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                                "error %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         int i, rc = 0;
 420         struct lmv_desc *desc;
 421         struct obd_uuid *uuids;
 422         struct lmv_tgt_desc *tgts;
 423         struct lustre_cfg *lcfg = buf;
 424         struct lmv_obd *lmv = &obd->u.lmv;
 425         ENTRY;
 426
 427         if (lcfg->lcfg_inllen1 < 1) {
 428                 CERROR("LMV setup requires a descriptor\n");
 429                 RETURN(-EINVAL);
 430         }
 431
 432         if (lcfg->lcfg_inllen2 < 1) {
 433                 CERROR("LMV setup requires an OST UUID list\n");
 434                 RETURN(-EINVAL);
 435         }
 436
 437         desc = (struct lmv_desc *)lcfg->lcfg_inlbuf1;
 438         if (sizeof(*desc) > lcfg->lcfg_inllen1) {
 439                 CERROR("descriptor size wrong: %d > %d\n",
 440                        (int)sizeof(*desc), lcfg->lcfg_inllen1);
 441                 RETURN(-EINVAL);
 442         }
 443
 444         uuids = (struct obd_uuid *)lcfg->lcfg_inlbuf2;
 445         if (sizeof(*uuids) * desc->ld_tgt_count != lcfg->lcfg_inllen2) {
 446                 CERROR("UUID array size wrong: %u * %u != %u\n",
 447                        sizeof(*uuids), desc->ld_tgt_count, lcfg->lcfg_inllen2);
 448                 RETURN(-EINVAL);
 449         }
 450
 451         lmv->bufsize = sizeof(struct lmv_tgt_desc) * desc->ld_tgt_count;
 452         OBD_ALLOC(lmv->tgts, lmv->bufsize);
 453         if (lmv->tgts == NULL) {
 454                 CERROR("Out of memory\n");
 455                 RETURN(-ENOMEM);
 456         }
 457
 458         lmv->desc = *desc;
 459         spin_lock_init(&lmv->lmv_lock);
 460
 461         for (i = 0, tgts = lmv->tgts; i < desc->ld_tgt_count; i++, tgts++)
 462                 tgts->uuid = uuids[i];
 463
 464         lmv->max_cookiesize = 0;
 465
 466         lmv->max_easize = sizeof(struct ll_fid) *
 467                 desc->ld_tgt_count + sizeof(struct mea);
 468
 469         rc = lmv_setup_mgr(obd);
 470         if (rc) {
 471                 CERROR("Can't setup LMV object manager, "
 472                        "error %d.\n", rc);
 473                 OBD_FREE(lmv->tgts, lmv->bufsize);
 474         }
 475
 476         RETURN(rc);
 477 }
 478
 479 static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs,
 480                       unsigned long max_age)
 481 {
 482         struct lmv_obd *lmv = &obd->u.lmv;
 483         struct obd_statfs temp;
 484         int rc = 0, i;
 485         ENTRY;
 486
 487         rc = lmv_check_connect(obd);
 488         if (rc)
 489                 RETURN(rc);
 490
 491         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
 492                 rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, &temp, max_age);
 493                 if (rc) {
 494                         CERROR("can't stat MDS #%d (%s)\n", i,
 495                                lmv->tgts[i].ltd_exp->exp_obd->obd_name);
 496                         RETURN(rc);
 497                 }
 498                 if (i == 0) {
 499                         memcpy(osfs, &temp, sizeof(temp));
 500                 } else {
 501                         osfs->os_bavail += temp.os_bavail;
 502                         osfs->os_blocks += temp.os_blocks;
 503                         osfs->os_ffree += temp.os_ffree;
 504                         osfs->os_files += temp.os_files;
 505                 }
 506         }
 507         RETURN(rc);
 508 }
 509
 510 static int lmv_cleanup(struct obd_device *obd, int flags)
 511 {
 512         struct lmv_obd *lmv = &obd->u.lmv;
 513         ENTRY;
 514         lmv_cleanup_mgr(obd);
 515         OBD_FREE(lmv->tgts, lmv->bufsize);
 516         RETURN(0);
 517 }
 518
 519 static int lmv_getstatus(struct obd_export *exp, struct ll_fid *fid)
 520 {
 521         struct obd_device *obd = exp->exp_obd;
 522         struct lmv_obd *lmv = &obd->u.lmv;
 523         int rc;
 524         ENTRY;
 525         rc = lmv_check_connect(obd);
 526         if (rc)
 527                 RETURN(rc);
 528         rc = md_getstatus(lmv->tgts[0].ltd_exp, fid);
 529         fid->mds = 0;
 530         RETURN(rc);
 531 }
 532
 533 static int lmv_getattr(struct obd_export *exp, struct ll_fid *fid,
 534                        unsigned long valid, unsigned int ea_size,
 535                        struct ptlrpc_request **request)
 536 {
 537         struct obd_device *obd = exp->exp_obd;
 538         struct lmv_obd *lmv = &obd->u.lmv;
 539         int rc, i = fid->mds;
 540         struct lmv_obj *obj;
 541         ENTRY;
 542
 543         rc = lmv_check_connect(obd);
 544         if (rc)
 545                 RETURN(rc);
 546
 547         LASSERT(i < lmv->desc.ld_tgt_count);
 548
 549         rc = md_getattr(lmv->tgts[i].ltd_exp, fid, valid,
 550                         ea_size, request);
 551         if (rc)
 552                 RETURN(rc);
 553
 554         obj = lmv_grab_obj(obd, fid);
 555
 556         CDEBUG(D_OTHER, "GETATTR for %lu/%lu/%lu %s\n",
 557                (unsigned long)fid->mds, (unsigned long)fid->id,
 558                (unsigned long)fid->generation, obj ? "(splitted)" : "");
 559
 560         /* if object is splitted, then we loop over all the slaves and gather
 561          * size attribute. In ideal world we would have to gather also mds field
 562          * from all slaves, as object is spread over the cluster and this is
 563          * definitely interesting information and it is not good to loss it,
 564          * but...*/
 565         if (obj) {
 566                 struct mds_body *body;
 567
 568                 if (*request == NULL) {
 569                         lmv_put_obj(obj);
 570                         RETURN(rc);
 571                 }
 572
 573                 body = lustre_msg_buf((*request)->rq_repmsg, 0,
 574                                       sizeof(*body));
 575                 LASSERT(body != NULL);
 576
 577                 lmv_lock_obj(obj);
 578
 579                 for (i = 0; i < obj->objcount; i++) {
 580
 581                         /* skip master obj. */
 582                         if (fid_equal(&obj->fid, &obj->objs[i].fid))
 583                                 continue;
 584
 585                         body->size += obj->objs[i].size;
 586                 }
 587
 588                 lmv_unlock_obj(obj);
 589                 lmv_put_obj(obj);
 590         }
 591
 592         RETURN(rc);
 593 }
 594
 595 static int lmv_change_cbdata(struct obd_export *exp, struct ll_fid *fid,
 596                              ldlm_iterator_t it, void *data)
 597 {
 598         struct obd_device *obd = exp->exp_obd;
 599         struct lmv_obd *lmv = &obd->u.lmv;
 600         int rc = 0;
 601         ENTRY;
 602
 603         rc = lmv_check_connect(obd);
 604         if (rc)
 605                 RETURN(rc);
 606
 607         CDEBUG(D_OTHER, "CBDATA for %lu/%lu/%lu\n", (unsigned long)fid->mds,
 608                (unsigned long)fid->id, (unsigned long)fid->generation);
 609
 610         LASSERT(fid->mds < lmv->desc.ld_tgt_count);
 611
 612         rc = md_change_cbdata(lmv->tgts[fid->mds].ltd_exp,
 613                               fid, it, data);
 614
 615         RETURN(rc);
 616 }
 617
 618 static int lmv_change_cbdata_name(struct obd_export *exp, struct ll_fid *pfid,
 619                                   char *name, int len, struct ll_fid *cfid,
 620                                   ldlm_iterator_t it, void *data)
 621 {
 622         struct obd_device *obd = exp->exp_obd;
 623         struct lmv_obd *lmv = &obd->u.lmv;
 624         struct lmv_obj *obj;
 625         int rc = 0, mds;
 626         ENTRY;
 627
 628         rc = lmv_check_connect(obd);
 629         if (rc)
 630                 RETURN(rc);
 631
 632         LASSERT(pfid->mds < lmv->desc.ld_tgt_count);
 633         LASSERT(cfid->mds < lmv->desc.ld_tgt_count);
 634
 635         CDEBUG(D_OTHER, "CBDATA for %lu/%lu/%lu:%*s -> %lu/%lu/%lu\n",
 636                (unsigned long)pfid->mds, (unsigned long)pfid->id,
 637                (unsigned long)pfid->generation, len, name,
 638                (unsigned long)cfid->mds, (unsigned long)cfid->id,
 639                (unsigned long)cfid->generation);
 640
 641         /* this is default mds for directory name belongs to. */
 642         mds = pfid->mds;
 643         obj = lmv_grab_obj(obd, pfid);
 644         if (obj) {
 645                 /* directory is splitted. look for right mds for this name. */
 646                 mds = raw_name2idx(obj->objcount, name, len);
 647                 lmv_put_obj(obj);
 648         }
 649         rc = md_change_cbdata(lmv->tgts[mds].ltd_exp, cfid, it, data);
 650         RETURN(rc);
 651 }
 652
 653 static int lmv_valid_attrs(struct obd_export *exp, struct ll_fid *fid)
 654 {
 655         struct obd_device *obd = exp->exp_obd;
 656         struct lmv_obd *lmv = &obd->u.lmv;
 657         int rc = 0;
 658         ENTRY;
 659         rc = lmv_check_connect(obd);
 660         if (rc)
 661                 RETURN(rc);
 662         CDEBUG(D_OTHER, "validate %lu/%lu/%lu\n", (unsigned long) fid->mds,
 663                (unsigned long) fid->id, (unsigned long) fid->generation);
 664         LASSERT(fid->mds < lmv->desc.ld_tgt_count);
 665         rc = md_valid_attrs(lmv->tgts[fid->mds].ltd_exp, fid);
 666         RETURN(rc);
 667 }
 668
 669 int lmv_close(struct obd_export *exp, struct obdo *obdo,
 670                   struct obd_client_handle *och,
 671                   struct ptlrpc_request **request)
 672 {
 673         struct obd_device *obd = exp->exp_obd;
 674         struct lmv_obd *lmv = &obd->u.lmv;
 675         int rc, i = obdo->o_mds;
 676         ENTRY;
 677         rc = lmv_check_connect(obd);
 678         if (rc)
 679                 RETURN(rc);
 680         LASSERT(i < lmv->desc.ld_tgt_count);
 681         CDEBUG(D_OTHER, "CLOSE %lu/%lu/%lu\n", (unsigned long) obdo->o_mds,
 682                (unsigned long) obdo->o_id, (unsigned long) obdo->o_generation);
 683         rc = md_close(lmv->tgts[i].ltd_exp, obdo, och, request);
 684         RETURN(rc);
 685 }
 686
 687 int lmv_get_mea_and_update_object(struct obd_export *exp, struct ll_fid *fid)
 688 {
 689         struct obd_device *obd = exp->exp_obd;
 690         struct lmv_obd *lmv = &obd->u.lmv;
 691         struct ptlrpc_request *req = NULL;
 692         struct lmv_obj *obj;
 693         struct lustre_md md;
 694         unsigned long valid;
 695         int mealen, rc;
 696
 697         md.mea = NULL;
 698         mealen = MEA_SIZE_LMV(lmv);
 699
 700         valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA;
 701
 702         /* time to update mea of parent fid */
 703         rc = md_getattr(lmv->tgts[fid->mds].ltd_exp, fid,
 704                         valid, mealen, &req);
 705         if (rc) {
 706                 CERROR("md_getattr() failed, error %d\n", rc);
 707                 GOTO(cleanup, rc);
 708         }
 709
 710         rc = mdc_req2lustre_md(exp, req, 0, NULL, &md);
 711         if (rc) {
 712                 CERROR("mdc_req2lustre_md() failed, error %d\n", rc);
 713                 GOTO(cleanup, rc);
 714         }
 715
 716         if (md.mea == NULL)
 717                 GOTO(cleanup, rc = -ENODATA);
 718
 719         obj = lmv_create_obj(exp, fid, md.mea);
 720         if (IS_ERR(obj))
 721                 rc = PTR_ERR(obj);
 722
 723         obd_free_memmd(exp, (struct lov_stripe_md **)&md.mea);
 724
 725 cleanup:
 726         if (req)
 727                 ptlrpc_req_finished(req);
 728         RETURN(rc);
 729 }
 730
 731 int lmv_create(struct obd_export *exp, struct mdc_op_data *op_data,
 732                const void *data, int datalen, int mode, __u32 uid,
 733                __u32 gid, __u64 rdev, struct ptlrpc_request **request)
 734 {
 735         struct obd_device *obd = exp->exp_obd;
 736         struct lmv_obd *lmv = &obd->u.lmv;
 737         struct mds_body *body;
 738         struct lmv_obj *obj;
 739         int rc, mds;
 740         ENTRY;
 741
 742         rc = lmv_check_connect(obd);
 743         if (rc)
 744                 RETURN(rc);
 745
 746         if (!lmv->desc.ld_active_tgt_count)
 747                 RETURN(-EIO);
 748 repeat:
 749         obj = lmv_grab_obj(obd, &op_data->fid1);
 750         if (obj) {
 751                 mds = raw_name2idx(obj->objcount, op_data->name,
 752                                    op_data->namelen);
 753                 op_data->fid1 = obj->objs[mds].fid;
 754                 lmv_put_obj(obj);
 755         }
 756
 757         CDEBUG(D_OTHER, "CREATE '%*s' on %lu/%lu/%lu\n", op_data->namelen,
 758                op_data->name, (unsigned long)op_data->fid1.mds,
 759                (unsigned long)op_data->fid1.id,
 760                (unsigned long)op_data->fid1.generation);
 761
 762         rc = md_create(lmv->tgts[op_data->fid1.mds].ltd_exp, op_data, data,
 763                        datalen, mode, uid, gid, rdev, request);
 764         if (rc == 0) {
 765                 if (*request == NULL)
 766                         RETURN(rc);
 767
 768                 body = lustre_msg_buf((*request)->rq_repmsg, 0,
 769                                       sizeof(*body));
 770                 LASSERT(body != NULL);
 771
 772                 CDEBUG(D_OTHER, "created. id = %lu, generation = %lu, "
 773                        "mds = %d\n", (unsigned long)body->fid1.id,
 774                        (unsigned long)body->fid1.generation, op_data->fid1.mds);
 775
 776                 LASSERT(body->valid & OBD_MD_MDS ||
 777                         body->mds == op_data->fid1.mds);
 778         } else if (rc == -ERESTART) {
 779                 /* directory got splitted. time to update local object and
 780                  * repeat the request with proper MDS */
 781                 rc = lmv_get_mea_and_update_object(exp, &op_data->fid1);
 782                 if (rc == 0) {
 783                         ptlrpc_req_finished(*request);
 784                         goto repeat;
 785                 }
 786         }
 787         RETURN(rc);
 788 }
 789
 790 int lmv_done_writing(struct obd_export *exp, struct obdo *obdo)
 791 {
 792         struct obd_device *obd = exp->exp_obd;
 793         struct lmv_obd *lmv = &obd->u.lmv;
 794         int rc;
 795         ENTRY;
 796         rc = lmv_check_connect(obd);
 797         if (rc)
 798                 RETURN(rc);
 799
 800         /* FIXME: choose right MDC here */
 801         CWARN("this method isn't implemented yet\n");
 802         rc = md_done_writing(lmv->tgts[0].ltd_exp, obdo);
 803         RETURN(rc);
 804 }
 805
 806 int lmv_enqueue_slaves(struct obd_export *exp, int locktype,
 807                        struct lookup_intent *it, int lockmode,
 808                        struct mdc_op_data *data, struct lustre_handle *lockh,
 809                        void *lmm, int lmmsize, ldlm_completion_callback cb_completion,
 810                        ldlm_blocking_callback cb_blocking, void *cb_data)
 811 {
 812         struct obd_device *obd = exp->exp_obd;
 813         struct lmv_obd *lmv = &obd->u.lmv;
 814         struct mea *mea = data->mea1;
 815         struct mdc_op_data data2;
 816         int i, rc, mds;
 817         ENTRY;
 818
 819         LASSERT(mea != NULL);
 820         for (i = 0; i < mea->mea_count; i++) {
 821                 if (lmv->tgts[i].ltd_exp == NULL)
 822                         continue;
 823
 824                 memset(&data2, 0, sizeof(data2));
 825                 data2.fid1 = mea->mea_fids[i];
 826                 mds = data2.fid1.mds;
 827
 828                 rc = md_enqueue(lmv->tgts[mds].ltd_exp, locktype, it, lockmode,
 829                                 &data2, lockh + i, lmm, lmmsize, cb_completion,
 830                                 cb_blocking, cb_data);
 831
 832                 CDEBUG(D_OTHER, "take lock on slave %lu/%lu/%lu -> %d/%d\n",
 833                        (unsigned long)mea->mea_fids[i].mds,
 834                        (unsigned long)mea->mea_fids[i].id,
 835                        (unsigned long)mea->mea_fids[i].generation,
 836                        rc, it->d.lustre.it_status);
 837                 if (rc)
 838                         GOTO(cleanup, rc);
 839                 if (it->d.lustre.it_data) {
 840                         struct ptlrpc_request *req;
 841                         req = (struct ptlrpc_request *) it->d.lustre.it_data;
 842                         ptlrpc_req_finished(req);
 843                 }
 844
 845                 if (it->d.lustre.it_status)
 846                         GOTO(cleanup, rc = it->d.lustre.it_status);
 847         }
 848         RETURN(0);
 849
 850 cleanup:
 851         /* drop all taken locks */
 852         while (--i >= 0) {
 853                 if (lockh[i].cookie)
 854                         ldlm_lock_decref(lockh + i, lockmode);
 855                 lockh[i].cookie = 0;
 856         }
 857         RETURN(rc);
 858 }
 859
 860 int lmv_enqueue(struct obd_export *exp, int lock_type,
 861                 struct lookup_intent *it, int lock_mode,
 862                 struct mdc_op_data *data, struct lustre_handle *lockh,
 863                 void *lmm, int lmmsize, ldlm_completion_callback cb_completion,
 864                 ldlm_blocking_callback cb_blocking, void *cb_data)
 865 {
 866         struct obd_device *obd = exp->exp_obd;
 867         struct lmv_obd *lmv = &obd->u.lmv;
 868         struct lmv_obj *obj;
 869         int rc, mds;
 870         ENTRY;
 871
 872         rc = lmv_check_connect(obd);
 873         if (rc)
 874                 RETURN(rc);
 875
 876         if (it->it_op == IT_UNLINK) {
 877                 rc = lmv_enqueue_slaves(exp, lock_type, it, lock_mode,
 878                                         data, lockh, lmm, lmmsize,
 879                                         cb_completion, cb_blocking, cb_data);
 880                 RETURN(rc);
 881         }
 882
 883         if (data->namelen) {
 884                 obj = lmv_grab_obj(obd, &data->fid1);
 885                 if (obj) {
 886                         /* directory is splitted. look for right mds for this
 887                          * name */
 888                         mds = raw_name2idx(obj->objcount, (char *)data->name,
 889                                            data->namelen);
 890                         data->fid1 = obj->objs[mds].fid;
 891                         lmv_put_obj(obj);
 892                 }
 893         }
 894         CDEBUG(D_OTHER, "ENQUEUE '%s' on %lu/%lu\n", LL_IT2STR(it),
 895                (unsigned long)data->fid1.id, (unsigned long)data->fid1.generation);
 896
 897         rc = md_enqueue(lmv->tgts[data->fid1.mds].ltd_exp, lock_type, it,
 898                         lock_mode, data, lockh, lmm, lmmsize, cb_completion,
 899                         cb_blocking, cb_data);
 900
 901         RETURN(rc);
 902 }
 903
 904 int lmv_getattr_name(struct obd_export *exp, struct ll_fid *fid,
 905                      char *filename, int namelen, unsigned long valid,
 906                      unsigned int ea_size, struct ptlrpc_request **request)
 907 {
 908         struct obd_device *obd = exp->exp_obd;
 909         struct lmv_obd *lmv = &obd->u.lmv;
 910         struct ll_fid rfid = *fid;
 911         int rc, mds = fid->mds;
 912         struct mds_body *body;
 913         struct lmv_obj *obj;
 914         ENTRY;
 915         rc = lmv_check_connect(obd);
 916         if (rc)
 917                 RETURN(rc);
 918 repeat:
 919         obj = lmv_grab_obj(obd, fid);
 920         if (obj) {
 921                 /* directory is splitted. look for right mds for this name */
 922                 mds = raw_name2idx(obj->objcount, filename, namelen - 1);
 923                 rfid = obj->objs[mds].fid;
 924                 lmv_put_obj(obj);
 925         }
 926
 927         CDEBUG(D_OTHER, "getattr_name for %*s on %lu/%lu/%lu -> %lu/%lu/%lu\n",
 928                namelen, filename, (unsigned long)fid->mds,
 929                (unsigned long)fid->id, (unsigned long)fid->generation,
 930                (unsigned long)rfid.mds, (unsigned long)rfid.id,
 931                (unsigned long)rfid.generation);
 932
 933         rc = md_getattr_name(lmv->tgts[mds].ltd_exp, &rfid, filename,
 934                              namelen, valid, ea_size, request);
 935         if (rc == 0) {
 936                 /* this could be cross-node reference. in this case all we have
 937                  * right now is mds/ino/generation triple. we'd like to find
 938                  * other attributes */
 939                 body = lustre_msg_buf((*request)->rq_repmsg, 0, sizeof(*body));
 940                 LASSERT(body != NULL);
 941                 if (body->valid & OBD_MD_MDS) {
 942                         struct ptlrpc_request *req = NULL;
 943                         rfid = body->fid1;
 944                         CDEBUG(D_OTHER, "request attrs for %lu/%lu/%lu\n",
 945                                (unsigned long) rfid.mds,
 946                                (unsigned long) rfid.id,
 947                                (unsigned long) rfid.generation);
 948                         rc = md_getattr_name(lmv->tgts[rfid.mds].ltd_exp, &rfid,
 949                                              NULL, 1, valid, ea_size, &req);
 950                         ptlrpc_req_finished(*request);
 951                         *request = req;
 952                 }
 953         } else if (rc == -ERESTART) {
 954                 /* directory got splitted. time to update local object and
 955                  * repeat the request with proper MDS */
 956                 rc = lmv_get_mea_and_update_object(exp, &rfid);
 957                 if (rc == 0) {
 958                         ptlrpc_req_finished(*request);
 959                         goto repeat;
 960                 }
 961         }
 962         RETURN(rc);
 963 }
 964
 965
 966 /*
 967  * llite passes fid of an target inode in data->fid1 and fid of directory in
 968  * data->fid2
 969  */
 970 int lmv_link(struct obd_export *exp, struct mdc_op_data *data,
 971              struct ptlrpc_request **request)
 972 {
 973         struct obd_device *obd = exp->exp_obd;
 974         struct lmv_obd *lmv = &obd->u.lmv;
 975         struct lmv_obj *obj;
 976         int rc;
 977         ENTRY;
 978
 979         rc = lmv_check_connect(obd);
 980         if (rc)
 981                 RETURN(rc);
 982
 983         if (data->namelen != 0) {
 984                 /* usual link request */
 985                 obj = lmv_grab_obj(obd, &data->fid1);
 986                 if (obj) {
 987                         rc = raw_name2idx(obj->objcount, data->name,
 988                                           data->namelen);
 989                         data->fid1 = obj->objs[rc].fid;
 990                         lmv_put_obj(obj);
 991                 }
 992
 993                 CDEBUG(D_OTHER,"link %lu/%lu/%lu:%*s to %lu/%lu/%lu mds %lu\n",
 994                        (unsigned long)data->fid2.mds,
 995                        (unsigned long)data->fid2.id,
 996                        (unsigned long)data->fid2.generation,
 997                        data->namelen, data->name,
 998                        (unsigned long)data->fid1.mds,
 999                        (unsigned long)data->fid1.id,
1000                        (unsigned long)data->fid1.generation,
1001                        (unsigned long)data->fid1.mds);
1002         } else {
1003                 /* request from MDS to acquire i_links for inode by fid1 */
1004                 CDEBUG(D_OTHER, "inc i_nlinks for %lu/%lu/%lu\n",
1005                        (unsigned long)data->fid1.mds,
1006                        (unsigned long)data->fid1.id,
1007                        (unsigned long)data->fid1.generation);
1008         }
1009
1010         rc = md_link(lmv->tgts[data->fid1.mds].ltd_exp, data, request);
1011         RETURN(rc);
1012 }
1013
1014 int lmv_rename(struct obd_export *exp, struct mdc_op_data *data,
1015                const char *old, int oldlen, const char *new, int newlen,
1016                struct ptlrpc_request **request)
1017 {
1018         struct obd_device *obd = exp->exp_obd;
1019         struct lmv_obd *lmv = &obd->u.lmv;
1020         struct lmv_obj *obj;
1021         int rc, mds;
1022         ENTRY;
1023
1024         CDEBUG(D_OTHER, "rename %*s in %lu/%lu/%lu to %*s in %lu/%lu/%lu\n",
1025                oldlen, old, (unsigned long)data->fid1.mds,
1026                (unsigned long)data->fid1.id,
1027                (unsigned long)data->fid1.generation,
1028                newlen, new, (unsigned long) data->fid2.mds,
1029                (unsigned long) data->fid2.id,
1030                (unsigned long) data->fid2.generation);
1031
1032         if (!fid_equal(&data->fid1, &data->fid2))
1033                 CWARN("cross-node rename %lu/%lu/%lu:%*s to %lu/%lu/%lu:%*s\n",
1034                       (unsigned long)data->fid1.mds,
1035                       (unsigned long)data->fid1.id,
1036                       (unsigned long)data->fid1.generation, oldlen, old,
1037                       (unsigned long)data->fid2.mds,
1038                       (unsigned long)data->fid2.id,
1039                       (unsigned long)data->fid2.generation, newlen, new);
1040
1041         rc = lmv_check_connect(obd);
1042         if (rc)
1043                 RETURN(rc);
1044
1045         if (oldlen == 0) {
1046                 /* MDS with old dir entry is asking another MDS to create name
1047                  * there */
1048                 CDEBUG(D_OTHER,
1049                        "create %*s(%d/%d) in %lu/%lu/%lu pointing to %lu/%lu/%lu\n",
1050                        newlen, new, oldlen, newlen,
1051                        (unsigned long)data->fid2.mds,
1052                        (unsigned long)data->fid2.id,
1053                        (unsigned long)data->fid2.generation,
1054                        (unsigned long)data->fid1.mds,
1055                        (unsigned long)data->fid1.id,
1056                        (unsigned long)data->fid1.generation);
1057                 mds = data->fid2.mds;
1058                 goto request;
1059         }
1060
1061         obj = lmv_grab_obj(obd, &data->fid1);
1062         if (obj) {
1063                 /* directory is already splitted, so we have to forward request
1064                  * to the right MDS */
1065                 mds = raw_name2idx(obj->objcount, (char *)old, oldlen);
1066                 data->fid1 = obj->objs[mds].fid;
1067                 CDEBUG(D_OTHER, "forward to MDS #%u (%lu/%lu/%lu)\n", mds,
1068                        (unsigned long)obj->objs[mds].fid.mds,
1069                        (unsigned long)obj->objs[mds].fid.id,
1070                        (unsigned long)obj->objs[mds].fid.generation);
1071                 lmv_put_obj(obj);
1072         }
1073
1074         obj = lmv_grab_obj(obd, &data->fid2);
1075         if (obj) {
1076                 /* directory is already splitted, so we have to forward request
1077                  * to the right MDS */
1078                 mds = raw_name2idx(obj->objcount, (char *)new, newlen);
1079                 data->fid2 = obj->objs[mds].fid;
1080                 CDEBUG(D_OTHER, "forward to MDS #%u (%lu/%lu/%lu)\n", mds,
1081                        (unsigned long)obj->objs[mds].fid.mds,
1082                        (unsigned long)obj->objs[mds].fid.id,
1083                        (unsigned long)obj->objs[mds].fid.generation);
1084                 lmv_put_obj(obj);
1085         }
1086
1087         mds = data->fid1.mds;
1088
1089 request:
1090         rc = md_rename(lmv->tgts[mds].ltd_exp, data, old, oldlen,
1091                        new, newlen, request);
1092         RETURN(rc);
1093 }
1094
1095 int lmv_setattr(struct obd_export *exp, struct mdc_op_data *data,
1096                 struct iattr *iattr, void *ea, int ealen, void *ea2,
1097                 int ea2len, struct ptlrpc_request **request)
1098 {
1099         struct obd_device *obd = exp->exp_obd;
1100         struct lmv_obd *lmv = &obd->u.lmv;
1101         int rc = 0, i = data->fid1.mds;
1102         struct ptlrpc_request *req;
1103         struct mds_body *body;
1104         struct lmv_obj *obj;
1105         ENTRY;
1106
1107         rc = lmv_check_connect(obd);
1108         if (rc)
1109                 RETURN(rc);
1110
1111         obj = lmv_grab_obj(obd, &data->fid1);
1112
1113         CDEBUG(D_OTHER, "SETATTR for %lu/%lu/%lu, valid 0x%x%s\n",
1114                (unsigned long)data->fid1.mds, (unsigned long)data->fid1.id,
1115                (unsigned long)data->fid1.generation, iattr->ia_valid,
1116                obj ? ", splitted" : "");
1117
1118         if (obj) {
1119                 for (i = 0; i < obj->objcount; i++) {
1120                         data->fid1 = obj->objs[i].fid;
1121
1122                         rc = md_setattr(lmv->tgts[i].ltd_exp, data, iattr,
1123                                         ea, ealen, ea2, ea2len, &req);
1124
1125                         if (fid_equal(&obj->fid, &obj->objs[i].fid)) {
1126                                 /* this is master object and this request should
1127                                  * be returned back to llite */
1128                                 *request = req;
1129                         } else {
1130                                 ptlrpc_req_finished(req);
1131                         }
1132
1133                         if (rc)
1134                                 break;
1135                 }
1136                 lmv_put_obj(obj);
1137         } else {
1138                 LASSERT(i < lmv->desc.ld_tgt_count);
1139                 rc = md_setattr(lmv->tgts[i].ltd_exp, data, iattr, ea,
1140                                 ealen, ea2, ea2len, request);
1141                 if (rc == 0) {
1142                         body = lustre_msg_buf((*request)->rq_repmsg, 0,
1143                                               sizeof(*body));
1144                         LASSERT(body != NULL);
1145                         LASSERT(body->mds == i);
1146                 }
1147         }
1148         RETURN(rc);
1149 }
1150
1151 int lmv_sync(struct obd_export *exp, struct ll_fid *fid,
1152              struct ptlrpc_request **request)
1153 {
1154         struct obd_device *obd = exp->exp_obd;
1155         struct lmv_obd *lmv = &obd->u.lmv;
1156         int rc;
1157         ENTRY;
1158
1159         rc = lmv_check_connect(obd);
1160         if (rc)
1161                 RETURN(rc);
1162
1163         rc = md_sync(lmv->tgts[0].ltd_exp, fid, request);
1164         RETURN(rc);
1165 }
1166
1167 int lmv_dirobj_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
1168                             void *data, int flag)
1169 {
1170         struct lustre_handle lockh;
1171         struct lmv_obj *obj;
1172         int rc;
1173         ENTRY;
1174
1175         switch (flag) {
1176         case LDLM_CB_BLOCKING:
1177                 ldlm_lock2handle(lock, &lockh);
1178                 rc = ldlm_cli_cancel(&lockh);
1179                 if (rc < 0) {
1180                         CDEBUG(D_INODE, "ldlm_cli_cancel: %d\n", rc);
1181                         RETURN(rc);
1182                 }
1183                 break;
1184         case LDLM_CB_CANCELING:
1185                 /* time to drop cached attrs for dirobj */
1186                 obj = lock->l_ast_data;
1187                 if (obj) {
1188                         CDEBUG(D_OTHER, "cancel %s on %lu/%lu, master %lu/%lu/%lu\n",
1189                                lock->l_resource->lr_name.name[3] == 1 ? "LOOKUP" : "UPDATE",
1190                                (unsigned long)lock->l_resource->lr_name.name[0],
1191                                (unsigned long)lock->l_resource->lr_name.name[1],
1192                                (unsigned long)obj->fid.mds, (unsigned long)obj->fid.id,
1193                                (unsigned long)obj->fid.generation);
1194                         lmv_put_obj(obj);
1195                 }
1196                 break;
1197         default:
1198                 LBUG();
1199         }
1200         RETURN(0);
1201 }
1202
1203 void lmv_remove_dots(struct page *page)
1204 {
1205         char *kaddr = page_address(page);
1206         unsigned limit = PAGE_CACHE_SIZE;
1207         unsigned offs, rec_len;
1208         struct ext2_dir_entry_2 *p;
1209
1210         for (offs = 0; offs <= limit - EXT2_DIR_REC_LEN(1); offs += rec_len) {
1211                 p = (struct ext2_dir_entry_2 *)(kaddr + offs);
1212                 rec_len = le16_to_cpu(p->rec_len);
1213
1214                 if ((p->name_len == 1 && p->name[0] == '.') ||
1215                     (p->name_len == 2 && p->name[0] == '.' && p->name[1] == '.'))
1216                         p->inode = 0;
1217         }
1218 }
1219
1220 int lmv_readpage(struct obd_export *exp, struct ll_fid *mdc_fid,
1221                  __u64 offset, struct page *page,
1222                  struct ptlrpc_request **request)
1223 {
1224         struct obd_device *obd = exp->exp_obd;
1225         struct lmv_obd *lmv = &obd->u.lmv;
1226         struct ll_fid rfid = *mdc_fid;
1227         struct lmv_obj *obj;
1228         int rc, i;
1229         ENTRY;
1230
1231         rc = lmv_check_connect(obd);
1232         if (rc)
1233                 RETURN(rc);
1234
1235         LASSERT(mdc_fid->mds < lmv->desc.ld_tgt_count);
1236         CDEBUG(D_OTHER, "READPAGE at %llu from %lu/%lu/%lu\n",
1237                offset, (unsigned long) rfid.mds,
1238                (unsigned long) rfid.id,
1239                (unsigned long) rfid.generation);
1240
1241         obj = lmv_grab_obj(obd, mdc_fid);
1242         if (obj) {
1243                 lmv_lock_obj(obj);
1244
1245                 /* find dirobj containing page with requested offset. */
1246                 for (i = 0; i < obj->objcount; i++) {
1247                         if (offset < obj->objs[i].size)
1248                                 break;
1249                         offset -= obj->objs[i].size;
1250                 }
1251                 rfid = obj->objs[i].fid;
1252
1253                 lmv_unlock_obj(obj);
1254                 lmv_put_obj(obj);
1255
1256                 CDEBUG(D_OTHER, "forward to %lu/%lu/%lu with offset %lu\n",
1257                        (unsigned long)rfid.mds, (unsigned long)rfid.id,
1258                        (unsigned long)rfid.generation, (unsigned long)offset);
1259         }
1260         rc = md_readpage(lmv->tgts[rfid.mds].ltd_exp, &rfid, offset,
1261                          page, request);
1262
1263         if (rc == 0 && !fid_equal(&rfid, mdc_fid))
1264                 /* this page isn't from master object. To avoid "." and ".."
1265                  * duplication in directory, we have to remove them from all
1266                  * slave objects */
1267                 lmv_remove_dots(page);
1268
1269         RETURN(rc);
1270 }
1271
1272 int lmv_unlink_slaves(struct obd_export *exp, struct mdc_op_data *data,
1273                       struct ptlrpc_request **req)
1274 {
1275         struct obd_device *obd = exp->exp_obd;
1276         struct lmv_obd *lmv = &obd->u.lmv;
1277         struct mea *mea = data->mea1;
1278         struct mdc_op_data data2;
1279         int i, rc = 0, mds;
1280         ENTRY;
1281
1282         LASSERT(mea != NULL);
1283         for (i = 0; i < mea->mea_count; i++) {
1284                 if (lmv->tgts[i].ltd_exp == NULL)
1285                         continue;
1286
1287                 memset(&data2, 0, sizeof(data2));
1288                 data2.fid1 = mea->mea_fids[i];
1289                 data2.create_mode = MDS_MODE_DONT_LOCK | S_IFDIR;
1290                 mds = data2.fid1.mds;
1291                 rc = md_unlink(lmv->tgts[mds].ltd_exp, &data2, req);
1292                 CDEBUG(D_OTHER, "unlink slave %lu/%lu/%lu -> %d\n",
1293                        (unsigned long) mea->mea_fids[i].mds,
1294                        (unsigned long) mea->mea_fids[i].id,
1295                        (unsigned long) mea->mea_fids[i].generation, rc);
1296                 if (*req) {
1297                         ptlrpc_req_finished(*req);
1298                         *req = NULL;
1299                 }
1300                 if (rc)
1301                         break;
1302         }
1303         RETURN(rc);
1304 }
1305
1306 int lmv_delete_object(struct obd_export *exp, struct ll_fid *fid)
1307 {
1308         ENTRY;
1309
1310         if (!lmv_delete_obj(exp, fid)) {
1311                 CDEBUG(D_OTHER, "Object %lu/%lu/%lu is not found.\n",
1312                        (unsigned long)fid->mds, (unsigned long)fid->id,
1313                        (unsigned long)fid->generation);
1314         }
1315
1316         RETURN(0);
1317 }
1318
1319 int lmv_unlink(struct obd_export *exp, struct mdc_op_data *data,
1320                struct ptlrpc_request **request)
1321 {
1322         struct obd_device *obd = exp->exp_obd;
1323         struct lmv_obd *lmv = &obd->u.lmv;
1324         int rc, i = 0;
1325         ENTRY;
1326
1327         rc = lmv_check_connect(obd);
1328         if (rc)
1329                 RETURN(rc);
1330
1331         if (data->namelen == 0 && data->mea1 != NULL) {
1332                 /* mds asks to remove slave objects */
1333                 rc = lmv_unlink_slaves(exp, data, request);
1334                 RETURN(rc);
1335         } else if (data->namelen != 0) {
1336                 struct lmv_obj *obj;
1337
1338                 obj = lmv_grab_obj(obd, &data->fid1);
1339                 if (obj) {
1340                         i = raw_name2idx(obj->objcount, data->name,
1341                                          data->namelen);
1342                         data->fid1 = obj->objs[i].fid;
1343                         lmv_put_obj(obj);
1344                 }
1345                 CDEBUG(D_OTHER, "unlink '%*s' in %lu/%lu/%lu -> %u\n",
1346                        data->namelen, data->name,
1347                        (unsigned long) data->fid1.mds,
1348                        (unsigned long) data->fid1.id,
1349                        (unsigned long) data->fid1.generation, i);
1350         } else {
1351                 CDEBUG(D_OTHER, "drop i_nlink on %lu/%lu/%lu\n",
1352                        (unsigned long) data->fid1.mds,
1353                        (unsigned long) data->fid1.id,
1354                        (unsigned long) data->fid1.generation);
1355         }
1356         rc = md_unlink(lmv->tgts[data->fid1.mds].ltd_exp, data, request);
1357         RETURN(rc);
1358 }
1359
1360 struct obd_device *lmv_get_real_obd(struct obd_export *exp,
1361                                     char *name, int len)
1362 {
1363         struct obd_device *obd = exp->exp_obd;
1364         struct lmv_obd *lmv = &obd->u.lmv;
1365         int rc;
1366         ENTRY;
1367
1368         rc = lmv_check_connect(obd);
1369         if (rc)
1370                 RETURN(ERR_PTR(rc));
1371         obd = lmv->tgts[0].ltd_exp->exp_obd;
1372         EXIT;
1373         return obd;
1374 }
1375
1376 int lmv_init_ea_size(struct obd_export *exp, int easize, int cookiesize)
1377 {
1378         struct obd_device *obd = exp->exp_obd;
1379         struct lmv_obd *lmv = &obd->u.lmv;
1380         int i, rc = 0, change = 0;
1381         ENTRY;
1382
1383         if (lmv->max_easize < easize) {
1384                 lmv->max_easize = easize;
1385                 change = 1;
1386         }
1387         if (lmv->max_cookiesize < cookiesize) {
1388                 lmv->max_cookiesize = cookiesize;
1389                 change = 1;
1390         }
1391         if (change == 0)
1392                 RETURN(0);
1393
1394         if (lmv->connected == 0)
1395                 RETURN(0);
1396
1397         for (i = 0; i < lmv->desc.ld_tgt_count; i++) {
1398                 rc = obd_init_ea_size(lmv->tgts[i].ltd_exp, easize, cookiesize);
1399                 if (rc) {
1400                         CERROR("obd_init_ea_size() failed on MDT target %d, "
1401                                "error %d.\n", i, rc);
1402                         break;
1403                 }
1404         }
1405         RETURN(rc);
1406 }
1407
1408 int lmv_obd_create_single(struct obd_export *exp, struct obdo *oa,
1409                           struct lov_stripe_md **ea, struct obd_trans_info *oti)
1410 {
1411         struct obd_device *obd = exp->exp_obd;
1412         struct lmv_obd *lmv = &obd->u.lmv;
1413         struct lov_stripe_md obj_md;
1414         struct lov_stripe_md *obj_mdp = &obj_md;
1415         int rc = 0;
1416         ENTRY;
1417
1418         rc = lmv_check_connect(obd);
1419         if (rc)
1420                 RETURN(rc);
1421
1422         LASSERT(ea == NULL);
1423         LASSERT(oa->o_mds < lmv->desc.ld_tgt_count);
1424
1425         rc = obd_create(lmv->tgts[oa->o_mds].ltd_exp, oa, &obj_mdp, oti);
1426
1427         RETURN(rc);
1428 }
1429
1430 /*
1431  * to be called from MDS only
1432  */
1433 int lmv_obd_create(struct obd_export *exp, struct obdo *oa,
1434                    struct lov_stripe_md **ea, struct obd_trans_info *oti)
1435 {
1436         struct obd_device *obd = exp->exp_obd;
1437         struct lmv_obd *lmv = &obd->u.lmv;
1438         struct mea *mea;
1439         int i, c, rc = 0;
1440         struct ll_fid mfid;
1441         ENTRY;
1442
1443         rc = lmv_check_connect(obd);
1444         if (rc)
1445                 RETURN(rc);
1446
1447         LASSERT(oa != NULL);
1448
1449         if (ea == NULL) {
1450                 rc = lmv_obd_create_single(exp, oa, NULL, oti);
1451                 RETURN(rc);
1452         }
1453
1454         if (*ea == NULL) {
1455                 rc = obd_alloc_diskmd(exp, (struct lov_mds_md **)ea);
1456                 if (rc < 0) {
1457                         CERROR("obd_alloc_diskmd() failed, error %d\n",
1458                                rc);
1459                         RETURN(rc);
1460                 }
1461
1462                 if (*ea == NULL)
1463                         RETURN(-EINVAL);
1464         }
1465
1466         rc = 0;
1467         mfid.id = oa->o_id;
1468         mfid.generation = oa->o_generation;
1469
1470         mea = (struct mea *)*ea;
1471         if (!mea->mea_count || mea->mea_count > lmv->desc.ld_tgt_count)
1472                 mea->mea_count = lmv->desc.ld_tgt_count;
1473
1474         mea->mea_master = -1;
1475
1476         for (i = 0, c = 0; c < mea->mea_count &&
1477                 i < lmv->desc.ld_tgt_count; i++) {
1478                 struct lov_stripe_md obj_md;
1479                 struct lov_stripe_md *obj_mdp = &obj_md;
1480
1481                 if (lmv->tgts[i].ltd_exp == NULL) {
1482                         /* this is master MDS */
1483                         mea->mea_fids[c].id = mfid.id;
1484                         mea->mea_fids[c].generation = mfid.generation;
1485                         mea->mea_fids[c].mds = i;
1486                         mea->mea_master = i;
1487                         c++;
1488                         continue;
1489                 }
1490
1491                 /* "master" MDS should always be part of stripped dir, so scan
1492                    for it. */
1493                 if (mea->mea_master == -1 && c == mea->mea_count - 1)
1494                         continue;
1495
1496                 oa->o_valid = OBD_MD_FLGENER | OBD_MD_FLTYPE | OBD_MD_FLMODE
1497                         | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLID;
1498
1499                 rc = obd_create(lmv->tgts[c].ltd_exp, oa, &obj_mdp, oti);
1500                 if (rc) {
1501                         CERROR("obd_create() failed on MDT target %d, "
1502                                "error %d\n", c, rc);
1503                         RETURN(rc);
1504                 }
1505
1506                 mea->mea_fids[c].id = oa->o_id;
1507                 mea->mea_fids[c].generation = oa->o_generation;
1508                 mea->mea_fids[c].mds = i;
1509                 c++;
1510                 CDEBUG(D_OTHER, "dirobj at mds %d: "LPU64"/%u\n",
1511                        i, oa->o_id, oa->o_generation);
1512         }
1513         LASSERT(c == mea->mea_count);
1514         CDEBUG(D_OTHER, "%d dirobjects created\n", (int) mea->mea_count);
1515
1516         RETURN(rc);
1517 }
1518
1519 static int lmv_get_info(struct obd_export *exp, __u32 keylen,
1520                         void *key, __u32 *vallen, void *val)
1521 {
1522         struct obd_device *obd;
1523         struct lmv_obd *lmv;
1524         ENTRY;
1525
1526         obd = class_exp2obd(exp);
1527         if (obd == NULL) {
1528                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1529                        exp->exp_handle.h_cookie);
1530                 RETURN(-EINVAL);
1531         }
1532
1533         lmv = &obd->u.lmv;
1534         if (keylen == 6 && memcmp(key, "mdsize", 6) == 0) {
1535                 __u32 *mdsize = val;
1536                 *vallen = sizeof(__u32);
1537                 *mdsize = sizeof(struct ll_fid) * lmv->desc.ld_tgt_count
1538                                 + sizeof(struct mea);
1539                 RETURN(0);
1540         } else if (keylen == 6 && memcmp(key, "mdsnum", 6) == 0) {
1541                 struct obd_uuid *cluuid = &lmv->cluuid;
1542                 struct lmv_tgt_desc *tgts;
1543                 __u32 *mdsnum = val;
1544                 int i;
1545
1546                 for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) {
1547                         if (obd_uuid_equals(&tgts->uuid, cluuid)) {
1548                                 *vallen = sizeof(__u32);
1549                                 *mdsnum = i;
1550                                 RETURN(0);
1551                         }
1552                 }
1553                 LASSERT(0);
1554         }
1555
1556         CDEBUG(D_IOCTL, "invalid key\n");
1557         RETURN(-EINVAL);
1558 }
1559
1560 int lmv_set_info(struct obd_export *exp, obd_count keylen,
1561                  void *key, obd_count vallen, void *val)
1562 {
1563         struct obd_device *obd;
1564         struct lmv_obd *lmv;
1565         ENTRY;
1566
1567         obd = class_exp2obd(exp);
1568         if (obd == NULL) {
1569                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1570                        exp->exp_handle.h_cookie);
1571                 RETURN(-EINVAL);
1572         }
1573         lmv = &obd->u.lmv;
1574
1575         if (keylen >= strlen("client") && strcmp(key, "client") == 0) {
1576                 struct lmv_tgt_desc *tgts;
1577                 int i, rc;
1578
1579                 rc = lmv_check_connect(obd);
1580                 if (rc)
1581                         RETURN(rc);
1582
1583                 for (i = 0, tgts = lmv->tgts;
1584                         i < lmv->desc.ld_tgt_count; i++, tgts++) {
1585                         rc = obd_set_info(tgts->ltd_exp, keylen, key, vallen, val);
1586                         if (rc)
1587                                 RETURN(rc);
1588                 }
1589                 RETURN(0);
1590         } else if (keylen >= strlen("inter_mds") && strcmp(key, "inter_mds") == 0) {
1591                 lmv->server_timeout = 1;
1592                 lmv_set_timeouts(obd);
1593                 RETURN(0);
1594         }
1595
1596         RETURN(-EINVAL);
1597 }
1598
1599 int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
1600                struct lov_stripe_md *lsm)
1601 {
1602         struct obd_device *obd = class_exp2obd(exp);
1603         struct lmv_obd *lmv = &obd->u.lmv;
1604         int mea_size;
1605         ENTRY;
1606
1607         mea_size = sizeof(struct ll_fid) *
1608                 lmv->desc.ld_tgt_count + sizeof(struct mea);
1609         if (!lmmp)
1610                 RETURN(mea_size);
1611
1612         if (*lmmp && !lsm) {
1613                 OBD_FREE(*lmmp, mea_size);
1614                 *lmmp = NULL;
1615                 RETURN(0);
1616         }
1617
1618         if (*lmmp == NULL) {
1619                 OBD_ALLOC(*lmmp, mea_size);
1620                 if (*lmmp == NULL)
1621                         RETURN(-ENOMEM);
1622         }
1623
1624         if (!lsm)
1625                 RETURN(mea_size);
1626
1627 #warning "MEA packing/convertation must be here! -bzzz"
1628         memcpy(*lmmp, lsm, mea_size);
1629         RETURN(mea_size);
1630 }
1631
1632 int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **mem_tgt,
1633                  struct lov_mds_md *disk_src, int mdsize)
1634 {
1635         struct obd_device *obd = class_exp2obd(exp);
1636         struct lmv_obd *lmv = &obd->u.lmv;
1637         struct mea **tmea = (struct mea **) mem_tgt;
1638         struct mea *mea = (void *) disk_src;
1639         int mea_size;
1640         ENTRY;
1641
1642         mea_size = sizeof(struct ll_fid) *
1643                 lmv->desc.ld_tgt_count + sizeof(struct mea);
1644         if (mem_tgt == NULL)
1645                 return mea_size;
1646
1647         if (*mem_tgt != NULL && disk_src == NULL) {
1648                 OBD_FREE(*tmea, mea_size);
1649                 RETURN(0);
1650         }
1651
1652         LASSERT(mea_size == mdsize);
1653
1654         OBD_ALLOC(*tmea, mea_size);
1655         if (*tmea == NULL)
1656                 RETURN(-ENOMEM);
1657
1658         if (!disk_src)
1659                 RETURN(mea_size);
1660
1661 #warning "MEA unpacking/convertation must be here! -bzzz"
1662         memcpy(*tmea, mea, mdsize);
1663         RETURN(mea_size);
1664 }
1665
1666 int lmv_brw(int rw, struct obd_export *exp, struct obdo *oa,
1667             struct lov_stripe_md *ea, obd_count oa_bufs,
1668             struct brw_page *pgarr, struct obd_trans_info *oti)
1669 {
1670         struct obd_device *obd = exp->exp_obd;
1671         struct lmv_obd *lmv = &obd->u.lmv;
1672         struct mea *mea = (struct mea *) ea;
1673         int err;
1674
1675         LASSERT(oa != NULL);
1676         LASSERT(ea != NULL);
1677         LASSERT(pgarr != NULL);
1678         LASSERT(oa->o_mds < lmv->desc.ld_tgt_count);
1679
1680         oa->o_gr = mea->mea_fids[oa->o_mds].generation;
1681         oa->o_id = mea->mea_fids[oa->o_mds].id;
1682         oa->o_valid =  OBD_MD_FLID | OBD_MD_FLGROUP;
1683         err = obd_brw(rw, lmv->tgts[oa->o_mds].ltd_exp, oa,
1684                       NULL, oa_bufs, pgarr, oti);
1685         RETURN(err);
1686 }
1687
1688 struct obd_ops lmv_obd_ops = {
1689         .o_owner                = THIS_MODULE,
1690         .o_attach               = lmv_attach,
1691         .o_detach               = lmv_detach,
1692         .o_setup                = lmv_setup,
1693         .o_cleanup              = lmv_cleanup,
1694         .o_connect              = lmv_connect,
1695         .o_disconnect           = lmv_disconnect,
1696         .o_statfs               = lmv_statfs,
1697         .o_get_info             = lmv_get_info,
1698         .o_set_info             = lmv_set_info,
1699         .o_create               = lmv_obd_create,
1700         .o_packmd               = lmv_packmd,
1701         .o_unpackmd             = lmv_unpackmd,
1702         .o_brw                  = lmv_brw,
1703         .o_init_ea_size         = lmv_init_ea_size,
1704         .o_notify               = lmv_notify,
1705         .o_iocontrol            = lmv_iocontrol,
1706 };
1707
1708 struct md_ops lmv_md_ops = {
1709         .m_getstatus            = lmv_getstatus,
1710         .m_getattr              = lmv_getattr,
1711         .m_change_cbdata        = lmv_change_cbdata,
1712         .m_change_cbdata_name   = lmv_change_cbdata_name,
1713         .m_close                = lmv_close,
1714         .m_create               = lmv_create,
1715         .m_done_writing         = lmv_done_writing,
1716         .m_enqueue              = lmv_enqueue,
1717         .m_getattr_name         = lmv_getattr_name,
1718         .m_intent_lock          = lmv_intent_lock,
1719         .m_link                 = lmv_link,
1720         .m_rename               = lmv_rename,
1721         .m_setattr              = lmv_setattr,
1722         .m_sync                 = lmv_sync,
1723         .m_readpage             = lmv_readpage,
1724         .m_unlink               = lmv_unlink,
1725         .m_get_real_obd         = lmv_get_real_obd,
1726         .m_valid_attrs          = lmv_valid_attrs,
1727         .m_delete_object        = lmv_delete_object,
1728 };
1729
1730 int __init lmv_init(void)
1731 {
1732         struct lprocfs_static_vars lvars;
1733         int rc;
1734
1735         lprocfs_init_vars(lmv, &lvars);
1736         rc = class_register_type(&lmv_obd_ops, &lmv_md_ops,
1737                                  lvars.module_vars, OBD_LMV_DEVICENAME);
1738         RETURN(rc);
1739 }
1740
1741 #ifdef __KERNEL__
1742 static void lmv_exit(void)
1743 {
1744         class_unregister_type(OBD_LMV_DEVICENAME);
1745 }
1746
1747 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1748 MODULE_DESCRIPTION("Lustre Logical Metadata Volume OBD driver");
1749 MODULE_LICENSE("GPL");
1750
1751 module_init(lmv_init);
1752 module_exit(lmv_exit);
1753 #endif