lustre/lov/lov_obd.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright (C) 2002-2006 Cluster File Systems, Inc.
   5  * Author: Phil Schwan <phil@clusterfs.com>
   6  *         Peter Braam <braam@clusterfs.com>
   7  *         Mike Shaver <shaver@clusterfs.com>
   8  *         Nathan Rutman <nathan@clusterfs.com>
   9  *
  10  *   This file is part of the Lustre file system, http://www.lustre.org
  11  *   Lustre is a trademark of Cluster File Systems, Inc.
  12  *
  13  *   You may have signed or agreed to another license before downloading
  14  *   this software.  If so, you are bound by the terms and conditions
  15  *   of that agreement, and the following does not apply to you.  See the
  16  *   LICENSE file included with this distribution for more information.
  17  *
  18  *   If you did not agree to a different license, then this copy of Lustre
  19  *   is open source software; you can redistribute it and/or modify it
  20  *   under the terms of version 2 of the GNU General Public License as
  21  *   published by the Free Software Foundation.
  22  *
  23  *   In either case, Lustre is distributed in the hope that it will be
  24  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  25  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  26  *   license text for more details.
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_LOV
  33 #ifdef __KERNEL__
  34 #include <libcfs/libcfs.h>
  35 #else
  36 #include <liblustre.h>
  37 #endif
  38
  39 #include <obd_support.h>
  40 #include <lustre_lib.h>
  41 #include <lustre_net.h>
  42 #include <lustre/lustre_idl.h>
  43 #include <lustre_dlm.h>
  44 #include <lustre_mds.h>
  45 #include <lustre_debug.h>
  46 #include <obd_class.h>
  47 #include <obd_lov.h>
  48 #include <obd_ost.h>
  49 #include <lprocfs_status.h>
  50 #include <lustre_param.h>
  51
  52 #include "lov_internal.h"
  53
  54
  55 /* Keep a refcount of lov->tgt usage to prevent racing with addition/deletion.
  56    Any function that expects lov_tgts to remain stationary must take a ref. */
  57 void lov_getref(struct obd_device *obd)
  58 {
  59         struct lov_obd *lov = &obd->u.lov;
  60
  61         /* nobody gets through here until lov_putref is done */
  62         mutex_down(&lov->lov_lock);
  63         atomic_inc(&lov->lov_refcount);
  64         mutex_up(&lov->lov_lock);
  65         return;
  66 }
  67
  68 static void __lov_del_obd(struct obd_device *obd, __u32 index);
  69
  70 void lov_putref(struct obd_device *obd)
  71 {
  72         struct lov_obd *lov = &obd->u.lov;
  73         mutex_down(&lov->lov_lock);
  74         /* ok to dec to 0 more than once -- ltd_exp's will be null */
  75         if (atomic_dec_and_test(&lov->lov_refcount) && lov->lov_death_row) {
  76                 int i;
  77                 CDEBUG(D_CONFIG, "destroying %d lov targets\n",
  78                        lov->lov_death_row);
  79                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
  80                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_reap)
  81                                 continue;
  82                         /* Disconnect and delete from list */
  83                         __lov_del_obd(obd, i);
  84                         lov->lov_death_row--;
  85                 }
  86         }
  87         mutex_up(&lov->lov_lock);
  88 }
  89
  90 #define MAX_STRING_SIZE 128
  91 static int lov_connect_obd(struct obd_device *obd, __u32 index, int activate,
  92                            struct obd_connect_data *data)
  93 {
  94         struct lov_obd *lov = &obd->u.lov;
  95         struct obd_uuid tgt_uuid;
  96         struct obd_device *tgt_obd;
  97         struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" };
  98         struct lustre_handle conn = {0, };
  99         struct obd_import *imp;
 100 #ifdef __KERNEL__
 101         cfs_proc_dir_entry_t *lov_proc_dir;
 102 #endif
 103         int rc;
 104         ENTRY;
 105
 106         if (!lov->lov_tgts[index])
 107                 RETURN(-EINVAL);
 108
 109         tgt_uuid = lov->lov_tgts[index]->ltd_uuid;
 110         tgt_obd = class_find_client_obd(&tgt_uuid, LUSTRE_OSC_NAME,
 111                                         &obd->obd_uuid);
 112
 113         if (!tgt_obd) {
 114                 CERROR("Target %s not attached\n", obd_uuid2str(&tgt_uuid));
 115                 RETURN(-EINVAL);
 116         }
 117
 118         if (!tgt_obd->obd_set_up) {
 119                 CERROR("Target %s not set up\n", obd_uuid2str(&tgt_uuid));
 120                 RETURN(-EINVAL);
 121         }
 122
 123         if (data && (data->ocd_connect_flags & OBD_CONNECT_INDEX))
 124                 data->ocd_index = index;
 125
 126         /*
 127          * Divine LOV knows that OBDs under it are OSCs.
 128          */
 129         imp = tgt_obd->u.cli.cl_import;
 130
 131         if (activate) {
 132                 tgt_obd->obd_no_recov = 0;
 133                 /* FIXME this is probably supposed to be
 134                    ptlrpc_set_import_active.  Horrible naming. */
 135                 ptlrpc_activate_import(imp);
 136         }
 137
 138         if (imp->imp_invalid) {
 139                 CERROR("not connecting OSC %s; administratively "
 140                        "disabled\n", obd_uuid2str(&tgt_uuid));
 141                 rc = obd_register_observer(tgt_obd, obd);
 142                 if (rc) {
 143                         CERROR("Target %s register_observer error %d; "
 144                                "will not be able to reactivate\n",
 145                                obd_uuid2str(&tgt_uuid), rc);
 146                 }
 147                 RETURN(0);
 148         }
 149
 150         rc = obd_connect(&conn, tgt_obd, &lov_osc_uuid, data);
 151         if (rc) {
 152                 CERROR("Target %s connect error %d\n",
 153                        obd_uuid2str(&tgt_uuid), rc);
 154                 RETURN(rc);
 155         }
 156         lov->lov_tgts[index]->ltd_exp = class_conn2export(&conn);
 157         if (!lov->lov_tgts[index]->ltd_exp) {
 158                 CERROR("Target %s: null export!\n", obd_uuid2str(&tgt_uuid));
 159                 RETURN(-ENODEV);
 160         }
 161
 162         rc = obd_register_observer(tgt_obd, obd);
 163         if (rc) {
 164                 CERROR("Target %s register_observer error %d\n",
 165                        obd_uuid2str(&tgt_uuid), rc);
 166                 obd_disconnect(lov->lov_tgts[index]->ltd_exp);
 167                 lov->lov_tgts[index]->ltd_exp = NULL;
 168                 RETURN(rc);
 169         }
 170
 171         lov->lov_tgts[index]->ltd_reap = 0;
 172         if (activate) {
 173                 lov->lov_tgts[index]->ltd_active = 1;
 174                 lov->desc.ld_active_tgt_count++;
 175         }
 176         CDEBUG(D_CONFIG, "Connected tgt idx %d %s (%s) %sactive\n", index,
 177                obd_uuid2str(&tgt_uuid), tgt_obd->obd_name, activate ? "":"in");
 178
 179 #ifdef __KERNEL__
 180         lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 181         if (lov_proc_dir) {
 182                 struct obd_device *osc_obd = class_conn2obd(&conn);
 183                 cfs_proc_dir_entry_t *osc_symlink;
 184                 char name[MAX_STRING_SIZE];
 185
 186                 LASSERT(osc_obd != NULL);
 187                 LASSERT(osc_obd->obd_magic == OBD_DEVICE_MAGIC);
 188                 LASSERT(osc_obd->obd_type->typ_name != NULL);
 189                 snprintf(name, MAX_STRING_SIZE, "../../../%s/%s",
 190                          osc_obd->obd_type->typ_name,
 191                          osc_obd->obd_name);
 192                 osc_symlink = proc_symlink(osc_obd->obd_name, lov_proc_dir,
 193                                            name);
 194                 if (osc_symlink == NULL) {
 195                         CERROR("could not register LOV target "
 196                                "/proc/fs/lustre/%s/%s/target_obds/%s.",
 197                                obd->obd_type->typ_name, obd->obd_name,
 198                                osc_obd->obd_name);
 199                         lprocfs_remove(&lov_proc_dir);
 200                 }
 201         }
 202 #endif
 203
 204         rc = qos_add_tgt(obd, index);
 205         if (rc)
 206                 CERROR("qos_add_tgt failed %d\n", rc);
 207
 208         RETURN(0);
 209 }
 210
 211 static int lov_connect(struct lustre_handle *conn, struct obd_device *obd,
 212                        struct obd_uuid *cluuid, struct obd_connect_data *data)
 213 {
 214         struct lov_obd *lov = &obd->u.lov;
 215         struct lov_tgt_desc *tgt;
 216         int i, rc;
 217         ENTRY;
 218
 219         CDEBUG(D_CONFIG, "connect #%d\n", lov->lov_connects);
 220
 221         rc = class_connect(conn, obd, cluuid);
 222         if (rc)
 223                 RETURN(rc);
 224
 225         /* Why should there ever be more than 1 connect? */
 226         lov->lov_connects++;
 227         LASSERT(lov->lov_connects == 1);
 228
 229         memset(&lov->lov_ocd, 0, sizeof(lov->lov_ocd));
 230         if (data)
 231                 lov->lov_ocd = *data;
 232
 233         lov_getref(obd);
 234         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 235                 tgt = lov->lov_tgts[i];
 236                 if (!tgt || obd_uuid_empty(&tgt->ltd_uuid))
 237                         continue;
 238                 /* Flags will be lowest common denominator */
 239                 rc = lov_connect_obd(obd, i, lov->lov_tgts[i]->ltd_activate,
 240                                      &lov->lov_ocd);
 241                 if (rc) {
 242                         CERROR("%s: lov connect tgt %d failed: %d\n",
 243                                obd->obd_name, i, rc);
 244                         continue;
 245                 }
 246         }
 247         lov_putref(obd);
 248
 249         RETURN(0);
 250 }
 251
 252 static int lov_disconnect_obd(struct obd_device *obd, __u32 index)
 253 {
 254         cfs_proc_dir_entry_t *lov_proc_dir;
 255         struct lov_obd *lov = &obd->u.lov;
 256         struct obd_device *osc_obd =
 257                 class_exp2obd(lov->lov_tgts[index]->ltd_exp);
 258         int rc;
 259         ENTRY;
 260
 261         CDEBUG(D_CONFIG, "%s: disconnecting target %s\n",
 262                obd->obd_name, osc_obd->obd_name);
 263
 264         if (lov->lov_tgts[index]->ltd_active) {
 265                 lov->lov_tgts[index]->ltd_active = 0;
 266                 lov->desc.ld_active_tgt_count--;
 267         }
 268
 269         lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds");
 270         if (lov_proc_dir) {
 271                 cfs_proc_dir_entry_t *osc_symlink;
 272
 273                 osc_symlink = lprocfs_srch(lov_proc_dir, osc_obd->obd_name);
 274                 if (osc_symlink) {
 275                         lprocfs_remove(&osc_symlink);
 276                 } else {
 277                         CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing.",
 278                                obd->obd_type->typ_name, obd->obd_name,
 279                                osc_obd->obd_name);
 280                 }
 281         }
 282
 283         if (obd->obd_no_recov) {
 284                 /* Pass it on to our clients.
 285                  * XXX This should be an argument to disconnect,
 286                  * XXX not a back-door flag on the OBD.  Ah well.
 287                  */
 288                 if (osc_obd)
 289                         osc_obd->obd_no_recov = 1;
 290         }
 291
 292         obd_register_observer(osc_obd, NULL);
 293
 294         rc = obd_disconnect(lov->lov_tgts[index]->ltd_exp);
 295         if (rc) {
 296                 CERROR("Target %s disconnect error %d\n",
 297                        lov_uuid2str(lov, index), rc);
 298                 rc = 0;
 299         }
 300
 301         qos_del_tgt(obd, index);
 302
 303         lov->lov_tgts[index]->ltd_exp = NULL;
 304         RETURN(0);
 305 }
 306
 307 static int lov_del_target(struct obd_device *obd, __u32 index,
 308                           struct obd_uuid *uuidp, int gen);
 309
 310 static int lov_disconnect(struct obd_export *exp)
 311 {
 312         struct obd_device *obd = class_exp2obd(exp);
 313         struct lov_obd *lov = &obd->u.lov;
 314         int i, rc;
 315         ENTRY;
 316
 317         if (!lov->lov_tgts)
 318                 goto out;
 319
 320         /* Only disconnect the underlying layers on the final disconnect. */
 321         lov->lov_connects--;
 322         if (lov->lov_connects != 0) {
 323                 /* why should there be more than 1 connect? */
 324                 CERROR("disconnect #%d\n", lov->lov_connects);
 325                 goto out;
 326         }
 327
 328         /* Let's hold another reference so lov_del_obd doesn't spin through
 329            putref every time */
 330         lov_getref(obd);
 331         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 332                 if (lov->lov_tgts[i] && lov->lov_tgts[i]->ltd_exp) {
 333                         /* Disconnection is the last we know about an obd */
 334                         lov_del_target(obd, i, 0, lov->lov_tgts[i]->ltd_gen);
 335                 }
 336         }
 337         lov_putref(obd);
 338
 339 out:
 340         rc = class_disconnect(exp); /* bz 9811 */
 341         RETURN(rc);
 342 }
 343
 344 /* Error codes:
 345  *
 346  *  -EINVAL  : UUID can't be found in the LOV's target list
 347  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
 348  *  -EBADF   : The UUID is found, but the OBD is the wrong type (!)
 349  */
 350 static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid,
 351                               int activate)
 352 {
 353         struct lov_obd *lov = &obd->u.lov;
 354         struct lov_tgt_desc *tgt;
 355         int i, rc = 0;
 356         ENTRY;
 357
 358         CDEBUG(D_INFO, "Searching in lov %p for uuid %s (activate=%d)\n",
 359                lov, uuid->uuid, activate);
 360
 361         lov_getref(obd);
 362         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 363                 tgt = lov->lov_tgts[i];
 364                 if (!tgt || !tgt->ltd_exp)
 365                         continue;
 366
 367                 CDEBUG(D_INFO, "lov idx %d is %s conn "LPX64"\n",
 368                        i, obd_uuid2str(&tgt->ltd_uuid),
 369                        tgt->ltd_exp->exp_handle.h_cookie);
 370                 if (obd_uuid_equals(uuid, &tgt->ltd_uuid))
 371                         break;
 372         }
 373
 374         if (i == lov->desc.ld_tgt_count)
 375                 GOTO(out, rc = -EINVAL);
 376
 377         if (lov->lov_tgts[i]->ltd_active == activate) {
 378                 CDEBUG(D_INFO, "OSC %s already %sactive!\n", uuid->uuid,
 379                        activate ? "" : "in");
 380                 GOTO(out, rc);
 381         }
 382
 383         CDEBUG(D_CONFIG, "Marking OSC %s %sactive\n", obd_uuid2str(uuid),
 384                activate ? "" : "in");
 385
 386         lov->lov_tgts[i]->ltd_active = activate;
 387         if (activate)
 388                 lov->desc.ld_active_tgt_count++;
 389         else
 390                 lov->desc.ld_active_tgt_count--;
 391         /* remove any old qos penalty */
 392         lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0;
 393
 394  out:
 395         lov_putref(obd);
 396         RETURN(rc);
 397 }
 398
 399 static int lov_notify(struct obd_device *obd, struct obd_device *watched,
 400                       enum obd_notify_event ev, void *data)
 401 {
 402         int rc = 0;
 403         ENTRY;
 404
 405         if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) {
 406                 struct obd_uuid *uuid;
 407
 408                 LASSERT(watched);
 409
 410                 if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) {
 411                         CERROR("unexpected notification of %s %s!\n",
 412                                watched->obd_type->typ_name,
 413                                watched->obd_name);
 414                         RETURN(-EINVAL);
 415                 }
 416                 uuid = &watched->u.cli.cl_target_uuid;
 417
 418                 /* Set OSC as active before notifying the observer, so the
 419                  * observer can use the OSC normally.
 420                  */
 421                 rc = lov_set_osc_active(obd, uuid, ev == OBD_NOTIFY_ACTIVE);
 422                 if (rc) {
 423                         CERROR("%sactivation of %s failed: %d\n",
 424                                (ev == OBD_NOTIFY_ACTIVE) ? "" : "de",
 425                                obd_uuid2str(uuid), rc);
 426                         RETURN(rc);
 427                 }
 428         }
 429
 430         /* Pass the notification up the chain. */
 431         if (watched) {
 432                 rc = obd_notify_observer(obd, watched, ev, data);
 433         } else {
 434                 /* NULL watched means all osc's in the lov (only for syncs) */
 435                 struct lov_obd *lov = &obd->u.lov;
 436                 struct obd_device *tgt_obd;
 437                 int i;
 438                 lov_getref(obd);
 439                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 440                         if (!lov->lov_tgts[i])
 441                                 continue;
 442                         tgt_obd = class_exp2obd(lov->lov_tgts[i]->ltd_exp);
 443                         rc = obd_notify_observer(obd, tgt_obd, ev, data);
 444                         if (rc) {
 445                                 CERROR("%s: notify %s of %s failed %d\n",
 446                                        obd->obd_name,
 447                                        obd->obd_observer->obd_name,
 448                                        tgt_obd->obd_name, rc);
 449                                 break;
 450                         }
 451                 }
 452                 lov_putref(obd);
 453         }
 454
 455         RETURN(rc);
 456 }
 457
 458 static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp,
 459                           __u32 index, int gen, int active)
 460 {
 461         struct lov_obd *lov = &obd->u.lov;
 462         struct lov_tgt_desc *tgt;
 463         int rc;
 464         ENTRY;
 465
 466         CDEBUG(D_CONFIG, "uuid:%s idx:%d gen:%d active:%d\n",
 467                uuidp->uuid, index, gen, active);
 468
 469         if (gen <= 0) {
 470                 CERROR("request to add OBD %s with invalid generation: %d\n",
 471                        uuidp->uuid, gen);
 472                 RETURN(-EINVAL);
 473         }
 474
 475         mutex_down(&lov->lov_lock);
 476
 477         if ((index < lov->lov_tgt_size) && (lov->lov_tgts[index] != NULL)) {
 478                 tgt = lov->lov_tgts[index];
 479                 CERROR("UUID %s already assigned at LOV target index %d\n",
 480                        obd_uuid2str(&tgt->ltd_uuid), index);
 481                 mutex_up(&lov->lov_lock);
 482                 RETURN(-EEXIST);
 483         }
 484
 485         if (index >= lov->lov_tgt_size) {
 486                 /* We need to reallocate the lov target array. */
 487                 struct lov_tgt_desc **newtgts, **old = NULL;
 488                 __u32 newsize, oldsize = 0;
 489
 490                 newsize = max(lov->lov_tgt_size, (__u32)2);
 491                 while (newsize < index + 1)
 492                         newsize = newsize << 1;
 493                 OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize);
 494                 if (newtgts == NULL) {
 495                         mutex_up(&lov->lov_lock);
 496                         RETURN(-ENOMEM);
 497                 }
 498
 499                 if (lov->lov_tgt_size) {
 500                         memcpy(newtgts, lov->lov_tgts, sizeof(*newtgts) *
 501                                lov->lov_tgt_size);
 502                         old = lov->lov_tgts;
 503                         oldsize = lov->lov_tgt_size;
 504                 }
 505
 506                 lov->lov_tgts = newtgts;
 507                 lov->lov_tgt_size = newsize;
 508 #ifdef __KERNEL__
 509                 smp_rmb();
 510 #endif
 511                 if (old)
 512                         OBD_FREE(old, sizeof(*old) * oldsize);
 513
 514                 CDEBUG(D_CONFIG, "tgts: %p size: %d\n",
 515                        lov->lov_tgts, lov->lov_tgt_size);
 516         }
 517
 518
 519         OBD_ALLOC_PTR(tgt);
 520         if (!tgt) {
 521                 mutex_up(&lov->lov_lock);
 522                 RETURN(-ENOMEM);
 523         }
 524
 525         memset(tgt, 0, sizeof(*tgt));
 526         tgt->ltd_uuid = *uuidp;
 527         /* XXX - add a sanity check on the generation number. */
 528         tgt->ltd_gen = gen;
 529         tgt->ltd_index = index;
 530         tgt->ltd_activate = active;
 531         lov->lov_tgts[index] = tgt;
 532         if (index >= lov->desc.ld_tgt_count)
 533                 lov->desc.ld_tgt_count = index + 1;
 534         mutex_up(&lov->lov_lock);
 535
 536         CDEBUG(D_CONFIG, "idx=%d ltd_gen=%d ld_tgt_count=%d\n",
 537                 index, tgt->ltd_gen, lov->desc.ld_tgt_count);
 538
 539         if (lov->lov_connects == 0) {
 540                 /* lov_connect hasn't been called yet. We'll do the
 541                    lov_connect_obd on this target when that fn first runs,
 542                    because we don't know the connect flags yet. */
 543                 RETURN(0);
 544         }
 545
 546         lov_getref(obd);
 547
 548         rc = lov_connect_obd(obd, index, active, &lov->lov_ocd);
 549         if (rc)
 550                 GOTO(out, rc);
 551
 552         rc = lov_notify(obd, tgt->ltd_exp->exp_obd,
 553                         active ? OBD_NOTIFY_ACTIVE : OBD_NOTIFY_INACTIVE,
 554                         (void *)&index);
 555
 556 out:
 557         if (rc) {
 558                 CERROR("add failed (%d), deleting %s\n", rc,
 559                        obd_uuid2str(&tgt->ltd_uuid));
 560                 lov_del_target(obd, index, 0, 0);
 561         }
 562         lov_putref(obd);
 563         RETURN(rc);
 564 }
 565
 566 /* Schedule a target for deletion */
 567 static int lov_del_target(struct obd_device *obd, __u32 index,
 568                           struct obd_uuid *uuidp, int gen)
 569 {
 570         struct lov_obd *lov = &obd->u.lov;
 571         int count = lov->desc.ld_tgt_count;
 572         int rc = 0;
 573         ENTRY;
 574
 575         if (index >= count) {
 576                 CERROR("LOV target index %d >= number of LOV OBDs %d.\n",
 577                        index, count);
 578                 RETURN(-EINVAL);
 579         }
 580
 581         lov_getref(obd);
 582
 583         if (!lov->lov_tgts[index]) {
 584                 CERROR("LOV target at index %d is not setup.\n", index);
 585                 GOTO(out, rc = -EINVAL);
 586         }
 587
 588         if (uuidp && !obd_uuid_equals(uuidp, &lov->lov_tgts[index]->ltd_uuid)) {
 589                 CERROR("LOV target UUID %s at index %d doesn't match %s.\n",
 590                        lov_uuid2str(lov, index), index,
 591                        obd_uuid2str(uuidp));
 592                 GOTO(out, rc = -EINVAL);
 593         }
 594
 595         CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d exp: %p active: %d\n",
 596                lov_uuid2str(lov, index), index,
 597                lov->lov_tgts[index]->ltd_gen, lov->lov_tgts[index]->ltd_exp,
 598                lov->lov_tgts[index]->ltd_active);
 599
 600         lov->lov_tgts[index]->ltd_reap = 1;
 601         lov->lov_death_row++;
 602         /* we really delete it from lov_putref */
 603 out:
 604         lov_putref(obd);
 605
 606         RETURN(rc);
 607 }
 608
 609 /* We are holding lov_lock */
 610 static void __lov_del_obd(struct obd_device *obd, __u32 index)
 611 {
 612         struct lov_obd *lov = &obd->u.lov;
 613         struct obd_device *osc_obd;
 614         struct lov_tgt_desc *tgt = lov->lov_tgts[index];
 615
 616         LASSERT(tgt);
 617         LASSERT(tgt->ltd_reap);
 618
 619         osc_obd = class_exp2obd(tgt->ltd_exp);
 620
 621         CDEBUG(D_CONFIG, "Removing tgt %s : %s\n",
 622                lov_uuid2str(lov, index),
 623                osc_obd ? osc_obd->obd_name : "<no obd>");
 624
 625         if (tgt->ltd_exp)
 626                 lov_disconnect_obd(obd, index);
 627
 628         /* XXX - right now there is a dependency on ld_tgt_count being the
 629          * maximum tgt index for computing the mds_max_easize. So we can't
 630          * shrink it. */
 631
 632         lov->lov_tgts[index] = NULL;
 633         OBD_FREE_PTR(tgt);
 634
 635         /* Manual cleanup - no cleanup logs to clean up the osc's.  We must
 636            do it ourselves. And we can't do it from lov_cleanup,
 637            because we just lost our only reference to it. */
 638         if (osc_obd) {
 639                 /* Use lov's force/fail flags. */
 640                 osc_obd->obd_force = obd->obd_force;
 641                 osc_obd->obd_fail = obd->obd_fail;
 642                 class_manual_cleanup(osc_obd);
 643         }
 644 }
 645
 646 void lov_fix_desc(struct lov_desc *desc)
 647 {
 648         if (desc->ld_default_stripe_size < PTLRPC_MAX_BRW_SIZE) {
 649                 LCONSOLE_WARN("Increasing default stripe size to min %u\n",
 650                               PTLRPC_MAX_BRW_SIZE);
 651                 desc->ld_default_stripe_size = PTLRPC_MAX_BRW_SIZE;
 652         } else if (desc->ld_default_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
 653                 desc->ld_default_stripe_size &= ~(LOV_MIN_STRIPE_SIZE - 1);
 654                 LCONSOLE_WARN("Changing default stripe size to "LPU64" (a "
 655                               "multiple of %u)\n",
 656                               desc->ld_default_stripe_size,LOV_MIN_STRIPE_SIZE);
 657         }
 658
 659         if (desc->ld_default_stripe_count == 0)
 660                 desc->ld_default_stripe_count = 1;
 661
 662         /* from lov_setstripe */
 663         if ((desc->ld_pattern != 0) &&
 664             (desc->ld_pattern != LOV_PATTERN_RAID0)) {
 665                 LCONSOLE_WARN("Unknown stripe pattern: %#x\n",desc->ld_pattern);
 666                 desc->ld_pattern = 0;
 667         }
 668         /* fix qos_maxage */
 669         if (desc->ld_qos_maxage == 0)
 670                 desc->ld_qos_maxage = QOS_DEFAULT_MAXAGE;
 671 }
 672
 673 static int lov_setup(struct obd_device *obd, obd_count len, void *buf)
 674 {
 675         struct lprocfs_static_vars lvars;
 676         struct lustre_cfg *lcfg = buf;
 677         struct lov_desc *desc;
 678         struct lov_obd *lov = &obd->u.lov;
 679         int count;
 680         ENTRY;
 681
 682         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 683                 CERROR("LOV setup requires a descriptor\n");
 684                 RETURN(-EINVAL);
 685         }
 686
 687         desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
 688
 689         if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
 690                 CERROR("descriptor size wrong: %d > %d\n",
 691                        (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
 692                 RETURN(-EINVAL);
 693         }
 694
 695         if (desc->ld_magic != LOV_DESC_MAGIC) {
 696                 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
 697                             CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
 698                                    obd->obd_name, desc);
 699                             lustre_swab_lov_desc(desc);
 700                 } else {
 701                         CERROR("%s: Bad lov desc magic: %#x\n",
 702                                obd->obd_name, desc->ld_magic);
 703                         RETURN(-EINVAL);
 704                 }
 705         }
 706
 707         lov_fix_desc(desc);
 708
 709         /* Because of 64-bit divide/mod operations only work with a 32-bit
 710          * divisor in a 32-bit kernel, we cannot support a stripe width
 711          * of 4GB or larger on 32-bit CPUs. */
 712         count = desc->ld_default_stripe_count;
 713         if ((count > 0 ? count : desc->ld_tgt_count) *
 714             desc->ld_default_stripe_size > 0xffffffff) {
 715                 CERROR("LOV: stripe width "LPU64"x%u > 4294967295 bytes\n",
 716                        desc->ld_default_stripe_size, count);
 717                 RETURN(-EINVAL);
 718         }
 719
 720         desc->ld_active_tgt_count = 0;
 721         lov->desc = *desc;
 722         lov->lov_tgt_size = 0;
 723         sema_init(&lov->lov_lock, 1);
 724         atomic_set(&lov->lov_refcount, 0);
 725         INIT_LIST_HEAD(&lov->lov_qos.lq_oss_list);
 726         init_rwsem(&lov->lov_qos.lq_rw_sem);
 727         lov->lov_qos.lq_dirty = 1;
 728         lov->lov_qos.lq_dirty_rr = 1;
 729         lov->lov_qos.lq_reset = 1;
 730         /* Default priority is toward free space balance */
 731         lov->lov_qos.lq_prio_free = 232;
 732
 733         lprocfs_init_vars(lov, &lvars);
 734         lprocfs_obd_setup(obd, lvars.obd_vars);
 735 #ifdef LPROCFS
 736         {
 737                 cfs_proc_dir_entry_t *entry;
 738
 739                 entry = create_proc_entry("target_obd", 0444,
 740                                           obd->obd_proc_entry);
 741                 if (entry != NULL) {
 742                         entry->proc_fops = &lov_proc_target_fops;
 743                         entry->data = obd;
 744                 }
 745         }
 746 #endif
 747
 748         RETURN(0);
 749 }
 750
 751 static int lov_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
 752 {
 753         int rc = 0;
 754         ENTRY;
 755
 756         switch (stage) {
 757         case OBD_CLEANUP_EARLY: {
 758                 struct lov_obd *lov = &obd->u.lov;
 759                 int i;
 760                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 761                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
 762                                 continue;
 763                         obd_precleanup(class_exp2obd(lov->lov_tgts[i]->ltd_exp),
 764                                        OBD_CLEANUP_EARLY);
 765                 }
 766                 break;
 767         }
 768         case OBD_CLEANUP_EXPORTS:
 769                 break;
 770         case OBD_CLEANUP_SELF_EXP:
 771                 rc = obd_llog_finish(obd, 0);
 772                 if (rc != 0)
 773                         CERROR("failed to cleanup llogging subsystems\n");
 774                 break;
 775         case OBD_CLEANUP_OBD:
 776                 break;
 777         }
 778         RETURN(rc);
 779 }
 780
 781 static int lov_cleanup(struct obd_device *obd)
 782 {
 783         struct lov_obd *lov = &obd->u.lov;
 784
 785         lprocfs_obd_cleanup(obd);
 786         if (lov->lov_tgts) {
 787                 int i;
 788                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 789                         if (lov->lov_tgts[i]) {
 790                                 /* Inactive targets may never have connected */
 791                                 if (lov->lov_tgts[i]->ltd_active ||
 792                                     atomic_read(&lov->lov_refcount))
 793                                         /* We should never get here - these
 794                                            should have been removed in the
 795                                            disconnect. */
 796                                         CERROR("lov tgt %d not cleaned!"
 797                                                " deathrow=%d, lovrc=%d\n",
 798                                                i, lov->lov_death_row,
 799                                                atomic_read(&lov->lov_refcount));
 800                                 lov_del_target(obd, i, 0, 0);
 801                         }
 802                 }
 803                 OBD_FREE(lov->lov_tgts, sizeof(*lov->lov_tgts) *
 804                          lov->lov_tgt_size);
 805                 lov->lov_tgt_size = 0;
 806         }
 807
 808         if (lov->lov_qos.lq_rr_size)
 809                 OBD_FREE(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size);
 810
 811         RETURN(0);
 812 }
 813
 814 static int lov_process_config(struct obd_device *obd, obd_count len, void *buf)
 815 {
 816         struct lustre_cfg *lcfg = buf;
 817         struct obd_uuid obd_uuid;
 818         int cmd;
 819         int rc = 0;
 820         ENTRY;
 821
 822         switch(cmd = lcfg->lcfg_command) {
 823         case LCFG_LOV_ADD_OBD:
 824         case LCFG_LOV_ADD_INA:
 825         case LCFG_LOV_DEL_OBD: {
 826                 __u32 index;
 827                 int gen;
 828                 /* lov_modify_tgts add  0:lov_mdsA  1:ost1_UUID  2:0  3:1 */
 829                 if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid))
 830                         GOTO(out, rc = -EINVAL);
 831
 832                 obd_str2uuid(&obd_uuid,  lustre_cfg_buf(lcfg, 1));
 833
 834                 if (sscanf(lustre_cfg_buf(lcfg, 2), "%d", &index) != 1)
 835                         GOTO(out, rc = -EINVAL);
 836                 if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1)
 837                         GOTO(out, rc = -EINVAL);
 838                 if (cmd == LCFG_LOV_ADD_OBD)
 839                         rc = lov_add_target(obd, &obd_uuid, index, gen, 1);
 840                 else if (cmd == LCFG_LOV_ADD_INA)
 841                         rc = lov_add_target(obd, &obd_uuid, index, gen, 0);
 842                 else
 843                         rc = lov_del_target(obd, index, &obd_uuid, gen);
 844                 GOTO(out, rc);
 845         }
 846         case LCFG_PARAM: {
 847                 struct lprocfs_static_vars lvars;
 848                 struct lov_desc *desc = &(obd->u.lov.desc);
 849
 850                 if (!desc)
 851                         GOTO(out, rc = -EINVAL);
 852
 853                 lprocfs_init_vars(lov, &lvars);
 854
 855                 rc = class_process_proc_param(PARAM_LOV, lvars.obd_vars,
 856                                               lcfg, obd);
 857                 GOTO(out, rc);
 858         }
 859         default: {
 860                 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
 861                 GOTO(out, rc = -EINVAL);
 862
 863         }
 864         }
 865 out:
 866         RETURN(rc);
 867 }
 868
 869 #ifndef log2
 870 #define log2(n) ffz(~(n))
 871 #endif
 872
 873 static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa,
 874                              struct lov_stripe_md **ea,
 875                              struct obd_trans_info *oti)
 876 {
 877         struct lov_obd *lov;
 878         struct obdo *tmp_oa;
 879         struct obd_uuid *ost_uuid = NULL;
 880         int rc = 0, i;
 881         ENTRY;
 882
 883         LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS &&
 884                 src_oa->o_flags == OBD_FL_DELORPHAN);
 885
 886         lov = &export->exp_obd->u.lov;
 887
 888         OBDO_ALLOC(tmp_oa);
 889         if (tmp_oa == NULL)
 890                 RETURN(-ENOMEM);
 891
 892         if (src_oa->o_valid & OBD_MD_FLINLINE) {
 893                 ost_uuid = (struct obd_uuid *)src_oa->o_inline;
 894                 CDEBUG(D_HA, "clearing orphans only for %s\n",
 895                        ost_uuid->uuid);
 896         }
 897
 898         lov_getref(export->exp_obd);
 899         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 900                 struct lov_stripe_md obj_md;
 901                 struct lov_stripe_md *obj_mdp = &obj_md;
 902                 struct lov_tgt_desc *tgt;
 903                 int err;
 904
 905                 tgt = lov->lov_tgts[i];
 906                 if (!tgt)
 907                         continue;
 908
 909                 /* if called for a specific target, we don't
 910                    care if it is not active. */
 911                 if (!lov->lov_tgts[i]->ltd_active && ost_uuid == NULL) {
 912                         CDEBUG(D_HA, "lov idx %d inactive\n", i);
 913                         continue;
 914                 }
 915
 916                 if (ost_uuid && !obd_uuid_equals(ost_uuid, &tgt->ltd_uuid))
 917                         continue;
 918
 919                 CDEBUG(D_CONFIG,"Clear orphans for %d:%s\n", i,
 920                        obd_uuid2str(ost_uuid));
 921
 922                 memcpy(tmp_oa, src_oa, sizeof(*tmp_oa));
 923
 924                 LASSERT(lov->lov_tgts[i]->ltd_exp);
 925                 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
 926                 err = obd_create(lov->lov_tgts[i]->ltd_exp,
 927                                  tmp_oa, &obj_mdp, oti);
 928                 if (err)
 929                         /* This export will be disabled until it is recovered,
 930                            and then orphan recovery will be completed. */
 931                         CERROR("error in orphan recovery on OST idx %d/%d: "
 932                                "rc = %d\n", i, lov->desc.ld_tgt_count, err);
 933
 934                 if (ost_uuid)
 935                         break;
 936         }
 937         lov_putref(export->exp_obd);
 938
 939         OBDO_FREE(tmp_oa);
 940         RETURN(rc);
 941 }
 942
 943 static int lov_recreate(struct obd_export *exp, struct obdo *src_oa,
 944                         struct lov_stripe_md **ea, struct obd_trans_info *oti)
 945 {
 946         struct lov_stripe_md *obj_mdp, *lsm;
 947         struct lov_obd *lov = &exp->exp_obd->u.lov;
 948         unsigned ost_idx;
 949         int rc, i;
 950         ENTRY;
 951
 952         LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS &&
 953                 src_oa->o_flags & OBD_FL_RECREATE_OBJS);
 954
 955         OBD_ALLOC(obj_mdp, sizeof(*obj_mdp));
 956         if (obj_mdp == NULL)
 957                 RETURN(-ENOMEM);
 958
 959         ost_idx = src_oa->o_nlink;
 960         lsm = *ea;
 961         if (lsm == NULL)
 962                 GOTO(out, rc = -EINVAL);
 963         if (ost_idx >= lov->desc.ld_tgt_count ||
 964             !lov->lov_tgts[ost_idx])
 965                 GOTO(out, rc = -EINVAL);
 966
 967         for (i = 0; i < lsm->lsm_stripe_count; i++) {
 968                 if (lsm->lsm_oinfo[i]->loi_ost_idx == ost_idx) {
 969                         if (lsm->lsm_oinfo[i]->loi_id != src_oa->o_id)
 970                                 GOTO(out, rc = -EINVAL);
 971                         break;
 972                 }
 973         }
 974         if (i == lsm->lsm_stripe_count)
 975                 GOTO(out, rc = -EINVAL);
 976
 977         rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp, src_oa, &obj_mdp, oti);
 978 out:
 979         OBD_FREE(obj_mdp, sizeof(*obj_mdp));
 980         RETURN(rc);
 981 }
 982
 983 /* the LOV expects oa->o_id to be set to the LOV object id */
 984 static int lov_create(struct obd_export *exp, struct obdo *src_oa,
 985                       struct lov_stripe_md **ea, struct obd_trans_info *oti)
 986 {
 987         struct lov_obd *lov;
 988         struct obd_info oinfo;
 989         struct lov_request_set *set = NULL;
 990         struct lov_request *req;
 991         struct obd_statfs osfs;
 992         __u64 maxage;
 993         int rc = 0;
 994         ENTRY;
 995
 996         LASSERT(ea != NULL);
 997         if (exp == NULL)
 998                 RETURN(-EINVAL);
 999
1000         if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
1001             src_oa->o_flags == OBD_FL_DELORPHAN) {
1002                 rc = lov_clear_orphans(exp, src_oa, ea, oti);
1003                 RETURN(rc);
1004         }
1005
1006         lov = &exp->exp_obd->u.lov;
1007         if (!lov->desc.ld_active_tgt_count)
1008                 RETURN(-EIO);
1009
1010         /* Recreate a specific object id at the given OST index */
1011         if ((src_oa->o_valid & OBD_MD_FLFLAGS) &&
1012             (src_oa->o_flags & OBD_FL_RECREATE_OBJS)) {
1013                  rc = lov_recreate(exp, src_oa, ea, oti);
1014                  RETURN(rc);
1015         }
1016
1017         maxage = cfs_time_shift_64(-lov->desc.ld_qos_maxage);
1018         obd_statfs_rqset(exp->exp_obd, &osfs, maxage);
1019
1020         rc = lov_prep_create_set(exp, &oinfo, ea, src_oa, oti, &set);
1021         if (rc)
1022                 RETURN(rc);
1023
1024         list_for_each_entry(req, &set->set_list, rq_link) {
1025                 /* XXX: LOV STACKING: use real "obj_mdp" sub-data */
1026                 rc = obd_create(lov->lov_tgts[req->rq_idx]->ltd_exp,
1027                                 req->rq_oi.oi_oa, &req->rq_oi.oi_md, oti);
1028                 lov_update_create_set(set, req, rc);
1029         }
1030         rc = lov_fini_create_set(set, ea);
1031         RETURN(rc);
1032 }
1033
1034 #define ASSERT_LSM_MAGIC(lsmp)                                                  \
1035 do {                                                                            \
1036         LASSERT((lsmp) != NULL);                                                \
1037         LASSERTF(((lsmp)->lsm_magic == LOV_MAGIC ||                             \
1038                  (lsmp)->lsm_magic == LOV_MAGIC_JOIN), "%p->lsm_magic=%x\n",    \
1039                  (lsmp), (lsmp)->lsm_magic);                                    \
1040 } while (0)
1041
1042 static int lov_destroy(struct obd_export *exp, struct obdo *oa,
1043                        struct lov_stripe_md *lsm, struct obd_trans_info *oti,
1044                        struct obd_export *md_exp)
1045 {
1046         struct lov_request_set *set;
1047         struct obd_info oinfo;
1048         struct lov_request *req;
1049         struct list_head *pos;
1050         struct lov_obd *lov;
1051         int rc = 0, err;
1052         ENTRY;
1053
1054         ASSERT_LSM_MAGIC(lsm);
1055
1056         if (!exp || !exp->exp_obd)
1057                 RETURN(-ENODEV);
1058
1059         if (oa->o_valid & OBD_MD_FLCOOKIE) {
1060                 LASSERT(oti);
1061                 LASSERT(oti->oti_logcookies);
1062         }
1063
1064         lov = &exp->exp_obd->u.lov;
1065         rc = lov_prep_destroy_set(exp, &oinfo, oa, lsm, oti, &set);
1066         if (rc)
1067                 RETURN(rc);
1068
1069         list_for_each (pos, &set->set_list) {
1070                 int err;
1071                 req = list_entry(pos, struct lov_request, rq_link);
1072
1073                 if (oa->o_valid & OBD_MD_FLCOOKIE)
1074                         oti->oti_logcookies = set->set_cookies + req->rq_stripe;
1075
1076                 err = obd_destroy(lov->lov_tgts[req->rq_idx]->ltd_exp,
1077                                   req->rq_oi.oi_oa, NULL, oti, NULL);
1078                 err = lov_update_common_set(set, req, err);
1079                 if (err) {
1080                         CERROR("error: destroying objid "LPX64" subobj "
1081                                LPX64" on OST idx %d: rc = %d\n",
1082                                oa->o_id, req->rq_oi.oi_oa->o_id,
1083                                req->rq_idx, err);
1084                         if (!rc)
1085                                 rc = err;
1086                 }
1087         }
1088
1089         if (rc == 0) {
1090                 LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
1091                 rc = lsm_op_find(lsm->lsm_magic)->lsm_destroy(lsm, oa, md_exp);
1092         }
1093         err = lov_fini_destroy_set(set);
1094         RETURN(rc ? rc : err);
1095 }
1096
1097 static int lov_getattr(struct obd_export *exp, struct obd_info *oinfo)
1098 {
1099         struct lov_request_set *set;
1100         struct lov_request *req;
1101         struct list_head *pos;
1102         struct lov_obd *lov;
1103         int err = 0, rc = 0;
1104         ENTRY;
1105
1106         LASSERT(oinfo);
1107         ASSERT_LSM_MAGIC(oinfo->oi_md);
1108
1109         if (!exp || !exp->exp_obd)
1110                 RETURN(-ENODEV);
1111
1112         lov = &exp->exp_obd->u.lov;
1113
1114         rc = lov_prep_getattr_set(exp, oinfo, &set);
1115         if (rc)
1116                 RETURN(rc);
1117
1118         list_for_each (pos, &set->set_list) {
1119                 req = list_entry(pos, struct lov_request, rq_link);
1120
1121                 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
1122                        "%u\n", oinfo->oi_oa->o_id, req->rq_stripe,
1123                        req->rq_oi.oi_oa->o_id, req->rq_idx);
1124
1125                 rc = obd_getattr(lov->lov_tgts[req->rq_idx]->ltd_exp,
1126                                  &req->rq_oi);
1127                 err = lov_update_common_set(set, req, rc);
1128                 if (err) {
1129                         CERROR("error: getattr objid "LPX64" subobj "
1130                                LPX64" on OST idx %d: rc = %d\n",
1131                                oinfo->oi_oa->o_id, req->rq_oi.oi_oa->o_id,
1132                                req->rq_idx, err);
1133                         break;
1134                 }
1135         }
1136
1137         rc = lov_fini_getattr_set(set);
1138         if (err)
1139                 rc = err;
1140         RETURN(rc);
1141 }
1142
1143 static int lov_getattr_interpret(struct ptlrpc_request_set *rqset,
1144                                  void *data, int rc)
1145 {
1146         struct lov_request_set *lovset = (struct lov_request_set *)data;
1147         int err;
1148         ENTRY;
1149
1150         /* don't do attribute merge if this aysnc op failed */
1151         if (rc)
1152                 lovset->set_completes = 0;
1153         err = lov_fini_getattr_set(lovset);
1154         RETURN(rc ? rc : err);
1155 }
1156
1157 static int lov_getattr_async(struct obd_export *exp, struct obd_info *oinfo,
1158                               struct ptlrpc_request_set *rqset)
1159 {
1160         struct lov_request_set *lovset;
1161         struct lov_obd *lov;
1162         struct list_head *pos;
1163         struct lov_request *req;
1164         int rc = 0, err;
1165         ENTRY;
1166
1167         LASSERT(oinfo);
1168         ASSERT_LSM_MAGIC(oinfo->oi_md);
1169
1170         if (!exp || !exp->exp_obd)
1171                 RETURN(-ENODEV);
1172
1173         lov = &exp->exp_obd->u.lov;
1174
1175         rc = lov_prep_getattr_set(exp, oinfo, &lovset);
1176         if (rc)
1177                 RETURN(rc);
1178
1179         CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1180                oinfo->oi_md->lsm_object_id, oinfo->oi_md->lsm_stripe_count,
1181                oinfo->oi_md->lsm_stripe_size);
1182
1183         list_for_each (pos, &lovset->set_list) {
1184                 req = list_entry(pos, struct lov_request, rq_link);
1185
1186                 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
1187                        "%u\n", oinfo->oi_oa->o_id, req->rq_stripe,
1188                        req->rq_oi.oi_oa->o_id, req->rq_idx);
1189                 rc = obd_getattr_async(lov->lov_tgts[req->rq_idx]->ltd_exp,
1190                                        &req->rq_oi, rqset);
1191                 if (rc) {
1192                         CERROR("error: getattr objid "LPX64" subobj "
1193                                LPX64" on OST idx %d: rc = %d\n",
1194                                oinfo->oi_oa->o_id, req->rq_oi.oi_oa->o_id,
1195                                req->rq_idx, rc);
1196                         GOTO(out, rc);
1197                 }
1198         }
1199
1200         if (!list_empty(&rqset->set_requests)) {
1201                 LASSERT(rc == 0);
1202                 LASSERT (rqset->set_interpret == NULL);
1203                 rqset->set_interpret = lov_getattr_interpret;
1204                 rqset->set_arg = (void *)lovset;
1205                 RETURN(rc);
1206         }
1207 out:
1208         if (rc)
1209                 lovset->set_completes = 0;
1210         err = lov_fini_getattr_set(lovset);
1211         RETURN(rc ? rc : err);
1212 }
1213
1214 static int lov_setattr(struct obd_export *exp, struct obd_info *oinfo,
1215                        struct obd_trans_info *oti)
1216 {
1217         struct lov_request_set *set;
1218         struct lov_obd *lov;
1219         struct list_head *pos;
1220         struct lov_request *req;
1221         int err = 0, rc = 0;
1222         ENTRY;
1223
1224         LASSERT(oinfo);
1225         ASSERT_LSM_MAGIC(oinfo->oi_md);
1226
1227         if (!exp || !exp->exp_obd)
1228                 RETURN(-ENODEV);
1229
1230         /* for now, we only expect the following updates here */
1231         LASSERT(!(oinfo->oi_oa->o_valid & ~(OBD_MD_FLID | OBD_MD_FLTYPE |
1232                                             OBD_MD_FLMODE | OBD_MD_FLATIME |
1233                                             OBD_MD_FLMTIME | OBD_MD_FLCTIME |
1234                                             OBD_MD_FLFLAGS | OBD_MD_FLSIZE |
1235                                             OBD_MD_FLGROUP | OBD_MD_FLUID |
1236                                             OBD_MD_FLGID | OBD_MD_FLINLINE |
1237                                             OBD_MD_FLFID | OBD_MD_FLGENER)));
1238         lov = &exp->exp_obd->u.lov;
1239         rc = lov_prep_setattr_set(exp, oinfo, oti, &set);
1240         if (rc)
1241                 RETURN(rc);
1242
1243         list_for_each (pos, &set->set_list) {
1244                 req = list_entry(pos, struct lov_request, rq_link);
1245
1246                 rc = obd_setattr(lov->lov_tgts[req->rq_idx]->ltd_exp,
1247                                  &req->rq_oi, NULL);
1248                 err = lov_update_setattr_set(set, req, rc);
1249                 if (err) {
1250                         CERROR("error: setattr objid "LPX64" subobj "
1251                                LPX64" on OST idx %d: rc = %d\n",
1252                                set->set_oi->oi_oa->o_id,
1253                                req->rq_oi.oi_oa->o_id, req->rq_idx, err);
1254                         if (!rc)
1255                                 rc = err;
1256                 }
1257         }
1258         err = lov_fini_setattr_set(set);
1259         if (!rc)
1260                 rc = err;
1261         RETURN(rc);
1262 }
1263
1264 static int lov_setattr_interpret(struct ptlrpc_request_set *rqset,
1265                                  void *data, int rc)
1266 {
1267         struct lov_request_set *lovset = (struct lov_request_set *)data;
1268         int err;
1269         ENTRY;
1270
1271         if (rc)
1272                 lovset->set_completes = 0;
1273         err = lov_fini_setattr_set(lovset);
1274         RETURN(rc ? rc : err);
1275 }
1276
1277 /* If @oti is given, the request goes from MDS and responses from OSTs are not
1278    needed. Otherwise, a client is waiting for responses. */
1279 static int lov_setattr_async(struct obd_export *exp, struct obd_info *oinfo,
1280                              struct obd_trans_info *oti,
1281                              struct ptlrpc_request_set *rqset)
1282 {
1283         struct lov_request_set *set;
1284         struct lov_request *req;
1285         struct list_head *pos;
1286         struct lov_obd *lov;
1287         int rc = 0;
1288         ENTRY;
1289
1290         LASSERT(oinfo);
1291         ASSERT_LSM_MAGIC(oinfo->oi_md);
1292         if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE) {
1293                 LASSERT(oti);
1294                 LASSERT(oti->oti_logcookies);
1295         }
1296
1297         if (!exp || !exp->exp_obd)
1298                 RETURN(-ENODEV);
1299
1300         lov = &exp->exp_obd->u.lov;
1301         rc = lov_prep_setattr_set(exp, oinfo, oti, &set);
1302         if (rc)
1303                 RETURN(rc);
1304
1305         CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n",
1306                oinfo->oi_md->lsm_object_id, oinfo->oi_md->lsm_stripe_count,
1307                oinfo->oi_md->lsm_stripe_size);
1308
1309         list_for_each (pos, &set->set_list) {
1310                 req = list_entry(pos, struct lov_request, rq_link);
1311
1312                 if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE)
1313                         oti->oti_logcookies = set->set_cookies + req->rq_stripe;
1314
1315                 CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx "
1316                        "%u\n", oinfo->oi_oa->o_id, req->rq_stripe,
1317                        req->rq_oi.oi_oa->o_id, req->rq_idx);
1318
1319                 rc = obd_setattr_async(lov->lov_tgts[req->rq_idx]->ltd_exp,
1320                                        &req->rq_oi, oti, rqset);
1321                 if (rc) {
1322                         CERROR("error: setattr objid "LPX64" subobj "
1323                                LPX64" on OST idx %d: rc = %d\n",
1324                                set->set_oi->oi_oa->o_id,
1325                                req->rq_oi.oi_oa->o_id,
1326                                req->rq_idx, rc);
1327                         break;
1328                 }
1329         }
1330
1331         /* If we are not waiting for responses on async requests, return. */
1332         if (rc || !rqset || list_empty(&rqset->set_requests)) {
1333                 int err;
1334                 if (rc)
1335                         set->set_completes = 0;
1336                 err = lov_fini_setattr_set(set);
1337                 RETURN(rc ? rc : err);
1338         }
1339
1340         LASSERT(rqset->set_interpret == NULL);
1341         rqset->set_interpret = lov_setattr_interpret;
1342         rqset->set_arg = (void *)set;
1343
1344         RETURN(0);
1345 }
1346
1347 static int lov_punch_interpret(struct ptlrpc_request_set *rqset,
1348                                void *data, int rc)
1349 {
1350         struct lov_request_set *lovset = (struct lov_request_set *)data;
1351         int err;
1352         ENTRY;
1353
1354         if (rc)
1355                 lovset->set_completes = 0;
1356         err = lov_fini_punch_set(lovset);
1357         RETURN(rc ? rc : err);
1358 }
1359
1360 /* FIXME: maybe we'll just make one node the authoritative attribute node, then
1361  * we can send this 'punch' to just the authoritative node and the nodes
1362  * that the punch will affect. */
1363 static int lov_punch(struct obd_export *exp, struct obd_info *oinfo,
1364                      struct obd_trans_info *oti,
1365                      struct ptlrpc_request_set *rqset)
1366 {
1367         struct lov_request_set *set;
1368         struct lov_obd *lov;
1369         struct list_head *pos;
1370         struct lov_request *req;
1371         int rc = 0;
1372         ENTRY;
1373
1374         LASSERT(oinfo);
1375         ASSERT_LSM_MAGIC(oinfo->oi_md);
1376
1377         if (!exp || !exp->exp_obd)
1378                 RETURN(-ENODEV);
1379
1380         lov = &exp->exp_obd->u.lov;
1381         rc = lov_prep_punch_set(exp, oinfo, oti, &set);
1382         if (rc)
1383                 RETURN(rc);
1384
1385         list_for_each (pos, &set->set_list) {
1386                 req = list_entry(pos, struct lov_request, rq_link);
1387
1388                 rc = obd_punch(lov->lov_tgts[req->rq_idx]->ltd_exp,
1389                                &req->rq_oi, NULL, rqset);
1390                 if (rc) {
1391                         CERROR("error: punch objid "LPX64" subobj "LPX64
1392                                " on OST idx %d: rc = %d\n",
1393                                set->set_oi->oi_oa->o_id,
1394                                req->rq_oi.oi_oa->o_id, req->rq_idx, rc);
1395                         break;
1396                 }
1397         }
1398
1399         if (rc || list_empty(&rqset->set_requests)) {
1400                 int err;
1401                 err = lov_fini_punch_set(set);
1402                 RETURN(rc ? rc : err);
1403         }
1404
1405         LASSERT(rqset->set_interpret == NULL);
1406         rqset->set_interpret = lov_punch_interpret;
1407         rqset->set_arg = (void *)set;
1408
1409         RETURN(0);
1410 }
1411
1412 static int lov_sync(struct obd_export *exp, struct obdo *oa,
1413                     struct lov_stripe_md *lsm, obd_off start, obd_off end)
1414 {
1415         struct lov_request_set *set;
1416         struct obd_info oinfo;
1417         struct lov_obd *lov;
1418         struct list_head *pos;
1419         struct lov_request *req;
1420         int err = 0, rc = 0;
1421         ENTRY;
1422
1423         ASSERT_LSM_MAGIC(lsm);
1424
1425         if (!exp->exp_obd)
1426                 RETURN(-ENODEV);
1427
1428         lov = &exp->exp_obd->u.lov;
1429         rc = lov_prep_sync_set(exp, &oinfo, oa, lsm, start, end, &set);
1430         if (rc)
1431                 RETURN(rc);
1432
1433         list_for_each (pos, &set->set_list) {
1434                 req = list_entry(pos, struct lov_request, rq_link);
1435
1436                 rc = obd_sync(lov->lov_tgts[req->rq_idx]->ltd_exp,
1437                               req->rq_oi.oi_oa, NULL,
1438                               req->rq_oi.oi_policy.l_extent.start,
1439                               req->rq_oi.oi_policy.l_extent.end);
1440                 err = lov_update_common_set(set, req, rc);
1441                 if (err) {
1442                         CERROR("error: fsync objid "LPX64" subobj "LPX64
1443                                " on OST idx %d: rc = %d\n",
1444                                set->set_oi->oi_oa->o_id,
1445                                req->rq_oi.oi_oa->o_id, req->rq_idx, rc);
1446                         if (!rc)
1447                                 rc = err;
1448                 }
1449         }
1450         err = lov_fini_sync_set(set);
1451         if (!rc)
1452                 rc = err;
1453         RETURN(rc);
1454 }
1455
1456 static int lov_brw_check(struct lov_obd *lov, struct obd_info *lov_oinfo,
1457                          obd_count oa_bufs, struct brw_page *pga)
1458 {
1459         struct obd_info oinfo = { { { 0 } } };
1460         int i, rc = 0;
1461
1462         oinfo.oi_oa = lov_oinfo->oi_oa;
1463
1464         /* The caller just wants to know if there's a chance that this
1465          * I/O can succeed */
1466         for (i = 0; i < oa_bufs; i++) {
1467                 int stripe = lov_stripe_number(lov_oinfo->oi_md, pga[i].off);
1468                 int ost = lov_oinfo->oi_md->lsm_oinfo[stripe]->loi_ost_idx;
1469                 obd_off start, end;
1470
1471                 if (!lov_stripe_intersects(lov_oinfo->oi_md, i, pga[i].off,
1472                                            pga[i].off + pga[i].count,
1473                                            &start, &end))
1474                         continue;
1475
1476                 if (!lov->lov_tgts[ost] || !lov->lov_tgts[ost]->ltd_active) {
1477                         CDEBUG(D_HA, "lov idx %d inactive\n", ost);
1478                         return -EIO;
1479                 }
1480
1481                 rc = obd_brw(OBD_BRW_CHECK, lov->lov_tgts[ost]->ltd_exp, &oinfo,
1482                              1, &pga[i], NULL);
1483                 if (rc)
1484                         break;
1485         }
1486         return rc;
1487 }
1488
1489 static int lov_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo,
1490                    obd_count oa_bufs, struct brw_page *pga,
1491                    struct obd_trans_info *oti)
1492 {
1493         struct lov_request_set *set;
1494         struct lov_request *req;
1495         struct list_head *pos;
1496         struct lov_obd *lov = &exp->exp_obd->u.lov;
1497         int err, rc = 0;
1498         ENTRY;
1499
1500         ASSERT_LSM_MAGIC(oinfo->oi_md);
1501
1502         if (cmd == OBD_BRW_CHECK) {
1503                 rc = lov_brw_check(lov, oinfo, oa_bufs, pga);
1504                 RETURN(rc);
1505         }
1506
1507         rc = lov_prep_brw_set(exp, oinfo, oa_bufs, pga, oti, &set);
1508         if (rc)
1509                 RETURN(rc);
1510
1511         list_for_each (pos, &set->set_list) {
1512                 struct obd_export *sub_exp;
1513                 struct brw_page *sub_pga;
1514                 req = list_entry(pos, struct lov_request, rq_link);
1515
1516                 sub_exp = lov->lov_tgts[req->rq_idx]->ltd_exp;
1517                 sub_pga = set->set_pga + req->rq_pgaidx;
1518                 rc = obd_brw(cmd, sub_exp, &req->rq_oi, req->rq_oabufs,
1519                              sub_pga, oti);
1520                 if (rc)
1521                         break;
1522                 lov_update_common_set(set, req, rc);
1523         }
1524
1525         err = lov_fini_brw_set(set);
1526         if (!rc)
1527                 rc = err;
1528         RETURN(rc);
1529 }
1530
1531 static int lov_brw_interpret(struct ptlrpc_request_set *reqset, void *data,
1532                              int rc)
1533 {
1534         struct lov_request_set *lovset = (struct lov_request_set *)data;
1535         ENTRY;
1536
1537         if (rc) {
1538                 lovset->set_completes = 0;
1539                 lov_fini_brw_set(lovset);
1540         } else {
1541                 rc = lov_fini_brw_set(lovset);
1542         }
1543
1544         RETURN(rc);
1545 }
1546
1547 static int lov_brw_async(int cmd, struct obd_export *exp,
1548                          struct obd_info *oinfo, obd_count oa_bufs,
1549                          struct brw_page *pga, struct obd_trans_info *oti,
1550                          struct ptlrpc_request_set *set)
1551 {
1552         struct lov_request_set *lovset;
1553         struct lov_request *req;
1554         struct list_head *pos;
1555         struct lov_obd *lov = &exp->exp_obd->u.lov;
1556         int rc = 0;
1557         ENTRY;
1558
1559         LASSERT(oinfo);
1560         ASSERT_LSM_MAGIC(oinfo->oi_md);
1561
1562         if (cmd == OBD_BRW_CHECK) {
1563                 rc = lov_brw_check(lov, oinfo, oa_bufs, pga);
1564                 RETURN(rc);
1565         }
1566
1567         rc = lov_prep_brw_set(exp, oinfo, oa_bufs, pga, oti, &lovset);
1568         if (rc)
1569                 RETURN(rc);
1570
1571         list_for_each (pos, &lovset->set_list) {
1572                 struct obd_export *sub_exp;
1573                 struct brw_page *sub_pga;
1574                 req = list_entry(pos, struct lov_request, rq_link);
1575
1576                 sub_exp = lov->lov_tgts[req->rq_idx]->ltd_exp;
1577                 sub_pga = lovset->set_pga + req->rq_pgaidx;
1578                 rc = obd_brw_async(cmd, sub_exp, &req->rq_oi, req->rq_oabufs,
1579                                    sub_pga, oti, set);
1580                 if (rc)
1581                         GOTO(out, rc);
1582                 lov_update_common_set(lovset, req, rc);
1583         }
1584         LASSERT(rc == 0);
1585         LASSERT(set->set_interpret == NULL);
1586         LASSERT(set->set_arg == NULL);
1587         rc = ptlrpc_set_add_cb(set, lov_brw_interpret, lovset);
1588         if (rc)
1589                 GOTO(out, rc);
1590
1591         RETURN(rc);
1592 out:
1593         lov_fini_brw_set(lovset);
1594         RETURN(rc);
1595 }
1596
1597 static int lov_ap_make_ready(void *data, int cmd)
1598 {
1599         struct lov_async_page *lap = LAP_FROM_COOKIE(data);
1600
1601         return lap->lap_caller_ops->ap_make_ready(lap->lap_caller_data, cmd);
1602 }
1603
1604 static int lov_ap_refresh_count(void *data, int cmd)
1605 {
1606         struct lov_async_page *lap = LAP_FROM_COOKIE(data);
1607
1608         return lap->lap_caller_ops->ap_refresh_count(lap->lap_caller_data,
1609                                                      cmd);
1610 }
1611
1612 static void lov_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
1613 {
1614         struct lov_async_page *lap = LAP_FROM_COOKIE(data);
1615
1616         lap->lap_caller_ops->ap_fill_obdo(lap->lap_caller_data, cmd, oa);
1617         /* XXX woah, shouldn't we be altering more here?  size? */
1618         oa->o_id = lap->lap_loi_id;
1619         oa->o_stripe_idx = lap->lap_stripe;
1620 }
1621
1622 static void lov_ap_update_obdo(void *data, int cmd, struct obdo *oa,
1623                                obd_valid valid)
1624 {
1625         struct lov_async_page *lap = LAP_FROM_COOKIE(data);
1626
1627         lap->lap_caller_ops->ap_update_obdo(lap->lap_caller_data, cmd,oa,valid);
1628 }
1629
1630 static int lov_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
1631 {
1632         struct lov_async_page *lap = LAP_FROM_COOKIE(data);
1633
1634         /* in a raid1 regime this would down a count of many ios
1635          * in flight, onl calling the caller_ops completion when all
1636          * the raid1 ios are complete */
1637         rc = lap->lap_caller_ops->ap_completion(lap->lap_caller_data,cmd,oa,rc);
1638         return rc;
1639 }
1640
1641 static struct obd_async_page_ops lov_async_page_ops = {
1642         .ap_make_ready =        lov_ap_make_ready,
1643         .ap_refresh_count =     lov_ap_refresh_count,
1644         .ap_fill_obdo =         lov_ap_fill_obdo,
1645         .ap_update_obdo =       lov_ap_update_obdo,
1646         .ap_completion =        lov_ap_completion,
1647 };
1648
1649 int lov_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1650                            struct lov_oinfo *loi, cfs_page_t *page,
1651                            obd_off offset, struct obd_async_page_ops *ops,
1652                            void *data, void **res)
1653 {
1654         struct lov_obd *lov = &exp->exp_obd->u.lov;
1655         struct lov_async_page *lap;
1656         int rc = 0;
1657         ENTRY;
1658
1659         if (!page) {
1660                 int i = 0;
1661                 /* Find an existing osc so we can get it's stupid sizeof(*oap).
1662                    Only because of this layering limitation will a client
1663                    mount with no osts fail */
1664                 while (!lov->lov_tgts || !lov->lov_tgts[i] ||
1665                        !lov->lov_tgts[i]->ltd_exp) {
1666                         i++;
1667                         if (i >= lov->desc.ld_tgt_count)
1668                                 RETURN(-ENOMEDIUM);
1669                 }
1670                 rc = size_round(sizeof(*lap)) +
1671                         obd_prep_async_page(lov->lov_tgts[i]->ltd_exp, NULL,
1672                                             NULL, NULL, 0, NULL, NULL, NULL);
1673                 RETURN(rc);
1674         }
1675         ASSERT_LSM_MAGIC(lsm);
1676         LASSERT(loi == NULL);
1677
1678         lap = *res;
1679         lap->lap_magic = LOV_AP_MAGIC;
1680         lap->lap_caller_ops = ops;
1681         lap->lap_caller_data = data;
1682
1683         /* for now only raid 0 which passes through */
1684         lap->lap_stripe = lov_stripe_number(lsm, offset);
1685         lov_stripe_offset(lsm, offset, lap->lap_stripe, &lap->lap_sub_offset);
1686         loi = lsm->lsm_oinfo[lap->lap_stripe];
1687
1688         /* so the callback doesn't need the lsm */
1689         lap->lap_loi_id = loi->loi_id;
1690
1691         lap->lap_sub_cookie = (void *)lap + size_round(sizeof(*lap));
1692
1693         rc = obd_prep_async_page(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
1694                                  lsm, loi, page, lap->lap_sub_offset,
1695                                  &lov_async_page_ops, lap,
1696                                  &lap->lap_sub_cookie);
1697         if (rc)
1698                 RETURN(rc);
1699         CDEBUG(D_CACHE, "lap %p page %p cookie %p off "LPU64"\n", lap, page,
1700                lap->lap_sub_cookie, offset);
1701         RETURN(0);
1702 }
1703
1704 static int lov_queue_async_io(struct obd_export *exp,
1705                               struct lov_stripe_md *lsm,
1706                               struct lov_oinfo *loi, void *cookie,
1707                               int cmd, obd_off off, int count,
1708                               obd_flag brw_flags, obd_flag async_flags)
1709 {
1710         struct lov_obd *lov = &exp->exp_obd->u.lov;
1711         struct lov_async_page *lap;
1712         int rc;
1713
1714         LASSERT(loi == NULL);
1715
1716         ASSERT_LSM_MAGIC(lsm);
1717
1718         lap = LAP_FROM_COOKIE(cookie);
1719
1720         loi = lsm->lsm_oinfo[lap->lap_stripe];
1721
1722         rc = obd_queue_async_io(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp, lsm,
1723                                 loi, lap->lap_sub_cookie, cmd, off, count,
1724                                 brw_flags, async_flags);
1725         RETURN(rc);
1726 }
1727
1728 static int lov_set_async_flags(struct obd_export *exp,
1729                                struct lov_stripe_md *lsm,
1730                                struct lov_oinfo *loi, void *cookie,
1731                                obd_flag async_flags)
1732 {
1733         struct lov_obd *lov = &exp->exp_obd->u.lov;
1734         struct lov_async_page *lap;
1735         int rc;
1736
1737         LASSERT(loi == NULL);
1738
1739         ASSERT_LSM_MAGIC(lsm);
1740
1741         lap = LAP_FROM_COOKIE(cookie);
1742
1743         loi = lsm->lsm_oinfo[lap->lap_stripe];
1744
1745         rc = obd_set_async_flags(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
1746                                  lsm, loi, lap->lap_sub_cookie, async_flags);
1747         RETURN(rc);
1748 }
1749
1750 static int lov_queue_group_io(struct obd_export *exp,
1751                               struct lov_stripe_md *lsm,
1752                               struct lov_oinfo *loi,
1753                               struct obd_io_group *oig, void *cookie,
1754                               int cmd, obd_off off, int count,
1755                               obd_flag brw_flags, obd_flag async_flags)
1756 {
1757         struct lov_obd *lov = &exp->exp_obd->u.lov;
1758         struct lov_async_page *lap;
1759         int rc;
1760
1761         LASSERT(loi == NULL);
1762
1763         ASSERT_LSM_MAGIC(lsm);
1764
1765         lap = LAP_FROM_COOKIE(cookie);
1766
1767         loi = lsm->lsm_oinfo[lap->lap_stripe];
1768
1769         rc = obd_queue_group_io(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp, lsm,
1770                                 loi, oig, lap->lap_sub_cookie, cmd, off, count,
1771                                 brw_flags, async_flags);
1772         RETURN(rc);
1773 }
1774
1775 /* this isn't exactly optimal.  we may have queued sync io in oscs on
1776  * all stripes, but we don't record that fact at queue time.  so we
1777  * trigger sync io on all stripes. */
1778 static int lov_trigger_group_io(struct obd_export *exp,
1779                                 struct lov_stripe_md *lsm,
1780                                 struct lov_oinfo *loi,
1781                                 struct obd_io_group *oig)
1782 {
1783         struct lov_obd *lov = &exp->exp_obd->u.lov;
1784         int rc = 0, i, err;
1785
1786         LASSERT(loi == NULL);
1787
1788         ASSERT_LSM_MAGIC(lsm);
1789
1790         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1791                 loi = lsm->lsm_oinfo[i];
1792                 if (!lov->lov_tgts[loi->loi_ost_idx] ||
1793                     !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
1794                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
1795                         continue;
1796                 }
1797
1798                 err = obd_trigger_group_io(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
1799                                            lsm, loi, oig);
1800                 if (rc == 0 && err != 0)
1801                         rc = err;
1802         };
1803         RETURN(rc);
1804 }
1805
1806 static int lov_teardown_async_page(struct obd_export *exp,
1807                                    struct lov_stripe_md *lsm,
1808                                    struct lov_oinfo *loi, void *cookie)
1809 {
1810         struct lov_obd *lov = &exp->exp_obd->u.lov;
1811         struct lov_async_page *lap;
1812         int rc;
1813
1814         LASSERT(loi == NULL);
1815
1816         ASSERT_LSM_MAGIC(lsm);
1817
1818         lap = LAP_FROM_COOKIE(cookie);
1819
1820         loi = lsm->lsm_oinfo[lap->lap_stripe];
1821
1822         rc = obd_teardown_async_page(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
1823                                      lsm, loi, lap->lap_sub_cookie);
1824         if (rc) {
1825                 CERROR("unable to teardown sub cookie %p: %d\n",
1826                        lap->lap_sub_cookie, rc);
1827                 RETURN(rc);
1828         }
1829         RETURN(rc);
1830 }
1831
1832 static int lov_enqueue_interpret(struct ptlrpc_request_set *rqset,
1833                                  void *data, int rc)
1834 {
1835         struct lov_request_set *lovset = (struct lov_request_set *)data;
1836         ENTRY;
1837         rc = lov_fini_enqueue_set(lovset, lovset->set_ei->ei_mode, rc, rqset);
1838         RETURN(rc);
1839 }
1840
1841 static int lov_enqueue(struct obd_export *exp, struct obd_info *oinfo,
1842                        struct ldlm_enqueue_info *einfo,
1843                        struct ptlrpc_request_set *rqset)
1844 {
1845         struct lov_request_set *set;
1846         struct lov_request *req;
1847         struct list_head *pos;
1848         struct lov_obd *lov;
1849         ldlm_error_t rc;
1850         ENTRY;
1851
1852         LASSERT(oinfo);
1853         ASSERT_LSM_MAGIC(oinfo->oi_md);
1854
1855         /* we should never be asked to replay a lock this way. */
1856         LASSERT((oinfo->oi_flags & LDLM_FL_REPLAY) == 0);
1857
1858         if (!exp || !exp->exp_obd)
1859                 RETURN(-ENODEV);
1860
1861         lov = &exp->exp_obd->u.lov;
1862         rc = lov_prep_enqueue_set(exp, oinfo, einfo, &set);
1863         if (rc)
1864                 RETURN(rc);
1865
1866         list_for_each (pos, &set->set_list) {
1867                 req = list_entry(pos, struct lov_request, rq_link);
1868
1869                 rc = obd_enqueue(lov->lov_tgts[req->rq_idx]->ltd_exp,
1870                                  &req->rq_oi, einfo, rqset);
1871                 if (rc != ELDLM_OK)
1872                         GOTO(out, rc);
1873         }
1874
1875         if (rqset && !list_empty(&rqset->set_requests)) {
1876                 LASSERT(rc == 0);
1877                 LASSERT(rqset->set_interpret == NULL);
1878                 rqset->set_interpret = lov_enqueue_interpret;
1879                 rqset->set_arg = (void *)set;
1880                 RETURN(rc);
1881         }
1882 out:
1883         rc = lov_fini_enqueue_set(set, einfo->ei_mode, rc, rqset);
1884         RETURN(rc);
1885 }
1886
1887 static int lov_match(struct obd_export *exp, struct lov_stripe_md *lsm,
1888                      __u32 type, ldlm_policy_data_t *policy, __u32 mode,
1889                      int *flags, void *data, struct lustre_handle *lockh)
1890 {
1891         struct lov_request_set *set;
1892         struct obd_info oinfo;
1893         struct lov_request *req;
1894         struct list_head *pos;
1895         struct lov_obd *lov = &exp->exp_obd->u.lov;
1896         struct lustre_handle *lov_lockhp;
1897         int lov_flags, rc = 0;
1898         ENTRY;
1899
1900         ASSERT_LSM_MAGIC(lsm);
1901
1902         if (!exp || !exp->exp_obd)
1903                 RETURN(-ENODEV);
1904
1905         lov = &exp->exp_obd->u.lov;
1906         rc = lov_prep_match_set(exp, &oinfo, lsm, policy, mode, lockh, &set);
1907         if (rc)
1908                 RETURN(rc);
1909
1910         list_for_each (pos, &set->set_list) {
1911                 ldlm_policy_data_t sub_policy;
1912                 req = list_entry(pos, struct lov_request, rq_link);
1913                 lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
1914                 LASSERT(lov_lockhp);
1915
1916                 lov_flags = *flags;
1917                 sub_policy.l_extent = req->rq_oi.oi_policy.l_extent;
1918
1919                 rc = obd_match(lov->lov_tgts[req->rq_idx]->ltd_exp,
1920                                req->rq_oi.oi_md, type, &sub_policy,
1921                                mode, &lov_flags, data, lov_lockhp);
1922                 rc = lov_update_match_set(set, req, rc);
1923                 if (rc != 1)
1924                         break;
1925         }
1926         lov_fini_match_set(set, mode, *flags);
1927         RETURN(rc);
1928 }
1929
1930 static int lov_change_cbdata(struct obd_export *exp,
1931                              struct lov_stripe_md *lsm, ldlm_iterator_t it,
1932                              void *data)
1933 {
1934         struct lov_obd *lov;
1935         struct lov_oinfo *loi;
1936         int rc = 0, i;
1937         ENTRY;
1938
1939         ASSERT_LSM_MAGIC(lsm);
1940
1941         if (!exp || !exp->exp_obd)
1942                 RETURN(-ENODEV);
1943
1944         lov = &exp->exp_obd->u.lov;
1945         for (i = 0; i < lsm->lsm_stripe_count; i++) {
1946                 struct lov_stripe_md submd;
1947
1948                 loi = lsm->lsm_oinfo[i];
1949                 if (!lov->lov_tgts[loi->loi_ost_idx]) {
1950                         CDEBUG(D_HA, "lov idx %d NULL \n", loi->loi_ost_idx);
1951                         continue;
1952                 }
1953                 submd.lsm_object_id = loi->loi_id;
1954                 submd.lsm_stripe_count = 0;
1955                 rc = obd_change_cbdata(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
1956                                        &submd, it, data);
1957         }
1958         RETURN(rc);
1959 }
1960
1961 static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm,
1962                       __u32 mode, struct lustre_handle *lockh)
1963 {
1964         struct lov_request_set *set;
1965         struct obd_info oinfo;
1966         struct lov_request *req;
1967         struct list_head *pos;
1968         struct lov_obd *lov = &exp->exp_obd->u.lov;
1969         struct lustre_handle *lov_lockhp;
1970         ldlm_mode_t this_mode;
1971         int err = 0, rc = 0;
1972         ENTRY;
1973
1974         ASSERT_LSM_MAGIC(lsm);
1975
1976         if (!exp || !exp->exp_obd)
1977                 RETURN(-ENODEV);
1978
1979         LASSERT(lockh);
1980         lov = &exp->exp_obd->u.lov;
1981         rc = lov_prep_cancel_set(exp, &oinfo, lsm, mode, lockh, &set);
1982         if (rc)
1983                 RETURN(rc);
1984
1985         list_for_each (pos, &set->set_list) {
1986                 req = list_entry(pos, struct lov_request, rq_link);
1987                 lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe;
1988
1989                 /* If this lock was used for a write or truncate, the object
1990                  * will have been recreated by the OST, cancel the lock
1991                  * (setting LCK_GROUP incidentally causes immediate cancel). */
1992                 if (OST_LVB_IS_ERR(lsm->lsm_oinfo[req->rq_stripe]->loi_lvb.lvb_blocks) &&
1993                     (mode == LCK_PW || mode == LCK_CW))
1994                         this_mode = LCK_GROUP;
1995                 else
1996                         this_mode = mode;
1997
1998                 rc = obd_cancel(lov->lov_tgts[req->rq_idx]->ltd_exp,
1999                                 req->rq_oi.oi_md, this_mode, lov_lockhp);
2000                 rc = lov_update_common_set(set, req, rc);
2001                 if (rc) {
2002                         CERROR("error: cancel objid "LPX64" subobj "
2003                                LPX64" on OST idx %d: rc = %d\n",
2004                                lsm->lsm_object_id,
2005                                req->rq_oi.oi_md->lsm_object_id,
2006                                req->rq_idx, rc);
2007                         err = rc;
2008                 }
2009
2010         }
2011         lov_fini_cancel_set(set);
2012         RETURN(err);
2013 }
2014
2015 static int lov_cancel_unused(struct obd_export *exp,
2016                              struct lov_stripe_md *lsm, int flags, void *opaque)
2017 {
2018         struct lov_obd *lov;
2019         struct lov_oinfo *loi;
2020         int rc = 0, i;
2021         ENTRY;
2022
2023         if (!exp || !exp->exp_obd)
2024                 RETURN(-ENODEV);
2025
2026         lov = &exp->exp_obd->u.lov;
2027         if (lsm == NULL) {
2028                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2029                         int err;
2030                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
2031                                 continue;
2032
2033                         err = obd_cancel_unused(lov->lov_tgts[i]->ltd_exp, NULL,
2034                                                 flags, opaque);
2035                         if (!rc)
2036                                 rc = err;
2037                 }
2038                 RETURN(rc);
2039         }
2040
2041         ASSERT_LSM_MAGIC(lsm);
2042
2043         for (i = 0; i < lsm->lsm_stripe_count; i++) {
2044                 struct lov_stripe_md submd;
2045                 int err;
2046
2047                 loi = lsm->lsm_oinfo[i];
2048                 if (!lov->lov_tgts[loi->loi_ost_idx]) {
2049                         CDEBUG(D_HA, "lov idx %d NULL\n", loi->loi_ost_idx);
2050                         continue;
2051                 }
2052
2053                 if (!lov->lov_tgts[loi->loi_ost_idx]->ltd_active)
2054                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2055
2056                 submd.lsm_object_id = loi->loi_id;
2057                 submd.lsm_stripe_count = 0;
2058                 err = obd_cancel_unused(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
2059                                         &submd, flags, opaque);
2060                 if (err && lov->lov_tgts[loi->loi_ost_idx]->ltd_active) {
2061                         CERROR("error: cancel unused objid "LPX64" subobj "LPX64
2062                                " on OST idx %d: rc = %d\n", lsm->lsm_object_id,
2063                                loi->loi_id, loi->loi_ost_idx, err);
2064                         if (!rc)
2065                                 rc = err;
2066                 }
2067         }
2068         RETURN(rc);
2069 }
2070
2071 static int lov_join_lru(struct obd_export *exp,
2072                         struct lov_stripe_md *lsm, int join)
2073 {
2074         struct lov_obd *lov;
2075         struct lov_oinfo *loi;
2076         int i, count = 0;
2077         ENTRY;
2078
2079         ASSERT_LSM_MAGIC(lsm);
2080         if (!exp || !exp->exp_obd)
2081                 RETURN(-ENODEV);
2082
2083         lov = &exp->exp_obd->u.lov;
2084         for (i = 0; i < lsm->lsm_stripe_count; i++) {
2085                 struct lov_stripe_md submd;
2086                 int rc = 0;
2087
2088                 loi = lsm->lsm_oinfo[i];
2089                 if (!lov->lov_tgts[loi->loi_ost_idx]) {
2090                         CDEBUG(D_HA, "lov idx %d NULL\n", loi->loi_ost_idx);
2091                         continue;
2092                 }
2093
2094                 if (!lov->lov_tgts[loi->loi_ost_idx]->ltd_active)
2095                         CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx);
2096
2097                 submd.lsm_object_id = loi->loi_id;
2098                 submd.lsm_stripe_count = 0;
2099                 rc = obd_join_lru(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp,
2100                                   &submd, join);
2101                 if (rc < 0) {
2102                         CERROR("join lru failed. objid: "LPX64" subobj: "LPX64
2103                                " ostidx: %d rc: %d\n", lsm->lsm_object_id,
2104                                loi->loi_id, loi->loi_ost_idx, rc);
2105                         return rc;
2106                 } else {
2107                         count += rc;
2108                 }
2109         }
2110         RETURN(count);
2111 }
2112
2113 static int lov_statfs_interpret(struct ptlrpc_request_set *rqset,
2114                                 void *data, int rc)
2115 {
2116         struct lov_request_set *lovset = (struct lov_request_set *)data;
2117         int err;
2118         ENTRY;
2119
2120         if (rc)
2121                 lovset->set_completes = 0;
2122
2123         err = lov_fini_statfs_set(lovset);
2124         RETURN(rc ? rc : err);
2125 }
2126
2127 static int lov_statfs_async(struct obd_device *obd, struct obd_info *oinfo,
2128                             __u64 max_age, struct ptlrpc_request_set *rqset)
2129 {
2130         struct lov_request_set *set;
2131         struct lov_request *req;
2132         struct list_head *pos;
2133         struct lov_obd *lov;
2134         int rc = 0;
2135         ENTRY;
2136
2137         LASSERT(oinfo != NULL);
2138         LASSERT(oinfo->oi_osfs != NULL);
2139
2140         lov = &obd->u.lov;
2141         rc = lov_prep_statfs_set(obd, oinfo, &set);
2142         if (rc)
2143                 RETURN(rc);
2144
2145         list_for_each (pos, &set->set_list) {
2146                 struct obd_device *osc_obd;
2147
2148                 req = list_entry(pos, struct lov_request, rq_link);
2149
2150                 osc_obd = class_exp2obd(lov->lov_tgts[req->rq_idx]->ltd_exp);
2151                 rc = obd_statfs_async(osc_obd, &req->rq_oi, max_age, rqset);
2152                 if (rc)
2153                         break;
2154         }
2155
2156         if (rc || list_empty(&rqset->set_requests)) {
2157                 int err;
2158                 if (rc)
2159                         set->set_completes = 0;
2160                 err = lov_fini_statfs_set(set);
2161                 RETURN(rc ? rc : err);
2162         }
2163
2164         LASSERT(rqset->set_interpret == NULL);
2165         rqset->set_interpret = lov_statfs_interpret;
2166         rqset->set_arg = (void *)set;
2167         RETURN(0);
2168 }
2169
2170 static int lov_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2171                       __u64 max_age)
2172 {
2173         struct lov_obd *lov = &obd->u.lov;
2174         struct obd_statfs lov_sfs;
2175         int set = 0;
2176         int rc = 0, err;
2177         int i;
2178         ENTRY;
2179
2180         /* We only get block data from the OBD */
2181         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2182                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active) {
2183                         CDEBUG(D_HA, "lov idx %d inactive\n", i);
2184                         continue;
2185                 }
2186
2187                 err = obd_statfs(class_exp2obd(lov->lov_tgts[i]->ltd_exp),
2188                                  &lov_sfs, max_age);
2189                 if (err) {
2190                         if (lov->lov_tgts[i]->ltd_active && !rc)
2191                                 rc = err;
2192                         continue;
2193                 }
2194                 lov_update_statfs(class_exp2obd(lov->lov_tgts[i]->ltd_exp),
2195                                   osfs, &lov_sfs, set);
2196                 set++;
2197         }
2198
2199         err = lov_fini_statfs(obd, osfs, set);
2200         qos_update(lov);
2201
2202         RETURN(rc ? rc : err);
2203 }
2204
2205 static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2206                          void *karg, void *uarg)
2207 {
2208         struct obd_device *obddev = class_exp2obd(exp);
2209         struct lov_obd *lov = &obddev->u.lov;
2210         int i, rc, count = lov->desc.ld_tgt_count;
2211         struct obd_uuid *uuidp;
2212         ENTRY;
2213
2214         switch (cmd) {
2215         case OBD_IOC_LOV_GET_CONFIG: {
2216                 struct obd_ioctl_data *data;
2217                 struct lov_desc *desc;
2218                 char *buf = NULL;
2219                 __u32 *genp;
2220
2221                 len = 0;
2222                 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2223                         RETURN(-EINVAL);
2224
2225                 data = (struct obd_ioctl_data *)buf;
2226
2227                 if (sizeof(*desc) > data->ioc_inllen1) {
2228                         obd_ioctl_freedata(buf, len);
2229                         RETURN(-EINVAL);
2230                 }
2231
2232                 if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) {
2233                         obd_ioctl_freedata(buf, len);
2234                         RETURN(-EINVAL);
2235                 }
2236
2237                 if (sizeof(__u32) * count > data->ioc_inllen3) {
2238                         obd_ioctl_freedata(buf, len);
2239                         RETURN(-EINVAL);
2240                 }
2241
2242                 desc = (struct lov_desc *)data->ioc_inlbuf1;
2243                 memcpy(desc, &(lov->desc), sizeof(*desc));
2244
2245                 uuidp = (struct obd_uuid *)data->ioc_inlbuf2;
2246                 genp = (__u32 *)data->ioc_inlbuf3;
2247                 /* the uuid will be empty for deleted OSTs */
2248                 for (i = 0; i < count; i++, uuidp++, genp++) {
2249                         if (!lov->lov_tgts[i])
2250                                 continue;
2251                         *uuidp = lov->lov_tgts[i]->ltd_uuid;
2252                         *genp = lov->lov_tgts[i]->ltd_gen;
2253                 }
2254
2255                 rc = copy_to_user((void *)uarg, buf, len);
2256                 if (rc)
2257                         rc = -EFAULT;
2258                 obd_ioctl_freedata(buf, len);
2259                 break;
2260         }
2261         case LL_IOC_LOV_SETSTRIPE:
2262                 rc = lov_setstripe(exp, karg, uarg);
2263                 break;
2264         case LL_IOC_LOV_GETSTRIPE:
2265                 rc = lov_getstripe(exp, karg, uarg);
2266                 break;
2267         case LL_IOC_LOV_SETEA:
2268                 rc = lov_setea(exp, karg, uarg);
2269                 break;
2270         default: {
2271                 int set = 0;
2272
2273                 if (count == 0)
2274                         RETURN(-ENOTTY);
2275
2276                 rc = 0;
2277                 for (i = 0; i < count; i++) {
2278                         int err;
2279
2280                         /* OST was disconnected */
2281                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
2282                                 continue;
2283
2284                         err = obd_iocontrol(cmd, lov->lov_tgts[i]->ltd_exp,
2285                                             len, karg, uarg);
2286                         if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) {
2287                                 RETURN(err);
2288                         } else if (err) {
2289                                 if (lov->lov_tgts[i]->ltd_active) {
2290                                         CDEBUG(err == -ENOTTY ?
2291                                                D_IOCTL : D_WARNING,
2292                                                "iocontrol OSC %s on OST "
2293                                                "idx %d cmd %x: err = %d\n",
2294                                                lov_uuid2str(lov, i),
2295                                                i, cmd, err);
2296                                         if (!rc)
2297                                                 rc = err;
2298                                 }
2299                         } else {
2300                                 set = 1;
2301                         }
2302                 }
2303                 if (!set && !rc)
2304                         rc = -EIO;
2305         }
2306         }
2307
2308         RETURN(rc);
2309 }
2310
2311 static int lov_get_info(struct obd_export *exp, __u32 keylen,
2312                         void *key, __u32 *vallen, void *val)
2313 {
2314         struct obd_device *obddev = class_exp2obd(exp);
2315         struct lov_obd *lov = &obddev->u.lov;
2316         int i, rc;
2317         ENTRY;
2318
2319         if (!vallen || !val)
2320                 RETURN(-EFAULT);
2321
2322         lov_getref(obddev);
2323
2324         if (KEY_IS("lock_to_stripe")) {
2325                 struct {
2326                         char name[16];
2327                         struct ldlm_lock *lock;
2328                         struct lov_stripe_md *lsm;
2329                 } *data = key;
2330                 struct ldlm_res_id *res_id = &data->lock->l_resource->lr_name;
2331                 struct lov_oinfo *loi;
2332                 __u32 *stripe = val;
2333
2334                 if (*vallen < sizeof(*stripe))
2335                         GOTO(out, rc = -EFAULT);
2336                 *vallen = sizeof(*stripe);
2337
2338                 /* XXX This is another one of those bits that will need to
2339                  * change if we ever actually support nested LOVs.  It uses
2340                  * the lock's export to find out which stripe it is. */
2341                 /* XXX - it's assumed all the locks for deleted OSTs have
2342                  * been cancelled. Also, the export for deleted OSTs will
2343                  * be NULL and won't match the lock's export. */
2344                 for (i = 0; i < data->lsm->lsm_stripe_count; i++) {
2345                         loi = data->lsm->lsm_oinfo[i];
2346                         if (!lov->lov_tgts[loi->loi_ost_idx])
2347                                 continue;
2348                         if (lov->lov_tgts[loi->loi_ost_idx]->ltd_exp ==
2349                             data->lock->l_conn_export &&
2350                             loi->loi_id == res_id->name[0] &&
2351                             loi->loi_gr == res_id->name[1]) {
2352                                 *stripe = i;
2353                                 GOTO(out, rc = 0);
2354                         }
2355                 }
2356                 LDLM_ERROR(data->lock, "lock on inode without such object");
2357                 dump_lsm(D_ERROR, data->lsm);
2358                 GOTO(out, rc = -ENXIO);
2359         } else if (KEY_IS("last_id")) {
2360                 obd_id *ids = val;
2361                 int size = sizeof(obd_id);
2362                 for (i = 0; i < lov->desc.ld_tgt_count; i++) {
2363                         if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active)
2364                                 continue;
2365                         rc = obd_get_info(lov->lov_tgts[i]->ltd_exp,
2366                                           keylen, key, &size, ids + i);
2367                         if (rc != 0)
2368                                 GOTO(out, rc);
2369                 }
2370                 GOTO(out, rc = 0);
2371         } else if (KEY_IS(KEY_LOVDESC)) {
2372                 struct lov_desc *desc_ret = val;
2373                 *desc_ret = lov->desc;
2374
2375                 GOTO(out, rc = 0);
2376         }
2377
2378         rc = -EINVAL;
2379 out:
2380         lov_putref(obddev);
2381         RETURN(rc);
2382 }
2383
2384 static int lov_set_info_async(struct obd_export *exp, obd_count keylen,
2385                               void *key, obd_count vallen, void *val,
2386                               struct ptlrpc_request_set *set)
2387 {
2388         struct obd_device *obddev = class_exp2obd(exp);
2389         struct lov_obd *lov = &obddev->u.lov;
2390         obd_count count;
2391         int i, rc = 0, err, incr = 0, check_uuid = 0, do_inactive = 0;
2392         int no_set = !set;
2393         ENTRY;
2394
2395         if (no_set) {
2396                 set = ptlrpc_prep_set();
2397                 if (!set)
2398                         RETURN(-ENOMEM);
2399         }
2400
2401         lov_getref(obddev);
2402         count = lov->desc.ld_tgt_count;
2403
2404         if (KEY_IS(KEY_NEXT_ID)) {
2405                 /* We must use mds's idea of # osts for indexing into
2406                    mds->mds_lov_objids */
2407                 count = vallen / sizeof(obd_id);
2408                 LASSERT(count <= lov->desc.ld_tgt_count);
2409                 vallen = sizeof(obd_id);
2410                 incr = sizeof(obd_id);
2411                 do_inactive = 1;
2412         } else if (KEY_IS("checksum")) {
2413                 do_inactive = 1;
2414         } else if (KEY_IS(KEY_MDS_CONN) || KEY_IS("unlinked")) {
2415                 check_uuid = val ? 1 : 0;
2416         } else if (KEY_IS("evict_by_nid")) {
2417                 /* use defaults:
2418                 do_inactive = incr = 0;
2419                  */
2420         }
2421
2422         for (i = 0; i < count; i++, val = (char *)val + incr) {
2423                 /* OST was disconnected */
2424                 if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp)
2425                         continue;
2426
2427                 /* OST is inactive and we don't want inactive OSCs */
2428                 if (!lov->lov_tgts[i]->ltd_active && !do_inactive)
2429                         continue;
2430
2431                 /* Only want a specific OSC */
2432                 if (check_uuid &&
2433                     !obd_uuid_equals(val, &lov->lov_tgts[i]->ltd_uuid))
2434                         continue;
2435
2436                 err = obd_set_info_async(lov->lov_tgts[i]->ltd_exp,
2437                                          keylen, key, vallen, val, set);
2438                 if (!rc)
2439                         rc = err;
2440         }
2441         lov_putref(obddev);
2442         if (no_set) {
2443                 err = ptlrpc_set_wait(set);
2444                 if (!rc)
2445                         rc = err;
2446                 ptlrpc_set_destroy(set);
2447         }
2448         RETURN(rc);
2449 }
2450
2451 static int lov_checkmd(struct obd_export *exp, struct obd_export *md_exp,
2452                        struct lov_stripe_md *lsm)
2453 {
2454         int rc;
2455         ENTRY;
2456
2457         if (!lsm)
2458                 RETURN(0);
2459         LASSERT(md_exp);
2460         LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
2461         rc = lsm_op_find(lsm->lsm_magic)->lsm_revalidate(lsm, md_exp->exp_obd);
2462
2463         RETURN(rc);
2464 }
2465
2466 int lov_test_and_clear_async_rc(struct lov_stripe_md *lsm)
2467 {
2468         struct lov_oinfo *loi;
2469         int i, rc = 0;
2470         ENTRY;
2471
2472         for (i = 0; i < lsm->lsm_stripe_count; i++) {
2473                 loi = lsm->lsm_oinfo[i];
2474                 if (loi->loi_ar.ar_rc && !rc)
2475                         rc = loi->loi_ar.ar_rc;
2476                 loi->loi_ar.ar_rc = 0;
2477         }
2478         RETURN(rc);
2479 }
2480 EXPORT_SYMBOL(lov_test_and_clear_async_rc);
2481
2482
2483 static int lov_extent_calc(struct obd_export *exp, struct lov_stripe_md *lsm,
2484                            int cmd, __u64 *offset)
2485 {
2486         __u64 start;
2487         __u32 ssize  = lsm->lsm_stripe_size;
2488
2489         start = *offset;
2490         do_div(start, ssize);
2491         start = start * ssize;
2492
2493         CDEBUG(D_DLMTRACE, "offset "LPU64", stripe %u, start "LPU64
2494                ", end "LPU64"\n", *offset, ssize, start, start + ssize - 1);
2495         if (cmd == OBD_CALC_STRIPE_END) {
2496                 *offset = start + ssize - 1;
2497         } else if (cmd == OBD_CALC_STRIPE_START) {
2498                 *offset = start;
2499         } else {
2500                 LBUG();
2501         }
2502
2503         RETURN(0);
2504 }
2505
2506 #if 0
2507 struct lov_multi_wait {
2508         struct ldlm_lock *lock;
2509         wait_queue_t      wait;
2510         int               completed;
2511         int               generation;
2512 };
2513
2514 int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm,
2515                       struct lustre_handle *lockh)
2516 {
2517         struct lov_lock_handles *lov_lockh = NULL;
2518         struct lustre_handle *lov_lockhp;
2519         struct lov_obd *lov;
2520         struct lov_oinfo *loi;
2521         struct lov_multi_wait *queues;
2522         int rc = 0, i;
2523         ENTRY;
2524
2525         ASSERT_LSM_MAGIC(lsm);
2526
2527         if (!exp || !exp->exp_obd)
2528                 RETURN(-ENODEV);
2529
2530         LASSERT(lockh != NULL);
2531         if (lsm->lsm_stripe_count > 1) {
2532                 lov_lockh = lov_handle2llh(lockh);
2533                 if (lov_lockh == NULL) {
2534                         CERROR("LOV: invalid lov lock handle %p\n", lockh);
2535                         RETURN(-EINVAL);
2536                 }
2537
2538                 lov_lockhp = lov_lockh->llh_handles;
2539         } else {
2540                 lov_lockhp = lockh;
2541         }
2542
2543         OBD_ALLOC(queues, lsm->lsm_stripe_count * sizeof(*queues));
2544         if (queues == NULL)
2545                 GOTO(out, rc = -ENOMEM);
2546
2547         lov = &exp->exp_obd->u.lov;
2548         for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count;
2549              i++, loi++, lov_lockhp++) {
2550                 struct ldlm_lock *lock;
2551                 struct obd_device *obd;
2552
2553                 lock = ldlm_handle2lock(lov_lockhp);
2554                 if (lock == NULL) {
2555                         CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n",
2556                                loi->loi_ost_idx, loi->loi_id);
2557                         queues[i].completed = 1;
2558                         continue;
2559                 }
2560
2561                 queues[i].lock = lock;
2562                 init_waitqueue_entry(&(queues[i].wait), current);
2563                 add_wait_queue(lock->l_waitq, &(queues[i].wait));
2564
2565                 obd = class_exp2obd(lock->l_conn_export);
2566                 if (obd != NULL)
2567                         imp = obd->u.cli.cl_import;
2568                 if (imp != NULL) {
2569                         spin_lock(&imp->imp_lock);
2570                         queues[i].generation = imp->imp_generation;
2571                         spin_unlock(&imp->imp_lock);
2572                 }
2573         }
2574
2575         lwi = LWI_TIMEOUT_INTR(obd_timeout * HZ, ldlm_expired_completion_wait,
2576                                interrupted_completion_wait, &lwd);
2577         rc = l_wait_event_added(check_multi_complete(queues, lsm), &lwi);
2578
2579         for (i = 0; i < lsm->lsm_stripe_count; i++)
2580                 remove_wait_queue(lock->l_waitq, &(queues[i].wait));
2581
2582         if (rc == -EINTR || rc == -ETIMEDOUT) {
2583
2584
2585         }
2586
2587  out:
2588         if (lov_lockh != NULL)
2589                 lov_llh_put(lov_lockh);
2590         RETURN(rc);
2591 }
2592 #endif
2593
2594 void lov_stripe_lock(struct lov_stripe_md *md)
2595 {
2596         LASSERT(md->lsm_lock_owner != cfs_current());
2597         spin_lock(&md->lsm_lock);
2598         LASSERT(md->lsm_lock_owner == NULL);
2599         md->lsm_lock_owner = cfs_current();
2600 }
2601 EXPORT_SYMBOL(lov_stripe_lock);
2602
2603 void lov_stripe_unlock(struct lov_stripe_md *md)
2604 {
2605         LASSERT(md->lsm_lock_owner == cfs_current());
2606         md->lsm_lock_owner = NULL;
2607         spin_unlock(&md->lsm_lock);
2608 }
2609 EXPORT_SYMBOL(lov_stripe_unlock);
2610
2611 struct obd_ops lov_obd_ops = {
2612         .o_owner               = THIS_MODULE,
2613         .o_setup               = lov_setup,
2614         .o_precleanup          = lov_precleanup,
2615         .o_cleanup             = lov_cleanup,
2616         .o_process_config      = lov_process_config,
2617         .o_connect             = lov_connect,
2618         .o_disconnect          = lov_disconnect,
2619         .o_statfs              = lov_statfs,
2620         .o_statfs_async        = lov_statfs_async,
2621         .o_packmd              = lov_packmd,
2622         .o_unpackmd            = lov_unpackmd,
2623         .o_checkmd             = lov_checkmd,
2624         .o_create              = lov_create,
2625         .o_destroy             = lov_destroy,
2626         .o_getattr             = lov_getattr,
2627         .o_getattr_async       = lov_getattr_async,
2628         .o_setattr             = lov_setattr,
2629         .o_setattr_async       = lov_setattr_async,
2630         .o_brw                 = lov_brw,
2631         .o_brw_async           = lov_brw_async,
2632         .o_prep_async_page     = lov_prep_async_page,
2633         .o_queue_async_io      = lov_queue_async_io,
2634         .o_set_async_flags     = lov_set_async_flags,
2635         .o_queue_group_io      = lov_queue_group_io,
2636         .o_trigger_group_io    = lov_trigger_group_io,
2637         .o_teardown_async_page = lov_teardown_async_page,
2638         .o_merge_lvb           = lov_merge_lvb,
2639         .o_adjust_kms          = lov_adjust_kms,
2640         .o_punch               = lov_punch,
2641         .o_sync                = lov_sync,
2642         .o_enqueue             = lov_enqueue,
2643         .o_match               = lov_match,
2644         .o_change_cbdata       = lov_change_cbdata,
2645         .o_cancel              = lov_cancel,
2646         .o_cancel_unused       = lov_cancel_unused,
2647         .o_join_lru            = lov_join_lru,
2648         .o_iocontrol           = lov_iocontrol,
2649         .o_get_info            = lov_get_info,
2650         .o_set_info_async      = lov_set_info_async,
2651         .o_extent_calc         = lov_extent_calc,
2652         .o_llog_init           = lov_llog_init,
2653         .o_llog_finish         = lov_llog_finish,
2654         .o_notify              = lov_notify,
2655 };
2656
2657 static quota_interface_t *quota_interface;
2658 extern quota_interface_t lov_quota_interface;
2659
2660 cfs_mem_cache_t *lov_oinfo_slab;
2661
2662 int __init lov_init(void)
2663 {
2664         struct lprocfs_static_vars lvars;
2665         int rc, rc2;
2666         ENTRY;
2667
2668         lov_oinfo_slab = cfs_mem_cache_create("lov_oinfo",
2669                                               sizeof(struct lov_oinfo),
2670                                               0, SLAB_HWCACHE_ALIGN);
2671         if (lov_oinfo_slab == NULL)
2672                 return -ENOMEM;
2673         lprocfs_init_vars(lov, &lvars);
2674
2675         request_module("lquota");
2676         quota_interface = PORTAL_SYMBOL_GET(lov_quota_interface);
2677         init_obd_quota_ops(quota_interface, &lov_obd_ops);
2678
2679         rc = class_register_type(&lov_obd_ops, lvars.module_vars,
2680                                  LUSTRE_LOV_NAME);
2681         if (rc) {
2682                 if (quota_interface)
2683                         PORTAL_SYMBOL_PUT(lov_quota_interface);
2684                 rc2 = cfs_mem_cache_destroy(lov_oinfo_slab);
2685                 LASSERT(rc2 == 0);
2686         }
2687
2688         RETURN(rc);
2689 }
2690
2691 #ifdef __KERNEL__
2692 static void /*__exit*/ lov_exit(void)
2693 {
2694         int rc;
2695
2696         if (quota_interface)
2697                 PORTAL_SYMBOL_PUT(lov_quota_interface);
2698
2699         class_unregister_type(LUSTRE_LOV_NAME);
2700         rc = cfs_mem_cache_destroy(lov_oinfo_slab);
2701         LASSERT(rc == 0);
2702 }
2703
2704 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2705 MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver");
2706 MODULE_LICENSE("GPL");
2707
2708 cfs_module(lov, LUSTRE_VERSION_STRING, lov_init, lov_exit);
2709 #endif