lustre/mds/mds_fs.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  mds/mds_fs.c
   5  *  Lustre Metadata Server (MDS) filesystem interface code
   6  *
   7  *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   8  *   Author: Andreas Dilger <adilger@clusterfs.com>
   9  *
  10  *   This file is part of Lustre, http://www.lustre.org.
  11  *
  12  *   Lustre is free software; you can redistribute it and/or
  13  *   modify it under the terms of version 2 of the GNU General Public
  14  *   License as published by the Free Software Foundation.
  15  *
  16  *   Lustre is distributed in the hope that it will be useful,
  17  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  *   GNU General Public License for more details.
  20  *
  21  *   You should have received a copy of the GNU General Public License
  22  *   along with Lustre; if not, write to the Free Software
  23  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  24  */
  25
  26 #ifndef EXPORT_SYMTAB
  27 # define EXPORT_SYMTAB
  28 #endif
  29 #define DEBUG_SUBSYSTEM S_MDS
  30
  31 #include <linux/module.h>
  32 #include <linux/kmod.h>
  33 #include <linux/version.h>
  34 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  35 #include <linux/mount.h>
  36 #endif
  37 #include <linux/lustre_mds.h>
  38 #include <linux/obd_class.h>
  39 #include <linux/obd_support.h>
  40 #include <linux/lustre_lib.h>
  41 #include <linux/lustre_fsfilt.h>
  42 #include <portals/list.h>
  43
  44 #include "mds_internal.h"
  45
  46 /* This limit is arbitrary, but for now we fit it in 1 page (32k clients) */
  47 #define MDS_MAX_CLIENTS (PAGE_SIZE * 8)
  48 #define MDS_MAX_CLIENT_WORDS (MDS_MAX_CLIENTS / sizeof(unsigned long))
  49
  50 #define LAST_RCVD "last_rcvd"
  51 #define LOV_OBJID "lov_objid"
  52
  53 /* Add client data to the MDS.  We use a bitmap to locate a free space
  54  * in the last_rcvd file if cl_off is -1 (i.e. a new client).
  55  * Otherwise, we have just read the data from the last_rcvd file and
  56  * we know its offset.
  57  */
  58 int mds_client_add(struct obd_device *obd, struct mds_obd *mds,
  59                    struct mds_export_data *med, int cl_idx)
  60 {
  61         unsigned long *bitmap = mds->mds_client_bitmap;
  62         int new_client = (cl_idx == -1);
  63         ENTRY;
  64
  65         LASSERT(bitmap != NULL);
  66
  67         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
  68         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
  69                 RETURN(0);
  70
  71         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
  72          * there's no need for extra complication here
  73          */
  74         if (new_client) {
  75                 cl_idx = find_first_zero_bit(bitmap, MDS_MAX_CLIENTS);
  76         repeat:
  77                 if (cl_idx >= MDS_MAX_CLIENTS) {
  78                         CERROR("no room for clients - fix MDS_MAX_CLIENTS\n");
  79                         return -ENOMEM;
  80                 }
  81                 if (test_and_set_bit(cl_idx, bitmap)) {
  82                         cl_idx = find_next_zero_bit(bitmap, MDS_MAX_CLIENTS,
  83                                                     cl_idx);
  84                         goto repeat;
  85                 }
  86         } else {
  87                 if (test_and_set_bit(cl_idx, bitmap)) {
  88                         CERROR("MDS client %d: bit already set in bitmap!!\n",
  89                                cl_idx);
  90                         LBUG();
  91                 }
  92         }
  93
  94         CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
  95                cl_idx, med->med_mcd->mcd_uuid);
  96
  97         med->med_idx = cl_idx;
  98         med->med_off = MDS_LR_CLIENT_START + (cl_idx * MDS_LR_CLIENT_SIZE);
  99
 100         if (new_client) {
 101                 struct obd_run_ctxt saved;
 102                 loff_t off = med->med_off;
 103                 struct file *file = mds->mds_rcvd_filp;
 104                 int rc;
 105
 106                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
 107                 rc = fsfilt_write_record(obd, file, med->med_mcd,
 108                                          sizeof(*med->med_mcd), &off, 1);
 109                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 110
 111                 if (rc)
 112                         return rc;
 113                 CDEBUG(D_INFO, "wrote client mcd at idx %u off %llu (len %u)\n",
 114                        med->med_idx, med->med_off,
 115                        (unsigned int)sizeof(*med->med_mcd));
 116         }
 117         return 0;
 118 }
 119
 120 int mds_client_free(struct obd_export *exp, int clear_client)
 121 {
 122         struct mds_export_data *med = &exp->exp_mds_data;
 123         struct mds_obd *mds = &exp->exp_obd->u.mds;
 124         struct obd_device *obd = exp->exp_obd;
 125         struct mds_client_data zero_mcd;
 126         struct obd_run_ctxt saved;
 127         int rc;
 128         unsigned long *bitmap = mds->mds_client_bitmap;
 129
 130         LASSERT(bitmap);
 131         if (!med->med_mcd)
 132                 RETURN(0);
 133
 134         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 135         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
 136                 GOTO(free_and_out, 0);
 137
 138         CDEBUG(D_INFO, "freeing client at idx %u (%lld)with UUID '%s'\n",
 139                med->med_idx, med->med_off, med->med_mcd->mcd_uuid);
 140
 141         /* Clear the bit _after_ zeroing out the client so we don't
 142            race with mds_client_add and zero out new clients.*/
 143         if (!test_bit(med->med_idx, bitmap)) {
 144                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 145                        med->med_idx);
 146                 LBUG();
 147         }
 148
 149         if (clear_client) {
 150                 memset(&zero_mcd, 0, sizeof zero_mcd);
 151                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
 152                 rc = fsfilt_write_record(obd, mds->mds_rcvd_filp, &zero_mcd,
 153                                          sizeof(zero_mcd), &med->med_off, 1);
 154                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 155
 156                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 157                        "zeroing out client %s idx %u in %s rc %d\n",
 158                        med->med_mcd->mcd_uuid, med->med_idx, LAST_RCVD, rc);
 159         }
 160
 161         if (!test_and_clear_bit(med->med_idx, bitmap)) {
 162                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 163                        med->med_idx);
 164                 LBUG();
 165         }
 166
 167  free_and_out:
 168         OBD_FREE(med->med_mcd, sizeof(*med->med_mcd));
 169
 170         return 0;
 171 }
 172
 173 static int mds_server_free_data(struct mds_obd *mds)
 174 {
 175         OBD_FREE(mds->mds_client_bitmap,
 176                  MDS_MAX_CLIENT_WORDS * sizeof(unsigned long));
 177         OBD_FREE(mds->mds_server_data, sizeof(*mds->mds_server_data));
 178         mds->mds_server_data = NULL;
 179
 180         return 0;
 181 }
 182
 183 static int mds_read_last_rcvd(struct obd_device *obd, struct file *file)
 184 {
 185         struct mds_obd *mds = &obd->u.mds;
 186         struct mds_server_data *msd;
 187         struct mds_client_data *mcd = NULL;
 188         loff_t off = 0;
 189         unsigned long last_rcvd_size = file->f_dentry->d_inode->i_size;
 190         __u64 mount_count;
 191         int cl_idx, rc = 0;
 192         ENTRY;
 193
 194         /* ensure padding in the struct is the correct size */
 195         LASSERT (offsetof(struct mds_server_data, msd_padding) +
 196                  sizeof(msd->msd_padding) == MDS_LR_SERVER_SIZE);
 197         LASSERT (offsetof(struct mds_client_data, mcd_padding) +
 198                  sizeof(mcd->mcd_padding) == MDS_LR_CLIENT_SIZE);
 199
 200         OBD_ALLOC_WAIT(msd, sizeof(*msd));
 201         if (!msd)
 202                 RETURN(-ENOMEM);
 203
 204         OBD_ALLOC_WAIT(mds->mds_client_bitmap,
 205                   MDS_MAX_CLIENT_WORDS * sizeof(unsigned long));
 206         if (!mds->mds_client_bitmap) {
 207                 OBD_FREE(msd, sizeof(*msd));
 208                 RETURN(-ENOMEM);
 209         }
 210
 211         mds->mds_server_data = msd;
 212
 213         if (last_rcvd_size == 0) {
 214                 CWARN("%s: initializing new %s\n", obd->obd_name, LAST_RCVD);
 215
 216                 memcpy(msd->msd_uuid, obd->obd_uuid.uuid,sizeof(msd->msd_uuid));
 217                 msd->msd_last_transno = 0;
 218                 mount_count = msd->msd_mount_count = 0;
 219                 msd->msd_server_size = cpu_to_le32(MDS_LR_SERVER_SIZE);
 220                 msd->msd_client_start = cpu_to_le32(MDS_LR_CLIENT_START);
 221                 msd->msd_client_size = cpu_to_le16(MDS_LR_CLIENT_SIZE);
 222                 msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
 223         } else {
 224                 rc = fsfilt_read_record(obd, file, msd, sizeof(*msd), &off);
 225                 if (rc) {
 226                         CERROR("error reading MDS %s: rc = %d\n", LAST_RCVD, rc);
 227                         GOTO(err_msd, rc);
 228                 }
 229                 if (strcmp(msd->msd_uuid, obd->obd_uuid.uuid) != 0) {
 230                         CERROR("OBD UUID %s does not match last_rcvd UUID %s\n",
 231                                obd->obd_uuid.uuid, msd->msd_uuid);
 232                         GOTO(err_msd, rc = -EINVAL);
 233                 }
 234                 mount_count = le64_to_cpu(msd->msd_mount_count);
 235         }
 236         if (msd->msd_feature_incompat & ~cpu_to_le32(MDS_INCOMPAT_SUPP)) {
 237                 CERROR("unsupported incompat feature %x\n",
 238                        le32_to_cpu(msd->msd_feature_incompat) &
 239                        ~MDS_INCOMPAT_SUPP);
 240                 GOTO(err_msd, rc = -EINVAL);
 241         }
 242         /* XXX updating existing b_devel fs only, can be removed in future */
 243         msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
 244         if (msd->msd_feature_rocompat & ~cpu_to_le32(MDS_ROCOMPAT_SUPP)) {
 245                 CERROR("unsupported read-only feature %x\n",
 246                        le32_to_cpu(msd->msd_feature_rocompat) &
 247                        ~MDS_ROCOMPAT_SUPP);
 248                 /* Do something like remount filesystem read-only */
 249                 GOTO(err_msd, rc = -EINVAL);
 250         }
 251
 252         mds->mds_last_transno = le64_to_cpu(msd->msd_last_transno);
 253
 254         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 255                obd->obd_name, mds->mds_last_transno);
 256         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 257                obd->obd_name, mount_count + 1);
 258         CDEBUG(D_INODE, "%s: server data size: %u\n",
 259                obd->obd_name, le32_to_cpu(msd->msd_server_size));
 260         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 261                obd->obd_name, le32_to_cpu(msd->msd_client_start));
 262         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 263                obd->obd_name, le32_to_cpu(msd->msd_client_size));
 264         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 265                obd->obd_name, last_rcvd_size);
 266         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 267                last_rcvd_size <= MDS_LR_CLIENT_START ? 0 :
 268                (last_rcvd_size - MDS_LR_CLIENT_START) / MDS_LR_CLIENT_SIZE);
 269
 270         /* When we do a clean MDS shutdown, we save the last_transno into
 271          * the header.  If we find clients with higher last_transno values
 272          * then those clients may need recovery done. */
 273         for (cl_idx = 0, off = le32_to_cpu(msd->msd_client_start);
 274              off < last_rcvd_size; cl_idx++) {
 275                 __u64 last_transno;
 276                 int mount_age;
 277
 278                 if (!mcd) {
 279                         OBD_ALLOC_WAIT(mcd, sizeof(*mcd));
 280                         if (!mcd)
 281                                 GOTO(err_client, rc = -ENOMEM);
 282                 }
 283
 284                 /* Don't assume off is incremented properly by
 285                  * fsfilt_read_record(), in case sizeof(*mcd)
 286                  * isn't the same as msd->msd_client_size.  */
 287                 off = le32_to_cpu(msd->msd_client_start) +
 288                         cl_idx * le16_to_cpu(msd->msd_client_size);
 289                 rc = fsfilt_read_record(obd, file, mcd, sizeof(*mcd), &off);
 290                 if (rc) {
 291                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 292                                LAST_RCVD, cl_idx, off, rc);
 293                         break; /* read error shouldn't cause startup to fail */
 294                 }
 295
 296                 if (mcd->mcd_uuid[0] == '\0') {
 297                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 298                                cl_idx);
 299                         continue;
 300                 }
 301
 302                 last_transno = le64_to_cpu(mcd->mcd_last_transno);
 303
 304                 /* These exports are cleaned up by mds_disconnect(), so they
 305                  * need to be set up like real exports as mds_connect() does.
 306                  */
 307                 mount_age = mount_count - le64_to_cpu(mcd->mcd_mount_count);
 308                 if (mount_age < MDS_MOUNT_RECOV) {
 309                         struct obd_export *exp = class_new_export(obd);
 310                         struct mds_export_data *med;
 311                         CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 312                                " srv lr: "LPU64" mnt: "LPU64" last mount: "
 313                                LPU64"\n", mcd->mcd_uuid, cl_idx,
 314                                last_transno, le64_to_cpu(msd->msd_last_transno),
 315                                le64_to_cpu(mcd->mcd_mount_count), mount_count);
 316                         if (exp == NULL)
 317                                 GOTO(err_client, rc = -ENOMEM);
 318
 319                         memcpy(&exp->exp_client_uuid.uuid, mcd->mcd_uuid,
 320                                sizeof exp->exp_client_uuid.uuid);
 321                         med = &exp->exp_mds_data;
 322                         med->med_mcd = mcd;
 323                         mds_client_add(obd, mds, med, cl_idx);
 324                         /* create helper if export init gets more complex */
 325                         INIT_LIST_HEAD(&med->med_open_head);
 326                         spin_lock_init(&med->med_open_lock);
 327
 328                         mcd = NULL;
 329                         obd->obd_recoverable_clients++;
 330                         obd->obd_max_recoverable_clients++;
 331                         class_export_put(exp);
 332                 } else {
 333                         CDEBUG(D_INFO, "discarded client %d, UUID '%s', count "
 334                                LPU64"\n", cl_idx, mcd->mcd_uuid,
 335                                le64_to_cpu(mcd->mcd_mount_count));
 336                 }
 337
 338                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
 339                        cl_idx, last_transno);
 340
 341                 if (last_transno > mds->mds_last_transno)
 342                        mds->mds_last_transno = last_transno;
 343         }
 344
 345         obd->obd_last_committed = mds->mds_last_transno;
 346         if (obd->obd_recoverable_clients) {
 347                 CWARN("RECOVERY: service %s, %d recoverable clients, "
 348                       "last_transno "LPU64"\n", obd->obd_name,
 349                       obd->obd_recoverable_clients, mds->mds_last_transno);
 350                 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
 351                 obd->obd_recovering = 1;
 352         }
 353
 354         if (mcd)
 355                 OBD_FREE(mcd, sizeof(*mcd));
 356
 357         mds->mds_mount_count = mount_count + 1;
 358         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
 359
 360         /* save it, so mount count and last_transno is current */
 361         rc = mds_update_server_data(obd, 1);
 362
 363         RETURN(rc);
 364
 365 err_client:
 366         class_disconnect_exports(obd, 0);
 367 err_msd:
 368         mds_server_free_data(mds);
 369         RETURN(rc);
 370 }
 371
 372 int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
 373 {
 374         struct mds_obd *mds = &obd->u.mds;
 375         struct obd_run_ctxt saved;
 376         struct dentry *dentry;
 377         struct file *file;
 378         int rc;
 379         ENTRY;
 380
 381
 382         /* Get rid of unneeded supplementary groups */
 383         current->ngroups = 0;
 384         memset(current->groups, 0, sizeof(current->groups));
 385
 386         mds->mds_vfsmnt = mnt;
 387         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 388
 389         fsfilt_setup(obd, mds->mds_sb);
 390
 391         OBD_SET_CTXT_MAGIC(&obd->obd_ctxt);
 392         obd->obd_ctxt.pwdmnt = mnt;
 393         obd->obd_ctxt.pwd = mnt->mnt_root;
 394         obd->obd_ctxt.fs = get_ds();
 395         obd->obd_ctxt.cb_ops = mds_lvfs_ops;
 396
 397         /* setup the directory tree */
 398         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 399         dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755);
 400         if (IS_ERR(dentry)) {
 401                 rc = PTR_ERR(dentry);
 402                 CERROR("cannot create ROOT directory: rc = %d\n", rc);
 403                 GOTO(err_pop, rc);
 404         }
 405
 406         mds->mds_rootfid.id = dentry->d_inode->i_ino;
 407         mds->mds_rootfid.generation = dentry->d_inode->i_generation;
 408         mds->mds_rootfid.f_type = S_IFDIR;
 409
 410         dput(dentry);
 411
 412         dentry = lookup_one_len("__iopen__", current->fs->pwd,
 413                                 strlen("__iopen__"));
 414         if (IS_ERR(dentry) || !dentry->d_inode) {
 415                 rc = (IS_ERR(dentry)) ? PTR_ERR(dentry): -ENOENT;
 416                 CERROR("cannot open iopen FH directory: rc = %d\n", rc);
 417                 GOTO(err_pop, rc);
 418         }
 419         mds->mds_fid_de = dentry;
 420
 421         dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777);
 422         if (IS_ERR(dentry)) {
 423                 rc = PTR_ERR(dentry);
 424                 CERROR("cannot create PENDING directory: rc = %d\n", rc);
 425                 GOTO(err_fid, rc);
 426         }
 427         mds->mds_pending_dir = dentry;
 428
 429         dentry = simple_mkdir(current->fs->pwd, "LOGS", 0777);
 430         if (IS_ERR(dentry)) {
 431                 rc = PTR_ERR(dentry);
 432                 CERROR("cannot create LOGS directory: rc = %d\n", rc);
 433                 GOTO(err_pending, rc);
 434         }
 435         mds->mds_logs_dir = dentry;
 436
 437         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777);
 438         if (IS_ERR(dentry)) {
 439                 rc = PTR_ERR(dentry);
 440                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
 441                 GOTO(err_logs, rc);
 442         }
 443         mds->mds_objects_dir = dentry;
 444
 445         /* open and test the last rcvd file */
 446         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
 447         if (IS_ERR(file)) {
 448                 rc = PTR_ERR(file);
 449                 CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
 450                 GOTO(err_objects, rc = PTR_ERR(file));
 451         }
 452         mds->mds_rcvd_filp = file;
 453         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 454                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 455                        file->f_dentry->d_inode->i_mode);
 456                 GOTO(err_last_rcvd, rc = -ENOENT);
 457         }
 458
 459         rc = mds_read_last_rcvd(obd, file);
 460         if (rc) {
 461                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
 462                 GOTO(err_last_rcvd, rc);
 463         }
 464
 465         /* open and test the lov objd file */
 466         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
 467         if (IS_ERR(file)) {
 468                 rc = PTR_ERR(file);
 469                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
 470                 GOTO(err_client, rc = PTR_ERR(file));
 471         }
 472         mds->mds_lov_objid_filp = file;
 473         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 474                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
 475                        file->f_dentry->d_inode->i_mode);
 476                 GOTO(err_lov_objid, rc = -ENOENT);
 477         }
 478 err_pop:
 479         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 480
 481         return rc;
 482
 483 err_lov_objid:
 484         if (mds->mds_lov_objid_filp && filp_close(mds->mds_lov_objid_filp, 0))
 485                 CERROR("can't close %s after error\n", LOV_OBJID);
 486 err_client:
 487         class_disconnect_exports(obd, 0);
 488 err_last_rcvd:
 489         if (mds->mds_rcvd_filp && filp_close(mds->mds_rcvd_filp, 0))
 490                 CERROR("can't close %s after error\n", LAST_RCVD);
 491 err_objects:
 492         dput(mds->mds_objects_dir);
 493 err_logs:
 494         dput(mds->mds_logs_dir);
 495 err_pending:
 496         dput(mds->mds_pending_dir);
 497 err_fid:
 498         dput(mds->mds_fid_de);
 499         goto err_pop;
 500 }
 501
 502
 503 int mds_fs_cleanup(struct obd_device *obd, int flags)
 504 {
 505         struct mds_obd *mds = &obd->u.mds;
 506         struct obd_run_ctxt saved;
 507         int rc = 0;
 508
 509         if (flags & OBD_OPT_FAILOVER)
 510                 CERROR("%s: shutting down for failover; client state will"
 511                        " be preserved.\n", obd->obd_name);
 512
 513         class_disconnect_exports(obd, flags); /* cleans up client info too */
 514         mds_server_free_data(mds);
 515
 516         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 517         if (mds->mds_rcvd_filp) {
 518                 rc = filp_close(mds->mds_rcvd_filp, 0);
 519                 mds->mds_rcvd_filp = NULL;
 520                 if (rc)
 521                         CERROR("%s file won't close, rc=%d\n", LAST_RCVD, rc);
 522         }
 523         if (mds->mds_lov_objid_filp) {
 524                 rc = filp_close(mds->mds_lov_objid_filp, 0);
 525                 mds->mds_lov_objid_filp = NULL;
 526                 if (rc)
 527                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
 528         }
 529         if (mds->mds_objects_dir != NULL) {
 530                 l_dput(mds->mds_objects_dir);
 531                 mds->mds_objects_dir = NULL;
 532         }
 533         if (mds->mds_logs_dir) {
 534                 l_dput(mds->mds_logs_dir);
 535                 mds->mds_logs_dir = NULL;
 536         }
 537         if (mds->mds_pending_dir) {
 538                 l_dput(mds->mds_pending_dir);
 539                 mds->mds_pending_dir = NULL;
 540         }
 541         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 542         shrink_dcache_parent(mds->mds_fid_de);
 543         dput(mds->mds_fid_de);
 544
 545         return rc;
 546 }
 547
 548 /* Creates an object with the same name as its fid.  Because this is not at all
 549  * performance sensitive, it is accomplished by creating a file, checking the
 550  * fid, and renaming it. */
 551 int mds_obd_create(struct obd_export *exp, struct obdo *oa,
 552                       struct lov_stripe_md **ea, struct obd_trans_info *oti)
 553 {
 554         struct mds_obd *mds = &exp->exp_obd->u.mds;
 555         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 556         unsigned int tmpname = ll_insecure_random_int();
 557         struct file *filp;
 558         struct dentry *new_child;
 559         struct obd_run_ctxt saved;
 560         char fidname[LL_FID_NAMELEN];
 561         void *handle;
 562         int rc = 0, err, namelen;
 563         ENTRY;
 564
 565         push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
 566
 567         sprintf(fidname, "OBJECTS/%u", tmpname);
 568         filp = filp_open(fidname, O_CREAT | O_EXCL, 0644);
 569         if (IS_ERR(filp)) {
 570                 rc = PTR_ERR(filp);
 571                 if (rc == -EEXIST) {
 572                         CERROR("impossible object name collision %u\n",
 573                                tmpname);
 574                         LBUG();
 575                 }
 576                 CERROR("error creating tmp object %u: rc %d\n", tmpname, rc);
 577                 GOTO(out_pop, rc);
 578         }
 579
 580         LASSERT(mds->mds_objects_dir == filp->f_dentry->d_parent);
 581
 582         oa->o_id = filp->f_dentry->d_inode->i_ino;
 583         oa->o_generation = filp->f_dentry->d_inode->i_generation;
 584         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 585
 586         down(&parent_inode->i_sem);
 587         new_child = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 588
 589         if (IS_ERR(new_child)) {
 590                 CERROR("getting neg dentry for obj rename: %d\n", rc);
 591                 GOTO(out_close, rc = PTR_ERR(new_child));
 592         }
 593         if (new_child->d_inode != NULL) {
 594                 CERROR("impossible non-negative obj dentry " LPU64":%u!\n",
 595                        oa->o_id, oa->o_generation);
 596                 LBUG();
 597         }
 598
 599         handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
 600                               FSFILT_OP_RENAME, NULL);
 601         if (IS_ERR(handle))
 602                 GOTO(out_dput, rc = PTR_ERR(handle));
 603
 604         lock_kernel();
 605         rc = vfs_rename(mds->mds_objects_dir->d_inode, filp->f_dentry,
 606                         mds->mds_objects_dir->d_inode, new_child);
 607         unlock_kernel();
 608         if (rc)
 609                 CERROR("error renaming new object "LPU64":%u: rc %d\n",
 610                        oa->o_id, oa->o_generation, rc);
 611
 612         err = fsfilt_commit(exp->exp_obd, mds->mds_objects_dir->d_inode,
 613                             handle, 0);
 614         if (!err)
 615                 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
 616         else if (!rc)
 617                 rc = err;
 618 out_dput:
 619         dput(new_child);
 620 out_close:
 621         up(&parent_inode->i_sem);
 622         err = filp_close(filp, 0);
 623         if (err) {
 624                 CERROR("closing tmpfile %u: rc %d\n", tmpname, rc);
 625                 if (!rc)
 626                         rc = err;
 627         }
 628 out_pop:
 629         pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
 630         RETURN(rc);
 631 }
 632
 633 int mds_obd_destroy(struct obd_export *exp, struct obdo *oa,
 634                     struct lov_stripe_md *ea, struct obd_trans_info *oti)
 635 {
 636         struct mds_obd *mds = &exp->exp_obd->u.mds;
 637         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 638         struct obd_device *obd = exp->exp_obd;
 639         struct obd_run_ctxt saved;
 640         char fidname[LL_FID_NAMELEN];
 641         struct dentry *de;
 642         void *handle;
 643         int err, namelen, rc = 0;
 644         ENTRY;
 645
 646         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 647
 648         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 649
 650         down(&parent_inode->i_sem);
 651         de = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 652         if (de == NULL || de->d_inode == NULL) {
 653                 CERROR("destroying non-existent object "LPU64"\n", oa->o_id);
 654                 GOTO(out, rc = IS_ERR(de) ? PTR_ERR(de) : -ENOENT);
 655         }
 656
 657         handle = fsfilt_start(obd, mds->mds_objects_dir->d_inode,
 658                               FSFILT_OP_UNLINK_LOG, oti);
 659         if (IS_ERR(handle)) {
 660                 GOTO(out_dput, rc = PTR_ERR(handle));
 661         }
 662
 663         rc = vfs_unlink(mds->mds_objects_dir->d_inode, de);
 664         if (rc)
 665                 CERROR("error destroying object "LPU64":%u: rc %d\n",
 666                        oa->o_id, oa->o_generation, rc);
 667
 668         err = fsfilt_commit(obd, mds->mds_objects_dir->d_inode, handle, 0);
 669         if (err && !rc)
 670                 rc = err;
 671 out_dput:
 672         l_dput(de);
 673 out:
 674         up(&parent_inode->i_sem);
 675         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 676         RETURN(rc);
 677 }