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 index %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 index %u (%lld)with UUID '%s'\n",
 139                med->med_idx, med->med_off, med->med_mcd->mcd_uuid);
 140
 141         if (!test_and_clear_bit(med->med_idx, bitmap)) {
 142                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 143                        med->med_idx);
 144                 LBUG();
 145         }
 146
 147         if (clear_client) {
 148                 memset(&zero_mcd, 0, sizeof zero_mcd);
 149                 push_ctxt(&saved, &obd->obd_ctxt, NULL);
 150                 rc = fsfilt_write_record(obd, mds->mds_rcvd_filp, &zero_mcd,
 151                                          sizeof(zero_mcd), &med->med_off, 1);
 152                 pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 153
 154                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 155                        "zeroing out client %s off %u in %s rc %d\n",
 156                        med->med_mcd->mcd_uuid, med->med_idx, LAST_RCVD, rc);
 157         }
 158
 159  free_and_out:
 160         OBD_FREE(med->med_mcd, sizeof(*med->med_mcd));
 161
 162         return 0;
 163 }
 164
 165 static int mds_server_free_data(struct mds_obd *mds)
 166 {
 167         OBD_FREE(mds->mds_client_bitmap,
 168                  MDS_MAX_CLIENT_WORDS * sizeof(unsigned long));
 169         OBD_FREE(mds->mds_server_data, sizeof(*mds->mds_server_data));
 170         mds->mds_server_data = NULL;
 171
 172         return 0;
 173 }
 174
 175 static int mds_read_last_rcvd(struct obd_device *obd, struct file *file)
 176 {
 177         struct mds_obd *mds = &obd->u.mds;
 178         struct mds_server_data *msd;
 179         struct mds_client_data *mcd = NULL;
 180         loff_t off = 0;
 181         unsigned long last_rcvd_size = file->f_dentry->d_inode->i_size;
 182         __u64 last_transno = 0;
 183         __u64 mount_count;
 184         int cl_idx, rc = 0;
 185         ENTRY;
 186
 187         LASSERT(sizeof(struct mds_client_data) == MDS_LR_CLIENT_SIZE);
 188         LASSERT(sizeof(struct mds_server_data) <= MDS_LR_SERVER_SIZE);
 189
 190         OBD_ALLOC_WAIT(msd, sizeof(*msd));
 191         if (!msd)
 192                 RETURN(-ENOMEM);
 193
 194         OBD_ALLOC_WAIT(mds->mds_client_bitmap,
 195                   MDS_MAX_CLIENT_WORDS * sizeof(unsigned long));
 196         if (!mds->mds_client_bitmap) {
 197                 OBD_FREE(msd, sizeof(*msd));
 198                 RETURN(-ENOMEM);
 199         }
 200
 201         mds->mds_server_data = msd;
 202
 203         if (last_rcvd_size == 0) {
 204                 CWARN("%s: initializing new %s\n", obd->obd_name, LAST_RCVD);
 205                 memcpy(msd->msd_uuid, obd->obd_uuid.uuid,sizeof(msd->msd_uuid));
 206                 msd->msd_server_size = cpu_to_le32(MDS_LR_SERVER_SIZE);
 207                 msd->msd_client_start = cpu_to_le32(MDS_LR_CLIENT_START);
 208                 msd->msd_client_size = cpu_to_le16(MDS_LR_CLIENT_SIZE);
 209                 msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
 210                 rc = fsfilt_write_record(obd, file, msd, sizeof(*msd), &off, 1);
 211
 212                 if (rc == 0)
 213                         RETURN(0);
 214
 215                 CERROR("%s: error writing new MSD: %d\n", obd->obd_name, rc);
 216                 GOTO(err_msd, rc);
 217         }
 218
 219         rc = fsfilt_read_record(obd, file, msd, sizeof(*msd), &off);
 220         if (rc) {
 221                 CERROR("error reading MDS %s: rc = %d\n", LAST_RCVD, rc);
 222                 GOTO(err_msd, rc);
 223         }
 224         if (!msd->msd_server_size)
 225                 msd->msd_server_size = cpu_to_le32(MDS_LR_SERVER_SIZE);
 226         if (!msd->msd_client_start)
 227                 msd->msd_client_start = cpu_to_le32(MDS_LR_CLIENT_START);
 228         if (!msd->msd_client_size)
 229                 msd->msd_client_size = cpu_to_le16(MDS_LR_CLIENT_SIZE);
 230
 231         if (msd->msd_feature_incompat & ~cpu_to_le32(MDS_INCOMPAT_SUPP)) {
 232                 CERROR("unsupported incompat feature %x\n",
 233                        le32_to_cpu(msd->msd_feature_incompat) &
 234                        ~MDS_INCOMPAT_SUPP);
 235                 GOTO(err_msd, rc = -EINVAL);
 236         }
 237         /* XXX updating existing b_devel fs only, can be removed in future */
 238         msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
 239         if (msd->msd_feature_rocompat & ~cpu_to_le32(MDS_ROCOMPAT_SUPP)) {
 240                 CERROR("unsupported read-only feature %x\n",
 241                        le32_to_cpu(msd->msd_feature_rocompat) &
 242                        ~MDS_ROCOMPAT_SUPP);
 243                 /* Do something like remount filesystem read-only */
 244                 GOTO(err_msd, rc = -EINVAL);
 245         }
 246
 247         last_transno = le64_to_cpu(msd->msd_last_transno);
 248         mds->mds_last_transno = last_transno;
 249
 250         mount_count = le64_to_cpu(msd->msd_mount_count);
 251         mds->mds_mount_count = mount_count;
 252
 253         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 254                obd->obd_name, last_transno);
 255         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 256                obd->obd_name, mount_count);
 257         CDEBUG(D_INODE, "%s: server data size: %u\n",
 258                obd->obd_name, le32_to_cpu(msd->msd_server_size));
 259         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 260                obd->obd_name, le32_to_cpu(msd->msd_client_start));
 261         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 262                obd->obd_name, le32_to_cpu(msd->msd_client_size));
 263         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 264                obd->obd_name, last_rcvd_size);
 265         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 266                last_rcvd_size <= MDS_LR_CLIENT_START ? 0 :
 267                (last_rcvd_size - MDS_LR_CLIENT_START) / MDS_LR_CLIENT_SIZE);
 268
 269         /* When we do a clean MDS shutdown, we save the last_transno into
 270          * the header.  If we find clients with higher last_transno values
 271          * then those clients may need recovery done. */
 272         for (cl_idx = 0, off = le32_to_cpu(msd->msd_client_start);
 273              off < last_rcvd_size; cl_idx++) {
 274                 __u64 last_transno;
 275                 int mount_age;
 276
 277                 if (!mcd) {
 278                         OBD_ALLOC_WAIT(mcd, sizeof(*mcd));
 279                         if (!mcd)
 280                                 GOTO(err_client, rc = -ENOMEM);
 281                 }
 282
 283                 /* Don't assume off is incremented properly by
 284                  * fsfilt_read_record(), in case sizeof(*mcd)
 285                  * isn't the same as msd->msd_client_size.  */
 286                 off = le32_to_cpu(msd->msd_client_start) +
 287                         cl_idx * le16_to_cpu(msd->msd_client_size);
 288                 rc = fsfilt_read_record(obd, file, mcd, sizeof(*mcd), &off);
 289                 if (rc) {
 290                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 291                                LAST_RCVD, cl_idx, off, rc);
 292                         break; /* read error shouldn't cause startup to fail */
 293                 }
 294
 295                 if (mcd->mcd_uuid[0] == '\0') {
 296                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 297                                cl_idx);
 298                         continue;
 299                 }
 300
 301                 last_transno = le64_to_cpu(mcd->mcd_last_transno);
 302
 303                 /* These exports are cleaned up by mds_disconnect(), so they
 304                  * need to be set up like real exports as mds_connect() does.
 305                  */
 306                 mount_age = mount_count - le64_to_cpu(mcd->mcd_mount_count);
 307                 if (mount_age < MDS_MOUNT_RECOV) {
 308                         struct obd_export *exp = class_new_export(obd);
 309                         struct mds_export_data *med;
 310                         CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 311                                " srv lr: "LPU64" mnt: "LPU64" last mount: "
 312                                LPU64"\n", mcd->mcd_uuid, cl_idx,
 313                                last_transno, le64_to_cpu(msd->msd_last_transno),
 314                                le64_to_cpu(mcd->mcd_mount_count), mount_count);
 315                         if (exp == NULL)
 316                                 GOTO(err_client, rc = -ENOMEM);
 317
 318                         memcpy(&exp->exp_client_uuid.uuid, mcd->mcd_uuid,
 319                                sizeof exp->exp_client_uuid.uuid);
 320                         med = &exp->exp_mds_data;
 321                         med->med_mcd = mcd;
 322                         mds_client_add(obd, mds, med, cl_idx);
 323                         /* create helper if export init gets more complex */
 324                         INIT_LIST_HEAD(&med->med_open_head);
 325                         spin_lock_init(&med->med_open_lock);
 326
 327                         mcd = NULL;
 328                         obd->obd_recoverable_clients++;
 329                         obd->obd_max_recoverable_clients++;
 330                         class_export_put(exp);
 331                 } else {
 332                         CDEBUG(D_INFO, "discarded client %d, UUID '%s', count "
 333                                LPU64"\n", cl_idx, mcd->mcd_uuid,
 334                                le64_to_cpu(mcd->mcd_mount_count));
 335                 }
 336
 337                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
 338                        cl_idx, last_transno);
 339
 340                 if (last_transno > mds->mds_last_transno)
 341                         mds->mds_last_transno = last_transno;
 342         }
 343
 344         obd->obd_last_committed = mds->mds_last_transno;
 345         if (obd->obd_recoverable_clients) {
 346                 CWARN("RECOVERY: %d recoverable clients, last_transno "
 347                        LPU64"\n", obd->obd_recoverable_clients,
 348                        mds->mds_last_transno);
 349                 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
 350                 obd->obd_recovering = 1;
 351         }
 352
 353         if (mcd)
 354                 OBD_FREE(mcd, sizeof(*mcd));
 355
 356         return 0;
 357
 358 err_client:
 359         class_disconnect_exports(obd, 0);
 360 err_msd:
 361         mds_server_free_data(mds);
 362         return rc;
 363 }
 364
 365 int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
 366 {
 367         struct mds_obd *mds = &obd->u.mds;
 368         struct obd_run_ctxt saved;
 369         struct dentry *dentry;
 370         struct file *file;
 371         int rc;
 372         ENTRY;
 373
 374
 375         /* Get rid of unneeded supplementary groups */
 376         current->ngroups = 0;
 377         memset(current->groups, 0, sizeof(current->groups));
 378
 379         mds->mds_vfsmnt = mnt;
 380         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 381
 382         fsfilt_setup(obd, mds->mds_sb);
 383
 384         OBD_SET_CTXT_MAGIC(&obd->obd_ctxt);
 385         obd->obd_ctxt.pwdmnt = mnt;
 386         obd->obd_ctxt.pwd = mnt->mnt_root;
 387         obd->obd_ctxt.fs = get_ds();
 388         obd->obd_ctxt.cb_ops = mds_lvfs_ops;
 389
 390         /* setup the directory tree */
 391         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 392         dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755);
 393         if (IS_ERR(dentry)) {
 394                 rc = PTR_ERR(dentry);
 395                 CERROR("cannot create ROOT directory: rc = %d\n", rc);
 396                 GOTO(err_pop, rc);
 397         }
 398
 399         mds->mds_rootfid.id = dentry->d_inode->i_ino;
 400         mds->mds_rootfid.generation = dentry->d_inode->i_generation;
 401         mds->mds_rootfid.f_type = S_IFDIR;
 402
 403         dput(dentry);
 404
 405         dentry = lookup_one_len("__iopen__", current->fs->pwd,
 406                                 strlen("__iopen__"));
 407         if (IS_ERR(dentry) || !dentry->d_inode) {
 408                 rc = (IS_ERR(dentry)) ? PTR_ERR(dentry): -ENOENT;
 409                 CERROR("cannot open iopen FH directory: rc = %d\n", rc);
 410                 GOTO(err_pop, rc);
 411         }
 412         mds->mds_fid_de = dentry;
 413
 414         dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777);
 415         if (IS_ERR(dentry)) {
 416                 rc = PTR_ERR(dentry);
 417                 CERROR("cannot create PENDING directory: rc = %d\n", rc);
 418                 GOTO(err_fid, rc);
 419         }
 420         mds->mds_pending_dir = dentry;
 421
 422         dentry = simple_mkdir(current->fs->pwd, "LOGS", 0777);
 423         if (IS_ERR(dentry)) {
 424                 rc = PTR_ERR(dentry);
 425                 CERROR("cannot create LOGS directory: rc = %d\n", rc);
 426                 GOTO(err_pending, rc);
 427         }
 428         mds->mds_logs_dir = dentry;
 429
 430         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777);
 431         if (IS_ERR(dentry)) {
 432                 rc = PTR_ERR(dentry);
 433                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
 434                 GOTO(err_logs, rc);
 435         }
 436         mds->mds_objects_dir = dentry;
 437
 438         /* open and test the last rcvd file */
 439         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
 440         if (IS_ERR(file)) {
 441                 rc = PTR_ERR(file);
 442                 CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
 443                 GOTO(err_objects, rc = PTR_ERR(file));
 444         }
 445         mds->mds_rcvd_filp = file;
 446         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 447                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 448                        file->f_dentry->d_inode->i_mode);
 449                 GOTO(err_last_rcvd, rc = -ENOENT);
 450         }
 451
 452         rc = mds_read_last_rcvd(obd, file);
 453         if (rc) {
 454                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
 455                 GOTO(err_last_rcvd, rc);
 456         }
 457
 458         /* open and test the lov objd file */
 459         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
 460         if (IS_ERR(file)) {
 461                 rc = PTR_ERR(file);
 462                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
 463                 GOTO(err_client, rc = PTR_ERR(file));
 464         }
 465         mds->mds_lov_objid_filp = file;
 466         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 467                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
 468                        file->f_dentry->d_inode->i_mode);
 469                 GOTO(err_lov_objid, rc = -ENOENT);
 470         }
 471 err_pop:
 472         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 473
 474         return rc;
 475
 476 err_lov_objid:
 477         if (mds->mds_lov_objid_filp && filp_close(mds->mds_lov_objid_filp, 0))
 478                 CERROR("can't close %s after error\n", LOV_OBJID);
 479 err_client:
 480         class_disconnect_exports(obd, 0);
 481 err_last_rcvd:
 482         if (mds->mds_rcvd_filp && filp_close(mds->mds_rcvd_filp, 0))
 483                 CERROR("can't close %s after error\n", LAST_RCVD);
 484 err_objects:
 485         dput(mds->mds_objects_dir);
 486 err_logs:
 487         dput(mds->mds_logs_dir);
 488 err_pending:
 489         dput(mds->mds_pending_dir);
 490 err_fid:
 491         dput(mds->mds_fid_de);
 492         goto err_pop;
 493 }
 494
 495
 496 int mds_fs_cleanup(struct obd_device *obd, int flags)
 497 {
 498         struct mds_obd *mds = &obd->u.mds;
 499         struct obd_run_ctxt saved;
 500         int rc = 0;
 501
 502         if (flags & OBD_OPT_FAILOVER)
 503                 CERROR("%s: shutting down for failover; client state will"
 504                        " be preserved.\n", obd->obd_name);
 505
 506         class_disconnect_exports(obd, flags); /* cleans up client info too */
 507         mds_server_free_data(mds);
 508
 509         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 510         if (mds->mds_rcvd_filp) {
 511                 rc = filp_close(mds->mds_rcvd_filp, 0);
 512                 mds->mds_rcvd_filp = NULL;
 513                 if (rc)
 514                         CERROR("%s file won't close, rc=%d\n", LAST_RCVD, rc);
 515         }
 516         if (mds->mds_lov_objid_filp) {
 517                 rc = filp_close(mds->mds_lov_objid_filp, 0);
 518                 mds->mds_lov_objid_filp = NULL;
 519                 if (rc)
 520                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
 521         }
 522         if (mds->mds_objects_dir != NULL) {
 523                 l_dput(mds->mds_objects_dir);
 524                 mds->mds_objects_dir = NULL;
 525         }
 526         if (mds->mds_logs_dir) {
 527                 l_dput(mds->mds_logs_dir);
 528                 mds->mds_logs_dir = NULL;
 529         }
 530         if (mds->mds_pending_dir) {
 531                 l_dput(mds->mds_pending_dir);
 532                 mds->mds_pending_dir = NULL;
 533         }
 534         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 535         shrink_dcache_parent(mds->mds_fid_de);
 536         dput(mds->mds_fid_de);
 537
 538         return rc;
 539 }
 540
 541 /* Creates an object with the same name as its fid.  Because this is not at all
 542  * performance sensitive, it is accomplished by creating a file, checking the
 543  * fid, and renaming it. */
 544 int mds_obd_create(struct obd_export *exp, struct obdo *oa,
 545                       struct lov_stripe_md **ea, struct obd_trans_info *oti)
 546 {
 547         struct mds_obd *mds = &exp->exp_obd->u.mds;
 548         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 549         unsigned int tmpname = ll_insecure_random_int();
 550         struct file *filp;
 551         struct dentry *new_child;
 552         struct obd_run_ctxt saved;
 553         char fidname[LL_FID_NAMELEN];
 554         void *handle;
 555         int rc = 0, err, namelen;
 556         ENTRY;
 557
 558         push_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
 559
 560         sprintf(fidname, "OBJECTS/%u", tmpname);
 561         filp = filp_open(fidname, O_CREAT | O_EXCL, 0644);
 562         if (IS_ERR(filp)) {
 563                 rc = PTR_ERR(filp);
 564                 if (rc == -EEXIST) {
 565                         CERROR("impossible object name collision %u\n",
 566                                tmpname);
 567                         LBUG();
 568                 }
 569                 CERROR("error creating tmp object %u: rc %d\n", tmpname, rc);
 570                 GOTO(out_pop, rc);
 571         }
 572
 573         LASSERT(mds->mds_objects_dir == filp->f_dentry->d_parent);
 574
 575         oa->o_id = filp->f_dentry->d_inode->i_ino;
 576         oa->o_generation = filp->f_dentry->d_inode->i_generation;
 577         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 578
 579         down(&parent_inode->i_sem);
 580         new_child = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 581
 582         if (IS_ERR(new_child)) {
 583                 CERROR("getting neg dentry for obj rename: %d\n", rc);
 584                 GOTO(out_close, rc = PTR_ERR(new_child));
 585         }
 586         if (new_child->d_inode != NULL) {
 587                 CERROR("impossible non-negative obj dentry " LPU64":%u!\n",
 588                        oa->o_id, oa->o_generation);
 589                 LBUG();
 590         }
 591
 592         handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
 593                               FSFILT_OP_RENAME, NULL);
 594         if (IS_ERR(handle))
 595                 GOTO(out_dput, rc = PTR_ERR(handle));
 596
 597         lock_kernel();
 598         rc = vfs_rename(mds->mds_objects_dir->d_inode, filp->f_dentry,
 599                         mds->mds_objects_dir->d_inode, new_child);
 600         unlock_kernel();
 601         if (rc)
 602                 CERROR("error renaming new object "LPU64":%u: rc %d\n",
 603                        oa->o_id, oa->o_generation, rc);
 604
 605         err = fsfilt_commit(exp->exp_obd, mds->mds_objects_dir->d_inode,
 606                             handle, 0);
 607         if (!err)
 608                 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
 609         else if (!rc)
 610                 rc = err;
 611 out_dput:
 612         dput(new_child);
 613 out_close:
 614         up(&parent_inode->i_sem);
 615         err = filp_close(filp, 0);
 616         if (err) {
 617                 CERROR("closing tmpfile %u: rc %d\n", tmpname, rc);
 618                 if (!rc)
 619                         rc = err;
 620         }
 621 out_pop:
 622         pop_ctxt(&saved, &exp->exp_obd->obd_ctxt, NULL);
 623         RETURN(rc);
 624 }
 625
 626 int mds_obd_destroy(struct obd_export *exp, struct obdo *oa,
 627                     struct lov_stripe_md *ea, struct obd_trans_info *oti)
 628 {
 629         struct mds_obd *mds = &exp->exp_obd->u.mds;
 630         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 631         struct obd_device *obd = exp->exp_obd;
 632         struct obd_run_ctxt saved;
 633         char fidname[LL_FID_NAMELEN];
 634         struct dentry *de;
 635         void *handle;
 636         int err, namelen, rc = 0;
 637         ENTRY;
 638
 639         push_ctxt(&saved, &obd->obd_ctxt, NULL);
 640
 641         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 642
 643         down(&parent_inode->i_sem);
 644         de = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 645         if (de == NULL || de->d_inode == NULL) {
 646                 CERROR("destroying non-existent object "LPU64"\n", oa->o_id);
 647                 GOTO(out, rc = IS_ERR(de) ? PTR_ERR(de) : -ENOENT);
 648         }
 649
 650         handle = fsfilt_start(obd, mds->mds_objects_dir->d_inode,
 651                               FSFILT_OP_UNLINK_LOG, oti);
 652         if (IS_ERR(handle)) {
 653                 GOTO(out_dput, rc = PTR_ERR(handle));
 654         }
 655
 656         rc = vfs_unlink(mds->mds_objects_dir->d_inode, de);
 657         if (rc)
 658                 CERROR("error destroying object "LPU64":%u: rc %d\n",
 659                        oa->o_id, oa->o_generation, rc);
 660
 661         err = fsfilt_commit(obd, mds->mds_objects_dir->d_inode, handle, 0);
 662         if (err && !rc)
 663                 rc = err;
 664 out_dput:
 665         l_dput(de);
 666 out:
 667         up(&parent_inode->i_sem);
 668         pop_ctxt(&saved, &obd->obd_ctxt, NULL);
 669         RETURN(rc);
 670 }