lustre/mds/handler.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  lustre/mds/handler.c
   5  *  Lustre Metadata Server (mds) request handler
   6  *
   7  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   8  *   Author: Peter Braam <braam@clusterfs.com>
   9  *   Author: Andreas Dilger <adilger@clusterfs.com>
  10  *   Author: Phil Schwan <phil@clusterfs.com>
  11  *   Author: Mike Shaver <shaver@clusterfs.com>
  12  *
  13  *   This file is part of Lustre, http://www.lustre.org.
  14  *
  15  *   Lustre is free software; you can redistribute it and/or
  16  *   modify it under the terms of version 2 of the GNU General Public
  17  *   License as published by the Free Software Foundation.
  18  *
  19  *   Lustre is distributed in the hope that it will be useful,
  20  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   GNU General Public License for more details.
  23  *
  24  *   You should have received a copy of the GNU General Public License
  25  *   along with Lustre; if not, write to the Free Software
  26  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_MDS
  33
  34 #include <linux/module.h>
  35 #include <linux/lustre_mds.h>
  36 #include <linux/lustre_dlm.h>
  37 #include <linux/init.h>
  38 #include <linux/obd_class.h>
  39 #include <linux/random.h>
  40 #include <linux/fs.h>
  41 #include <linux/jbd.h>
  42 #include <linux/ext3_fs.h>
  43 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  44 # include <linux/smp_lock.h>
  45 # include <linux/buffer_head.h>
  46 # include <linux/workqueue.h>
  47 # include <linux/mount.h>
  48 #else
  49 # include <linux/locks.h>
  50 #endif
  51 #include <linux/obd_lov.h>
  52 #include <linux/obd_ost.h>
  53 #include <linux/lustre_mds.h>
  54 #include <linux/lustre_fsfilt.h>
  55 #include <linux/lprocfs_status.h>
  56 #include <linux/lustre_commit_confd.h>
  57
  58 #include "mds_internal.h"
  59
  60 static int mds_intent_policy(struct ldlm_namespace *ns,
  61                              struct ldlm_lock **lockp, void *req_cookie,
  62                              ldlm_mode_t mode, int flags, void *data);
  63 static int mds_postsetup(struct obd_device *obd);
  64 static int mds_cleanup(struct obd_device *obd, int flags);
  65
  66
  67 /* Assumes caller has already pushed into the kernel filesystem context */
  68 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  69                         loff_t offset, int count)
  70 {
  71         struct ptlrpc_bulk_desc *desc;
  72         struct l_wait_info lwi;
  73         struct page **pages;
  74         int rc = 0, npages, i, tmpcount, tmpsize = 0;
  75         ENTRY;
  76
  77         LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
  78
  79         npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
  80         OBD_ALLOC(pages, sizeof(*pages) * npages);
  81         if (!pages)
  82                 GOTO(out, rc = -ENOMEM);
  83
  84         desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
  85                                     MDS_BULK_PORTAL);
  86         if (desc == NULL)
  87                 GOTO(out_free, rc = -ENOMEM);
  88
  89         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
  90                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
  91
  92                 pages[i] = alloc_pages(GFP_KERNEL, 0);
  93                 if (pages[i] == NULL)
  94                         GOTO(cleanup_buf, rc = -ENOMEM);
  95
  96                 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
  97         }
  98
  99         for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
 100                 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
 101                 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
 102                        tmpsize, offset, file->f_dentry->d_inode->i_ino,
 103                        file->f_dentry->d_inode->i_size);
 104
 105                 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
 106                                      kmap(pages[i]), tmpsize, &offset);
 107                 kunmap(pages[i]);
 108
 109                 if (rc != tmpsize)
 110                         GOTO(cleanup_buf, rc = -EIO);
 111         }
 112
 113         LASSERT(desc->bd_nob == count);
 114
 115         rc = ptlrpc_start_bulk_transfer(desc);
 116         if (rc)
 117                 GOTO(cleanup_buf, rc);
 118
 119         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
 120                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
 121                        OBD_FAIL_MDS_SENDPAGE, rc = -EIO);
 122                 GOTO(abort_bulk, rc);
 123         }
 124
 125         lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
 126         rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
 127         LASSERT (rc == 0 || rc == -ETIMEDOUT);
 128
 129         if (rc == 0) {
 130                 if (desc->bd_success &&
 131                     desc->bd_nob_transferred == count)
 132                         GOTO(cleanup_buf, rc);
 133
 134                 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
 135         }
 136
 137         DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
 138                   (rc == -ETIMEDOUT) ? "timeout" : "network error",
 139                   desc->bd_nob_transferred, count,
 140                   req->rq_export->exp_client_uuid.uuid,
 141                   req->rq_export->exp_connection->c_remote_uuid.uuid);
 142
 143         ptlrpc_fail_export(req->rq_export);
 144
 145         EXIT;
 146  abort_bulk:
 147         ptlrpc_abort_bulk (desc);
 148  cleanup_buf:
 149         for (i = 0; i < npages; i++)
 150                 if (pages[i])
 151                         __free_pages(pages[i], 0);
 152
 153         ptlrpc_free_bulk(desc);
 154  out_free:
 155         OBD_FREE(pages, sizeof(*pages) * npages);
 156  out:
 157         return rc;
 158 }
 159
 160 int mds_lock_mode_for_dir(struct obd_device *obd,
 161                           struct dentry *dentry, int mode)
 162 {
 163         int ret_mode, split;
 164
 165         /* any dir access needs couple locks:
 166          * 1) on part of dir we gonna lookup/modify in
 167          * 2) on a whole dir to protect it from concurrent splitting
 168          *    and to flush client's cache for readdir()
 169          * so, for a given mode and dentry this routine decides what
 170          * lock mode to use for lock #2:
 171          * 1) if caller's gonna lookup in dir then we need to protect
 172          *    dir from being splitted only - LCK_CR
 173          * 2) if caller's gonna modify dir then we need to protect
 174          *    dir from being splitted and to flush cache - LCK_CW
 175          * 3) if caller's gonna modify dir and that dir seems ready
 176          *    for splitting then we need to protect it from any
 177          *    type of access (lookup/modify/split) - LCK_EX -bzzz */
 178
 179         split = mds_splitting_expected(obd, dentry);
 180
 181         /*
 182          * it is important to check here only for MDS_NO_SPLITTABLE. The reason
 183          * is that MDS_NO_SPLITTABLE means dir is not splittable in principle
 184          * and another thread will not split it on the quiet. But if we have
 185          * MDS_NO_SPLIT_EXPECTED, this means, that dir may be splitted anytime,
 186          * but not now (for current thread) and we should consider that it can
 187          * happen soon and go that branch which can yield LCK_EX to protect from
 188          * possible splitting.
 189          */
 190         if (split == MDS_NO_SPLITTABLE) {
 191                 /*
 192                  * this inode won't be splitted. so we need not to protect from
 193                  * just flush client's cache on modification.
 194                  */
 195                 ret_mode = 0;
 196                 if (mode == LCK_PW)
 197                         ret_mode = LCK_CW;
 198         } else {
 199                 if (mode == LCK_EX)
 200                         return LCK_EX;
 201
 202                 if (mode == LCK_PR) {
 203                         ret_mode = LCK_CR;
 204                 } else if (mode == LCK_PW) {
 205                         /*
 206                          * caller gonna modify directory.we use concurrent write
 207                          * lock here to retract client's cache for readdir.
 208                          */
 209                         ret_mode = LCK_CW;
 210                         if (split == MDS_EXPECT_SPLIT) {
 211                                 /*
 212                                  * splitting possible. serialize any access the
 213                                  * idea is that first one seen dir is splittable
 214                                  * is given exclusive lock and split
 215                                  * directory. caller passes lock mode to
 216                                  * mds_try_to_split_dir() and splitting would be
 217                                  * done with exclusive lock only -bzzz.
 218                                  */
 219                                 CDEBUG(D_OTHER, "%s: gonna split %lu/%lu\n",
 220                                        obd->obd_name,
 221                                        (unsigned long)dentry->d_inode->i_ino,
 222                                        (unsigned long)dentry->d_inode->i_generation);
 223                                 ret_mode = LCK_EX;
 224                         }
 225                 }
 226         }
 227         return ret_mode;
 228 }
 229
 230 /* only valid locked dentries or errors should be returned */
 231 struct dentry *mds_id2locked_dentry(struct obd_device *obd, struct lustre_id *id,
 232                                     struct vfsmount **mnt, int lock_mode,
 233                                     struct lustre_handle *lockh, int *mode,
 234                                     char *name, int namelen, __u64 lockpart)
 235 {
 236         struct dentry *de = mds_id2dentry(obd, id, mnt), *retval = de;
 237         ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
 238         struct ldlm_res_id res_id = { .name = {0} };
 239         int flags = 0, rc;
 240         ENTRY;
 241
 242         if (IS_ERR(de))
 243                 RETURN(de);
 244
 245         lockh[1].cookie = 0;
 246         res_id.name[0] = id_fid(id);
 247         res_id.name[1] = id_group(id);
 248
 249 #ifdef S_PDIROPS
 250         if (name && IS_PDIROPS(de->d_inode)) {
 251                 ldlm_policy_data_t cpolicy =
 252                         { .l_inodebits = { MDS_INODELOCK_UPDATE } };
 253                 LASSERT(mode != NULL);
 254                 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
 255                 if (*mode) {
 256                         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
 257                                               res_id, LDLM_IBITS,
 258                                               &cpolicy, *mode, &flags,
 259                                               mds_blocking_ast,
 260                                               ldlm_completion_ast, NULL, NULL,
 261                                               NULL, 0, NULL, lockh + 1);
 262                         if (rc != ELDLM_OK) {
 263                                 l_dput(de);
 264                                 RETURN(ERR_PTR(-ENOLCK));
 265                         }
 266                 }
 267                 flags = 0;
 268
 269                 res_id.name[2] = full_name_hash(name, namelen);
 270
 271                 CDEBUG(D_INFO, "take lock on "DLID4":"LPX64"\n",
 272                        OLID4(id), res_id.name[2]);
 273         }
 274 #else
 275 #warning "No PDIROPS support in the kernel"
 276 #endif
 277         rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
 278                               LDLM_IBITS, &policy, lock_mode, &flags,
 279                               mds_blocking_ast, ldlm_completion_ast,
 280                               NULL, NULL, NULL, 0, NULL, lockh);
 281         if (rc != ELDLM_OK) {
 282                 l_dput(de);
 283                 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
 284 #ifdef S_PDIROPS
 285                 if (lockh[1].cookie)
 286                         ldlm_lock_decref(lockh + 1, *mode);
 287 #endif
 288         }
 289
 290         RETURN(retval);
 291 }
 292
 293 #ifndef DCACHE_DISCONNECTED
 294 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
 295 #endif
 296
 297
 298 /* Look up an entry by inode number. This function ONLY returns valid dget'd
 299  * dentries with an initialized inode or errors */
 300 struct dentry *mds_id2dentry(struct obd_device *obd, struct lustre_id *id,
 301                              struct vfsmount **mnt)
 302 {
 303         unsigned long ino = (unsigned long)id_ino(id);
 304         __u32 generation = (__u32)id_gen(id);
 305         struct mds_obd *mds = &obd->u.mds;
 306         struct dentry *result;
 307         struct inode *inode;
 308         char idname[32];
 309
 310         if (ino == 0)
 311                 RETURN(ERR_PTR(-ESTALE));
 312
 313         snprintf(idname, sizeof(idname), "0x%lx", ino);
 314
 315         CDEBUG(D_DENTRY, "--> mds_id2dentry: ino/gen %lu/%u, sb %p\n",
 316                ino, generation, mds->mds_sb);
 317
 318         /* under ext3 this is neither supposed to return bad inodes nor NULL
 319            inodes. */
 320         result = ll_lookup_one_len(idname, mds->mds_id_de,
 321                                    strlen(idname));
 322         if (IS_ERR(result))
 323                 RETURN(result);
 324
 325         inode = result->d_inode;
 326         if (!inode)
 327                 RETURN(ERR_PTR(-ENOENT));
 328
 329         if (is_bad_inode(inode)) {
 330                 CERROR("bad inode returned %lu/%u\n",
 331                        inode->i_ino, inode->i_generation);
 332                 dput(result);
 333                 RETURN(ERR_PTR(-ENOENT));
 334         }
 335
 336         /* here we disabled generation check, as root inode i_generation
 337          * of cache mds and real mds are different. */
 338         if (inode->i_ino != id_ino(&mds->mds_rootid) && generation &&
 339             inode->i_generation != generation) {
 340                 /* we didn't find the right inode.. */
 341                 CERROR("bad inode %lu, link: %lu, ct: %d, generation %u/%u\n",
 342                        inode->i_ino, (unsigned long)inode->i_nlink,
 343                        atomic_read(&inode->i_count), inode->i_generation,
 344                        generation);
 345                 dput(result);
 346                 RETURN(ERR_PTR(-ENOENT));
 347         }
 348
 349         if (mnt) {
 350                 *mnt = mds->mds_vfsmnt;
 351                 mntget(*mnt);
 352         }
 353
 354         RETURN(result);
 355 }
 356
 357
 358 /* Establish a connection to the MDS.
 359  *
 360  * This will set up an export structure for the client to hold state data about
 361  * that client, like open files, the last operation number it did on the server,
 362  * etc.
 363  */
 364 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 365                        struct obd_uuid *cluuid, unsigned long flags)
 366 {
 367         struct mds_export_data *med;
 368         struct mds_client_data *mcd;
 369         struct obd_export *exp;
 370         int rc;
 371         ENTRY;
 372
 373         if (!conn || !obd || !cluuid)
 374                 RETURN(-EINVAL);
 375
 376         /* XXX There is a small race between checking the list and adding a new
 377          * connection for the same UUID, but the real threat (list corruption
 378          * when multiple different clients connect) is solved.
 379          *
 380          * There is a second race between adding the export to the list, and
 381          * filling in the client data below.  Hence skipping the case of NULL
 382          * mcd above.  We should already be controlling multiple connects at the
 383          * client, and we can't hold the spinlock over memory allocations
 384          * without risk of deadlocking.
 385          */
 386         rc = class_connect(conn, obd, cluuid);
 387         if (rc)
 388                 RETURN(rc);
 389         exp = class_conn2export(conn);
 390
 391         LASSERT(exp != NULL);
 392         med = &exp->exp_mds_data;
 393
 394         OBD_ALLOC(mcd, sizeof(*mcd));
 395         if (!mcd) {
 396                 CERROR("%s: out of memory for client data.\n",
 397                         obd->obd_name);
 398                 GOTO(out, rc = -ENOMEM);
 399         }
 400
 401         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 402         med->med_mcd = mcd;
 403
 404         rc = mds_client_add(obd, &obd->u.mds, med, -1);
 405         if (rc)
 406                 GOTO(out, rc);
 407
 408         EXIT;
 409 out:
 410         if (rc) {
 411                 if (mcd)
 412                         OBD_FREE(mcd, sizeof(*mcd));
 413                 class_disconnect(exp, 0);
 414         } else {
 415                 class_export_put(exp);
 416         }
 417         return rc;
 418 }
 419
 420 static int mds_connect_post(struct obd_export *exp, unsigned long flags)
 421 {
 422         struct obd_device *obd = exp->exp_obd;
 423         struct mds_obd *mds = &obd->u.mds;
 424         int rc = 0;
 425         ENTRY;
 426
 427         if (!(flags & OBD_OPT_MDS_CONNECTION)) {
 428                 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
 429                         atomic_inc(&mds->mds_real_clients);
 430                         CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
 431                                obd->obd_name, exp->exp_client_uuid.uuid,
 432                                atomic_read(&mds->mds_real_clients));
 433                         exp->exp_flags |= OBD_OPT_REAL_CLIENT;
 434                 }
 435                 if (mds->mds_lmv_name)
 436                         rc = mds_lmv_connect(obd, mds->mds_lmv_name);
 437         }
 438         RETURN(rc);
 439 }
 440
 441 static int mds_init_export(struct obd_export *exp)
 442 {
 443         struct mds_export_data *med = &exp->exp_mds_data;
 444
 445         INIT_LIST_HEAD(&med->med_open_head);
 446         spin_lock_init(&med->med_open_lock);
 447         return 0;
 448 }
 449
 450 static int mds_destroy_export(struct obd_export *export)
 451 {
 452         struct mds_export_data *med;
 453         struct obd_device *obd = export->exp_obd;
 454         struct lvfs_run_ctxt saved;
 455         int rc = 0;
 456         ENTRY;
 457
 458         med = &export->exp_mds_data;
 459         target_destroy_export(export);
 460
 461         if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
 462                 GOTO(out, 0);
 463
 464         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 465
 466         /* Close any open files (which may also cause orphan unlinking). */
 467         spin_lock(&med->med_open_lock);
 468         while (!list_empty(&med->med_open_head)) {
 469                 struct list_head *tmp = med->med_open_head.next;
 470                 struct mds_file_data *mfd =
 471                         list_entry(tmp, struct mds_file_data, mfd_list);
 472                 BDEVNAME_DECLARE_STORAGE(btmp);
 473
 474                 /* bug 1579: fix force-closing for 2.5 */
 475                 struct dentry *dentry = mfd->mfd_dentry;
 476
 477                 list_del(&mfd->mfd_list);
 478                 spin_unlock(&med->med_open_lock);
 479
 480                 /* If you change this message, be sure to update
 481                  * replay_single:test_46 */
 482                 CERROR("force closing client file handle for %*s (%s:%lu)\n",
 483                        dentry->d_name.len, dentry->d_name.name,
 484                        ll_bdevname(dentry->d_inode->i_sb, btmp),
 485                        dentry->d_inode->i_ino);
 486                 /* child inode->i_alloc_sem protects orphan_dec_test and
 487                  * is_orphan race, mds_mfd_close drops it */
 488                 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
 489                 rc = mds_mfd_close(NULL, 0, obd, mfd,
 490                                    !(export->exp_flags & OBD_OPT_FAILOVER));
 491                 if (rc)
 492                         CDEBUG(D_INODE, "Error closing file: %d\n", rc);
 493                 spin_lock(&med->med_open_lock);
 494         }
 495         spin_unlock(&med->med_open_lock);
 496         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 497
 498         EXIT;
 499 out:
 500         mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
 501         return rc;
 502 }
 503
 504 static int mds_disconnect(struct obd_export *exp, unsigned long flags)
 505 {
 506         unsigned long irqflags;
 507         struct obd_device *obd;
 508         struct mds_obd *mds;
 509         int rc;
 510         ENTRY;
 511
 512         LASSERT(exp != NULL);
 513         obd = class_exp2obd(exp);
 514         if (obd == NULL) {
 515                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
 516                        exp->exp_handle.h_cookie);
 517                 RETURN(-EINVAL);
 518         }
 519         mds = &obd->u.mds;
 520
 521         /*
 522          * suppress any inter-mds requests durring disconnecting lmv if this is
 523          * detected --force mode. This is needed to avoid endless recovery.
 524          */
 525         if (atomic_read(&mds->mds_real_clients) > 0 &&
 526             !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
 527                 flags |= OBD_OPT_FORCE;
 528
 529         if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
 530             && !atomic_read(&mds->mds_real_clients)) {
 531                 /* there was no client at all */
 532                 mds_lmv_disconnect(obd, flags);
 533         }
 534
 535         if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
 536             && atomic_dec_and_test(&mds->mds_real_clients)) {
 537                 /* time to drop LMV connections */
 538                 CDEBUG(D_OTHER, "%s: last real client %s disconnected.  "
 539                        "Disconnnect from LMV now\n",
 540                        obd->obd_name, exp->exp_client_uuid.uuid);
 541                 mds_lmv_disconnect(obd, flags);
 542         }
 543
 544         spin_lock_irqsave(&exp->exp_lock, irqflags);
 545         exp->exp_flags = flags;
 546         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 547
 548         /* disconnect early so that clients can't keep using export */
 549         rc = class_disconnect(exp, flags);
 550         ldlm_cancel_locks_for_export(exp);
 551
 552         /* complete all outstanding replies */
 553         spin_lock_irqsave(&exp->exp_lock, irqflags);
 554         while (!list_empty(&exp->exp_outstanding_replies)) {
 555                 struct ptlrpc_reply_state *rs =
 556                         list_entry(exp->exp_outstanding_replies.next,
 557                                    struct ptlrpc_reply_state, rs_exp_list);
 558                 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
 559
 560                 spin_lock(&svc->srv_lock);
 561                 list_del_init(&rs->rs_exp_list);
 562                 ptlrpc_schedule_difficult_reply(rs);
 563                 spin_unlock(&svc->srv_lock);
 564         }
 565         spin_unlock_irqrestore(&exp->exp_lock, irqflags);
 566         RETURN(rc);
 567 }
 568
 569 static int mds_getstatus(struct ptlrpc_request *req)
 570 {
 571         struct mds_obd *mds = mds_req2mds(req);
 572         struct mds_body *body;
 573         int rc, size;
 574         ENTRY;
 575
 576         size = sizeof(*body);
 577
 578         rc = lustre_pack_reply(req, 1, &size, NULL);
 579         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 580                 CERROR("mds: out of memory for message: size=%d\n", size);
 581                 req->rq_status = -ENOMEM;       /* superfluous? */
 582                 RETURN(-ENOMEM);
 583         }
 584
 585         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
 586
 587         body->valid |= OBD_MD_FID;
 588         memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
 589
 590         /*
 591          * the last_committed and last_xid fields are filled in for all replies
 592          * already - no need to do so here also.
 593          */
 594         RETURN(0);
 595 }
 596
 597 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 598                      void *data, int flag)
 599 {
 600         int do_ast;
 601         ENTRY;
 602
 603         if (flag == LDLM_CB_CANCELING) {
 604                 /* Don't need to do anything here. */
 605                 RETURN(0);
 606         }
 607
 608         /* XXX layering violation!  -phil */
 609         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 610
 611         /*
 612          * get this: if mds_blocking_ast is racing with mds_intent_policy, such
 613          * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
 614          * then by the time we get the lock, we might not be the correct
 615          * blocking function anymore.  So check, and return early, if so.
 616          */
 617         if (lock->l_blocking_ast != mds_blocking_ast) {
 618                 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 619                 RETURN(0);
 620         }
 621
 622         lock->l_flags |= LDLM_FL_CBPENDING;
 623         do_ast = (!lock->l_readers && !lock->l_writers);
 624         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 625
 626         if (do_ast) {
 627                 struct lustre_handle lockh;
 628                 int rc;
 629
 630                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 631                 ldlm_lock2handle(lock, &lockh);
 632                 rc = ldlm_cli_cancel(&lockh);
 633                 if (rc < 0)
 634                         CERROR("ldlm_cli_cancel: %d\n", rc);
 635         } else {
 636                 LDLM_DEBUG(lock, "Lock still has references, will be "
 637                            "cancelled later");
 638         }
 639         RETURN(0);
 640 }
 641
 642 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
 643                int *size, int lock)
 644 {
 645         int lmm_size;
 646         int rc = 0;
 647         ENTRY;
 648
 649         if (lock)
 650                 down(&inode->i_sem);
 651         rc = fsfilt_get_md(obd, inode, md, *size);
 652         if (lock)
 653                 up(&inode->i_sem);
 654
 655         if (rc < 0) {
 656                 CERROR("Error %d reading eadata for ino %lu\n",
 657                        rc, inode->i_ino);
 658         } else if (rc > 0) {
 659                 lmm_size = rc;
 660
 661                 if (S_ISREG(inode->i_mode))
 662                         rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
 663                 if (S_ISDIR(inode->i_mode))
 664                         rc = mds_convert_mea_ea(obd, inode, md, lmm_size);
 665
 666                 if (rc == 0) {
 667                         *size = lmm_size;
 668                         rc = lmm_size;
 669                 } else if (rc > 0) {
 670                         *size = rc;
 671                 }
 672         }
 673
 674         RETURN (rc);
 675 }
 676
 677
 678 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
 679  * Call with lock=0 if the caller has already taken the i_sem. */
 680 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
 681                 struct mds_body *body, struct inode *inode, int lock)
 682 {
 683         struct mds_obd *mds = &obd->u.mds;
 684         void *lmm;
 685         int lmm_size;
 686         int rc;
 687         ENTRY;
 688
 689         lmm = lustre_msg_buf(msg, offset, 0);
 690         if (lmm == NULL) {
 691                 /* Some problem with getting eadata when I sized the reply
 692                  * buffer... */
 693                 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
 694                        inode->i_ino);
 695                 RETURN(0);
 696         }
 697         lmm_size = msg->buflens[offset];
 698
 699         /* I don't really like this, but it is a sanity check on the client
 700          * MD request.  However, if the client doesn't know how much space
 701          * to reserve for the MD, it shouldn't be bad to have too much space.
 702          */
 703         if (lmm_size > mds->mds_max_mdsize) {
 704                 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
 705                        inode->i_ino, lmm_size, mds->mds_max_mdsize);
 706                 // RETURN(-EINVAL);
 707         }
 708
 709         rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
 710         if (rc > 0) {
 711                 if (S_ISDIR(inode->i_mode))
 712                         body->valid |= OBD_MD_FLDIREA;
 713                 else
 714                         body->valid |= OBD_MD_FLEASIZE;
 715                 body->eadatasize = lmm_size;
 716                 rc = 0;
 717         }
 718
 719         RETURN(rc);
 720 }
 721
 722 void mds_squash_root(struct mds_obd *mds, struct mds_req_sec_desc *rsd,
 723                      ptl_nid_t *peernid)
 724 {
 725         if (!mds->mds_squash_uid ||
 726             (rsd->rsd_uid && rsd->rsd_fsuid))
 727                 return;
 728
 729         if (*peernid == mds->mds_nosquash_nid)
 730                 return;
 731
 732         CDEBUG(D_OTHER, "squash req from 0x%llx, (%d:%d/%x)=>(%d:%d/%x)\n",
 733                 *peernid, rsd->rsd_fsuid, rsd->rsd_fsgid, rsd->rsd_cap,
 734                 mds->mds_squash_uid, mds->mds_squash_gid,
 735                 (rsd->rsd_cap & ~CAP_FS_MASK));
 736
 737         rsd->rsd_uid = mds->mds_squash_uid;
 738         rsd->rsd_fsuid = mds->mds_squash_uid;
 739         rsd->rsd_fsgid = mds->mds_squash_gid;
 740
 741         /* XXX should we remove all capabilities? */
 742         rsd->rsd_cap &= ~CAP_FS_MASK;
 743 }
 744
 745 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
 746                                 struct ptlrpc_request *req, struct mds_body *reqbody,
 747                                 int reply_off)
 748 {
 749         struct inode *inode = dentry->d_inode;
 750         struct mds_body *body;
 751         int rc = 0;
 752         ENTRY;
 753
 754         if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
 755                 RETURN(-ENOENT);
 756
 757         body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
 758         LASSERT(body != NULL);                 /* caller prepped reply */
 759
 760         if (dentry->d_flags & DCACHE_CROSS_REF) {
 761                 mds_pack_dentry2body(obd, body, dentry,
 762                                      (reqbody->valid & OBD_MD_FID) ? 1 : 0);
 763                 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
 764                        OLID4(&body->id1));
 765                 RETURN(0);
 766         }
 767
 768         mds_pack_inode2body(obd, body, inode,
 769                             (reqbody->valid & OBD_MD_FID) ? 1 : 0);
 770
 771         if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
 772             (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
 773                 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
 774                                  inode, 1);
 775
 776                 /* if we have LOV EA data, the OST holds size, atime, mtime. */
 777                 if (!(body->valid & OBD_MD_FLEASIZE) &&
 778                     !(body->valid & OBD_MD_FLDIREA))
 779                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 780                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
 781         } else if (S_ISLNK(inode->i_mode) &&
 782                    (reqbody->valid & OBD_MD_LINKNAME) != 0) {
 783                 int len = req->rq_repmsg->buflens[reply_off + 1];
 784                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1, 0);
 785
 786                 LASSERT(symname != NULL);       /* caller prepped reply */
 787
 788                 if (!inode->i_op->readlink) {
 789                         rc = -ENOSYS;
 790                 } else {
 791                         rc = inode->i_op->readlink(dentry, symname, len);
 792                         if (rc < 0) {
 793                                 CERROR("readlink failed: %d\n", rc);
 794                         } else if (rc != len - 1) {
 795                                 CERROR("Unexpected readlink rc %d: expecting %d\n",
 796                                         rc, len - 1);
 797                                 rc = -EINVAL;
 798                         } else {
 799                                 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 800                                 body->valid |= OBD_MD_LINKNAME;
 801                                 body->eadatasize = rc + 1;
 802                                 symname[rc] = 0;
 803                                 rc = 0;
 804                         }
 805                 }
 806         }
 807
 808         RETURN(rc);
 809 }
 810
 811 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
 812                                    struct dentry *dentry,
 813                                    int offset)
 814 {
 815         int rc = 0, size[1] = {sizeof(struct mds_body)};
 816         ENTRY;
 817
 818         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 819                 CERROR("failed MDS_GETATTR_PACK test\n");
 820                 req->rq_status = -ENOMEM;
 821                 RETURN(-ENOMEM);
 822         }
 823
 824         rc = lustre_pack_reply(req, 1, size, NULL);
 825         if (rc) {
 826                 CERROR("lustre_pack_reply failed: rc %d\n", rc);
 827                 GOTO(out, req->rq_status = rc);
 828         }
 829
 830         EXIT;
 831 out:
 832         return rc;
 833 }
 834
 835 static int mds_getattr_pack_msg(struct ptlrpc_request *req,
 836                                 struct inode *inode,
 837                                 int offset)
 838 {
 839         struct mds_obd *mds = mds_req2mds(req);
 840         struct mds_body *body;
 841         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 842         ENTRY;
 843
 844         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
 845         LASSERT(body != NULL);                 /* checked by caller */
 846         LASSERT_REQSWABBED(req, offset);       /* swabbed by caller */
 847
 848         if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
 849             (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
 850                 int rc;
 851                 down(&inode->i_sem);
 852                 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
 853                 up(&inode->i_sem);
 854                 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
 855                        rc, inode->i_ino);
 856                 if (rc < 0) {
 857                         if (rc != -ENODATA)
 858                                 CERROR("error getting inode %lu MD: rc = %d\n",
 859                                        inode->i_ino, rc);
 860                         size[bufcount] = 0;
 861                 } else if (rc > mds->mds_max_mdsize) {
 862                         size[bufcount] = 0;
 863                         CERROR("MD size %d larger than maximum possible %u\n",
 864                                rc, mds->mds_max_mdsize);
 865                 } else {
 866                         size[bufcount] = rc;
 867                 }
 868                 bufcount++;
 869         } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
 870                 if (inode->i_size + 1 != body->eadatasize)
 871                         CERROR("symlink size: %Lu, reply space: %d\n",
 872                                inode->i_size + 1, body->eadatasize);
 873                 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
 874                 bufcount++;
 875                 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
 876                        inode->i_size + 1, body->eadatasize);
 877         }
 878
 879         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 880                 CERROR("failed MDS_GETATTR_PACK test\n");
 881                 req->rq_status = -ENOMEM;
 882                 GOTO(out, rc = -ENOMEM);
 883         }
 884
 885         rc = lustre_pack_reply(req, bufcount, size, NULL);
 886         if (rc) {
 887                 CERROR("out of memory\n");
 888                 GOTO(out, req->rq_status = rc);
 889         }
 890
 891         EXIT;
 892  out:
 893         return rc;
 894 }
 895
 896 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
 897                       char *name, int namelen)
 898 {
 899         struct mea *mea = NULL;
 900         int mea_size, rc = 0;
 901         ENTRY;
 902
 903         rc = mds_get_lmv_attr(obd, inode, &mea, &mea_size);
 904         if (rc)
 905                 RETURN(rc);
 906         if (mea != NULL) {
 907                 /*
 908                  * dir is already splitted, check if requested filename should
 909                  * live at this MDS or at another one.
 910                  */
 911                 int i = mea_name2idx(mea, name, namelen - 1);
 912                 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
 913                         CDEBUG(D_OTHER,
 914                                "inapropriate MDS(%d) for %s. should be "
 915                                "%lu(%d)\n", mea->mea_master, name,
 916                                (unsigned long)id_group(&mea->mea_ids[i]), i);
 917                         rc = -ERESTART;
 918                 }
 919         }
 920
 921         if (mea)
 922                 OBD_FREE(mea, mea_size);
 923         RETURN(rc);
 924 }
 925
 926 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
 927                             struct lustre_handle *child_lockh, int child_part)
 928 {
 929         struct obd_device *obd = req->rq_export->exp_obd;
 930         struct mds_obd *mds = &obd->u.mds;
 931         struct ldlm_reply *rep = NULL;
 932         struct lvfs_run_ctxt saved;
 933         struct mds_req_sec_desc *rsd;
 934         struct mds_body *body;
 935         struct dentry *dparent = NULL, *dchild = NULL;
 936         struct lvfs_ucred uc;
 937         struct lustre_handle parent_lockh[2] = {{0}, {0}};
 938         unsigned int namesize;
 939         int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
 940         char *name = NULL;
 941         ENTRY;
 942
 943         LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
 944         MD_COUNTER_INCREMENT(obd, getattr_lock);
 945
 946         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
 947         if (!rsd) {
 948                 CERROR("Can't unpack security desc\n");
 949                 RETURN(-EFAULT);
 950         }
 951         mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
 952
 953         /* swab now, before anyone looks inside the request. */
 954         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
 955                                   lustre_swab_mds_body);
 956         if (body == NULL) {
 957                 CERROR("Can't swab mds_body\n");
 958                 GOTO(cleanup, rc = -EFAULT);
 959         }
 960
 961         LASSERT_REQSWAB(req, offset + 1);
 962         name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
 963         if (name == NULL) {
 964                 CERROR("Can't unpack name\n");
 965                 GOTO(cleanup, rc = -EFAULT);
 966         }
 967         namesize = req->rq_reqmsg->buflens[offset + 1];
 968
 969         /* namesize less than 2 means we have empty name, probably came from
 970            revalidate by cfid, so no point in having name to be set */
 971         if (namesize <= 1)
 972                 name = NULL;
 973
 974         LASSERT (offset == 1 || offset == 3);
 975         if (offset == 3) {
 976                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
 977                 reply_offset = 1;
 978         } else {
 979                 reply_offset = 0;
 980         }
 981
 982         rc = mds_init_ucred(&uc, rsd);
 983         if (rc) {
 984                 CERROR("can't init ucred\n");
 985                 GOTO(cleanup, rc);
 986         }
 987
 988         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
 989         cleanup_phase = 1; /* kernel context */
 990         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
 991
 992         LASSERT(namesize > 0);
 993         if (child_lockh->cookie != 0) {
 994                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
 995                 resent_req = 1;
 996         }
 997 #if 0
 998 #if HAVE_LOOKUP_RAW
 999         if (body->valid == OBD_MD_FLID) {
1000                 struct mds_body *mds_reply;
1001                 int size = sizeof(*mds_reply);
1002                 struct inode *dir;
1003                 ino_t inum;
1004
1005                 dparent = mds_id2dentry(obd, &body->id1, NULL);
1006                 if (IS_ERR(dparent)) {
1007                         rc = PTR_ERR(dparent);
1008                         GOTO(cleanup, rc);
1009                 }
1010
1011                 /*
1012                  * the user requested ONLY the inode number, so do a raw lookup.
1013                  */
1014                 rc = lustre_pack_reply(req, 1, &size, NULL);
1015                 if (rc) {
1016                         CERROR("out of memory\n");
1017                         l_dput(dparent);
1018                         GOTO(cleanup, rc);
1019                 }
1020                 dir  = dparent->d_inode;
1021                 LASSERT(dir->i_op->lookup_raw != NULL);
1022                 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1023                 l_dput(dparent);
1024                 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1025                                            sizeof(*mds_reply));
1026
1027                 id_ino(&mds_reply->id1) = inum;
1028                 mds_reply->valid = OBD_MD_FLID;
1029                 GOTO(cleanup, rc);
1030         }
1031 #endif
1032 #endif
1033         if (resent_req == 0) {
1034                 LASSERT(id_fid(&body->id1) != 0);
1035                 if (name) {
1036                         rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1037                                                          parent_lockh, &dparent,
1038                                                          LCK_PR,
1039                                                          MDS_INODELOCK_UPDATE,
1040                                                          &update_mode,
1041                                                          name, namesize,
1042                                                          child_lockh, &dchild,
1043                                                          LCK_PR, child_part);
1044                         if (rc)
1045                                 GOTO(cleanup, rc);
1046
1047                         /*
1048                          * let's make sure this name should leave on this mds
1049                          * node.
1050                          */
1051                         rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1052                         if (rc)
1053                                 GOTO(cleanup, rc);
1054                 } else {
1055                         /* we have no dentry here, drop LOOKUP bit */
1056                         /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1057                         child_part &= ~MDS_INODELOCK_LOOKUP;
1058                         CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1059                                obd->obd_name, OLID4(&body->id1));
1060
1061                         dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1062                                                       LCK_PR, parent_lockh,
1063                                                       &update_mode,
1064                                                       NULL, 0,
1065                                                       MDS_INODELOCK_UPDATE);
1066                         if (IS_ERR(dchild)) {
1067                                 CERROR("can't find inode with id "DLID4", err = %d\n",
1068                                        OLID4(&body->id1), (int)PTR_ERR(dchild));
1069                                 GOTO(cleanup, rc = PTR_ERR(dchild));
1070                         }
1071                         memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1072                 }
1073         } else {
1074                 struct ldlm_lock *granted_lock;
1075
1076                 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1077                 granted_lock = ldlm_handle2lock(child_lockh);
1078
1079                 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1080                          id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1081                          child_lockh->cookie);
1082
1083                 dparent = mds_id2dentry(obd, &body->id1, NULL);
1084                 LASSERT(dparent);
1085
1086                 dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1087                 LASSERT(dchild);
1088                 LDLM_LOCK_PUT(granted_lock);
1089         }
1090
1091         cleanup_phase = 2; /* dchild, dparent, locks */
1092
1093         if (!DENTRY_VALID(dchild)) {
1094                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1095                 /*
1096                  * in the intent case, the policy clears this error: the
1097                  * disposition is enough.
1098                  */
1099                 rc = -ENOENT;
1100                 GOTO(cleanup, rc);
1101         } else {
1102                 intent_set_disposition(rep, DISP_LOOKUP_POS);
1103         }
1104
1105         if (req->rq_repmsg == NULL) {
1106                 if (dchild->d_flags & DCACHE_CROSS_REF)
1107                         rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1108                 else
1109                         rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
1110                 if (rc != 0) {
1111                         CERROR ("mds_getattr_pack_msg: %d\n", rc);
1112                         GOTO (cleanup, rc);
1113                 }
1114         }
1115
1116         rc = mds_getattr_internal(obd, dchild, req, body, reply_offset);
1117         GOTO(cleanup, rc); /* returns the lock to the client */
1118
1119  cleanup:
1120         switch (cleanup_phase) {
1121         case 2:
1122                 if (resent_req == 0) {
1123                         if (rc && DENTRY_VALID(dchild))
1124                                 ldlm_lock_decref(child_lockh, LCK_PR);
1125                         if (name)
1126                                 ldlm_lock_decref(parent_lockh, LCK_PR);
1127 #ifdef S_PDIROPS
1128                         if (parent_lockh[1].cookie != 0)
1129                                 ldlm_lock_decref(parent_lockh + 1, update_mode);
1130 #endif
1131                         if (dparent)
1132                                 l_dput(dparent);
1133                 }
1134                 l_dput(dchild);
1135         case 1:
1136                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1137                 mds_exit_ucred(&uc);
1138         }
1139         return rc;
1140 }
1141
1142 static int mds_getattr(struct ptlrpc_request *req, int offset)
1143 {
1144         struct obd_device *obd = req->rq_export->exp_obd;
1145         struct lvfs_run_ctxt saved;
1146         struct dentry *de;
1147         struct mds_req_sec_desc *rsd;
1148         struct mds_body *body;
1149         struct lvfs_ucred uc;
1150         int rc = 0;
1151         ENTRY;
1152
1153         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1154         if (!rsd) {
1155                 CERROR("Can't unpack security desc\n");
1156                 RETURN(-EFAULT);
1157         }
1158
1159         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1160                                   lustre_swab_mds_body);
1161         if (body == NULL) {
1162                 CERROR ("Can't unpack body\n");
1163                 RETURN (-EFAULT);
1164         }
1165
1166         MD_COUNTER_INCREMENT(obd, getattr);
1167
1168         rc = mds_init_ucred(&uc, rsd);
1169         if (rc) {
1170                 CERROR("can't init ucred\n");
1171                 RETURN(rc);
1172         }
1173
1174         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1175         de = mds_id2dentry(obd, &body->id1, NULL);
1176         if (IS_ERR(de)) {
1177                 rc = req->rq_status = PTR_ERR(de);
1178                 GOTO(out_pop, rc);
1179         }
1180
1181         rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1182         if (rc != 0) {
1183                 CERROR("mds_getattr_pack_msg: %d\n", rc);
1184                 GOTO(out_pop, rc);
1185         }
1186
1187         req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
1188
1189         l_dput(de);
1190
1191         EXIT;
1192 out_pop:
1193         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1194         mds_exit_ucred(&uc);
1195         return rc;
1196 }
1197
1198 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1199                           unsigned long max_age)
1200 {
1201         int rc;
1202         ENTRY;
1203
1204         spin_lock(&obd->obd_osfs_lock);
1205         rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1206         if (rc == 0)
1207                 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1208         spin_unlock(&obd->obd_osfs_lock);
1209
1210         RETURN(rc);
1211 }
1212
1213 static int mds_statfs(struct ptlrpc_request *req)
1214 {
1215         struct obd_device *obd = req->rq_export->exp_obd;
1216         int rc, size = sizeof(struct obd_statfs);
1217         ENTRY;
1218
1219         rc = lustre_pack_reply(req, 1, &size, NULL);
1220         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1221                 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1222                 GOTO(out, rc);
1223         }
1224
1225         OBD_COUNTER_INCREMENT(obd, statfs);
1226
1227         /* We call this so that we can cache a bit - 1 jiffie worth */
1228         rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1229                             jiffies - HZ);
1230         if (rc) {
1231                 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1232                 GOTO(out, rc);
1233         }
1234
1235         EXIT;
1236 out:
1237         req->rq_status = rc;
1238         return 0;
1239 }
1240
1241 static int mds_sync(struct ptlrpc_request *req, int offset)
1242 {
1243         struct obd_device *obd = req->rq_export->exp_obd;
1244         struct mds_obd *mds = &obd->u.mds;
1245         struct mds_body *body;
1246         int rc, size = sizeof(*body);
1247         ENTRY;
1248
1249         body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1250         if (body == NULL)
1251                 GOTO(out, rc = -EPROTO);
1252
1253         rc = lustre_pack_reply(req, 1, &size, NULL);
1254         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1255                 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1256                 GOTO(out, rc);
1257         }
1258
1259         if (id_ino(&body->id1) == 0) {
1260                 /* an id of zero is taken to mean "sync whole filesystem" */
1261                 rc = fsfilt_sync(obd, mds->mds_sb);
1262                 if (rc)
1263                         GOTO(out, rc);
1264         } else {
1265                 /* just any file to grab fsync method - "file" arg unused */
1266                 struct file *file = mds->mds_rcvd_filp;
1267                 struct dentry *de;
1268
1269                 de = mds_id2dentry(obd, &body->id1, NULL);
1270                 if (IS_ERR(de))
1271                         GOTO(out, rc = PTR_ERR(de));
1272
1273                 rc = file->f_op->fsync(NULL, de, 1);
1274                 if (rc)
1275                         GOTO(out, rc);
1276
1277                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1278                 mds_pack_inode2body(obd, body, de->d_inode, 0);
1279                 l_dput(de);
1280         }
1281
1282         EXIT;
1283 out:
1284         req->rq_status = rc;
1285         return 0;
1286 }
1287
1288 /* mds_readpage does not take a DLM lock on the inode, because the client must
1289  * already have a PR lock.
1290  *
1291  * If we were to take another one here, a deadlock will result, if another
1292  * thread is already waiting for a PW lock. */
1293 static int mds_readpage(struct ptlrpc_request *req, int offset)
1294 {
1295         struct obd_device *obd = req->rq_export->exp_obd;
1296         struct mds_obd *mds = &obd->u.mds;
1297         struct vfsmount *mnt;
1298         struct dentry *de;
1299         struct file *file;
1300         struct mds_req_sec_desc *rsd;
1301         struct mds_body *body, *repbody;
1302         struct lvfs_run_ctxt saved;
1303         int rc, size = sizeof(*repbody);
1304         struct lvfs_ucred uc;
1305         ENTRY;
1306
1307         rc = lustre_pack_reply(req, 1, &size, NULL);
1308         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1309                 CERROR("mds: out of memory\n");
1310                 GOTO(out, rc = -ENOMEM);
1311         }
1312
1313         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1314         if (!rsd) {
1315                 CERROR("Can't unpack security desc\n");
1316                 GOTO (out, rc = -EFAULT);
1317         }
1318         mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
1319
1320         body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1321                                   lustre_swab_mds_body);
1322         if (body == NULL) {
1323                 CERROR("Can't unpack body\n");
1324                 GOTO (out, rc = -EFAULT);
1325         }
1326
1327         rc = mds_init_ucred(&uc, rsd);
1328         if (rc) {
1329                 CERROR("can't init ucred\n");
1330                 GOTO(out, rc);
1331         }
1332
1333         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1334         de = mds_id2dentry(obd, &body->id1, &mnt);
1335         if (IS_ERR(de))
1336                 GOTO(out_pop, rc = PTR_ERR(de));
1337
1338         CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1339
1340         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1341         /* note: in case of an error, dentry_open puts dentry */
1342         if (IS_ERR(file))
1343                 GOTO(out_pop, rc = PTR_ERR(file));
1344
1345         /* body->size is actually the offset -eeb */
1346         if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1347                 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1348                        body->size, de->d_inode->i_blksize);
1349                 GOTO(out_file, rc = -EFAULT);
1350         }
1351
1352         /* body->nlink is actually the #bytes to read -eeb */
1353         if (body->nlink & (de->d_inode->i_blksize - 1)) {
1354                 CERROR("size %u is not multiple of blocksize %lu\n",
1355                        body->nlink, de->d_inode->i_blksize);
1356                 GOTO(out_file, rc = -EFAULT);
1357         }
1358
1359         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1360         repbody->size = file->f_dentry->d_inode->i_size;
1361         repbody->valid = OBD_MD_FLSIZE;
1362
1363         /* to make this asynchronous make sure that the handling function
1364            doesn't send a reply when this function completes. Instead a
1365            callback function would send the reply */
1366         /* body->size is actually the offset -eeb */
1367         rc = mds_sendpage(req, file, body->size, body->nlink);
1368
1369         EXIT;
1370 out_file:
1371         filp_close(file, 0);
1372 out_pop:
1373         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1374         mds_exit_ucred(&uc);
1375 out:
1376         req->rq_status = rc;
1377         return 0;
1378 }
1379
1380 /* update master MDS ID, which is stored in local inode EA. */
1381 int mds_update_mid(struct obd_device *obd, struct lustre_id *id,
1382                    void *data, int data_len)
1383 {
1384         struct mds_obd *mds = &obd->u.mds;
1385         struct dentry *dentry;
1386         void *handle;
1387         int rc = 0;
1388         ENTRY;
1389
1390         LASSERT(id);
1391         LASSERT(obd);
1392
1393         dentry = mds_id2dentry(obd, id, NULL);
1394         if (IS_ERR(dentry))
1395                 GOTO(out, rc = PTR_ERR(dentry));
1396
1397         if (!dentry->d_inode) {
1398                 CERROR("Can't find object "DLID4".\n",
1399                        OLID4(id));
1400                 GOTO(out_dentry, rc = -EINVAL);
1401         }
1402
1403         handle = fsfilt_start(obd, dentry->d_inode,
1404                               FSFILT_OP_SETATTR, NULL);
1405         if (IS_ERR(handle))
1406                 GOTO(out_dentry, rc = PTR_ERR(handle));
1407
1408         rc = mds_update_inode_mid(obd, dentry->d_inode, handle,
1409                                   (struct lustre_id *)data);
1410         if (rc) {
1411                 CERROR("Can't update inode "DLID4" master id, "
1412                        "error = %d.\n", OLID4(id), rc);
1413                 GOTO(out_commit, rc);
1414         }
1415
1416         EXIT;
1417 out_commit:
1418         fsfilt_commit(obd, mds->mds_sb, dentry->d_inode,
1419                       handle, 0);
1420 out_dentry:
1421         l_dput(dentry);
1422 out:
1423         return rc;
1424 }
1425 EXPORT_SYMBOL(mds_update_mid);
1426
1427 /* read master MDS ID, which is stored in local inode EA. */
1428 int mds_read_mid(struct obd_device *obd, struct lustre_id *id,
1429                  void *data, int data_len)
1430 {
1431         struct dentry *dentry;
1432         int rc = 0;
1433         ENTRY;
1434
1435         LASSERT(id);
1436         LASSERT(obd);
1437
1438         dentry = mds_id2dentry(obd, id, NULL);
1439         if (IS_ERR(dentry))
1440                 GOTO(out, rc = PTR_ERR(dentry));
1441
1442         if (!dentry->d_inode) {
1443                 CERROR("Can't find object "DLID4".\n",
1444                        OLID4(id));
1445                 GOTO(out_dentry, rc = -EINVAL);
1446         }
1447
1448         down(&dentry->d_inode->i_sem);
1449         rc = mds_read_inode_mid(obd, dentry->d_inode,
1450                                 (struct lustre_id *)data);
1451         up(&dentry->d_inode->i_sem);
1452         if (rc) {
1453                 CERROR("Can't read inode "DLID4" master id, "
1454                        "error = %d.\n", OLID4(id), rc);
1455                 GOTO(out_dentry, rc);
1456         }
1457
1458         EXIT;
1459 out_dentry:
1460         l_dput(dentry);
1461 out:
1462         return rc;
1463 }
1464 EXPORT_SYMBOL(mds_read_mid);
1465
1466 int mds_reint(struct ptlrpc_request *req, int offset,
1467               struct lustre_handle *lockh)
1468 {
1469         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1470         struct mds_update_record *rec;
1471         struct mds_req_sec_desc *rsd;
1472         int rc;
1473         ENTRY;
1474
1475         OBD_ALLOC(rec, sizeof(*rec));
1476         if (rec == NULL)
1477                 RETURN(-ENOMEM);
1478
1479         rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1480         if (!rsd) {
1481                 CERROR("Can't unpack security desc\n");
1482                 GOTO(out, rc = -EFAULT);
1483         }
1484         mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
1485
1486         rc = mds_update_unpack(req, offset, rec);
1487         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1488                 CERROR("invalid record\n");
1489                 GOTO(out, req->rq_status = -EINVAL);
1490         }
1491
1492         rc = mds_init_ucred(&rec->ur_uc, rsd);
1493         if (rc) {
1494                 CERROR("can't init ucred\n");
1495                 GOTO(out, rc);
1496         }
1497
1498         /* rc will be used to interrupt a for loop over multiple records */
1499         rc = mds_reint_rec(rec, offset, req, lockh);
1500         mds_exit_ucred(&rec->ur_uc);
1501         EXIT;
1502  out:
1503         OBD_FREE(rec, sizeof(*rec));
1504         return rc;
1505 }
1506
1507 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1508                                        struct obd_device *obd, int *process)
1509 {
1510         switch (req->rq_reqmsg->opc) {
1511         case MDS_CONNECT: /* This will never get here, but for completeness. */
1512         case OST_CONNECT: /* This will never get here, but for completeness. */
1513         case MDS_DISCONNECT:
1514         case OST_DISCONNECT:
1515                *process = 1;
1516                RETURN(0);
1517
1518         case MDS_CLOSE:
1519         case MDS_SYNC: /* used in unmounting */
1520         case OBD_PING:
1521         case MDS_REINT:
1522         case LDLM_ENQUEUE:
1523         case OST_CREATE:
1524                 *process = target_queue_recovery_request(req, obd);
1525                 RETURN(0);
1526
1527         default:
1528                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1529                 *process = 0;
1530                 /* XXX what should we set rq_status to here? */
1531                 req->rq_status = -EAGAIN;
1532                 RETURN(ptlrpc_error(req));
1533         }
1534 }
1535
1536 static char *reint_names[] = {
1537         [REINT_SETATTR] "setattr",
1538         [REINT_CREATE]  "create",
1539         [REINT_LINK]    "link",
1540         [REINT_UNLINK]  "unlink",
1541         [REINT_RENAME]  "rename",
1542         [REINT_OPEN]    "open",
1543 };
1544
1545 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER  | \
1546                             OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
1547                             OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
1548                             OBD_MD_FLID)
1549
1550 static void reconstruct_create(struct ptlrpc_request *req)
1551 {
1552         struct mds_export_data *med = &req->rq_export->exp_mds_data;
1553         struct mds_client_data *mcd = med->med_mcd;
1554         struct dentry *dentry;
1555         struct ost_body *body;
1556         struct lustre_id id;
1557         int rc;
1558         ENTRY;
1559
1560         /* copy rc, transno and disp; steal locks */
1561         mds_req_from_mcd(req, mcd);
1562         if (req->rq_status) {
1563                 EXIT;
1564                 return;
1565         }
1566
1567         id_gen(&id) = 0;
1568         id_group(&id) = 0;
1569
1570         id_ino(&id) = mcd->mcd_last_data;
1571         LASSERT(id_ino(&id) != 0);
1572
1573         dentry = mds_id2dentry(req2obd(req), &id, NULL);
1574         if (IS_ERR(dentry)) {
1575                 CERROR("can't find inode "LPU64"\n", id_ino(&id));
1576                 req->rq_status = PTR_ERR(dentry);
1577                 EXIT;
1578                 return;
1579         }
1580
1581         CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
1582               req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
1583               dentry->d_inode->i_generation);
1584
1585         body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1586         obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
1587         body->oa.o_id = dentry->d_inode->i_ino;
1588         body->oa.o_generation = dentry->d_inode->i_generation;
1589         body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1590
1591         down(&dentry->d_inode->i_sem);
1592         rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
1593         up(&dentry->d_inode->i_sem);
1594         if (rc) {
1595                 CERROR("Can't read inode self id, inode %lu, "
1596                        "rc %d\n", dentry->d_inode->i_ino, rc);
1597                 id_fid(&id) = 0;
1598         }
1599
1600         body->oa.o_fid = id_fid(&id);
1601         body->oa.o_mds = id_group(&id);
1602         l_dput(dentry);
1603
1604         EXIT;
1605 }
1606
1607 static int mdt_obj_create(struct ptlrpc_request *req)
1608 {
1609         struct obd_device *obd = req->rq_export->exp_obd;
1610         struct mds_obd *mds = &obd->u.mds;
1611         struct ost_body *body, *repbody;
1612         char idname[LL_ID_NAMELEN];
1613         int size = sizeof(*repbody);
1614         struct inode *parent_inode;
1615         struct lvfs_run_ctxt saved;
1616         int rc, cleanup_phase = 0;
1617         struct dentry *new = NULL;
1618         struct dentry_params dp;
1619         int mealen, flags = 0;
1620         struct lvfs_ucred uc;
1621         struct lustre_id id;
1622         struct mea *mea;
1623         void *handle = NULL;
1624         unsigned long cr_inum = 0;
1625         ENTRY;
1626
1627         DEBUG_REQ(D_HA, req, "create remote object");
1628         parent_inode = mds->mds_unnamed_dir->d_inode;
1629
1630         body = lustre_swab_reqbuf(req, 0, sizeof(*body),
1631                                   lustre_swab_ost_body);
1632         if (body == NULL)
1633                 RETURN(-EFAULT);
1634
1635         rc = lustre_pack_reply(req, 1, &size, NULL);
1636         if (rc)
1637                 RETURN(rc);
1638
1639         MDS_CHECK_RESENT(req, reconstruct_create(req));
1640
1641         /*
1642          * this only serve to inter-mds request, don't need check group database
1643          * here. --ericm.
1644          */
1645         uc.luc_ghash = NULL;
1646         uc.luc_ginfo = NULL;
1647         uc.luc_uid = body->oa.o_uid;
1648         uc.luc_fsuid = body->oa.o_uid;
1649         uc.luc_fsgid = body->oa.o_gid;
1650
1651         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1652         repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
1653
1654         /* in REPLAY case inum should be given (client or other MDS fills it) */
1655         if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
1656             (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
1657                 /*
1658                  * this is re-create request from MDS holding directory name.
1659                  * we have to lookup given ino/gen first. if it exists (good
1660                  * case) then there is nothing to do. if it does not then we
1661                  * have to recreate it.
1662                  */
1663                 id_ino(&id) = body->oa.o_id;
1664                 id_gen(&id) = body->oa.o_generation;
1665
1666                 new = mds_id2dentry(obd, &id, NULL);
1667                 if (!IS_ERR(new) && new->d_inode) {
1668                         struct lustre_id sid;
1669
1670                         CWARN("mkdir() repairing is on its way: %lu/%lu\n",
1671                               (unsigned long)id_ino(&id), (unsigned long)id_gen(&id));
1672
1673                         obdo_from_inode(&repbody->oa, new->d_inode,
1674                                         FILTER_VALID_FLAGS);
1675
1676                         repbody->oa.o_id = new->d_inode->i_ino;
1677                         repbody->oa.o_generation = new->d_inode->i_generation;
1678                         repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1679                         cleanup_phase = 1;
1680
1681                         down(&new->d_inode->i_sem);
1682                         rc = mds_read_inode_sid(obd, new->d_inode, &sid);
1683                         up(&new->d_inode->i_sem);
1684                         if (rc) {
1685                                 CERROR("Can't read inode self id "
1686                                        "inode %lu, rc %d.\n",
1687                                        new->d_inode->i_ino, rc);
1688                                 GOTO(cleanup, rc);
1689                         }
1690
1691                         repbody->oa.o_fid = id_fid(&sid);
1692                         repbody->oa.o_mds = id_group(&sid);
1693                         LASSERT(id_fid(&sid) != 0);
1694
1695                         /*
1696                          * here we could use fid passed in body->oa.o_fid and
1697                          * thus avoid mds_read_inode_sid().
1698                          */
1699                         cr_inum = new->d_inode->i_ino;
1700                         GOTO(cleanup, rc = 0);
1701                 }
1702         }
1703
1704         down(&parent_inode->i_sem);
1705         handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
1706         if (IS_ERR(handle)) {
1707                 up(&parent_inode->i_sem);
1708                 CERROR("fsfilt_start() failed, rc = %d\n",
1709                        (int)PTR_ERR(handle));
1710                 GOTO(cleanup, rc = PTR_ERR(handle));
1711         }
1712         cleanup_phase = 1; /* transaction */
1713
1714 repeat:
1715         rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
1716         new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
1717         if (IS_ERR(new)) {
1718                 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
1719                        obd->obd_name, idname, (int) PTR_ERR(new));
1720                 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
1721                 up(&parent_inode->i_sem);
1722                 RETURN(PTR_ERR(new));
1723         } else if (new->d_inode) {
1724                 CERROR("%s: name exists. repeat\n", obd->obd_name);
1725                 goto repeat;
1726         }
1727
1728         new->d_fsdata = (void *)&dp;
1729         dp.p_inum = 0;
1730         dp.p_ptr = req;
1731
1732         if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
1733             (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
1734                 LASSERT(body->oa.o_id != 0);
1735                 dp.p_inum = body->oa.o_id;
1736                 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
1737                           (unsigned long)body->oa.o_id,
1738                           (unsigned long)body->oa.o_generation);
1739         }
1740
1741         rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
1742         if (rc == 0) {
1743                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
1744                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1745                         new->d_inode->i_generation = body->oa.o_generation;
1746                         mark_inode_dirty(new->d_inode);
1747
1748                         /*
1749                          * avoiding asserts in cache flush case, as
1750                          * @body->oa.o_id should be zero.
1751                          */
1752                         if (body->oa.o_id) {
1753                                 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
1754                                          "BUG 3550: failed to recreate obj "
1755                                          LPU64" -> %lu\n", body->oa.o_id,
1756                                          new->d_inode->i_ino);
1757
1758                                 LASSERTF(body->oa.o_generation ==
1759                                          new->d_inode->i_generation,
1760                                          "BUG 3550: failed to recreate obj/gen "
1761                                          LPU64"/%u -> %lu/%u\n", body->oa.o_id,
1762                                          body->oa.o_generation,
1763                                          new->d_inode->i_ino,
1764                                          new->d_inode->i_generation);
1765                         }
1766                 }
1767
1768                 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
1769                 repbody->oa.o_id = new->d_inode->i_ino;
1770                 repbody->oa.o_generation = new->d_inode->i_generation;
1771                 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
1772
1773                 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
1774                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1775                         id_group(&id) = mds->mds_num;
1776
1777                         LASSERT(body->oa.o_fid != 0);
1778                         id_fid(&id) = body->oa.o_fid;
1779
1780                         LASSERT(body->oa.o_id != 0);
1781                         id_ino(&id) = repbody->oa.o_id;
1782                         id_gen(&id) = repbody->oa.o_generation;
1783
1784                         down(&new->d_inode->i_sem);
1785                         rc = mds_update_inode_sid(obd, new->d_inode, handle, &id);
1786                         up(&new->d_inode->i_sem);
1787
1788                         /*
1789                          * make sure, that fid is up-to-date.
1790                          */
1791                         mds_set_last_fid(obd, id_fid(&id));
1792                 } else {
1793                         /*
1794                          * allocate new sid, as object is created from scratch
1795                          * and this is not replay.
1796                          */
1797                         down(&new->d_inode->i_sem);
1798                         rc = mds_alloc_inode_sid(obd, new->d_inode, handle, &id);
1799                         up(&new->d_inode->i_sem);
1800                 }
1801                 if (rc) {
1802                         CERROR("Can't update lustre ID for inode %lu, "
1803                                "error = %d\n", new->d_inode->i_ino, rc);
1804                         GOTO(cleanup, rc);
1805                 }
1806
1807                 /* initializing o_fid after it is allocated. */
1808                 repbody->oa.o_fid = id_fid(&id);
1809                 repbody->oa.o_mds = id_group(&id);
1810
1811                 rc = fsfilt_del_dir_entry(obd, new);
1812                 up(&parent_inode->i_sem);
1813                 if (rc) {
1814                         CERROR("can't remove name for object: %d\n", rc);
1815                         GOTO(cleanup, rc);
1816                 }
1817
1818                 cleanup_phase = 2; /* created directory object */
1819
1820                 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
1821                        (unsigned long)new->d_inode->i_ino,
1822                        (unsigned long)new->d_inode->i_generation,
1823                        (unsigned)new->d_inode->i_mode);
1824                 cr_inum = new->d_inode->i_ino;
1825         } else {
1826                 up(&parent_inode->i_sem);
1827                 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
1828                 GOTO(cleanup, rc);
1829         }
1830
1831         if (body->oa.o_valid & OBD_MD_FLID) {
1832                 /* this is new object for splitted dir. We have to prevent
1833                  * recursive splitting on it -bzzz */
1834                 mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL);
1835
1836                 OBD_ALLOC(mea, mealen);
1837                 if (mea == NULL)
1838                         GOTO(cleanup, rc = -ENOMEM);
1839
1840                 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
1841                 mea->mea_master = 0;
1842                 mea->mea_count = 0;
1843
1844                 down(&new->d_inode->i_sem);
1845                 rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
1846                 up(&new->d_inode->i_sem);
1847                 if (rc)
1848                         CERROR("fsfilt_set_md() failed, rc = %d\n", rc);
1849
1850                 OBD_FREE(mea, mealen);
1851                 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
1852                        obd->obd_name, new->d_inode->i_ino,
1853                        new->d_inode->i_generation, flags);
1854         } else if (body->oa.o_easize) {
1855                 /* we pass LCK_EX to split routine to signal that we have
1856                  * exclusive access to the directory. simple because nobody
1857                  * knows it already exists -bzzz */
1858                 rc = mds_try_to_split_dir(obd, new, NULL,
1859                                           body->oa.o_easize, LCK_EX);
1860                 if (rc < 0) {
1861                         CERROR("Can't split directory %lu, error = %d.\n",
1862                                new->d_inode->i_ino, rc);
1863                 } else {
1864                         rc = 0;
1865                 }
1866         }
1867
1868         EXIT;
1869 cleanup:
1870         switch (cleanup_phase) {
1871         case 2: /* object has been created, but we'll may want to replay it later */
1872                 if (rc == 0)
1873                         ptlrpc_require_repack(req);
1874         case 1: /* transaction */
1875                 rc = mds_finish_transno(mds, parent_inode, handle,
1876                                         req, rc, cr_inum);
1877         }
1878
1879         l_dput(new);
1880         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1881         mds_put_group_entry(mds, uc.luc_ghash);
1882         return rc;
1883 }
1884
1885 static int mdt_get_info(struct ptlrpc_request *req)
1886 {
1887         struct obd_export *exp = req->rq_export;
1888         int keylen, rc = 0;
1889         char *key;
1890         ENTRY;
1891
1892         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1893         if (key == NULL) {
1894                 DEBUG_REQ(D_HA, req, "no get_info key");
1895                 RETURN(-EFAULT);
1896         }
1897         keylen = req->rq_reqmsg->buflens[0];
1898
1899         if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
1900             (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
1901             (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
1902                 RETURN(-EPROTO);
1903
1904         if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
1905                 struct lustre_id *reply;
1906                 int size = sizeof(*reply);
1907
1908                 rc = lustre_pack_reply(req, 1, &size, NULL);
1909                 if (rc)
1910                         RETURN(rc);
1911
1912                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
1913                 rc = obd_get_info(exp, keylen, key, &size, reply);
1914         } else {
1915                 obd_id *reply;
1916                 int size = sizeof(*reply);
1917
1918                 rc = lustre_pack_reply(req, 1, &size, NULL);
1919                 if (rc)
1920                         RETURN(rc);
1921
1922                 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
1923                 rc = obd_get_info(exp, keylen, key, &size, reply);
1924         }
1925
1926         req->rq_repmsg->status = 0;
1927         RETURN(rc);
1928 }
1929
1930 static int mds_set_info(struct obd_export *exp, __u32 keylen,
1931                         void *key, __u32 vallen, void *val)
1932 {
1933         struct obd_device *obd;
1934         struct mds_obd *mds;
1935         int rc = 0;
1936         ENTRY;
1937
1938         obd = class_exp2obd(exp);
1939         if (obd == NULL) {
1940                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1941                        exp->exp_handle.h_cookie);
1942                 RETURN(-EINVAL);
1943         }
1944
1945         mds = &obd->u.mds;
1946         if (keylen >= strlen("mds_type") &&
1947              memcmp(key, "mds_type", keylen) == 0) {
1948                 int valsize;
1949                 __u32 group;
1950
1951                 CDEBUG(D_IOCTL, "set mds type to %x\n", *(int*)val);
1952
1953                 mds->mds_obd_type = *(int*)val;
1954                 group = FILTER_GROUP_FIRST_MDS + mds->mds_obd_type;
1955                 valsize = sizeof(group);
1956
1957                 /* mds number has been changed, so the corresponding obdfilter
1958                  * exp need to be changed too. */
1959                 rc = obd_set_info(mds->mds_lov_exp, strlen("mds_conn"),
1960                                   "mds_conn", valsize, &group);
1961                 RETURN(rc);
1962         }
1963         CDEBUG(D_IOCTL, "invalid key\n");
1964         RETURN(-EINVAL);
1965 }
1966
1967 static int mdt_set_info(struct ptlrpc_request *req)
1968 {
1969         char *key, *val;
1970         struct obd_export *exp = req->rq_export;
1971         int keylen, rc = 0, vallen;
1972         ENTRY;
1973
1974         key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1975         if (key == NULL) {
1976                 DEBUG_REQ(D_HA, req, "no set_info key");
1977                 RETURN(-EFAULT);
1978         }
1979         keylen = req->rq_reqmsg->buflens[0];
1980
1981         if (keylen == strlen("mds_type") &&
1982             memcmp(key, "mds_type", keylen) == 0) {
1983                 rc = lustre_pack_reply(req, 0, NULL, NULL);
1984                 if (rc)
1985                         RETURN(rc);
1986
1987                 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
1988                 vallen = req->rq_reqmsg->buflens[1];
1989
1990                 rc = obd_set_info(exp, keylen, key, vallen, val);
1991                 req->rq_repmsg->status = 0;
1992                 RETURN(rc);
1993         }
1994         CDEBUG(D_IOCTL, "invalid key\n");
1995         RETURN(-EINVAL);
1996 }
1997
1998 static int mds_msg_check_version(struct lustre_msg *msg)
1999 {
2000         int rc;
2001
2002         switch (msg->opc) {
2003         case MDS_CONNECT:
2004         case MDS_DISCONNECT:
2005         case OBD_PING:
2006                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2007                 if (rc)
2008                         CERROR("bad opc %u version %08x, expecting %08x\n",
2009                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2010                 break;
2011         case MDS_STATFS:
2012         case MDS_GETSTATUS:
2013         case MDS_GETATTR:
2014         case MDS_GETATTR_LOCK:
2015         case MDS_READPAGE:
2016         case MDS_REINT:
2017         case MDS_CLOSE:
2018         case MDS_DONE_WRITING:
2019         case MDS_PIN:
2020         case MDS_SYNC:
2021                 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2022                 if (rc)
2023                         CERROR("bad opc %u version %08x, expecting %08x\n",
2024                                msg->opc, msg->version, LUSTRE_MDS_VERSION);
2025                 break;
2026         case LDLM_ENQUEUE:
2027         case LDLM_CONVERT:
2028         case LDLM_BL_CALLBACK:
2029         case LDLM_CP_CALLBACK:
2030                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2031                 if (rc)
2032                         CERROR("bad opc %u version %08x, expecting %08x\n",
2033                                msg->opc, msg->version, LUSTRE_DLM_VERSION);
2034                 break;
2035         case OBD_LOG_CANCEL:
2036         case LLOG_ORIGIN_HANDLE_OPEN:
2037         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2038         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2039         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2040         case LLOG_ORIGIN_HANDLE_CLOSE:
2041         case LLOG_CATINFO:
2042                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2043                 if (rc)
2044                         CERROR("bad opc %u version %08x, expecting %08x\n",
2045                                msg->opc, msg->version, LUSTRE_LOG_VERSION);
2046                 break;
2047         case OST_CREATE:
2048         case OST_WRITE:
2049         case OST_GET_INFO:
2050         case OST_SET_INFO:
2051                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2052                 if (rc)
2053                         CERROR("bad opc %u version %08x, expecting %08x\n",
2054                                msg->opc, msg->version, LUSTRE_OBD_VERSION);
2055                 break;
2056         default:
2057                 CERROR("MDS unknown opcode %d\n", msg->opc);
2058                 rc = -ENOTSUPP;
2059                 break;
2060         }
2061
2062         return rc;
2063 }
2064
2065 static char str[PTL_NALFMT_SIZE];
2066
2067 int mds_handle(struct ptlrpc_request *req)
2068 {
2069         int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2070         struct obd_device *obd = NULL;
2071         struct mds_obd *mds = NULL; /* quell gcc overwarning */
2072         int rc = 0;
2073         ENTRY;
2074
2075         OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2076
2077         rc = mds_msg_check_version(req->rq_reqmsg);
2078         if (rc) {
2079                 CERROR("MDS drop mal-formed request\n");
2080                 RETURN(rc);
2081         }
2082
2083         LASSERT(current->journal_info == NULL);
2084         /* XXX identical to OST */
2085         if (req->rq_reqmsg->opc != MDS_CONNECT) {
2086                 struct mds_export_data *med;
2087                 int recovering;
2088
2089                 if (req->rq_export == NULL) {
2090                         CERROR("operation %d on unconnected MDS from NID %s\n",
2091                                req->rq_reqmsg->opc,
2092                                ptlrpc_peernid2str(&req->rq_peer, str));
2093                         req->rq_status = -ENOTCONN;
2094                         GOTO(out, rc = -ENOTCONN);
2095                 }
2096
2097                 med = &req->rq_export->exp_mds_data;
2098                 obd = req->rq_export->exp_obd;
2099                 mds = &obd->u.mds;
2100
2101                 /* sanity check: if the xid matches, the request must
2102                  * be marked as a resent or replayed */
2103                 if (req->rq_xid == med->med_mcd->mcd_last_xid) {
2104                         LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
2105                                  (MSG_RESENT | MSG_REPLAY),
2106                                  "rq_xid "LPU64" matches last_xid, "
2107                                  "expected RESENT flag\n",
2108                                  req->rq_xid);
2109                 }
2110                 /* else: note the opposite is not always true; a
2111                  * RESENT req after a failover will usually not match
2112                  * the last_xid, since it was likely never
2113                  * committed. A REPLAYed request will almost never
2114                  * match the last xid, however it could for a
2115                  * committed, but still retained, open. */
2116
2117                 spin_lock_bh(&obd->obd_processing_task_lock);
2118                 recovering = obd->obd_recovering;
2119                 spin_unlock_bh(&obd->obd_processing_task_lock);
2120                 if (recovering) {
2121                         rc = mds_filter_recovery_request(req, obd,
2122                                                          &should_process);
2123                         if (rc || should_process == 0) {
2124                                 RETURN(rc);
2125                         } else if (should_process < 0) {
2126                                 req->rq_status = should_process;
2127                                 rc = ptlrpc_error(req);
2128                                 RETURN(rc);
2129                         }
2130                 }
2131         }
2132
2133         switch (req->rq_reqmsg->opc) {
2134         case MDS_CONNECT:
2135                 DEBUG_REQ(D_INODE, req, "connect");
2136                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
2137                 rc = target_handle_connect(req);
2138                 if (!rc)
2139                         /* Now that we have an export, set mds. */
2140                         mds = mds_req2mds(req);
2141                 break;
2142
2143         case MDS_DISCONNECT:
2144                 DEBUG_REQ(D_INODE, req, "disconnect");
2145                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
2146                 rc = target_handle_disconnect(req);
2147                 req->rq_status = rc;            /* superfluous? */
2148                 break;
2149
2150         case MDS_GETSTATUS:
2151                 DEBUG_REQ(D_INODE, req, "getstatus");
2152                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
2153                 rc = mds_getstatus(req);
2154                 break;
2155
2156         case MDS_GETATTR:
2157                 DEBUG_REQ(D_INODE, req, "getattr");
2158                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
2159                 rc = mds_getattr(req, MDS_REQ_REC_OFF);
2160                 break;
2161
2162         case MDS_GETATTR_LOCK: {
2163                 struct lustre_handle lockh;
2164                 DEBUG_REQ(D_INODE, req, "getattr_lock");
2165                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
2166
2167                 /* If this request gets a reconstructed reply, we won't be
2168                  * acquiring any new locks in mds_getattr_lock, so we don't
2169                  * want to cancel.
2170                  */
2171                 lockh.cookie = 0;
2172                 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
2173                                       MDS_INODELOCK_UPDATE);
2174                 /* this non-intent call (from an ioctl) is special */
2175                 req->rq_status = rc;
2176                 if (rc == 0 && lockh.cookie)
2177                         ldlm_lock_decref(&lockh, LCK_PR);
2178                 break;
2179         }
2180         case MDS_STATFS:
2181                 DEBUG_REQ(D_INODE, req, "statfs");
2182                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
2183                 rc = mds_statfs(req);
2184                 break;
2185
2186         case MDS_READPAGE:
2187                 DEBUG_REQ(D_INODE, req, "readpage");
2188                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
2189                 rc = mds_readpage(req, MDS_REQ_REC_OFF);
2190
2191                 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
2192                         if (req->rq_reply_state) {
2193                                 lustre_free_reply_state (req->rq_reply_state);
2194                                 req->rq_reply_state = NULL;
2195                         }
2196                         RETURN(0);
2197                 }
2198
2199                 break;
2200         case MDS_REINT: {
2201                 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
2202                                              sizeof (*opcp));
2203                 __u32  opc;
2204                 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
2205                                mds->mds_max_cookiesize};
2206                 int bufcount;
2207
2208                 /* NB only peek inside req now; mds_reint() will swab it */
2209                 if (opcp == NULL) {
2210                         CERROR ("Can't inspect opcode\n");
2211                         rc = -EINVAL;
2212                         break;
2213                 }
2214                 opc = *opcp;
2215                 if (lustre_msg_swabbed (req->rq_reqmsg))
2216                         __swab32s(&opc);
2217
2218                 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
2219                           (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
2220                            reint_names[opc] == NULL) ? reint_names[opc] :
2221                                                        "unknown opcode");
2222
2223                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
2224
2225                 if (opc == REINT_UNLINK || opc == REINT_RENAME)
2226                         bufcount = 3;
2227                 else if (opc == REINT_OPEN)
2228                         bufcount = 2;
2229                 else
2230                         bufcount = 1;
2231
2232                 rc = lustre_pack_reply(req, bufcount, size, NULL);
2233                 if (rc)
2234                         break;
2235
2236                 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
2237                 fail = OBD_FAIL_MDS_REINT_NET_REP;
2238                 break;
2239         }
2240
2241         case MDS_CLOSE:
2242                 DEBUG_REQ(D_INODE, req, "close");
2243                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
2244                 rc = mds_close(req, MDS_REQ_REC_OFF);
2245                 break;
2246
2247         case MDS_DONE_WRITING:
2248                 DEBUG_REQ(D_INODE, req, "done_writing");
2249                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
2250                 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
2251                 break;
2252
2253         case MDS_PIN:
2254                 DEBUG_REQ(D_INODE, req, "pin");
2255                 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
2256                 rc = mds_pin(req, MDS_REQ_REC_OFF);
2257                 break;
2258
2259         case MDS_SYNC:
2260                 DEBUG_REQ(D_INODE, req, "sync");
2261                 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
2262                 rc = mds_sync(req, MDS_REQ_REC_OFF);
2263                 break;
2264
2265         case OBD_PING:
2266                 DEBUG_REQ(D_INODE, req, "ping");
2267                 rc = target_handle_ping(req);
2268                 break;
2269
2270         case OBD_LOG_CANCEL:
2271                 CDEBUG(D_INODE, "log cancel\n");
2272                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
2273                 rc = -ENOTSUPP; /* la la la */
2274                 break;
2275
2276         case LDLM_ENQUEUE:
2277                 DEBUG_REQ(D_INODE, req, "enqueue");
2278                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
2279                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2280                                          ldlm_server_blocking_ast, NULL);
2281                 fail = OBD_FAIL_LDLM_REPLY;
2282                 break;
2283         case LDLM_CONVERT:
2284                 DEBUG_REQ(D_INODE, req, "convert");
2285                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
2286                 rc = ldlm_handle_convert(req);
2287                 break;
2288         case LDLM_BL_CALLBACK:
2289         case LDLM_CP_CALLBACK:
2290                 DEBUG_REQ(D_INODE, req, "callback");
2291                 CERROR("callbacks should not happen on MDS\n");
2292                 LBUG();
2293                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
2294                 break;
2295         case LLOG_ORIGIN_HANDLE_OPEN:
2296                 DEBUG_REQ(D_INODE, req, "llog_init");
2297                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2298                 rc = llog_origin_handle_open(req);
2299                 break;
2300         case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2301                 DEBUG_REQ(D_INODE, req, "llog next block");
2302                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2303                 rc = llog_origin_handle_next_block(req);
2304                 break;
2305         case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2306                 DEBUG_REQ(D_INODE, req, "llog prev block");
2307                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2308                 rc = llog_origin_handle_prev_block(req);
2309                 break;
2310         case LLOG_ORIGIN_HANDLE_READ_HEADER:
2311                 DEBUG_REQ(D_INODE, req, "llog read header");
2312                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2313                 rc = llog_origin_handle_read_header(req);
2314                 break;
2315         case LLOG_ORIGIN_HANDLE_CLOSE:
2316                 DEBUG_REQ(D_INODE, req, "llog close");
2317                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2318                 rc = llog_origin_handle_close(req);
2319                 break;
2320         case OST_CREATE:
2321                 DEBUG_REQ(D_INODE, req, "ost_create");
2322                 rc = mdt_obj_create(req);
2323                 break;
2324         case OST_GET_INFO:
2325                 DEBUG_REQ(D_INODE, req, "get_info");
2326                 rc = mdt_get_info(req);
2327                 break;
2328         case OST_SET_INFO:
2329                 DEBUG_REQ(D_INODE, req, "set_info");
2330                 rc = mdt_set_info(req);
2331                 break;
2332         case OST_WRITE:
2333                 CDEBUG(D_INODE, "write\n");
2334                 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
2335                 rc = ost_brw_write(req, NULL);
2336                 LASSERT(current->journal_info == NULL);
2337                 /* mdt_brw sends its own replies */
2338                 RETURN(rc);
2339                 break;
2340         case LLOG_CATINFO:
2341                 DEBUG_REQ(D_INODE, req, "llog catinfo");
2342                 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2343                 rc = llog_catinfo(req);
2344                 break;
2345         default:
2346                 req->rq_status = -ENOTSUPP;
2347                 rc = ptlrpc_error(req);
2348                 RETURN(rc);
2349         }
2350
2351         LASSERT(current->journal_info == NULL);
2352
2353         EXIT;
2354
2355         /* If we're DISCONNECTing, the mds_export_data is already freed */
2356         if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
2357                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2358                 struct obd_device *obd = list_entry(mds, struct obd_device,
2359                                                     u.mds);
2360                 req->rq_repmsg->last_xid =
2361                         le64_to_cpu(med->med_mcd->mcd_last_xid);
2362
2363                 if (!obd->obd_no_transno) {
2364                         req->rq_repmsg->last_committed =
2365                                 obd->obd_last_committed;
2366                 } else {
2367                         DEBUG_REQ(D_IOCTL, req,
2368                                   "not sending last_committed update");
2369                 }
2370                 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
2371                        ", xid "LPU64"\n",
2372                        mds->mds_last_transno, obd->obd_last_committed,
2373                        req->rq_xid);
2374         }
2375  out:
2376
2377         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
2378                 if (obd && obd->obd_recovering) {
2379                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
2380                         return target_queue_final_reply(req, rc);
2381                 }
2382                 /* Lost a race with recovery; let the error path DTRT. */
2383                 rc = req->rq_status = -ENOTCONN;
2384         }
2385
2386         target_send_reply(req, rc, fail);
2387         return 0;
2388 }
2389
2390 /* Update the server data on disk.  This stores the new mount_count and also the
2391  * last_rcvd value to disk.  If we don't have a clean shutdown, then the server
2392  * last_rcvd value may be less than that of the clients.  This will alert us
2393  * that we may need to do client recovery.
2394  *
2395  * Also assumes for mds_last_transno that we are not modifying it (no locking).
2396  */
2397 int mds_update_server_data(struct obd_device *obd, int force_sync)
2398 {
2399         struct mds_obd *mds = &obd->u.mds;
2400         struct mds_server_data *msd = mds->mds_server_data;
2401         struct file *filp = mds->mds_rcvd_filp;
2402         struct lvfs_run_ctxt saved;
2403         loff_t off = 0;
2404         int rc;
2405         ENTRY;
2406
2407         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2408         msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
2409
2410         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
2411                mds->mds_mount_count, mds->mds_last_transno);
2412         rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
2413         if (rc)
2414                 CERROR("error writing MDS server data: rc = %d\n", rc);
2415         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2416
2417         RETURN(rc);
2418 }
2419
2420 /* saves last allocated fid counter to file. */
2421 int mds_update_last_fid(struct obd_device *obd, void *handle,
2422                         int force_sync)
2423 {
2424         struct mds_obd *mds = &obd->u.mds;
2425         struct file *filp = mds->mds_fid_filp;
2426         struct lvfs_run_ctxt saved;
2427         loff_t off = 0;
2428         __u64 last_fid;
2429         int rc = 0;
2430         ENTRY;
2431
2432         spin_lock(&mds->mds_last_fid_lock);
2433         last_fid = mds->mds_last_fid;
2434         spin_unlock(&mds->mds_last_fid_lock);
2435
2436         CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
2437                last_fid);
2438
2439         if (handle) {
2440                 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
2441                                       handle, mds_commit_last_fid_cb,
2442                                       NULL);
2443         }
2444
2445         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2446         rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
2447                                  &off, force_sync);
2448         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2449
2450         if (rc) {
2451                 CERROR("error writing MDS last_fid #"LPU64
2452                        ", err = %d\n", last_fid, rc);
2453                 RETURN(rc);
2454         }
2455
2456         CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
2457                "%llu: err = %d\n", last_fid, off, rc);
2458
2459         RETURN(rc);
2460 }
2461
2462 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
2463 {
2464         struct mds_obd *mds = &obd->u.mds;
2465
2466         spin_lock(&mds->mds_last_fid_lock);
2467         if (fid > mds->mds_last_fid)
2468                 mds->mds_last_fid = fid;
2469         spin_unlock(&mds->mds_last_fid_lock);
2470 }
2471
2472 void mds_commit_last_transno_cb(struct obd_device *obd,
2473                                 __u64 transno, void *data,
2474                                 int error)
2475 {
2476         obd_transno_commit_cb(obd, transno, error);
2477 }
2478
2479 void mds_commit_last_fid_cb(struct obd_device *obd,
2480                             __u64 fid, void *data,
2481                             int error)
2482 {
2483         if (error) {
2484                 CERROR("%s: fid "LPD64" commit error: %d\n",
2485                        obd->obd_name, fid, error);
2486                 return;
2487         }
2488
2489         CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
2490                obd->obd_name, fid);
2491 }
2492
2493 __u64 mds_alloc_fid(struct obd_device *obd)
2494 {
2495         struct mds_obd *mds = &obd->u.mds;
2496         __u64 fid;
2497
2498         spin_lock(&mds->mds_last_fid_lock);
2499         fid = ++mds->mds_last_fid;
2500         spin_unlock(&mds->mds_last_fid_lock);
2501
2502         return fid;
2503 }
2504
2505 /*
2506  * allocates new lustre_id on passed @inode and saves it to inode EA.
2507  */
2508 int mds_alloc_inode_sid(struct obd_device *obd, struct inode *inode,
2509                         void *handle, struct lustre_id *id)
2510 {
2511         struct mds_obd *mds = &obd->u.mds;
2512         int alloc = 0, rc = 0;
2513         ENTRY;
2514
2515         LASSERT(obd != NULL);
2516         LASSERT(inode != NULL);
2517
2518         if (id == NULL) {
2519                 OBD_ALLOC(id, sizeof(*id));
2520                 if (id == NULL)
2521                         RETURN(-ENOMEM);
2522                 alloc = 1;
2523         }
2524
2525         id_group(id) = mds->mds_num;
2526         id_fid(id) = mds_alloc_fid(obd);
2527         id_ino(id) = inode->i_ino;
2528         id_gen(id) = inode->i_generation;
2529         id_type(id) = (S_IFMT & inode->i_mode);
2530
2531         rc = mds_update_inode_sid(obd, inode, handle, id);
2532         if (rc) {
2533                 CERROR("Can't update inode FID EA, "
2534                        "rc = %d\n", rc);
2535         }
2536
2537         if (alloc)
2538                 OBD_FREE(id, sizeof(*id));
2539         RETURN(rc);
2540 }
2541
2542 /*
2543  * reads inode self id from inode EA. Probably later this should be replaced by
2544  * caching inode self id to avoid raeding it every time it is needed.
2545  */
2546 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
2547                        struct lustre_id *id)
2548 {
2549         int rc;
2550         ENTRY;
2551
2552         LASSERT(id != NULL);
2553         LASSERT(obd != NULL);
2554         LASSERT(inode != NULL);
2555
2556         rc = fsfilt_get_sid(obd, inode, &id->li_fid,
2557                             sizeof(id->li_fid));
2558         if (rc < 0) {
2559                 CERROR("fsfilt_get_sid() failed, "
2560                        "rc = %d\n", rc);
2561                 RETURN(rc);
2562         } else if (!rc) {
2563                 rc = -ENODATA;
2564                 RETURN(rc);
2565         } else {
2566                 rc = 0;
2567         }
2568
2569         RETURN(rc);
2570 }
2571
2572 /* updates inode self id in EA. */
2573 int mds_update_inode_sid(struct obd_device *obd, struct inode *inode,
2574                          void *handle, struct lustre_id *id)
2575 {
2576         int rc = 0;
2577         ENTRY;
2578
2579         LASSERT(id != NULL);
2580         LASSERT(obd != NULL);
2581         LASSERT(inode != NULL);
2582
2583         rc = fsfilt_set_sid(obd, inode, handle, &id->li_fid,
2584                             sizeof(id->li_fid));
2585         if (rc) {
2586                 CERROR("fsfilt_set_sid() failed, rc = %d\n", rc);
2587                 RETURN(rc);
2588         }
2589
2590         RETURN(rc);
2591 }
2592
2593 /*
2594  * reads inode id on master MDS. This is usualy done by CMOBD to update requests
2595  * to master MDS by correct store cookie, needed to find inode on master MDS
2596  * quickly.
2597  */
2598 int mds_read_inode_mid(struct obd_device *obd, struct inode *inode,
2599                        struct lustre_id *id)
2600 {
2601         int rc;
2602         ENTRY;
2603
2604         LASSERT(id != NULL);
2605         LASSERT(obd != NULL);
2606         LASSERT(inode != NULL);
2607
2608         rc = fsfilt_get_mid(obd, inode, id, sizeof(*id));
2609         if (rc < 0) {
2610                 CERROR("fsfilt_get_mid() failed, "
2611                        "rc = %d\n", rc);
2612                 RETURN(rc);
2613         } else if (!rc) {
2614                 rc = -ENODATA;
2615                 RETURN(rc);
2616         } else {
2617                 rc = 0;
2618         }
2619
2620         RETURN(rc);
2621 }
2622
2623 /*
2624  * updates master inode id. Usualy this is done by CMOBD after an inode is
2625  * created and relationship between cache MDS and master one should be
2626  * established.
2627  */
2628 int mds_update_inode_mid(struct obd_device *obd, struct inode *inode,
2629                          void *handle, struct lustre_id *id)
2630 {
2631         int rc = 0;
2632         ENTRY;
2633
2634         LASSERT(id != NULL);
2635         LASSERT(obd != NULL);
2636         LASSERT(inode != NULL);
2637
2638         rc = fsfilt_set_mid(obd, inode, handle, id, sizeof(*id));
2639         if (rc) {
2640                 CERROR("fsfilt_set_mid() failed, rc = %d\n", rc);
2641                 RETURN(rc);
2642         }
2643
2644         RETURN(rc);
2645 }
2646
2647 /* mount the file system (secretly) */
2648 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
2649 {
2650         struct lustre_cfg* lcfg = buf;
2651         struct mds_obd *mds = &obd->u.mds;
2652         char *options = NULL;
2653         struct vfsmount *mnt;
2654         char ns_name[48];
2655         unsigned long page;
2656         int rc = 0;
2657         ENTRY;
2658
2659         dev_clear_rdonly(2);
2660
2661         if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
2662                 RETURN(rc = -EINVAL);
2663
2664         obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
2665         if (IS_ERR(obd->obd_fsops))
2666                 RETURN(rc = PTR_ERR(obd->obd_fsops));
2667
2668         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
2669
2670         page = __get_free_page(GFP_KERNEL);
2671         if (!page)
2672                 RETURN(-ENOMEM);
2673
2674         options = (char *)page;
2675         memset(options, 0, PAGE_SIZE);
2676
2677         /*
2678          * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
2679          * and the rest of options are passed by mount options. Probably this
2680          * should be moved to somewhere else like startup scripts or lconf.
2681          */
2682         sprintf(options, "iopen_nopriv");
2683
2684         if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
2685                 sprintf(options + strlen(options), ",%s",
2686                         lcfg->lcfg_inlbuf4);
2687
2688         /* we have to know mdsnum before touching underlying fs -bzzz */
2689         sema_init(&mds->mds_lmv_sem, 1);
2690         mds->mds_lmv_connected = 0;
2691         mds->mds_lmv_name = NULL;
2692
2693         if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
2694             strncmp(lcfg->lcfg_inlbuf5, "dumb", lcfg->lcfg_inllen5)) {
2695                 class_uuid_t uuid;
2696
2697                 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
2698                        obd->obd_name, lcfg->lcfg_inlbuf5);
2699
2700                 generate_random_uuid(uuid);
2701                 class_uuid_unparse(uuid, &mds->mds_lmv_uuid);
2702
2703                 OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen5);
2704                 if (mds->mds_lmv_name == NULL)
2705                         RETURN(rc = -ENOMEM);
2706
2707                 memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf5,
2708                        lcfg->lcfg_inllen5);
2709
2710                 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
2711                 if (rc) {
2712                         OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen5);
2713                         GOTO(err_ops, rc);
2714                 }
2715         }
2716
2717         mds->mds_obd_type = MDS_MASTER_OBD;
2718
2719         if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 &&
2720             strncmp(lcfg->lcfg_inlbuf6, "dumb", lcfg->lcfg_inllen6)) {
2721                 if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master"))) {
2722                         mds->mds_obd_type = MDS_MASTER_OBD;
2723                 } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache"))) {
2724                         mds->mds_obd_type = MDS_CACHE_OBD;
2725                 }
2726         }
2727
2728         mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0, lcfg->lcfg_inlbuf1, options);
2729         free_page(page);
2730
2731         if (IS_ERR(mnt)) {
2732                 rc = PTR_ERR(mnt);
2733                 CERROR("do_kern_mount failed: rc = %d\n", rc);
2734                 GOTO(err_ops, rc);
2735         }
2736
2737         CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
2738
2739         sema_init(&mds->mds_epoch_sem, 1);
2740         atomic_set(&mds->mds_real_clients, 0);
2741         spin_lock_init(&mds->mds_transno_lock);
2742         spin_lock_init(&mds->mds_last_fid_lock);
2743         sema_init(&mds->mds_orphan_recovery_sem, 1);
2744         mds->mds_max_cookiesize = sizeof(struct llog_cookie);
2745
2746         sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
2747         obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
2748
2749         if (obd->obd_namespace == NULL) {
2750                 mds_cleanup(obd, 0);
2751                 GOTO(err_put, rc = -ENOMEM);
2752         }
2753         ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
2754
2755         rc = mds_fs_setup(obd, mnt);
2756         if (rc) {
2757                 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
2758                        obd->obd_name, rc);
2759                 GOTO(err_ns, rc);
2760         }
2761
2762         rc = llog_start_commit_thread();
2763         if (rc < 0)
2764                 GOTO(err_fs, rc);
2765
2766         if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
2767             strncmp(lcfg->lcfg_inlbuf3, "dumb", lcfg->lcfg_inllen3)) {
2768                 class_uuid_t uuid;
2769
2770                 generate_random_uuid(uuid);
2771                 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
2772
2773                 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
2774                 if (mds->mds_profile == NULL)
2775                         GOTO(err_fs, rc = -ENOMEM);
2776
2777                 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
2778                        lcfg->lcfg_inllen3);
2779         }
2780
2781         /*
2782          * setup root dir and files ID dir if lmv already connected, or there is
2783          * not lmv at all.
2784          */
2785         if (mds->mds_lmv_exp || (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
2786                                  strncmp(lcfg->lcfg_inlbuf3, "dumb", lcfg->lcfg_inllen3)))
2787         {
2788                 rc = mds_fs_setup_rootid(obd);
2789                 if (rc)
2790                         GOTO(err_fs, rc);
2791
2792                 rc = mds_fs_setup_virtid(obd);
2793                 if (rc)
2794                         GOTO(err_fs, rc);
2795         }
2796
2797         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2798                            "mds_ldlm_client", &obd->obd_ldlm_client);
2799         obd->obd_replayable = 1;
2800
2801         rc = mds_postsetup(obd);
2802         if (rc)
2803                 GOTO(err_fs, rc);
2804
2805         RETURN(0);
2806
2807 err_fs:
2808         /* No extra cleanup needed for llog_init_commit_thread() */
2809         mds_fs_cleanup(obd, 0);
2810 err_ns:
2811         ldlm_namespace_free(obd->obd_namespace, 0);
2812         obd->obd_namespace = NULL;
2813 err_put:
2814         unlock_kernel();
2815         mntput(mds->mds_vfsmnt);
2816         mds->mds_sb = 0;
2817         lock_kernel();
2818 err_ops:
2819         fsfilt_put_ops(obd->obd_fsops);
2820         return rc;
2821 }
2822
2823 static int mds_postsetup(struct obd_device *obd)
2824 {
2825         struct mds_obd *mds = &obd->u.mds;
2826         int rc = 0;
2827         ENTRY;
2828
2829         rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
2830                             obd, 0, NULL, &llog_lvfs_ops);
2831         if (rc)
2832                 RETURN(rc);
2833
2834         if (mds->mds_profile) {
2835                 struct llog_ctxt *lgctxt;
2836                 struct lvfs_run_ctxt saved;
2837                 struct lustre_profile *lprof;
2838                 struct config_llog_instance cfg;
2839
2840                 cfg.cfg_instance = NULL;
2841                 cfg.cfg_uuid = mds->mds_lov_uuid;
2842                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2843
2844                 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
2845                 if (!lgctxt) {
2846                         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2847                         GOTO(err_llog, rc = -EINVAL);
2848                 }
2849
2850                 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
2851                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2852
2853                 if (rc)
2854                         GOTO(err_llog, rc);
2855
2856                 lprof = class_get_profile(mds->mds_profile);
2857                 if (lprof == NULL) {
2858                         CERROR("No profile found: %s\n", mds->mds_profile);
2859                         GOTO(err_cleanup, rc = -ENOENT);
2860                 }
2861                 rc = mds_lov_connect(obd, lprof->lp_lov);
2862                 if (rc)
2863                         GOTO(err_cleanup, rc);
2864
2865                 rc = mds_lmv_postsetup(obd);
2866                 if (rc)
2867                         GOTO(err_cleanup, rc);
2868         }
2869
2870         RETURN(rc);
2871 err_cleanup:
2872         mds_lov_clean(obd);
2873 err_llog:
2874         obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
2875                                           LLOG_CONFIG_ORIG_CTXT));
2876         return rc;
2877 }
2878
2879 int mds_postrecov(struct obd_device *obd)
2880 {
2881         struct mds_obd *mds = &obd->u.mds;
2882         struct llog_ctxt *ctxt;
2883         int rc, item = 0, valsize;
2884          __u32 group;
2885         ENTRY;
2886
2887         LASSERT(!obd->obd_recovering);
2888         ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
2889         LASSERT(ctxt != NULL);
2890
2891         /* set nextid first, so we are sure it happens */
2892         rc = mds_lov_set_nextid(obd);
2893         if (rc) {
2894                 CERROR("%s: mds_lov_set_nextid failed\n", obd->obd_name);
2895                 GOTO(out, rc);
2896         }
2897
2898         /* clean PENDING dir */
2899         rc = mds_cleanup_orphans(obd);
2900         if (rc < 0)
2901                 GOTO(out, rc);
2902         item = rc;
2903
2904         group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
2905         valsize = sizeof(group);
2906         rc = obd_set_info(mds->mds_lov_exp, strlen("mds_conn"), "mds_conn",
2907                           valsize, &group);
2908         if (rc)
2909                 GOTO(out, rc);
2910
2911         rc = llog_connect(ctxt, obd->u.mds.mds_lov_desc.ld_tgt_count,
2912                           NULL, NULL, NULL);
2913         if (rc) {
2914                 CERROR("%s: failed at llog_origin_connect: %d\n",
2915                        obd->obd_name, rc);
2916                 GOTO(out, rc);
2917         }
2918
2919         /* remove the orphaned precreated objects */
2920         rc = mds_lov_clearorphans(mds, NULL /* all OSTs */);
2921         if (rc)
2922                 GOTO(err_llog, rc);
2923
2924 out:
2925         RETURN(rc < 0 ? rc : item);
2926
2927 err_llog:
2928         /* cleanup all llogging subsystems */
2929         rc = obd_llog_finish(obd, &obd->obd_llogs,
2930                              mds->mds_lov_desc.ld_tgt_count);
2931         if (rc)
2932                 CERROR("%s: failed to cleanup llogging subsystems\n",
2933                         obd->obd_name);
2934         goto out;
2935 }
2936
2937 int mds_lov_clean(struct obd_device *obd)
2938 {
2939         struct mds_obd *mds = &obd->u.mds;
2940         ENTRY;
2941
2942         if (mds->mds_profile) {
2943                 char * cln_prof;
2944                 struct llog_ctxt *llctx;
2945                 struct lvfs_run_ctxt saved;
2946                 struct config_llog_instance cfg;
2947                 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
2948
2949                 OBD_ALLOC(cln_prof, len);
2950                 sprintf(cln_prof, "%s-clean", mds->mds_profile);
2951
2952                 cfg.cfg_instance = NULL;
2953                 cfg.cfg_uuid = mds->mds_lov_uuid;
2954
2955                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2956                 llctx = llog_get_context(&obd->obd_llogs,
2957                                          LLOG_CONFIG_ORIG_CTXT);
2958                 class_config_process_llog(llctx, cln_prof, &cfg);
2959                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2960
2961                 OBD_FREE(cln_prof, len);
2962                 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2963                 mds->mds_profile = NULL;
2964         }
2965         RETURN(0);
2966 }
2967
2968 int mds_lmv_clean(struct obd_device *obd)
2969 {
2970         struct mds_obd *mds = &obd->u.mds;
2971         ENTRY;
2972
2973         if (mds->mds_lmv_name) {
2974                 OBD_FREE(mds->mds_lmv_name, strlen(mds->mds_lmv_name) + 1);
2975                 mds->mds_lmv_name = NULL;
2976         }
2977         RETURN(0);
2978 }
2979
2980 static int mds_precleanup(struct obd_device *obd, int flags)
2981 {
2982         int rc = 0;
2983         ENTRY;
2984
2985         mds_lmv_clean(obd);
2986         mds_lov_disconnect(obd, flags);
2987         mds_lov_clean(obd);
2988         obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
2989         RETURN(rc);
2990 }
2991
2992 static int mds_cleanup(struct obd_device *obd, int flags)
2993 {
2994         struct mds_obd *mds = &obd->u.mds;
2995         ENTRY;
2996
2997         if (mds->mds_sb == NULL)
2998                 RETURN(0);
2999
3000         mds_update_server_data(obd, 1);
3001         mds_update_last_fid(obd, NULL, 1);
3002
3003         if (mds->mds_lov_objids != NULL) {
3004                 int size = mds->mds_lov_desc.ld_tgt_count *
3005                         sizeof(obd_id);
3006                 OBD_FREE(mds->mds_lov_objids, size);
3007         }
3008         mds_fs_cleanup(obd, flags);
3009
3010         unlock_kernel();
3011
3012         /*
3013          * 2 seems normal on mds, (may_umount() also expects 2 fwiw), but we
3014          * only see 1 at this point in obdfilter.
3015          */
3016         if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
3017                 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
3018                        atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
3019
3020         mntput(mds->mds_vfsmnt);
3021         mds->mds_sb = 0;
3022
3023         ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3024
3025         spin_lock_bh(&obd->obd_processing_task_lock);
3026         if (obd->obd_recovering) {
3027                 target_cancel_recovery_timer(obd);
3028                 obd->obd_recovering = 0;
3029         }
3030         spin_unlock_bh(&obd->obd_processing_task_lock);
3031
3032         lock_kernel();
3033         dev_clear_rdonly(2);
3034         fsfilt_put_ops(obd->obd_fsops);
3035
3036         RETURN(0);
3037 }
3038
3039 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
3040                                         int offset,
3041                                         struct ldlm_lock *new_lock,
3042                                         struct ldlm_lock **old_lock,
3043                                         struct lustre_handle *lockh)
3044 {
3045         struct obd_export *exp = req->rq_export;
3046         struct obd_device *obd = exp->exp_obd;
3047         struct ldlm_request *dlmreq =
3048                 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
3049         struct lustre_handle remote_hdl = dlmreq->lock_handle1;
3050         struct list_head *iter;
3051
3052         if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
3053                 return;
3054
3055         l_lock(&obd->obd_namespace->ns_lock);
3056         list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
3057                 struct ldlm_lock *lock;
3058                 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
3059                 if (lock == new_lock)
3060                         continue;
3061                 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
3062                         lockh->cookie = lock->l_handle.h_cookie;
3063                         LDLM_DEBUG(lock, "restoring lock cookie");
3064                         DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
3065                                   lockh->cookie);
3066                         if (old_lock)
3067                                 *old_lock = LDLM_LOCK_GET(lock);
3068                         l_unlock(&obd->obd_namespace->ns_lock);
3069                         return;
3070                 }
3071         }
3072         l_unlock(&obd->obd_namespace->ns_lock);
3073
3074         /* If the xid matches, then we know this is a resent request,
3075          * and allow it. (It's probably an OPEN, for which we don't
3076          * send a lock */
3077         if (req->rq_xid == exp->exp_mds_data.med_mcd->mcd_last_xid)
3078                 return;
3079
3080         /* This remote handle isn't enqueued, so we never received or
3081          * processed this request.  Clear MSG_RESENT, because it can
3082          * be handled like any normal request now. */
3083
3084         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3085
3086         DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
3087                   remote_hdl.cookie);
3088 }
3089
3090 int intent_disposition(struct ldlm_reply *rep, int flag)
3091 {
3092         if (!rep)
3093                 return 0;
3094         return (rep->lock_policy_res1 & flag);
3095 }
3096
3097 void intent_set_disposition(struct ldlm_reply *rep, int flag)
3098 {
3099         if (!rep)
3100                 return;
3101         rep->lock_policy_res1 |= flag;
3102 }
3103
3104 static int mds_intent_policy(struct ldlm_namespace *ns,
3105                              struct ldlm_lock **lockp, void *req_cookie,
3106                              ldlm_mode_t mode, int flags, void *data)
3107 {
3108         struct ptlrpc_request *req = req_cookie;
3109         struct ldlm_lock *lock = *lockp;
3110         struct ldlm_intent *it;
3111         struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
3112         struct ldlm_reply *rep;
3113         struct lustre_handle lockh[2] = {{0}, {0}};
3114         struct ldlm_lock *new_lock = NULL;
3115         int getattr_part = MDS_INODELOCK_UPDATE;
3116         int rc, repsize[4] = { sizeof(struct ldlm_reply),
3117                                sizeof(struct mds_body),
3118                                mds->mds_max_mdsize,
3119                                mds->mds_max_cookiesize };
3120         int offset = MDS_REQ_INTENT_REC_OFF;
3121         ENTRY;
3122
3123         LASSERT(req != NULL);
3124         MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
3125
3126         if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
3127                 /* No intent was provided */
3128                 int size = sizeof(struct ldlm_reply);
3129                 rc = lustre_pack_reply(req, 1, &size, NULL);
3130                 LASSERT(rc == 0);
3131                 RETURN(0);
3132         }
3133
3134         it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
3135                                 lustre_swab_ldlm_intent);
3136         if (it == NULL) {
3137                 CERROR("Intent missing\n");
3138                 RETURN(req->rq_status = -EFAULT);
3139         }
3140
3141         LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
3142
3143         rc = lustre_pack_reply(req, 3, repsize, NULL);
3144         if (rc)
3145                 RETURN(req->rq_status = rc);
3146
3147         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
3148         LASSERT(rep != NULL);
3149
3150         intent_set_disposition(rep, DISP_IT_EXECD);
3151
3152         /* execute policy */
3153         switch ((long)it->opc) {
3154         case IT_OPEN:
3155         case IT_CREAT|IT_OPEN:
3156                 /* XXX swab here to assert that an mds_open reint
3157                  * packet is following */
3158                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3159                                             lock, NULL, lockh);
3160                 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
3161 #if 0
3162                 /* We abort the lock if the lookup was negative and
3163                  * we did not make it to the OPEN portion */
3164                 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
3165                         RETURN(ELDLM_LOCK_ABORTED);
3166                 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
3167                     !intent_disposition(rep, DISP_OPEN_OPEN))
3168 #endif
3169                 /* IT_OPEN may return lock on cross-node dentry
3170                  * that we want to hold during attr retrival -bzzz */
3171                 if (rc != 0 || lockh[0].cookie == 0)
3172                         RETURN(ELDLM_LOCK_ABORTED);
3173                 break;
3174         case IT_LOOKUP:
3175                 getattr_part = MDS_INODELOCK_LOOKUP;
3176         case IT_CHDIR:
3177         case IT_GETATTR:
3178                 getattr_part |= MDS_INODELOCK_LOOKUP;
3179         case IT_READDIR:
3180                 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
3181                                             lock, &new_lock, lockh);
3182                 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
3183                                                          getattr_part);
3184                 /* FIXME: LDLM can set req->rq_status. MDS sets
3185                    policy_res{1,2} with disposition and status.
3186                    - replay: returns 0 & req->status is old status
3187                    - otherwise: returns req->status */
3188                 if (intent_disposition(rep, DISP_LOOKUP_NEG))
3189                         rep->lock_policy_res2 = 0;
3190                 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
3191                     rep->lock_policy_res2)
3192                         RETURN(ELDLM_LOCK_ABORTED);
3193                 if (req->rq_status != 0) {
3194                         LBUG();
3195                         rep->lock_policy_res2 = req->rq_status;
3196                         RETURN(ELDLM_LOCK_ABORTED);
3197                 }
3198                 break;
3199         case IT_UNLINK:
3200                 rc = mds_lock_and_check_slave(offset, req, lockh);
3201                 if ((rep->lock_policy_res2 = rc)) {
3202                         if (rc == ENOLCK)
3203                                 rep->lock_policy_res2 = 0;
3204                         RETURN(ELDLM_LOCK_ABORTED);
3205                 }
3206                 break;
3207         default:
3208                 CERROR("Unhandled intent "LPD64"\n", it->opc);
3209                 LBUG();
3210         }
3211
3212         /* By this point, whatever function we called above must have either
3213          * filled in 'lockh', been an intent replay, or returned an error.  We
3214          * want to allow replayed RPCs to not get a lock, since we would just
3215          * drop it below anyways because lock replay is done separately by the
3216          * client afterwards.  For regular RPCs we want to give the new lock to
3217          * the client instead of whatever lock it was about to get. */
3218         if (new_lock == NULL)
3219                 new_lock = ldlm_handle2lock(&lockh[0]);
3220         if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
3221                 RETURN(0);
3222
3223         LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
3224                  it->opc, lockh[0].cookie);
3225
3226         /* If we've already given this lock to a client once, then we should
3227          * have no readers or writers.  Otherwise, we should have one reader
3228          * _or_ writer ref (which will be zeroed below) before returning the
3229          * lock to a client. */
3230         if (new_lock->l_export == req->rq_export) {
3231                 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
3232         } else {
3233                 LASSERT(new_lock->l_export == NULL);
3234                 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
3235         }
3236
3237         *lockp = new_lock;
3238
3239         if (new_lock->l_export == req->rq_export) {
3240                 /* Already gave this to the client, which means that we
3241                  * reconstructed a reply. */
3242                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
3243                         MSG_RESENT);
3244                 RETURN(ELDLM_LOCK_REPLACED);
3245         }
3246
3247         /* Fixup the lock to be given to the client */
3248         l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
3249         new_lock->l_readers = 0;
3250         new_lock->l_writers = 0;
3251
3252         new_lock->l_export = class_export_get(req->rq_export);
3253         list_add(&new_lock->l_export_chain,
3254                  &new_lock->l_export->exp_ldlm_data.led_held_locks);
3255
3256         new_lock->l_blocking_ast = lock->l_blocking_ast;
3257         new_lock->l_completion_ast = lock->l_completion_ast;
3258
3259         memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
3260                sizeof(lock->l_remote_handle));
3261
3262         new_lock->l_flags &= ~LDLM_FL_LOCAL;
3263
3264         LDLM_LOCK_PUT(new_lock);
3265         l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
3266
3267         RETURN(ELDLM_LOCK_REPLACED);
3268 }
3269
3270 int mds_attach(struct obd_device *dev, obd_count len, void *data)
3271 {
3272         struct lprocfs_static_vars lvars;
3273         int rc = 0;
3274
3275         lprocfs_init_multi_vars(0, &lvars);
3276
3277         rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3278         if (rc)
3279                 return rc;
3280
3281         return lprocfs_alloc_md_stats(dev, 0);
3282 }
3283
3284 int mds_detach(struct obd_device *dev)
3285 {
3286         lprocfs_free_md_stats(dev);
3287         return lprocfs_obd_detach(dev);
3288 }
3289
3290 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
3291 {
3292         struct lprocfs_static_vars lvars;
3293
3294         lprocfs_init_multi_vars(1, &lvars);
3295         return lprocfs_obd_attach(dev, lvars.obd_vars);
3296 }
3297
3298 int mdt_detach(struct obd_device *dev)
3299 {
3300         return lprocfs_obd_detach(dev);
3301 }
3302
3303 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
3304 {
3305         struct mds_obd *mds = &obd->u.mds;
3306         int rc = 0;
3307         ENTRY;
3308
3309         mds->mds_service =
3310                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3311                                 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
3312                                 mds_handle, "mds", obd->obd_proc_entry);
3313
3314         if (!mds->mds_service) {
3315                 CERROR("failed to start service\n");
3316                 RETURN(-ENOMEM);
3317         }
3318
3319         rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
3320                                     "ll_mdt");
3321         if (rc)
3322                 GOTO(err_thread, rc);
3323
3324         mds->mds_setattr_service =
3325                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3326                                 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
3327                                 mds_handle, "mds_setattr",
3328                                 obd->obd_proc_entry);
3329         if (!mds->mds_setattr_service) {
3330                 CERROR("failed to start getattr service\n");
3331                 GOTO(err_thread, rc = -ENOMEM);
3332         }
3333
3334         rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
3335                                     MDT_NUM_THREADS, "ll_mdt_attr");
3336         if (rc)
3337                 GOTO(err_thread2, rc);
3338
3339         mds->mds_readpage_service =
3340                 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
3341                                 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
3342                                 mds_handle, "mds_readpage",
3343                                 obd->obd_proc_entry);
3344         if (!mds->mds_readpage_service) {
3345                 CERROR("failed to start readpage service\n");
3346                 GOTO(err_thread2, rc = -ENOMEM);
3347         }
3348
3349         rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
3350                                     MDT_NUM_THREADS, "ll_mdt_rdpg");
3351
3352         if (rc)
3353                 GOTO(err_thread3, rc);
3354
3355         RETURN(0);
3356
3357 err_thread3:
3358         ptlrpc_unregister_service(mds->mds_readpage_service);
3359 err_thread2:
3360         ptlrpc_unregister_service(mds->mds_setattr_service);
3361 err_thread:
3362         ptlrpc_unregister_service(mds->mds_service);
3363         return rc;
3364 }
3365
3366 static int mdt_cleanup(struct obd_device *obd, int flags)
3367 {
3368         struct mds_obd *mds = &obd->u.mds;
3369         ENTRY;
3370
3371         ptlrpc_stop_all_threads(mds->mds_readpage_service);
3372         ptlrpc_unregister_service(mds->mds_readpage_service);
3373
3374         ptlrpc_stop_all_threads(mds->mds_setattr_service);
3375         ptlrpc_unregister_service(mds->mds_setattr_service);
3376
3377         ptlrpc_stop_all_threads(mds->mds_service);
3378         ptlrpc_unregister_service(mds->mds_service);
3379
3380         RETURN(0);
3381 }
3382
3383 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
3384                                          __u64 gr, void *data)
3385 {
3386         struct lustre_id id;
3387         struct obd_device *obd = data;
3388
3389         id_ino(&id) = ino;
3390         id_gen(&id) = gen;
3391         return mds_id2dentry(obd, &id, NULL);
3392 }
3393
3394 static int mds_get_info(struct obd_export *exp, __u32 keylen,
3395                         void *key, __u32 *valsize, void *val)
3396 {
3397         struct obd_device *obd;
3398         struct mds_obd *mds;
3399         ENTRY;
3400
3401         obd = class_exp2obd(exp);
3402         mds = &obd->u.mds;
3403
3404         if (obd == NULL) {
3405                 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
3406                        exp->exp_handle.h_cookie);
3407                 RETURN(-EINVAL);
3408         }
3409
3410         if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
3411                 /* get log_context handle. */
3412                 unsigned long *llh_handle = val;
3413                 *valsize = sizeof(unsigned long);
3414                 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
3415                 RETURN(0);
3416         }
3417         if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
3418                 /* get log_context handle. */
3419                 unsigned long *sb = val;
3420                 *valsize = sizeof(unsigned long);
3421                 *sb = (unsigned long)obd->u.mds.mds_sb;
3422                 RETURN(0);
3423         }
3424
3425         if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
3426                 __u32 *mdsize = val;
3427                 *valsize = sizeof(*mdsize);
3428                 *mdsize = mds->mds_max_mdsize;
3429                 RETURN(0);
3430         }
3431
3432         if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
3433                 __u32 *mdsnum = val;
3434                 *valsize = sizeof(*mdsnum);
3435                 *mdsnum = mds->mds_num;
3436                 RETURN(0);
3437         }
3438
3439         if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
3440                 struct lustre_id *rootid = val;
3441                 *valsize = sizeof(struct lustre_id);
3442                 *rootid = mds->mds_rootid;
3443                 RETURN(0);
3444         }
3445
3446         CDEBUG(D_IOCTL, "invalid key\n");
3447         RETURN(-EINVAL);
3448
3449 }
3450 struct lvfs_callback_ops mds_lvfs_ops = {
3451         l_id2dentry:     mds_lvfs_id2dentry,
3452 };
3453
3454 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
3455                 int objcount, struct obd_ioobj *obj,
3456                 int niocount, struct niobuf_remote *nb,
3457                 struct niobuf_local *res,
3458                 struct obd_trans_info *oti);
3459
3460 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
3461                  int objcount, struct obd_ioobj *obj, int niocount,
3462                  struct niobuf_local *res, struct obd_trans_info *oti,
3463                  int rc);
3464
3465 /* use obd ops to offer management infrastructure */
3466 static struct obd_ops mds_obd_ops = {
3467         .o_owner           = THIS_MODULE,
3468         .o_attach          = mds_attach,
3469         .o_detach          = mds_detach,
3470         .o_connect         = mds_connect,
3471         .o_connect_post    = mds_connect_post,
3472         .o_init_export     = mds_init_export,
3473         .o_destroy_export  = mds_destroy_export,
3474         .o_disconnect      = mds_disconnect,
3475         .o_setup           = mds_setup,
3476         .o_precleanup      = mds_precleanup,
3477         .o_cleanup         = mds_cleanup,
3478         .o_postrecov       = mds_postrecov,
3479         .o_statfs          = mds_obd_statfs,
3480         .o_iocontrol       = mds_iocontrol,
3481         .o_create          = mds_obd_create,
3482         .o_destroy         = mds_obd_destroy,
3483         .o_llog_init       = mds_llog_init,
3484         .o_llog_finish     = mds_llog_finish,
3485         .o_notify          = mds_notify,
3486         .o_get_info        = mds_get_info,
3487         .o_set_info        = mds_set_info,
3488         .o_preprw          = mds_preprw,
3489         .o_commitrw        = mds_commitrw,
3490 };
3491
3492 static struct obd_ops mdt_obd_ops = {
3493         .o_owner           = THIS_MODULE,
3494         .o_attach          = mdt_attach,
3495         .o_detach          = mdt_detach,
3496         .o_setup           = mdt_setup,
3497         .o_cleanup         = mdt_cleanup,
3498 };
3499
3500 static int __init mds_init(void)
3501 {
3502         struct lprocfs_static_vars lvars;
3503
3504         mds_group_hash_init();
3505
3506         lprocfs_init_multi_vars(0, &lvars);
3507         class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
3508                             LUSTRE_MDS_NAME);
3509         lprocfs_init_multi_vars(1, &lvars);
3510         class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
3511                             LUSTRE_MDT_NAME);
3512
3513         return 0;
3514 }
3515
3516 static void /*__exit*/ mds_exit(void)
3517 {
3518         mds_group_hash_cleanup();
3519
3520         class_unregister_type(LUSTRE_MDS_NAME);
3521         class_unregister_type(LUSTRE_MDT_NAME);
3522 }
3523
3524 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3525 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
3526 MODULE_LICENSE("GPL");
3527
3528 module_init(mds_init);
3529 module_exit(mds_exit);