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 #include <linux/seq_file.h>
  34 #else
  35 #include <liblustre.h>
  36 #endif
  37 #include <linux/ext2_fs.h>
  38
  39 #include <linux/obd_support.h>
  40 #include <linux/lustre_lib.h>
  41 #include <linux/lustre_net.h>
  42 #include <linux/lustre_idl.h>
  43 #include <linux/lustre_dlm.h>
  44 #include <linux/lustre_mds.h>
  45 #include <linux/obd_class.h>
  46 #include <linux/obd_ost.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 /* Error codes:
  53  *
  54  *  -EINVAL  : UUID can't be found in the LMV's target list
  55  *  -ENOTCONN: The UUID is found, but the target connection is bad (!)
  56  *  -EBADF   : The UUID is found, but the OBD of the wrong type (!)
  57  */
  58 static int lmv_set_mdc_active(struct lmv_obd *lmv, struct obd_uuid *uuid,
  59                               int activate)
  60 {
  61         struct obd_device *obd;
  62         struct lmv_tgt_desc *tgt;
  63         int i, rc = 0;
  64         ENTRY;
  65
  66         CDEBUG(D_INFO, "Searching in lmv %p for uuid %s (activate=%d)\n",
  67                lmv, uuid->uuid, activate);
  68
  69         spin_lock(&lmv->lmv_lock);
  70         for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) {
  71                 if (tgt->ltd_exp == NULL)
  72                         continue;
  73
  74                 CDEBUG(D_INFO, "lmv idx %d is %s conn "LPX64"\n",
  75                        i, tgt->uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie);
  76                 if (strncmp(uuid->uuid, tgt->uuid.uuid, sizeof uuid->uuid) == 0)
  77                         break;
  78         }
  79
  80         if (i == lmv->desc.ld_tgt_count)
  81                 GOTO(out, rc = -EINVAL);
  82
  83         obd = class_exp2obd(tgt->ltd_exp);
  84         if (obd == NULL)
  85                 GOTO(out, rc = -ENOTCONN);
  86
  87         CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n",
  88                obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd,
  89                obd->obd_type->typ_name, i);
  90         LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0);
  91
  92         if (tgt->active == activate) {
  93                 CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd,
  94                        activate ? "" : "in");
  95                 GOTO(out, rc);
  96         }
  97
  98         CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd, activate ? "" : "in");
  99
 100         tgt->active = activate;
 101         if (activate)
 102                 lmv->desc.ld_active_tgt_count++;
 103         else
 104                 lmv->desc.ld_active_tgt_count--;
 105
 106         EXIT;
 107  out:
 108         spin_unlock(&lmv->lmv_lock);
 109         return rc;
 110 }
 111
 112 static int lmv_notify(struct obd_device *obd, struct obd_device *watched,
 113                       int active)
 114 {
 115         int rc;
 116         struct obd_uuid *uuid;
 117
 118         if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) {
 119                 CERROR("unexpected notification of %s %s!\n",
 120                        watched->obd_type->typ_name,
 121                        watched->obd_name);
 122                 return -EINVAL;
 123         }
 124         uuid = &watched->u.cli.cl_import->imp_target_uuid;
 125
 126         /* Set MDC as active before notifying the observer, so the
 127          * observer can use the MDC normally.
 128          */
 129         rc = lmv_set_mdc_active(&obd->u.lmv, uuid, active);
 130         if (rc) {
 131                 CERROR("%sactivation of %s failed: %d\n",
 132                        active ? "" : "de", uuid->uuid, rc);
 133                 RETURN(rc);
 134         }
 135
 136         if (obd->obd_observer)
 137                 /* Pass the notification up the chain. */
 138                 rc = obd_notify(obd->obd_observer, watched, active);
 139
 140         RETURN(rc);
 141 }
 142
 143 int lmv_attach(struct obd_device *dev, obd_count len, void *data)
 144 {
 145         struct lprocfs_static_vars lvars;
 146         int rc;
 147         ENTRY;
 148
 149         lprocfs_init_vars(lmv, &lvars);
 150         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
 151         if (rc == 0) {
 152 #ifdef __KERNEL__
 153                 struct proc_dir_entry *entry;
 154
 155                 entry = create_proc_entry("target_obd", 0444, dev->obd_proc_entry);
 156                 if (entry == NULL)
 157                         RETURN(-ENOMEM);
 158                 /* entry->proc_fops = &lmv_proc_target_fops; */
 159                 entry->data = dev;
 160 #endif
 161        }
 162         RETURN (rc);
 163 }
 164
 165 int lmv_detach(struct obd_device *dev)
 166 {
 167         return lprocfs_obd_detach(dev);
 168 }
 169
 170 /* This is fake connect function. Its purpose is to initialize lmv and
 171  * say caller that everything is okay. Real connection will be performed
 172  * later. */
 173 static int lmv_connect(struct lustre_handle *conn, struct obd_device *obd,
 174                        struct obd_uuid *cluuid)
 175 {
 176         struct lmv_obd *lmv = &obd->u.lmv;
 177         struct obd_export *exp;
 178         int rc;
 179         ENTRY;
 180
 181         rc = class_connect(conn, obd, cluuid);
 182         if (rc) {
 183                 CERROR("class_connection() returned %d\n", rc);
 184                 RETURN(rc);
 185         }
 186
 187         exp = class_conn2export(conn);
 188         /* We don't want to actually do the underlying connections more than
 189          * once, so keep track. */
 190         lmv->refcount++;
 191         if (lmv->refcount > 1) {
 192                 class_export_put(exp);
 193                 RETURN(0);
 194         }
 195
 196         lmv->cluuid = *cluuid;
 197         lmv->connected = 0;
 198         lmv->exp = exp;
 199
 200         RETURN(0);
 201 }
 202
 203 void lmv_set_timeouts(struct obd_device *obd)
 204 {
 205         struct lmv_tgt_desc *tgts;
 206         struct lmv_obd *lmv;
 207         int i;
 208
 209         lmv = &obd->u.lmv;
 210         if (lmv->server_timeout == 0)
 211                 return;
 212
 213         if (lmv->connected == 0)
 214                 return;
 215
 216         for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
 217                 if (tgts->ltd_exp == NULL)
 218                         continue;
 219                 obd_set_info(tgts->ltd_exp, strlen("inter_mds"),
 220                              "inter_mds", 0, NULL);
 221         }
 222 }
 223
 224 /* Performs a check if passed obd is connected. If no - connect it. */
 225 int lmv_check_connect(struct obd_device *obd) {
 226         struct lmv_obd *lmv = &obd->u.lmv;
 227         struct obd_uuid *cluuid;
 228         struct lmv_tgt_desc *tgts;
 229         struct obd_export *exp;
 230         int rc, rc2, i;
 231
 232         if (lmv->connected)
 233                 return 0;
 234
 235         lmv->connected = 1;
 236         cluuid = &lmv->cluuid;
 237         exp = lmv->exp;
 238
 239         CDEBUG(D_OTHER, "time to connect %s to %s\n",
 240                cluuid->uuid, obd->obd_name);
 241
 242         for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
 243                 struct obd_device *tgt_obd;
 244                 struct obd_uuid lmv_osc_uuid = { "LMV_OSC_UUID" };
 245                 struct lustre_handle conn = {0, };
 246
 247                 LASSERT(tgts != NULL);
 248
 249                 tgt_obd = class_find_client_obd(&tgts->uuid, LUSTRE_MDC_NAME,
 250                                                 &obd->obd_uuid);
 251                 if (!tgt_obd) {
 252                         CERROR("Target %s not attached\n", tgts->uuid.uuid);
 253                         GOTO(out_disc, rc = -EINVAL);
 254                 }
 255
 256                 /* for MDS: don't connect to yourself */
 257                 if (obd_uuid_equals(&tgts->uuid, cluuid)) {
 258                         CDEBUG(D_OTHER, "don't connect back to %s\n",
 259                                cluuid->uuid);
 260                         tgts->ltd_exp = NULL;
 261                         continue;
 262                 }
 263
 264                 CDEBUG(D_OTHER, "connect to %s(%s) - %s, %s FOR %s\n",
 265                         tgt_obd->obd_name, tgt_obd->obd_uuid.uuid,
 266                         tgts->uuid.uuid, obd->obd_uuid.uuid,
 267                         cluuid->uuid);
 268
 269                 if (!tgt_obd->obd_set_up) {
 270                         CERROR("Target %s not set up\n", tgts->uuid.uuid);
 271                         GOTO(out_disc, rc = -EINVAL);
 272                 }
 273
 274                 rc = obd_connect(&conn, tgt_obd, &lmv_osc_uuid);
 275                 if (rc) {
 276                         CERROR("Target %s connect error %d\n",
 277                                 tgts->uuid.uuid, rc);
 278                         GOTO(out_disc, rc);
 279                 }
 280                 tgts->ltd_exp = class_conn2export(&conn);
 281
 282                 obd_init_ea_size(tgts->ltd_exp, lmv->max_easize,
 283                                  lmv->max_cookiesize);
 284
 285                 rc = obd_register_observer(tgt_obd, obd);
 286                 if (rc) {
 287                         CERROR("Target %s register_observer error %d\n",
 288                                tgts->uuid.uuid, rc);
 289                         obd_disconnect(tgts->ltd_exp, 0);
 290                         GOTO(out_disc, rc);
 291                 }
 292
 293                 lmv->desc.ld_active_tgt_count++;
 294                 tgts->active = 1;
 295
 296                 CDEBUG(D_OTHER, "connected to %s(%s) successfully (%d)\n",
 297                         tgt_obd->obd_name, tgt_obd->obd_uuid.uuid,
 298                         atomic_read(&obd->obd_refcount));
 299         }
 300
 301         lmv_set_timeouts(obd);
 302
 303         class_export_put(exp);
 304         return 0;
 305
 306  out_disc:
 307         while (i-- > 0) {
 308                 struct obd_uuid uuid;
 309                 --tgts;
 310                 --lmv->desc.ld_active_tgt_count;
 311                 tgts->active = 0;
 312                 /* save for CERROR below; (we know it's terminated) */
 313                 uuid = tgts->uuid;
 314                 rc2 = obd_disconnect(tgts->ltd_exp, 0);
 315                 if (rc2)
 316                         CERROR("error: LMV target %s disconnect on MDT idx %d: "
 317                                "rc = %d\n", uuid.uuid, i, rc2);
 318         }
 319         class_disconnect(exp, 0);
 320         RETURN (rc);
 321 }
 322
 323 static int lmv_disconnect(struct obd_export *exp, int flags)
 324 {
 325         struct obd_device *obd = class_exp2obd(exp);
 326         struct lmv_obd *lmv = &obd->u.lmv;
 327         int rc, i;
 328         ENTRY;
 329
 330         if (!lmv->tgts)
 331                 goto out_local;
 332
 333         /* Only disconnect the underlying layers on the final disconnect. */
 334         lmv->refcount--;
 335         if (lmv->refcount != 0)
 336                 goto out_local;
 337
 338         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 339                 if (lmv->tgts[i].ltd_exp == NULL)
 340                         continue;
 341
 342                 if (obd->obd_no_recov) {
 343                         /* Pass it on to our clients.
 344                          * XXX This should be an argument to disconnect,
 345                          * XXX not a back-door flag on the OBD.  Ah well.
 346                          */
 347                         struct obd_device *mdc_obd;
 348                         mdc_obd = class_exp2obd(lmv->tgts[i].ltd_exp);
 349                         if (mdc_obd)
 350                                 mdc_obd->obd_no_recov = 1;
 351                 }
 352
 353                 CDEBUG(D_OTHER, "disconnected from %s(%s) successfully\n",
 354                         lmv->tgts[i].ltd_exp->exp_obd->obd_name,
 355                         lmv->tgts[i].ltd_exp->exp_obd->obd_uuid.uuid);
 356
 357                 obd_register_observer(lmv->tgts[i].ltd_exp->exp_obd, NULL);
 358
 359                 rc = obd_disconnect(lmv->tgts[i].ltd_exp, flags);
 360                 if (rc) {
 361                         if (lmv->tgts[i].active) {
 362                                 CERROR("Target %s disconnect error %d\n",
 363                                        lmv->tgts[i].uuid.uuid, rc);
 364                         }
 365                         rc = 0;
 366                 }
 367                 if (lmv->tgts[i].active) {
 368                         lmv->desc.ld_active_tgt_count--;
 369                         lmv->tgts[i].active = 0;
 370                 }
 371                 lmv->tgts[i].ltd_exp = NULL;
 372         }
 373
 374 out_local:
 375         /* This is the case when no real connection is established by
 376          * lmv_check_connect(). */
 377         if (!lmv->connected)
 378                 class_export_put(exp);
 379         rc = class_disconnect(exp, 0);
 380         RETURN(rc);
 381 }
 382
 383 static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp,
 384                          int len, void *karg, void *uarg)
 385 {
 386         struct obd_device *obddev = class_exp2obd(exp);
 387         struct lmv_obd *lmv = &obddev->u.lmv;
 388         int i, rc = 0, set = 0;
 389
 390         ENTRY;
 391
 392         if (lmv->desc.ld_tgt_count == 0)
 393                 RETURN(-ENOTTY);
 394
 395         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 396                 int err;
 397
 398                 err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp,
 399                                     len, karg, uarg);
 400                 if (err) {
 401                         if (lmv->tgts[i].active) {
 402                                 CERROR("error: iocontrol MDC %s on MDT"
 403                                        "idx %d: err = %d\n",
 404                                        lmv->tgts[i].uuid.uuid, i, err);
 405                                 if (!rc)
 406                                         rc = err;
 407                         }
 408                 } else
 409                         set = 1;
 410         }
 411         if (!set && !rc)
 412                 rc = -EIO;
 413
 414         RETURN(rc);
 415 }
 416
 417 static int lmv_setup(struct obd_device *obd, obd_count len, void *buf)
 418 {
 419         struct lustre_cfg *lcfg = buf;
 420         struct lmv_desc *desc;
 421         struct lmv_obd *lmv = &obd->u.lmv;
 422         struct obd_uuid *uuids;
 423         struct lmv_tgt_desc *tgts;
 424         int i;
 425         int rc = 0;
 426         ENTRY;
 427
 428         if (lcfg->lcfg_inllen1 < 1) {
 429                 CERROR("LMV setup requires a descriptor\n");
 430                 RETURN(-EINVAL);
 431         }
 432
 433         if (lcfg->lcfg_inllen2 < 1) {
 434                 CERROR("LMV setup requires an OST UUID list\n");
 435                 RETURN(-EINVAL);
 436         }
 437
 438         desc = (struct lmv_desc *)lcfg->lcfg_inlbuf1;
 439         if (sizeof(*desc) > lcfg->lcfg_inllen1) {
 440                 CERROR("descriptor size wrong: %d > %d\n",
 441                        (int)sizeof(*desc), lcfg->lcfg_inllen1);
 442                 RETURN(-EINVAL);
 443         }
 444
 445         uuids = (struct obd_uuid *)lcfg->lcfg_inlbuf2;
 446         if (sizeof(*uuids) * desc->ld_tgt_count != lcfg->lcfg_inllen2) {
 447                 CERROR("UUID array size wrong: %u * %u != %u\n",
 448                        sizeof(*uuids), desc->ld_tgt_count, lcfg->lcfg_inllen2);
 449                 RETURN(-EINVAL);
 450         }
 451
 452         lmv->bufsize = sizeof(struct lmv_tgt_desc) * desc->ld_tgt_count;
 453         OBD_ALLOC(lmv->tgts, lmv->bufsize);
 454         if (lmv->tgts == NULL) {
 455                 CERROR("Out of memory\n");
 456                 RETURN(-EINVAL);
 457         }
 458
 459         lmv->desc = *desc;
 460         spin_lock_init(&lmv->lmv_lock);
 461
 462         for (i = 0, tgts = lmv->tgts; i < desc->ld_tgt_count; i++, tgts++)
 463                 tgts->uuid = uuids[i];
 464
 465         lmv->max_easize = sizeof(struct ll_fid) *
 466                 desc->ld_tgt_count + sizeof(struct mea);
 467
 468         lmv->max_cookiesize = 0;
 469
 470         RETURN(rc);
 471 }
 472
 473 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
 474                       unsigned long max_age)
 475 {
 476         struct lmv_obd *lmv = &obd->u.lmv;
 477         struct obd_statfs temp;
 478         int rc = 0, i;
 479         ENTRY;
 480
 481         rc = lmv_check_connect(obd);
 482         if (rc)
 483                 RETURN(rc);
 484
 485         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 486                 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, &temp, max_age);
 487                 if (rc) {
 488                         CERROR("can't stat MDS #%d (%s)\n", i,
 489                                lmv->tgts[i].ltd_exp->exp_obd->obd_name);
 490                         RETURN(rc);
 491                 }
 492                 if (i == 0) {
 493                         memcpy(osfs, &temp, sizeof(temp));
 494                 } else {
 495                         osfs->os_bavail += temp.os_bavail;
 496                         osfs->os_blocks += temp.os_blocks;
 497                         osfs->os_ffree += temp.os_ffree;
 498                         osfs->os_files += temp.os_files;
 499                 }
 500         }
 501         RETURN(rc);
 502 }
 503
 504 static int lmv_cleanup(struct obd_device *obd, int flags)
 505 {
 506         struct lmv_obd *lmv = &obd->u.lmv;
 507         ENTRY;
 508         lmv_cleanup_objs(obd);
 509         OBD_FREE(lmv->tgts, lmv->bufsize);
 510         RETURN(0);
 511 }
 512
 513 static int lmv_getstatus(struct obd_export *exp, struct ll_fid *fid)
 514 {
 515         struct obd_device *obd = exp->exp_obd;
 516         struct lmv_obd *lmv = &obd->u.lmv;
 517         int rc;
 518         ENTRY;
 519         rc = lmv_check_connect(obd);
 520         if (rc)
 521                 RETURN(rc);
 522         rc = md_getstatus(lmv->tgts[0].ltd_exp, fid);
 523         fid->mds = 0;
 524         RETURN(rc);
 525 }
 526
 527 static int lmv_getattr(struct obd_export *exp, struct ll_fid *fid,
 528                 unsigned long valid, unsigned int ea_size,
 529                 struct ptlrpc_request **request)
 530 {
 531         struct obd_device *obd = exp->exp_obd;
 532         struct lmv_obd *lmv = &obd->u.lmv;
 533         int rc, i = fid->mds;
 534         struct lmv_obj *obj;
 535         ENTRY;
 536         rc = lmv_check_connect(obd);
 537         if (rc)
 538                 RETURN(rc);
 539         obj = lmv_grab_obj(obd, fid, 0);
 540         CDEBUG(D_OTHER, "GETATTR for %lu/%lu/%lu %s\n",
 541                (unsigned long) fid->mds,
 542                (unsigned long) fid->id,
 543                (unsigned long) fid->generation,
 544                obj ? "(splitted)" : "");
 545
 546         LASSERT(fid->mds < lmv->desc.ld_tgt_count);
 547         rc = md_getattr(lmv->tgts[i].ltd_exp, fid,
 548                              valid, ea_size, request);
 549         if (rc == 0 && obj) {
 550                 /* we have to loop over dirobjs here and gather attrs
 551                  * for all the slaves */
 552 #warning "attrs gathering here"
 553         }
 554         lmv_put_obj(obj);
 555         RETURN(rc);
 556 }
 557
 558 static int lmv_change_cbdata(struct obd_export *exp,
 559                                  struct ll_fid *fid,
 560                                  ldlm_iterator_t it, void *data)
 561 {
 562         struct obd_device *obd = exp->exp_obd;
 563         struct lmv_obd *lmv = &obd->u.lmv;
 564         int rc = 0;
 565         ENTRY;
 566
 567         rc = lmv_check_connect(obd);
 568         if (rc)
 569                 RETURN(rc);
 570         CDEBUG(D_OTHER, "CBDATA for %lu/%lu/%lu\n",
 571                (unsigned long) fid->mds,
 572                (unsigned long) fid->id,
 573                (unsigned long) fid->generation);
 574         LASSERT(fid->mds < lmv->desc.ld_tgt_count);
 575         rc = md_change_cbdata(lmv->tgts[fid->mds].ltd_exp, fid, it, data);
 576         RETURN(rc);
 577 }
 578
 579 static int lmv_change_cbdata_name(struct obd_export *exp, struct ll_fid *pfid,
 580                                   char *name, int len, struct ll_fid *cfid,
 581                                   ldlm_iterator_t it, void *data)
 582 {
 583         struct obd_device *obd = exp->exp_obd;
 584         struct lmv_obd *lmv = &obd->u.lmv;
 585         struct lmv_obj *obj;
 586         int rc = 0, mds;
 587         ENTRY;
 588         rc = lmv_check_connect(obd);
 589         if (rc)
 590                 RETURN(rc);
 591         LASSERT(pfid->mds < lmv->desc.ld_tgt_count);
 592         LASSERT(cfid->mds < lmv->desc.ld_tgt_count);
 593         CDEBUG(D_OTHER, "CBDATA for %lu/%lu/%lu:%*s -> %lu/%lu/%lu\n",
 594                (unsigned long) pfid->mds, (unsigned long) pfid->id,
 595                (unsigned long) pfid->generation, len, name,
 596                (unsigned long) cfid->mds, (unsigned long) cfid->id,
 597                (unsigned long) cfid->generation);
 598
 599         /* this is default mds for directory name belongs to */
 600         mds = pfid->mds;
 601         obj = lmv_grab_obj(obd, pfid, 0);
 602         if (obj) {
 603                 /* directory is splitted. look for right mds for this name */
 604                 mds = raw_name2idx(obj->objcount, name, len);
 605                 lmv_put_obj(obj);
 606         }
 607         rc = md_change_cbdata(lmv->tgts[mds].ltd_exp, cfid, it, data);
 608         RETURN(rc);
 609 }
 610
 611 static int lmv_valid_attrs(struct obd_export *exp, struct ll_fid *fid)
 612 {
 613         struct obd_device *obd = exp->exp_obd;
 614         struct lmv_obd *lmv = &obd->u.lmv;
 615         int rc = 0;
 616         ENTRY;
 617         rc = lmv_check_connect(obd);
 618         if (rc)
 619                 RETURN(rc);
 620         CDEBUG(D_OTHER, "validate %lu/%lu/%lu\n",
 621                (unsigned long) fid->mds,
 622                (unsigned long) fid->id,
 623                (unsigned long) fid->generation);
 624         LASSERT(fid->mds < lmv->desc.ld_tgt_count);
 625         rc = md_valid_attrs(lmv->tgts[fid->mds].ltd_exp, fid);
 626         RETURN(rc);
 627 }
 628
 629 int lmv_close(struct obd_export *exp, struct obdo *obdo,
 630                   struct obd_client_handle *och,
 631                   struct ptlrpc_request **request)
 632 {
 633         struct obd_device *obd = exp->exp_obd;
 634         struct lmv_obd *lmv = &obd->u.lmv;
 635         int rc, i = obdo->o_mds;
 636         ENTRY;
 637         rc = lmv_check_connect(obd);
 638         if (rc)
 639                 RETURN(rc);
 640         LASSERT(i < lmv->desc.ld_tgt_count);
 641         CDEBUG(D_OTHER, "CLOSE %lu/%lu/%lu\n", (unsigned long) obdo->o_mds,
 642                (unsigned long) obdo->o_id, (unsigned long) obdo->o_generation);
 643         rc = md_close(lmv->tgts[i].ltd_exp, obdo, och, request);
 644         RETURN(rc);
 645 }
 646
 647 int lmv_get_mea_and_update_object(struct obd_export *exp, struct ll_fid *fid)
 648 {
 649         struct obd_device *obd = exp->exp_obd;
 650         struct lmv_obd *lmv = &obd->u.lmv;
 651         struct ptlrpc_request *req = NULL;
 652         struct lustre_md md;
 653         unsigned long valid;
 654         int mealen, rc;
 655
 656         md.mea = NULL;
 657         mealen = MEA_SIZE_LMV(lmv);
 658
 659         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
 660
 661         /* time to update mea of parent fid */
 662         rc = md_getattr(lmv->tgts[fid->mds].ltd_exp, fid,
 663                         valid, mealen, &req);
 664         if (rc) {
 665                 CERROR("md_getattr() failed, rc = %d\n", rc);
 666                 GOTO(cleanup, rc);
 667         }
 668
 669         rc = mdc_req2lustre_md(exp, req, 0, NULL, &md);
 670         if (rc) {
 671                 CERROR("mdc_req2lustre_md() failed, rc = %d\n", rc);
 672                 GOTO(cleanup, rc);
 673         }
 674
 675         if (md.mea == NULL)
 676                 GOTO(cleanup, rc = -ENODATA);
 677
 678         rc = lmv_create_obj_from_attrs(exp, fid, md.mea);
 679         obd_free_memmd(exp, (struct lov_stripe_md **) &md.mea);
 680
 681 cleanup:
 682         if (req)
 683                 ptlrpc_req_finished(req);
 684         RETURN(rc);
 685 }
 686
 687 int lmv_create(struct obd_export *exp, struct mdc_op_data *op_data,
 688                const void *data, int datalen, int mode, __u32 uid,
 689                __u32 gid, __u64 rdev, struct ptlrpc_request **request)
 690 {
 691         struct obd_device *obd = exp->exp_obd;
 692         struct lmv_obd *lmv = &obd->u.lmv;
 693         struct mds_body *mds_body;
 694         struct lmv_obj *obj;
 695         int rc, mds;
 696         ENTRY;
 697
 698         rc = lmv_check_connect(obd);
 699         if (rc)
 700                 RETURN(rc);
 701
 702         if (!lmv->desc.ld_active_tgt_count)
 703                 RETURN(-EIO);
 704 repeat:
 705         obj = lmv_grab_obj(obd, &op_data->fid1, 0);
 706         if (obj) {
 707                 mds = raw_name2idx(obj->objcount, op_data->name,
 708                                    op_data->namelen);
 709                 op_data->fid1 = obj->objs[mds].fid;
 710                 lmv_put_obj(obj);
 711         }
 712
 713         CDEBUG(D_OTHER, "CREATE '%*s' on %lu/%lu/%lu\n",
 714                         op_data->namelen, op_data->name,
 715                         (unsigned long) op_data->fid1.mds,
 716                         (unsigned long) op_data->fid1.id,
 717                         (unsigned long) op_data->fid1.generation);
 718         rc = md_create(lmv->tgts[op_data->fid1.mds].ltd_exp, op_data, data,
 719                        datalen, mode, uid, gid, rdev, request);
 720         if (rc == 0) {
 721                 if (*request == NULL)
 722                      RETURN(rc);
 723                 mds_body = lustre_msg_buf((*request)->rq_repmsg, 0,
 724                                           sizeof(*mds_body));
 725                 LASSERT(mds_body != NULL);
 726                 CDEBUG(D_OTHER, "created. id = %lu, generation = %lu, mds = %d\n",
 727                        (unsigned long) mds_body->fid1.id,
 728                        (unsigned long) mds_body->fid1.generation,
 729                        op_data->fid1.mds);
 730                 LASSERT(mds_body->valid & OBD_MD_MDS ||
 731                         mds_body->mds == op_data->fid1.mds);
 732         } else if (rc == -ERESTART) {
 733                 /* directory got splitted. time to update local object
 734                  * and repeat the request with proper MDS */
 735                 rc = lmv_get_mea_and_update_object(exp, &op_data->fid1);
 736                 if (rc == 0) {
 737                         ptlrpc_req_finished(*request);
 738                         goto repeat;
 739                 }
 740         }
 741         RETURN(rc);
 742 }
 743
 744 int lmv_done_writing(struct obd_export *exp, struct obdo *obdo)
 745 {
 746         struct obd_device *obd = exp->exp_obd;
 747         struct lmv_obd *lmv = &obd->u.lmv;
 748         int rc;
 749         ENTRY;
 750         rc = lmv_check_connect(obd);
 751         if (rc)
 752                 RETURN(rc);
 753
 754         /* FIXME: choose right MDC here */
 755         CWARN("this method isn't implemented yet\n");
 756         rc = md_done_writing(lmv->tgts[0].ltd_exp, obdo);
 757         RETURN(rc);
 758 }
 759
 760 int lmv_enqueue_slaves(struct obd_export *exp, int locktype,
 761                          struct lookup_intent *it, int lockmode,
 762                          struct mdc_op_data *data, struct lustre_handle *lockh,
 763                          void *lmm, int lmmsize,
 764                          ldlm_completion_callback cb_completion,
 765                          ldlm_blocking_callback cb_blocking, void *cb_data)
 766 {
 767         struct obd_device *obd = exp->exp_obd;
 768         struct lmv_obd *lmv = &obd->u.lmv;
 769         struct mea *mea = data->mea1;
 770         struct mdc_op_data data2;
 771         int i, rc, mds;
 772         ENTRY;
 773
 774         LASSERT(mea != NULL);
 775         for (i = 0; i < mea->mea_count; i++) {
 776                 if (lmv->tgts[i].ltd_exp == NULL)
 777                         continue;
 778
 779                 memset(&data2, 0, sizeof(data2));
 780                 data2.fid1 = mea->mea_fids[i];
 781                 mds = data2.fid1.mds;
 782                 rc = md_enqueue(lmv->tgts[mds].ltd_exp, locktype, it, lockmode,
 783                                 &data2, lockh + i, lmm, lmmsize, cb_completion,
 784                                 cb_blocking, cb_data);
 785                 CDEBUG(D_OTHER, "take lock on slave %lu/%lu/%lu -> %d/%d\n",
 786                        (unsigned long) mea->mea_fids[i].mds,
 787                        (unsigned long) mea->mea_fids[i].id,
 788                        (unsigned long) mea->mea_fids[i].generation,
 789                        rc, it->d.lustre.it_status);
 790                 if (rc)
 791                         GOTO(cleanup, rc);
 792                 if (it->d.lustre.it_data) {
 793                         struct ptlrpc_request *req;
 794                         req = (struct ptlrpc_request *) it->d.lustre.it_data;
 795                         ptlrpc_req_finished(req);
 796                 }
 797
 798                 if (it->d.lustre.it_status)
 799                         GOTO(cleanup, rc = it->d.lustre.it_status);
 800         }
 801         RETURN(0);
 802
 803 cleanup:
 804         /* drop all taken locks */
 805         while (--i >= 0) {
 806                 if (lockh[i].cookie)
 807                         ldlm_lock_decref(lockh + i, lockmode);
 808                 lockh[i].cookie = 0;
 809         }
 810         RETURN(rc);
 811 }
 812
 813 int lmv_enqueue(struct obd_export *exp, int lock_type,
 814                 struct lookup_intent *it, int lock_mode,
 815                 struct mdc_op_data *data, struct lustre_handle *lockh,
 816                 void *lmm, int lmmsize,
 817                 ldlm_completion_callback cb_completion,
 818                 ldlm_blocking_callback cb_blocking, void *cb_data)
 819 {
 820         struct obd_device *obd = exp->exp_obd;
 821         struct lmv_obd *lmv = &obd->u.lmv;
 822         struct lmv_obj *obj;
 823         int rc, mds;
 824         ENTRY;
 825
 826         rc = lmv_check_connect(obd);
 827         if (rc)
 828                 RETURN(rc);
 829
 830         if (it->it_op == IT_UNLINK) {
 831                 rc = lmv_enqueue_slaves(exp, lock_type, it, lock_mode,
 832                                         data, lockh, lmm, lmmsize,
 833                                         cb_completion, cb_blocking, cb_data);
 834                 RETURN(rc);
 835         }
 836
 837         if (data->namelen) {
 838                 obj = lmv_grab_obj(obd, &data->fid1, 0);
 839                 if (obj) {
 840                         /* directory is splitted. look for
 841                          * right mds for this name */
 842                         mds = raw_name2idx(obj->objcount, (char *)data->name,
 843                                            data->namelen);
 844                         data->fid1 = obj->objs[mds].fid;
 845                         lmv_put_obj(obj);
 846                 }
 847         }
 848         CDEBUG(D_OTHER, "ENQUEUE '%s' on %lu/%lu\n",
 849                LL_IT2STR(it), (unsigned long) data->fid1.id,
 850                (unsigned long) data->fid1.generation);
 851         rc = md_enqueue(lmv->tgts[data->fid1.mds].ltd_exp, lock_type, it,
 852                         lock_mode, data, lockh, lmm, lmmsize, cb_completion,
 853                         cb_blocking, cb_data);
 854
 855         RETURN(rc);
 856 }
 857
 858 int lmv_getattr_name(struct obd_export *exp, struct ll_fid *fid,
 859                          char *filename, int namelen, unsigned long valid,
 860                          unsigned int ea_size, struct ptlrpc_request **request)
 861 {
 862         struct obd_device *obd = exp->exp_obd;
 863         struct lmv_obd *lmv = &obd->u.lmv;
 864         struct ll_fid rfid = *fid;
 865         int rc, mds = fid->mds;
 866         struct mds_body *body;
 867         struct lmv_obj *obj;
 868         ENTRY;
 869         rc = lmv_check_connect(obd);
 870         if (rc)
 871                 RETURN(rc);
 872 repeat:
 873         obj = lmv_grab_obj(obd, fid, 0);
 874         if (obj) {
 875                 /* directory is splitted. look for right mds for this name */
 876                 mds = raw_name2idx(obj->objcount, filename, namelen - 1);
 877                 rfid = obj->objs[mds].fid;
 878                 lmv_put_obj(obj);
 879         }
 880         CDEBUG(D_OTHER, "getattr_name for %*s on %lu/%lu/%lu -> %lu/%lu/%lu\n",
 881                namelen, filename, (unsigned long) fid->mds,
 882                (unsigned long) fid->id, (unsigned long) fid->generation,
 883                (unsigned long) rfid.mds, (unsigned long) rfid.id,
 884                (unsigned long) rfid.generation);
 885         rc = md_getattr_name(lmv->tgts[mds].ltd_exp, &rfid, filename, namelen,
 886                                   valid, ea_size, request);
 887         if (rc == 0) {
 888                 /* this could be cross-node reference. in this case all
 889                  * we have right now is mds/ino/generation triple. we'd
 890                  * like to find other attributes */
 891                 body = lustre_msg_buf((*request)->rq_repmsg, 0, sizeof(*body));
 892                 LASSERT(body != NULL);
 893                 if (body->valid & OBD_MD_MDS) {
 894                         struct ptlrpc_request *req = NULL;
 895                         rfid = body->fid1;
 896                         CDEBUG(D_OTHER, "request attrs for %lu/%lu/%lu\n",
 897                                (unsigned long) rfid.mds,
 898                                (unsigned long) rfid.id,
 899                                (unsigned long) rfid.generation);
 900                         rc = md_getattr_name(lmv->tgts[rfid.mds].ltd_exp, &rfid,
 901                                              NULL, 1, valid, ea_size, &req);
 902                         ptlrpc_req_finished(*request);
 903                         *request = req;
 904                 }
 905         } else if (rc == -ERESTART) {
 906                 /* directory got splitted. time to update local object
 907                  * and repeat the request with proper MDS */
 908                 rc = lmv_get_mea_and_update_object(exp, &rfid);
 909                 if (rc == 0) {
 910                         ptlrpc_req_finished(*request);
 911                         goto repeat;
 912                 }
 913         }
 914         RETURN(rc);
 915 }
 916
 917
 918 /*
 919  * llite passes fid of an target inode in data->fid1 and
 920  * fid of directory in data->fid2
 921  */
 922 int lmv_link(struct obd_export *exp, struct mdc_op_data *data,
 923              struct ptlrpc_request **request)
 924 {
 925         struct obd_device *obd = exp->exp_obd;
 926         struct lmv_obd *lmv = &obd->u.lmv;
 927         struct lmv_obj *obj;
 928         int rc;
 929         ENTRY;
 930         rc = lmv_check_connect(obd);
 931         if (rc)
 932                 RETURN(rc);
 933         if (data->namelen != 0) {
 934                 /* usual link request */
 935                 obj = lmv_grab_obj(obd, &data->fid1, 0);
 936                 if (obj) {
 937                         rc = raw_name2idx(obj->objcount, data->name,
 938                                          data->namelen);
 939                         data->fid1 = obj->objs[rc].fid;
 940                         lmv_put_obj(obj);
 941                 }
 942                 CDEBUG(D_OTHER,"link %u/%u/%u:%*s to %u/%u/%u mds %d\n",
 943                        (unsigned) data->fid2.mds, (unsigned) data->fid2.id,
 944                        (unsigned) data->fid2.generation, data->namelen,
 945                        data->name, (unsigned) data->fid1.mds,
 946                        (unsigned) data->fid1.id,
 947                        (unsigned) data->fid1.generation, data->fid1.mds);
 948         } else {
 949                 /* request from MDS to acquire i_links for inode by fid1 */
 950                 CDEBUG(D_OTHER, "inc i_nlinks for %u/%u/%u\n",
 951                        (unsigned) data->fid1.mds, (unsigned) data->fid1.id,
 952                        (unsigned) data->fid1.generation);
 953         }
 954
 955         rc = md_link(lmv->tgts[data->fid1.mds].ltd_exp, data, request);
 956         RETURN(rc);
 957 }
 958
 959 int lmv_rename(struct obd_export *exp, struct mdc_op_data *data,
 960                const char *old, int oldlen, const char *new, int newlen,
 961                struct ptlrpc_request **request)
 962 {
 963         struct obd_device *obd = exp->exp_obd;
 964         struct lmv_obd *lmv = &obd->u.lmv;
 965         struct lmv_obj *obj;
 966         int rc, mds;
 967         ENTRY;
 968
 969         CDEBUG(D_OTHER, "rename %*s in %lu/%lu/%lu to %*s in %lu/%lu/%lu\n",
 970                oldlen, old, (unsigned long) data->fid1.mds,
 971                (unsigned long) data->fid1.id,
 972                (unsigned long) data->fid1.generation,
 973                newlen, new, (unsigned long) data->fid2.mds,
 974                (unsigned long) data->fid2.id,
 975                (unsigned long) data->fid2.generation);
 976         if (!fid_equal(&data->fid1, &data->fid2))
 977                 CWARN("cross-node rename %lu/%lu/%lu:%*s to %lu/%lu/%lu:%*s\n",
 978                       (unsigned long) data->fid1.mds,
 979                       (unsigned long) data->fid1.id,
 980                       (unsigned long) data->fid1.generation, oldlen, old,
 981                       (unsigned long) data->fid2.mds,
 982                       (unsigned long) data->fid2.id,
 983                       (unsigned long) data->fid2.generation, newlen, new);
 984
 985         rc = lmv_check_connect(obd);
 986         if (rc)
 987                 RETURN(rc);
 988
 989         if (oldlen == 0) {
 990                 /* MDS with old dir entry is asking another MDS
 991                  * to create name there */
 992                 CDEBUG(D_OTHER,
 993                        "create %*s(%d/%d) in %lu/%lu/%lu pointing to %lu/%lu/%lu\n",
 994                        newlen, new, oldlen, newlen,
 995                        (unsigned long) data->fid2.mds,
 996                        (unsigned long) data->fid2.id,
 997                        (unsigned long) data->fid2.generation,
 998                        (unsigned long) data->fid1.mds,
 999                        (unsigned long) data->fid1.id,
1000                        (unsigned long) data->fid1.generation);
1001                 mds = data->fid2.mds;
1002                 goto request;
1003         }
1004
1005         obj = lmv_grab_obj(obd, &data->fid1, 0);
1006         if (obj) {
1007                 /* directory is already splitted, so we have to forward
1008                  * request to the right MDS */
1009                 mds = raw_name2idx(obj->objcount, (char *)old, oldlen);
1010                 data->fid1 = obj->objs[mds].fid;
1011                 CDEBUG(D_OTHER, "forward to MDS #%u (%lu/%lu/%lu)\n", mds,
1012                        (unsigned long) obj->objs[mds].fid.mds,
1013                        (unsigned long) obj->objs[mds].fid.id,
1014                        (unsigned long) obj->objs[mds].fid.generation);
1015         }
1016         lmv_put_obj(obj);
1017
1018         obj = lmv_grab_obj(obd, &data->fid2, 0);
1019         if (obj) {
1020                 /* directory is already splitted, so we have to forward
1021                  * request to the right MDS */
1022                 mds = raw_name2idx(obj->objcount, (char *)new, newlen);
1023                 data->fid2 = obj->objs[mds].fid;
1024                 CDEBUG(D_OTHER, "forward to MDS #%u (%lu/%lu/%lu)\n", mds,
1025                        (unsigned long) obj->objs[mds].fid.mds,
1026                        (unsigned long) obj->objs[mds].fid.id,
1027                        (unsigned long) obj->objs[mds].fid.generation);
1028         }
1029         lmv_put_obj(obj);
1030
1031         mds = data->fid1.mds;
1032
1033 request:
1034         rc = md_rename(lmv->tgts[mds].ltd_exp, data, old, oldlen,
1035                        new, newlen, request);
1036         RETURN(rc);
1037 }
1038
1039 int lmv_setattr(struct obd_export *exp, struct mdc_op_data *data,
1040                 struct iattr *iattr, void *ea, int ealen, void *ea2, int ea2len,
1041                 struct ptlrpc_request **request)
1042 {
1043         struct obd_device *obd = exp->exp_obd;
1044         struct lmv_obd *lmv = &obd->u.lmv;
1045         int rc = 0, i = data->fid1.mds;
1046         struct ptlrpc_request *req;
1047         struct mds_body *mds_body;
1048         struct lmv_obj *obj;
1049         ENTRY;
1050
1051         rc = lmv_check_connect(obd);
1052         if (rc)
1053                 RETURN(rc);
1054
1055         obj = lmv_grab_obj(obd, &data->fid1, 0);
1056         CDEBUG(D_OTHER, "SETATTR for %lu/%lu/%lu, valid 0x%x%s\n",
1057                (unsigned long) data->fid1.mds,
1058                (unsigned long) data->fid1.id,
1059                (unsigned long) data->fid1.generation, iattr->ia_valid,
1060                obj ? ", splitted" : "");
1061         if (obj) {
1062                 for (i = 0; i < obj->objcount; i++) {
1063                         data->fid1 = obj->objs[i].fid;
1064                         rc = md_setattr(lmv->tgts[i].ltd_exp, data, iattr, ea,
1065                                         ealen, ea2, ea2len, &req);
1066                         LASSERT(rc == 0);
1067                         if (fid_equal(&obj->fid, &obj->objs[i].fid)) {
1068                                 /* this is master object and this request
1069                                  * should be returned back to llite */
1070                                 *request = req;
1071                         } else {
1072                                 ptlrpc_req_finished(req);
1073                         }
1074                 }
1075                 lmv_put_obj(obj);
1076         } else {
1077                 LASSERT(data->fid1.mds < lmv->desc.ld_tgt_count);
1078                 rc = md_setattr(lmv->tgts[i].ltd_exp, data, iattr, ea, ealen,
1079                                 ea2, ea2len, request);
1080                 if (rc == 0) {
1081                         mds_body = lustre_msg_buf((*request)->rq_repmsg, 0,
1082                                         sizeof(*mds_body));
1083                         LASSERT(mds_body != NULL);
1084                         LASSERT(mds_body->mds == i);
1085                 }
1086         }
1087         RETURN(rc);
1088 }
1089
1090 int lmv_sync(struct obd_export *exp, struct ll_fid *fid,
1091              struct ptlrpc_request **request)
1092 {
1093         struct obd_device *obd = exp->exp_obd;
1094         struct lmv_obd *lmv = &obd->u.lmv;
1095         int rc;
1096         ENTRY;
1097
1098         rc = lmv_check_connect(obd);
1099         if (rc)
1100                 RETURN(rc);
1101
1102         rc = md_sync(lmv->tgts[0].ltd_exp, fid, request);
1103         RETURN(rc);
1104 }
1105
1106 int lmv_dirobj_blocking_ast(struct ldlm_lock *lock,
1107                             struct ldlm_lock_desc *desc, void *data, int flag)
1108 {
1109         struct lustre_handle lockh;
1110         struct lmv_obj *obj;
1111         int rc;
1112         ENTRY;
1113
1114         switch (flag) {
1115         case LDLM_CB_BLOCKING:
1116                 ldlm_lock2handle(lock, &lockh);
1117                 rc = ldlm_cli_cancel(&lockh);
1118                 if (rc < 0) {
1119                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
1120                         RETURN(rc);
1121                 }
1122                 break;
1123         case LDLM_CB_CANCELING:
1124                 /* time to drop cached attrs for dirobj */
1125                 obj = lock->l_ast_data;
1126                 if (!obj)
1127                         break;
1128
1129                 CDEBUG(D_OTHER, "cancel %s on %lu/%lu, master %lu/%lu/%lu\n",
1130                        lock->l_resource->lr_name.name[3] == 1 ?
1131                                 "LOOKUP" : "UPDATE",
1132                        (unsigned long) lock->l_resource->lr_name.name[0],
1133                        (unsigned long) lock->l_resource->lr_name.name[1],
1134                        (unsigned long) obj->fid.mds,
1135                        (unsigned long) obj->fid.id,
1136                        (unsigned long) obj->fid.generation);
1137                 break;
1138         default:
1139                 LBUG();
1140         }
1141         RETURN(0);
1142 }
1143
1144 void lmv_remove_dots(struct page *page)
1145 {
1146         char *kaddr = page_address(page);
1147         unsigned limit = PAGE_CACHE_SIZE;
1148         unsigned offs, rec_len;
1149         struct ext2_dir_entry_2 *p;
1150
1151         for (offs = 0; offs <= limit - EXT2_DIR_REC_LEN(1); offs += rec_len) {
1152                 p = (struct ext2_dir_entry_2 *)(kaddr + offs);
1153                 rec_len = le16_to_cpu(p->rec_len);
1154
1155                 if ((p->name_len == 1 && p->name[0] == '.') ||
1156                     (p->name_len == 2 && p->name[0] == '.' && p->name[1] == '.'))
1157                         p->inode = 0;
1158         }
1159 }
1160
1161 int lmv_readpage(struct obd_export *exp, struct ll_fid *mdc_fid,
1162                  __u64 offset, struct page *page,
1163                  struct ptlrpc_request **request)
1164 {
1165         struct obd_device *obd = exp->exp_obd;
1166         struct lmv_obd *lmv = &obd->u.lmv;
1167         struct ll_fid rfid = *mdc_fid;
1168         struct lmv_obj *obj;
1169         int rc, i;
1170         ENTRY;
1171
1172         rc = lmv_check_connect(obd);
1173         if (rc)
1174                 RETURN(rc);
1175
1176         LASSERT(mdc_fid->mds < lmv->desc.ld_tgt_count);
1177         CDEBUG(D_OTHER, "READPAGE at %llu from %lu/%lu/%lu\n",
1178                offset, (unsigned long) rfid.mds,
1179                (unsigned long) rfid.id,
1180                (unsigned long) rfid.generation);
1181
1182         obj = lmv_grab_obj(obd, mdc_fid, 0);
1183         if (obj) {
1184                 /* find dirobj containing page with requested offset */
1185                 /* FIXME: what about protecting cached attrs here? */
1186                 for (i = 0; i < obj->objcount; i++) {
1187                         if (offset < obj->objs[i].size)
1188                                 break;
1189                         offset -= obj->objs[i].size;
1190                 }
1191                 rfid = obj->objs[i].fid;
1192                 CDEBUG(D_OTHER, "forward to %lu/%lu/%lu with offset %lu\n",
1193                        (unsigned long) rfid.mds,
1194                        (unsigned long) rfid.id,
1195                        (unsigned long) rfid.generation,
1196                        (unsigned long) offset);
1197         }
1198         rc = md_readpage(lmv->tgts[rfid.mds].ltd_exp, &rfid, offset, page, request);
1199         if (rc == 0 && !fid_equal(&rfid, mdc_fid)) {
1200                 /* this page isn't from master object. to avoid
1201                  * ./.. duplication in directory, we have to remove them
1202                  * from all slave objects */
1203                 lmv_remove_dots(page);
1204         }
1205
1206         lmv_put_obj(obj);
1207
1208         RETURN(rc);
1209 }
1210
1211 int lmv_unlink_slaves(struct obd_export *exp, struct mdc_op_data *data,
1212                       struct ptlrpc_request **req)
1213 {
1214         struct obd_device *obd = exp->exp_obd;
1215         struct lmv_obd *lmv = &obd->u.lmv;
1216         struct mea *mea = data->mea1;
1217         struct mdc_op_data data2;
1218         int i, rc = 0, mds;
1219         ENTRY;
1220
1221         LASSERT(mea != NULL);
1222         for (i = 0; i < mea->mea_count; i++) {
1223                 if (lmv->tgts[i].ltd_exp == NULL)
1224                         continue;
1225
1226                 memset(&data2, 0, sizeof(data2));
1227                 data2.fid1 = mea->mea_fids[i];
1228                 data2.create_mode = MDS_MODE_DONT_LOCK | S_IFDIR;
1229                 mds = data2.fid1.mds;
1230                 rc = md_unlink(lmv->tgts[mds].ltd_exp, &data2, req);
1231                 CDEBUG(D_OTHER, "unlink slave %lu/%lu/%lu -> %d\n",
1232                        (unsigned long) mea->mea_fids[i].mds,
1233                        (unsigned long) mea->mea_fids[i].id,
1234                        (unsigned long) mea->mea_fids[i].generation, rc);
1235                 if (*req) {
1236                         ptlrpc_req_finished(*req);
1237                         *req = NULL;
1238                 }
1239                 if (rc)
1240                         break;
1241         }
1242         RETURN(rc);
1243 }
1244
1245 int lmv_unlink(struct obd_export *exp, struct mdc_op_data *data,
1246                struct ptlrpc_request **request)
1247 {
1248         struct obd_device *obd = exp->exp_obd;
1249         struct lmv_obd *lmv = &obd->u.lmv;
1250         int rc, i = 0;
1251         ENTRY;
1252         rc = lmv_check_connect(obd);
1253         if (rc)
1254                 RETURN(rc);
1255
1256         if (data->namelen == 0 && data->mea1 != NULL) {
1257                 /* mds asks to remove slave objects */
1258                 rc = lmv_unlink_slaves(exp, data, request);
1259                 RETURN(rc);
1260         } else if (data->namelen != 0) {
1261                 struct lmv_obj *obj;
1262                 obj = lmv_grab_obj(obd, &data->fid1, 0);
1263                 if (obj) {
1264                         i = raw_name2idx(obj->objcount, data->name,
1265                                          data->namelen);
1266                         data->fid1 = obj->objs[i].fid;
1267                         lmv_put_obj(obj);
1268                 }
1269                 CDEBUG(D_OTHER, "unlink '%*s' in %lu/%lu/%lu -> %u\n",
1270                        data->namelen, data->name,
1271                        (unsigned long) data->fid1.mds,
1272                        (unsigned long) data->fid1.id,
1273                        (unsigned long) data->fid1.generation, i);
1274         } else {
1275                 CDEBUG(D_OTHER, "drop i_nlink on %lu/%lu/%lu\n",
1276                        (unsigned long) data->fid1.mds,
1277                        (unsigned long) data->fid1.id,
1278                        (unsigned long) data->fid1.generation);
1279         }
1280         rc = md_unlink(lmv->tgts[data->fid1.mds].ltd_exp, data, request);
1281         RETURN(rc);
1282 }
1283
1284 struct obd_device *lmv_get_real_obd(struct obd_export *exp,
1285                                     char *name, int len)
1286 {
1287         struct obd_device *obd = exp->exp_obd;
1288         struct lmv_obd *lmv = &obd->u.lmv;
1289         int rc;
1290         ENTRY;
1291
1292         rc = lmv_check_connect(obd);
1293         if (rc)
1294                 RETURN(ERR_PTR(rc));
1295         obd = lmv->tgts[0].ltd_exp->exp_obd;
1296         EXIT;
1297         return obd;
1298 }
1299
1300 int lmv_init_ea_size(struct obd_export *exp, int easize, int cookiesize)
1301 {
1302         struct obd_device *obd = exp->exp_obd;
1303         struct lmv_obd *lmv = &obd->u.lmv;
1304         int i, rc = 0, change = 0;
1305         ENTRY;
1306
1307         if (lmv->max_easize < easize) {
1308                 lmv->max_easize = easize;
1309                 change = 1;
1310         }
1311         if (lmv->max_cookiesize < cookiesize) {
1312                 lmv->max_cookiesize = cookiesize;
1313                 change = 1;
1314         }
1315         if (change == 0)
1316                 RETURN(0);
1317
1318         if (lmv->connected == 0)
1319                 RETURN(0);
1320
1321         /* FIXME: error handling? */
1322         for (i = 0; i < lmv->desc.ld_tgt_count; i++)
1323                 rc = obd_init_ea_size(lmv->tgts[i].ltd_exp, easize, cookiesize);
1324         RETURN(rc);
1325 }
1326
1327 int lmv_obd_create_single(struct obd_export *exp, struct obdo *oa,
1328                           struct lov_stripe_md **ea, struct obd_trans_info *oti)
1329 {
1330         struct obd_device *obd = exp->exp_obd;
1331         struct lmv_obd *lmv = &obd->u.lmv;
1332         struct lov_stripe_md obj_md;
1333         struct lov_stripe_md *obj_mdp = &obj_md;
1334         int rc = 0;
1335         ENTRY;
1336
1337         rc = lmv_check_connect(obd);
1338         if (rc)
1339                 RETURN(rc);
1340
1341         LASSERT(ea == NULL);
1342         LASSERT(oa->o_mds < lmv->desc.ld_tgt_count);
1343
1344         rc = obd_create(lmv->tgts[oa->o_mds].ltd_exp, oa, &obj_mdp, oti);
1345         LASSERT(rc == 0);
1346
1347         RETURN(rc);
1348 }
1349
1350 /*
1351  * to be called from MDS only
1352  */
1353 int lmv_obd_create(struct obd_export *exp, struct obdo *oa,
1354                    struct lov_stripe_md **ea, struct obd_trans_info *oti)
1355 {
1356         struct obd_device *obd = exp->exp_obd;
1357         struct lmv_obd *lmv = &obd->u.lmv;
1358         struct mea *mea;
1359         int i, c, rc = 0;
1360         struct ll_fid mfid;
1361         ENTRY;
1362
1363         rc = lmv_check_connect(obd);
1364         if (rc)
1365                 RETURN(rc);
1366
1367         LASSERT(oa != NULL);
1368
1369         if (ea == NULL) {
1370                 rc = lmv_obd_create_single(exp, oa, NULL, oti);
1371                 RETURN(rc);
1372         }
1373
1374         if (*ea == NULL) {
1375                 rc = obd_alloc_diskmd(exp, (struct lov_mds_md **)ea);
1376                 LASSERT(*ea != NULL);
1377         }
1378
1379         mea = (struct mea *)*ea;
1380         mfid.id = oa->o_id;
1381         mfid.generation = oa->o_generation;
1382         rc = 0;
1383         if (!mea->mea_count || mea->mea_count > lmv->desc.ld_tgt_count)
1384                 mea->mea_count = lmv->desc.ld_tgt_count;
1385
1386         mea->mea_master = -1;
1387
1388         /* FIXME: error handling? */
1389         for (i = 0, c = 0; c < mea->mea_count &&
1390                 i < lmv->desc.ld_tgt_count; i++) {
1391                 struct lov_stripe_md obj_md;
1392                 struct lov_stripe_md *obj_mdp = &obj_md;
1393
1394                 if (lmv->tgts[i].ltd_exp == NULL) {
1395                         /* this is master MDS */
1396                         mea->mea_fids[c].id = mfid.id;
1397                         mea->mea_fids[c].generation = mfid.generation;
1398                         mea->mea_fids[c].mds = i;
1399                         mea->mea_master = i;
1400                         c++;
1401                         continue;
1402                 }
1403
1404                 /* "Master" MDS should always be part of stripped dir, so
1405                    scan for it */
1406                 if (mea->mea_master == -1 && c == mea->mea_count - 1)
1407                         continue;
1408
1409                 oa->o_valid = OBD_MD_FLGENER | OBD_MD_FLTYPE | OBD_MD_FLMODE
1410                                 | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLID;
1411
1412                 rc = obd_create(lmv->tgts[c].ltd_exp, oa, &obj_mdp, oti);
1413                 /* FIXME: error handling here */
1414                 LASSERT(rc == 0);
1415
1416                 mea->mea_fids[c].id = oa->o_id;
1417                 mea->mea_fids[c].generation = oa->o_generation;
1418                 mea->mea_fids[c].mds = i;
1419                 c++;
1420                 CDEBUG(D_OTHER, "dirobj at mds %d: "LPU64"/%u\n",
1421                        i, oa->o_id, oa->o_generation);
1422         }
1423         LASSERT(c == mea->mea_count);
1424         CDEBUG(D_OTHER, "%d dirobjects created\n", (int) mea->mea_count);
1425
1426         RETURN(rc);
1427 }
1428
1429 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
1430                         void *key, __u32 *vallen, void *val)
1431 {
1432         struct obd_device *obd;
1433         struct lmv_obd *lmv;
1434         ENTRY;
1435
1436         obd = class_exp2obd(exp);
1437         if (obd == NULL) {
1438                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1439                        exp->exp_handle.h_cookie);
1440                 RETURN(-EINVAL);
1441         }
1442
1443         lmv = &obd->u.lmv;
1444         if (keylen == 6 && memcmp(key, "mdsize", 6) == 0) {
1445                 __u32 *mdsize = val;
1446                 *vallen = sizeof(__u32);
1447                 *mdsize = sizeof(struct ll_fid) * lmv->desc.ld_tgt_count
1448                                 + sizeof(struct mea);
1449                 RETURN(0);
1450         } else if (keylen == 6 && memcmp(key, "mdsnum", 6) == 0) {
1451                 struct obd_uuid *cluuid = &lmv->cluuid;
1452                 struct lmv_tgt_desc *tgts;
1453                 __u32 *mdsnum = val;
1454                 int i;
1455
1456                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
1457                         if (obd_uuid_equals(&tgts->uuid, cluuid)) {
1458                                 *vallen = sizeof(__u32);
1459                                 *mdsnum = i;
1460                                 RETURN(0);
1461                         }
1462                 }
1463                 LASSERT(0);
1464         }
1465
1466         CDEBUG(D_IOCTL, "invalid key\n");
1467         RETURN(-EINVAL);
1468 }
1469
1470 int lmv_set_info(struct obd_export *exp, obd_count keylen,
1471                  void *key, obd_count vallen, void *val)
1472 {
1473         struct obd_device *obd;
1474         struct lmv_obd *lmv;
1475         ENTRY;
1476
1477         obd = class_exp2obd(exp);
1478         if (obd == NULL) {
1479                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1480                        exp->exp_handle.h_cookie);
1481                 RETURN(-EINVAL);
1482         }
1483         lmv = &obd->u.lmv;
1484
1485         if (keylen >= strlen("client") && strcmp(key, "client") == 0) {
1486                 struct lmv_tgt_desc *tgts;
1487                 int i, rc;
1488
1489                 rc = lmv_check_connect(obd);
1490                 if (rc)
1491                         RETURN(rc);
1492
1493                 for (i = 0, tgts = lmv->tgts;
1494                         i < lmv->desc.ld_tgt_count; i++, tgts++) {
1495                         rc = obd_set_info(tgts->ltd_exp, keylen, key, vallen, val);
1496                         if (rc)
1497                                 RETURN(rc);
1498                 }
1499                 RETURN(0);
1500         } else if (keylen >= strlen("inter_mds") && strcmp(key, "inter_mds") == 0) {
1501                 lmv->server_timeout = 1;
1502                 lmv_set_timeouts(obd);
1503                 RETURN(0);
1504         }
1505
1506         RETURN(-EINVAL);
1507 }
1508
1509 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
1510                struct lov_stripe_md *lsm)
1511 {
1512         struct obd_device *obd = class_exp2obd(exp);
1513         struct lmv_obd *lmv = &obd->u.lmv;
1514         int mea_size;
1515         ENTRY;
1516
1517         mea_size = sizeof(struct ll_fid) *
1518                 lmv->desc.ld_tgt_count + sizeof(struct mea);
1519         if (!lmmp)
1520                 RETURN(mea_size);
1521
1522         if (*lmmp && !lsm) {
1523                 OBD_FREE(*lmmp, mea_size);
1524                 *lmmp = NULL;
1525                 RETURN(0);
1526         }
1527
1528         if (!*lmmp) {
1529                 OBD_ALLOC(*lmmp, mea_size);
1530                 if (!*lmmp)
1531                         RETURN(-ENOMEM);
1532         }
1533
1534         if (!lsm)
1535                 RETURN(mea_size);
1536
1537 #warning "MEA packing/convertation must be here! -bzzz"
1538         memcpy(*lmmp, lsm, mea_size);
1539         RETURN(mea_size);
1540 }
1541
1542 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **mem_tgt,
1543                         struct lov_mds_md *disk_src, int mdsize)
1544 {
1545         struct obd_device *obd = class_exp2obd(exp);
1546         struct lmv_obd *lmv = &obd->u.lmv;
1547         struct mea **tmea = (struct mea **) mem_tgt;
1548         struct mea *mea = (void *) disk_src;
1549         int mea_size;
1550         ENTRY;
1551
1552         mea_size = sizeof(struct ll_fid) *
1553                 lmv->desc.ld_tgt_count + sizeof(struct mea);
1554         if (mem_tgt == NULL)
1555                 return mea_size;
1556
1557         if (*mem_tgt != NULL && disk_src == NULL) {
1558                 OBD_FREE(*tmea, mea_size);
1559                 RETURN(0);
1560         }
1561
1562         LASSERT(mea_size == mdsize);
1563
1564         OBD_ALLOC(*tmea, mea_size);
1565         /* FIXME: error handling here */
1566         LASSERT(*tmea != NULL);
1567
1568         if (!disk_src)
1569                 RETURN(mea_size);
1570
1571 #warning "MEA unpacking/convertation must be here! -bzzz"
1572         memcpy(*tmea, mea, mdsize);
1573         RETURN(mea_size);
1574 }
1575
1576 int lmv_brw(int rw, struct obd_export *exp, struct obdo *oa,
1577                 struct lov_stripe_md *ea, obd_count oa_bufs,
1578                 struct brw_page *pgarr, struct obd_trans_info *oti)
1579 {
1580         struct obd_device *obd = exp->exp_obd;
1581         struct lmv_obd *lmv = &obd->u.lmv;
1582         struct mea *mea = (struct mea *) ea;
1583         int err;
1584
1585         LASSERT(oa != NULL);
1586         LASSERT(ea != NULL);
1587         LASSERT(pgarr != NULL);
1588         LASSERT(oa->o_mds < lmv->desc.ld_tgt_count);
1589
1590         oa->o_gr = mea->mea_fids[oa->o_mds].generation;
1591         oa->o_id = mea->mea_fids[oa->o_mds].id;
1592         oa->o_valid =  OBD_MD_FLID | OBD_MD_FLGROUP;
1593         err = obd_brw(rw, lmv->tgts[oa->o_mds].ltd_exp, oa,
1594                       NULL, oa_bufs, pgarr, oti);
1595         RETURN(err);
1596 }
1597
1598 struct obd_ops lmv_obd_ops = {
1599         .o_owner                = THIS_MODULE,
1600         .o_attach               = lmv_attach,
1601         .o_detach               = lmv_detach,
1602         .o_setup                = lmv_setup,
1603         .o_cleanup              = lmv_cleanup,
1604         .o_connect              = lmv_connect,
1605         .o_disconnect           = lmv_disconnect,
1606         .o_statfs               = lmv_statfs,
1607         .o_get_info             = lmv_get_info,
1608         .o_set_info             = lmv_set_info,
1609         .o_create               = lmv_obd_create,
1610         .o_packmd               = lmv_packmd,
1611         .o_unpackmd             = lmv_unpackmd,
1612         .o_brw                  = lmv_brw,
1613         .o_init_ea_size         = lmv_init_ea_size,
1614         .o_notify               = lmv_notify,
1615         .o_iocontrol            = lmv_iocontrol,
1616 };
1617
1618 struct md_ops lmv_md_ops = {
1619         .m_getstatus            = lmv_getstatus,
1620         .m_getattr              = lmv_getattr,
1621         .m_change_cbdata        = lmv_change_cbdata,
1622         .m_change_cbdata_name   = lmv_change_cbdata_name,
1623         .m_close                = lmv_close,
1624         .m_create               = lmv_create,
1625         .m_done_writing         = lmv_done_writing,
1626         .m_enqueue              = lmv_enqueue,
1627         .m_getattr_name         = lmv_getattr_name,
1628         .m_intent_lock          = lmv_intent_lock,
1629         .m_link                 = lmv_link,
1630         .m_rename               = lmv_rename,
1631         .m_setattr              = lmv_setattr,
1632         .m_sync                 = lmv_sync,
1633         .m_readpage             = lmv_readpage,
1634         .m_unlink               = lmv_unlink,
1635         .m_get_real_obd         = lmv_get_real_obd,
1636         .m_valid_attrs          = lmv_valid_attrs,
1637 };
1638
1639 int __init lmv_init(void)
1640 {
1641         struct lprocfs_static_vars lvars;
1642         int rc;
1643
1644         lprocfs_init_vars(lmv, &lvars);
1645         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
1646                                  lvars.module_vars, OBD_LMV_DEVICENAME);
1647         RETURN(rc);
1648 }
1649
1650 #ifdef __KERNEL__
1651 static void lmv_exit(void)
1652 {
1653         class_unregister_type(OBD_LMV_DEVICENAME);
1654 }
1655
1656 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1657 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
1658 MODULE_LICENSE("GPL");
1659
1660 module_init(lmv_init);
1661 module_exit(lmv_exit);
1662 #endif