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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/lov/lov_obd.c
  33  *
  34  * Author: Phil Schwan <phil@clusterfs.com>
  35  * Author: Peter Braam <braam@clusterfs.com>
  36  * Author: Mike Shaver <shaver@clusterfs.com>
  37  * Author: Nathan Rutman <nathan@clusterfs.com>
  38  */
  39
  40 #define DEBUG_SUBSYSTEM S_LOV
  41 #include <libcfs/libcfs.h>
  42
  43 #include <lustre/lustre_idl.h>
  44
  45 #include <cl_object.h>
  46 #include <lustre_dlm.h>
  47 #include <lustre_fid.h>
  48 #include <lustre_ioctl.h>
  49 #include <lustre_lib.h>
  50 #include <lustre_mds.h>
  51 #include <lustre_net.h>
  52 #include <lustre_param.h>
  53 #include <lustre_swab.h>
  54 #include <lprocfs_status.h>
  55 #include <obd_class.h>
  56 #include <obd_support.h>
  57
  58 #include "lov_internal.h"
  59
  60 /* Keep a refcount of lov->tgt usage to prevent racing with addition/deletion.
  61    Any function that expects lov_tgts to remain stationary must take a ref. */
  62 static void lov_getref(struct obd_device *obd)
  63 {
  64         struct lov_obd *lov = &obd->u.lov;
  65
  66         /* nobody gets through here until lov_putref is done */
  67         mutex_lock(&lov->lov_lock);
  68         atomic_inc(&lov->lov_refcount);
  69         mutex_unlock(&lov->lov_lock);
  70         return;
  71 }
  72
  73 static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt);
  74
  75 static void lov_putref(struct obd_device *obd)
  76 {
  77         struct lov_obd *lov = &obd->u.lov;
  78
  79         mutex_lock(&lov->lov_lock);
  80         /* ok to dec to 0 more than once -- ltd_exp's will be null */
  81         if (atomic_dec_and_test(&lov->lov_refcount) && lov->lov_death_row) {
  82                 struct list_head kill = LIST_HEAD_INIT(kill);
  83                 struct lov_tgt_desc *tgt, *n;
  84                 int i;
  85
  86                 CDEBUG(D_CONFIG, "destroying %d lov targets\n",
  87                        lov->lov_death_row);
  88                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
  89                         tgt = lov->lov_tgts[i];
  90
  91                         if (!tgt || !tgt->ltd_reap)
  92                                 continue;
  93                         list_add(&tgt->ltd_kill, &kill);
  94                         /* XXX - right now there is a dependency on ld_tgt_count
  95                          * being the maximum tgt index for computing the
  96                          * mds_max_easize. So we can't shrink it. */
  97                         lov_ost_pool_remove(&lov->lov_packed, i);
  98                         lov->lov_tgts[i] = NULL;
  99                         lov->lov_death_row--;
 100                 }
 101                 mutex_unlock(&lov->lov_lock);
 102
 103                 list_for_each_entry_safe(tgt, n, &kill, ltd_kill) {
 104                         list_del(&tgt->ltd_kill);
 105                         /* Disconnect */
 106                         __lov_del_obd(obd, tgt);
 107                 }
 108         } else {
 109                 mutex_unlock(&lov->lov_lock);
 110         }
 111 }
 112
 113 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
 114                               enum obd_notify_event ev);
 115 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
 116                       enum obd_notify_event ev, void *data);
 117
 118 int lov_connect_obd(struct obd_device *obd, __u32 index, int activate,
 119                     struct obd_connect_data *data)
 120 {
 121         struct lov_obd *lov = &obd->u.lov;
 122         struct obd_uuid *tgt_uuid;
 123         struct obd_device *tgt_obd;
 124         static struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
 125         struct obd_import *imp;
 126         int rc;
 127         ENTRY;
 128
 129         if (lov->lov_tgts[index] == NULL)
 130                 RETURN(-EINVAL);
 131
 132         tgt_uuid = &lov->lov_tgts[index]->ltd_uuid;
 133         tgt_obd = lov->lov_tgts[index]->ltd_obd;
 134
 135         if (!tgt_obd->obd_set_up) {
 136                 CERROR("Target %s not set up\n", obd_uuid2str(tgt_uuid));
 137                 RETURN(-EINVAL);
 138         }
 139
 140         /* override the sp_me from lov */
 141         tgt_obd->u.cli.cl_sp_me = lov->lov_sp_me;
 142
 143         if (data && (data->ocd_connect_flags & OBD_CONNECT_INDEX))
 144                 data->ocd_index = index;
 145
 146         /*
 147          * Divine LOV knows that OBDs under it are OSCs.
 148          */
 149         imp = tgt_obd->u.cli.cl_import;
 150
 151         if (activate) {
 152                 tgt_obd->obd_no_recov = 0;
 153                 /* FIXME this is probably supposed to be
 154                    ptlrpc_set_import_active.  Horrible naming. */
 155                 ptlrpc_activate_import(imp);
 156         }
 157
 158         rc = obd_register_observer(tgt_obd, obd);
 159         if (rc) {
 160                 CERROR("Target %s register_observer error %d\n",
 161                        obd_uuid2str(tgt_uuid), rc);
 162                 RETURN(rc);
 163         }
 164
 165
 166         if (imp->imp_invalid) {
 167                 CDEBUG(D_CONFIG, "not connecting OSC %s; administratively "
 168                        "disabled\n", obd_uuid2str(tgt_uuid));
 169                 RETURN(0);
 170         }
 171
 172         rc = obd_connect(NULL, &lov->lov_tgts[index]->ltd_exp, tgt_obd,
 173                          &lov_osc_uuid, data, NULL);
 174         if (rc || !lov->lov_tgts[index]->ltd_exp) {
 175                 CERROR("Target %s connect error %d\n",
 176                        obd_uuid2str(tgt_uuid), rc);
 177                 RETURN(-ENODEV);
 178         }
 179
 180         lov->lov_tgts[index]->ltd_reap = 0;
 181
 182         CDEBUG(D_CONFIG, "Connected tgt idx %d %s (%s) %sactive\n", index,
 183                obd_uuid2str(tgt_uuid), tgt_obd->obd_name, activate ? "":"in");
 184
 185         if (lov->targets_proc_entry != NULL) {
 186                 struct proc_dir_entry *osc_symlink;
 187                 struct obd_device *osc_obd;
 188
 189                 osc_obd = lov->lov_tgts[index]->ltd_exp->exp_obd;
 190
 191                 LASSERT(osc_obd != NULL);
 192                 LASSERT(osc_obd->obd_magic == OBD_DEVICE_MAGIC);
 193                 LASSERT(osc_obd->obd_type->typ_name != NULL);
 194
 195                 osc_symlink = lprocfs_add_symlink(osc_obd->obd_name,
 196                                                   lov->targets_proc_entry,
 197                                                   "../../../%s/%s",
 198                                                   osc_obd->obd_type->typ_name,
 199                                                   osc_obd->obd_name);
 200                 if (osc_symlink == NULL) {
 201                         CERROR("cannot register LOV target "
 202                                "/proc/fs/lustre/%s/%s/target_obds/%s\n",
 203                                obd->obd_type->typ_name, obd->obd_name,
 204                                osc_obd->obd_name);
 205                 }
 206         }
 207         RETURN(0);
 208 }
 209
 210 static int lov_connect(const struct lu_env *env,
 211                        struct obd_export **exp, struct obd_device *obd,
 212                        struct obd_uuid *cluuid, struct obd_connect_data *data,
 213                        void *localdata)
 214 {
 215         struct lov_obd *lov = &obd->u.lov;
 216         struct lov_tgt_desc *tgt;
 217         struct lustre_handle conn;
 218         int i, rc;
 219         ENTRY;
 220
 221         CDEBUG(D_CONFIG, "connect #%d\n", lov->lov_connects);
 222
 223         rc = class_connect(&conn, obd, cluuid);
 224         if (rc)
 225                 RETURN(rc);
 226
 227         *exp = class_conn2export(&conn);
 228
 229         /* Why should there ever be more than 1 connect? */
 230         lov->lov_connects++;
 231         LASSERT(lov->lov_connects == 1);
 232
 233         memset(&lov->lov_ocd, 0, sizeof(lov->lov_ocd));
 234         if (data)
 235                 lov->lov_ocd = *data;
 236
 237         lov->targets_proc_entry = lprocfs_register("target_obds",
 238                                                    obd->obd_proc_entry,
 239                                                    NULL, NULL);
 240         if (IS_ERR(lov->targets_proc_entry)) {
 241                 CERROR("%s: cannot register "
 242                        "/proc/fs/lustre/%s/%s/target_obds\n",
 243                        obd->obd_name, obd->obd_type->typ_name, obd->obd_name);
 244                 lov->targets_proc_entry = NULL;
 245         }
 246
 247         obd_getref(obd);
 248         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 249                 tgt = lov->lov_tgts[i];
 250                 if (!tgt || obd_uuid_empty(&tgt->ltd_uuid))
 251                         continue;
 252                 /* Flags will be lowest common denominator */
 253                 rc = lov_connect_obd(obd, i, tgt->ltd_activate, &lov->lov_ocd);
 254                 if (rc) {
 255                         CERROR("%s: lov connect tgt %d failed: %d\n",
 256                                obd->obd_name, i, rc);
 257                         continue;
 258                 }
 259                 /* connect to administrative disabled ost */
 260                 if (!lov->lov_tgts[i]->ltd_exp)
 261                         continue;
 262
 263                 rc = lov_notify(obd, lov->lov_tgts[i]->ltd_exp->exp_obd,
 264                                 OBD_NOTIFY_CONNECT, (void *)&i);
 265                 if (rc) {
 266                         CERROR("%s error sending notify %d\n",
 267                                obd->obd_name, rc);
 268                 }
 269         }
 270         obd_putref(obd);
 271
 272         RETURN(0);
 273 }
 274
 275 static int lov_disconnect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
 276 {
 277         struct lov_obd *lov = &obd->u.lov;
 278         struct obd_device *osc_obd;
 279         int rc;
 280         ENTRY;
 281
 282         osc_obd = class_exp2obd(tgt->ltd_exp);
 283         CDEBUG(D_CONFIG, "%s: disconnecting target %s\n",
 284                obd->obd_name, osc_obd->obd_name);
 285
 286         if (tgt->ltd_active) {
 287                 tgt->ltd_active = 0;
 288                 lov->desc.ld_active_tgt_count--;
 289                 tgt->ltd_exp->exp_obd->obd_inactive = 1;
 290         }
 291
 292         if (osc_obd) {
 293                 /* Pass it on to our clients.
 294                  * XXX This should be an argument to disconnect,
 295                  * XXX not a back-door flag on the OBD.  Ah well.
 296                  */
 297                 osc_obd->obd_force = obd->obd_force;
 298                 osc_obd->obd_fail = obd->obd_fail;
 299                 osc_obd->obd_no_recov = obd->obd_no_recov;
 300
 301                 if (lov->targets_proc_entry != NULL)
 302                         lprocfs_remove_proc_entry(osc_obd->obd_name,
 303                                                   lov->targets_proc_entry);
 304         }
 305
 306         obd_register_observer(osc_obd, NULL);
 307
 308         rc = obd_disconnect(tgt->ltd_exp);
 309         if (rc) {
 310                 CERROR("Target %s disconnect error %d\n",
 311                        tgt->ltd_uuid.uuid, rc);
 312                 rc = 0;
 313         }
 314
 315         tgt->ltd_exp = NULL;
 316         RETURN(0);
 317 }
 318
 319 static int lov_disconnect(struct obd_export *exp)
 320 {
 321         struct obd_device *obd = class_exp2obd(exp);
 322         struct lov_obd *lov = &obd->u.lov;
 323         int i, rc;
 324         ENTRY;
 325
 326         if (!lov->lov_tgts)
 327                 goto out;
 328
 329         /* Only disconnect the underlying layers on the final disconnect. */
 330         lov->lov_connects--;
 331         if (lov->lov_connects != 0) {
 332                 /* why should there be more than 1 connect? */
 333                 CERROR("disconnect #%d\n", lov->lov_connects);
 334                 goto out;
 335         }
 336
 337         /* Let's hold another reference so lov_del_obd doesn't spin through
 338            putref every time */
 339         obd_getref(obd);
 340
 341         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 342                 if (lov->lov_tgts[i] && lov->lov_tgts[i]->ltd_exp) {
 343                         /* Disconnection is the last we know about an obd */
 344                         lov_del_target(obd, i, NULL, lov->lov_tgts[i]->ltd_gen);
 345                 }
 346         }
 347         obd_putref(obd);
 348
 349         if (lov->targets_proc_entry != NULL)
 350                 lprocfs_remove(&lov->targets_proc_entry);
 351
 352 out:
 353         rc = class_disconnect(exp); /* bz 9811 */
 354         RETURN(rc);
 355 }
 356
 357 /* Error codes:
 358  *
 359  *  -EINVAL  : UUID can't be found in the LOV's target list
 360  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
 361  *  -EBADF   : The UUID is found, but the OBD is the wrong type (!)
 362  *  any >= 0 : is log target index
 363  */
 364 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
 365                               enum obd_notify_event ev)
 366 {
 367         struct lov_obd *lov = &obd->u.lov;
 368         struct lov_tgt_desc *tgt;
 369         int index, activate, active;
 370         ENTRY;
 371
 372         CDEBUG(D_INFO, "Searching in lov %p for uuid %s event(%d)\n",
 373                lov, uuid->uuid, ev);
 374
 375         obd_getref(obd);
 376         for (index = 0; index < lov->desc.ld_tgt_count; index++) {
 377                 tgt = lov->lov_tgts[index];
 378                 if (!tgt)
 379                         continue;
 380                 /*
 381                  * LU-642, initially inactive OSC could miss the obd_connect,
 382                  * we make up for it here.
 383                  */
 384                 if (ev == OBD_NOTIFY_ACTIVATE && tgt->ltd_exp == NULL &&
 385                     obd_uuid_equals(uuid, &tgt->ltd_uuid)) {
 386                         struct obd_uuid lov_osc_uuid = {"LOV_OSC_UUID"};
 387
 388                         obd_connect(NULL, &tgt->ltd_exp, tgt->ltd_obd,
 389                                     &lov_osc_uuid, &lov->lov_ocd, NULL);
 390                 }
 391                 if (!tgt->ltd_exp)
 392                         continue;
 393
 394                 CDEBUG(D_INFO, "lov idx %d is %s conn %#llx\n",
 395                        index, obd_uuid2str(&tgt->ltd_uuid),
 396                        tgt->ltd_exp->exp_handle.h_cookie);
 397                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 398                         break;
 399         }
 400
 401         if (index == lov->desc.ld_tgt_count)
 402                 GOTO(out, index = -EINVAL);
 403
 404         if (ev == OBD_NOTIFY_DEACTIVATE || ev == OBD_NOTIFY_ACTIVATE) {
 405                 activate = (ev == OBD_NOTIFY_ACTIVATE) ? 1 : 0;
 406
 407                 if (lov->lov_tgts[index]->ltd_activate == activate) {
 408                         CDEBUG(D_INFO, "OSC %s already %sactivate!\n",
 409                                uuid->uuid, activate ? "" : "de");
 410                 } else {
 411                         lov->lov_tgts[index]->ltd_activate = activate;
 412                         CDEBUG(D_CONFIG, "%sactivate OSC %s\n",
 413                                activate ? "" : "de", obd_uuid2str(uuid));
 414                 }
 415
 416         } else if (ev == OBD_NOTIFY_INACTIVE || ev == OBD_NOTIFY_ACTIVE) {
 417                 active = (ev == OBD_NOTIFY_ACTIVE) ? 1 : 0;
 418
 419                 if (lov->lov_tgts[index]->ltd_active == active) {
 420                         CDEBUG(D_INFO, "OSC %s already %sactive!\n",
 421                                uuid->uuid, active ? "" : "in");
 422                         GOTO(out, index);
 423                 } else {
 424                         CDEBUG(D_CONFIG, "Marking OSC %s %sactive\n",
 425                                obd_uuid2str(uuid), active ? "" : "in");
 426                 }
 427
 428                 lov->lov_tgts[index]->ltd_active = active;
 429                 if (active) {
 430                         lov->desc.ld_active_tgt_count++;
 431                         lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 0;
 432                 } else {
 433                         lov->desc.ld_active_tgt_count--;
 434                         lov->lov_tgts[index]->ltd_exp->exp_obd->obd_inactive = 1;
 435                 }
 436         } else {
 437                 CERROR("%s: unknown event %d for uuid %s\n", obd->obd_name,
 438                        ev, uuid->uuid);
 439         }
 440
 441  out:
 442         obd_putref(obd);
 443         RETURN(index);
 444 }
 445
 446 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
 447                       enum obd_notify_event ev, void *data)
 448 {
 449         int rc = 0;
 450         struct lov_obd *lov = &obd->u.lov;
 451         ENTRY;
 452
 453         down_read(&lov->lov_notify_lock);
 454         if (!lov->lov_connects) {
 455                 up_read(&lov->lov_notify_lock);
 456                 RETURN(rc);
 457         }
 458
 459         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE ||
 460             ev == OBD_NOTIFY_ACTIVATE || ev == OBD_NOTIFY_DEACTIVATE) {
 461                 struct obd_uuid *uuid;
 462
 463                 LASSERT(watched);
 464
 465                 if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) {
 466                         up_read(&lov->lov_notify_lock);
 467                         CERROR("unexpected notification of %s %s!\n",
 468                                watched->obd_type->typ_name,
 469                                watched->obd_name);
 470                         RETURN(-EINVAL);
 471                 }
 472                 uuid = &watched->u.cli.cl_target_uuid;
 473
 474                 /* Set OSC as active before notifying the observer, so the
 475                  * observer can use the OSC normally.
 476                  */
 477                 rc = lov_set_osc_active(obd, uuid, ev);
 478                 if (rc < 0) {
 479                         up_read(&lov->lov_notify_lock);
 480                         CERROR("event(%d) of %s failed: %d\n", ev,
 481                                obd_uuid2str(uuid), rc);
 482                         RETURN(rc);
 483                 }
 484                 /* active event should be pass lov target index as data */
 485                 data = &rc;
 486         }
 487
 488         /* Pass the notification up the chain. */
 489         if (watched) {
 490                 rc = obd_notify_observer(obd, watched, ev, data);
 491         } else {
 492                 /* NULL watched means all osc's in the lov (only for syncs) */
 493                 /* sync event should be send lov idx as data */
 494                 struct lov_obd *lov = &obd->u.lov;
 495                 int i, is_sync;
 496
 497                 data = &i;
 498                 is_sync = (ev == OBD_NOTIFY_SYNC) ||
 499                           (ev == OBD_NOTIFY_SYNC_NONBLOCK);
 500
 501                 obd_getref(obd);
 502                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 503                         if (!lov->lov_tgts[i])
 504                                 continue;
 505
 506                         /* don't send sync event if target not
 507                          * connected/activated */
 508                         if (is_sync &&  !lov->lov_tgts[i]->ltd_active)
 509                                 continue;
 510
 511                         rc = obd_notify_observer(obd, lov->lov_tgts[i]->ltd_obd,
 512                                                  ev, data);
 513                         if (rc) {
 514                                 CERROR("%s: notify %s of %s failed %d\n",
 515                                        obd->obd_name,
 516                                        obd->obd_observer->obd_name,
 517                                        lov->lov_tgts[i]->ltd_obd->obd_name,
 518                                        rc);
 519                         }
 520                 }
 521                 obd_putref(obd);
 522         }
 523
 524         up_read(&lov->lov_notify_lock);
 525         RETURN(rc);
 526 }
 527
 528 static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 529                           __u32 index, int gen, int active)
 530 {
 531         struct lov_obd *lov = &obd->u.lov;
 532         struct lov_tgt_desc *tgt;
 533         struct obd_device *tgt_obd;
 534         int rc;
 535         ENTRY;
 536
 537         CDEBUG(D_CONFIG, "uuid:%s idx:%d gen:%d active:%d\n",
 538                uuidp->uuid, index, gen, active);
 539
 540         if (gen <= 0) {
 541                 CERROR("request to add OBD %s with invalid generation: %d\n",
 542                        uuidp->uuid, gen);
 543                 RETURN(-EINVAL);
 544         }
 545
 546         tgt_obd = class_find_client_obd(uuidp, LUSTRE_OSC_NAME,
 547                                         &obd->obd_uuid);
 548         if (tgt_obd == NULL)
 549                 RETURN(-EINVAL);
 550
 551         mutex_lock(&lov->lov_lock);
 552
 553         if ((index < lov->lov_tgt_size) && (lov->lov_tgts[index] != NULL)) {
 554                 tgt = lov->lov_tgts[index];
 555                 CERROR("UUID %s already assigned at LOV target index %d\n",
 556                        obd_uuid2str(&tgt->ltd_uuid), index);
 557                 mutex_unlock(&lov->lov_lock);
 558                 RETURN(-EEXIST);
 559         }
 560
 561         if (index >= lov->lov_tgt_size) {
 562                 /* We need to reallocate the lov target array. */
 563                 struct lov_tgt_desc **newtgts, **old = NULL;
 564                 __u32 newsize, oldsize = 0;
 565
 566                 newsize = max(lov->lov_tgt_size, (__u32)2);
 567                 while (newsize < index + 1)
 568                         newsize = newsize << 1;
 569                 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
 570                 if (newtgts == NULL) {
 571                         mutex_unlock(&lov->lov_lock);
 572                         RETURN(-ENOMEM);
 573                 }
 574
 575                 if (lov->lov_tgt_size) {
 576                         memcpy(newtgts, lov->lov_tgts, sizeof(*newtgts) *
 577                                lov->lov_tgt_size);
 578                         old = lov->lov_tgts;
 579                         oldsize = lov->lov_tgt_size;
 580                 }
 581
 582                 lov->lov_tgts = newtgts;
 583                 lov->lov_tgt_size = newsize;
 584                 smp_rmb();
 585                 if (old)
 586                         OBD_FREE(old, sizeof(*old) * oldsize);
 587
 588                 CDEBUG(D_CONFIG, "tgts: %p size: %d\n",
 589                        lov->lov_tgts, lov->lov_tgt_size);
 590         }
 591
 592         OBD_ALLOC_PTR(tgt);
 593         if (!tgt) {
 594                 mutex_unlock(&lov->lov_lock);
 595                 RETURN(-ENOMEM);
 596         }
 597
 598         rc = lov_ost_pool_add(&lov->lov_packed, index, lov->lov_tgt_size);
 599         if (rc) {
 600                 mutex_unlock(&lov->lov_lock);
 601                 OBD_FREE_PTR(tgt);
 602                 RETURN(rc);
 603         }
 604
 605         tgt->ltd_uuid = *uuidp;
 606         tgt->ltd_obd = tgt_obd;
 607         /* XXX - add a sanity check on the generation number. */
 608         tgt->ltd_gen = gen;
 609         tgt->ltd_index = index;
 610         tgt->ltd_activate = active;
 611         lov->lov_tgts[index] = tgt;
 612         if (index >= lov->desc.ld_tgt_count)
 613                 lov->desc.ld_tgt_count = index + 1;
 614
 615         mutex_unlock(&lov->lov_lock);
 616
 617         CDEBUG(D_CONFIG, "idx=%d ltd_gen=%d ld_tgt_count=%d\n",
 618                 index, tgt->ltd_gen, lov->desc.ld_tgt_count);
 619
 620         if (lov->lov_connects == 0) {
 621                 /* lov_connect hasn't been called yet. We'll do the
 622                    lov_connect_obd on this target when that fn first runs,
 623                    because we don't know the connect flags yet. */
 624                 RETURN(0);
 625         }
 626
 627         obd_getref(obd);
 628
 629         rc = lov_connect_obd(obd, index, active, &lov->lov_ocd);
 630         if (rc)
 631                 GOTO(out, rc);
 632
 633         /* connect to administrative disabled ost */
 634         if (!tgt->ltd_exp)
 635                 GOTO(out, rc = 0);
 636
 637         if (lov->lov_cache != NULL) {
 638                 rc = obd_set_info_async(NULL, tgt->ltd_exp,
 639                                 sizeof(KEY_CACHE_SET), KEY_CACHE_SET,
 640                                 sizeof(struct cl_client_cache), lov->lov_cache,
 641                                 NULL);
 642                 if (rc < 0)
 643                         GOTO(out, rc);
 644         }
 645
 646         rc = lov_notify(obd, tgt->ltd_exp->exp_obd,
 647                         active ? OBD_NOTIFY_CONNECT : OBD_NOTIFY_INACTIVE,
 648                         (void *)&index);
 649
 650 out:
 651         if (rc) {
 652                 CERROR("add failed (%d), deleting %s\n", rc,
 653                        obd_uuid2str(&tgt->ltd_uuid));
 654                 lov_del_target(obd, index, NULL, 0);
 655         }
 656         obd_putref(obd);
 657         RETURN(rc);
 658 }
 659
 660 /* Schedule a target for deletion */
 661 int lov_del_target(struct obd_device *obd, __u32 index,
 662                    struct obd_uuid *uuidp, int gen)
 663 {
 664         struct lov_obd *lov = &obd->u.lov;
 665         int count = lov->desc.ld_tgt_count;
 666         int rc = 0;
 667         ENTRY;
 668
 669         if (index >= count) {
 670                 CERROR("LOV target index %d >= number of LOV OBDs %d.\n",
 671                        index, count);
 672                 RETURN(-EINVAL);
 673         }
 674
 675         /* to make sure there's no ongoing lov_notify() now */
 676         down_write(&lov->lov_notify_lock);
 677         obd_getref(obd);
 678
 679         if (!lov->lov_tgts[index]) {
 680                 CERROR("LOV target at index %d is not setup.\n", index);
 681                 GOTO(out, rc = -EINVAL);
 682         }
 683
 684         if (uuidp && !obd_uuid_equals(uuidp, &lov->lov_tgts[index]->ltd_uuid)) {
 685                 CERROR("LOV target UUID %s at index %d doesn't match %s.\n",
 686                        lov_uuid2str(lov, index), index,
 687                        obd_uuid2str(uuidp));
 688                 GOTO(out, rc = -EINVAL);
 689         }
 690
 691         CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d exp: %p active: %d\n",
 692                lov_uuid2str(lov, index), index,
 693                lov->lov_tgts[index]->ltd_gen, lov->lov_tgts[index]->ltd_exp,
 694                lov->lov_tgts[index]->ltd_active);
 695
 696         lov->lov_tgts[index]->ltd_reap = 1;
 697         lov->lov_death_row++;
 698         /* we really delete it from obd_putref */
 699 out:
 700         obd_putref(obd);
 701         up_write(&lov->lov_notify_lock);
 702
 703         RETURN(rc);
 704 }
 705
 706 static void __lov_del_obd(struct obd_device *obd, struct lov_tgt_desc *tgt)
 707 {
 708         struct obd_device *osc_obd;
 709
 710         LASSERT(tgt);
 711         LASSERT(tgt->ltd_reap);
 712
 713         osc_obd = class_exp2obd(tgt->ltd_exp);
 714
 715         CDEBUG(D_CONFIG, "Removing tgt %s : %s\n",
 716                tgt->ltd_uuid.uuid,
 717                osc_obd ? osc_obd->obd_name : "<no obd>");
 718
 719         if (tgt->ltd_exp)
 720                 lov_disconnect_obd(obd, tgt);
 721
 722         OBD_FREE_PTR(tgt);
 723
 724         /* Manual cleanup - no cleanup logs to clean up the osc's.  We must
 725            do it ourselves. And we can't do it from lov_cleanup,
 726            because we just lost our only reference to it. */
 727         if (osc_obd)
 728                 class_manual_cleanup(osc_obd);
 729 }
 730
 731 void lov_fix_desc_stripe_size(__u64 *val)
 732 {
 733         if (*val < LOV_MIN_STRIPE_SIZE) {
 734                 if (*val != 0)
 735                         LCONSOLE_INFO("Increasing default stripe size to "
 736                                       "minimum %u\n",
 737                                       LOV_DESC_STRIPE_SIZE_DEFAULT);
 738                 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
 739         } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
 740                 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
 741                 LCONSOLE_WARN("Changing default stripe size to %llu (a "
 742                               "multiple of %u)\n",
 743                               *val, LOV_MIN_STRIPE_SIZE);
 744         }
 745 }
 746
 747 void lov_fix_desc_stripe_count(__u32 *val)
 748 {
 749         if (*val == 0)
 750                 *val = 1;
 751 }
 752
 753 void lov_fix_desc_pattern(__u32 *val)
 754 {
 755         /* from lov_setstripe */
 756         if ((*val != 0) && (*val != LOV_PATTERN_RAID0)) {
 757                 LCONSOLE_WARN("Unknown stripe pattern: %#x\n", *val);
 758                 *val = 0;
 759         }
 760 }
 761
 762 void lov_fix_desc_qos_maxage(__u32 *val)
 763 {
 764         if (*val == 0)
 765                 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
 766 }
 767
 768 void lov_fix_desc(struct lov_desc *desc)
 769 {
 770         lov_fix_desc_stripe_size(&desc->ld_default_stripe_size);
 771         lov_fix_desc_stripe_count(&desc->ld_default_stripe_count);
 772         lov_fix_desc_pattern(&desc->ld_pattern);
 773         lov_fix_desc_qos_maxage(&desc->ld_qos_maxage);
 774 }
 775
 776 int lov_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 777 {
 778         struct lov_desc *desc;
 779         struct lov_obd *lov = &obd->u.lov;
 780 #ifdef CONFIG_PROC_FS
 781         struct obd_type *type;
 782 #endif
 783         int rc;
 784         ENTRY;
 785
 786         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 787                 CERROR("LOV setup requires a descriptor\n");
 788                 RETURN(-EINVAL);
 789         }
 790
 791         desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
 792
 793         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
 794                 CERROR("descriptor size wrong: %d > %d\n",
 795                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
 796                 RETURN(-EINVAL);
 797         }
 798
 799         if (desc->ld_magic != LOV_DESC_MAGIC) {
 800                 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
 801                             CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
 802                                    obd->obd_name, desc);
 803                             lustre_swab_lov_desc(desc);
 804                 } else {
 805                         CERROR("%s: Bad lov desc magic: %#x\n",
 806                                obd->obd_name, desc->ld_magic);
 807                         RETURN(-EINVAL);
 808                 }
 809         }
 810
 811         lov_fix_desc(desc);
 812
 813         desc->ld_active_tgt_count = 0;
 814         lov->desc = *desc;
 815         lov->lov_tgt_size = 0;
 816
 817         mutex_init(&lov->lov_lock);
 818         atomic_set(&lov->lov_refcount, 0);
 819         lov->lov_sp_me = LUSTRE_SP_CLI;
 820
 821         init_rwsem(&lov->lov_notify_lock);
 822
 823         lov->lov_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,
 824                                                    HASH_POOLS_MAX_BITS,
 825                                                    HASH_POOLS_BKT_BITS, 0,
 826                                                    CFS_HASH_MIN_THETA,
 827                                                    CFS_HASH_MAX_THETA,
 828                                                    &pool_hash_operations,
 829                                                    CFS_HASH_DEFAULT);
 830         INIT_LIST_HEAD(&lov->lov_pool_list);
 831         lov->lov_pool_count = 0;
 832         rc = lov_ost_pool_init(&lov->lov_packed, 0);
 833         if (rc)
 834                 GOTO(out, rc);
 835
 836 #ifdef CONFIG_PROC_FS
 837         obd->obd_vars = lprocfs_lov_obd_vars;
 838         /* If this is true then both client (lov) and server
 839          * (lod) are on the same node. The lod layer if loaded
 840          * first will register the lov proc directory. In that
 841          * case obd->obd_type->typ_procroot will be not set.
 842          * Instead we use type->typ_procsym as the parent. */
 843         type = class_search_type(LUSTRE_LOD_NAME);
 844         if (type != NULL && type->typ_procsym != NULL) {
 845                 obd->obd_proc_entry = lprocfs_register(obd->obd_name,
 846                                                        type->typ_procsym,
 847                                                        obd->obd_vars, obd);
 848                 if (IS_ERR(obd->obd_proc_entry)) {
 849                         rc = PTR_ERR(obd->obd_proc_entry);
 850                         CERROR("error %d setting up lprocfs for %s\n", rc,
 851                                obd->obd_name);
 852                         obd->obd_proc_entry = NULL;
 853                 }
 854         } else {
 855                 rc = lprocfs_obd_setup(obd);
 856         }
 857
 858         if (rc == 0) {
 859                 rc = lprocfs_seq_create(obd->obd_proc_entry, "target_obd",
 860                                         0444, &lov_proc_target_fops, obd);
 861                 if (rc)
 862                         CWARN("Error adding the target_obd file\n");
 863
 864                 lov->lov_pool_proc_entry = lprocfs_register("pools",
 865                                                             obd->obd_proc_entry,
 866                                                             NULL, NULL);
 867                 if (IS_ERR(lov->lov_pool_proc_entry)) {
 868                         rc = PTR_ERR(lov->lov_pool_proc_entry);
 869                         CERROR("error %d setting up lprocfs for pools\n", rc);
 870                         lov->lov_pool_proc_entry = NULL;
 871                 }
 872         }
 873 #endif
 874         RETURN(0);
 875
 876 out:
 877         return rc;
 878 }
 879
 880 static int lov_cleanup(struct obd_device *obd)
 881 {
 882         struct lov_obd *lov = &obd->u.lov;
 883         struct list_head *pos, *tmp;
 884         struct pool_desc *pool;
 885         ENTRY;
 886
 887         list_for_each_safe(pos, tmp, &lov->lov_pool_list) {
 888                 pool = list_entry(pos, struct pool_desc, pool_list);
 889                 /* free pool structs */
 890                 CDEBUG(D_INFO, "delete pool %p\n", pool);
 891                 /* In the function below, .hs_keycmp resolves to
 892                  * pool_hashkey_keycmp() */
 893                 /* coverity[overrun-buffer-val] */
 894                 lov_pool_del(obd, pool->pool_name);
 895         }
 896         cfs_hash_putref(lov->lov_pools_hash_body);
 897         lov_ost_pool_free(&lov->lov_packed);
 898
 899         lprocfs_obd_cleanup(obd);
 900         if (lov->lov_tgts) {
 901                 int i;
 902                 obd_getref(obd);
 903                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 904                         if (!lov->lov_tgts[i])
 905                                 continue;
 906
 907                         /* Inactive targets may never have connected */
 908                         if (lov->lov_tgts[i]->ltd_active ||
 909                             atomic_read(&lov->lov_refcount))
 910                                 /* We should never get here - these
 911                                  * should have been removed in the
 912                                  * disconnect. */
 913                                 CERROR("%s: lov tgt %d not cleaned! "
 914                                        "deathrow=%d, lovrc=%d\n",
 915                                        obd->obd_name, i, lov->lov_death_row,
 916                                        atomic_read(&lov->lov_refcount));
 917                         lov_del_target(obd, i, NULL, 0);
 918                 }
 919                 obd_putref(obd);
 920                 OBD_FREE(lov->lov_tgts, sizeof(*lov->lov_tgts) *
 921                          lov->lov_tgt_size);
 922                 lov->lov_tgt_size = 0;
 923         }
 924
 925         if (lov->lov_cache != NULL) {
 926                 cl_cache_decref(lov->lov_cache);
 927                 lov->lov_cache = NULL;
 928         }
 929
 930         RETURN(0);
 931 }
 932
 933 int lov_process_config_base(struct obd_device *obd, struct lustre_cfg *lcfg,
 934                             __u32 *indexp, int *genp)
 935 {
 936         struct obd_uuid obd_uuid;
 937         int cmd;
 938         int rc = 0;
 939         ENTRY;
 940
 941         switch(cmd = lcfg->lcfg_command) {
 942         case LCFG_LOV_ADD_OBD:
 943         case LCFG_LOV_ADD_INA:
 944         case LCFG_LOV_DEL_OBD: {
 945                 __u32 index;
 946                 int gen;
 947                 /* lov_modify_tgts add  0:lov_mdsA  1:ost1_UUID  2:0  3:1 */
 948                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
 949                         GOTO(out, rc = -EINVAL);
 950
 951                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
 952
 953                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%u", indexp) != 1)
 954                         GOTO(out, rc = -EINVAL);
 955                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", genp) != 1)
 956                         GOTO(out, rc = -EINVAL);
 957                 index = *indexp;
 958                 gen = *genp;
 959                 if (cmd == LCFG_LOV_ADD_OBD)
 960                         rc = lov_add_target(obd, &obd_uuid, index, gen, 1);
 961                 else if (cmd == LCFG_LOV_ADD_INA)
 962                         rc = lov_add_target(obd, &obd_uuid, index, gen, 0);
 963                 else
 964                         rc = lov_del_target(obd, index, &obd_uuid, gen);
 965                 GOTO(out, rc);
 966         }
 967         case LCFG_PARAM: {
 968                 struct lov_desc *desc = &(obd->u.lov.desc);
 969
 970                 if (!desc)
 971                         GOTO(out, rc = -EINVAL);
 972
 973                 rc = class_process_proc_param(PARAM_LOV, obd->obd_vars,
 974                                               lcfg, obd);
 975                 if (rc > 0)
 976                         rc = 0;
 977                 GOTO(out, rc);
 978         }
 979         case LCFG_POOL_NEW:
 980         case LCFG_POOL_ADD:
 981         case LCFG_POOL_DEL:
 982         case LCFG_POOL_REM:
 983                 GOTO(out, rc);
 984
 985         default: {
 986                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
 987                 GOTO(out, rc = -EINVAL);
 988
 989         }
 990         }
 991 out:
 992         RETURN(rc);
 993 }
 994
 995 static int
 996 lov_statfs_interpret(struct ptlrpc_request_set *rqset, void *data, int rc)
 997 {
 998         struct lov_request_set *lovset = (struct lov_request_set *)data;
 999         int err;
1000         ENTRY;
1001
1002         if (rc)
1003                 atomic_set(&lovset->set_completes, 0);
1004
1005         err = lov_fini_statfs_set(lovset);
1006         RETURN(rc ? rc : err);
1007 }
1008
1009 static int lov_statfs_async(struct obd_export *exp, struct obd_info *oinfo,
1010                             __u64 max_age, struct ptlrpc_request_set *rqset)
1011 {
1012         struct obd_device      *obd = class_exp2obd(exp);
1013         struct lov_request_set *set;
1014         struct lov_request *req;
1015         struct list_head *pos;
1016         struct lov_obd *lov;
1017         int rc = 0;
1018         ENTRY;
1019
1020         LASSERT(oinfo != NULL);
1021         LASSERT(oinfo->oi_osfs != NULL);
1022
1023         lov = &obd->u.lov;
1024         rc = lov_prep_statfs_set(obd, oinfo, &set);
1025         if (rc)
1026                 RETURN(rc);
1027
1028         list_for_each(pos, &set->set_list) {
1029                 req = list_entry(pos, struct lov_request, rq_link);
1030                 rc = obd_statfs_async(lov->lov_tgts[req->rq_idx]->ltd_exp,
1031                                       &req->rq_oi, max_age, rqset);
1032                 if (rc)
1033                         break;
1034         }
1035
1036         if (rc || list_empty(&rqset->set_requests)) {
1037                 int err;
1038                 if (rc)
1039                         atomic_set(&set->set_completes, 0);
1040                 err = lov_fini_statfs_set(set);
1041                 RETURN(rc ? rc : err);
1042         }
1043
1044         LASSERT(rqset->set_interpret == NULL);
1045         rqset->set_interpret = lov_statfs_interpret;
1046         rqset->set_arg = (void *)set;
1047         RETURN(0);
1048 }
1049
1050 static int lov_statfs(const struct lu_env *env, struct obd_export *exp,
1051                       struct obd_statfs *osfs, __u64 max_age, __u32 flags)
1052 {
1053         struct ptlrpc_request_set *set = NULL;
1054         struct obd_info oinfo = {
1055                 .oi_osfs = osfs,
1056                 .oi_flags = flags,
1057         };
1058         int rc = 0;
1059
1060         ENTRY;
1061
1062         /* for obdclass we forbid using obd_statfs_rqset, but prefer using async
1063          * statfs requests */
1064         set = ptlrpc_prep_set();
1065         if (set == NULL)
1066                 RETURN(-ENOMEM);
1067
1068         rc = lov_statfs_async(exp, &oinfo, max_age, set);
1069         if (rc == 0)
1070                 rc = ptlrpc_set_wait(set);
1071
1072         ptlrpc_set_destroy(set);
1073
1074         RETURN(rc);
1075 }
1076
1077 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
1078                          void *karg, void __user *uarg)
1079 {
1080         struct obd_device *obddev = class_exp2obd(exp);
1081         struct lov_obd *lov = &obddev->u.lov;
1082         int i = 0, rc = 0, count = lov->desc.ld_tgt_count;
1083         struct obd_uuid *uuidp;
1084         ENTRY;
1085
1086         switch (cmd) {
1087         case IOC_OBD_STATFS: {
1088                 struct obd_ioctl_data *data = karg;
1089                 struct obd_device *osc_obd;
1090                 struct obd_statfs stat_buf = {0};
1091                 __u32 index;
1092                 __u32 flags;
1093
1094                 memcpy(&index, data->ioc_inlbuf2, sizeof(index));
1095                 if ((index >= count))
1096                         RETURN(-ENODEV);
1097
1098                 if (!lov->lov_tgts[index])
1099                         /* Try again with the next index */
1100                         RETURN(-EAGAIN);
1101                 if (!lov->lov_tgts[index]->ltd_active)
1102                         RETURN(-ENODATA);
1103
1104                 osc_obd = class_exp2obd(lov->lov_tgts[index]->ltd_exp);
1105                 if (!osc_obd)
1106                         RETURN(-EINVAL);
1107
1108                 /* copy UUID */
1109                 if (copy_to_user(data->ioc_pbuf2, obd2cli_tgt(osc_obd),
1110                                  min_t(unsigned long, data->ioc_plen2,
1111                                        sizeof(struct obd_uuid))))
1112                         RETURN(-EFAULT);
1113
1114                 memcpy(&flags, data->ioc_inlbuf1, sizeof(flags));
1115                 flags = flags & LL_STATFS_NODELAY ? OBD_STATFS_NODELAY : 0;
1116
1117                 /* got statfs data */
1118                 rc = obd_statfs(NULL, lov->lov_tgts[index]->ltd_exp, &stat_buf,
1119                                 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
1120                                 flags);
1121                 if (rc)
1122                         RETURN(rc);
1123                 if (copy_to_user(data->ioc_pbuf1, &stat_buf,
1124                                  min_t(unsigned long, data->ioc_plen1,
1125                                        sizeof(struct obd_statfs))))
1126                         RETURN(-EFAULT);
1127                 break;
1128         }
1129         case OBD_IOC_LOV_GET_CONFIG: {
1130                 struct obd_ioctl_data *data;
1131                 struct lov_desc *desc;
1132                 char *buf = NULL;
1133                 __u32 *genp;
1134
1135                 len = 0;
1136                 if (obd_ioctl_getdata(&buf, &len, uarg))
1137                         RETURN(-EINVAL);
1138
1139                 data = (struct obd_ioctl_data *)buf;
1140
1141                 if (sizeof(*desc) > data->ioc_inllen1) {
1142                         obd_ioctl_freedata(buf, len);
1143                         RETURN(-EINVAL);
1144                 }
1145
1146                 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
1147                         obd_ioctl_freedata(buf, len);
1148                         RETURN(-EINVAL);
1149                 }
1150
1151                 if (sizeof(__u32) * count > data->ioc_inllen3) {
1152                         obd_ioctl_freedata(buf, len);
1153                         RETURN(-EINVAL);
1154                 }
1155
1156                 desc = (struct lov_desc *)data->ioc_inlbuf1;
1157                 memcpy(desc, &(lov->desc), sizeof(*desc));
1158
1159                 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
1160                 genp = (__u32 *)data->ioc_inlbuf3;
1161                 /* the uuid will be empty for deleted OSTs */
1162                 for (i = 0; i < count; i++, uuidp++, genp++) {
1163                         if (!lov->lov_tgts[i])
1164                                 continue;
1165                         *uuidp = lov->lov_tgts[i]->ltd_uuid;
1166                         *genp = lov->lov_tgts[i]->ltd_gen;
1167                 }
1168
1169                 if (copy_to_user(uarg, buf, len))
1170                         rc = -EFAULT;
1171                 obd_ioctl_freedata(buf, len);
1172                 break;
1173         }
1174         case OBD_IOC_QUOTACTL: {
1175                 struct if_quotactl *qctl = karg;
1176                 struct lov_tgt_desc *tgt = NULL;
1177                 struct obd_quotactl *oqctl;
1178
1179                 if (qctl->qc_valid == QC_OSTIDX) {
1180                         if (count <= qctl->qc_idx)
1181                                 RETURN(-EINVAL);
1182
1183                         tgt = lov->lov_tgts[qctl->qc_idx];
1184                         if (!tgt || !tgt->ltd_exp)
1185                                 RETURN(-EINVAL);
1186                 } else if (qctl->qc_valid == QC_UUID) {
1187                         for (i = 0; i < count; i++) {
1188                                 tgt = lov->lov_tgts[i];
1189                                 if (!tgt ||
1190                                     !obd_uuid_equals(&tgt->ltd_uuid,
1191                                                      &qctl->obd_uuid))
1192                                         continue;
1193
1194                                 if (tgt->ltd_exp == NULL)
1195                                         RETURN(-EINVAL);
1196
1197                                 break;
1198                         }
1199                 } else {
1200                         RETURN(-EINVAL);
1201                 }
1202
1203                 if (i >= count)
1204                         RETURN(-EAGAIN);
1205
1206                 LASSERT(tgt && tgt->ltd_exp);
1207                 OBD_ALLOC_PTR(oqctl);
1208                 if (!oqctl)
1209                         RETURN(-ENOMEM);
1210
1211                 QCTL_COPY(oqctl, qctl);
1212                 rc = obd_quotactl(tgt->ltd_exp, oqctl);
1213                 if (rc == 0) {
1214                         QCTL_COPY(qctl, oqctl);
1215                         qctl->qc_valid = QC_OSTIDX;
1216                         qctl->obd_uuid = tgt->ltd_uuid;
1217                 }
1218                 OBD_FREE_PTR(oqctl);
1219                 break;
1220         }
1221         default: {
1222                 int set = 0;
1223
1224                 if (count == 0)
1225                         RETURN(-ENOTTY);
1226
1227                 for (i = 0; i < count; i++) {
1228                         int err;
1229                         struct obd_device *osc_obd;
1230
1231                         /* OST was disconnected */
1232                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
1233                                 continue;
1234
1235                         /* ll_umount_begin() sets force flag but for lov, not
1236                          * osc. Let's pass it through */
1237                         osc_obd = class_exp2obd(lov->lov_tgts[i]->ltd_exp);
1238                         osc_obd->obd_force = obddev->obd_force;
1239                         err = obd_iocontrol(cmd, lov->lov_tgts[i]->ltd_exp,
1240                                             len, karg, uarg);
1241                         if (err) {
1242                                 if (lov->lov_tgts[i]->ltd_active) {
1243                                         CDEBUG(err == -ENOTTY ?
1244                                                D_IOCTL : D_WARNING,
1245                                                "iocontrol OSC %s on OST "
1246                                                "idx %d cmd %x: err = %d\n",
1247                                                lov_uuid2str(lov, i),
1248                                                i, cmd, err);
1249                                         if (!rc)
1250                                                 rc = err;
1251                                 }
1252                         } else {
1253                                 set = 1;
1254                         }
1255                 }
1256                 if (!set && !rc)
1257                         rc = -EIO;
1258         }
1259         }
1260
1261         RETURN(rc);
1262 }
1263
1264 static int lov_get_info(const struct lu_env *env, struct obd_export *exp,
1265                         __u32 keylen, void *key, __u32 *vallen, void *val)
1266 {
1267         struct obd_device *obddev = class_exp2obd(exp);
1268         struct lov_obd *lov = &obddev->u.lov;
1269         struct lov_desc *ld = &lov->desc;
1270         int rc = 0;
1271         ENTRY;
1272
1273         if (vallen == NULL || val == NULL)
1274                 RETURN(-EFAULT);
1275
1276         obd_getref(obddev);
1277
1278         if (KEY_IS(KEY_MAX_EASIZE)) {
1279                 u32 max_stripe_count = min_t(u32, ld->ld_active_tgt_count,
1280                                              LOV_MAX_STRIPE_COUNT);
1281
1282                 *((u32 *)val) = lov_mds_md_size(max_stripe_count, LOV_MAGIC_V3);
1283         } else if (KEY_IS(KEY_DEFAULT_EASIZE)) {
1284                 u32 def_stripe_count = min_t(u32, ld->ld_default_stripe_count,
1285                                              LOV_MAX_STRIPE_COUNT);
1286
1287                 *((u32 *)val) = lov_mds_md_size(def_stripe_count, LOV_MAGIC_V3);
1288         } else if (KEY_IS(KEY_TGT_COUNT)) {
1289                 *((int *)val) = lov->desc.ld_tgt_count;
1290         } else {
1291                 rc = -EINVAL;
1292         }
1293
1294         obd_putref(obddev);
1295
1296         RETURN(rc);
1297 }
1298
1299 static int lov_set_info_async(const struct lu_env *env, struct obd_export *exp,
1300                               __u32 keylen, void *key,
1301                               __u32 vallen, void *val,
1302                               struct ptlrpc_request_set *set)
1303 {
1304         struct obd_device *obddev = class_exp2obd(exp);
1305         struct lov_obd *lov = &obddev->u.lov;
1306         struct lov_tgt_desc *tgt;
1307         int do_inactive = 0;
1308         int no_set = 0;
1309         u32 count;
1310         u32 i;
1311         int rc = 0;
1312         int err;
1313         ENTRY;
1314
1315         if (set == NULL) {
1316                 no_set = 1;
1317                 set = ptlrpc_prep_set();
1318                 if (!set)
1319                         RETURN(-ENOMEM);
1320         }
1321
1322         obd_getref(obddev);
1323         count = lov->desc.ld_tgt_count;
1324
1325         if (KEY_IS(KEY_CHECKSUM)) {
1326                 do_inactive = 1;
1327         } else if (KEY_IS(KEY_CACHE_SET)) {
1328                 LASSERT(lov->lov_cache == NULL);
1329                 lov->lov_cache = val;
1330                 do_inactive = 1;
1331                 cl_cache_incref(lov->lov_cache);
1332         }
1333
1334         for (i = 0; i < count; i++) {
1335                 tgt = lov->lov_tgts[i];
1336
1337                 /* OST was disconnected */
1338                 if (!tgt || !tgt->ltd_exp)
1339                         continue;
1340
1341                 /* OST is inactive and we don't want inactive OSCs */
1342                 if (!tgt->ltd_active && !do_inactive)
1343                         continue;
1344
1345                 err = obd_set_info_async(env, tgt->ltd_exp, keylen, key,
1346                                          vallen, val, set);
1347                 if (!rc)
1348                         rc = err;
1349         }
1350
1351         obd_putref(obddev);
1352         if (no_set) {
1353                 err = ptlrpc_set_wait(set);
1354                 if (!rc)
1355                         rc = err;
1356                 ptlrpc_set_destroy(set);
1357         }
1358         RETURN(rc);
1359 }
1360
1361 void lov_stripe_lock(struct lov_stripe_md *md)
1362 __acquires(&md->lsm_lock)
1363 {
1364         LASSERT(md->lsm_lock_owner != current_pid());
1365         spin_lock(&md->lsm_lock);
1366         LASSERT(md->lsm_lock_owner == 0);
1367         md->lsm_lock_owner = current_pid();
1368 }
1369
1370 void lov_stripe_unlock(struct lov_stripe_md *md)
1371 __releases(&md->lsm_lock)
1372 {
1373         LASSERT(md->lsm_lock_owner == current_pid());
1374         md->lsm_lock_owner = 0;
1375         spin_unlock(&md->lsm_lock);
1376 }
1377
1378 static int lov_quotactl(struct obd_device *obd, struct obd_export *exp,
1379                         struct obd_quotactl *oqctl)
1380 {
1381         struct lov_obd      *lov = &obd->u.lov;
1382         struct lov_tgt_desc *tgt;
1383         __u64                curspace = 0;
1384         __u64                bhardlimit = 0;
1385         int                  i, rc = 0;
1386         ENTRY;
1387
1388         if (oqctl->qc_cmd != Q_GETOQUOTA &&
1389             oqctl->qc_cmd != LUSTRE_Q_SETQUOTA) {
1390                 CERROR("%s: bad quota opc %x for lov obd\n",
1391                        obd->obd_name, oqctl->qc_cmd);
1392                 RETURN(-EFAULT);
1393         }
1394
1395         /* for lov tgt */
1396         obd_getref(obd);
1397         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
1398                 int err;
1399
1400                 tgt = lov->lov_tgts[i];
1401
1402                 if (!tgt)
1403                         continue;
1404
1405                 if (!tgt->ltd_active || tgt->ltd_reap) {
1406                         if (oqctl->qc_cmd == Q_GETOQUOTA &&
1407                             lov->lov_tgts[i]->ltd_activate) {
1408                                 rc = -ENETDOWN;
1409                                 CERROR("ost %d is inactive\n", i);
1410                         } else {
1411                                 CDEBUG(D_HA, "ost %d is inactive\n", i);
1412                         }
1413                         continue;
1414                 }
1415
1416                 err = obd_quotactl(tgt->ltd_exp, oqctl);
1417                 if (err) {
1418                         if (tgt->ltd_active && !rc)
1419                                 rc = err;
1420                         continue;
1421                 }
1422
1423                 if (oqctl->qc_cmd == Q_GETOQUOTA) {
1424                         curspace += oqctl->qc_dqblk.dqb_curspace;
1425                         bhardlimit += oqctl->qc_dqblk.dqb_bhardlimit;
1426                 }
1427         }
1428         obd_putref(obd);
1429
1430         if (oqctl->qc_cmd == Q_GETOQUOTA) {
1431                 oqctl->qc_dqblk.dqb_curspace = curspace;
1432                 oqctl->qc_dqblk.dqb_bhardlimit = bhardlimit;
1433         }
1434         RETURN(rc);
1435 }
1436
1437 static struct obd_ops lov_obd_ops = {
1438         .o_owner                = THIS_MODULE,
1439         .o_setup                = lov_setup,
1440         .o_cleanup              = lov_cleanup,
1441         .o_connect              = lov_connect,
1442         .o_disconnect           = lov_disconnect,
1443         .o_statfs               = lov_statfs,
1444         .o_statfs_async         = lov_statfs_async,
1445         .o_iocontrol            = lov_iocontrol,
1446         .o_get_info             = lov_get_info,
1447         .o_set_info_async       = lov_set_info_async,
1448         .o_notify               = lov_notify,
1449         .o_pool_new             = lov_pool_new,
1450         .o_pool_rem             = lov_pool_remove,
1451         .o_pool_add             = lov_pool_add,
1452         .o_pool_del             = lov_pool_del,
1453         .o_getref               = lov_getref,
1454         .o_putref               = lov_putref,
1455         .o_quotactl             = lov_quotactl,
1456 };
1457
1458 struct kmem_cache *lov_oinfo_slab;
1459
1460 static int __init lov_init(void)
1461 {
1462         bool enable_proc = true;
1463         struct obd_type *type;
1464         int rc;
1465         ENTRY;
1466
1467         /* print an address of _any_ initialized kernel symbol from this
1468          * module, to allow debugging with gdb that doesn't support data
1469          * symbols from modules.*/
1470         CDEBUG(D_INFO, "Lustre LOV module (%p).\n", &lov_caches);
1471
1472         rc = lu_kmem_init(lov_caches);
1473         if (rc)
1474                 return rc;
1475
1476         lov_oinfo_slab = kmem_cache_create("lov_oinfo",
1477                                            sizeof(struct lov_oinfo), 0,
1478                                            SLAB_HWCACHE_ALIGN, NULL);
1479         if (lov_oinfo_slab == NULL) {
1480                 lu_kmem_fini(lov_caches);
1481                 return -ENOMEM;
1482         }
1483
1484         type = class_search_type(LUSTRE_LOD_NAME);
1485         if (type != NULL && type->typ_procsym != NULL)
1486                 enable_proc = false;
1487
1488         rc = class_register_type(&lov_obd_ops, NULL, enable_proc, NULL,
1489                                  LUSTRE_LOV_NAME, &lov_device_type);
1490
1491         if (rc) {
1492                 kmem_cache_destroy(lov_oinfo_slab);
1493                 lu_kmem_fini(lov_caches);
1494         }
1495
1496         RETURN(rc);
1497 }
1498
1499 static void __exit lov_exit(void)
1500 {
1501         class_unregister_type(LUSTRE_LOV_NAME);
1502         kmem_cache_destroy(lov_oinfo_slab);
1503         lu_kmem_fini(lov_caches);
1504 }
1505
1506 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
1507 MODULE_DESCRIPTION("Lustre Logical Object Volume");
1508 MODULE_VERSION(LUSTRE_VERSION_STRING);
1509 MODULE_LICENSE("GPL");
1510
1511 module_init(lov_init);
1512 module_exit(lov_exit);