lustre/lmv/lmv_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, 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  */
  21
  22 #ifndef EXPORT_SYMTAB
  23 # define EXPORT_SYMTAB
  24 #endif
  25 #define DEBUG_SUBSYSTEM S_LMV
  26 #ifdef __KERNEL__
  27 #include <linux/slab.h>
  28 #include <linux/module.h>
  29 #include <linux/init.h>
  30 #include <linux/slab.h>
  31 #include <linux/pagemap.h>
  32 #include <asm/div64.h>
  33 #else
  34 #include <liblustre.h>
  35 #endif
  36 #include <linux/ext2_fs.h>
  37
  38 #include <linux/obd_support.h>
  39 #include <linux/lustre_lib.h>
  40 #include <linux/lustre_net.h>
  41 #include <linux/lustre_idl.h>
  42 #include <linux/lustre_dlm.h>
  43 #include <linux/lustre_mds.h>
  44 #include <linux/obd_class.h>
  45 #include <linux/obd_ost.h>
  46 #include <linux/seq_file.h>
  47 #include <linux/lprocfs_status.h>
  48 #include <linux/lustre_fsfilt.h>
  49 #include <linux/obd_lmv.h>
  50 #include "lmv_internal.h"
  51
  52 int lmv_attach(struct obd_device *dev, obd_count len, void *data)
  53 {
  54         struct lprocfs_static_vars lvars;
  55         struct proc_dir_entry *entry;
  56         int rc;
  57         ENTRY;
  58
  59         lprocfs_init_vars(lmv, &lvars);
  60         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
  61         if (rc)
  62                 RETURN (rc);
  63
  64         entry = create_proc_entry("target_obd", 0444, dev->obd_proc_entry);
  65         if (entry == NULL)
  66                 RETURN(-ENOMEM);
  67         /* entry->proc_fops = &lmv_proc_target_fops; */
  68         entry->data = dev;
  69
  70         RETURN (rc);
  71 }
  72
  73 int lmv_detach(struct obd_device *dev)
  74 {
  75         return lprocfs_obd_detach(dev);
  76 }
  77
  78 static int lmv_connect_fake(struct lustre_handle *conn,
  79                             struct obd_device *obd,
  80                             struct obd_uuid *cluuid)
  81 {
  82         struct lmv_obd *lmv = &obd->u.lmv;
  83         struct obd_export *exp;
  84         int rc;
  85         ENTRY;
  86
  87         rc = class_connect(conn, obd, cluuid);
  88         if (rc) {
  89                 CERROR("class_connection() returned %d\n", rc);
  90                 RETURN(rc);
  91         }
  92
  93         exp = class_conn2export(conn);
  94         /* We don't want to actually do the underlying connections more than
  95          * once, so keep track. */
  96         lmv->refcount++;
  97         if (lmv->refcount > 1) {
  98                 class_export_put(exp);
  99                 RETURN(0);
 100         }
 101
 102         lmv->cluuid = *cluuid;
 103         lmv->connected = 0;
 104         lmv->exp = exp;
 105
 106         RETURN(0);
 107 }
 108
 109 int lmv_connect(struct obd_device *obd)
 110 {
 111         struct lmv_obd *lmv = &obd->u.lmv;
 112         struct obd_uuid *cluuid;
 113         struct lmv_tgt_desc *tgts;
 114         struct obd_export *exp;
 115         int rc, i;
 116         ENTRY;
 117
 118         if (lmv->connected)
 119                 RETURN(0);
 120
 121         lmv->connected = 1;
 122         cluuid = &lmv->cluuid;
 123         exp = lmv->exp;
 124         CDEBUG(D_OTHER, "time to connect %s to %s\n",
 125                         cluuid->uuid, obd->obd_name);
 126
 127         for (i = 0, tgts = lmv->tgts; i < lmv->count; i++, tgts++) {
 128                 struct obd_device *tgt_obd;
 129                 struct obd_uuid lmv_osc_uuid = { "LMV_OSC_UUID" };
 130                 struct lustre_handle conn = {0, };
 131
 132                 LASSERT(tgts != NULL);
 133
 134                 tgt_obd = class_find_client_obd(&tgts->uuid, LUSTRE_MDC_NAME,
 135                                                 &obd->obd_uuid);
 136                 if (!tgt_obd) {
 137                         CERROR("Target %s not attached\n", tgts->uuid.uuid);
 138                         GOTO(out_disc, rc = -EINVAL);
 139                 }
 140
 141                 /* for MDS: don't connect to yourself */
 142                 if (obd_uuid_equals(&tgts->uuid, cluuid)) {
 143                         CDEBUG(D_OTHER, "don't connect back to %s\n",
 144                                cluuid->uuid);
 145                         tgts->exp = NULL;
 146                         continue;
 147                 }
 148
 149                 CDEBUG(D_OTHER, "connect to %s(%s) - %s, %s FOR %s\n",
 150                         tgt_obd->obd_name, tgt_obd->obd_uuid.uuid,
 151                         tgts->uuid.uuid, obd->obd_uuid.uuid,
 152                         cluuid->uuid);
 153
 154                 if (!tgt_obd->obd_set_up) {
 155                         CERROR("Target %s not set up\n", tgts->uuid.uuid);
 156                         GOTO(out_disc, rc = -EINVAL);
 157                 }
 158
 159                 rc = obd_connect(&conn, tgt_obd, &lmv_osc_uuid);
 160                 if (rc) {
 161                         CERROR("Target %s connect error %d\n",
 162                                 tgts->uuid.uuid, rc);
 163                         GOTO(out_disc, rc);
 164                 }
 165                 tgts->exp = class_conn2export(&conn);
 166
 167                 obd_init_ea_size(tgts->exp, lmv->max_easize,
 168                                         lmv->max_cookiesize);
 169
 170                 rc = obd_register_observer(tgt_obd, obd);
 171                 if (rc) {
 172                         CERROR("Target %s register_observer error %d\n",
 173                                tgts->uuid.uuid, rc);
 174                         obd_disconnect(tgts->exp, 0);
 175                         GOTO(out_disc, rc);
 176                 }
 177
 178                 CDEBUG(D_OTHER, "connected to %s(%s) successfully (%d)\n",
 179                         tgt_obd->obd_name, tgt_obd->obd_uuid.uuid,
 180                         atomic_read(&obd->obd_refcount));
 181         }
 182
 183         class_export_put(exp);
 184         RETURN (0);
 185
 186  out_disc:
 187         /* FIXME: cleanup here */
 188         class_disconnect(exp, 0);
 189         RETURN (rc);
 190 }
 191
 192 static int lmv_disconnect(struct obd_export *exp, int flags)
 193 {
 194         struct obd_device *obd = class_exp2obd(exp);
 195         struct lmv_obd *lmv = &obd->u.lmv;
 196         int rc, i;
 197         ENTRY;
 198
 199         if (!lmv->tgts)
 200                 goto out_local;
 201
 202         /* Only disconnect the underlying layers on the final disconnect. */
 203         lmv->refcount--;
 204         if (lmv->refcount != 0)
 205                 goto out_local;
 206
 207         for (i = 0; i < lmv->count; i++) {
 208                 if (lmv->tgts[i].exp == NULL)
 209                         continue;
 210
 211                 if (obd->obd_no_recov) {
 212                         /* Pass it on to our clients.
 213                          * XXX This should be an argument to disconnect,
 214                          * XXX not a back-door flag on the OBD.  Ah well.
 215                          */
 216                         struct obd_device *mdc_obd;
 217                         mdc_obd = class_exp2obd(lmv->tgts[i].exp);
 218                         if (mdc_obd)
 219                                 mdc_obd->obd_no_recov = 1;
 220                 }
 221
 222                 CDEBUG(D_OTHER, "disconnected from %s(%s) successfully\n",
 223                         lmv->tgts[i].exp->exp_obd->obd_name,
 224                         lmv->tgts[i].exp->exp_obd->obd_uuid.uuid);
 225
 226                 obd_register_observer(lmv->tgts[i].exp->exp_obd, NULL);
 227
 228                 rc = obd_disconnect(lmv->tgts[i].exp, flags);
 229                 lmv->tgts[i].exp = NULL;
 230         }
 231
 232  out_local:
 233         /* FIXME: cleanup here */
 234         if (!lmv->connected)
 235                 class_export_put(exp);
 236         rc = class_disconnect(exp, 0);
 237         RETURN(rc);
 238 }
 239
 240 static int lmv_setup(struct obd_device *obd, obd_count len, void *buf)
 241 {
 242         struct lustre_cfg *lcfg = buf;
 243         struct lmv_desc *desc;
 244         struct lmv_obd *lmv = &obd->u.lmv;
 245         struct obd_uuid *uuids;
 246         struct lmv_tgt_desc *tgts;
 247         int i;
 248         int count;
 249         int rc = 0;
 250         ENTRY;
 251
 252         if (lcfg->lcfg_inllen1 < 1) {
 253                 CERROR("LMV setup requires a descriptor\n");
 254                 RETURN(-EINVAL);
 255         }
 256
 257         if (lcfg->lcfg_inllen2 < 1) {
 258                 CERROR("LMV setup requires an OST UUID list\n");
 259                 RETURN(-EINVAL);
 260         }
 261
 262         desc = (struct lmv_desc *)lcfg->lcfg_inlbuf1;
 263         if (sizeof(*desc) > lcfg->lcfg_inllen1) {
 264                 CERROR("descriptor size wrong: %d > %d\n",
 265                        (int)sizeof(*desc), lcfg->lcfg_inllen1);
 266                 RETURN(-EINVAL);
 267         }
 268
 269         count = desc->ld_count;
 270         uuids = (struct obd_uuid *)lcfg->lcfg_inlbuf2;
 271         if (sizeof(*uuids) * count != lcfg->lcfg_inllen2) {
 272                 CERROR("UUID array size wrong: %u * %u != %u\n",
 273                        sizeof(*uuids), count, lcfg->lcfg_inllen2);
 274                 RETURN(-EINVAL);
 275         }
 276
 277         lmv->bufsize = sizeof(struct lmv_tgt_desc) * count;
 278         OBD_ALLOC(lmv->tgts, lmv->bufsize);
 279         if (lmv->tgts == NULL) {
 280                 CERROR("Out of memory\n");
 281                 RETURN(-EINVAL);
 282         }
 283
 284         for (i = 0, tgts = lmv->tgts; i < count; i++, tgts++) {
 285                 tgts->uuid = uuids[i];
 286                 lmv->count++;
 287         }
 288
 289         lmv->max_easize = sizeof(struct ll_fid) * lmv->count
 290                                         + sizeof(struct mea);
 291         lmv->max_cookiesize = 0;
 292
 293         RETURN(rc);
 294 }
 295
 296 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
 297                       unsigned long max_age)
 298 {
 299         struct lmv_obd *lmv = &obd->u.lmv;
 300         struct obd_statfs temp;
 301         int rc = 0, i;
 302         ENTRY;
 303         lmv_connect(obd);
 304         for (i = 0; i < lmv->count; i++) {
 305                 rc = obd_statfs(lmv->tgts[i].exp->exp_obd, &temp, max_age);
 306                 if (rc) {
 307                         CERROR("can't stat MDS #%d (%s)\n", i,
 308                                lmv->tgts[i].exp->exp_obd->obd_name);
 309                         RETURN(rc);
 310                 }
 311                 if (i == 0) {
 312                         memcpy(osfs, &temp, sizeof(temp));
 313                 } else {
 314                         osfs->os_bavail += temp.os_bavail;
 315                         osfs->os_blocks += temp.os_blocks;
 316                         osfs->os_ffree += temp.os_ffree;
 317                         osfs->os_files += temp.os_files;
 318                 }
 319         }
 320         RETURN(rc);
 321 }
 322
 323 static int lmv_cleanup(struct obd_device *obd, int flags)
 324 {
 325         struct lmv_obd *lmv = &obd->u.lmv;
 326         ENTRY;
 327         lmv_cleanup_objs(obd);
 328         OBD_FREE(lmv->tgts, lmv->bufsize);
 329         RETURN(0);
 330 }
 331
 332 static int lmv_getstatus(struct obd_export *exp, struct ll_fid *fid)
 333 {
 334         struct obd_device *obd = exp->exp_obd;
 335         struct lmv_obd *lmv = &obd->u.lmv;
 336         int rc;
 337         ENTRY;
 338         lmv_connect(obd);
 339         rc = md_getstatus(lmv->tgts[0].exp, fid);
 340         fid->mds = 0;
 341         RETURN(rc);
 342 }
 343
 344 static int lmv_getattr(struct obd_export *exp, struct ll_fid *fid,
 345                 unsigned long valid, unsigned int ea_size,
 346                 struct ptlrpc_request **request)
 347 {
 348         struct obd_device *obd = exp->exp_obd;
 349         struct lmv_obd *lmv = &obd->u.lmv;
 350         int rc, i = fid->mds;
 351         struct lmv_obj *obj;
 352         ENTRY;
 353         lmv_connect(obd);
 354         obj = lmv_grab_obj(obd, fid, 0);
 355         CDEBUG(D_OTHER, "GETATTR for %lu/%lu/%lu %s\n",
 356                (unsigned long) fid->mds,
 357                (unsigned long) fid->id,
 358                (unsigned long) fid->generation,
 359                obj ? "(splitted)" : "");
 360
 361         LASSERT(fid->mds < lmv->count);
 362         rc = md_getattr(lmv->tgts[i].exp, fid,
 363                              valid, ea_size, request);
 364         if (rc == 0 && obj) {
 365                 /* we have to loop over dirobjs here and gather attrs
 366                  * for all the slaves */
 367 #warning "attrs gathering here"
 368         }
 369         lmv_put_obj(obj);
 370         RETURN(rc);
 371 }
 372
 373 static int lmv_change_cbdata(struct obd_export *exp,
 374                                  struct ll_fid *fid,
 375                                  ldlm_iterator_t it, void *data)
 376 {
 377         struct obd_device *obd = exp->exp_obd;
 378         struct lmv_obd *lmv = &obd->u.lmv;
 379         int rc = 0;
 380         ENTRY;
 381         lmv_connect(obd);
 382         CDEBUG(D_OTHER, "CBDATA for %lu/%lu/%lu\n",
 383                (unsigned long) fid->mds,
 384                (unsigned long) fid->id,
 385                (unsigned long) fid->generation);
 386         LASSERT(fid->mds < lmv->count);
 387         rc = md_change_cbdata(lmv->tgts[fid->mds].exp, fid, it, data);
 388         RETURN(rc);
 389 }
 390
 391 static int lmv_change_cbdata_name(struct obd_export *exp, struct ll_fid *pfid,
 392                                   char *name, int len, struct ll_fid *cfid,
 393                                   ldlm_iterator_t it, void *data)
 394 {
 395         struct obd_device *obd = exp->exp_obd;
 396         struct lmv_obd *lmv = &obd->u.lmv;
 397         struct lmv_obj *obj;
 398         int rc = 0, mds;
 399         ENTRY;
 400         lmv_connect(obd);
 401         LASSERT(pfid->mds < lmv->count);
 402         LASSERT(cfid->mds < lmv->count);
 403         CDEBUG(D_OTHER, "CBDATA for %lu/%lu/%lu:%*s -> %lu/%lu/%lu\n",
 404                (unsigned long) pfid->mds, (unsigned long) pfid->id,
 405                (unsigned long) pfid->generation, len, name,
 406                (unsigned long) cfid->mds, (unsigned long) cfid->id,
 407                (unsigned long) cfid->generation);
 408
 409         /* this is default mds for directory name belongs to */
 410         mds = pfid->mds;
 411         obj = lmv_grab_obj(obd, pfid, 0);
 412         if (obj) {
 413                 /* directory is splitted. look for right mds for this name */
 414                 mds = raw_name2idx(obj->objcount, name, len);
 415                 lmv_put_obj(obj);
 416         }
 417         rc = md_change_cbdata(lmv->tgts[mds].exp, cfid, it, data);
 418         RETURN(rc);
 419 }
 420
 421 static int lmv_valid_attrs(struct obd_export *exp, struct ll_fid *fid)
 422 {
 423         struct obd_device *obd = exp->exp_obd;
 424         struct lmv_obd *lmv = &obd->u.lmv;
 425         int rc = 0;
 426         ENTRY;
 427         lmv_connect(obd);
 428         CDEBUG(D_OTHER, "validate %lu/%lu/%lu\n",
 429                (unsigned long) fid->mds,
 430                (unsigned long) fid->id,
 431                (unsigned long) fid->generation);
 432         LASSERT(fid->mds < lmv->count);
 433         rc = md_valid_attrs(lmv->tgts[fid->mds].exp, fid);
 434         RETURN(rc);
 435 }
 436
 437 int lmv_close(struct obd_export *exp, struct obdo *obdo,
 438                   struct obd_client_handle *och,
 439                   struct ptlrpc_request **request)
 440 {
 441         struct obd_device *obd = exp->exp_obd;
 442         struct lmv_obd *lmv = &obd->u.lmv;
 443         int rc, i = obdo->o_mds;
 444         ENTRY;
 445         lmv_connect(obd);
 446         LASSERT(i < lmv->count);
 447         CDEBUG(D_OTHER, "CLOSE %lu/%lu/%lu\n", (unsigned long) obdo->o_mds,
 448                (unsigned long) obdo->o_id, (unsigned long) obdo->o_generation);
 449         rc = md_close(lmv->tgts[i].exp, obdo, och, request);
 450         RETURN(rc);
 451 }
 452
 453 int lmv_get_mea_and_update_object(struct obd_export *exp, struct ll_fid *fid)
 454 {
 455         struct obd_device *obd = exp->exp_obd;
 456         struct lmv_obd *lmv = &obd->u.lmv;
 457         struct ptlrpc_request *req = NULL;
 458         struct lustre_md md;
 459         int mealen, rc;
 460
 461         md.mea = NULL;
 462         mealen = MEA_SIZE_LMV(lmv);
 463
 464         /* time to update mea of parent fid */
 465         rc = md_getattr(lmv->tgts[fid->mds].exp, fid,
 466                         OBD_MD_FLEASIZE, mealen, &req);
 467         if (rc)
 468                 GOTO(cleanup, rc);
 469         rc = mdc_req2lustre_md(req, 0, NULL, exp, &md);
 470         if (rc)
 471                 GOTO(cleanup, rc);
 472         if (md.mea == NULL)
 473                 GOTO(cleanup, rc = -ENODATA);
 474         rc = lmv_create_obj_from_attrs(exp, fid, md.mea);
 475         obd_free_memmd(exp, (struct lov_stripe_md **) &md.mea);
 476
 477 cleanup:
 478         if (req)
 479                 ptlrpc_req_finished(req);
 480         RETURN(rc);
 481 }
 482
 483 int lmv_create(struct obd_export *exp, struct mdc_op_data *op_data,
 484                    const void *data, int datalen, int mode, __u32 uid,
 485                    __u32 gid, __u64 rdev, struct ptlrpc_request **request)
 486 {
 487         struct obd_device *obd = exp->exp_obd;
 488         struct lmv_obd *lmv = &obd->u.lmv;
 489         struct mds_body *mds_body;
 490         struct lmv_obj *obj;
 491         int rc, mds;
 492         ENTRY;
 493
 494         lmv_connect(obd);
 495 repeat:
 496         obj = lmv_grab_obj(obd, &op_data->fid1, 0);
 497         if (obj) {
 498                 mds = raw_name2idx(obj->objcount, op_data->name,
 499                                         op_data->namelen);
 500                 op_data->fid1 = obj->objs[mds].fid;
 501                 lmv_put_obj(obj);
 502         }
 503
 504         CDEBUG(D_OTHER, "CREATE '%*s' on %lu/%lu/%lu\n",
 505                         op_data->namelen, op_data->name,
 506                         (unsigned long) op_data->fid1.mds,
 507                         (unsigned long) op_data->fid1.id,
 508                         (unsigned long) op_data->fid1.generation);
 509         rc = md_create(lmv->tgts[op_data->fid1.mds].exp, op_data, data,
 510                        datalen, mode, uid, gid, rdev, request);
 511         if (rc == 0) {
 512                 if (*request == NULL)
 513                      RETURN(rc);
 514                 mds_body = lustre_msg_buf((*request)->rq_repmsg, 0,
 515                                           sizeof(*mds_body));
 516                 LASSERT(mds_body != NULL);
 517                 CDEBUG(D_OTHER, "created. id = %lu, generation = %lu, mds = %d\n",
 518                        (unsigned long) mds_body->fid1.id,
 519                        (unsigned long) mds_body->fid1.generation,
 520                        op_data->fid1.mds);
 521                 LASSERT(mds_body->valid & OBD_MD_MDS ||
 522                                 mds_body->mds == op_data->fid1.mds);
 523         } else if (rc == -ERESTART) {
 524                 /* directory got splitted. time to update local object
 525                  * and repeat the request with proper MDS */
 526                 rc = lmv_get_mea_and_update_object(exp, &op_data->fid1);
 527                 if (rc == 0) {
 528                         ptlrpc_req_finished(*request);
 529                         goto repeat;
 530                 }
 531         }
 532         RETURN(rc);
 533 }
 534
 535 int lmv_done_writing(struct obd_export *exp, struct obdo *obdo)
 536 {
 537         struct obd_device *obd = exp->exp_obd;
 538         struct lmv_obd *lmv = &obd->u.lmv;
 539         int rc;
 540         ENTRY;
 541         lmv_connect(obd);
 542         /* FIXME: choose right MDC here */
 543         rc = md_done_writing(lmv->tgts[0].exp, obdo);
 544         RETURN(rc);
 545 }
 546
 547 int lmv_enqueue(struct obd_export *exp, int lock_type,
 548                     struct lookup_intent *it, int lock_mode,
 549                     struct mdc_op_data *data, struct lustre_handle *lockh,
 550                     void *lmm, int lmmsize,
 551                     ldlm_completion_callback cb_completion,
 552                     ldlm_blocking_callback cb_blocking, void *cb_data)
 553 {
 554         struct obd_device *obd = exp->exp_obd;
 555         struct lmv_obd *lmv = &obd->u.lmv;
 556         struct lmv_obj *obj;
 557         int rc, mds;
 558         ENTRY;
 559         lmv_connect(obd);
 560         if (data->namelen) {
 561                 obj = lmv_grab_obj(obd, &data->fid1, 0);
 562                 if (obj) {
 563                         /* directory is splitted. look for
 564                          * right mds for this name */
 565                         mds = raw_name2idx(obj->objcount, data->name,
 566                                                 data->namelen);
 567                         data->fid1 = obj->objs[mds].fid;
 568                         lmv_put_obj(obj);
 569                 }
 570         }
 571         CDEBUG(D_OTHER, "ENQUEUE '%s' on %lu/%lu\n",
 572                LL_IT2STR(it), (unsigned long) data->fid1.id,
 573                (unsigned long) data->fid1.generation);
 574         rc = md_enqueue(lmv->tgts[data->fid1.mds].exp, lock_type, it,
 575                         lock_mode, data, lockh, lmm, lmmsize, cb_completion,
 576                         cb_blocking, cb_data);
 577
 578         RETURN(rc);
 579 }
 580
 581 int lmv_getattr_name(struct obd_export *exp, struct ll_fid *fid,
 582                          char *filename, int namelen, unsigned long valid,
 583                          unsigned int ea_size, struct ptlrpc_request **request)
 584 {
 585         struct obd_device *obd = exp->exp_obd;
 586         struct lmv_obd *lmv = &obd->u.lmv;
 587         struct ll_fid rfid = *fid;
 588         int rc, mds = fid->mds;
 589         struct mds_body *body;
 590         struct lmv_obj *obj;
 591         ENTRY;
 592         lmv_connect(obd);
 593         obj = lmv_grab_obj(obd, fid, 0);
 594         if (obj) {
 595                 /* directory is splitted. look for right mds for this name */
 596                 mds = raw_name2idx(obj->objcount, filename, namelen - 1);
 597                 rfid = obj->objs[mds].fid;
 598                 lmv_put_obj(obj);
 599         }
 600         CDEBUG(D_OTHER, "getattr_name for %*s on %lu/%lu/%lu -> %lu/%lu/%lu\n",
 601                namelen, filename, (unsigned long) fid->mds,
 602                (unsigned long) fid->id, (unsigned long) fid->generation,
 603                (unsigned long) rfid.mds, (unsigned long) rfid.id,
 604                (unsigned long) rfid.generation);
 605         rc = md_getattr_name(lmv->tgts[mds].exp, &rfid, filename, namelen,
 606                                   valid, ea_size, request);
 607         if (rc == 0) {
 608                 /* this could be cross-node reference. in this case all
 609                  * we have right now is mds/ino/generation triple. we'd
 610                  * like to find other attributes */
 611                 body = lustre_msg_buf((*request)->rq_repmsg, 0, sizeof(*body));
 612                 LASSERT(body != NULL);
 613                 if (body->valid & OBD_MD_MDS) {
 614                         struct ptlrpc_request *req = NULL;
 615                         rfid = body->fid1;
 616                         CDEBUG(D_OTHER, "request attrs for %lu/%lu/%lu\n",
 617                                (unsigned long) rfid.mds,
 618                                (unsigned long) rfid.id,
 619                                (unsigned long) rfid.generation);
 620                         rc = md_getattr_name(lmv->tgts[rfid.mds].exp, &rfid,
 621                                              NULL, 1, valid, ea_size, &req);
 622                         ptlrpc_req_finished(*request);
 623                         *request = req;
 624                 }
 625         }
 626         RETURN(rc);
 627 }
 628
 629
 630 /*
 631  * llite passes fid of an target inode in data->fid1 and
 632  * fid of directory in data->fid2
 633  */
 634 int lmv_link(struct obd_export *exp, struct mdc_op_data *data,
 635              struct ptlrpc_request **request)
 636 {
 637         struct obd_device *obd = exp->exp_obd;
 638         struct lmv_obd *lmv = &obd->u.lmv;
 639         struct lmv_obj *obj;
 640         int rc;
 641         ENTRY;
 642         lmv_connect(obd);
 643         if (data->namelen != 0) {
 644                 /* usual link request */
 645                 obj = lmv_grab_obj(obd, &data->fid1, 0);
 646                 if (obj) {
 647                         rc = raw_name2idx(obj->objcount, data->name,
 648                                          data->namelen);
 649                         data->fid1 = obj->objs[rc].fid;
 650                         lmv_put_obj(obj);
 651                 }
 652                 CDEBUG(D_OTHER,"link %u/%u/%u:%*s to %u/%u/%u mds %d\n",
 653                        (unsigned) data->fid2.mds, (unsigned) data->fid2.id,
 654                        (unsigned) data->fid2.generation, data->namelen,
 655                        data->name, (unsigned) data->fid1.mds,
 656                        (unsigned) data->fid1.id,
 657                        (unsigned) data->fid1.generation, data->fid1.mds);
 658         } else {
 659                 /* request from MDS to acquire i_links for inode by fid1 */
 660                 CDEBUG(D_OTHER, "inc i_nlinks for %u/%u/%u\n",
 661                        (unsigned) data->fid1.mds, (unsigned) data->fid1.id,
 662                        (unsigned) data->fid1.generation);
 663         }
 664
 665         rc = md_link(lmv->tgts[data->fid1.mds].exp, data, request);
 666         RETURN(rc);
 667 }
 668
 669 int lmv_rename(struct obd_export *exp, struct mdc_op_data *data,
 670                const char *old, int oldlen, const char *new, int newlen,
 671                struct ptlrpc_request **request)
 672 {
 673         struct obd_device *obd = exp->exp_obd;
 674         struct lmv_obd *lmv = &obd->u.lmv;
 675         struct lmv_obj *obj;
 676         int rc, mds;
 677         ENTRY;
 678
 679         CDEBUG(D_OTHER, "rename %*s in %lu/%lu/%lu to %*s in %lu/%lu/%lu\n",
 680                oldlen, old, (unsigned long) data->fid1.mds,
 681                (unsigned long) data->fid1.id,
 682                (unsigned long) data->fid1.generation,
 683                newlen, new, (unsigned long) data->fid2.mds,
 684                (unsigned long) data->fid2.id,
 685                (unsigned long) data->fid2.generation);
 686         if (!fid_equal(&data->fid1, &data->fid2))
 687                 CWARN("cross-node rename %lu/%lu/%lu:%*s to %lu/%lu/%lu:%*s\n",
 688                       (unsigned long) data->fid1.mds,
 689                       (unsigned long) data->fid1.id,
 690                       (unsigned long) data->fid1.generation, oldlen, old,
 691                       (unsigned long) data->fid2.mds,
 692                       (unsigned long) data->fid2.id,
 693                       (unsigned long) data->fid2.generation, newlen, new);
 694
 695         lmv_connect(obd);
 696
 697         if (oldlen == 0) {
 698                 /* MDS with old dir entry is asking another MDS
 699                  * to create name there */
 700                 CDEBUG(D_OTHER,
 701                        "create %*s(%d/%d) in %lu/%lu/%lu pointing to %lu/%lu/%lu\n",
 702                        newlen, new, oldlen, newlen,
 703                        (unsigned long) data->fid2.mds,
 704                        (unsigned long) data->fid2.id,
 705                        (unsigned long) data->fid2.generation,
 706                        (unsigned long) data->fid1.mds,
 707                        (unsigned long) data->fid1.id,
 708                        (unsigned long) data->fid1.generation);
 709                 mds = data->fid2.mds;
 710                 goto request;
 711         }
 712
 713         obj = lmv_grab_obj(obd, &data->fid1, 0);
 714         if (obj) {
 715                 /* directory is already splitted, so we have to forward
 716                  * request to the right MDS */
 717                 mds = raw_name2idx(obj->objcount, old, oldlen);
 718                 data->fid1 = obj->objs[mds].fid;
 719                 CDEBUG(D_OTHER, "forward to MDS #%u (%lu/%lu/%lu)\n", mds,
 720                        (unsigned long) obj->objs[mds].fid.mds,
 721                        (unsigned long) obj->objs[mds].fid.id,
 722                        (unsigned long) obj->objs[mds].fid.generation);
 723         }
 724         lmv_put_obj(obj);
 725
 726         obj = lmv_grab_obj(obd, &data->fid2, 0);
 727         if (obj) {
 728                 /* directory is already splitted, so we have to forward
 729                  * request to the right MDS */
 730                 mds = raw_name2idx(obj->objcount, new, newlen);
 731                 data->fid2 = obj->objs[mds].fid;
 732                 CDEBUG(D_OTHER, "forward to MDS #%u (%lu/%lu/%lu)\n", mds,
 733                        (unsigned long) obj->objs[mds].fid.mds,
 734                        (unsigned long) obj->objs[mds].fid.id,
 735                        (unsigned long) obj->objs[mds].fid.generation);
 736         }
 737         lmv_put_obj(obj);
 738
 739         mds = data->fid1.mds;
 740
 741 request:
 742         rc = md_rename(lmv->tgts[mds].exp, data, old, oldlen,
 743                             new, newlen, request);
 744         RETURN(rc);
 745 }
 746
 747 int lmv_setattr(struct obd_export *exp, struct mdc_op_data *data,
 748                 struct iattr *iattr, void *ea, int ealen, void *ea2, int ea2len,
 749                 struct ptlrpc_request **request)
 750 {
 751         struct obd_device *obd = exp->exp_obd;
 752         struct lmv_obd *lmv = &obd->u.lmv;
 753         int rc = 0, i = data->fid1.mds;
 754         struct ptlrpc_request *req;
 755         struct mds_body *mds_body;
 756         struct lmv_obj *obj;
 757         ENTRY;
 758         lmv_connect(obd);
 759         obj = lmv_grab_obj(obd, &data->fid1, 0);
 760         CDEBUG(D_OTHER, "SETATTR for %lu/%lu/%lu, valid 0x%x%s\n",
 761                (unsigned long) data->fid1.mds,
 762                (unsigned long) data->fid1.id,
 763                (unsigned long) data->fid1.generation, iattr->ia_valid,
 764                obj ? ", splitted" : "");
 765         if (obj) {
 766                 for (i = 0; i < obj->objcount; i++) {
 767                         data->fid1 = obj->objs[i].fid;
 768                         rc = md_setattr(lmv->tgts[i].exp, data, iattr, ea,
 769                                         ealen, ea2, ea2len, &req);
 770                         LASSERT(rc == 0);
 771                         if (fid_equal(&obj->fid, &obj->objs[i].fid)) {
 772                                 /* this is master object and this request
 773                                  * should be returned back to llite */
 774                                 *request = req;
 775                         } else {
 776                                 ptlrpc_req_finished(req);
 777                         }
 778                 }
 779                 lmv_put_obj(obj);
 780         } else {
 781                 LASSERT(data->fid1.mds < lmv->count);
 782                 rc = md_setattr(lmv->tgts[i].exp, data, iattr, ea, ealen,
 783                                 ea2, ea2len, request);
 784                 if (rc == 0) {
 785                         mds_body = lustre_msg_buf((*request)->rq_repmsg, 0,
 786                                         sizeof(*mds_body));
 787                         LASSERT(mds_body != NULL);
 788                         LASSERT(mds_body->mds == i);
 789                 }
 790         }
 791         RETURN(rc);
 792 }
 793
 794 int lmv_sync(struct obd_export *exp, struct ll_fid *fid,
 795              struct ptlrpc_request **request)
 796 {
 797         struct obd_device *obd = exp->exp_obd;
 798         struct lmv_obd *lmv = &obd->u.lmv;
 799         int rc;
 800         ENTRY;
 801         lmv_connect(obd);
 802         rc = md_sync(lmv->tgts[0].exp, fid, request);
 803         RETURN(rc);
 804 }
 805
 806 int lmv_dirobj_blocking_ast(struct ldlm_lock *lock,
 807                             struct ldlm_lock_desc *desc, void *data, int flag)
 808 {
 809         struct lustre_handle lockh;
 810         struct lmv_obj *obj;
 811         int rc;
 812         ENTRY;
 813
 814         switch (flag) {
 815         case LDLM_CB_BLOCKING:
 816                 ldlm_lock2handle(lock, &lockh);
 817                 rc = ldlm_cli_cancel(&lockh);
 818                 if (rc < 0) {
 819                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
 820                         RETURN(rc);
 821                 }
 822                 break;
 823         case LDLM_CB_CANCELING:
 824                 /* time to drop cached attrs for dirobj */
 825                 obj = lock->l_ast_data;
 826                 if (!obj)
 827                         break;
 828
 829                 CDEBUG(D_OTHER, "cancel %s on %lu/%lu, master %lu/%lu/%lu\n",
 830                        lock->l_resource->lr_name.name[3] == 1 ?
 831                                 "LOOKUP" : "UPDATE",
 832                        (unsigned long) lock->l_resource->lr_name.name[0],
 833                        (unsigned long) lock->l_resource->lr_name.name[1],
 834                        (unsigned long) obj->fid.mds,
 835                        (unsigned long) obj->fid.id,
 836                        (unsigned long) obj->fid.generation);
 837                 break;
 838         default:
 839                 LBUG();
 840         }
 841         RETURN(0);
 842 }
 843
 844 void lmv_remove_dots(struct page *page)
 845 {
 846         char *kaddr = page_address(page);
 847         unsigned limit = PAGE_CACHE_SIZE;
 848         unsigned offs, rec_len;
 849         struct ext2_dir_entry_2 *p;
 850
 851         for (offs = 0; offs <= limit - EXT2_DIR_REC_LEN(1); offs += rec_len) {
 852                 p = (struct ext2_dir_entry_2 *)(kaddr + offs);
 853                 rec_len = le16_to_cpu(p->rec_len);
 854
 855                 if ((p->name_len == 1 && p->name[0] == '.') ||
 856                     (p->name_len == 2 && p->name[0] == '.' && p->name[1] == '.'))
 857                         p->inode = 0;
 858         }
 859 }
 860
 861 int lmv_readpage(struct obd_export *exp, struct ll_fid *mdc_fid,
 862                  __u64 offset, struct page *page,
 863                  struct ptlrpc_request **request)
 864 {
 865         struct obd_device *obd = exp->exp_obd;
 866         struct lmv_obd *lmv = &obd->u.lmv;
 867         struct ll_fid rfid = *mdc_fid;
 868         struct lmv_obj *obj;
 869         int rc, i;
 870         ENTRY;
 871         lmv_connect(obd);
 872
 873         LASSERT(mdc_fid->mds < lmv->count);
 874         CDEBUG(D_OTHER, "READPAGE at %llu from %lu/%lu/%lu\n",
 875                offset, (unsigned long) rfid.mds,
 876                (unsigned long) rfid.id,
 877                (unsigned long) rfid.generation);
 878
 879         obj = lmv_grab_obj(obd, mdc_fid, 0);
 880         if (obj) {
 881                 /* find dirobj containing page with requested offset */
 882                 /* FIXME: what about protecting cached attrs here? */
 883                 for (i = 0; i < obj->objcount; i++) {
 884                         if (offset < obj->objs[i].size)
 885                                 break;
 886                         offset -= obj->objs[i].size;
 887                 }
 888                 rfid = obj->objs[i].fid;
 889                 CDEBUG(D_OTHER, "forward to %lu/%lu/%lu with offset %lu\n",
 890                        (unsigned long) rfid.mds,
 891                        (unsigned long) rfid.id,
 892                        (unsigned long) rfid.generation,
 893                        (unsigned long) offset);
 894         }
 895         rc = md_readpage(lmv->tgts[rfid.mds].exp, &rfid, offset, page, request);
 896         if (rc == 0 && !fid_equal(&rfid, mdc_fid)) {
 897                 /* this page isn't from master object. to avoid
 898                  * ./.. duplication in directory, we have to remove them
 899                  * from all slave objects */
 900                 lmv_remove_dots(page);
 901         }
 902
 903         lmv_put_obj(obj);
 904
 905         RETURN(rc);
 906 }
 907
 908 int lmv_unlink(struct obd_export *exp, struct mdc_op_data *data,
 909                struct ptlrpc_request **request)
 910 {
 911         struct obd_device *obd = exp->exp_obd;
 912         struct lmv_obd *lmv = &obd->u.lmv;
 913         int rc, i = 0;
 914         ENTRY;
 915         lmv_connect(obd);
 916         if (data->namelen != 0) {
 917                 struct lmv_obj *obj;
 918                 obj = lmv_grab_obj(obd, &data->fid1, 0);
 919                 if (obj) {
 920                         i = raw_name2idx(obj->objcount, data->name,
 921                                          data->namelen);
 922                         data->fid1 = obj->objs[i].fid;
 923                         lmv_put_obj(obj);
 924                 }
 925                 CDEBUG(D_OTHER, "unlink '%*s' in %lu/%lu/%lu -> %u\n",
 926                        data->namelen, data->name,
 927                        (unsigned long) data->fid1.mds,
 928                        (unsigned long) data->fid1.id,
 929                        (unsigned long) data->fid1.generation, i);
 930         } else {
 931                 CDEBUG(D_OTHER, "drop i_nlink on %lu/%lu/%lu\n",
 932                        (unsigned long) data->fid1.mds,
 933                        (unsigned long) data->fid1.id,
 934                        (unsigned long) data->fid1.generation);
 935         }
 936         rc = md_unlink(lmv->tgts[data->fid1.mds].exp, data, request);
 937         RETURN(rc);
 938 }
 939
 940 struct obd_device *lmv_get_real_obd(struct obd_export *exp,
 941                                         char *name, int len)
 942 {
 943         struct obd_device *obd = exp->exp_obd;
 944         struct lmv_obd *lmv = &obd->u.lmv;
 945         ENTRY;
 946         lmv_connect(obd);
 947         obd = lmv->tgts[0].exp->exp_obd;
 948         EXIT;
 949         return obd;
 950 }
 951
 952 int lmv_init_ea_size(struct obd_export *exp, int easize, int cookiesize)
 953 {
 954         struct obd_device *obd = exp->exp_obd;
 955         struct lmv_obd *lmv = &obd->u.lmv;
 956         int i, rc = 0, change = 0;
 957         ENTRY;
 958
 959         if (lmv->max_easize < easize) {
 960                 lmv->max_easize = easize;
 961                 change = 1;
 962         }
 963         if (lmv->max_cookiesize < cookiesize) {
 964                 lmv->max_cookiesize = cookiesize;
 965                 change = 1;
 966         }
 967         if (change == 0)
 968                 RETURN(0);
 969
 970         if (lmv->connected == 0)
 971                 RETURN(0);
 972
 973         /* FIXME: error handling? */
 974         for (i = 0; i < lmv->count; i++)
 975                 rc = obd_init_ea_size(lmv->tgts[i].exp, easize, cookiesize);
 976         RETURN(rc);
 977 }
 978
 979 int lmv_obd_create_single(struct obd_export *exp, struct obdo *oa,
 980                           struct lov_stripe_md **ea, struct obd_trans_info *oti)
 981 {
 982         struct obd_device *obd = exp->exp_obd;
 983         struct lmv_obd *lmv = &obd->u.lmv;
 984         struct lov_stripe_md obj_md;
 985         struct lov_stripe_md *obj_mdp = &obj_md;
 986         int rc = 0;
 987         ENTRY;
 988         lmv_connect(obd);
 989
 990         LASSERT(ea == NULL);
 991         LASSERT(oa->o_mds < lmv->count);
 992
 993         rc = obd_create(lmv->tgts[oa->o_mds].exp, oa, &obj_mdp, oti);
 994         LASSERT(rc == 0);
 995
 996         RETURN(rc);
 997 }
 998
 999 /*
1000  * to be called from MDS only
1001  */
1002 int lmv_obd_create(struct obd_export *exp, struct obdo *oa,
1003                struct lov_stripe_md **ea, struct obd_trans_info *oti)
1004 {
1005         struct obd_device *obd = exp->exp_obd;
1006         struct lmv_obd *lmv = &obd->u.lmv;
1007         struct mea *mea;
1008         int i, c, rc = 0;
1009         struct ll_fid mfid;
1010         ENTRY;
1011         lmv_connect(obd);
1012
1013         LASSERT(oa != NULL);
1014
1015         if (ea == NULL) {
1016                 rc = lmv_obd_create_single(exp, oa, NULL, oti);
1017                 RETURN(rc);
1018         }
1019
1020         if (*ea == NULL) {
1021                 rc = obd_alloc_diskmd(exp, (struct lov_mds_md **) ea);
1022                 LASSERT(*ea != NULL);
1023         }
1024
1025         mea = (struct mea *) *ea;
1026         mfid.id = oa->o_id;
1027         mfid.generation = oa->o_generation;
1028         rc = 0;
1029         if (!mea->mea_count || mea->mea_count > lmv->count)
1030                 mea->mea_count = lmv->count;
1031
1032         mea->mea_master = -1;
1033
1034         /* FIXME: error handling? */
1035         for (i = 0, c = 0; c < mea->mea_count && i < lmv->count; i++) {
1036                 struct lov_stripe_md obj_md;
1037                 struct lov_stripe_md *obj_mdp = &obj_md;
1038
1039                 if (lmv->tgts[i].exp == NULL) {
1040                         /* this is master MDS */
1041                         mea->mea_fids[c].id = mfid.id;
1042                         mea->mea_fids[c].generation = mfid.generation;
1043                         mea->mea_fids[c].mds = i;
1044                         mea->mea_master = i;
1045                         c++;
1046                         continue;
1047                 }
1048
1049                 /* "Master" MDS should always be part of stripped dir, so
1050                    scan for it */
1051                 if (mea->mea_master == -1 && c == mea->mea_count - 1)
1052                         continue;
1053
1054                 oa->o_valid = OBD_MD_FLGENER | OBD_MD_FLTYPE | OBD_MD_FLMODE
1055                                 | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLID;
1056
1057                 rc = obd_create(lmv->tgts[c].exp, oa, &obj_mdp, oti);
1058                 /* FIXME: error handling here */
1059                 LASSERT(rc == 0);
1060
1061                 mea->mea_fids[c].id = oa->o_id;
1062                 mea->mea_fids[c].generation = oa->o_generation;
1063                 mea->mea_fids[c].mds = i;
1064                 c++;
1065                 CDEBUG(D_OTHER, "dirobj at mds %d: "LPU64"/%u\n",
1066                        i, oa->o_id, oa->o_generation);
1067         }
1068         LASSERT(c == mea->mea_count);
1069         CDEBUG(D_OTHER, "%d dirobjects created\n", (int) mea->mea_count);
1070
1071         RETURN(rc);
1072 }
1073
1074 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
1075                            void *key, __u32 *vallen, void *val)
1076 {
1077         struct obd_device *obd;
1078         struct lmv_obd *lmv;
1079         ENTRY;
1080
1081         obd = class_exp2obd(exp);
1082         if (obd == NULL) {
1083                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1084                        exp->exp_handle.h_cookie);
1085                 RETURN(-EINVAL);
1086         }
1087
1088         lmv = &obd->u.lmv;
1089         if (keylen == 6 && memcmp(key, "mdsize", 6) == 0) {
1090                 __u32 *mdsize = val;
1091                 *vallen = sizeof(__u32);
1092                 *mdsize = sizeof(struct ll_fid) * lmv->count
1093                                 + sizeof(struct mea);
1094                 RETURN(0);
1095         } else if (keylen == 6 && memcmp(key, "mdsnum", 6) == 0) {
1096                 struct obd_uuid *cluuid = &lmv->cluuid;
1097                 struct lmv_tgt_desc *tgts;
1098                 __u32 *mdsnum = val;
1099                 int i;
1100
1101                 for (i = 0, tgts = lmv->tgts; i < lmv->count; i++, tgts++) {
1102                         if (obd_uuid_equals(&tgts->uuid, cluuid)) {
1103                                 *vallen = sizeof(__u32);
1104                                 *mdsnum = i;
1105                                 RETURN(0);
1106                         }
1107                 }
1108                 LASSERT(0);
1109         }
1110
1111         CDEBUG(D_IOCTL, "invalid key\n");
1112         RETURN(-EINVAL);
1113 }
1114
1115 int lmv_set_info(struct obd_export *exp, obd_count keylen,
1116                  void *key, obd_count vallen, void *val)
1117 {
1118         struct obd_device *obd;
1119         struct lmv_obd *lmv;
1120         ENTRY;
1121
1122         obd = class_exp2obd(exp);
1123         if (obd == NULL) {
1124                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1125                        exp->exp_handle.h_cookie);
1126                 RETURN(-EINVAL);
1127         }
1128         lmv = &obd->u.lmv;
1129         lmv_connect(obd);
1130
1131         if (keylen >= strlen("client") && strcmp(key, "client") == 0) {
1132                 struct lmv_tgt_desc *tgts;
1133                 int i, rc;
1134
1135                 for (i = 0, tgts = lmv->tgts; i < lmv->count; i++, tgts++) {
1136                         rc = obd_set_info(tgts->exp, keylen, key, vallen, val);
1137                         if (rc)
1138                                 RETURN(rc);
1139                 }
1140                 RETURN(0);
1141         }
1142
1143         RETURN(-EINVAL);
1144 }
1145
1146 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
1147                struct lov_stripe_md *lsm)
1148 {
1149         struct obd_device *obd = class_exp2obd(exp);
1150         struct lmv_obd *lmv = &obd->u.lmv;
1151         int mea_size;
1152         ENTRY;
1153
1154         mea_size = sizeof(struct ll_fid) * lmv->count + sizeof(struct mea);
1155         if (!lmmp)
1156                 RETURN(mea_size);
1157
1158         if (*lmmp && !lsm) {
1159                 OBD_FREE(*lmmp, mea_size);
1160                 *lmmp = NULL;
1161                 RETURN(0);
1162         }
1163
1164         if (!*lmmp) {
1165                 OBD_ALLOC(*lmmp, mea_size);
1166                 if (!*lmmp)
1167                         RETURN(-ENOMEM);
1168         }
1169
1170         if (!lsm)
1171                 RETURN(mea_size);
1172
1173 #warning "MEA packing/convertation must be here! -bzzz"
1174         memcpy(*lmmp, lsm, mea_size);
1175         RETURN(mea_size);
1176 }
1177
1178 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **mem_tgt,
1179                         struct lov_mds_md *disk_src, int mdsize)
1180 {
1181         struct obd_device *obd = class_exp2obd(exp);
1182         struct lmv_obd *lmv = &obd->u.lmv;
1183         struct mea **tmea = (struct mea **) mem_tgt;
1184         struct mea *mea = (void *) disk_src;
1185         int mea_size;
1186         ENTRY;
1187
1188         mea_size = sizeof(struct ll_fid) * lmv->count + sizeof(struct mea);
1189         if (mem_tgt == NULL)
1190                 return mea_size;
1191
1192         if (*mem_tgt != NULL && disk_src == NULL) {
1193                 OBD_FREE(*tmea, mea_size);
1194                 RETURN(0);
1195         }
1196
1197         LASSERT(mea_size == mdsize);
1198
1199         OBD_ALLOC(*tmea, mea_size);
1200         /* FIXME: error handling here */
1201         LASSERT(*tmea != NULL);
1202
1203         if (!disk_src)
1204                 RETURN(mea_size);
1205
1206 #warning "MEA unpacking/convertation must be here! -bzzz"
1207         memcpy(*tmea, mea, mdsize);
1208         RETURN(mea_size);
1209 }
1210
1211 int lmv_brw(int rw, struct obd_export *exp, struct obdo *oa,
1212                 struct lov_stripe_md *ea, obd_count oa_bufs,
1213                 struct brw_page *pgarr, struct obd_trans_info *oti)
1214 {
1215         struct obd_device *obd = exp->exp_obd;
1216         struct lmv_obd *lmv = &obd->u.lmv;
1217         struct mea *mea = (struct mea *) ea;
1218         int err;
1219
1220         LASSERT(oa != NULL);
1221         LASSERT(ea != NULL);
1222         LASSERT(pgarr != NULL);
1223         LASSERT(oa->o_mds < lmv->count);
1224
1225         oa->o_gr = mea->mea_fids[oa->o_mds].generation;
1226         oa->o_id = mea->mea_fids[oa->o_mds].id;
1227         oa->o_valid =  OBD_MD_FLID | OBD_MD_FLGROUP;
1228         err = obd_brw(rw, lmv->tgts[oa->o_mds].exp, oa,
1229                         NULL, oa_bufs, pgarr, oti);
1230         RETURN(err);
1231 }
1232
1233 struct obd_ops lmv_obd_ops = {
1234         o_owner:                THIS_MODULE,
1235         o_attach:               lmv_attach,
1236         o_detach:               lmv_detach,
1237         o_setup:                lmv_setup,
1238         o_cleanup:              lmv_cleanup,
1239         o_connect:              lmv_connect_fake,
1240         o_disconnect:           lmv_disconnect,
1241         o_statfs:               lmv_statfs,
1242         o_get_info:             lmv_get_info,
1243         o_set_info:             lmv_set_info,
1244         o_create:               lmv_obd_create,
1245         o_packmd:               lmv_packmd,
1246         o_unpackmd:             lmv_unpackmd,
1247         o_brw:                  lmv_brw,
1248         o_init_ea_size:         lmv_init_ea_size,
1249 };
1250
1251 struct md_ops lmv_md_ops = {
1252         m_getstatus:            lmv_getstatus,
1253         m_getattr:              lmv_getattr,
1254         m_change_cbdata:        lmv_change_cbdata,
1255         m_change_cbdata_name:   lmv_change_cbdata_name,
1256         m_close:                lmv_close,
1257         m_create:               lmv_create,
1258         m_done_writing:         lmv_done_writing,
1259         m_enqueue:              lmv_enqueue,
1260         m_getattr_name:         lmv_getattr_name,
1261         m_intent_lock:          lmv_intent_lock,
1262         m_link:                 lmv_link,
1263         m_rename:               lmv_rename,
1264         m_setattr:              lmv_setattr,
1265         m_sync:                 lmv_sync,
1266         m_readpage:             lmv_readpage,
1267         m_unlink:               lmv_unlink,
1268         m_get_real_obd:         lmv_get_real_obd,
1269         m_valid_attrs:          lmv_valid_attrs,
1270 };
1271
1272 //#ifndef LPROCFS
1273 static struct lprocfs_vars lprocfs_module_vars[] = { {0} };
1274 static struct lprocfs_vars lprocfs_obd_vars[] = { {0} };
1275 //#else
1276 LPROCFS_INIT_VARS(lmv, lprocfs_module_vars, lprocfs_obd_vars)
1277
1278 int __init lmv_init(void)
1279 {
1280         struct lprocfs_static_vars lvars;
1281         int rc;
1282
1283         lprocfs_init_vars(lmv, &lvars);
1284         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
1285                                  lvars.module_vars, OBD_LMV_DEVICENAME);
1286         RETURN(rc);
1287 }
1288
1289 static void lmv_exit(void)
1290 {
1291         class_unregister_type(OBD_LMV_DEVICENAME);
1292 }
1293
1294 #ifdef __KERNEL__
1295 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1296 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
1297 MODULE_LICENSE("GPL");
1298
1299 module_init(lmv_init);
1300 module_exit(lmv_exit);
1301 #endif
1302