
3141                              struct obd_client_handle *och,
3142                              struct lookup_intent *it)
3143 {
3144         struct obd_device       *obd = exp->exp_obd;
3145         struct lmv_obd          *lmv = &obd->u.lmv;
3146         struct lmv_tgt_desc     *tgt;
3147         ENTRY;
3148
3149         tgt = lmv_find_target(lmv, &och->och_fid);
3150         if (IS_ERR(tgt))
3151                 RETURN(PTR_ERR(tgt));
3152
3153         RETURN(md_set_open_replay_data(tgt->ltd_exp, och, it));
3154 }
3155
3156 int lmv_clear_open_replay_data(struct obd_export *exp,
3157                                struct obd_client_handle *och)
3158 {
3159         struct obd_device       *obd = exp->exp_obd;
3160         struct lmv_obd          *lmv = &obd->u.lmv;
3161         struct lmv_tgt_desc     *tgt;
3162         ENTRY;
3163
3164         tgt = lmv_find_target(lmv, &och->och_fid);
3165         if (IS_ERR(tgt))
3166                 RETURN(PTR_ERR(tgt));
3167
3168         RETURN(md_clear_open_replay_data(tgt->ltd_exp, och));
3169 }
3170
3171 int lmv_intent_getattr_async(struct obd_export *exp,
3172                              struct md_enqueue_info *minfo)
3173 {
3174         struct md_op_data *op_data = &minfo->mi_data;
3175         struct obd_device *obd = exp->exp_obd;
3176         struct lmv_obd *lmv = &obd->u.lmv;
3177         struct lmv_tgt_desc *tgt = NULL;
3178         int rc;
3179         ENTRY;
3180
3181         if (!fid_is_sane(&op_data->op_fid2))
3182                 RETURN(-EINVAL);
3183
3184         tgt = lmv_locate_mds(lmv, op_data, &op_data->op_fid1);
3185         if (IS_ERR(tgt))
3186                 RETURN(PTR_ERR(tgt));
3187
3188         /*
3189          * no special handle for remote dir, which needs to fetch both LOOKUP
3190          * lock on parent, and then UPDATE lock on child MDT, which makes all
3191          * complicated because this is done async. So only LOOKUP lock is
3192          * fetched for remote dir, but considering remote dir is rare case,
3193          * and not supporting it in statahead won't cause any issue, just leave
3194          * it as is.
3195          */
3196
3197         rc = md_intent_getattr_async(tgt->ltd_exp, minfo);
3198         RETURN(rc);
3199 }
3200
3201 int lmv_revalidate_lock(struct obd_export *exp, struct lookup_intent *it,
3202                         struct lu_fid *fid, __u64 *bits)
3203 {
3204         struct obd_device       *obd = exp->exp_obd;
3205         struct lmv_obd          *lmv = &obd->u.lmv;
3206         struct lmv_tgt_desc     *tgt;
3207         int                      rc;
3208         ENTRY;
3209
3210         tgt = lmv_find_target(lmv, fid);
3211         if (IS_ERR(tgt))
3212                 RETURN(PTR_ERR(tgt));
3213
3214         rc = md_revalidate_lock(tgt->ltd_exp, it, fid, bits);
3215         RETURN(rc);
3216 }
3217
3218 int lmv_get_fid_from_lsm(struct obd_export *exp,
3219                          const struct lmv_stripe_md *lsm,
3220                          const char *name, int namelen, struct lu_fid *fid)
3221 {
3222         const struct lmv_oinfo *oinfo;
3223
3224         LASSERT(lsm != NULL);
3225         oinfo = lsm_name_to_stripe_info(lsm, name, namelen);
3226         if (IS_ERR(oinfo))
3227                 return PTR_ERR(oinfo);
3228
3229         *fid = oinfo->lmo_fid;
3230
3231         RETURN(0);
3232 }
3233
3234 /**
3235  * For lmv, only need to send request to master MDT, and the master MDT will
3236  * process with other slave MDTs. The only exception is Q_GETOQUOTA for which
3237  * we directly fetch data from the slave MDTs.
3238  */
3239 int lmv_quotactl(struct obd_device *unused, struct obd_export *exp,
3240                  struct obd_quotactl *oqctl)
3241 {
3242         struct obd_device   *obd = class_exp2obd(exp);
3243         struct lmv_obd      *lmv = &obd->u.lmv;
3244         struct lmv_tgt_desc *tgt = lmv->tgts[0];
3245         int                  rc = 0;
3246         __u32                i;
3247         __u64                curspace, curinodes;
3248         ENTRY;
3249
3250         if (tgt == NULL ||
3251             tgt->ltd_exp == NULL ||
3252             !tgt->ltd_active ||
3253             lmv->desc.ld_tgt_count == 0) {
3254                 CERROR("master lmv inactive\n");
3255                 RETURN(-EIO);
3256         }
3257
3258         if (oqctl->qc_cmd != Q_GETOQUOTA) {
3259                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
3260                 RETURN(rc);
3261         }
3262
3263         curspace = curinodes = 0;
3264         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
3265                 int err;
3266                 tgt = lmv->tgts[i];
3267
3268                 if (tgt == NULL || tgt->ltd_exp == NULL || !tgt->ltd_active)
3269                         continue;
3270
3271                 err = obd_quotactl(tgt->ltd_exp, oqctl);
3272                 if (err) {
3273                         CERROR("getquota on mdt %d failed. %d\n", i, err);
3274                         if (!rc)
3275                                 rc = err;
3276                 } else {
3277                         curspace += oqctl->qc_dqblk.dqb_curspace;
3278                         curinodes += oqctl->qc_dqblk.dqb_curinodes;
3279                 }
3280         }
3281         oqctl->qc_dqblk.dqb_curspace = curspace;
3282         oqctl->qc_dqblk.dqb_curinodes = curinodes;
3283
3284         RETURN(rc);
3285 }
3286
3287 static int lmv_merge_attr(struct obd_export *exp,
3288                           const struct lmv_stripe_md *lsm,
3289                           struct cl_attr *attr,
3290                           ldlm_blocking_callback cb_blocking)
3291 {
3292         int rc;
3293         int i;
3294
3295         rc = lmv_revalidate_slaves(exp, lsm, cb_blocking, 0);
3296         if (rc < 0)
3297                 return rc;
3298
3299         for (i = 0; i < lsm->lsm_md_stripe_count; i++) {
3300                 struct inode *inode = lsm->lsm_md_oinfo[i].lmo_root;
3301
3302                 CDEBUG(D_INFO, ""DFID" size %llu, blocks %llu nlink %u,"
3303                        " atime %lu ctime %lu, mtime %lu.\n",
3304                        PFID(&lsm->lsm_md_oinfo[i].lmo_fid),
3305                        i_size_read(inode), (unsigned long long)inode->i_blocks,
3306                        inode->i_nlink, LTIME_S(inode->i_atime),
3307                        LTIME_S(inode->i_ctime), LTIME_S(inode->i_mtime));
3308
3309                 /* for slave stripe, it needs to subtract nlink for . and .. */
3310                 if (i != 0)
3311                         attr->cat_nlink += inode->i_nlink - 2;
3312                 else
3313                         attr->cat_nlink = inode->i_nlink;
3314
3315                 attr->cat_size += i_size_read(inode);
3316                 attr->cat_blocks += inode->i_blocks;
3317
3318                 if (attr->cat_atime < LTIME_S(inode->i_atime))
3319                         attr->cat_atime = LTIME_S(inode->i_atime);
3320
3321                 if (attr->cat_ctime < LTIME_S(inode->i_ctime))
3322                         attr->cat_ctime = LTIME_S(inode->i_ctime);
3323
3324                 if (attr->cat_mtime < LTIME_S(inode->i_mtime))
3325                         attr->cat_mtime = LTIME_S(inode->i_mtime);
3326         }
3327         return 0;
3328 }
3329
3330 struct obd_ops lmv_obd_ops = {
3331         .o_owner                = THIS_MODULE,
3332         .o_setup                = lmv_setup,
3333         .o_cleanup              = lmv_cleanup,
3334         .o_precleanup           = lmv_precleanup,
3335         .o_process_config       = lmv_process_config,
3336         .o_connect              = lmv_connect,
3337         .o_disconnect           = lmv_disconnect,
3338         .o_statfs               = lmv_statfs,
3339         .o_get_info             = lmv_get_info,
3340         .o_set_info_async       = lmv_set_info_async,
3341         .o_notify               = lmv_notify,
3342         .o_get_uuid             = lmv_get_uuid,
3343         .o_iocontrol            = lmv_iocontrol,
3344         .o_quotactl             = lmv_quotactl
3345 };
3346
3347 struct md_ops lmv_md_ops = {
3348         .m_get_root             = lmv_get_root,
3349         .m_null_inode           = lmv_null_inode,
3350         .m_close                = lmv_close,
3351         .m_create               = lmv_create,
3352         .m_enqueue              = lmv_enqueue,
3353         .m_getattr              = lmv_getattr,
3354         .m_getxattr             = lmv_getxattr,
3355         .m_getattr_name         = lmv_getattr_name,
3356         .m_intent_lock          = lmv_intent_lock,
3357         .m_link                 = lmv_link,
3358         .m_rename               = lmv_rename,
3359         .m_setattr              = lmv_setattr,
3360         .m_setxattr             = lmv_setxattr,
3361         .m_fsync                = lmv_fsync,
3362         .m_file_resync          = lmv_file_resync,
3363         .m_read_page            = lmv_read_page,
3364         .m_unlink               = lmv_unlink,
3365         .m_init_ea_size         = lmv_init_ea_size,
3366         .m_cancel_unused        = lmv_cancel_unused,
3367         .m_set_lock_data        = lmv_set_lock_data,
3368         .m_lock_match           = lmv_lock_match,
3369         .m_get_lustre_md        = lmv_get_lustre_md,
3370         .m_free_lustre_md       = lmv_free_lustre_md,
3371         .m_merge_attr           = lmv_merge_attr,
3372         .m_set_open_replay_data = lmv_set_open_replay_data,
3373         .m_clear_open_replay_data = lmv_clear_open_replay_data,
3374         .m_intent_getattr_async = lmv_intent_getattr_async,
3375         .m_revalidate_lock      = lmv_revalidate_lock,
3376         .m_get_fid_from_lsm     = lmv_get_fid_from_lsm,
3377         .m_unpackmd             = lmv_unpackmd,
3378 };
3379
3380 static int __init lmv_init(void)
3381 {
3382         return class_register_type(&lmv_obd_ops, &lmv_md_ops, true, NULL,
3383                                    LUSTRE_LMV_NAME, NULL);
3384 }
3385
3386 static void __exit lmv_exit(void)
3387 {
3388         class_unregister_type(LUSTRE_LMV_NAME);
3389 }
3390
3391 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
3392 MODULE_DESCRIPTION("Lustre Logical Metadata Volume");
3393 MODULE_VERSION(LUSTRE_VERSION_STRING);
3394 MODULE_LICENSE("GPL");
3395
3396 module_init(lmv_init);
3397 module_exit(lmv_exit);