lustre/lov/lov_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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/lov/lov_obd.c
  37  *
  38  * Author: Phil Schwan <phil@clusterfs.com>
  39  * Author: Peter Braam <braam@clusterfs.com>
  40  * Author: Mike Shaver <shaver@clusterfs.com>
  41  * Author: Nathan Rutman <nathan@clusterfs.com>
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_LOV
  45 #ifdef __KERNEL__
  46 #include <libcfs/libcfs.h>
  47 #else
  48 #include <liblustre.h>
  49 #endif
  50
  51 #include <obd_support.h>
  52 #include <lustre_ioctl.h>
  53 #include <lustre_lib.h>
  54 #include <lustre_net.h>
  55 #include <lustre/lustre_idl.h>
  56 #include <lustre_dlm.h>
  57 #include <lustre_mds.h>
  58 #include <obd_class.h>
  59 #include <lprocfs_status.h>
  60 #include <lustre_param.h>
  61 #include <cl_object.h>
  62 #include <lclient.h>
  63 #include <lustre/ll_fiemap.h>
  64 #include <lustre_fid.h>
  65
  66 #include "lov_internal.h"
  67
  68 /* Keep a refcount of lov->tgt usage to prevent racing with addition/deletion.
  69    Any function that expects lov_tgts to remain stationary must take a ref. */
  70 static void lov_getref(struct obd_device *obd)
  71 {
  72         struct lov_obd *lov = &obd->u.lov;
  73
  74         /* nobody gets through here until lov_putref is done */
  75         mutex_lock(&lov->lov_lock);
  76         atomic_inc(&lov->lov_refcount);
  77         mutex_unlock(&lov->lov_lock);
  78         return;
  79 }
  80
  81 static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt);
  82
  83 static void lov_putref(struct obd_device *obd)
  84 {
  85         struct lov_obd *lov = &obd->u.lov;
  86
  87         mutex_lock(&lov->lov_lock);
  88         /* ok to dec to 0 more than once -- ltd_exp's will be null */
  89         if (atomic_dec_and_test(&lov->lov_refcount) && lov->lov_death_row) {
  90                 struct list_head kill = LIST_HEAD_INIT(kill);
  91                 struct lov_tgt_desc *tgt, *n;
  92                 int i;
  93
  94                 CDEBUG(D_CONFIG, "destroying %d lov targets\n",
  95                        lov->lov_death_row);
  96                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
  97                         tgt = lov->lov_tgts[i];
  98
  99                         if (!tgt || !tgt->ltd_reap)
 100                                 continue;
 101                         list_add(&tgt->ltd_kill, &kill);
 102                         /* XXX - right now there is a dependency on ld_tgt_count
 103                          * being the maximum tgt index for computing the
 104                          * mds_max_easize. So we can't shrink it. */
 105                         lov_ost_pool_remove(&lov->lov_packed, i);
 106                         lov->lov_tgts[i] = NULL;
 107                         lov->lov_death_row--;
 108                 }
 109                 mutex_unlock(&lov->lov_lock);
 110
 111                 list_for_each_entry_safe(tgt, n, &kill, ltd_kill) {
 112                         list_del(&tgt->ltd_kill);
 113                         /* Disconnect */
 114                         __lov_del_obd(obd, tgt);
 115                 }
 116         } else {
 117                 mutex_unlock(&lov->lov_lock);
 118         }
 119 }
 120
 121 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
 122                               enum obd_notify_event ev);
 123 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
 124                       enum obd_notify_event ev, void *data);
 125
 126 int lov_connect_obd(struct obd_device *obd, __u32 index, int activate,
 127                     struct obd_connect_data *data)
 128 {
 129         struct lov_obd *lov = &obd->u.lov;
 130         struct obd_uuid *tgt_uuid;
 131         struct obd_device *tgt_obd;
 132         static struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
 133         struct obd_import *imp;
 134         int rc;
 135         ENTRY;
 136
 137         if (lov->lov_tgts[index] == NULL)
 138                 RETURN(-EINVAL);
 139
 140         tgt_uuid = &lov->lov_tgts[index]->ltd_uuid;
 141         tgt_obd = lov->lov_tgts[index]->ltd_obd;
 142
 143         if (!tgt_obd->obd_set_up) {
 144                 CERROR("Target %s not set up\n", obd_uuid2str(tgt_uuid));
 145                 RETURN(-EINVAL);
 146         }
 147
 148         /* override the sp_me from lov */
 149         tgt_obd->u.cli.cl_sp_me = lov->lov_sp_me;
 150
 151         if (data && (data->ocd_connect_flags & OBD_CONNECT_INDEX))
 152                 data->ocd_index = index;
 153
 154         /*
 155          * Divine LOV knows that OBDs under it are OSCs.
 156          */
 157         imp = tgt_obd->u.cli.cl_import;
 158
 159         if (activate) {
 160                 tgt_obd->obd_no_recov = 0;
 161                 /* FIXME this is probably supposed to be
 162                    ptlrpc_set_import_active.  Horrible naming. */
 163                 ptlrpc_activate_import(imp);
 164         }
 165
 166         rc = obd_register_observer(tgt_obd, obd);
 167         if (rc) {
 168                 CERROR("Target %s register_observer error %d\n",
 169                        obd_uuid2str(tgt_uuid), rc);
 170                 RETURN(rc);
 171         }
 172
 173
 174         if (imp->imp_invalid) {
 175                 CDEBUG(D_CONFIG, "not connecting OSC %s; administratively "
 176                        "disabled\n", obd_uuid2str(tgt_uuid));
 177                 RETURN(0);
 178         }
 179
 180         rc = obd_connect(NULL, &lov->lov_tgts[index]->ltd_exp, tgt_obd,
 181                          &lov_osc_uuid, data, NULL);
 182         if (rc || !lov->lov_tgts[index]->ltd_exp) {
 183                 CERROR("Target %s connect error %d\n",
 184                        obd_uuid2str(tgt_uuid), rc);
 185                 RETURN(-ENODEV);
 186         }
 187
 188         lov->lov_tgts[index]->ltd_reap = 0;
 189
 190         CDEBUG(D_CONFIG, "Connected tgt idx %d %s (%s) %sactive\n", index,
 191                obd_uuid2str(tgt_uuid), tgt_obd->obd_name, activate ? "":"in");
 192
 193         if (lov->targets_proc_entry != NULL) {
 194                 struct proc_dir_entry *osc_symlink;
 195                 struct obd_device *osc_obd;
 196
 197                 osc_obd = lov->lov_tgts[index]->ltd_exp->exp_obd;
 198
 199                 LASSERT(osc_obd != NULL);
 200                 LASSERT(osc_obd->obd_magic == OBD_DEVICE_MAGIC);
 201                 LASSERT(osc_obd->obd_type->typ_name != NULL);
 202
 203                 osc_symlink = lprocfs_add_symlink(osc_obd->obd_name,
 204                                                   lov->targets_proc_entry,
 205                                                   "../../../%s/%s",
 206                                                   osc_obd->obd_type->typ_name,
 207                                                   osc_obd->obd_name);
 208                 if (osc_symlink == NULL) {
 209                         CERROR("could not register LOV target "
 210                                "/proc/fs/lustre/%s/%s/target_obds/%s.",
 211                                obd->obd_type->typ_name, obd->obd_name,
 212                                osc_obd->obd_name);
 213                 }
 214         }
 215         RETURN(0);
 216 }
 217
 218 static int lov_connect(const struct lu_env *env,
 219                        struct obd_export **exp, struct obd_device *obd,
 220                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 221                        void *localdata)
 222 {
 223         struct lov_obd *lov = &obd->u.lov;
 224         struct lov_tgt_desc *tgt;
 225         struct lustre_handle conn;
 226         int i, rc;
 227         ENTRY;
 228
 229         CDEBUG(D_CONFIG, "connect #%d\n", lov->lov_connects);
 230
 231         rc = class_connect(&conn, obd, cluuid);
 232         if (rc)
 233                 RETURN(rc);
 234
 235         *exp = class_conn2export(&conn);
 236
 237         /* Why should there ever be more than 1 connect? */
 238         lov->lov_connects++;
 239         LASSERT(lov->lov_connects == 1);
 240
 241         memset(&lov->lov_ocd, 0, sizeof(lov->lov_ocd));
 242         if (data)
 243                 lov->lov_ocd = *data;
 244
 245         lov->targets_proc_entry = lprocfs_seq_register("target_obds",
 246                                                        obd->obd_proc_entry,
 247                                                        NULL, NULL);
 248         if (IS_ERR(lov->targets_proc_entry)) {
 249                 CERROR("%s: could not register /proc/fs/lustre/%s/%s/target_obds.",
 250                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
 251                 lov->targets_proc_entry = NULL;
 252         }
 253
 254         obd_getref(obd);
 255         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 256                 tgt = lov->lov_tgts[i];
 257                 if (!tgt || obd_uuid_empty(&tgt->ltd_uuid))
 258                         continue;
 259                 /* Flags will be lowest common denominator */
 260                 rc = lov_connect_obd(obd, i, tgt->ltd_activate, &lov->lov_ocd);
 261                 if (rc) {
 262                         CERROR("%s: lov connect tgt %d failed: %d\n",
 263                                obd->obd_name, i, rc);
 264                         continue;
 265                 }
 266                 /* connect to administrative disabled ost */
 267                 if (!lov->lov_tgts[i]->ltd_exp)
 268                         continue;
 269
 270                 rc = lov_notify(obd, lov->lov_tgts[i]->ltd_exp->exp_obd,
 271                                 OBD_NOTIFY_CONNECT, (void *)&i);
 272                 if (rc) {
 273                         CERROR("%s error sending notify %d\n",
 274                                obd->obd_name, rc);
 275                 }
 276         }
 277         obd_putref(obd);
 278
 279         RETURN(0);
 280 }
 281
 282 static int lov_disconnect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
 283 {
 284         struct lov_obd *lov = &obd->u.lov;
 285         struct obd_device *osc_obd;
 286         int rc;
 287         ENTRY;
 288
 289         osc_obd = class_exp2obd(tgt->ltd_exp);
 290         CDEBUG(D_CONFIG, "%s: disconnecting target %s\n",
 291                obd->obd_name, osc_obd->obd_name);
 292
 293         if (tgt->ltd_active) {
 294                 tgt->ltd_active = 0;
 295                 lov->desc.ld_active_tgt_count--;
 296                 tgt->ltd_exp->exp_obd->obd_inactive = 1;
 297         }
 298
 299         if (osc_obd) {
 300                 /* Pass it on to our clients.
 301                  * XXX This should be an argument to disconnect,
 302                  * XXX not a back-door flag on the OBD.  Ah well.
 303                  */
 304                 osc_obd->obd_force = obd->obd_force;
 305                 osc_obd->obd_fail = obd->obd_fail;
 306                 osc_obd->obd_no_recov = obd->obd_no_recov;
 307
 308                 if (lov->targets_proc_entry != NULL)
 309                         lprocfs_remove_proc_entry(osc_obd->obd_name,
 310                                                   lov->targets_proc_entry);
 311         }
 312
 313         obd_register_observer(osc_obd, NULL);
 314
 315         rc = obd_disconnect(tgt->ltd_exp);
 316         if (rc) {
 317                 CERROR("Target %s disconnect error %d\n",
 318                        tgt->ltd_uuid.uuid, rc);
 319                 rc = 0;
 320         }
 321
 322         tgt->ltd_exp = NULL;
 323         RETURN(0);
 324 }
 325
 326 static int lov_disconnect(struct obd_export *exp)
 327 {
 328         struct obd_device *obd = class_exp2obd(exp);
 329         struct lov_obd *lov = &obd->u.lov;
 330         int i, rc;
 331         ENTRY;
 332
 333         if (!lov->lov_tgts)
 334                 goto out;
 335
 336         /* Only disconnect the underlying layers on the final disconnect. */
 337         lov->lov_connects--;
 338         if (lov->lov_connects != 0) {
 339                 /* why should there be more than 1 connect? */
 340                 CERROR("disconnect #%d\n", lov->lov_connects);
 341                 goto out;
 342         }
 343
 344         /* Let's hold another reference so lov_del_obd doesn't spin through
 345            putref every time */
 346         obd_getref(obd);
 347
 348         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 349                 if (lov->lov_tgts[i] && lov->lov_tgts[i]->ltd_exp) {
 350                         /* Disconnection is the last we know about an obd */
 351                         lov_del_target(obd, i, 0, lov->lov_tgts[i]->ltd_gen);
 352                 }
 353         }
 354         obd_putref(obd);
 355
 356         if (lov->targets_proc_entry != NULL)
 357                 lprocfs_remove(&lov->targets_proc_entry);
 358
 359 out:
 360         rc = class_disconnect(exp); /* bz 9811 */
 361         RETURN(rc);
 362 }
 363
 364 /* Error codes:
 365  *
 366  *  -EINVAL  : UUID can't be found in the LOV's target list
 367  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
 368  *  -EBADF   : The UUID is found, but the OBD is the wrong type (!)
 369  *  any >= 0 : is log target index
 370  */
 371 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
 372                               enum obd_notify_event ev)
 373 {
 374         struct lov_obd *lov = &obd->u.lov;
 375         struct lov_tgt_desc *tgt;
 376         int index, activate, active;
 377         ENTRY;
 378
 379         CDEBUG(D_INFO, "Searching in lov %p for uuid %s event(%d)\n",
 380                lov, uuid->uuid, ev);
 381
 382         obd_getref(obd);
 383         for (index = 0; index < lov->desc.ld_tgt_count; index++) {
 384                 tgt = lov->lov_tgts[index];
 385                 if (!tgt)
 386                         continue;
 387                 /*
 388                  * LU-642, initially inactive OSC could miss the obd_connect,
 389                  * we make up for it here.
 390                  */
 391                 if (ev == OBD_NOTIFY_ACTIVATE && tgt->ltd_exp == NULL &&
 392                     obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
 393                         struct obd_uuid lov_osc_uuid = {"LOV_OSC_UUID"};
 394
 395                         obd_connect(NULL, &tgt->ltd_exp, tgt->ltd_obd,
 396                                     &lov_osc_uuid, &lov->lov_ocd, NULL);
 397                 }
 398                 if (!tgt->ltd_exp)
 399                         continue;
 400
 401                 CDEBUG(D_INFO, "lov idx %d is %s conn "LPX64"\n",
 402                        index, obd_uuid2str(&tgt->ltd_uuid),
 403                        tgt->ltd_exp->exp_handle.h_cookie);
 404                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 405                         break;
 406         }
 407
 408         if (index == lov->desc.ld_tgt_count)
 409                 GOTO(out, index = -EINVAL);
 410
 411         if (ev == OBD_NOTIFY_DEACTIVATE || ev == OBD_NOTIFY_ACTIVATE) {
 412                 activate = (ev == OBD_NOTIFY_ACTIVATE) ? 1 : 0;
 413
 414                 if (lov->lov_tgts[index]->ltd_activate == activate) {
 415                         CDEBUG(D_INFO, "OSC %s already %sactivate!\n",
 416                                uuid->uuid, activate ? "" : "de");
 417                 } else {
 418                         lov->lov_tgts[index]->ltd_activate = activate;
 419                         CDEBUG(D_CONFIG, "%sactivate OSC %s\n",
 420                                activate ? "" : "de", obd_uuid2str(uuid));
 421                 }
 422
 423         } else if (ev == OBD_NOTIFY_INACTIVE || ev == OBD_NOTIFY_ACTIVE) {
 424                 active = (ev == OBD_NOTIFY_ACTIVE) ? 1 : 0;
 425
 426                 if (lov->lov_tgts[index]->ltd_active == active) {
 427                         CDEBUG(D_INFO, "OSC %s already %sactive!\n",
 428                                uuid->uuid, active ? "" : "in");
 429                         GOTO(out, index);
 430                 } else {
 431                         CDEBUG(D_CONFIG, "Marking OSC %s %sactive\n",
 432                                obd_uuid2str(uuid), active ? "" : "in");
 433                 }
 434
 435                 lov->lov_tgts[index]->ltd_active = active;
 436                 if (active) {
 437                         lov->desc.ld_active_tgt_count++;
 438                         lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 0;
 439                 } else {
 440                         lov->desc.ld_active_tgt_count--;
 441                         lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 1;
 442                 }
 443         } else {
 444                 CERROR("Unknown event(%d) for uuid %s", ev, uuid->uuid);
 445         }
 446
 447  out:
 448         obd_putref(obd);
 449         RETURN(index);
 450 }
 451
 452 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
 453                       enum obd_notify_event ev, void *data)
 454 {
 455         int rc = 0;
 456         struct lov_obd *lov = &obd->u.lov;
 457         ENTRY;
 458
 459         down_read(&lov->lov_notify_lock);
 460         if (!lov->lov_connects) {
 461                 up_read(&lov->lov_notify_lock);
 462                 RETURN(rc);
 463         }
 464
 465         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE ||
 466             ev == OBD_NOTIFY_ACTIVATE || ev == OBD_NOTIFY_DEACTIVATE) {
 467                 struct obd_uuid *uuid;
 468
 469                 LASSERT(watched);
 470
 471                 if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) {
 472                         up_read(&lov->lov_notify_lock);
 473                         CERROR("unexpected notification of %s %s!\n",
 474                                watched->obd_type->typ_name,
 475                                watched->obd_name);
 476                         RETURN(-EINVAL);
 477                 }
 478                 uuid = &watched->u.cli.cl_target_uuid;
 479
 480                 /* Set OSC as active before notifying the observer, so the
 481                  * observer can use the OSC normally.
 482                  */
 483                 rc = lov_set_osc_active(obd, uuid, ev);
 484                 if (rc < 0) {
 485                         up_read(&lov->lov_notify_lock);
 486                         CERROR("event(%d) of %s failed: %d\n", ev,
 487                                obd_uuid2str(uuid), rc);
 488                         RETURN(rc);
 489                 }
 490                 /* active event should be pass lov target index as data */
 491                 data = &rc;
 492         }
 493
 494         /* Pass the notification up the chain. */
 495         if (watched) {
 496                 rc = obd_notify_observer(obd, watched, ev, data);
 497         } else {
 498                 /* NULL watched means all osc's in the lov (only for syncs) */
 499                 /* sync event should be send lov idx as data */
 500                 struct lov_obd *lov = &obd->u.lov;
 501                 int i, is_sync;
 502
 503                 data = &i;
 504                 is_sync = (ev == OBD_NOTIFY_SYNC) ||
 505                           (ev == OBD_NOTIFY_SYNC_NONBLOCK);
 506
 507                 obd_getref(obd);
 508                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 509                         if (!lov->lov_tgts[i])
 510                                 continue;
 511
 512                         /* don't send sync event if target not
 513                          * connected/activated */
 514                         if (is_sync &&  !lov->lov_tgts[i]->ltd_active)
 515                                 continue;
 516
 517                         rc = obd_notify_observer(obd, lov->lov_tgts[i]->ltd_obd,
 518                                                  ev, data);
 519                         if (rc) {
 520                                 CERROR("%s: notify %s of %s failed %d\n",
 521                                        obd->obd_name,
 522                                        obd->obd_observer->obd_name,
 523                                        lov->lov_tgts[i]->ltd_obd->obd_name,
 524                                        rc);
 525                         }
 526                 }
 527                 obd_putref(obd);
 528         }
 529
 530         up_read(&lov->lov_notify_lock);
 531         RETURN(rc);
 532 }
 533
 534 static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 535                           __u32 index, int gen, int active)
 536 {
 537         struct lov_obd *lov = &obd->u.lov;
 538         struct lov_tgt_desc *tgt;
 539         struct obd_device *tgt_obd;
 540         int rc;
 541         ENTRY;
 542
 543         CDEBUG(D_CONFIG, "uuid:%s idx:%d gen:%d active:%d\n",
 544                uuidp->uuid, index, gen, active);
 545
 546         if (gen <= 0) {
 547                 CERROR("request to add OBD %s with invalid generation: %d\n",
 548                        uuidp->uuid, gen);
 549                 RETURN(-EINVAL);
 550         }
 551
 552         tgt_obd = class_find_client_obd(uuidp, LUSTRE_OSC_NAME,
 553                                         &obd->obd_uuid);
 554         if (tgt_obd == NULL)
 555                 RETURN(-EINVAL);
 556
 557         mutex_lock(&lov->lov_lock);
 558
 559         if ((index < lov->lov_tgt_size) && (lov->lov_tgts[index] != NULL)) {
 560                 tgt = lov->lov_tgts[index];
 561                 CERROR("UUID %s already assigned at LOV target index %d\n",
 562                        obd_uuid2str(&tgt->ltd_uuid), index);
 563                 mutex_unlock(&lov->lov_lock);
 564                 RETURN(-EEXIST);
 565         }
 566
 567         if (index >= lov->lov_tgt_size) {
 568                 /* We need to reallocate the lov target array. */
 569                 struct lov_tgt_desc **newtgts, **old = NULL;
 570                 __u32 newsize, oldsize = 0;
 571
 572                 newsize = max(lov->lov_tgt_size, (__u32)2);
 573                 while (newsize < index + 1)
 574                         newsize = newsize << 1;
 575                 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
 576                 if (newtgts == NULL) {
 577                         mutex_unlock(&lov->lov_lock);
 578                         RETURN(-ENOMEM);
 579                 }
 580
 581                 if (lov->lov_tgt_size) {
 582                         memcpy(newtgts, lov->lov_tgts, sizeof(*newtgts) *
 583                                lov->lov_tgt_size);
 584                         old = lov->lov_tgts;
 585                         oldsize = lov->lov_tgt_size;
 586                 }
 587
 588                 lov->lov_tgts = newtgts;
 589                 lov->lov_tgt_size = newsize;
 590                 smp_rmb();
 591                 if (old)
 592                         OBD_FREE(old, sizeof(*old) * oldsize);
 593
 594                 CDEBUG(D_CONFIG, "tgts: %p size: %d\n",
 595                        lov->lov_tgts, lov->lov_tgt_size);
 596         }
 597
 598         OBD_ALLOC_PTR(tgt);
 599         if (!tgt) {
 600                 mutex_unlock(&lov->lov_lock);
 601                 RETURN(-ENOMEM);
 602         }
 603
 604         rc = lov_ost_pool_add(&lov->lov_packed, index, lov->lov_tgt_size);
 605         if (rc) {
 606                 mutex_unlock(&lov->lov_lock);
 607                 OBD_FREE_PTR(tgt);
 608                 RETURN(rc);
 609         }
 610
 611         tgt->ltd_uuid = *uuidp;
 612         tgt->ltd_obd = tgt_obd;
 613         /* XXX - add a sanity check on the generation number. */
 614         tgt->ltd_gen = gen;
 615         tgt->ltd_index = index;
 616         tgt->ltd_activate = active;
 617         lov->lov_tgts[index] = tgt;
 618         if (index >= lov->desc.ld_tgt_count)
 619                 lov->desc.ld_tgt_count = index + 1;
 620
 621         mutex_unlock(&lov->lov_lock);
 622
 623         CDEBUG(D_CONFIG, "idx=%d ltd_gen=%d ld_tgt_count=%d\n",
 624                 index, tgt->ltd_gen, lov->desc.ld_tgt_count);
 625
 626         rc = obd_notify(obd, tgt_obd, OBD_NOTIFY_CREATE, &index);
 627
 628         if (lov->lov_connects == 0) {
 629                 /* lov_connect hasn't been called yet. We'll do the
 630                    lov_connect_obd on this target when that fn first runs,
 631                    because we don't know the connect flags yet. */
 632                 RETURN(0);
 633         }
 634
 635         obd_getref(obd);
 636
 637         rc = lov_connect_obd(obd, index, active, &lov->lov_ocd);
 638         if (rc)
 639                 GOTO(out, rc);
 640
 641         /* connect to administrative disabled ost */
 642         if (!tgt->ltd_exp)
 643                 GOTO(out, rc = 0);
 644
 645         if (lov->lov_cache != NULL) {
 646                 rc = obd_set_info_async(NULL, tgt->ltd_exp,
 647                                 sizeof(KEY_CACHE_SET), KEY_CACHE_SET,
 648                                 sizeof(struct cl_client_cache), lov->lov_cache,
 649                                 NULL);
 650                 if (rc < 0)
 651                         GOTO(out, rc);
 652         }
 653
 654         rc = lov_notify(obd, tgt->ltd_exp->exp_obd,
 655                         active ? OBD_NOTIFY_CONNECT : OBD_NOTIFY_INACTIVE,
 656                         (void *)&index);
 657
 658 out:
 659         if (rc) {
 660                 CERROR("add failed (%d), deleting %s\n", rc,
 661                        obd_uuid2str(&tgt->ltd_uuid));
 662                 lov_del_target(obd, index, 0, 0);
 663         }
 664         obd_putref(obd);
 665         RETURN(rc);
 666 }
 667
 668 /* Schedule a target for deletion */
 669 int lov_del_target(struct obd_device *obd, __u32 index,
 670                    struct obd_uuid *uuidp, int gen)
 671 {
 672         struct lov_obd *lov = &obd->u.lov;
 673         int count = lov->desc.ld_tgt_count;
 674         int rc = 0;
 675         ENTRY;
 676
 677         if (index >= count) {
 678                 CERROR("LOV target index %d >= number of LOV OBDs %d.\n",
 679                        index, count);
 680                 RETURN(-EINVAL);
 681         }
 682
 683         /* to make sure there's no ongoing lov_notify() now */
 684         down_write(&lov->lov_notify_lock);
 685         obd_getref(obd);
 686
 687         if (!lov->lov_tgts[index]) {
 688                 CERROR("LOV target at index %d is not setup.\n", index);
 689                 GOTO(out, rc = -EINVAL);
 690         }
 691
 692         if (uuidp && !obd_uuid_equals(uuidp, &lov->lov_tgts[index]->ltd_uuid)) {
 693                 CERROR("LOV target UUID %s at index %d doesn't match %s.\n",
 694                        lov_uuid2str(lov, index), index,
 695                        obd_uuid2str(uuidp));
 696                 GOTO(out, rc = -EINVAL);
 697         }
 698
 699         CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d exp: %p active: %d\n",
 700                lov_uuid2str(lov, index), index,
 701                lov->lov_tgts[index]->ltd_gen, lov->lov_tgts[index]->ltd_exp,
 702                lov->lov_tgts[index]->ltd_active);
 703
 704         lov->lov_tgts[index]->ltd_reap = 1;
 705         lov->lov_death_row++;
 706         /* we really delete it from obd_putref */
 707 out:
 708         obd_putref(obd);
 709         up_write(&lov->lov_notify_lock);
 710
 711         RETURN(rc);
 712 }
 713
 714 static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
 715 {
 716         struct obd_device *osc_obd;
 717
 718         LASSERT(tgt);
 719         LASSERT(tgt->ltd_reap);
 720
 721         osc_obd = class_exp2obd(tgt->ltd_exp);
 722
 723         CDEBUG(D_CONFIG, "Removing tgt %s : %s\n",
 724                tgt->ltd_uuid.uuid,
 725                osc_obd ? osc_obd->obd_name : "<no obd>");
 726
 727         if (tgt->ltd_exp)
 728                 lov_disconnect_obd(obd, tgt);
 729
 730         OBD_FREE_PTR(tgt);
 731
 732         /* Manual cleanup - no cleanup logs to clean up the osc's.  We must
 733            do it ourselves. And we can't do it from lov_cleanup,
 734            because we just lost our only reference to it. */
 735         if (osc_obd)
 736                 class_manual_cleanup(osc_obd);
 737 }
 738
 739 void lov_fix_desc_stripe_size(__u64 *val)
 740 {
 741         if (*val < LOV_MIN_STRIPE_SIZE) {
 742                 if (*val != 0)
 743                         LCONSOLE_INFO("Increasing default stripe size to "
 744                                       "minimum %u\n",
 745                                       LOV_DESC_STRIPE_SIZE_DEFAULT);
 746                 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
 747         } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
 748                 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
 749                 LCONSOLE_WARN("Changing default stripe size to "LPU64" (a "
 750                               "multiple of %u)\n",
 751                               *val, LOV_MIN_STRIPE_SIZE);
 752         }
 753 }
 754
 755 void lov_fix_desc_stripe_count(__u32 *val)
 756 {
 757         if (*val == 0)
 758                 *val = 1;
 759 }
 760
 761 void lov_fix_desc_pattern(__u32 *val)
 762 {
 763         /* from lov_setstripe */
 764         if ((*val != 0) && (*val != LOV_PATTERN_RAID0)) {
 765                 LCONSOLE_WARN("Unknown stripe pattern: %#x\n", *val);
 766                 *val = 0;
 767         }
 768 }
 769
 770 void lov_fix_desc_qos_maxage(__u32 *val)
 771 {
 772         if (*val == 0)
 773                 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
 774 }
 775
 776 void lov_fix_desc(struct lov_desc *desc)
 777 {
 778         lov_fix_desc_stripe_size(&desc->ld_default_stripe_size);
 779         lov_fix_desc_stripe_count(&desc->ld_default_stripe_count);
 780         lov_fix_desc_pattern(&desc->ld_pattern);
 781         lov_fix_desc_qos_maxage(&desc->ld_qos_maxage);
 782 }
 783
 784 int lov_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 785 {
 786         struct lov_desc *desc;
 787         struct lov_obd *lov = &obd->u.lov;
 788 #ifdef LPROCFS
 789         struct obd_type *type;
 790 #endif
 791         int rc;
 792         ENTRY;
 793
 794         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 795                 CERROR("LOV setup requires a descriptor\n");
 796                 RETURN(-EINVAL);
 797         }
 798
 799         desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
 800
 801         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
 802                 CERROR("descriptor size wrong: %d > %d\n",
 803                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
 804                 RETURN(-EINVAL);
 805         }
 806
 807         if (desc->ld_magic != LOV_DESC_MAGIC) {
 808                 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
 809                             CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
 810                                    obd->obd_name, desc);
 811                             lustre_swab_lov_desc(desc);
 812                 } else {
 813                         CERROR("%s: Bad lov desc magic: %#x\n",
 814                                obd->obd_name, desc->ld_magic);
 815                         RETURN(-EINVAL);
 816                 }
 817         }
 818
 819         lov_fix_desc(desc);
 820
 821         desc->ld_active_tgt_count = 0;
 822         lov->desc = *desc;
 823         lov->lov_tgt_size = 0;
 824
 825         mutex_init(&lov->lov_lock);
 826         atomic_set(&lov->lov_refcount, 0);
 827         lov->lov_sp_me = LUSTRE_SP_CLI;
 828
 829         init_rwsem(&lov->lov_notify_lock);
 830
 831         lov->lov_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,
 832                                                    HASH_POOLS_MAX_BITS,
 833                                                    HASH_POOLS_BKT_BITS, 0,
 834                                                    CFS_HASH_MIN_THETA,
 835                                                    CFS_HASH_MAX_THETA,
 836                                                    &pool_hash_operations,
 837                                                    CFS_HASH_DEFAULT);
 838         INIT_LIST_HEAD(&lov->lov_pool_list);
 839         lov->lov_pool_count = 0;
 840         rc = lov_ost_pool_init(&lov->lov_packed, 0);
 841         if (rc)
 842                 GOTO(out, rc);
 843
 844 #ifdef LPROCFS
 845         obd->obd_vars = lprocfs_lov_obd_vars;
 846         /* If this is true then both client (lov) and server
 847          * (lod) are on the same node. The lod layer if loaded
 848          * first will register the lov proc directory. In that
 849          * case obd->obd_type->typ_procroot will be not set.
 850          * Instead we use type->typ_procsym as the parent. */
 851         type = class_search_type(LUSTRE_LOD_NAME);
 852         if (type != NULL && type->typ_procsym != NULL) {
 853                 obd->obd_proc_entry = lprocfs_seq_register(obd->obd_name,
 854                                                            type->typ_procsym,
 855                                                            obd->obd_vars, obd);
 856                 if (IS_ERR(obd->obd_proc_entry)) {
 857                         rc = PTR_ERR(obd->obd_proc_entry);
 858                         CERROR("error %d setting up lprocfs for %s\n", rc,
 859                                obd->obd_name);
 860                         obd->obd_proc_entry = NULL;
 861                 }
 862         } else {
 863                 rc = lprocfs_seq_obd_setup(obd);
 864         }
 865
 866         if (rc == 0) {
 867                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd",
 868                                         0444, &lov_proc_target_fops, obd);
 869                 if (rc)
 870                         CWARN("Error adding the target_obd file\n");
 871
 872                 lov->lov_pool_proc_entry = lprocfs_seq_register("pools",
 873                                                         obd->obd_proc_entry,
 874                                                         NULL, NULL);
 875                 if (IS_ERR(lov->lov_pool_proc_entry)) {
 876                         rc = PTR_ERR(lov->lov_pool_proc_entry);
 877                         CERROR("error %d setting up lprocfs for pools\n", rc);
 878                         lov->lov_pool_proc_entry = NULL;
 879                 }
 880         }
 881 #endif
 882         RETURN(0);
 883
 884 out:
 885         return rc;
 886 }
 887
 888 static int lov_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 889 {
 890         int rc = 0;
 891         struct lov_obd *lov = &obd->u.lov;
 892
 893         ENTRY;
 894
 895         switch (stage) {
 896         case OBD_CLEANUP_EARLY: {
 897                 int i;
 898                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 899                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 900                                 continue;
 901                         obd_precleanup(class_exp2obd(lov->lov_tgts[i]->ltd_exp),
 902                                        OBD_CLEANUP_EARLY);
 903                 }
 904                 break;
 905         }
 906         default:
 907                 break;
 908         }
 909
 910         RETURN(rc);
 911 }
 912
 913 static int lov_cleanup(struct obd_device *obd)
 914 {
 915         struct lov_obd *lov = &obd->u.lov;
 916         struct list_head *pos, *tmp;
 917         struct pool_desc *pool;
 918         ENTRY;
 919
 920         list_for_each_safe(pos, tmp, &lov->lov_pool_list) {
 921                 pool = list_entry(pos, struct pool_desc, pool_list);
 922                 /* free pool structs */
 923                 CDEBUG(D_INFO, "delete pool %p\n", pool);
 924                 /* In the function below, .hs_keycmp resolves to
 925                  * pool_hashkey_keycmp() */
 926                 /* coverity[overrun-buffer-val] */
 927                 lov_pool_del(obd, pool->pool_name);
 928         }
 929         cfs_hash_putref(lov->lov_pools_hash_body);
 930         lov_ost_pool_free(&lov->lov_packed);
 931
 932         lprocfs_obd_cleanup(obd);
 933         if (lov->lov_tgts) {
 934                 int i;
 935                 obd_getref(obd);
 936                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 937                         if (!lov->lov_tgts[i])
 938                                 continue;
 939
 940                         /* Inactive targets may never have connected */
 941                         if (lov->lov_tgts[i]->ltd_active ||
 942                             atomic_read(&lov->lov_refcount))
 943                                 /* We should never get here - these
 944                                  * should have been removed in the
 945                                  * disconnect. */
 946                                 CERROR("%s: lov tgt %d not cleaned! "
 947                                        "deathrow=%d, lovrc=%d\n",
 948                                        obd->obd_name, i, lov->lov_death_row,
 949                                        atomic_read(&lov->lov_refcount));
 950                         lov_del_target(obd, i, 0, 0);
 951                 }
 952                 obd_putref(obd);
 953                 OBD_FREE(lov->lov_tgts, sizeof(*lov->lov_tgts) *
 954                          lov->lov_tgt_size);
 955                 lov->lov_tgt_size = 0;
 956         }
 957         RETURN(0);
 958 }
 959
 960 int lov_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg,
 961                             __u32 *indexp, int *genp)
 962 {
 963         struct obd_uuid obd_uuid;
 964         int cmd;
 965         int rc = 0;
 966         ENTRY;
 967
 968         switch(cmd = lcfg->lcfg_command) {
 969         case LCFG_LOV_ADD_OBD:
 970         case LCFG_LOV_ADD_INA:
 971         case LCFG_LOV_DEL_OBD: {
 972                 __u32 index;
 973                 int gen;
 974                 /* lov_modify_tgts add  0:lov_mdsA  1:ost1_UUID  2:0  3:1 */
 975                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
 976                         GOTO(out, rc = -EINVAL);
 977
 978                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
 979
 980                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", indexp) != 1)
 981                         GOTO(out, rc = -EINVAL);
 982                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", genp) != 1)
 983                         GOTO(out, rc = -EINVAL);
 984                 index = *indexp;
 985                 gen = *genp;
 986                 if (cmd == LCFG_LOV_ADD_OBD)
 987                         rc = lov_add_target(obd, &obd_uuid, index, gen, 1);
 988                 else if (cmd == LCFG_LOV_ADD_INA)
 989                         rc = lov_add_target(obd, &obd_uuid, index, gen, 0);
 990                 else
 991                         rc = lov_del_target(obd, index, &obd_uuid, gen);
 992                 GOTO(out, rc);
 993         }
 994         case LCFG_PARAM: {
 995                 struct lov_desc *desc = &(obd->u.lov.desc);
 996
 997                 if (!desc)
 998                         GOTO(out, rc = -EINVAL);
 999
1000                 rc = class_process_proc_seq_param(PARAM_LOV, obd->obd_vars,
1001                                                   lcfg, obd);
1002                 if (rc > 0)
1003                         rc = 0;
1004                 GOTO(out, rc);
1005         }
1006         case LCFG_POOL_NEW:
1007         case LCFG_POOL_ADD:
1008         case LCFG_POOL_DEL:
1009         case LCFG_POOL_REM:
1010                 GOTO(out, rc);
1011
1012         default: {
1013                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1014                 GOTO(out, rc = -EINVAL);
1015
1016         }
1017         }
1018 out:
1019         RETURN(rc);
1020 }
1021
1022 static int lov_recreate(struct obd_export *exp, struct obdo *src_oa,
1023                         struct lov_stripe_md **ea, struct obd_trans_info *oti)
1024 {
1025         struct lov_stripe_md *obj_mdp, *lsm;
1026         struct lov_obd *lov = &exp->exp_obd->u.lov;
1027         unsigned ost_idx;
1028         int rc, i;
1029         ENTRY;
1030
1031         LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS &&
1032                 src_oa->o_flags & OBD_FL_RECREATE_OBJS);
1033
1034         OBD_ALLOC(obj_mdp, sizeof(*obj_mdp));
1035         if (obj_mdp == NULL)
1036                 RETURN(-ENOMEM);
1037
1038         ost_idx = src_oa->o_nlink;
1039         lsm = *ea;
1040         if (lsm == NULL)
1041                 GOTO(out, rc = -EINVAL);
1042         if (ost_idx >= lov->desc.ld_tgt_count ||
1043             !lov->lov_tgts[ost_idx])
1044                 GOTO(out, rc = -EINVAL);
1045
1046         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1047                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
1048
1049                 if (lov_oinfo_is_dummy(loi))
1050                         continue;
1051
1052                 if (loi->loi_ost_idx == ost_idx) {
1053                         if (ostid_id(&loi->loi_oi) !=
1054                                         ostid_id(&src_oa->o_oi))
1055                                 GOTO(out, rc = -EINVAL);
1056                         break;
1057                 }
1058         }
1059         if (i == lsm->lsm_stripe_count)
1060                 GOTO(out, rc = -EINVAL);
1061
1062         rc = obd_create(NULL, lov->lov_tgts[ost_idx]->ltd_exp,
1063                         src_oa, &obj_mdp, oti);
1064 out:
1065         OBD_FREE(obj_mdp, sizeof(*obj_mdp));
1066         RETURN(rc);
1067 }
1068
1069 /* the LOV expects oa->o_id to be set to the LOV object id */
1070 static int lov_create(const struct lu_env *env, struct obd_export *exp,
1071                       struct obdo *src_oa, struct lov_stripe_md **ea,
1072                       struct obd_trans_info *oti)
1073 {
1074         struct lov_obd *lov;
1075         int rc = 0;
1076         ENTRY;
1077
1078         LASSERT(ea != NULL);
1079         if (exp == NULL)
1080                 RETURN(-EINVAL);
1081
1082         if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
1083             src_oa->o_flags == OBD_FL_DELORPHAN) {
1084                 /* should be used with LOV anymore */
1085                 LBUG();
1086         }
1087
1088         lov = &exp->exp_obd->u.lov;
1089         if (!lov->desc.ld_active_tgt_count)
1090                 RETURN(-EIO);
1091
1092         obd_getref(exp->exp_obd);
1093         /* Recreate a specific object id at the given OST index */
1094         if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
1095             (src_oa->o_flags & OBD_FL_RECREATE_OBJS)) {
1096                  rc = lov_recreate(exp, src_oa, ea, oti);
1097         }
1098
1099         obd_putref(exp->exp_obd);
1100         RETURN(rc);
1101 }
1102
1103 #define ASSERT_LSM_MAGIC(lsmp)                                                  \
1104 do {                                                                            \
1105         LASSERT((lsmp) != NULL);                                                \
1106         LASSERTF(((lsmp)->lsm_magic == LOV_MAGIC_V1 ||                          \
1107                  (lsmp)->lsm_magic == LOV_MAGIC_V3),                            \
1108                  "%p->lsm_magic=%x\n", (lsmp), (lsmp)->lsm_magic);              \
1109 } while (0)
1110
1111 static int lov_destroy(const struct lu_env *env, struct obd_export *exp,
1112                        struct obdo *oa, struct lov_stripe_md *lsm,
1113                        struct obd_trans_info *oti, struct obd_export *md_exp,
1114                        void *capa)
1115 {
1116         struct lov_request_set *set;
1117         struct obd_info oinfo;
1118         struct lov_request *req;
1119         struct list_head *pos;
1120         struct lov_obd *lov;
1121         int rc = 0, err = 0;
1122         ENTRY;
1123
1124         ASSERT_LSM_MAGIC(lsm);
1125
1126         if (!exp || !exp->exp_obd)
1127                 RETURN(-ENODEV);
1128
1129         if (oa->o_valid & OBD_MD_FLCOOKIE) {
1130                 LASSERT(oti);
1131                 LASSERT(oti->oti_logcookies);
1132         }
1133
1134         lov = &exp->exp_obd->u.lov;
1135         obd_getref(exp->exp_obd);
1136         rc = lov_prep_destroy_set(exp, &oinfo, oa, lsm, oti, &set);
1137         if (rc)
1138                 GOTO(out, rc);
1139
1140         list_for_each(pos, &set->set_list) {
1141                 req = list_entry(pos, struct lov_request, rq_link);
1142
1143                 if (oa->o_valid & OBD_MD_FLCOOKIE)
1144                         oti->oti_logcookies = set->set_cookies + req->rq_stripe;
1145
1146                 err = obd_destroy(env, lov->lov_tgts[req->rq_idx]->ltd_exp,
1147                                   req->rq_oi.oi_oa, NULL, oti, NULL, capa);
1148                 err = lov_update_common_set(set, req, err);
1149                 if (err) {
1150                         CERROR("%s: destroying objid "DOSTID" subobj "
1151                                DOSTID" on OST idx %d: rc = %d\n",
1152                                exp->exp_obd->obd_name, POSTID(&oa->o_oi),
1153                                POSTID(&req->rq_oi.oi_oa->o_oi),
1154                                req->rq_idx, err);
1155                         if (!rc)
1156                                 rc = err;
1157                 }
1158         }
1159
1160         if (rc == 0) {
1161                 LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
1162                 rc = lsm_op_find(lsm->lsm_magic)->lsm_destroy(lsm, oa, md_exp);
1163         }
1164         err = lov_fini_destroy_set(set);
1165 out:
1166         obd_putref(exp->exp_obd);
1167         RETURN(rc ? rc : err);
1168 }
1169
1170 static int lov_getattr_interpret(struct ptlrpc_request_set *rqset,
1171                                  void *data, int rc)
1172 {
1173         struct lov_request_set *lovset = (struct lov_request_set *)data;
1174         int err;
1175         ENTRY;
1176
1177         /* don't do attribute merge if this aysnc op failed */
1178         if (rc)
1179                 atomic_set(&lovset->set_completes, 0);
1180         err = lov_fini_getattr_set(lovset);
1181         RETURN(rc ? rc : err);
1182 }
1183
1184 static int lov_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
1185                               struct ptlrpc_request_set *rqset)
1186 {
1187         struct lov_request_set *lovset;
1188         struct lov_obd *lov;
1189         struct list_head *pos;
1190         struct lov_request *req;
1191         int rc = 0, err;
1192         ENTRY;
1193
1194         LASSERT(oinfo);
1195         ASSERT_LSM_MAGIC(oinfo->oi_md);
1196
1197         if (!exp || !exp->exp_obd)
1198                 RETURN(-ENODEV);
1199
1200         lov = &exp->exp_obd->u.lov;
1201
1202         rc = lov_prep_getattr_set(exp, oinfo, &lovset);
1203         if (rc)
1204                 RETURN(rc);
1205
1206         CDEBUG(D_INFO, "objid "DOSTID": %ux%u byte stripes\n",
1207                POSTID(&oinfo->oi_md->lsm_oi), oinfo->oi_md->lsm_stripe_count,
1208                oinfo->oi_md->lsm_stripe_size);
1209
1210         list_for_each(pos, &lovset->set_list) {
1211                 req = list_entry(pos, struct lov_request, rq_link);
1212
1213                 CDEBUG(D_INFO, "objid "DOSTID"[%d] has subobj "DOSTID" at idx"
1214                        "%u\n", POSTID(&oinfo->oi_oa->o_oi), req->rq_stripe,
1215                        POSTID(&req->rq_oi.oi_oa->o_oi), req->rq_idx);
1216                 rc = obd_getattr_async(lov->lov_tgts[req->rq_idx]->ltd_exp,
1217                                        &req->rq_oi, rqset);
1218                 if (rc) {
1219                         CERROR("%s: getattr objid "DOSTID" subobj"
1220                                DOSTID" on OST idx %d: rc = %d\n",
1221                                exp->exp_obd->obd_name,
1222                                POSTID(&oinfo->oi_oa->o_oi),
1223                                POSTID(&req->rq_oi.oi_oa->o_oi),
1224                                req->rq_idx, rc);
1225                         GOTO(out, rc);
1226                 }
1227         }
1228
1229         if (!list_empty(&rqset->set_requests)) {
1230                 LASSERT(rc == 0);
1231                 LASSERT (rqset->set_interpret == NULL);
1232                 rqset->set_interpret = lov_getattr_interpret;
1233                 rqset->set_arg = (void *)lovset;
1234                 RETURN(rc);
1235         }
1236 out:
1237         if (rc)
1238                 atomic_set(&lovset->set_completes, 0);
1239         err = lov_fini_getattr_set(lovset);
1240         RETURN(rc ? rc : err);
1241 }
1242
1243 static int lov_setattr_interpret(struct ptlrpc_request_set *rqset,
1244                                  void *data, int rc)
1245 {
1246         struct lov_request_set *lovset = (struct lov_request_set *)data;
1247         int err;
1248         ENTRY;
1249
1250         if (rc)
1251                 atomic_set(&lovset->set_completes, 0);
1252         err = lov_fini_setattr_set(lovset);
1253         RETURN(rc ? rc : err);
1254 }
1255
1256 /* If @oti is given, the request goes from MDS and responses from OSTs are not
1257    needed. Otherwise, a client is waiting for responses. */
1258 static int lov_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
1259                              struct obd_trans_info *oti,
1260                              struct ptlrpc_request_set *rqset)
1261 {
1262         struct lov_request_set *set;
1263         struct lov_request *req;
1264         struct list_head *pos;
1265         struct lov_obd *lov;
1266         int rc = 0;
1267         ENTRY;
1268
1269         LASSERT(oinfo);
1270         ASSERT_LSM_MAGIC(oinfo->oi_md);
1271         if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE) {
1272                 LASSERT(oti);
1273                 LASSERT(oti->oti_logcookies);
1274         }
1275
1276         if (!exp || !exp->exp_obd)
1277                 RETURN(-ENODEV);
1278
1279         lov = &exp->exp_obd->u.lov;
1280         rc = lov_prep_setattr_set(exp, oinfo, oti, &set);
1281         if (rc)
1282                 RETURN(rc);
1283
1284         CDEBUG(D_INFO, "objid "DOSTID": %ux%u byte stripes\n",
1285                POSTID(&oinfo->oi_md->lsm_oi),
1286                oinfo->oi_md->lsm_stripe_count,
1287                oinfo->oi_md->lsm_stripe_size);
1288
1289         list_for_each(pos, &set->set_list) {
1290                 req = list_entry(pos, struct lov_request, rq_link);
1291
1292                 if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
1293                         oti->oti_logcookies = set->set_cookies + req->rq_stripe;
1294
1295                 CDEBUG(D_INFO, "objid "DOSTID"[%d] has subobj "DOSTID" at idx"
1296                        "%u\n", POSTID(&oinfo->oi_oa->o_oi), req->rq_stripe,
1297                        POSTID(&req->rq_oi.oi_oa->o_oi), req->rq_idx);
1298
1299                 rc = obd_setattr_async(lov->lov_tgts[req->rq_idx]->ltd_exp,
1300                                        &req->rq_oi, oti, rqset);
1301                 if (rc) {
1302                         CERROR("error: setattr objid "DOSTID" subobj"
1303                                DOSTID" on OST idx %d: rc = %d\n",
1304                                POSTID(&set->set_oi->oi_oa->o_oi),
1305                                POSTID(&req->rq_oi.oi_oa->o_oi),
1306                                req->rq_idx, rc);
1307                         break;
1308                 }
1309         }
1310
1311         /* If we are not waiting for responses on async requests, return. */
1312         if (rc || !rqset || list_empty(&rqset->set_requests)) {
1313                 int err;
1314                 if (rc)
1315                         atomic_set(&set->set_completes, 0);
1316                 err = lov_fini_setattr_set(set);
1317                 RETURN(rc ? rc : err);
1318         }
1319
1320         LASSERT(rqset->set_interpret == NULL);
1321         rqset->set_interpret = lov_setattr_interpret;
1322         rqset->set_arg = (void *)set;
1323
1324         RETURN(0);
1325 }
1326
1327 static int lov_change_cbdata(struct obd_export *exp,
1328                              struct lov_stripe_md *lsm, ldlm_iterator_t it,
1329                              void *data)
1330 {
1331         struct lov_obd *lov;
1332         int rc = 0, i;
1333         ENTRY;
1334
1335         ASSERT_LSM_MAGIC(lsm);
1336
1337         if (!exp || !exp->exp_obd)
1338                 RETURN(-ENODEV);
1339
1340         lov = &exp->exp_obd->u.lov;
1341         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1342                 struct lov_stripe_md submd;
1343                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
1344
1345                 if (lov_oinfo_is_dummy(loi))
1346                         continue;
1347
1348                 if (!lov->lov_tgts[loi->loi_ost_idx]) {
1349                         CDEBUG(D_HA, "lov idx %d NULL\n", loi->loi_ost_idx);
1350                         continue;
1351                 }
1352
1353                 submd.lsm_oi = loi->loi_oi;
1354                 submd.lsm_stripe_count = 0;
1355                 rc = obd_change_cbdata(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
1356                                        &submd, it, data);
1357         }
1358         RETURN(rc);
1359 }
1360
1361 /* find any ldlm lock of the inode in lov
1362  * return 0    not find
1363  *        1    find one
1364  *      < 0    error */
1365 static int lov_find_cbdata(struct obd_export *exp,
1366                            struct lov_stripe_md *lsm, ldlm_iterator_t it,
1367                            void *data)
1368 {
1369         struct lov_obd *lov;
1370         int rc = 0, i;
1371         ENTRY;
1372
1373         ASSERT_LSM_MAGIC(lsm);
1374
1375         if (!exp || !exp->exp_obd)
1376                 RETURN(-ENODEV);
1377
1378         lov = &exp->exp_obd->u.lov;
1379         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1380                 struct lov_stripe_md submd;
1381                 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
1382
1383                 if (lov_oinfo_is_dummy(loi))
1384                         continue;
1385
1386                 if (!lov->lov_tgts[loi->loi_ost_idx]) {
1387                         CDEBUG(D_HA, "lov idx %d NULL\n", loi->loi_ost_idx);
1388                         continue;
1389                 }
1390
1391                 submd.lsm_oi = loi->loi_oi;
1392                 submd.lsm_stripe_count = 0;
1393                 rc = obd_find_cbdata(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
1394                                      &submd, it, data);
1395                 if (rc != 0)
1396                         RETURN(rc);
1397         }
1398         RETURN(rc);
1399 }
1400
1401 int lov_statfs_interpret(struct ptlrpc_request_set *rqset, void *data, int rc)
1402 {
1403         struct lov_request_set *lovset = (struct lov_request_set *)data;
1404         int err;
1405         ENTRY;
1406
1407         if (rc)
1408                 atomic_set(&lovset->set_completes, 0);
1409
1410         err = lov_fini_statfs_set(lovset);
1411         RETURN(rc ? rc : err);
1412 }
1413
1414 static int lov_statfs_async(struct obd_export *exp, struct obd_info *oinfo,
1415                             __u64 max_age, struct ptlrpc_request_set *rqset)
1416 {
1417         struct obd_device      *obd = class_exp2obd(exp);
1418         struct lov_request_set *set;
1419         struct lov_request *req;
1420         struct list_head *pos;
1421         struct lov_obd *lov;
1422         int rc = 0;
1423         ENTRY;
1424
1425         LASSERT(oinfo != NULL);
1426         LASSERT(oinfo->oi_osfs != NULL);
1427
1428         lov = &obd->u.lov;
1429         rc = lov_prep_statfs_set(obd, oinfo, &set);
1430         if (rc)
1431                 RETURN(rc);
1432
1433         list_for_each(pos, &set->set_list) {
1434                 req = list_entry(pos, struct lov_request, rq_link);
1435                 rc = obd_statfs_async(lov->lov_tgts[req->rq_idx]->ltd_exp,
1436                                       &req->rq_oi, max_age, rqset);
1437                 if (rc)
1438                         break;
1439         }
1440
1441         if (rc || list_empty(&rqset->set_requests)) {
1442                 int err;
1443                 if (rc)
1444                         atomic_set(&set->set_completes, 0);
1445                 err = lov_fini_statfs_set(set);
1446                 RETURN(rc ? rc : err);
1447         }
1448
1449         LASSERT(rqset->set_interpret == NULL);
1450         rqset->set_interpret = lov_statfs_interpret;
1451         rqset->set_arg = (void *)set;
1452         RETURN(0);
1453 }
1454
1455 static int lov_statfs(const struct lu_env *env, struct obd_export *exp,
1456                       struct obd_statfs *osfs, __u64 max_age, __u32 flags)
1457 {
1458         struct ptlrpc_request_set *set = NULL;
1459         struct obd_info oinfo = { { { 0 } } };
1460         int rc = 0;
1461         ENTRY;
1462
1463
1464         /* for obdclass we forbid using obd_statfs_rqset, but prefer using async
1465          * statfs requests */
1466         set = ptlrpc_prep_set();
1467         if (set == NULL)
1468                 RETURN(-ENOMEM);
1469
1470         oinfo.oi_osfs = osfs;
1471         oinfo.oi_flags = flags;
1472         rc = lov_statfs_async(exp, &oinfo, max_age, set);
1473         if (rc == 0)
1474                 rc = ptlrpc_set_wait(set);
1475         ptlrpc_set_destroy(set);
1476
1477         RETURN(rc);
1478 }
1479
1480 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1481                          void *karg, void *uarg)
1482 {
1483         struct obd_device *obddev = class_exp2obd(exp);
1484         struct lov_obd *lov = &obddev->u.lov;
1485         int i = 0, rc = 0, count = lov->desc.ld_tgt_count;
1486         struct obd_uuid *uuidp;
1487         ENTRY;
1488
1489         switch (cmd) {
1490         case IOC_OBD_STATFS: {
1491                 struct obd_ioctl_data *data = karg;
1492                 struct obd_device *osc_obd;
1493                 struct obd_statfs stat_buf = {0};
1494                 __u32 index;
1495                 __u32 flags;
1496
1497                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
1498                 if ((index >= count))
1499                         RETURN(-ENODEV);
1500
1501                 if (!lov->lov_tgts[index])
1502                         /* Try again with the next index */
1503                         RETURN(-EAGAIN);
1504                 if (!lov->lov_tgts[index]->ltd_active)
1505                         RETURN(-ENODATA);
1506
1507                 osc_obd = class_exp2obd(lov->lov_tgts[index]->ltd_exp);
1508                 if (!osc_obd)
1509                         RETURN(-EINVAL);
1510
1511                 /* copy UUID */
1512                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(osc_obd),
1513                                  min((int)data->ioc_plen2,
1514                                      (int)sizeof(struct obd_uuid))))
1515                         RETURN(-EFAULT);
1516
1517                 flags = uarg ? *(__u32*)uarg : 0;
1518                 /* got statfs data */
1519                 rc = obd_statfs(NULL, lov->lov_tgts[index]->ltd_exp, &stat_buf,
1520                                 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
1521                                 flags);
1522                 if (rc)
1523                         RETURN(rc);
1524                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
1525                                      min((int) data->ioc_plen1,
1526                                          (int) sizeof(stat_buf))))
1527                         RETURN(-EFAULT);
1528                 break;
1529         }
1530         case OBD_IOC_LOV_GET_CONFIG: {
1531                 struct obd_ioctl_data *data;
1532                 struct lov_desc *desc;
1533                 char *buf = NULL;
1534                 __u32 *genp;
1535
1536                 len = 0;
1537                 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
1538                         RETURN(-EINVAL);
1539
1540                 data = (struct obd_ioctl_data *)buf;
1541
1542                 if (sizeof(*desc) > data->ioc_inllen1) {
1543                         obd_ioctl_freedata(buf, len);
1544                         RETURN(-EINVAL);
1545                 }
1546
1547                 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
1548                         obd_ioctl_freedata(buf, len);
1549                         RETURN(-EINVAL);
1550                 }
1551
1552                 if (sizeof(__u32) * count > data->ioc_inllen3) {
1553                         obd_ioctl_freedata(buf, len);
1554                         RETURN(-EINVAL);
1555                 }
1556
1557                 desc = (struct lov_desc *)data->ioc_inlbuf1;
1558                 memcpy(desc, &(lov->desc), sizeof(*desc));
1559
1560                 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
1561                 genp = (__u32 *)data->ioc_inlbuf3;
1562                 /* the uuid will be empty for deleted OSTs */
1563                 for (i = 0; i < count; i++, uuidp++, genp++) {
1564                         if (!lov->lov_tgts[i])
1565                                 continue;
1566                         *uuidp = lov->lov_tgts[i]->ltd_uuid;
1567                         *genp = lov->lov_tgts[i]->ltd_gen;
1568                 }
1569
1570                 if (copy_to_user((void *)uarg, buf, len))
1571                         rc = -EFAULT;
1572                 obd_ioctl_freedata(buf, len);
1573                 break;
1574         }
1575         case LL_IOC_LOV_GETSTRIPE:
1576                 rc = lov_getstripe(exp, karg, uarg);
1577                 break;
1578         case OBD_IOC_QUOTACTL: {
1579                 struct if_quotactl *qctl = karg;
1580                 struct lov_tgt_desc *tgt = NULL;
1581                 struct obd_quotactl *oqctl;
1582
1583                 if (qctl->qc_valid == QC_OSTIDX) {
1584                         if (count <= qctl->qc_idx)
1585                                 RETURN(-EINVAL);
1586
1587                         tgt = lov->lov_tgts[qctl->qc_idx];
1588                         if (!tgt || !tgt->ltd_exp)
1589                                 RETURN(-EINVAL);
1590                 } else if (qctl->qc_valid == QC_UUID) {
1591                         for (i = 0; i < count; i++) {
1592                                 tgt = lov->lov_tgts[i];
1593                                 if (!tgt ||
1594                                     !obd_uuid_equals(&tgt->ltd_uuid,
1595                                                      &qctl->obd_uuid))
1596                                         continue;
1597
1598                                 if (tgt->ltd_exp == NULL)
1599                                         RETURN(-EINVAL);
1600
1601                                 break;
1602                         }
1603                 } else {
1604                         RETURN(-EINVAL);
1605                 }
1606
1607                 if (i >= count)
1608                         RETURN(-EAGAIN);
1609
1610                 LASSERT(tgt && tgt->ltd_exp);
1611                 OBD_ALLOC_PTR(oqctl);
1612                 if (!oqctl)
1613                         RETURN(-ENOMEM);
1614
1615                 QCTL_COPY(oqctl, qctl);
1616                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
1617                 if (rc == 0) {
1618                         QCTL_COPY(qctl, oqctl);
1619                         qctl->qc_valid = QC_OSTIDX;
1620                         qctl->obd_uuid = tgt->ltd_uuid;
1621                 }
1622                 OBD_FREE_PTR(oqctl);
1623                 break;
1624         }
1625         default: {
1626                 int set = 0;
1627
1628                 if (count == 0)
1629                         RETURN(-ENOTTY);
1630
1631                 for (i = 0; i < count; i++) {
1632                         int err;
1633                         struct obd_device *osc_obd;
1634
1635                         /* OST was disconnected */
1636                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
1637                                 continue;
1638
1639                         /* ll_umount_begin() sets force flag but for lov, not
1640                          * osc. Let's pass it through */
1641                         osc_obd = class_exp2obd(lov->lov_tgts[i]->ltd_exp);
1642                         osc_obd->obd_force = obddev->obd_force;
1643                         err = obd_iocontrol(cmd, lov->lov_tgts[i]->ltd_exp,
1644                                             len, karg, uarg);
1645                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
1646                                 RETURN(err);
1647                         } else if (err) {
1648                                 if (lov->lov_tgts[i]->ltd_active) {
1649                                         CDEBUG(err == -ENOTTY ?
1650                                                D_IOCTL : D_WARNING,
1651                                                "iocontrol OSC %s on OST "
1652                                                "idx %d cmd %x: err = %d\n",
1653                                                lov_uuid2str(lov, i),
1654                                                i, cmd, err);
1655                                         if (!rc)
1656                                                 rc = err;
1657                                 }
1658                         } else {
1659                                 set = 1;
1660                         }
1661                 }
1662                 if (!set && !rc)
1663                         rc = -EIO;
1664         }
1665         }
1666
1667         RETURN(rc);
1668 }
1669
1670 #define FIEMAP_BUFFER_SIZE 4096
1671
1672 /**
1673  * Non-zero fe_logical indicates that this is a continuation FIEMAP
1674  * call. The local end offset and the device are sent in the first
1675  * fm_extent. This function calculates the stripe number from the index.
1676  * This function returns a stripe_no on which mapping is to be restarted.
1677  *
1678  * This function returns fm_end_offset which is the in-OST offset at which
1679  * mapping should be restarted. If fm_end_offset=0 is returned then caller
1680  * will re-calculate proper offset in next stripe.
1681  * Note that the first extent is passed to lov_get_info via the value field.
1682  *
1683  * \param fiemap fiemap request header
1684  * \param lsm striping information for the file
1685  * \param fm_start logical start of mapping
1686  * \param fm_end logical end of mapping
1687  * \param start_stripe starting stripe will be returned in this
1688  */
1689 obd_size fiemap_calc_fm_end_offset(struct ll_user_fiemap *fiemap,
1690                                    struct lov_stripe_md *lsm, obd_size fm_start,
1691                                    obd_size fm_end, int *start_stripe)
1692 {
1693         obd_size local_end = fiemap->fm_extents[0].fe_logical;
1694         obd_off lun_start, lun_end;
1695         obd_size fm_end_offset;
1696         int stripe_no = -1, i;
1697
1698         if (fiemap->fm_extent_count == 0 ||
1699             fiemap->fm_extents[0].fe_logical == 0)
1700                 return 0;
1701
1702         /* Find out stripe_no from ost_index saved in the fe_device */
1703         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1704                 struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
1705
1706                 if (lov_oinfo_is_dummy(oinfo))
1707                         continue;
1708
1709                 if (oinfo->loi_ost_idx == fiemap->fm_extents[0].fe_device) {
1710                         stripe_no = i;
1711                         break;
1712                 }
1713         }
1714
1715         if (stripe_no == -1)
1716                 return -EINVAL;
1717
1718         /* If we have finished mapping on previous device, shift logical
1719          * offset to start of next device */
1720         if ((lov_stripe_intersects(lsm, stripe_no, fm_start, fm_end,
1721                                    &lun_start, &lun_end)) != 0 &&
1722                                    local_end < lun_end) {
1723                 fm_end_offset = local_end;
1724                 *start_stripe = stripe_no;
1725         } else {
1726                 /* This is a special value to indicate that caller should
1727                  * calculate offset in next stripe. */
1728                 fm_end_offset = 0;
1729                 *start_stripe = (stripe_no + 1) % lsm->lsm_stripe_count;
1730         }
1731
1732         return fm_end_offset;
1733 }
1734
1735 /**
1736  * We calculate on which OST the mapping will end. If the length of mapping
1737  * is greater than (stripe_size * stripe_count) then the last_stripe will
1738  * will be one just before start_stripe. Else we check if the mapping
1739  * intersects each OST and find last_stripe.
1740  * This function returns the last_stripe and also sets the stripe_count
1741  * over which the mapping is spread
1742  *
1743  * \param lsm striping information for the file
1744  * \param fm_start logical start of mapping
1745  * \param fm_end logical end of mapping
1746  * \param start_stripe starting stripe of the mapping
1747  * \param stripe_count the number of stripes across which to map is returned
1748  *
1749  * \retval last_stripe return the last stripe of the mapping
1750  */
1751 int fiemap_calc_last_stripe(struct lov_stripe_md *lsm, obd_size fm_start,
1752                             obd_size fm_end, int start_stripe,
1753                             int *stripe_count)
1754 {
1755         int last_stripe;
1756         obd_off obd_start, obd_end;
1757         int i, j;
1758
1759         if (fm_end - fm_start > lsm->lsm_stripe_size * lsm->lsm_stripe_count) {
1760                 last_stripe = (start_stripe < 1 ? lsm->lsm_stripe_count - 1 :
1761                                                               start_stripe - 1);
1762                 *stripe_count = lsm->lsm_stripe_count;
1763         } else {
1764                 for (j = 0, i = start_stripe; j < lsm->lsm_stripe_count;
1765                      i = (i + 1) % lsm->lsm_stripe_count, j++) {
1766                         if ((lov_stripe_intersects(lsm, i, fm_start, fm_end,
1767                                                    &obd_start, &obd_end)) == 0)
1768                                 break;
1769                 }
1770                 *stripe_count = j;
1771                 last_stripe = (start_stripe + j - 1) %lsm->lsm_stripe_count;
1772         }
1773
1774         return last_stripe;
1775 }
1776
1777 /**
1778  * Set fe_device and copy extents from local buffer into main return buffer.
1779  *
1780  * \param fiemap fiemap request header
1781  * \param lcl_fm_ext array of local fiemap extents to be copied
1782  * \param ost_index OST index to be written into the fm_device field for each
1783                     extent
1784  * \param ext_count number of extents to be copied
1785  * \param current_extent where to start copying in main extent array
1786  */
1787 void fiemap_prepare_and_copy_exts(struct ll_user_fiemap *fiemap,
1788                                   struct ll_fiemap_extent *lcl_fm_ext,
1789                                   int ost_index, unsigned int ext_count,
1790                                   int current_extent)
1791 {
1792         char *to;
1793         int ext;
1794
1795         for (ext = 0; ext < ext_count; ext++) {
1796                 lcl_fm_ext[ext].fe_device = ost_index;
1797                 lcl_fm_ext[ext].fe_flags |= FIEMAP_EXTENT_NET;
1798         }
1799
1800         /* Copy fm_extent's from fm_local to return buffer */
1801         to = (char *)fiemap + fiemap_count_to_size(current_extent);
1802         memcpy(to, lcl_fm_ext, ext_count * sizeof(struct ll_fiemap_extent));
1803 }
1804
1805 /**
1806  * Break down the FIEMAP request and send appropriate calls to individual OSTs.
1807  * This also handles the restarting of FIEMAP calls in case mapping overflows
1808  * the available number of extents in single call.
1809  */
1810 static int lov_fiemap(struct lov_obd *lov, __u32 keylen, void *key,
1811                       __u32 *vallen, void *val, struct lov_stripe_md *lsm)
1812 {
1813         struct ll_fiemap_info_key *fm_key = key;
1814         struct ll_user_fiemap *fiemap = val;
1815         struct ll_user_fiemap *fm_local = NULL;
1816         struct ll_fiemap_extent *lcl_fm_ext;
1817         int count_local;
1818         unsigned int get_num_extents = 0;
1819         int ost_index = 0, actual_start_stripe, start_stripe;
1820         obd_size fm_start, fm_end, fm_length, fm_end_offset;
1821         obd_size curr_loc;
1822         int current_extent = 0, rc = 0, i;
1823         int ost_eof = 0; /* EOF for object */
1824         int ost_done = 0; /* done with required mapping for this OST? */
1825         int last_stripe;
1826         int cur_stripe = 0, cur_stripe_wrap = 0, stripe_count;
1827         unsigned int buffer_size = FIEMAP_BUFFER_SIZE;
1828
1829         if (!lsm_has_objects(lsm)) {
1830                 if (lsm && lsm_is_released(lsm) && (fm_key->fiemap.fm_start <
1831                     fm_key->oa.o_size)) {
1832                         /* released file, return a minimal FIEMAP if
1833                          * request fits in file-size.
1834                          */
1835                         fiemap->fm_mapped_extents = 1;
1836                         fiemap->fm_extents[0].fe_logical =
1837                                                 fm_key->fiemap.fm_start;
1838                         if (fm_key->fiemap.fm_start + fm_key->fiemap.fm_length <
1839                             fm_key->oa.o_size)
1840                                 fiemap->fm_extents[0].fe_length =
1841                                                 fm_key->fiemap.fm_length;
1842                         else
1843                                 fiemap->fm_extents[0].fe_length =
1844                                                 fm_key->oa.o_size -
1845                                                 fm_key->fiemap.fm_start;
1846                         fiemap->fm_extents[0].fe_flags |=
1847                                                 (FIEMAP_EXTENT_UNKNOWN |
1848                                                  FIEMAP_EXTENT_LAST);
1849                 }
1850                 GOTO(out, rc = 0);
1851         }
1852
1853         if (fiemap_count_to_size(fm_key->fiemap.fm_extent_count) < buffer_size)
1854                 buffer_size = fiemap_count_to_size(fm_key->fiemap.fm_extent_count);
1855
1856         OBD_ALLOC_LARGE(fm_local, buffer_size);
1857         if (fm_local == NULL)
1858                 GOTO(out, rc = -ENOMEM);
1859         lcl_fm_ext = &fm_local->fm_extents[0];
1860
1861         count_local = fiemap_size_to_count(buffer_size);
1862
1863         memcpy(fiemap, &fm_key->fiemap, sizeof(*fiemap));
1864         fm_start = fiemap->fm_start;
1865         fm_length = fiemap->fm_length;
1866         /* Calculate start stripe, last stripe and length of mapping */
1867         actual_start_stripe = start_stripe = lov_stripe_number(lsm, fm_start);
1868         fm_end = (fm_length == ~0ULL ? fm_key->oa.o_size :
1869                                                 fm_start + fm_length - 1);
1870         /* If fm_length != ~0ULL but fm_start+fm_length-1 exceeds file size */
1871         if (fm_end > fm_key->oa.o_size)
1872                 fm_end = fm_key->oa.o_size;
1873
1874         last_stripe = fiemap_calc_last_stripe(lsm, fm_start, fm_end,
1875                                             actual_start_stripe, &stripe_count);
1876
1877         fm_end_offset = fiemap_calc_fm_end_offset(fiemap, lsm, fm_start,
1878                                                   fm_end, &start_stripe);
1879         if (fm_end_offset == -EINVAL)
1880                 GOTO(out, rc = -EINVAL);
1881
1882         if (fiemap_count_to_size(fiemap->fm_extent_count) > *vallen)
1883                 fiemap->fm_extent_count = fiemap_size_to_count(*vallen);
1884         if (fiemap->fm_extent_count == 0) {
1885                 get_num_extents = 1;
1886                 count_local = 0;
1887         }
1888         /* Check each stripe */
1889         for (cur_stripe = start_stripe, i = 0; i < stripe_count;
1890              i++, cur_stripe = (cur_stripe + 1) % lsm->lsm_stripe_count) {
1891                 obd_size req_fm_len; /* Stores length of required mapping */
1892                 obd_size len_mapped_single_call;
1893                 obd_off lun_start, lun_end, obd_object_end;
1894                 unsigned int ext_count;
1895
1896                 cur_stripe_wrap = cur_stripe;
1897
1898                 /* Find out range of mapping on this stripe */
1899                 if ((lov_stripe_intersects(lsm, cur_stripe, fm_start, fm_end,
1900                                            &lun_start, &obd_object_end)) == 0)
1901                         continue;
1902
1903                 if (lov_oinfo_is_dummy(lsm->lsm_oinfo[cur_stripe]))
1904                         GOTO(out, rc = -EIO);
1905
1906                 /* If this is a continuation FIEMAP call and we are on
1907                  * starting stripe then lun_start needs to be set to
1908                  * fm_end_offset */
1909                 if (fm_end_offset != 0 && cur_stripe == start_stripe)
1910                         lun_start = fm_end_offset;
1911
1912                 if (fm_length != ~0ULL) {
1913                         /* Handle fm_start + fm_length overflow */
1914                         if (fm_start + fm_length < fm_start)
1915                                 fm_length = ~0ULL - fm_start;
1916                         lun_end = lov_size_to_stripe(lsm, fm_start + fm_length,
1917                                                      cur_stripe);
1918                 } else {
1919                         lun_end = ~0ULL;
1920                 }
1921
1922                 if (lun_start == lun_end)
1923                         continue;
1924
1925                 req_fm_len = obd_object_end - lun_start;
1926                 fm_local->fm_length = 0;
1927                 len_mapped_single_call = 0;
1928
1929                 /* If the output buffer is very large and the objects have many
1930                  * extents we may need to loop on a single OST repeatedly */
1931                 ost_eof = 0;
1932                 ost_done = 0;
1933                 do {
1934                         if (get_num_extents == 0) {
1935                                 /* Don't get too many extents. */
1936                                 if (current_extent + count_local >
1937                                     fiemap->fm_extent_count)
1938                                         count_local = fiemap->fm_extent_count -
1939                                                                  current_extent;
1940                         }
1941
1942                         lun_start += len_mapped_single_call;
1943                         fm_local->fm_length = req_fm_len - len_mapped_single_call;
1944                         req_fm_len = fm_local->fm_length;
1945                         fm_local->fm_extent_count = count_local;
1946                         fm_local->fm_mapped_extents = 0;
1947                         fm_local->fm_flags = fiemap->fm_flags;
1948
1949                         fm_key->oa.o_oi = lsm->lsm_oinfo[cur_stripe]->loi_oi;
1950                         ost_index = lsm->lsm_oinfo[cur_stripe]->loi_ost_idx;
1951
1952                         if (ost_index < 0 || ost_index >=lov->desc.ld_tgt_count)
1953                                 GOTO(out, rc = -EINVAL);
1954
1955                         /* If OST is inactive, return extent with UNKNOWN flag */
1956                         if (!lov->lov_tgts[ost_index]->ltd_active) {
1957                                 fm_local->fm_flags |= FIEMAP_EXTENT_LAST;
1958                                 fm_local->fm_mapped_extents = 1;
1959
1960                                 lcl_fm_ext[0].fe_logical = lun_start;
1961                                 lcl_fm_ext[0].fe_length = obd_object_end -
1962                                                                       lun_start;
1963                                 lcl_fm_ext[0].fe_flags |= FIEMAP_EXTENT_UNKNOWN;
1964
1965                                 goto inactive_tgt;
1966                         }
1967
1968                         fm_local->fm_start = lun_start;
1969                         fm_local->fm_flags &= ~FIEMAP_FLAG_DEVICE_ORDER;
1970                         memcpy(&fm_key->fiemap, fm_local, sizeof(*fm_local));
1971                         *vallen=fiemap_count_to_size(fm_local->fm_extent_count);
1972                         rc = obd_get_info(NULL,
1973                                           lov->lov_tgts[ost_index]->ltd_exp,
1974                                           keylen, key, vallen, fm_local, lsm);
1975                         if (rc != 0)
1976                                 GOTO(out, rc);
1977
1978 inactive_tgt:
1979                         ext_count = fm_local->fm_mapped_extents;
1980                         if (ext_count == 0) {
1981                                 ost_done = 1;
1982                                 /* If last stripe has hole at the end,
1983                                  * then we need to return */
1984                                 if (cur_stripe_wrap == last_stripe) {
1985                                         fiemap->fm_mapped_extents = 0;
1986                                         goto finish;
1987                                 }
1988                                 break;
1989                         }
1990
1991                         /* If we just need num of extents then go to next device */
1992                         if (get_num_extents) {
1993                                 current_extent += ext_count;
1994                                 break;
1995                         }
1996
1997                         len_mapped_single_call = lcl_fm_ext[ext_count-1].fe_logical -
1998                                   lun_start + lcl_fm_ext[ext_count - 1].fe_length;
1999
2000                         /* Have we finished mapping on this device? */
2001                         if (req_fm_len <= len_mapped_single_call)
2002                                 ost_done = 1;
2003
2004                         /* Clear the EXTENT_LAST flag which can be present on
2005                          * last extent */
2006                         if (lcl_fm_ext[ext_count-1].fe_flags & FIEMAP_EXTENT_LAST)
2007                                 lcl_fm_ext[ext_count - 1].fe_flags &=
2008                                                             ~FIEMAP_EXTENT_LAST;
2009
2010                         curr_loc = lov_stripe_size(lsm,
2011                                            lcl_fm_ext[ext_count - 1].fe_logical+
2012                                            lcl_fm_ext[ext_count - 1].fe_length,
2013                                            cur_stripe);
2014                         if (curr_loc >= fm_key->oa.o_size)
2015                                 ost_eof = 1;
2016
2017                         fiemap_prepare_and_copy_exts(fiemap, lcl_fm_ext,
2018                                                      ost_index, ext_count,
2019                                                      current_extent);
2020
2021                         current_extent += ext_count;
2022
2023                         /* Ran out of available extents? */
2024                         if (current_extent >= fiemap->fm_extent_count)
2025                                 goto finish;
2026                 } while (ost_done == 0 && ost_eof == 0);
2027
2028                 if (cur_stripe_wrap == last_stripe)
2029                         goto finish;
2030         }
2031
2032 finish:
2033         /* Indicate that we are returning device offsets unless file just has
2034          * single stripe */
2035         if (lsm->lsm_stripe_count > 1)
2036                 fiemap->fm_flags |= FIEMAP_FLAG_DEVICE_ORDER;
2037
2038         if (get_num_extents)
2039                 goto skip_last_device_calc;
2040
2041         /* Check if we have reached the last stripe and whether mapping for that
2042          * stripe is done. */
2043         if (cur_stripe_wrap == last_stripe) {
2044                 if (ost_done || ost_eof)
2045                         fiemap->fm_extents[current_extent - 1].fe_flags |=
2046                                                              FIEMAP_EXTENT_LAST;
2047         }
2048
2049 skip_last_device_calc:
2050         fiemap->fm_mapped_extents = current_extent;
2051
2052 out:
2053         if (fm_local)
2054                 OBD_FREE_LARGE(fm_local, buffer_size);
2055         return rc;
2056 }
2057
2058 static int lov_get_info(const struct lu_env *env, struct obd_export *exp,
2059                         __u32 keylen, void *key, __u32 *vallen, void *val,
2060                         struct lov_stripe_md *lsm)
2061 {
2062         struct obd_device *obddev = class_exp2obd(exp);
2063         struct lov_obd *lov = &obddev->u.lov;
2064         int i, rc;
2065         ENTRY;
2066
2067         if (!vallen || !val)
2068                 RETURN(-EFAULT);
2069
2070         obd_getref(obddev);
2071
2072         if (KEY_IS(KEY_LOCK_TO_STRIPE)) {
2073                 struct {
2074                         char name[16];
2075                         struct ldlm_lock *lock;
2076                 } *data = key;
2077                 struct ldlm_res_id *res_id = &data->lock->l_resource->lr_name;
2078                 struct lov_oinfo *loi;
2079                 __u32 *stripe = val;
2080
2081                 if (*vallen < sizeof(*stripe))
2082                         GOTO(out, rc = -EFAULT);
2083                 *vallen = sizeof(*stripe);
2084
2085                 /* XXX This is another one of those bits that will need to
2086                  * change if we ever actually support nested LOVs.  It uses
2087                  * the lock's export to find out which stripe it is. */
2088                 /* XXX - it's assumed all the locks for deleted OSTs have
2089                  * been cancelled. Also, the export for deleted OSTs will
2090                  * be NULL and won't match the lock's export. */
2091                 for (i = 0; i < lsm->lsm_stripe_count; i++) {
2092                         loi = lsm->lsm_oinfo[i];
2093                         if (lov_oinfo_is_dummy(loi))
2094                                 continue;
2095
2096                         if (!lov->lov_tgts[loi->loi_ost_idx])
2097                                 continue;
2098                         if (lov->lov_tgts[loi->loi_ost_idx]->ltd_exp ==
2099                             data->lock->l_conn_export &&
2100                             ostid_res_name_eq(&loi->loi_oi, res_id)) {
2101                                 *stripe = i;
2102                                 GOTO(out, rc = 0);
2103                         }
2104                 }
2105                 LDLM_ERROR(data->lock, "lock on inode without such object");
2106                 dump_lsm(D_ERROR, lsm);
2107                 GOTO(out, rc = -ENXIO);
2108         } else if (KEY_IS(KEY_LAST_ID)) {
2109                 struct obd_id_info *info = val;
2110                 __u32 size = sizeof(obd_id);
2111                 struct lov_tgt_desc *tgt;
2112
2113                 LASSERT(*vallen == sizeof(struct obd_id_info));
2114                 tgt = lov->lov_tgts[info->idx];
2115
2116                 if (!tgt || !tgt->ltd_active)
2117                         GOTO(out, rc = -ESRCH);
2118
2119                 rc = obd_get_info(env, tgt->ltd_exp, keylen, key,
2120                                   &size, info->data, NULL);
2121                 GOTO(out, rc = 0);
2122         } else if (KEY_IS(KEY_LOVDESC)) {
2123                 struct lov_desc *desc_ret = val;
2124                 *desc_ret = lov->desc;
2125
2126                 GOTO(out, rc = 0);
2127         } else if (KEY_IS(KEY_FIEMAP)) {
2128                 rc = lov_fiemap(lov, keylen, key, vallen, val, lsm);
2129                 GOTO(out, rc);
2130         } else if (KEY_IS(KEY_CONNECT_FLAG)) {
2131                 struct lov_tgt_desc *tgt;
2132                 __u64 ost_idx = *((__u64*)val);
2133
2134                 LASSERT(*vallen == sizeof(__u64));
2135                 LASSERT(ost_idx < lov->desc.ld_tgt_count);
2136                 tgt = lov->lov_tgts[ost_idx];
2137
2138                 if (!tgt || !tgt->ltd_exp)
2139                         GOTO(out, rc = -ESRCH);
2140
2141                 *((__u64 *)val) = exp_connect_flags(tgt->ltd_exp);
2142                 GOTO(out, rc = 0);
2143         } else if (KEY_IS(KEY_TGT_COUNT)) {
2144                 *((int *)val) = lov->desc.ld_tgt_count;
2145                 GOTO(out, rc = 0);
2146         }
2147
2148         rc = -EINVAL;
2149
2150 out:
2151         obd_putref(obddev);
2152         RETURN(rc);
2153 }
2154
2155 static int lov_set_info_async(const struct lu_env *env, struct obd_export *exp,
2156                               obd_count keylen, void *key, obd_count vallen,
2157                               void *val, struct ptlrpc_request_set *set)
2158 {
2159         struct obd_device *obddev = class_exp2obd(exp);
2160         struct lov_obd *lov = &obddev->u.lov;
2161         obd_count count;
2162         int i, rc = 0, err;
2163         struct lov_tgt_desc *tgt;
2164         unsigned incr, check_uuid,
2165                  do_inactive, no_set;
2166         unsigned next_id = 0,  mds_con = 0, capa = 0;
2167         ENTRY;
2168
2169         incr = check_uuid = do_inactive = no_set = 0;
2170         if (set == NULL) {
2171                 no_set = 1;
2172                 set = ptlrpc_prep_set();
2173                 if (!set)
2174                         RETURN(-ENOMEM);
2175         }
2176
2177         obd_getref(obddev);
2178         count = lov->desc.ld_tgt_count;
2179
2180         if (KEY_IS(KEY_NEXT_ID)) {
2181                 count = vallen / sizeof(struct obd_id_info);
2182                 vallen = sizeof(obd_id);
2183                 incr = sizeof(struct obd_id_info);
2184                 do_inactive = 1;
2185                 next_id = 1;
2186         } else if (KEY_IS(KEY_CHECKSUM)) {
2187                 do_inactive = 1;
2188         } else if (KEY_IS(KEY_EVICT_BY_NID)) {
2189                 /* use defaults:  do_inactive = incr = 0; */
2190         } else if (KEY_IS(KEY_MDS_CONN)) {
2191                 mds_con = 1;
2192         } else if (KEY_IS(KEY_CAPA_KEY)) {
2193                 capa = 1;
2194         } else if (KEY_IS(KEY_CACHE_SET)) {
2195                 LASSERT(lov->lov_cache == NULL);
2196                 lov->lov_cache = val;
2197                 do_inactive = 1;
2198         }
2199
2200         for (i = 0; i < count; i++, val = (char *)val + incr) {
2201                 if (next_id) {
2202                         tgt = lov->lov_tgts[((struct obd_id_info*)val)->idx];
2203                 } else {
2204                         tgt = lov->lov_tgts[i];
2205                 }
2206                 /* OST was disconnected */
2207                 if (!tgt || !tgt->ltd_exp)
2208                         continue;
2209
2210                 /* OST is inactive and we don't want inactive OSCs */
2211                 if (!tgt->ltd_active && !do_inactive)
2212                         continue;
2213
2214                 if (mds_con) {
2215                         struct mds_group_info *mgi;
2216
2217                         LASSERT(vallen == sizeof(*mgi));
2218                         mgi = (struct mds_group_info *)val;
2219
2220                         /* Only want a specific OSC */
2221                         if (mgi->uuid && !obd_uuid_equals(mgi->uuid,
2222                                                 &tgt->ltd_uuid))
2223                                 continue;
2224
2225                         err = obd_set_info_async(env, tgt->ltd_exp,
2226                                          keylen, key, sizeof(int),
2227                                          &mgi->group, set);
2228                 } else if (next_id) {
2229                         err = obd_set_info_async(env, tgt->ltd_exp,
2230                                          keylen, key, vallen,
2231                                          ((struct obd_id_info*)val)->data, set);
2232                 } else if (capa) {
2233                         struct mds_capa_info *info = (struct mds_capa_info*)val;
2234
2235                         LASSERT(vallen == sizeof(*info));
2236
2237                          /* Only want a specific OSC */
2238                         if (info->uuid &&
2239                             !obd_uuid_equals(info->uuid, &tgt->ltd_uuid))
2240                                 continue;
2241
2242                         err = obd_set_info_async(env, tgt->ltd_exp, keylen,
2243                                                  key, sizeof(*info->capa),
2244                                                  info->capa, set);
2245                 } else {
2246                         /* Only want a specific OSC */
2247                         if (check_uuid &&
2248                             !obd_uuid_equals(val, &tgt->ltd_uuid))
2249                                 continue;
2250
2251                         err = obd_set_info_async(env, tgt->ltd_exp,
2252                                          keylen, key, vallen, val, set);
2253                 }
2254
2255                 if (!rc)
2256                         rc = err;
2257         }
2258
2259         obd_putref(obddev);
2260         if (no_set) {
2261                 err = ptlrpc_set_wait(set);
2262                 if (!rc)
2263                         rc = err;
2264                 ptlrpc_set_destroy(set);
2265         }
2266         RETURN(rc);
2267 }
2268
2269 void lov_stripe_lock(struct lov_stripe_md *md)
2270 {
2271         LASSERT(md->lsm_lock_owner != current_pid());
2272         spin_lock(&md->lsm_lock);
2273         LASSERT(md->lsm_lock_owner == 0);
2274         md->lsm_lock_owner = current_pid();
2275 }
2276
2277 void lov_stripe_unlock(struct lov_stripe_md *md)
2278 {
2279         LASSERT(md->lsm_lock_owner == current_pid());
2280         md->lsm_lock_owner = 0;
2281         spin_unlock(&md->lsm_lock);
2282 }
2283
2284 static int lov_quotactl(struct obd_device *obd, struct obd_export *exp,
2285                         struct obd_quotactl *oqctl)
2286 {
2287         struct lov_obd      *lov = &obd->u.lov;
2288         struct lov_tgt_desc *tgt;
2289         __u64                curspace = 0;
2290         __u64                bhardlimit = 0;
2291         int                  i, rc = 0;
2292         ENTRY;
2293
2294         if (oqctl->qc_cmd != LUSTRE_Q_QUOTAON &&
2295             oqctl->qc_cmd != LUSTRE_Q_QUOTAOFF &&
2296             oqctl->qc_cmd != Q_GETOQUOTA &&
2297             oqctl->qc_cmd != Q_INITQUOTA &&
2298             oqctl->qc_cmd != LUSTRE_Q_SETQUOTA &&
2299             oqctl->qc_cmd != Q_FINVALIDATE) {
2300                 CERROR("bad quota opc %x for lov obd", oqctl->qc_cmd);
2301                 RETURN(-EFAULT);
2302         }
2303
2304         /* for lov tgt */
2305         obd_getref(obd);
2306         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2307                 int err;
2308
2309                 tgt = lov->lov_tgts[i];
2310
2311                 if (!tgt)
2312                         continue;
2313
2314                 if (!tgt->ltd_active || tgt->ltd_reap) {
2315                         if (oqctl->qc_cmd == Q_GETOQUOTA &&
2316                             lov->lov_tgts[i]->ltd_activate) {
2317                                 rc = -ENETDOWN;
2318                                 CERROR("ost %d is inactive\n", i);
2319                         } else {
2320                                 CDEBUG(D_HA, "ost %d is inactive\n", i);
2321                         }
2322                         continue;
2323                 }
2324
2325                 err = obd_quotactl(tgt->ltd_exp, oqctl);
2326                 if (err) {
2327                         if (tgt->ltd_active && !rc)
2328                                 rc = err;
2329                         continue;
2330                 }
2331
2332                 if (oqctl->qc_cmd == Q_GETOQUOTA) {
2333                         curspace += oqctl->qc_dqblk.dqb_curspace;
2334                         bhardlimit += oqctl->qc_dqblk.dqb_bhardlimit;
2335                 }
2336         }
2337         obd_putref(obd);
2338
2339         if (oqctl->qc_cmd == Q_GETOQUOTA) {
2340                 oqctl->qc_dqblk.dqb_curspace = curspace;
2341                 oqctl->qc_dqblk.dqb_bhardlimit = bhardlimit;
2342         }
2343         RETURN(rc);
2344 }
2345
2346 static int lov_quotacheck(struct obd_device *obd, struct obd_export *exp,
2347                           struct obd_quotactl *oqctl)
2348 {
2349         struct lov_obd *lov = &obd->u.lov;
2350         int             i, rc = 0;
2351         ENTRY;
2352
2353         obd_getref(obd);
2354
2355         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2356                 if (!lov->lov_tgts[i])
2357                         continue;
2358
2359                 /* Skip quota check on the administratively disabled OSTs. */
2360                 if (!lov->lov_tgts[i]->ltd_activate) {
2361                         CWARN("lov idx %d was administratively disabled, "
2362                               "skip quotacheck on it.\n", i);
2363                         continue;
2364                 }
2365
2366                 if (!lov->lov_tgts[i]->ltd_active) {
2367                         CERROR("lov idx %d inactive\n", i);
2368                         rc = -EIO;
2369                         goto out;
2370                 }
2371         }
2372
2373         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2374                 int err;
2375
2376                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_activate)
2377                         continue;
2378
2379                 err = obd_quotacheck(lov->lov_tgts[i]->ltd_exp, oqctl);
2380                 if (err && !rc)
2381                         rc = err;
2382         }
2383
2384 out:
2385         obd_putref(obd);
2386
2387         RETURN(rc);
2388 }
2389
2390 static struct obd_ops lov_obd_ops = {
2391         .o_owner                = THIS_MODULE,
2392         .o_setup                = lov_setup,
2393         .o_precleanup           = lov_precleanup,
2394         .o_cleanup              = lov_cleanup,
2395         .o_connect              = lov_connect,
2396         .o_disconnect           = lov_disconnect,
2397         .o_statfs               = lov_statfs,
2398         .o_statfs_async         = lov_statfs_async,
2399         .o_packmd               = lov_packmd,
2400         .o_unpackmd             = lov_unpackmd,
2401         .o_create               = lov_create,
2402         .o_destroy              = lov_destroy,
2403         .o_getattr_async        = lov_getattr_async,
2404         .o_setattr_async        = lov_setattr_async,
2405         .o_change_cbdata        = lov_change_cbdata,
2406         .o_find_cbdata          = lov_find_cbdata,
2407         .o_iocontrol            = lov_iocontrol,
2408         .o_get_info             = lov_get_info,
2409         .o_set_info_async       = lov_set_info_async,
2410         .o_notify               = lov_notify,
2411         .o_pool_new             = lov_pool_new,
2412         .o_pool_rem             = lov_pool_remove,
2413         .o_pool_add             = lov_pool_add,
2414         .o_pool_del             = lov_pool_del,
2415         .o_getref               = lov_getref,
2416         .o_putref               = lov_putref,
2417         .o_quotactl             = lov_quotactl,
2418         .o_quotacheck           = lov_quotacheck,
2419 };
2420
2421 struct kmem_cache *lov_oinfo_slab;
2422
2423 extern struct lu_kmem_descr lov_caches[];
2424
2425 int __init lov_init(void)
2426 {
2427         bool enable_proc = true;
2428         struct obd_type *type;
2429         int rc;
2430         ENTRY;
2431
2432         /* print an address of _any_ initialized kernel symbol from this
2433          * module, to allow debugging with gdb that doesn't support data
2434          * symbols from modules.*/
2435         CDEBUG(D_INFO, "Lustre LOV module (%p).\n", &lov_caches);
2436
2437         rc = lu_kmem_init(lov_caches);
2438         if (rc)
2439                 return rc;
2440
2441         lov_oinfo_slab = kmem_cache_create("lov_oinfo",
2442                                            sizeof(struct lov_oinfo), 0,
2443                                            SLAB_HWCACHE_ALIGN, NULL);
2444         if (lov_oinfo_slab == NULL) {
2445                 lu_kmem_fini(lov_caches);
2446                 return -ENOMEM;
2447         }
2448
2449         type = class_search_type(LUSTRE_LOD_NAME);
2450         if (type != NULL && type->typ_procsym != NULL)
2451                 enable_proc = false;
2452
2453         rc = class_register_type(&lov_obd_ops, NULL, enable_proc, NULL,
2454 #ifndef HAVE_ONLY_PROCFS_SEQ
2455                                  NULL,
2456 #endif
2457                                  LUSTRE_LOV_NAME, &lov_device_type);
2458
2459         if (rc) {
2460                 kmem_cache_destroy(lov_oinfo_slab);
2461                 lu_kmem_fini(lov_caches);
2462         }
2463
2464         RETURN(rc);
2465 }
2466
2467 #ifdef __KERNEL__
2468 static void /*__exit*/ lov_exit(void)
2469 {
2470         class_unregister_type(LUSTRE_LOV_NAME);
2471         kmem_cache_destroy(lov_oinfo_slab);
2472         lu_kmem_fini(lov_caches);
2473 }
2474
2475 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2476 MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver");
2477 MODULE_LICENSE("GPL");
2478
2479 cfs_module(lov, LUSTRE_VERSION_STRING, lov_init, lov_exit);
2480 #endif