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, 2015, 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 #include <libcfs/libcfs.h>
  46
  47 #include <lustre/lustre_idl.h>
  48
  49 #include <cl_object.h>
  50 #include <lustre_dlm.h>
  51 #include <lustre_fid.h>
  52 #include <lustre_ioctl.h>
  53 #include <lustre_lib.h>
  54 #include <lustre_mds.h>
  55 #include <lustre_net.h>
  56 #include <lustre_param.h>
  57 #include <lustre_swab.h>
  58 #include <lprocfs_status.h>
  59 #include <obd_class.h>
  60 #include <obd_support.h>
  61
  62 #include "lov_internal.h"
  63
  64 /* Keep a refcount of lov->tgt usage to prevent racing with addition/deletion.
  65    Any function that expects lov_tgts to remain stationary must take a ref. */
  66 static void lov_getref(struct obd_device *obd)
  67 {
  68         struct lov_obd *lov = &obd->u.lov;
  69
  70         /* nobody gets through here until lov_putref is done */
  71         mutex_lock(&lov->lov_lock);
  72         atomic_inc(&lov->lov_refcount);
  73         mutex_unlock(&lov->lov_lock);
  74         return;
  75 }
  76
  77 static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt);
  78
  79 static void lov_putref(struct obd_device *obd)
  80 {
  81         struct lov_obd *lov = &obd->u.lov;
  82
  83         mutex_lock(&lov->lov_lock);
  84         /* ok to dec to 0 more than once -- ltd_exp's will be null */
  85         if (atomic_dec_and_test(&lov->lov_refcount) && lov->lov_death_row) {
  86                 struct list_head kill = LIST_HEAD_INIT(kill);
  87                 struct lov_tgt_desc *tgt, *n;
  88                 int i;
  89
  90                 CDEBUG(D_CONFIG, "destroying %d lov targets\n",
  91                        lov->lov_death_row);
  92                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
  93                         tgt = lov->lov_tgts[i];
  94
  95                         if (!tgt || !tgt->ltd_reap)
  96                                 continue;
  97                         list_add(&tgt->ltd_kill, &kill);
  98                         /* XXX - right now there is a dependency on ld_tgt_count
  99                          * being the maximum tgt index for computing the
 100                          * mds_max_easize. So we can't shrink it. */
 101                         lov_ost_pool_remove(&lov->lov_packed, i);
 102                         lov->lov_tgts[i] = NULL;
 103                         lov->lov_death_row--;
 104                 }
 105                 mutex_unlock(&lov->lov_lock);
 106
 107                 list_for_each_entry_safe(tgt, n, &kill, ltd_kill) {
 108                         list_del(&tgt->ltd_kill);
 109                         /* Disconnect */
 110                         __lov_del_obd(obd, tgt);
 111                 }
 112         } else {
 113                 mutex_unlock(&lov->lov_lock);
 114         }
 115 }
 116
 117 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
 118                               enum obd_notify_event ev);
 119 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
 120                       enum obd_notify_event ev, void *data);
 121
 122 int lov_connect_obd(struct obd_device *obd, __u32 index, int activate,
 123                     struct obd_connect_data *data)
 124 {
 125         struct lov_obd *lov = &obd->u.lov;
 126         struct obd_uuid *tgt_uuid;
 127         struct obd_device *tgt_obd;
 128         static struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
 129         struct obd_import *imp;
 130         int rc;
 131         ENTRY;
 132
 133         if (lov->lov_tgts[index] == NULL)
 134                 RETURN(-EINVAL);
 135
 136         tgt_uuid = &lov->lov_tgts[index]->ltd_uuid;
 137         tgt_obd = lov->lov_tgts[index]->ltd_obd;
 138
 139         if (!tgt_obd->obd_set_up) {
 140                 CERROR("Target %s not set up\n", obd_uuid2str(tgt_uuid));
 141                 RETURN(-EINVAL);
 142         }
 143
 144         /* override the sp_me from lov */
 145         tgt_obd->u.cli.cl_sp_me = lov->lov_sp_me;
 146
 147         if (data && (data->ocd_connect_flags & OBD_CONNECT_INDEX))
 148                 data->ocd_index = index;
 149
 150         /*
 151          * Divine LOV knows that OBDs under it are OSCs.
 152          */
 153         imp = tgt_obd->u.cli.cl_import;
 154
 155         if (activate) {
 156                 tgt_obd->obd_no_recov = 0;
 157                 /* FIXME this is probably supposed to be
 158                    ptlrpc_set_import_active.  Horrible naming. */
 159                 ptlrpc_activate_import(imp);
 160         }
 161
 162         rc = obd_register_observer(tgt_obd, obd);
 163         if (rc) {
 164                 CERROR("Target %s register_observer error %d\n",
 165                        obd_uuid2str(tgt_uuid), rc);
 166                 RETURN(rc);
 167         }
 168
 169
 170         if (imp->imp_invalid) {
 171                 CDEBUG(D_CONFIG, "not connecting OSC %s; administratively "
 172                        "disabled\n", obd_uuid2str(tgt_uuid));
 173                 RETURN(0);
 174         }
 175
 176         rc = obd_connect(NULL, &lov->lov_tgts[index]->ltd_exp, tgt_obd,
 177                          &lov_osc_uuid, data, NULL);
 178         if (rc || !lov->lov_tgts[index]->ltd_exp) {
 179                 CERROR("Target %s connect error %d\n",
 180                        obd_uuid2str(tgt_uuid), rc);
 181                 RETURN(-ENODEV);
 182         }
 183
 184         lov->lov_tgts[index]->ltd_reap = 0;
 185
 186         CDEBUG(D_CONFIG, "Connected tgt idx %d %s (%s) %sactive\n", index,
 187                obd_uuid2str(tgt_uuid), tgt_obd->obd_name, activate ? "":"in");
 188
 189         if (lov->targets_proc_entry != NULL) {
 190                 struct proc_dir_entry *osc_symlink;
 191                 struct obd_device *osc_obd;
 192
 193                 osc_obd = lov->lov_tgts[index]->ltd_exp->exp_obd;
 194
 195                 LASSERT(osc_obd != NULL);
 196                 LASSERT(osc_obd->obd_magic == OBD_DEVICE_MAGIC);
 197                 LASSERT(osc_obd->obd_type->typ_name != NULL);
 198
 199                 osc_symlink = lprocfs_add_symlink(osc_obd->obd_name,
 200                                                   lov->targets_proc_entry,
 201                                                   "../../../%s/%s",
 202                                                   osc_obd->obd_type->typ_name,
 203                                                   osc_obd->obd_name);
 204                 if (osc_symlink == NULL) {
 205                         CERROR("cannot register LOV target "
 206                                "/proc/fs/lustre/%s/%s/target_obds/%s\n",
 207                                obd->obd_type->typ_name, obd->obd_name,
 208                                osc_obd->obd_name);
 209                 }
 210         }
 211         RETURN(0);
 212 }
 213
 214 static int lov_connect(const struct lu_env *env,
 215                        struct obd_export **exp, struct obd_device *obd,
 216                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 217                        void *localdata)
 218 {
 219         struct lov_obd *lov = &obd->u.lov;
 220         struct lov_tgt_desc *tgt;
 221         struct lustre_handle conn;
 222         int i, rc;
 223         ENTRY;
 224
 225         CDEBUG(D_CONFIG, "connect #%d\n", lov->lov_connects);
 226
 227         rc = class_connect(&conn, obd, cluuid);
 228         if (rc)
 229                 RETURN(rc);
 230
 231         *exp = class_conn2export(&conn);
 232
 233         /* Why should there ever be more than 1 connect? */
 234         lov->lov_connects++;
 235         LASSERT(lov->lov_connects == 1);
 236
 237         memset(&lov->lov_ocd, 0, sizeof(lov->lov_ocd));
 238         if (data)
 239                 lov->lov_ocd = *data;
 240
 241         lov->targets_proc_entry = lprocfs_register("target_obds",
 242                                                    obd->obd_proc_entry,
 243                                                    NULL, NULL);
 244         if (IS_ERR(lov->targets_proc_entry)) {
 245                 CERROR("%s: cannot register "
 246                        "/proc/fs/lustre/%s/%s/target_obds\n",
 247                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
 248                 lov->targets_proc_entry = NULL;
 249         }
 250
 251         obd_getref(obd);
 252         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 253                 tgt = lov->lov_tgts[i];
 254                 if (!tgt || obd_uuid_empty(&tgt->ltd_uuid))
 255                         continue;
 256                 /* Flags will be lowest common denominator */
 257                 rc = lov_connect_obd(obd, i, tgt->ltd_activate, &lov->lov_ocd);
 258                 if (rc) {
 259                         CERROR("%s: lov connect tgt %d failed: %d\n",
 260                                obd->obd_name, i, rc);
 261                         continue;
 262                 }
 263                 /* connect to administrative disabled ost */
 264                 if (!lov->lov_tgts[i]->ltd_exp)
 265                         continue;
 266
 267                 rc = lov_notify(obd, lov->lov_tgts[i]->ltd_exp->exp_obd,
 268                                 OBD_NOTIFY_CONNECT, (void *)&i);
 269                 if (rc) {
 270                         CERROR("%s error sending notify %d\n",
 271                                obd->obd_name, rc);
 272                 }
 273         }
 274         obd_putref(obd);
 275
 276         RETURN(0);
 277 }
 278
 279 static int lov_disconnect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
 280 {
 281         struct lov_obd *lov = &obd->u.lov;
 282         struct obd_device *osc_obd;
 283         int rc;
 284         ENTRY;
 285
 286         osc_obd = class_exp2obd(tgt->ltd_exp);
 287         CDEBUG(D_CONFIG, "%s: disconnecting target %s\n",
 288                obd->obd_name, osc_obd->obd_name);
 289
 290         if (tgt->ltd_active) {
 291                 tgt->ltd_active = 0;
 292                 lov->desc.ld_active_tgt_count--;
 293                 tgt->ltd_exp->exp_obd->obd_inactive = 1;
 294         }
 295
 296         if (osc_obd) {
 297                 /* Pass it on to our clients.
 298                  * XXX This should be an argument to disconnect,
 299                  * XXX not a back-door flag on the OBD.  Ah well.
 300                  */
 301                 osc_obd->obd_force = obd->obd_force;
 302                 osc_obd->obd_fail = obd->obd_fail;
 303                 osc_obd->obd_no_recov = obd->obd_no_recov;
 304
 305                 if (lov->targets_proc_entry != NULL)
 306                         lprocfs_remove_proc_entry(osc_obd->obd_name,
 307                                                   lov->targets_proc_entry);
 308         }
 309
 310         obd_register_observer(osc_obd, NULL);
 311
 312         rc = obd_disconnect(tgt->ltd_exp);
 313         if (rc) {
 314                 CERROR("Target %s disconnect error %d\n",
 315                        tgt->ltd_uuid.uuid, rc);
 316                 rc = 0;
 317         }
 318
 319         tgt->ltd_exp = NULL;
 320         RETURN(0);
 321 }
 322
 323 static int lov_disconnect(struct obd_export *exp)
 324 {
 325         struct obd_device *obd = class_exp2obd(exp);
 326         struct lov_obd *lov = &obd->u.lov;
 327         int i, rc;
 328         ENTRY;
 329
 330         if (!lov->lov_tgts)
 331                 goto out;
 332
 333         /* Only disconnect the underlying layers on the final disconnect. */
 334         lov->lov_connects--;
 335         if (lov->lov_connects != 0) {
 336                 /* why should there be more than 1 connect? */
 337                 CERROR("disconnect #%d\n", lov->lov_connects);
 338                 goto out;
 339         }
 340
 341         /* Let's hold another reference so lov_del_obd doesn't spin through
 342            putref every time */
 343         obd_getref(obd);
 344
 345         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 346                 if (lov->lov_tgts[i] && lov->lov_tgts[i]->ltd_exp) {
 347                         /* Disconnection is the last we know about an obd */
 348                         lov_del_target(obd, i, NULL, lov->lov_tgts[i]->ltd_gen);
 349                 }
 350         }
 351         obd_putref(obd);
 352
 353         if (lov->targets_proc_entry != NULL)
 354                 lprocfs_remove(&lov->targets_proc_entry);
 355
 356 out:
 357         rc = class_disconnect(exp); /* bz 9811 */
 358         RETURN(rc);
 359 }
 360
 361 /* Error codes:
 362  *
 363  *  -EINVAL  : UUID can't be found in the LOV's target list
 364  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
 365  *  -EBADF   : The UUID is found, but the OBD is the wrong type (!)
 366  *  any >= 0 : is log target index
 367  */
 368 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
 369                               enum obd_notify_event ev)
 370 {
 371         struct lov_obd *lov = &obd->u.lov;
 372         struct lov_tgt_desc *tgt;
 373         int index, activate, active;
 374         ENTRY;
 375
 376         CDEBUG(D_INFO, "Searching in lov %p for uuid %s event(%d)\n",
 377                lov, uuid->uuid, ev);
 378
 379         obd_getref(obd);
 380         for (index = 0; index < lov->desc.ld_tgt_count; index++) {
 381                 tgt = lov->lov_tgts[index];
 382                 if (!tgt)
 383                         continue;
 384                 /*
 385                  * LU-642, initially inactive OSC could miss the obd_connect,
 386                  * we make up for it here.
 387                  */
 388                 if (ev == OBD_NOTIFY_ACTIVATE && tgt->ltd_exp == NULL &&
 389                     obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
 390                         struct obd_uuid lov_osc_uuid = {"LOV_OSC_UUID"};
 391
 392                         obd_connect(NULL, &tgt->ltd_exp, tgt->ltd_obd,
 393                                     &lov_osc_uuid, &lov->lov_ocd, NULL);
 394                 }
 395                 if (!tgt->ltd_exp)
 396                         continue;
 397
 398                 CDEBUG(D_INFO, "lov idx %d is %s conn "LPX64"\n",
 399                        index, obd_uuid2str(&tgt->ltd_uuid),
 400                        tgt->ltd_exp->exp_handle.h_cookie);
 401                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 402                         break;
 403         }
 404
 405         if (index == lov->desc.ld_tgt_count)
 406                 GOTO(out, index = -EINVAL);
 407
 408         if (ev == OBD_NOTIFY_DEACTIVATE || ev == OBD_NOTIFY_ACTIVATE) {
 409                 activate = (ev == OBD_NOTIFY_ACTIVATE) ? 1 : 0;
 410
 411                 if (lov->lov_tgts[index]->ltd_activate == activate) {
 412                         CDEBUG(D_INFO, "OSC %s already %sactivate!\n",
 413                                uuid->uuid, activate ? "" : "de");
 414                 } else {
 415                         lov->lov_tgts[index]->ltd_activate = activate;
 416                         CDEBUG(D_CONFIG, "%sactivate OSC %s\n",
 417                                activate ? "" : "de", obd_uuid2str(uuid));
 418                 }
 419
 420         } else if (ev == OBD_NOTIFY_INACTIVE || ev == OBD_NOTIFY_ACTIVE) {
 421                 active = (ev == OBD_NOTIFY_ACTIVE) ? 1 : 0;
 422
 423                 if (lov->lov_tgts[index]->ltd_active == active) {
 424                         CDEBUG(D_INFO, "OSC %s already %sactive!\n",
 425                                uuid->uuid, active ? "" : "in");
 426                         GOTO(out, index);
 427                 } else {
 428                         CDEBUG(D_CONFIG, "Marking OSC %s %sactive\n",
 429                                obd_uuid2str(uuid), active ? "" : "in");
 430                 }
 431
 432                 lov->lov_tgts[index]->ltd_active = active;
 433                 if (active) {
 434                         lov->desc.ld_active_tgt_count++;
 435                         lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 0;
 436                 } else {
 437                         lov->desc.ld_active_tgt_count--;
 438                         lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 1;
 439                 }
 440         } else {
 441                 CERROR("%s: unknown event %d for uuid %s\n", obd->obd_name,
 442                        ev, uuid->uuid);
 443         }
 444
 445  out:
 446         obd_putref(obd);
 447         RETURN(index);
 448 }
 449
 450 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
 451                       enum obd_notify_event ev, void *data)
 452 {
 453         int rc = 0;
 454         struct lov_obd *lov = &obd->u.lov;
 455         ENTRY;
 456
 457         down_read(&lov->lov_notify_lock);
 458         if (!lov->lov_connects) {
 459                 up_read(&lov->lov_notify_lock);
 460                 RETURN(rc);
 461         }
 462
 463         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE ||
 464             ev == OBD_NOTIFY_ACTIVATE || ev == OBD_NOTIFY_DEACTIVATE) {
 465                 struct obd_uuid *uuid;
 466
 467                 LASSERT(watched);
 468
 469                 if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) {
 470                         up_read(&lov->lov_notify_lock);
 471                         CERROR("unexpected notification of %s %s!\n",
 472                                watched->obd_type->typ_name,
 473                                watched->obd_name);
 474                         RETURN(-EINVAL);
 475                 }
 476                 uuid = &watched->u.cli.cl_target_uuid;
 477
 478                 /* Set OSC as active before notifying the observer, so the
 479                  * observer can use the OSC normally.
 480                  */
 481                 rc = lov_set_osc_active(obd, uuid, ev);
 482                 if (rc < 0) {
 483                         up_read(&lov->lov_notify_lock);
 484                         CERROR("event(%d) of %s failed: %d\n", ev,
 485                                obd_uuid2str(uuid), rc);
 486                         RETURN(rc);
 487                 }
 488                 /* active event should be pass lov target index as data */
 489                 data = &rc;
 490         }
 491
 492         /* Pass the notification up the chain. */
 493         if (watched) {
 494                 rc = obd_notify_observer(obd, watched, ev, data);
 495         } else {
 496                 /* NULL watched means all osc's in the lov (only for syncs) */
 497                 /* sync event should be send lov idx as data */
 498                 struct lov_obd *lov = &obd->u.lov;
 499                 int i, is_sync;
 500
 501                 data = &i;
 502                 is_sync = (ev == OBD_NOTIFY_SYNC) ||
 503                           (ev == OBD_NOTIFY_SYNC_NONBLOCK);
 504
 505                 obd_getref(obd);
 506                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 507                         if (!lov->lov_tgts[i])
 508                                 continue;
 509
 510                         /* don't send sync event if target not
 511                          * connected/activated */
 512                         if (is_sync &&  !lov->lov_tgts[i]->ltd_active)
 513                                 continue;
 514
 515                         rc = obd_notify_observer(obd, lov->lov_tgts[i]->ltd_obd,
 516                                                  ev, data);
 517                         if (rc) {
 518                                 CERROR("%s: notify %s of %s failed %d\n",
 519                                        obd->obd_name,
 520                                        obd->obd_observer->obd_name,
 521                                        lov->lov_tgts[i]->ltd_obd->obd_name,
 522                                        rc);
 523                         }
 524                 }
 525                 obd_putref(obd);
 526         }
 527
 528         up_read(&lov->lov_notify_lock);
 529         RETURN(rc);
 530 }
 531
 532 static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 533                           __u32 index, int gen, int active)
 534 {
 535         struct lov_obd *lov = &obd->u.lov;
 536         struct lov_tgt_desc *tgt;
 537         struct obd_device *tgt_obd;
 538         int rc;
 539         ENTRY;
 540
 541         CDEBUG(D_CONFIG, "uuid:%s idx:%d gen:%d active:%d\n",
 542                uuidp->uuid, index, gen, active);
 543
 544         if (gen <= 0) {
 545                 CERROR("request to add OBD %s with invalid generation: %d\n",
 546                        uuidp->uuid, gen);
 547                 RETURN(-EINVAL);
 548         }
 549
 550         tgt_obd = class_find_client_obd(uuidp, LUSTRE_OSC_NAME,
 551                                         &obd->obd_uuid);
 552         if (tgt_obd == NULL)
 553                 RETURN(-EINVAL);
 554
 555         mutex_lock(&lov->lov_lock);
 556
 557         if ((index < lov->lov_tgt_size) && (lov->lov_tgts[index] != NULL)) {
 558                 tgt = lov->lov_tgts[index];
 559                 CERROR("UUID %s already assigned at LOV target index %d\n",
 560                        obd_uuid2str(&tgt->ltd_uuid), index);
 561                 mutex_unlock(&lov->lov_lock);
 562                 RETURN(-EEXIST);
 563         }
 564
 565         if (index >= lov->lov_tgt_size) {
 566                 /* We need to reallocate the lov target array. */
 567                 struct lov_tgt_desc **newtgts, **old = NULL;
 568                 __u32 newsize, oldsize = 0;
 569
 570                 newsize = max(lov->lov_tgt_size, (__u32)2);
 571                 while (newsize < index + 1)
 572                         newsize = newsize << 1;
 573                 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
 574                 if (newtgts == NULL) {
 575                         mutex_unlock(&lov->lov_lock);
 576                         RETURN(-ENOMEM);
 577                 }
 578
 579                 if (lov->lov_tgt_size) {
 580                         memcpy(newtgts, lov->lov_tgts, sizeof(*newtgts) *
 581                                lov->lov_tgt_size);
 582                         old = lov->lov_tgts;
 583                         oldsize = lov->lov_tgt_size;
 584                 }
 585
 586                 lov->lov_tgts = newtgts;
 587                 lov->lov_tgt_size = newsize;
 588                 smp_rmb();
 589                 if (old)
 590                         OBD_FREE(old, sizeof(*old) * oldsize);
 591
 592                 CDEBUG(D_CONFIG, "tgts: %p size: %d\n",
 593                        lov->lov_tgts, lov->lov_tgt_size);
 594         }
 595
 596         OBD_ALLOC_PTR(tgt);
 597         if (!tgt) {
 598                 mutex_unlock(&lov->lov_lock);
 599                 RETURN(-ENOMEM);
 600         }
 601
 602         rc = lov_ost_pool_add(&lov->lov_packed, index, lov->lov_tgt_size);
 603         if (rc) {
 604                 mutex_unlock(&lov->lov_lock);
 605                 OBD_FREE_PTR(tgt);
 606                 RETURN(rc);
 607         }
 608
 609         tgt->ltd_uuid = *uuidp;
 610         tgt->ltd_obd = tgt_obd;
 611         /* XXX - add a sanity check on the generation number. */
 612         tgt->ltd_gen = gen;
 613         tgt->ltd_index = index;
 614         tgt->ltd_activate = active;
 615         lov->lov_tgts[index] = tgt;
 616         if (index >= lov->desc.ld_tgt_count)
 617                 lov->desc.ld_tgt_count = index + 1;
 618
 619         mutex_unlock(&lov->lov_lock);
 620
 621         CDEBUG(D_CONFIG, "idx=%d ltd_gen=%d ld_tgt_count=%d\n",
 622                 index, tgt->ltd_gen, lov->desc.ld_tgt_count);
 623
 624         rc = obd_notify(obd, tgt_obd, OBD_NOTIFY_CREATE, &index);
 625
 626         if (lov->lov_connects == 0) {
 627                 /* lov_connect hasn't been called yet. We'll do the
 628                    lov_connect_obd on this target when that fn first runs,
 629                    because we don't know the connect flags yet. */
 630                 RETURN(0);
 631         }
 632
 633         obd_getref(obd);
 634
 635         rc = lov_connect_obd(obd, index, active, &lov->lov_ocd);
 636         if (rc)
 637                 GOTO(out, rc);
 638
 639         /* connect to administrative disabled ost */
 640         if (!tgt->ltd_exp)
 641                 GOTO(out, rc = 0);
 642
 643         if (lov->lov_cache != NULL) {
 644                 rc = obd_set_info_async(NULL, tgt->ltd_exp,
 645                                 sizeof(KEY_CACHE_SET), KEY_CACHE_SET,
 646                                 sizeof(struct cl_client_cache), lov->lov_cache,
 647                                 NULL);
 648                 if (rc < 0)
 649                         GOTO(out, rc);
 650         }
 651
 652         rc = lov_notify(obd, tgt->ltd_exp->exp_obd,
 653                         active ? OBD_NOTIFY_CONNECT : OBD_NOTIFY_INACTIVE,
 654                         (void *)&index);
 655
 656 out:
 657         if (rc) {
 658                 CERROR("add failed (%d), deleting %s\n", rc,
 659                        obd_uuid2str(&tgt->ltd_uuid));
 660                 lov_del_target(obd, index, NULL, 0);
 661         }
 662         obd_putref(obd);
 663         RETURN(rc);
 664 }
 665
 666 /* Schedule a target for deletion */
 667 int lov_del_target(struct obd_device *obd, __u32 index,
 668                    struct obd_uuid *uuidp, int gen)
 669 {
 670         struct lov_obd *lov = &obd->u.lov;
 671         int count = lov->desc.ld_tgt_count;
 672         int rc = 0;
 673         ENTRY;
 674
 675         if (index >= count) {
 676                 CERROR("LOV target index %d >= number of LOV OBDs %d.\n",
 677                        index, count);
 678                 RETURN(-EINVAL);
 679         }
 680
 681         /* to make sure there's no ongoing lov_notify() now */
 682         down_write(&lov->lov_notify_lock);
 683         obd_getref(obd);
 684
 685         if (!lov->lov_tgts[index]) {
 686                 CERROR("LOV target at index %d is not setup.\n", index);
 687                 GOTO(out, rc = -EINVAL);
 688         }
 689
 690         if (uuidp && !obd_uuid_equals(uuidp, &lov->lov_tgts[index]->ltd_uuid)) {
 691                 CERROR("LOV target UUID %s at index %d doesn't match %s.\n",
 692                        lov_uuid2str(lov, index), index,
 693                        obd_uuid2str(uuidp));
 694                 GOTO(out, rc = -EINVAL);
 695         }
 696
 697         CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d exp: %p active: %d\n",
 698                lov_uuid2str(lov, index), index,
 699                lov->lov_tgts[index]->ltd_gen, lov->lov_tgts[index]->ltd_exp,
 700                lov->lov_tgts[index]->ltd_active);
 701
 702         lov->lov_tgts[index]->ltd_reap = 1;
 703         lov->lov_death_row++;
 704         /* we really delete it from obd_putref */
 705 out:
 706         obd_putref(obd);
 707         up_write(&lov->lov_notify_lock);
 708
 709         RETURN(rc);
 710 }
 711
 712 static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
 713 {
 714         struct obd_device *osc_obd;
 715
 716         LASSERT(tgt);
 717         LASSERT(tgt->ltd_reap);
 718
 719         osc_obd = class_exp2obd(tgt->ltd_exp);
 720
 721         CDEBUG(D_CONFIG, "Removing tgt %s : %s\n",
 722                tgt->ltd_uuid.uuid,
 723                osc_obd ? osc_obd->obd_name : "<no obd>");
 724
 725         if (tgt->ltd_exp)
 726                 lov_disconnect_obd(obd, tgt);
 727
 728         OBD_FREE_PTR(tgt);
 729
 730         /* Manual cleanup - no cleanup logs to clean up the osc's.  We must
 731            do it ourselves. And we can't do it from lov_cleanup,
 732            because we just lost our only reference to it. */
 733         if (osc_obd)
 734                 class_manual_cleanup(osc_obd);
 735 }
 736
 737 void lov_fix_desc_stripe_size(__u64 *val)
 738 {
 739         if (*val < LOV_MIN_STRIPE_SIZE) {
 740                 if (*val != 0)
 741                         LCONSOLE_INFO("Increasing default stripe size to "
 742                                       "minimum %u\n",
 743                                       LOV_DESC_STRIPE_SIZE_DEFAULT);
 744                 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
 745         } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
 746                 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
 747                 LCONSOLE_WARN("Changing default stripe size to "LPU64" (a "
 748                               "multiple of %u)\n",
 749                               *val, LOV_MIN_STRIPE_SIZE);
 750         }
 751 }
 752
 753 void lov_fix_desc_stripe_count(__u32 *val)
 754 {
 755         if (*val == 0)
 756                 *val = 1;
 757 }
 758
 759 void lov_fix_desc_pattern(__u32 *val)
 760 {
 761         /* from lov_setstripe */
 762         if ((*val != 0) && (*val != LOV_PATTERN_RAID0)) {
 763                 LCONSOLE_WARN("Unknown stripe pattern: %#x\n", *val);
 764                 *val = 0;
 765         }
 766 }
 767
 768 void lov_fix_desc_qos_maxage(__u32 *val)
 769 {
 770         if (*val == 0)
 771                 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
 772 }
 773
 774 void lov_fix_desc(struct lov_desc *desc)
 775 {
 776         lov_fix_desc_stripe_size(&desc->ld_default_stripe_size);
 777         lov_fix_desc_stripe_count(&desc->ld_default_stripe_count);
 778         lov_fix_desc_pattern(&desc->ld_pattern);
 779         lov_fix_desc_qos_maxage(&desc->ld_qos_maxage);
 780 }
 781
 782 int lov_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 783 {
 784         struct lov_desc *desc;
 785         struct lov_obd *lov = &obd->u.lov;
 786 #ifdef CONFIG_PROC_FS
 787         struct obd_type *type;
 788 #endif
 789         int rc;
 790         ENTRY;
 791
 792         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 793                 CERROR("LOV setup requires a descriptor\n");
 794                 RETURN(-EINVAL);
 795         }
 796
 797         desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
 798
 799         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
 800                 CERROR("descriptor size wrong: %d > %d\n",
 801                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
 802                 RETURN(-EINVAL);
 803         }
 804
 805         if (desc->ld_magic != LOV_DESC_MAGIC) {
 806                 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
 807                             CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
 808                                    obd->obd_name, desc);
 809                             lustre_swab_lov_desc(desc);
 810                 } else {
 811                         CERROR("%s: Bad lov desc magic: %#x\n",
 812                                obd->obd_name, desc->ld_magic);
 813                         RETURN(-EINVAL);
 814                 }
 815         }
 816
 817         lov_fix_desc(desc);
 818
 819         desc->ld_active_tgt_count = 0;
 820         lov->desc = *desc;
 821         lov->lov_tgt_size = 0;
 822
 823         mutex_init(&lov->lov_lock);
 824         atomic_set(&lov->lov_refcount, 0);
 825         lov->lov_sp_me = LUSTRE_SP_CLI;
 826
 827         init_rwsem(&lov->lov_notify_lock);
 828
 829         lov->lov_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,
 830                                                    HASH_POOLS_MAX_BITS,
 831                                                    HASH_POOLS_BKT_BITS, 0,
 832                                                    CFS_HASH_MIN_THETA,
 833                                                    CFS_HASH_MAX_THETA,
 834                                                    &pool_hash_operations,
 835                                                    CFS_HASH_DEFAULT);
 836         INIT_LIST_HEAD(&lov->lov_pool_list);
 837         lov->lov_pool_count = 0;
 838         rc = lov_ost_pool_init(&lov->lov_packed, 0);
 839         if (rc)
 840                 GOTO(out, rc);
 841
 842 #ifdef CONFIG_PROC_FS
 843         obd->obd_vars = lprocfs_lov_obd_vars;
 844         /* If this is true then both client (lov) and server
 845          * (lod) are on the same node. The lod layer if loaded
 846          * first will register the lov proc directory. In that
 847          * case obd->obd_type->typ_procroot will be not set.
 848          * Instead we use type->typ_procsym as the parent. */
 849         type = class_search_type(LUSTRE_LOD_NAME);
 850         if (type != NULL && type->typ_procsym != NULL) {
 851                 obd->obd_proc_entry = lprocfs_register(obd->obd_name,
 852                                                        type->typ_procsym,
 853                                                        obd->obd_vars, obd);
 854                 if (IS_ERR(obd->obd_proc_entry)) {
 855                         rc = PTR_ERR(obd->obd_proc_entry);
 856                         CERROR("error %d setting up lprocfs for %s\n", rc,
 857                                obd->obd_name);
 858                         obd->obd_proc_entry = NULL;
 859                 }
 860         } else {
 861                 rc = lprocfs_obd_setup(obd);
 862         }
 863
 864         if (rc == 0) {
 865                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd",
 866                                         0444, &lov_proc_target_fops, obd);
 867                 if (rc)
 868                         CWARN("Error adding the target_obd file\n");
 869
 870                 lov->lov_pool_proc_entry = lprocfs_register("pools",
 871                                                             obd->obd_proc_entry,
 872                                                             NULL, NULL);
 873                 if (IS_ERR(lov->lov_pool_proc_entry)) {
 874                         rc = PTR_ERR(lov->lov_pool_proc_entry);
 875                         CERROR("error %d setting up lprocfs for pools\n", rc);
 876                         lov->lov_pool_proc_entry = NULL;
 877                 }
 878         }
 879 #endif
 880         RETURN(0);
 881
 882 out:
 883         return rc;
 884 }
 885
 886 static int lov_cleanup(struct obd_device *obd)
 887 {
 888         struct lov_obd *lov = &obd->u.lov;
 889         struct list_head *pos, *tmp;
 890         struct pool_desc *pool;
 891         ENTRY;
 892
 893         list_for_each_safe(pos, tmp, &lov->lov_pool_list) {
 894                 pool = list_entry(pos, struct pool_desc, pool_list);
 895                 /* free pool structs */
 896                 CDEBUG(D_INFO, "delete pool %p\n", pool);
 897                 /* In the function below, .hs_keycmp resolves to
 898                  * pool_hashkey_keycmp() */
 899                 /* coverity[overrun-buffer-val] */
 900                 lov_pool_del(obd, pool->pool_name);
 901         }
 902         cfs_hash_putref(lov->lov_pools_hash_body);
 903         lov_ost_pool_free(&lov->lov_packed);
 904
 905         lprocfs_obd_cleanup(obd);
 906         if (lov->lov_tgts) {
 907                 int i;
 908                 obd_getref(obd);
 909                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 910                         if (!lov->lov_tgts[i])
 911                                 continue;
 912
 913                         /* Inactive targets may never have connected */
 914                         if (lov->lov_tgts[i]->ltd_active ||
 915                             atomic_read(&lov->lov_refcount))
 916                                 /* We should never get here - these
 917                                  * should have been removed in the
 918                                  * disconnect. */
 919                                 CERROR("%s: lov tgt %d not cleaned! "
 920                                        "deathrow=%d, lovrc=%d\n",
 921                                        obd->obd_name, i, lov->lov_death_row,
 922                                        atomic_read(&lov->lov_refcount));
 923                         lov_del_target(obd, i, NULL, 0);
 924                 }
 925                 obd_putref(obd);
 926                 OBD_FREE(lov->lov_tgts, sizeof(*lov->lov_tgts) *
 927                          lov->lov_tgt_size);
 928                 lov->lov_tgt_size = 0;
 929         }
 930
 931         if (lov->lov_cache != NULL) {
 932                 cl_cache_decref(lov->lov_cache);
 933                 lov->lov_cache = NULL;
 934         }
 935
 936         RETURN(0);
 937 }
 938
 939 int lov_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg,
 940                             __u32 *indexp, int *genp)
 941 {
 942         struct obd_uuid obd_uuid;
 943         int cmd;
 944         int rc = 0;
 945         ENTRY;
 946
 947         switch(cmd = lcfg->lcfg_command) {
 948         case LCFG_LOV_ADD_OBD:
 949         case LCFG_LOV_ADD_INA:
 950         case LCFG_LOV_DEL_OBD: {
 951                 __u32 index;
 952                 int gen;
 953                 /* lov_modify_tgts add  0:lov_mdsA  1:ost1_UUID  2:0  3:1 */
 954                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
 955                         GOTO(out, rc = -EINVAL);
 956
 957                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
 958
 959                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", indexp) != 1)
 960                         GOTO(out, rc = -EINVAL);
 961                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", genp) != 1)
 962                         GOTO(out, rc = -EINVAL);
 963                 index = *indexp;
 964                 gen = *genp;
 965                 if (cmd == LCFG_LOV_ADD_OBD)
 966                         rc = lov_add_target(obd, &obd_uuid, index, gen, 1);
 967                 else if (cmd == LCFG_LOV_ADD_INA)
 968                         rc = lov_add_target(obd, &obd_uuid, index, gen, 0);
 969                 else
 970                         rc = lov_del_target(obd, index, &obd_uuid, gen);
 971                 GOTO(out, rc);
 972         }
 973         case LCFG_PARAM: {
 974                 struct lov_desc *desc = &(obd->u.lov.desc);
 975
 976                 if (!desc)
 977                         GOTO(out, rc = -EINVAL);
 978
 979                 rc = class_process_proc_param(PARAM_LOV, obd->obd_vars,
 980                                               lcfg, obd);
 981                 if (rc > 0)
 982                         rc = 0;
 983                 GOTO(out, rc);
 984         }
 985         case LCFG_POOL_NEW:
 986         case LCFG_POOL_ADD:
 987         case LCFG_POOL_DEL:
 988         case LCFG_POOL_REM:
 989                 GOTO(out, rc);
 990
 991         default: {
 992                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
 993                 GOTO(out, rc = -EINVAL);
 994
 995         }
 996         }
 997 out:
 998         RETURN(rc);
 999 }
1000
1001 int lov_statfs_interpret(struct ptlrpc_request_set *rqset, void *data, int rc)
1002 {
1003         struct lov_request_set *lovset = (struct lov_request_set *)data;
1004         int err;
1005         ENTRY;
1006
1007         if (rc)
1008                 atomic_set(&lovset->set_completes, 0);
1009
1010         err = lov_fini_statfs_set(lovset);
1011         RETURN(rc ? rc : err);
1012 }
1013
1014 static int lov_statfs_async(struct obd_export *exp, struct obd_info *oinfo,
1015                             __u64 max_age, struct ptlrpc_request_set *rqset)
1016 {
1017         struct obd_device      *obd = class_exp2obd(exp);
1018         struct lov_request_set *set;
1019         struct lov_request *req;
1020         struct list_head *pos;
1021         struct lov_obd *lov;
1022         int rc = 0;
1023         ENTRY;
1024
1025         LASSERT(oinfo != NULL);
1026         LASSERT(oinfo->oi_osfs != NULL);
1027
1028         lov = &obd->u.lov;
1029         rc = lov_prep_statfs_set(obd, oinfo, &set);
1030         if (rc)
1031                 RETURN(rc);
1032
1033         list_for_each(pos, &set->set_list) {
1034                 req = list_entry(pos, struct lov_request, rq_link);
1035                 rc = obd_statfs_async(lov->lov_tgts[req->rq_idx]->ltd_exp,
1036                                       &req->rq_oi, max_age, rqset);
1037                 if (rc)
1038                         break;
1039         }
1040
1041         if (rc || list_empty(&rqset->set_requests)) {
1042                 int err;
1043                 if (rc)
1044                         atomic_set(&set->set_completes, 0);
1045                 err = lov_fini_statfs_set(set);
1046                 RETURN(rc ? rc : err);
1047         }
1048
1049         LASSERT(rqset->set_interpret == NULL);
1050         rqset->set_interpret = lov_statfs_interpret;
1051         rqset->set_arg = (void *)set;
1052         RETURN(0);
1053 }
1054
1055 static int lov_statfs(const struct lu_env *env, struct obd_export *exp,
1056                       struct obd_statfs *osfs, __u64 max_age, __u32 flags)
1057 {
1058         struct ptlrpc_request_set *set = NULL;
1059         struct obd_info oinfo = {
1060                 .oi_osfs = osfs,
1061                 .oi_flags = flags,
1062         };
1063         int rc = 0;
1064
1065         ENTRY;
1066
1067         /* for obdclass we forbid using obd_statfs_rqset, but prefer using async
1068          * statfs requests */
1069         set = ptlrpc_prep_set();
1070         if (set == NULL)
1071                 RETURN(-ENOMEM);
1072
1073         rc = lov_statfs_async(exp, &oinfo, max_age, set);
1074         if (rc == 0)
1075                 rc = ptlrpc_set_wait(set);
1076
1077         ptlrpc_set_destroy(set);
1078
1079         RETURN(rc);
1080 }
1081
1082 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1083                          void *karg, void __user *uarg)
1084 {
1085         struct obd_device *obddev = class_exp2obd(exp);
1086         struct lov_obd *lov = &obddev->u.lov;
1087         int i = 0, rc = 0, count = lov->desc.ld_tgt_count;
1088         struct obd_uuid *uuidp;
1089         ENTRY;
1090
1091         switch (cmd) {
1092         case IOC_OBD_STATFS: {
1093                 struct obd_ioctl_data *data = karg;
1094                 struct obd_device *osc_obd;
1095                 struct obd_statfs stat_buf = {0};
1096                 __u32 index;
1097                 __u32 flags;
1098
1099                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
1100                 if ((index >= count))
1101                         RETURN(-ENODEV);
1102
1103                 if (!lov->lov_tgts[index])
1104                         /* Try again with the next index */
1105                         RETURN(-EAGAIN);
1106                 if (!lov->lov_tgts[index]->ltd_active)
1107                         RETURN(-ENODATA);
1108
1109                 osc_obd = class_exp2obd(lov->lov_tgts[index]->ltd_exp);
1110                 if (!osc_obd)
1111                         RETURN(-EINVAL);
1112
1113                 /* copy UUID */
1114                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(osc_obd),
1115                                  min_t(unsigned long, data->ioc_plen2,
1116                                        sizeof(struct obd_uuid))))
1117                         RETURN(-EFAULT);
1118
1119                 flags = uarg ? *(__u32 __user *)uarg : 0;
1120                 /* got statfs data */
1121                 rc = obd_statfs(NULL, lov->lov_tgts[index]->ltd_exp, &stat_buf,
1122                                 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
1123                                 flags);
1124                 if (rc)
1125                         RETURN(rc);
1126                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
1127                                  min_t(unsigned long, data->ioc_plen1,
1128                                        sizeof(struct obd_statfs))))
1129                         RETURN(-EFAULT);
1130                 break;
1131         }
1132         case OBD_IOC_LOV_GET_CONFIG: {
1133                 struct obd_ioctl_data *data;
1134                 struct lov_desc *desc;
1135                 char *buf = NULL;
1136                 __u32 *genp;
1137
1138                 len = 0;
1139                 if (obd_ioctl_getdata(&buf, &len, uarg))
1140                         RETURN(-EINVAL);
1141
1142                 data = (struct obd_ioctl_data *)buf;
1143
1144                 if (sizeof(*desc) > data->ioc_inllen1) {
1145                         obd_ioctl_freedata(buf, len);
1146                         RETURN(-EINVAL);
1147                 }
1148
1149                 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
1150                         obd_ioctl_freedata(buf, len);
1151                         RETURN(-EINVAL);
1152                 }
1153
1154                 if (sizeof(__u32) * count > data->ioc_inllen3) {
1155                         obd_ioctl_freedata(buf, len);
1156                         RETURN(-EINVAL);
1157                 }
1158
1159                 desc = (struct lov_desc *)data->ioc_inlbuf1;
1160                 memcpy(desc, &(lov->desc), sizeof(*desc));
1161
1162                 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
1163                 genp = (__u32 *)data->ioc_inlbuf3;
1164                 /* the uuid will be empty for deleted OSTs */
1165                 for (i = 0; i < count; i++, uuidp++, genp++) {
1166                         if (!lov->lov_tgts[i])
1167                                 continue;
1168                         *uuidp = lov->lov_tgts[i]->ltd_uuid;
1169                         *genp = lov->lov_tgts[i]->ltd_gen;
1170                 }
1171
1172                 if (copy_to_user(uarg, buf, len))
1173                         rc = -EFAULT;
1174                 obd_ioctl_freedata(buf, len);
1175                 break;
1176         }
1177         case OBD_IOC_QUOTACTL: {
1178                 struct if_quotactl *qctl = karg;
1179                 struct lov_tgt_desc *tgt = NULL;
1180                 struct obd_quotactl *oqctl;
1181
1182                 if (qctl->qc_valid == QC_OSTIDX) {
1183                         if (count <= qctl->qc_idx)
1184                                 RETURN(-EINVAL);
1185
1186                         tgt = lov->lov_tgts[qctl->qc_idx];
1187                         if (!tgt || !tgt->ltd_exp)
1188                                 RETURN(-EINVAL);
1189                 } else if (qctl->qc_valid == QC_UUID) {
1190                         for (i = 0; i < count; i++) {
1191                                 tgt = lov->lov_tgts[i];
1192                                 if (!tgt ||
1193                                     !obd_uuid_equals(&tgt->ltd_uuid,
1194                                                      &qctl->obd_uuid))
1195                                         continue;
1196
1197                                 if (tgt->ltd_exp == NULL)
1198                                         RETURN(-EINVAL);
1199
1200                                 break;
1201                         }
1202                 } else {
1203                         RETURN(-EINVAL);
1204                 }
1205
1206                 if (i >= count)
1207                         RETURN(-EAGAIN);
1208
1209                 LASSERT(tgt && tgt->ltd_exp);
1210                 OBD_ALLOC_PTR(oqctl);
1211                 if (!oqctl)
1212                         RETURN(-ENOMEM);
1213
1214                 QCTL_COPY(oqctl, qctl);
1215                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
1216                 if (rc == 0) {
1217                         QCTL_COPY(qctl, oqctl);
1218                         qctl->qc_valid = QC_OSTIDX;
1219                         qctl->obd_uuid = tgt->ltd_uuid;
1220                 }
1221                 OBD_FREE_PTR(oqctl);
1222                 break;
1223         }
1224         default: {
1225                 int set = 0;
1226
1227                 if (count == 0)
1228                         RETURN(-ENOTTY);
1229
1230                 for (i = 0; i < count; i++) {
1231                         int err;
1232                         struct obd_device *osc_obd;
1233
1234                         /* OST was disconnected */
1235                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
1236                                 continue;
1237
1238                         /* ll_umount_begin() sets force flag but for lov, not
1239                          * osc. Let's pass it through */
1240                         osc_obd = class_exp2obd(lov->lov_tgts[i]->ltd_exp);
1241                         osc_obd->obd_force = obddev->obd_force;
1242                         err = obd_iocontrol(cmd, lov->lov_tgts[i]->ltd_exp,
1243                                             len, karg, uarg);
1244                         if (err) {
1245                                 if (lov->lov_tgts[i]->ltd_active) {
1246                                         CDEBUG(err == -ENOTTY ?
1247                                                D_IOCTL : D_WARNING,
1248                                                "iocontrol OSC %s on OST "
1249                                                "idx %d cmd %x: err = %d\n",
1250                                                lov_uuid2str(lov, i),
1251                                                i, cmd, err);
1252                                         if (!rc)
1253                                                 rc = err;
1254                                 }
1255                         } else {
1256                                 set = 1;
1257                         }
1258                 }
1259                 if (!set && !rc)
1260                         rc = -EIO;
1261         }
1262         }
1263
1264         RETURN(rc);
1265 }
1266
1267 static int lov_get_info(const struct lu_env *env, struct obd_export *exp,
1268                         __u32 keylen, void *key, __u32 *vallen, void *val)
1269 {
1270         struct obd_device *obddev = class_exp2obd(exp);
1271         struct lov_obd *lov = &obddev->u.lov;
1272         struct lov_desc *ld = &lov->desc;
1273         int rc = 0;
1274         ENTRY;
1275
1276         if (vallen == NULL || val == NULL)
1277                 RETURN(-EFAULT);
1278
1279         obd_getref(obddev);
1280
1281         if (KEY_IS(KEY_MAX_EASIZE)) {
1282                 u32 max_stripe_count = min_t(u32, ld->ld_active_tgt_count,
1283                                              LOV_MAX_STRIPE_COUNT);
1284
1285                 *((u32 *)val) = lov_mds_md_size(max_stripe_count, LOV_MAGIC_V3);
1286         } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
1287                 u32 def_stripe_count = min_t(u32, ld->ld_default_stripe_count,
1288                                              LOV_MAX_STRIPE_COUNT);
1289
1290                 *((u32 *)val) = lov_mds_md_size(def_stripe_count, LOV_MAGIC_V3);
1291         } else if (KEY_IS(KEY_TGT_COUNT)) {
1292                 *((int *)val) = lov->desc.ld_tgt_count;
1293         } else {
1294                 rc = -EINVAL;
1295         }
1296
1297         obd_putref(obddev);
1298
1299         RETURN(rc);
1300 }
1301
1302 static int lov_set_info_async(const struct lu_env *env, struct obd_export *exp,
1303                               __u32 keylen, void *key,
1304                               __u32 vallen, void *val,
1305                               struct ptlrpc_request_set *set)
1306 {
1307         struct obd_device *obddev = class_exp2obd(exp);
1308         struct lov_obd *lov = &obddev->u.lov;
1309         struct lov_tgt_desc *tgt;
1310         int do_inactive = 0;
1311         int no_set = 0;
1312         u32 count;
1313         u32 i;
1314         int rc = 0;
1315         int err;
1316         ENTRY;
1317
1318         if (set == NULL) {
1319                 no_set = 1;
1320                 set = ptlrpc_prep_set();
1321                 if (!set)
1322                         RETURN(-ENOMEM);
1323         }
1324
1325         obd_getref(obddev);
1326         count = lov->desc.ld_tgt_count;
1327
1328         if (KEY_IS(KEY_CHECKSUM)) {
1329                 do_inactive = 1;
1330         } else if (KEY_IS(KEY_CACHE_SET)) {
1331                 LASSERT(lov->lov_cache == NULL);
1332                 lov->lov_cache = val;
1333                 do_inactive = 1;
1334                 cl_cache_incref(lov->lov_cache);
1335         }
1336
1337         for (i = 0; i < count; i++) {
1338                 tgt = lov->lov_tgts[i];
1339
1340                 /* OST was disconnected */
1341                 if (!tgt || !tgt->ltd_exp)
1342                         continue;
1343
1344                 /* OST is inactive and we don't want inactive OSCs */
1345                 if (!tgt->ltd_active && !do_inactive)
1346                         continue;
1347
1348                 err = obd_set_info_async(env, tgt->ltd_exp, keylen, key,
1349                                          vallen, val, set);
1350                 if (!rc)
1351                         rc = err;
1352         }
1353
1354         obd_putref(obddev);
1355         if (no_set) {
1356                 err = ptlrpc_set_wait(set);
1357                 if (!rc)
1358                         rc = err;
1359                 ptlrpc_set_destroy(set);
1360         }
1361         RETURN(rc);
1362 }
1363
1364 void lov_stripe_lock(struct lov_stripe_md *md)
1365 __acquires(&md->lsm_lock)
1366 {
1367         LASSERT(md->lsm_lock_owner != current_pid());
1368         spin_lock(&md->lsm_lock);
1369         LASSERT(md->lsm_lock_owner == 0);
1370         md->lsm_lock_owner = current_pid();
1371 }
1372
1373 void lov_stripe_unlock(struct lov_stripe_md *md)
1374 __releases(&md->lsm_lock)
1375 {
1376         LASSERT(md->lsm_lock_owner == current_pid());
1377         md->lsm_lock_owner = 0;
1378         spin_unlock(&md->lsm_lock);
1379 }
1380
1381 static int lov_quotactl(struct obd_device *obd, struct obd_export *exp,
1382                         struct obd_quotactl *oqctl)
1383 {
1384         struct lov_obd      *lov = &obd->u.lov;
1385         struct lov_tgt_desc *tgt;
1386         __u64                curspace = 0;
1387         __u64                bhardlimit = 0;
1388         int                  i, rc = 0;
1389         ENTRY;
1390
1391         if (oqctl->qc_cmd != Q_GETOQUOTA &&
1392             oqctl->qc_cmd != LUSTRE_Q_SETQUOTA) {
1393                 CERROR("%s: bad quota opc %x for lov obd\n",
1394                        obd->obd_name, oqctl->qc_cmd);
1395                 RETURN(-EFAULT);
1396         }
1397
1398         /* for lov tgt */
1399         obd_getref(obd);
1400         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1401                 int err;
1402
1403                 tgt = lov->lov_tgts[i];
1404
1405                 if (!tgt)
1406                         continue;
1407
1408                 if (!tgt->ltd_active || tgt->ltd_reap) {
1409                         if (oqctl->qc_cmd == Q_GETOQUOTA &&
1410                             lov->lov_tgts[i]->ltd_activate) {
1411                                 rc = -ENETDOWN;
1412                                 CERROR("ost %d is inactive\n", i);
1413                         } else {
1414                                 CDEBUG(D_HA, "ost %d is inactive\n", i);
1415                         }
1416                         continue;
1417                 }
1418
1419                 err = obd_quotactl(tgt->ltd_exp, oqctl);
1420                 if (err) {
1421                         if (tgt->ltd_active && !rc)
1422                                 rc = err;
1423                         continue;
1424                 }
1425
1426                 if (oqctl->qc_cmd == Q_GETOQUOTA) {
1427                         curspace += oqctl->qc_dqblk.dqb_curspace;
1428                         bhardlimit += oqctl->qc_dqblk.dqb_bhardlimit;
1429                 }
1430         }
1431         obd_putref(obd);
1432
1433         if (oqctl->qc_cmd == Q_GETOQUOTA) {
1434                 oqctl->qc_dqblk.dqb_curspace = curspace;
1435                 oqctl->qc_dqblk.dqb_bhardlimit = bhardlimit;
1436         }
1437         RETURN(rc);
1438 }
1439
1440 static struct obd_ops lov_obd_ops = {
1441         .o_owner                = THIS_MODULE,
1442         .o_setup                = lov_setup,
1443         .o_cleanup              = lov_cleanup,
1444         .o_connect              = lov_connect,
1445         .o_disconnect           = lov_disconnect,
1446         .o_statfs               = lov_statfs,
1447         .o_statfs_async         = lov_statfs_async,
1448         .o_iocontrol            = lov_iocontrol,
1449         .o_get_info             = lov_get_info,
1450         .o_set_info_async       = lov_set_info_async,
1451         .o_notify               = lov_notify,
1452         .o_pool_new             = lov_pool_new,
1453         .o_pool_rem             = lov_pool_remove,
1454         .o_pool_add             = lov_pool_add,
1455         .o_pool_del             = lov_pool_del,
1456         .o_getref               = lov_getref,
1457         .o_putref               = lov_putref,
1458         .o_quotactl             = lov_quotactl,
1459 };
1460
1461 struct kmem_cache *lov_oinfo_slab;
1462
1463 static int __init lov_init(void)
1464 {
1465         bool enable_proc = true;
1466         struct obd_type *type;
1467         int rc;
1468         ENTRY;
1469
1470         /* print an address of _any_ initialized kernel symbol from this
1471          * module, to allow debugging with gdb that doesn't support data
1472          * symbols from modules.*/
1473         CDEBUG(D_INFO, "Lustre LOV module (%p).\n", &lov_caches);
1474
1475         rc = lu_kmem_init(lov_caches);
1476         if (rc)
1477                 return rc;
1478
1479         lov_oinfo_slab = kmem_cache_create("lov_oinfo",
1480                                            sizeof(struct lov_oinfo), 0,
1481                                            SLAB_HWCACHE_ALIGN, NULL);
1482         if (lov_oinfo_slab == NULL) {
1483                 lu_kmem_fini(lov_caches);
1484                 return -ENOMEM;
1485         }
1486
1487         type = class_search_type(LUSTRE_LOD_NAME);
1488         if (type != NULL && type->typ_procsym != NULL)
1489                 enable_proc = false;
1490
1491         rc = class_register_type(&lov_obd_ops, NULL, enable_proc, NULL,
1492                                  LUSTRE_LOV_NAME, &lov_device_type);
1493
1494         if (rc) {
1495                 kmem_cache_destroy(lov_oinfo_slab);
1496                 lu_kmem_fini(lov_caches);
1497         }
1498
1499         RETURN(rc);
1500 }
1501
1502 static void __exit lov_exit(void)
1503 {
1504         class_unregister_type(LUSTRE_LOV_NAME);
1505         kmem_cache_destroy(lov_oinfo_slab);
1506         lu_kmem_fini(lov_caches);
1507 }
1508
1509 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1510 MODULE_DESCRIPTION("Lustre Logical Object Volume");
1511 MODULE_VERSION(LUSTRE_VERSION_STRING);
1512 MODULE_LICENSE("GPL");
1513
1514 module_init(lov_init);
1515 module_exit(lov_exit);