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, 2014, 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 <obd_support.h>
  48 #include <lustre_ioctl.h>
  49 #include <lustre_lib.h>
  50 #include <lustre_net.h>
  51 #include <lustre/lustre_idl.h>
  52 #include <lustre_dlm.h>
  53 #include <lustre_mds.h>
  54 #include <obd_class.h>
  55 #include <lprocfs_status.h>
  56 #include <lustre_param.h>
  57 #include <cl_object.h>
  58 #include <lustre_fid.h>
  59
  60 #include "lov_internal.h"
  61
  62 /* Keep a refcount of lov->tgt usage to prevent racing with addition/deletion.
  63    Any function that expects lov_tgts to remain stationary must take a ref. */
  64 static void lov_getref(struct obd_device *obd)
  65 {
  66         struct lov_obd *lov = &obd->u.lov;
  67
  68         /* nobody gets through here until lov_putref is done */
  69         mutex_lock(&lov->lov_lock);
  70         atomic_inc(&lov->lov_refcount);
  71         mutex_unlock(&lov->lov_lock);
  72         return;
  73 }
  74
  75 static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt);
  76
  77 static void lov_putref(struct obd_device *obd)
  78 {
  79         struct lov_obd *lov = &obd->u.lov;
  80
  81         mutex_lock(&lov->lov_lock);
  82         /* ok to dec to 0 more than once -- ltd_exp's will be null */
  83         if (atomic_dec_and_test(&lov->lov_refcount) && lov->lov_death_row) {
  84                 struct list_head kill = LIST_HEAD_INIT(kill);
  85                 struct lov_tgt_desc *tgt, *n;
  86                 int i;
  87
  88                 CDEBUG(D_CONFIG, "destroying %d lov targets\n",
  89                        lov->lov_death_row);
  90                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
  91                         tgt = lov->lov_tgts[i];
  92
  93                         if (!tgt || !tgt->ltd_reap)
  94                                 continue;
  95                         list_add(&tgt->ltd_kill, &kill);
  96                         /* XXX - right now there is a dependency on ld_tgt_count
  97                          * being the maximum tgt index for computing the
  98                          * mds_max_easize. So we can't shrink it. */
  99                         lov_ost_pool_remove(&lov->lov_packed, i);
 100                         lov->lov_tgts[i] = NULL;
 101                         lov->lov_death_row--;
 102                 }
 103                 mutex_unlock(&lov->lov_lock);
 104
 105                 list_for_each_entry_safe(tgt, n, &kill, ltd_kill) {
 106                         list_del(&tgt->ltd_kill);
 107                         /* Disconnect */
 108                         __lov_del_obd(obd, tgt);
 109                 }
 110         } else {
 111                 mutex_unlock(&lov->lov_lock);
 112         }
 113 }
 114
 115 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
 116                               enum obd_notify_event ev);
 117 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
 118                       enum obd_notify_event ev, void *data);
 119
 120 int lov_connect_obd(struct obd_device *obd, __u32 index, int activate,
 121                     struct obd_connect_data *data)
 122 {
 123         struct lov_obd *lov = &obd->u.lov;
 124         struct obd_uuid *tgt_uuid;
 125         struct obd_device *tgt_obd;
 126         static struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
 127         struct obd_import *imp;
 128         int rc;
 129         ENTRY;
 130
 131         if (lov->lov_tgts[index] == NULL)
 132                 RETURN(-EINVAL);
 133
 134         tgt_uuid = &lov->lov_tgts[index]->ltd_uuid;
 135         tgt_obd = lov->lov_tgts[index]->ltd_obd;
 136
 137         if (!tgt_obd->obd_set_up) {
 138                 CERROR("Target %s not set up\n", obd_uuid2str(tgt_uuid));
 139                 RETURN(-EINVAL);
 140         }
 141
 142         /* override the sp_me from lov */
 143         tgt_obd->u.cli.cl_sp_me = lov->lov_sp_me;
 144
 145         if (data && (data->ocd_connect_flags & OBD_CONNECT_INDEX))
 146                 data->ocd_index = index;
 147
 148         /*
 149          * Divine LOV knows that OBDs under it are OSCs.
 150          */
 151         imp = tgt_obd->u.cli.cl_import;
 152
 153         if (activate) {
 154                 tgt_obd->obd_no_recov = 0;
 155                 /* FIXME this is probably supposed to be
 156                    ptlrpc_set_import_active.  Horrible naming. */
 157                 ptlrpc_activate_import(imp);
 158         }
 159
 160         rc = obd_register_observer(tgt_obd, obd);
 161         if (rc) {
 162                 CERROR("Target %s register_observer error %d\n",
 163                        obd_uuid2str(tgt_uuid), rc);
 164                 RETURN(rc);
 165         }
 166
 167
 168         if (imp->imp_invalid) {
 169                 CDEBUG(D_CONFIG, "not connecting OSC %s; administratively "
 170                        "disabled\n", obd_uuid2str(tgt_uuid));
 171                 RETURN(0);
 172         }
 173
 174         rc = obd_connect(NULL, &lov->lov_tgts[index]->ltd_exp, tgt_obd,
 175                          &lov_osc_uuid, data, NULL);
 176         if (rc || !lov->lov_tgts[index]->ltd_exp) {
 177                 CERROR("Target %s connect error %d\n",
 178                        obd_uuid2str(tgt_uuid), rc);
 179                 RETURN(-ENODEV);
 180         }
 181
 182         lov->lov_tgts[index]->ltd_reap = 0;
 183
 184         CDEBUG(D_CONFIG, "Connected tgt idx %d %s (%s) %sactive\n", index,
 185                obd_uuid2str(tgt_uuid), tgt_obd->obd_name, activate ? "":"in");
 186
 187         if (lov->targets_proc_entry != NULL) {
 188                 struct proc_dir_entry *osc_symlink;
 189                 struct obd_device *osc_obd;
 190
 191                 osc_obd = lov->lov_tgts[index]->ltd_exp->exp_obd;
 192
 193                 LASSERT(osc_obd != NULL);
 194                 LASSERT(osc_obd->obd_magic == OBD_DEVICE_MAGIC);
 195                 LASSERT(osc_obd->obd_type->typ_name != NULL);
 196
 197                 osc_symlink = lprocfs_add_symlink(osc_obd->obd_name,
 198                                                   lov->targets_proc_entry,
 199                                                   "../../../%s/%s",
 200                                                   osc_obd->obd_type->typ_name,
 201                                                   osc_obd->obd_name);
 202                 if (osc_symlink == NULL) {
 203                         CERROR("cannot register LOV target "
 204                                "/proc/fs/lustre/%s/%s/target_obds/%s\n",
 205                                obd->obd_type->typ_name, obd->obd_name,
 206                                osc_obd->obd_name);
 207                 }
 208         }
 209         RETURN(0);
 210 }
 211
 212 static int lov_connect(const struct lu_env *env,
 213                        struct obd_export **exp, struct obd_device *obd,
 214                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 215                        void *localdata)
 216 {
 217         struct lov_obd *lov = &obd->u.lov;
 218         struct lov_tgt_desc *tgt;
 219         struct lustre_handle conn;
 220         int i, rc;
 221         ENTRY;
 222
 223         CDEBUG(D_CONFIG, "connect #%d\n", lov->lov_connects);
 224
 225         rc = class_connect(&conn, obd, cluuid);
 226         if (rc)
 227                 RETURN(rc);
 228
 229         *exp = class_conn2export(&conn);
 230
 231         /* Why should there ever be more than 1 connect? */
 232         lov->lov_connects++;
 233         LASSERT(lov->lov_connects == 1);
 234
 235         memset(&lov->lov_ocd, 0, sizeof(lov->lov_ocd));
 236         if (data)
 237                 lov->lov_ocd = *data;
 238
 239         lov->targets_proc_entry = lprocfs_register("target_obds",
 240                                                    obd->obd_proc_entry,
 241                                                    NULL, NULL);
 242         if (IS_ERR(lov->targets_proc_entry)) {
 243                 CERROR("%s: cannot register "
 244                        "/proc/fs/lustre/%s/%s/target_obds\n",
 245                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
 246                 lov->targets_proc_entry = NULL;
 247         }
 248
 249         obd_getref(obd);
 250         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 251                 tgt = lov->lov_tgts[i];
 252                 if (!tgt || obd_uuid_empty(&tgt->ltd_uuid))
 253                         continue;
 254                 /* Flags will be lowest common denominator */
 255                 rc = lov_connect_obd(obd, i, tgt->ltd_activate, &lov->lov_ocd);
 256                 if (rc) {
 257                         CERROR("%s: lov connect tgt %d failed: %d\n",
 258                                obd->obd_name, i, rc);
 259                         continue;
 260                 }
 261                 /* connect to administrative disabled ost */
 262                 if (!lov->lov_tgts[i]->ltd_exp)
 263                         continue;
 264
 265                 rc = lov_notify(obd, lov->lov_tgts[i]->ltd_exp->exp_obd,
 266                                 OBD_NOTIFY_CONNECT, (void *)&i);
 267                 if (rc) {
 268                         CERROR("%s error sending notify %d\n",
 269                                obd->obd_name, rc);
 270                 }
 271         }
 272         obd_putref(obd);
 273
 274         RETURN(0);
 275 }
 276
 277 static int lov_disconnect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
 278 {
 279         struct lov_obd *lov = &obd->u.lov;
 280         struct obd_device *osc_obd;
 281         int rc;
 282         ENTRY;
 283
 284         osc_obd = class_exp2obd(tgt->ltd_exp);
 285         CDEBUG(D_CONFIG, "%s: disconnecting target %s\n",
 286                obd->obd_name, osc_obd->obd_name);
 287
 288         if (tgt->ltd_active) {
 289                 tgt->ltd_active = 0;
 290                 lov->desc.ld_active_tgt_count--;
 291                 tgt->ltd_exp->exp_obd->obd_inactive = 1;
 292         }
 293
 294         if (osc_obd) {
 295                 /* Pass it on to our clients.
 296                  * XXX This should be an argument to disconnect,
 297                  * XXX not a back-door flag on the OBD.  Ah well.
 298                  */
 299                 osc_obd->obd_force = obd->obd_force;
 300                 osc_obd->obd_fail = obd->obd_fail;
 301                 osc_obd->obd_no_recov = obd->obd_no_recov;
 302
 303                 if (lov->targets_proc_entry != NULL)
 304                         lprocfs_remove_proc_entry(osc_obd->obd_name,
 305                                                   lov->targets_proc_entry);
 306         }
 307
 308         obd_register_observer(osc_obd, NULL);
 309
 310         rc = obd_disconnect(tgt->ltd_exp);
 311         if (rc) {
 312                 CERROR("Target %s disconnect error %d\n",
 313                        tgt->ltd_uuid.uuid, rc);
 314                 rc = 0;
 315         }
 316
 317         tgt->ltd_exp = NULL;
 318         RETURN(0);
 319 }
 320
 321 static int lov_disconnect(struct obd_export *exp)
 322 {
 323         struct obd_device *obd = class_exp2obd(exp);
 324         struct lov_obd *lov = &obd->u.lov;
 325         int i, rc;
 326         ENTRY;
 327
 328         if (!lov->lov_tgts)
 329                 goto out;
 330
 331         /* Only disconnect the underlying layers on the final disconnect. */
 332         lov->lov_connects--;
 333         if (lov->lov_connects != 0) {
 334                 /* why should there be more than 1 connect? */
 335                 CERROR("disconnect #%d\n", lov->lov_connects);
 336                 goto out;
 337         }
 338
 339         /* Let's hold another reference so lov_del_obd doesn't spin through
 340            putref every time */
 341         obd_getref(obd);
 342
 343         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 344                 if (lov->lov_tgts[i] && lov->lov_tgts[i]->ltd_exp) {
 345                         /* Disconnection is the last we know about an obd */
 346                         lov_del_target(obd, i, NULL, lov->lov_tgts[i]->ltd_gen);
 347                 }
 348         }
 349         obd_putref(obd);
 350
 351         if (lov->targets_proc_entry != NULL)
 352                 lprocfs_remove(&lov->targets_proc_entry);
 353
 354 out:
 355         rc = class_disconnect(exp); /* bz 9811 */
 356         RETURN(rc);
 357 }
 358
 359 /* Error codes:
 360  *
 361  *  -EINVAL  : UUID can't be found in the LOV's target list
 362  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
 363  *  -EBADF   : The UUID is found, but the OBD is the wrong type (!)
 364  *  any >= 0 : is log target index
 365  */
 366 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
 367                               enum obd_notify_event ev)
 368 {
 369         struct lov_obd *lov = &obd->u.lov;
 370         struct lov_tgt_desc *tgt;
 371         int index, activate, active;
 372         ENTRY;
 373
 374         CDEBUG(D_INFO, "Searching in lov %p for uuid %s event(%d)\n",
 375                lov, uuid->uuid, ev);
 376
 377         obd_getref(obd);
 378         for (index = 0; index < lov->desc.ld_tgt_count; index++) {
 379                 tgt = lov->lov_tgts[index];
 380                 if (!tgt)
 381                         continue;
 382                 /*
 383                  * LU-642, initially inactive OSC could miss the obd_connect,
 384                  * we make up for it here.
 385                  */
 386                 if (ev == OBD_NOTIFY_ACTIVATE && tgt->ltd_exp == NULL &&
 387                     obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
 388                         struct obd_uuid lov_osc_uuid = {"LOV_OSC_UUID"};
 389
 390                         obd_connect(NULL, &tgt->ltd_exp, tgt->ltd_obd,
 391                                     &lov_osc_uuid, &lov->lov_ocd, NULL);
 392                 }
 393                 if (!tgt->ltd_exp)
 394                         continue;
 395
 396                 CDEBUG(D_INFO, "lov idx %d is %s conn "LPX64"\n",
 397                        index, obd_uuid2str(&tgt->ltd_uuid),
 398                        tgt->ltd_exp->exp_handle.h_cookie);
 399                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 400                         break;
 401         }
 402
 403         if (index == lov->desc.ld_tgt_count)
 404                 GOTO(out, index = -EINVAL);
 405
 406         if (ev == OBD_NOTIFY_DEACTIVATE || ev == OBD_NOTIFY_ACTIVATE) {
 407                 activate = (ev == OBD_NOTIFY_ACTIVATE) ? 1 : 0;
 408
 409                 if (lov->lov_tgts[index]->ltd_activate == activate) {
 410                         CDEBUG(D_INFO, "OSC %s already %sactivate!\n",
 411                                uuid->uuid, activate ? "" : "de");
 412                 } else {
 413                         lov->lov_tgts[index]->ltd_activate = activate;
 414                         CDEBUG(D_CONFIG, "%sactivate OSC %s\n",
 415                                activate ? "" : "de", obd_uuid2str(uuid));
 416                 }
 417
 418         } else if (ev == OBD_NOTIFY_INACTIVE || ev == OBD_NOTIFY_ACTIVE) {
 419                 active = (ev == OBD_NOTIFY_ACTIVE) ? 1 : 0;
 420
 421                 if (lov->lov_tgts[index]->ltd_active == active) {
 422                         CDEBUG(D_INFO, "OSC %s already %sactive!\n",
 423                                uuid->uuid, active ? "" : "in");
 424                         GOTO(out, index);
 425                 } else {
 426                         CDEBUG(D_CONFIG, "Marking OSC %s %sactive\n",
 427                                obd_uuid2str(uuid), active ? "" : "in");
 428                 }
 429
 430                 lov->lov_tgts[index]->ltd_active = active;
 431                 if (active) {
 432                         lov->desc.ld_active_tgt_count++;
 433                         lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 0;
 434                 } else {
 435                         lov->desc.ld_active_tgt_count--;
 436                         lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 1;
 437                 }
 438         } else {
 439                 CERROR("%s: unknown event %d for uuid %s\n", obd->obd_name,
 440                        ev, uuid->uuid);
 441         }
 442
 443  out:
 444         obd_putref(obd);
 445         RETURN(index);
 446 }
 447
 448 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
 449                       enum obd_notify_event ev, void *data)
 450 {
 451         int rc = 0;
 452         struct lov_obd *lov = &obd->u.lov;
 453         ENTRY;
 454
 455         down_read(&lov->lov_notify_lock);
 456         if (!lov->lov_connects) {
 457                 up_read(&lov->lov_notify_lock);
 458                 RETURN(rc);
 459         }
 460
 461         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE ||
 462             ev == OBD_NOTIFY_ACTIVATE || ev == OBD_NOTIFY_DEACTIVATE) {
 463                 struct obd_uuid *uuid;
 464
 465                 LASSERT(watched);
 466
 467                 if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) {
 468                         up_read(&lov->lov_notify_lock);
 469                         CERROR("unexpected notification of %s %s!\n",
 470                                watched->obd_type->typ_name,
 471                                watched->obd_name);
 472                         RETURN(-EINVAL);
 473                 }
 474                 uuid = &watched->u.cli.cl_target_uuid;
 475
 476                 /* Set OSC as active before notifying the observer, so the
 477                  * observer can use the OSC normally.
 478                  */
 479                 rc = lov_set_osc_active(obd, uuid, ev);
 480                 if (rc < 0) {
 481                         up_read(&lov->lov_notify_lock);
 482                         CERROR("event(%d) of %s failed: %d\n", ev,
 483                                obd_uuid2str(uuid), rc);
 484                         RETURN(rc);
 485                 }
 486                 /* active event should be pass lov target index as data */
 487                 data = &rc;
 488         }
 489
 490         /* Pass the notification up the chain. */
 491         if (watched) {
 492                 rc = obd_notify_observer(obd, watched, ev, data);
 493         } else {
 494                 /* NULL watched means all osc's in the lov (only for syncs) */
 495                 /* sync event should be send lov idx as data */
 496                 struct lov_obd *lov = &obd->u.lov;
 497                 int i, is_sync;
 498
 499                 data = &i;
 500                 is_sync = (ev == OBD_NOTIFY_SYNC) ||
 501                           (ev == OBD_NOTIFY_SYNC_NONBLOCK);
 502
 503                 obd_getref(obd);
 504                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 505                         if (!lov->lov_tgts[i])
 506                                 continue;
 507
 508                         /* don't send sync event if target not
 509                          * connected/activated */
 510                         if (is_sync &&  !lov->lov_tgts[i]->ltd_active)
 511                                 continue;
 512
 513                         rc = obd_notify_observer(obd, lov->lov_tgts[i]->ltd_obd,
 514                                                  ev, data);
 515                         if (rc) {
 516                                 CERROR("%s: notify %s of %s failed %d\n",
 517                                        obd->obd_name,
 518                                        obd->obd_observer->obd_name,
 519                                        lov->lov_tgts[i]->ltd_obd->obd_name,
 520                                        rc);
 521                         }
 522                 }
 523                 obd_putref(obd);
 524         }
 525
 526         up_read(&lov->lov_notify_lock);
 527         RETURN(rc);
 528 }
 529
 530 static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 531                           __u32 index, int gen, int active)
 532 {
 533         struct lov_obd *lov = &obd->u.lov;
 534         struct lov_tgt_desc *tgt;
 535         struct obd_device *tgt_obd;
 536         int rc;
 537         ENTRY;
 538
 539         CDEBUG(D_CONFIG, "uuid:%s idx:%d gen:%d active:%d\n",
 540                uuidp->uuid, index, gen, active);
 541
 542         if (gen <= 0) {
 543                 CERROR("request to add OBD %s with invalid generation: %d\n",
 544                        uuidp->uuid, gen);
 545                 RETURN(-EINVAL);
 546         }
 547
 548         tgt_obd = class_find_client_obd(uuidp, LUSTRE_OSC_NAME,
 549                                         &obd->obd_uuid);
 550         if (tgt_obd == NULL)
 551                 RETURN(-EINVAL);
 552
 553         mutex_lock(&lov->lov_lock);
 554
 555         if ((index < lov->lov_tgt_size) && (lov->lov_tgts[index] != NULL)) {
 556                 tgt = lov->lov_tgts[index];
 557                 CERROR("UUID %s already assigned at LOV target index %d\n",
 558                        obd_uuid2str(&tgt->ltd_uuid), index);
 559                 mutex_unlock(&lov->lov_lock);
 560                 RETURN(-EEXIST);
 561         }
 562
 563         if (index >= lov->lov_tgt_size) {
 564                 /* We need to reallocate the lov target array. */
 565                 struct lov_tgt_desc **newtgts, **old = NULL;
 566                 __u32 newsize, oldsize = 0;
 567
 568                 newsize = max(lov->lov_tgt_size, (__u32)2);
 569                 while (newsize < index + 1)
 570                         newsize = newsize << 1;
 571                 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
 572                 if (newtgts == NULL) {
 573                         mutex_unlock(&lov->lov_lock);
 574                         RETURN(-ENOMEM);
 575                 }
 576
 577                 if (lov->lov_tgt_size) {
 578                         memcpy(newtgts, lov->lov_tgts, sizeof(*newtgts) *
 579                                lov->lov_tgt_size);
 580                         old = lov->lov_tgts;
 581                         oldsize = lov->lov_tgt_size;
 582                 }
 583
 584                 lov->lov_tgts = newtgts;
 585                 lov->lov_tgt_size = newsize;
 586                 smp_rmb();
 587                 if (old)
 588                         OBD_FREE(old, sizeof(*old) * oldsize);
 589
 590                 CDEBUG(D_CONFIG, "tgts: %p size: %d\n",
 591                        lov->lov_tgts, lov->lov_tgt_size);
 592         }
 593
 594         OBD_ALLOC_PTR(tgt);
 595         if (!tgt) {
 596                 mutex_unlock(&lov->lov_lock);
 597                 RETURN(-ENOMEM);
 598         }
 599
 600         rc = lov_ost_pool_add(&lov->lov_packed, index, lov->lov_tgt_size);
 601         if (rc) {
 602                 mutex_unlock(&lov->lov_lock);
 603                 OBD_FREE_PTR(tgt);
 604                 RETURN(rc);
 605         }
 606
 607         tgt->ltd_uuid = *uuidp;
 608         tgt->ltd_obd = tgt_obd;
 609         /* XXX - add a sanity check on the generation number. */
 610         tgt->ltd_gen = gen;
 611         tgt->ltd_index = index;
 612         tgt->ltd_activate = active;
 613         lov->lov_tgts[index] = tgt;
 614         if (index >= lov->desc.ld_tgt_count)
 615                 lov->desc.ld_tgt_count = index + 1;
 616
 617         mutex_unlock(&lov->lov_lock);
 618
 619         CDEBUG(D_CONFIG, "idx=%d ltd_gen=%d ld_tgt_count=%d\n",
 620                 index, tgt->ltd_gen, lov->desc.ld_tgt_count);
 621
 622         rc = obd_notify(obd, tgt_obd, OBD_NOTIFY_CREATE, &index);
 623
 624         if (lov->lov_connects == 0) {
 625                 /* lov_connect hasn't been called yet. We'll do the
 626                    lov_connect_obd on this target when that fn first runs,
 627                    because we don't know the connect flags yet. */
 628                 RETURN(0);
 629         }
 630
 631         obd_getref(obd);
 632
 633         rc = lov_connect_obd(obd, index, active, &lov->lov_ocd);
 634         if (rc)
 635                 GOTO(out, rc);
 636
 637         /* connect to administrative disabled ost */
 638         if (!tgt->ltd_exp)
 639                 GOTO(out, rc = 0);
 640
 641         if (lov->lov_cache != NULL) {
 642                 rc = obd_set_info_async(NULL, tgt->ltd_exp,
 643                                 sizeof(KEY_CACHE_SET), KEY_CACHE_SET,
 644                                 sizeof(struct cl_client_cache), lov->lov_cache,
 645                                 NULL);
 646                 if (rc < 0)
 647                         GOTO(out, rc);
 648         }
 649
 650         rc = lov_notify(obd, tgt->ltd_exp->exp_obd,
 651                         active ? OBD_NOTIFY_CONNECT : OBD_NOTIFY_INACTIVE,
 652                         (void *)&index);
 653
 654 out:
 655         if (rc) {
 656                 CERROR("add failed (%d), deleting %s\n", rc,
 657                        obd_uuid2str(&tgt->ltd_uuid));
 658                 lov_del_target(obd, index, NULL, 0);
 659         }
 660         obd_putref(obd);
 661         RETURN(rc);
 662 }
 663
 664 /* Schedule a target for deletion */
 665 int lov_del_target(struct obd_device *obd, __u32 index,
 666                    struct obd_uuid *uuidp, int gen)
 667 {
 668         struct lov_obd *lov = &obd->u.lov;
 669         int count = lov->desc.ld_tgt_count;
 670         int rc = 0;
 671         ENTRY;
 672
 673         if (index >= count) {
 674                 CERROR("LOV target index %d >= number of LOV OBDs %d.\n",
 675                        index, count);
 676                 RETURN(-EINVAL);
 677         }
 678
 679         /* to make sure there's no ongoing lov_notify() now */
 680         down_write(&lov->lov_notify_lock);
 681         obd_getref(obd);
 682
 683         if (!lov->lov_tgts[index]) {
 684                 CERROR("LOV target at index %d is not setup.\n", index);
 685                 GOTO(out, rc = -EINVAL);
 686         }
 687
 688         if (uuidp && !obd_uuid_equals(uuidp, &lov->lov_tgts[index]->ltd_uuid)) {
 689                 CERROR("LOV target UUID %s at index %d doesn't match %s.\n",
 690                        lov_uuid2str(lov, index), index,
 691                        obd_uuid2str(uuidp));
 692                 GOTO(out, rc = -EINVAL);
 693         }
 694
 695         CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d exp: %p active: %d\n",
 696                lov_uuid2str(lov, index), index,
 697                lov->lov_tgts[index]->ltd_gen, lov->lov_tgts[index]->ltd_exp,
 698                lov->lov_tgts[index]->ltd_active);
 699
 700         lov->lov_tgts[index]->ltd_reap = 1;
 701         lov->lov_death_row++;
 702         /* we really delete it from obd_putref */
 703 out:
 704         obd_putref(obd);
 705         up_write(&lov->lov_notify_lock);
 706
 707         RETURN(rc);
 708 }
 709
 710 static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
 711 {
 712         struct obd_device *osc_obd;
 713
 714         LASSERT(tgt);
 715         LASSERT(tgt->ltd_reap);
 716
 717         osc_obd = class_exp2obd(tgt->ltd_exp);
 718
 719         CDEBUG(D_CONFIG, "Removing tgt %s : %s\n",
 720                tgt->ltd_uuid.uuid,
 721                osc_obd ? osc_obd->obd_name : "<no obd>");
 722
 723         if (tgt->ltd_exp)
 724                 lov_disconnect_obd(obd, tgt);
 725
 726         OBD_FREE_PTR(tgt);
 727
 728         /* Manual cleanup - no cleanup logs to clean up the osc's.  We must
 729            do it ourselves. And we can't do it from lov_cleanup,
 730            because we just lost our only reference to it. */
 731         if (osc_obd)
 732                 class_manual_cleanup(osc_obd);
 733 }
 734
 735 void lov_fix_desc_stripe_size(__u64 *val)
 736 {
 737         if (*val < LOV_MIN_STRIPE_SIZE) {
 738                 if (*val != 0)
 739                         LCONSOLE_INFO("Increasing default stripe size to "
 740                                       "minimum %u\n",
 741                                       LOV_DESC_STRIPE_SIZE_DEFAULT);
 742                 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
 743         } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
 744                 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
 745                 LCONSOLE_WARN("Changing default stripe size to "LPU64" (a "
 746                               "multiple of %u)\n",
 747                               *val, LOV_MIN_STRIPE_SIZE);
 748         }
 749 }
 750
 751 void lov_fix_desc_stripe_count(__u32 *val)
 752 {
 753         if (*val == 0)
 754                 *val = 1;
 755 }
 756
 757 void lov_fix_desc_pattern(__u32 *val)
 758 {
 759         /* from lov_setstripe */
 760         if ((*val != 0) && (*val != LOV_PATTERN_RAID0)) {
 761                 LCONSOLE_WARN("Unknown stripe pattern: %#x\n", *val);
 762                 *val = 0;
 763         }
 764 }
 765
 766 void lov_fix_desc_qos_maxage(__u32 *val)
 767 {
 768         if (*val == 0)
 769                 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
 770 }
 771
 772 void lov_fix_desc(struct lov_desc *desc)
 773 {
 774         lov_fix_desc_stripe_size(&desc->ld_default_stripe_size);
 775         lov_fix_desc_stripe_count(&desc->ld_default_stripe_count);
 776         lov_fix_desc_pattern(&desc->ld_pattern);
 777         lov_fix_desc_qos_maxage(&desc->ld_qos_maxage);
 778 }
 779
 780 int lov_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 781 {
 782         struct lov_desc *desc;
 783         struct lov_obd *lov = &obd->u.lov;
 784 #ifdef CONFIG_PROC_FS
 785         struct obd_type *type;
 786 #endif
 787         int rc;
 788         ENTRY;
 789
 790         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 791                 CERROR("LOV setup requires a descriptor\n");
 792                 RETURN(-EINVAL);
 793         }
 794
 795         desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
 796
 797         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
 798                 CERROR("descriptor size wrong: %d > %d\n",
 799                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
 800                 RETURN(-EINVAL);
 801         }
 802
 803         if (desc->ld_magic != LOV_DESC_MAGIC) {
 804                 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
 805                             CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
 806                                    obd->obd_name, desc);
 807                             lustre_swab_lov_desc(desc);
 808                 } else {
 809                         CERROR("%s: Bad lov desc magic: %#x\n",
 810                                obd->obd_name, desc->ld_magic);
 811                         RETURN(-EINVAL);
 812                 }
 813         }
 814
 815         lov_fix_desc(desc);
 816
 817         desc->ld_active_tgt_count = 0;
 818         lov->desc = *desc;
 819         lov->lov_tgt_size = 0;
 820
 821         mutex_init(&lov->lov_lock);
 822         atomic_set(&lov->lov_refcount, 0);
 823         lov->lov_sp_me = LUSTRE_SP_CLI;
 824
 825         init_rwsem(&lov->lov_notify_lock);
 826
 827         lov->lov_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,
 828                                                    HASH_POOLS_MAX_BITS,
 829                                                    HASH_POOLS_BKT_BITS, 0,
 830                                                    CFS_HASH_MIN_THETA,
 831                                                    CFS_HASH_MAX_THETA,
 832                                                    &pool_hash_operations,
 833                                                    CFS_HASH_DEFAULT);
 834         INIT_LIST_HEAD(&lov->lov_pool_list);
 835         lov->lov_pool_count = 0;
 836         rc = lov_ost_pool_init(&lov->lov_packed, 0);
 837         if (rc)
 838                 GOTO(out, rc);
 839
 840 #ifdef CONFIG_PROC_FS
 841         obd->obd_vars = lprocfs_lov_obd_vars;
 842         /* If this is true then both client (lov) and server
 843          * (lod) are on the same node. The lod layer if loaded
 844          * first will register the lov proc directory. In that
 845          * case obd->obd_type->typ_procroot will be not set.
 846          * Instead we use type->typ_procsym as the parent. */
 847         type = class_search_type(LUSTRE_LOD_NAME);
 848         if (type != NULL && type->typ_procsym != NULL) {
 849                 obd->obd_proc_entry = lprocfs_register(obd->obd_name,
 850                                                        type->typ_procsym,
 851                                                        obd->obd_vars, obd);
 852                 if (IS_ERR(obd->obd_proc_entry)) {
 853                         rc = PTR_ERR(obd->obd_proc_entry);
 854                         CERROR("error %d setting up lprocfs for %s\n", rc,
 855                                obd->obd_name);
 856                         obd->obd_proc_entry = NULL;
 857                 }
 858         } else {
 859                 rc = lprocfs_obd_setup(obd);
 860         }
 861
 862         if (rc == 0) {
 863                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd",
 864                                         0444, &lov_proc_target_fops, obd);
 865                 if (rc)
 866                         CWARN("Error adding the target_obd file\n");
 867
 868                 lov->lov_pool_proc_entry = lprocfs_register("pools",
 869                                                             obd->obd_proc_entry,
 870                                                             NULL, NULL);
 871                 if (IS_ERR(lov->lov_pool_proc_entry)) {
 872                         rc = PTR_ERR(lov->lov_pool_proc_entry);
 873                         CERROR("error %d setting up lprocfs for pools\n", rc);
 874                         lov->lov_pool_proc_entry = NULL;
 875                 }
 876         }
 877 #endif
 878         RETURN(0);
 879
 880 out:
 881         return rc;
 882 }
 883
 884 static int lov_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 885 {
 886         int rc = 0;
 887         struct lov_obd *lov = &obd->u.lov;
 888
 889         ENTRY;
 890
 891         switch (stage) {
 892         case OBD_CLEANUP_EARLY: {
 893                 int i;
 894                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 895                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 896                                 continue;
 897                         obd_precleanup(class_exp2obd(lov->lov_tgts[i]->ltd_exp),
 898                                        OBD_CLEANUP_EARLY);
 899                 }
 900                 break;
 901         }
 902         default:
 903                 break;
 904         }
 905
 906         RETURN(rc);
 907 }
 908
 909 static int lov_cleanup(struct obd_device *obd)
 910 {
 911         struct lov_obd *lov = &obd->u.lov;
 912         struct list_head *pos, *tmp;
 913         struct pool_desc *pool;
 914         ENTRY;
 915
 916         list_for_each_safe(pos, tmp, &lov->lov_pool_list) {
 917                 pool = list_entry(pos, struct pool_desc, pool_list);
 918                 /* free pool structs */
 919                 CDEBUG(D_INFO, "delete pool %p\n", pool);
 920                 /* In the function below, .hs_keycmp resolves to
 921                  * pool_hashkey_keycmp() */
 922                 /* coverity[overrun-buffer-val] */
 923                 lov_pool_del(obd, pool->pool_name);
 924         }
 925         cfs_hash_putref(lov->lov_pools_hash_body);
 926         lov_ost_pool_free(&lov->lov_packed);
 927
 928         lprocfs_obd_cleanup(obd);
 929         if (lov->lov_tgts) {
 930                 int i;
 931                 obd_getref(obd);
 932                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 933                         if (!lov->lov_tgts[i])
 934                                 continue;
 935
 936                         /* Inactive targets may never have connected */
 937                         if (lov->lov_tgts[i]->ltd_active ||
 938                             atomic_read(&lov->lov_refcount))
 939                                 /* We should never get here - these
 940                                  * should have been removed in the
 941                                  * disconnect. */
 942                                 CERROR("%s: lov tgt %d not cleaned! "
 943                                        "deathrow=%d, lovrc=%d\n",
 944                                        obd->obd_name, i, lov->lov_death_row,
 945                                        atomic_read(&lov->lov_refcount));
 946                         lov_del_target(obd, i, NULL, 0);
 947                 }
 948                 obd_putref(obd);
 949                 OBD_FREE(lov->lov_tgts, sizeof(*lov->lov_tgts) *
 950                          lov->lov_tgt_size);
 951                 lov->lov_tgt_size = 0;
 952         }
 953
 954         if (lov->lov_cache != NULL) {
 955                 cl_cache_decref(lov->lov_cache);
 956                 lov->lov_cache = NULL;
 957         }
 958
 959         RETURN(0);
 960 }
 961
 962 int lov_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg,
 963                             __u32 *indexp, int *genp)
 964 {
 965         struct obd_uuid obd_uuid;
 966         int cmd;
 967         int rc = 0;
 968         ENTRY;
 969
 970         switch(cmd = lcfg->lcfg_command) {
 971         case LCFG_LOV_ADD_OBD:
 972         case LCFG_LOV_ADD_INA:
 973         case LCFG_LOV_DEL_OBD: {
 974                 __u32 index;
 975                 int gen;
 976                 /* lov_modify_tgts add  0:lov_mdsA  1:ost1_UUID  2:0  3:1 */
 977                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
 978                         GOTO(out, rc = -EINVAL);
 979
 980                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
 981
 982                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", indexp) != 1)
 983                         GOTO(out, rc = -EINVAL);
 984                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", genp) != 1)
 985                         GOTO(out, rc = -EINVAL);
 986                 index = *indexp;
 987                 gen = *genp;
 988                 if (cmd == LCFG_LOV_ADD_OBD)
 989                         rc = lov_add_target(obd, &obd_uuid, index, gen, 1);
 990                 else if (cmd == LCFG_LOV_ADD_INA)
 991                         rc = lov_add_target(obd, &obd_uuid, index, gen, 0);
 992                 else
 993                         rc = lov_del_target(obd, index, &obd_uuid, gen);
 994                 GOTO(out, rc);
 995         }
 996         case LCFG_PARAM: {
 997                 struct lov_desc *desc = &(obd->u.lov.desc);
 998
 999                 if (!desc)
1000                         GOTO(out, rc = -EINVAL);
1001
1002                 rc = class_process_proc_param(PARAM_LOV, obd->obd_vars,
1003                                               lcfg, obd);
1004                 if (rc > 0)
1005                         rc = 0;
1006                 GOTO(out, rc);
1007         }
1008         case LCFG_POOL_NEW:
1009         case LCFG_POOL_ADD:
1010         case LCFG_POOL_DEL:
1011         case LCFG_POOL_REM:
1012                 GOTO(out, rc);
1013
1014         default: {
1015                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
1016                 GOTO(out, rc = -EINVAL);
1017
1018         }
1019         }
1020 out:
1021         RETURN(rc);
1022 }
1023
1024 #define ASSERT_LSM_MAGIC(lsmp)                                                  \
1025 do {                                                                            \
1026         LASSERT((lsmp) != NULL);                                                \
1027         LASSERTF(((lsmp)->lsm_magic == LOV_MAGIC_V1 ||                          \
1028                  (lsmp)->lsm_magic == LOV_MAGIC_V3),                            \
1029                  "%p->lsm_magic=%x\n", (lsmp), (lsmp)->lsm_magic);              \
1030 } while (0)
1031
1032 static int lov_getattr_interpret(struct ptlrpc_request_set *rqset,
1033                                  void *data, int rc)
1034 {
1035         struct lov_request_set *lovset = (struct lov_request_set *)data;
1036         int err;
1037         ENTRY;
1038
1039         /* don't do attribute merge if this aysnc op failed */
1040         if (rc)
1041                 atomic_set(&lovset->set_completes, 0);
1042         err = lov_fini_getattr_set(lovset);
1043         RETURN(rc ? rc : err);
1044 }
1045
1046 static int lov_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
1047                               struct ptlrpc_request_set *rqset)
1048 {
1049         struct lov_request_set *lovset;
1050         struct lov_obd *lov;
1051         struct list_head *pos;
1052         struct lov_request *req;
1053         int rc = 0, err;
1054         ENTRY;
1055
1056         LASSERT(oinfo);
1057         ASSERT_LSM_MAGIC(oinfo->oi_md);
1058
1059         if (!exp || !exp->exp_obd)
1060                 RETURN(-ENODEV);
1061
1062         lov = &exp->exp_obd->u.lov;
1063
1064         rc = lov_prep_getattr_set(exp, oinfo, &lovset);
1065         if (rc)
1066                 RETURN(rc);
1067
1068         CDEBUG(D_INFO, "objid "DOSTID": %ux%u byte stripes\n",
1069                POSTID(&oinfo->oi_md->lsm_oi), oinfo->oi_md->lsm_stripe_count,
1070                oinfo->oi_md->lsm_stripe_size);
1071
1072         list_for_each(pos, &lovset->set_list) {
1073                 req = list_entry(pos, struct lov_request, rq_link);
1074
1075                 CDEBUG(D_INFO, "objid "DOSTID"[%d] has subobj "DOSTID" at idx"
1076                        "%u\n", POSTID(&oinfo->oi_oa->o_oi), req->rq_stripe,
1077                        POSTID(&req->rq_oi.oi_oa->o_oi), req->rq_idx);
1078                 rc = obd_getattr_async(lov->lov_tgts[req->rq_idx]->ltd_exp,
1079                                        &req->rq_oi, rqset);
1080                 if (rc) {
1081                         CERROR("%s: getattr objid "DOSTID" subobj"
1082                                DOSTID" on OST idx %d: rc = %d\n",
1083                                exp->exp_obd->obd_name,
1084                                POSTID(&oinfo->oi_oa->o_oi),
1085                                POSTID(&req->rq_oi.oi_oa->o_oi),
1086                                req->rq_idx, rc);
1087                         GOTO(out, rc);
1088                 }
1089         }
1090
1091         if (!list_empty(&rqset->set_requests)) {
1092                 LASSERT(rc == 0);
1093                 LASSERT (rqset->set_interpret == NULL);
1094                 rqset->set_interpret = lov_getattr_interpret;
1095                 rqset->set_arg = (void *)lovset;
1096                 RETURN(rc);
1097         }
1098 out:
1099         if (rc)
1100                 atomic_set(&lovset->set_completes, 0);
1101         err = lov_fini_getattr_set(lovset);
1102         RETURN(rc ? rc : err);
1103 }
1104
1105 static int lov_setattr_interpret(struct ptlrpc_request_set *rqset,
1106                                  void *data, int rc)
1107 {
1108         struct lov_request_set *lovset = (struct lov_request_set *)data;
1109         int err;
1110         ENTRY;
1111
1112         if (rc)
1113                 atomic_set(&lovset->set_completes, 0);
1114         err = lov_fini_setattr_set(lovset);
1115         RETURN(rc ? rc : err);
1116 }
1117
1118 /* If @oti is given, the request goes from MDS and responses from OSTs are not
1119    needed. Otherwise, a client is waiting for responses. */
1120 static int lov_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
1121                              struct obd_trans_info *oti,
1122                              struct ptlrpc_request_set *rqset)
1123 {
1124         struct lov_request_set *set;
1125         struct lov_request *req;
1126         struct list_head *pos;
1127         struct lov_obd *lov;
1128         int rc = 0;
1129         ENTRY;
1130
1131         LASSERT(oinfo);
1132         ASSERT_LSM_MAGIC(oinfo->oi_md);
1133         if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE) {
1134                 LASSERT(oti);
1135                 LASSERT(oti->oti_logcookies);
1136         }
1137
1138         if (!exp || !exp->exp_obd)
1139                 RETURN(-ENODEV);
1140
1141         lov = &exp->exp_obd->u.lov;
1142         rc = lov_prep_setattr_set(exp, oinfo, oti, &set);
1143         if (rc)
1144                 RETURN(rc);
1145
1146         CDEBUG(D_INFO, "objid "DOSTID": %ux%u byte stripes\n",
1147                POSTID(&oinfo->oi_md->lsm_oi),
1148                oinfo->oi_md->lsm_stripe_count,
1149                oinfo->oi_md->lsm_stripe_size);
1150
1151         list_for_each(pos, &set->set_list) {
1152                 req = list_entry(pos, struct lov_request, rq_link);
1153
1154                 if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
1155                         oti->oti_logcookies = set->set_cookies + req->rq_stripe;
1156
1157                 CDEBUG(D_INFO, "objid "DOSTID"[%d] has subobj "DOSTID" at idx"
1158                        "%u\n", POSTID(&oinfo->oi_oa->o_oi), req->rq_stripe,
1159                        POSTID(&req->rq_oi.oi_oa->o_oi), req->rq_idx);
1160
1161                 rc = obd_setattr_async(lov->lov_tgts[req->rq_idx]->ltd_exp,
1162                                        &req->rq_oi, oti, rqset);
1163                 if (rc) {
1164                         CERROR("error: setattr objid "DOSTID" subobj"
1165                                DOSTID" on OST idx %d: rc = %d\n",
1166                                POSTID(&set->set_oi->oi_oa->o_oi),
1167                                POSTID(&req->rq_oi.oi_oa->o_oi),
1168                                req->rq_idx, rc);
1169                         break;
1170                 }
1171         }
1172
1173         /* If we are not waiting for responses on async requests, return. */
1174         if (rc || !rqset || list_empty(&rqset->set_requests)) {
1175                 int err;
1176                 if (rc)
1177                         atomic_set(&set->set_completes, 0);
1178                 err = lov_fini_setattr_set(set);
1179                 RETURN(rc ? rc : err);
1180         }
1181
1182         LASSERT(rqset->set_interpret == NULL);
1183         rqset->set_interpret = lov_setattr_interpret;
1184         rqset->set_arg = (void *)set;
1185
1186         RETURN(0);
1187 }
1188
1189 int lov_statfs_interpret(struct ptlrpc_request_set *rqset, void *data, int rc)
1190 {
1191         struct lov_request_set *lovset = (struct lov_request_set *)data;
1192         int err;
1193         ENTRY;
1194
1195         if (rc)
1196                 atomic_set(&lovset->set_completes, 0);
1197
1198         err = lov_fini_statfs_set(lovset);
1199         RETURN(rc ? rc : err);
1200 }
1201
1202 static int lov_statfs_async(struct obd_export *exp, struct obd_info *oinfo,
1203                             __u64 max_age, struct ptlrpc_request_set *rqset)
1204 {
1205         struct obd_device      *obd = class_exp2obd(exp);
1206         struct lov_request_set *set;
1207         struct lov_request *req;
1208         struct list_head *pos;
1209         struct lov_obd *lov;
1210         int rc = 0;
1211         ENTRY;
1212
1213         LASSERT(oinfo != NULL);
1214         LASSERT(oinfo->oi_osfs != NULL);
1215
1216         lov = &obd->u.lov;
1217         rc = lov_prep_statfs_set(obd, oinfo, &set);
1218         if (rc)
1219                 RETURN(rc);
1220
1221         list_for_each(pos, &set->set_list) {
1222                 req = list_entry(pos, struct lov_request, rq_link);
1223                 rc = obd_statfs_async(lov->lov_tgts[req->rq_idx]->ltd_exp,
1224                                       &req->rq_oi, max_age, rqset);
1225                 if (rc)
1226                         break;
1227         }
1228
1229         if (rc || list_empty(&rqset->set_requests)) {
1230                 int err;
1231                 if (rc)
1232                         atomic_set(&set->set_completes, 0);
1233                 err = lov_fini_statfs_set(set);
1234                 RETURN(rc ? rc : err);
1235         }
1236
1237         LASSERT(rqset->set_interpret == NULL);
1238         rqset->set_interpret = lov_statfs_interpret;
1239         rqset->set_arg = (void *)set;
1240         RETURN(0);
1241 }
1242
1243 static int lov_statfs(const struct lu_env *env, struct obd_export *exp,
1244                       struct obd_statfs *osfs, __u64 max_age, __u32 flags)
1245 {
1246         struct ptlrpc_request_set *set = NULL;
1247         struct obd_info oinfo = { { { 0 } } };
1248         int rc = 0;
1249         ENTRY;
1250
1251
1252         /* for obdclass we forbid using obd_statfs_rqset, but prefer using async
1253          * statfs requests */
1254         set = ptlrpc_prep_set();
1255         if (set == NULL)
1256                 RETURN(-ENOMEM);
1257
1258         oinfo.oi_osfs = osfs;
1259         oinfo.oi_flags = flags;
1260         rc = lov_statfs_async(exp, &oinfo, max_age, set);
1261         if (rc == 0)
1262                 rc = ptlrpc_set_wait(set);
1263         ptlrpc_set_destroy(set);
1264
1265         RETURN(rc);
1266 }
1267
1268 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1269                          void *karg, void __user *uarg)
1270 {
1271         struct obd_device *obddev = class_exp2obd(exp);
1272         struct lov_obd *lov = &obddev->u.lov;
1273         int i = 0, rc = 0, count = lov->desc.ld_tgt_count;
1274         struct obd_uuid *uuidp;
1275         ENTRY;
1276
1277         switch (cmd) {
1278         case IOC_OBD_STATFS: {
1279                 struct obd_ioctl_data *data = karg;
1280                 struct obd_device *osc_obd;
1281                 struct obd_statfs stat_buf = {0};
1282                 __u32 index;
1283                 __u32 flags;
1284
1285                 memcpy(&index, data->ioc_inlbuf2, sizeof(__u32));
1286                 if ((index >= count))
1287                         RETURN(-ENODEV);
1288
1289                 if (!lov->lov_tgts[index])
1290                         /* Try again with the next index */
1291                         RETURN(-EAGAIN);
1292                 if (!lov->lov_tgts[index]->ltd_active)
1293                         RETURN(-ENODATA);
1294
1295                 osc_obd = class_exp2obd(lov->lov_tgts[index]->ltd_exp);
1296                 if (!osc_obd)
1297                         RETURN(-EINVAL);
1298
1299                 /* copy UUID */
1300                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(osc_obd),
1301                                  min((int)data->ioc_plen2,
1302                                      (int)sizeof(struct obd_uuid))))
1303                         RETURN(-EFAULT);
1304
1305                 flags = uarg ? *(__u32 __user *)uarg : 0;
1306                 /* got statfs data */
1307                 rc = obd_statfs(NULL, lov->lov_tgts[index]->ltd_exp, &stat_buf,
1308                                 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
1309                                 flags);
1310                 if (rc)
1311                         RETURN(rc);
1312                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
1313                                      min((int) data->ioc_plen1,
1314                                          (int) sizeof(stat_buf))))
1315                         RETURN(-EFAULT);
1316                 break;
1317         }
1318         case OBD_IOC_LOV_GET_CONFIG: {
1319                 struct obd_ioctl_data *data;
1320                 struct lov_desc *desc;
1321                 char *buf = NULL;
1322                 __u32 *genp;
1323
1324                 len = 0;
1325                 if (obd_ioctl_getdata(&buf, &len, uarg))
1326                         RETURN(-EINVAL);
1327
1328                 data = (struct obd_ioctl_data *)buf;
1329
1330                 if (sizeof(*desc) > data->ioc_inllen1) {
1331                         obd_ioctl_freedata(buf, len);
1332                         RETURN(-EINVAL);
1333                 }
1334
1335                 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
1336                         obd_ioctl_freedata(buf, len);
1337                         RETURN(-EINVAL);
1338                 }
1339
1340                 if (sizeof(__u32) * count > data->ioc_inllen3) {
1341                         obd_ioctl_freedata(buf, len);
1342                         RETURN(-EINVAL);
1343                 }
1344
1345                 desc = (struct lov_desc *)data->ioc_inlbuf1;
1346                 memcpy(desc, &(lov->desc), sizeof(*desc));
1347
1348                 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
1349                 genp = (__u32 *)data->ioc_inlbuf3;
1350                 /* the uuid will be empty for deleted OSTs */
1351                 for (i = 0; i < count; i++, uuidp++, genp++) {
1352                         if (!lov->lov_tgts[i])
1353                                 continue;
1354                         *uuidp = lov->lov_tgts[i]->ltd_uuid;
1355                         *genp = lov->lov_tgts[i]->ltd_gen;
1356                 }
1357
1358                 if (copy_to_user(uarg, buf, len))
1359                         rc = -EFAULT;
1360                 obd_ioctl_freedata(buf, len);
1361                 break;
1362         }
1363         case OBD_IOC_QUOTACTL: {
1364                 struct if_quotactl *qctl = karg;
1365                 struct lov_tgt_desc *tgt = NULL;
1366                 struct obd_quotactl *oqctl;
1367
1368                 if (qctl->qc_valid == QC_OSTIDX) {
1369                         if (count <= qctl->qc_idx)
1370                                 RETURN(-EINVAL);
1371
1372                         tgt = lov->lov_tgts[qctl->qc_idx];
1373                         if (!tgt || !tgt->ltd_exp)
1374                                 RETURN(-EINVAL);
1375                 } else if (qctl->qc_valid == QC_UUID) {
1376                         for (i = 0; i < count; i++) {
1377                                 tgt = lov->lov_tgts[i];
1378                                 if (!tgt ||
1379                                     !obd_uuid_equals(&tgt->ltd_uuid,
1380                                                      &qctl->obd_uuid))
1381                                         continue;
1382
1383                                 if (tgt->ltd_exp == NULL)
1384                                         RETURN(-EINVAL);
1385
1386                                 break;
1387                         }
1388                 } else {
1389                         RETURN(-EINVAL);
1390                 }
1391
1392                 if (i >= count)
1393                         RETURN(-EAGAIN);
1394
1395                 LASSERT(tgt && tgt->ltd_exp);
1396                 OBD_ALLOC_PTR(oqctl);
1397                 if (!oqctl)
1398                         RETURN(-ENOMEM);
1399
1400                 QCTL_COPY(oqctl, qctl);
1401                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
1402                 if (rc == 0) {
1403                         QCTL_COPY(qctl, oqctl);
1404                         qctl->qc_valid = QC_OSTIDX;
1405                         qctl->obd_uuid = tgt->ltd_uuid;
1406                 }
1407                 OBD_FREE_PTR(oqctl);
1408                 break;
1409         }
1410         default: {
1411                 int set = 0;
1412
1413                 if (count == 0)
1414                         RETURN(-ENOTTY);
1415
1416                 for (i = 0; i < count; i++) {
1417                         int err;
1418                         struct obd_device *osc_obd;
1419
1420                         /* OST was disconnected */
1421                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
1422                                 continue;
1423
1424                         /* ll_umount_begin() sets force flag but for lov, not
1425                          * osc. Let's pass it through */
1426                         osc_obd = class_exp2obd(lov->lov_tgts[i]->ltd_exp);
1427                         osc_obd->obd_force = obddev->obd_force;
1428                         err = obd_iocontrol(cmd, lov->lov_tgts[i]->ltd_exp,
1429                                             len, karg, uarg);
1430                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
1431                                 RETURN(err);
1432                         } else if (err) {
1433                                 if (lov->lov_tgts[i]->ltd_active) {
1434                                         CDEBUG(err == -ENOTTY ?
1435                                                D_IOCTL : D_WARNING,
1436                                                "iocontrol OSC %s on OST "
1437                                                "idx %d cmd %x: err = %d\n",
1438                                                lov_uuid2str(lov, i),
1439                                                i, cmd, err);
1440                                         if (!rc)
1441                                                 rc = err;
1442                                 }
1443                         } else {
1444                                 set = 1;
1445                         }
1446                 }
1447                 if (!set && !rc)
1448                         rc = -EIO;
1449         }
1450         }
1451
1452         RETURN(rc);
1453 }
1454
1455 static int lov_get_info(const struct lu_env *env, struct obd_export *exp,
1456                         __u32 keylen, void *key,
1457                         __u32 *vallen, void *val,
1458                         struct lov_stripe_md *lsm)
1459 {
1460         struct obd_device *obddev = class_exp2obd(exp);
1461         struct lov_obd *lov = &obddev->u.lov;
1462         int rc;
1463         ENTRY;
1464
1465         if (!vallen || !val)
1466                 RETURN(-EFAULT);
1467
1468         obd_getref(obddev);
1469
1470         if (KEY_IS(KEY_LOVDESC)) {
1471                 struct lov_desc *desc_ret = val;
1472                 *desc_ret = lov->desc;
1473
1474                 GOTO(out, rc = 0);
1475         } else if (KEY_IS(KEY_TGT_COUNT)) {
1476                 *((int *)val) = lov->desc.ld_tgt_count;
1477                 GOTO(out, rc = 0);
1478         }
1479
1480         rc = -EINVAL;
1481
1482 out:
1483         obd_putref(obddev);
1484         RETURN(rc);
1485 }
1486
1487 static int lov_set_info_async(const struct lu_env *env, struct obd_export *exp,
1488                               __u32 keylen, void *key,
1489                               __u32 vallen, void *val,
1490                               struct ptlrpc_request_set *set)
1491 {
1492         struct obd_device *obddev = class_exp2obd(exp);
1493         struct lov_obd *lov = &obddev->u.lov;
1494         struct lov_tgt_desc *tgt;
1495         int do_inactive = 0;
1496         int no_set = 0;
1497         u32 count;
1498         u32 i;
1499         int rc = 0;
1500         int err;
1501         ENTRY;
1502
1503         if (set == NULL) {
1504                 no_set = 1;
1505                 set = ptlrpc_prep_set();
1506                 if (!set)
1507                         RETURN(-ENOMEM);
1508         }
1509
1510         obd_getref(obddev);
1511         count = lov->desc.ld_tgt_count;
1512
1513         if (KEY_IS(KEY_CHECKSUM)) {
1514                 do_inactive = 1;
1515         } else if (KEY_IS(KEY_CACHE_SET)) {
1516                 LASSERT(lov->lov_cache == NULL);
1517                 lov->lov_cache = val;
1518                 do_inactive = 1;
1519                 cl_cache_incref(lov->lov_cache);
1520         }
1521
1522         for (i = 0; i < count; i++) {
1523                 tgt = lov->lov_tgts[i];
1524
1525                 /* OST was disconnected */
1526                 if (!tgt || !tgt->ltd_exp)
1527                         continue;
1528
1529                 /* OST is inactive and we don't want inactive OSCs */
1530                 if (!tgt->ltd_active && !do_inactive)
1531                         continue;
1532
1533                 err = obd_set_info_async(env, tgt->ltd_exp, keylen, key,
1534                                          vallen, val, set);
1535                 if (!rc)
1536                         rc = err;
1537         }
1538
1539         obd_putref(obddev);
1540         if (no_set) {
1541                 err = ptlrpc_set_wait(set);
1542                 if (!rc)
1543                         rc = err;
1544                 ptlrpc_set_destroy(set);
1545         }
1546         RETURN(rc);
1547 }
1548
1549 void lov_stripe_lock(struct lov_stripe_md *md)
1550 __acquires(&md->lsm_lock)
1551 {
1552         LASSERT(md->lsm_lock_owner != current_pid());
1553         spin_lock(&md->lsm_lock);
1554         LASSERT(md->lsm_lock_owner == 0);
1555         md->lsm_lock_owner = current_pid();
1556 }
1557
1558 void lov_stripe_unlock(struct lov_stripe_md *md)
1559 __releases(&md->lsm_lock)
1560 {
1561         LASSERT(md->lsm_lock_owner == current_pid());
1562         md->lsm_lock_owner = 0;
1563         spin_unlock(&md->lsm_lock);
1564 }
1565
1566 static int lov_quotactl(struct obd_device *obd, struct obd_export *exp,
1567                         struct obd_quotactl *oqctl)
1568 {
1569         struct lov_obd      *lov = &obd->u.lov;
1570         struct lov_tgt_desc *tgt;
1571         __u64                curspace = 0;
1572         __u64                bhardlimit = 0;
1573         int                  i, rc = 0;
1574         ENTRY;
1575
1576         if (oqctl->qc_cmd != LUSTRE_Q_QUOTAON &&
1577             oqctl->qc_cmd != LUSTRE_Q_QUOTAOFF &&
1578             oqctl->qc_cmd != Q_GETOQUOTA &&
1579             oqctl->qc_cmd != Q_INITQUOTA &&
1580             oqctl->qc_cmd != LUSTRE_Q_SETQUOTA &&
1581             oqctl->qc_cmd != Q_FINVALIDATE) {
1582                 CERROR("%s: bad quota opc %x for lov obd\n",
1583                        obd->obd_name, oqctl->qc_cmd);
1584                 RETURN(-EFAULT);
1585         }
1586
1587         /* for lov tgt */
1588         obd_getref(obd);
1589         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1590                 int err;
1591
1592                 tgt = lov->lov_tgts[i];
1593
1594                 if (!tgt)
1595                         continue;
1596
1597                 if (!tgt->ltd_active || tgt->ltd_reap) {
1598                         if (oqctl->qc_cmd == Q_GETOQUOTA &&
1599                             lov->lov_tgts[i]->ltd_activate) {
1600                                 rc = -ENETDOWN;
1601                                 CERROR("ost %d is inactive\n", i);
1602                         } else {
1603                                 CDEBUG(D_HA, "ost %d is inactive\n", i);
1604                         }
1605                         continue;
1606                 }
1607
1608                 err = obd_quotactl(tgt->ltd_exp, oqctl);
1609                 if (err) {
1610                         if (tgt->ltd_active && !rc)
1611                                 rc = err;
1612                         continue;
1613                 }
1614
1615                 if (oqctl->qc_cmd == Q_GETOQUOTA) {
1616                         curspace += oqctl->qc_dqblk.dqb_curspace;
1617                         bhardlimit += oqctl->qc_dqblk.dqb_bhardlimit;
1618                 }
1619         }
1620         obd_putref(obd);
1621
1622         if (oqctl->qc_cmd == Q_GETOQUOTA) {
1623                 oqctl->qc_dqblk.dqb_curspace = curspace;
1624                 oqctl->qc_dqblk.dqb_bhardlimit = bhardlimit;
1625         }
1626         RETURN(rc);
1627 }
1628
1629 static int lov_quotacheck(struct obd_device *obd, struct obd_export *exp,
1630                           struct obd_quotactl *oqctl)
1631 {
1632         struct lov_obd *lov = &obd->u.lov;
1633         int             i, rc = 0;
1634         ENTRY;
1635
1636         obd_getref(obd);
1637
1638         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1639                 if (!lov->lov_tgts[i])
1640                         continue;
1641
1642                 /* Skip quota check on the administratively disabled OSTs. */
1643                 if (!lov->lov_tgts[i]->ltd_activate) {
1644                         CWARN("lov idx %d was administratively disabled, "
1645                               "skip quotacheck on it.\n", i);
1646                         continue;
1647                 }
1648
1649                 if (!lov->lov_tgts[i]->ltd_active) {
1650                         CERROR("lov idx %d inactive\n", i);
1651                         rc = -EIO;
1652                         goto out;
1653                 }
1654         }
1655
1656         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1657                 int err;
1658
1659                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_activate)
1660                         continue;
1661
1662                 err = obd_quotacheck(lov->lov_tgts[i]->ltd_exp, oqctl);
1663                 if (err && !rc)
1664                         rc = err;
1665         }
1666
1667 out:
1668         obd_putref(obd);
1669
1670         RETURN(rc);
1671 }
1672
1673 static struct obd_ops lov_obd_ops = {
1674         .o_owner                = THIS_MODULE,
1675         .o_setup                = lov_setup,
1676         .o_precleanup           = lov_precleanup,
1677         .o_cleanup              = lov_cleanup,
1678         .o_connect              = lov_connect,
1679         .o_disconnect           = lov_disconnect,
1680         .o_statfs               = lov_statfs,
1681         .o_statfs_async         = lov_statfs_async,
1682         .o_packmd               = lov_packmd,
1683         .o_unpackmd             = lov_unpackmd,
1684         .o_getattr_async        = lov_getattr_async,
1685         .o_setattr_async        = lov_setattr_async,
1686         .o_iocontrol            = lov_iocontrol,
1687         .o_get_info             = lov_get_info,
1688         .o_set_info_async       = lov_set_info_async,
1689         .o_notify               = lov_notify,
1690         .o_pool_new             = lov_pool_new,
1691         .o_pool_rem             = lov_pool_remove,
1692         .o_pool_add             = lov_pool_add,
1693         .o_pool_del             = lov_pool_del,
1694         .o_getref               = lov_getref,
1695         .o_putref               = lov_putref,
1696         .o_quotactl             = lov_quotactl,
1697         .o_quotacheck           = lov_quotacheck,
1698 };
1699
1700 struct kmem_cache *lov_oinfo_slab;
1701
1702 static int __init lov_init(void)
1703 {
1704         bool enable_proc = true;
1705         struct obd_type *type;
1706         int rc;
1707         ENTRY;
1708
1709         /* print an address of _any_ initialized kernel symbol from this
1710          * module, to allow debugging with gdb that doesn't support data
1711          * symbols from modules.*/
1712         CDEBUG(D_INFO, "Lustre LOV module (%p).\n", &lov_caches);
1713
1714         rc = lu_kmem_init(lov_caches);
1715         if (rc)
1716                 return rc;
1717
1718         lov_oinfo_slab = kmem_cache_create("lov_oinfo",
1719                                            sizeof(struct lov_oinfo), 0,
1720                                            SLAB_HWCACHE_ALIGN, NULL);
1721         if (lov_oinfo_slab == NULL) {
1722                 lu_kmem_fini(lov_caches);
1723                 return -ENOMEM;
1724         }
1725
1726         type = class_search_type(LUSTRE_LOD_NAME);
1727         if (type != NULL && type->typ_procsym != NULL)
1728                 enable_proc = false;
1729
1730         rc = class_register_type(&lov_obd_ops, NULL, enable_proc, NULL,
1731                                  LUSTRE_LOV_NAME, &lov_device_type);
1732
1733         if (rc) {
1734                 kmem_cache_destroy(lov_oinfo_slab);
1735                 lu_kmem_fini(lov_caches);
1736         }
1737
1738         RETURN(rc);
1739 }
1740
1741 static void /*__exit*/ lov_exit(void)
1742 {
1743         class_unregister_type(LUSTRE_LOV_NAME);
1744         kmem_cache_destroy(lov_oinfo_slab);
1745         lu_kmem_fini(lov_caches);
1746 }
1747
1748 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
1749 MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver");
1750 MODULE_LICENSE("GPL");
1751
1752 cfs_module(lov, LUSTRE_VERSION_STRING, lov_init, lov_exit);