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 <linux/lustre_smfs.h>
  45 #include "mds_internal.h"
  46
  47 /* This limit is arbitrary, but for now we fit it in 1 page (32k clients) */
  48 #define MDS_MAX_CLIENTS (PAGE_SIZE * 8)
  49 #define MDS_MAX_CLIENT_WORDS (MDS_MAX_CLIENTS / sizeof(unsigned long))
  50
  51 #define LAST_RCVD "last_rcvd"
  52 #define LOV_OBJID "lov_objid"
  53
  54 /* Add client data to the MDS.  We use a bitmap to locate a free space
  55  * in the last_rcvd file if cl_off is -1 (i.e. a new client).
  56  * Otherwise, we have just read the data from the last_rcvd file and
  57  * we know its offset.
  58  */
  59 int mds_client_add(struct obd_device *obd, struct mds_obd *mds,
  60                    struct mds_export_data *med, int cl_idx)
  61 {
  62         unsigned long *bitmap = mds->mds_client_bitmap;
  63         int new_client = (cl_idx == -1);
  64         ENTRY;
  65
  66         LASSERT(bitmap != NULL);
  67
  68         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
  69         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
  70                 RETURN(0);
  71
  72         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
  73          * there's no need for extra complication here
  74          */
  75         if (new_client) {
  76                 cl_idx = find_first_zero_bit(bitmap, MDS_MAX_CLIENTS);
  77         repeat:
  78                 if (cl_idx >= MDS_MAX_CLIENTS) {
  79                         CERROR("no room for clients - fix MDS_MAX_CLIENTS\n");
  80                         return -ENOMEM;
  81                 }
  82                 if (test_and_set_bit(cl_idx, bitmap)) {
  83                         cl_idx = find_next_zero_bit(bitmap, MDS_MAX_CLIENTS,
  84                                                     cl_idx);
  85                         goto repeat;
  86                 }
  87         } else {
  88                 if (test_and_set_bit(cl_idx, bitmap)) {
  89                         CERROR("MDS client %d: bit already set in bitmap!!\n",
  90                                cl_idx);
  91                         LBUG();
  92                 }
  93         }
  94
  95         CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
  96                cl_idx, med->med_mcd->mcd_uuid);
  97
  98         med->med_idx = cl_idx;
  99         med->med_off = le32_to_cpu(mds->mds_server_data->msd_client_start) +
 100                 (cl_idx * le16_to_cpu(mds->mds_server_data->msd_client_size));
 101
 102         if (new_client) {
 103                 struct lvfs_run_ctxt saved;
 104                 loff_t off = med->med_off;
 105                 struct file *file = mds->mds_rcvd_filp;
 106                 int rc;
 107
 108                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 109                 rc = fsfilt_write_record(obd, file, med->med_mcd,
 110                                          sizeof(*med->med_mcd), &off, 1);
 111                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 112
 113                 if (rc)
 114                         return rc;
 115                 CDEBUG(D_INFO, "wrote client mcd at idx %u off %llu (len %u)\n",
 116                        med->med_idx, med->med_off,
 117                        (unsigned int)sizeof(*med->med_mcd));
 118         }
 119         return 0;
 120 }
 121
 122 int mds_client_free(struct obd_export *exp, int clear_client)
 123 {
 124         struct mds_export_data *med = &exp->exp_mds_data;
 125         struct mds_obd *mds = &exp->exp_obd->u.mds;
 126         struct obd_device *obd = exp->exp_obd;
 127         struct mds_client_data zero_mcd;
 128         struct lvfs_run_ctxt saved;
 129         int rc;
 130         unsigned long *bitmap = mds->mds_client_bitmap;
 131
 132         if (!med->med_mcd)
 133                 RETURN(0);
 134
 135         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 136         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
 137                 GOTO(free_and_out, 0);
 138
 139         CDEBUG(D_INFO, "freeing client at idx %u (%lld)with UUID '%s'\n",
 140                med->med_idx, med->med_off, med->med_mcd->mcd_uuid);
 141
 142         LASSERT(bitmap);
 143
 144         /* Clear the bit _after_ zeroing out the client so we don't
 145            race with mds_client_add and zero out new clients.*/
 146         if (!test_bit(med->med_idx, bitmap)) {
 147                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 148                        med->med_idx);
 149                 LBUG();
 150         }
 151
 152         if (clear_client) {
 153                 memset(&zero_mcd, 0, sizeof zero_mcd);
 154                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 155                 rc = fsfilt_write_record(obd, mds->mds_rcvd_filp, &zero_mcd,
 156                                          sizeof(zero_mcd), &med->med_off, 1);
 157                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 158
 159                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 160                        "zeroing out client %s idx %u in %s rc %d\n",
 161                        med->med_mcd->mcd_uuid, med->med_idx, LAST_RCVD, rc);
 162         }
 163
 164         if (!test_and_clear_bit(med->med_idx, bitmap)) {
 165                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 166                        med->med_idx);
 167                 LBUG();
 168         }
 169
 170
 171         /* Make sure the server's last_transno is up to date. Do this
 172          * after the client is freed so we know all the client's
 173          * transactions have been committed. */
 174         mds_update_server_data(exp->exp_obd, 1);
 175
 176  free_and_out:
 177         OBD_FREE(med->med_mcd, sizeof(*med->med_mcd));
 178
 179         return 0;
 180 }
 181
 182 static int mds_server_free_data(struct mds_obd *mds)
 183 {
 184         OBD_FREE(mds->mds_client_bitmap,
 185                  MDS_MAX_CLIENT_WORDS * sizeof(unsigned long));
 186         OBD_FREE(mds->mds_server_data, sizeof(*mds->mds_server_data));
 187         mds->mds_server_data = NULL;
 188
 189         return 0;
 190 }
 191
 192 static int mds_read_last_rcvd(struct obd_device *obd, struct file *file)
 193 {
 194         struct mds_obd *mds = &obd->u.mds;
 195         struct mds_server_data *msd;
 196         struct mds_client_data *mcd = NULL;
 197         loff_t off = 0;
 198         unsigned long last_rcvd_size = file->f_dentry->d_inode->i_size;
 199         __u64 mount_count;
 200         int cl_idx, rc = 0;
 201         ENTRY;
 202
 203         /* ensure padding in the struct is the correct size */
 204         LASSERT(offsetof(struct mds_server_data, msd_padding) +
 205                 sizeof(msd->msd_padding) == MDS_LR_SERVER_SIZE);
 206         LASSERT(offsetof(struct mds_client_data, mcd_padding) +
 207                 sizeof(mcd->mcd_padding) == MDS_LR_CLIENT_SIZE);
 208
 209         OBD_ALLOC_WAIT(msd, sizeof(*msd));
 210         if (!msd)
 211                 RETURN(-ENOMEM);
 212
 213         OBD_ALLOC_WAIT(mds->mds_client_bitmap,
 214                   MDS_MAX_CLIENT_WORDS * sizeof(unsigned long));
 215         if (!mds->mds_client_bitmap) {
 216                 OBD_FREE(msd, sizeof(*msd));
 217                 RETURN(-ENOMEM);
 218         }
 219
 220         mds->mds_server_data = msd;
 221
 222         if (last_rcvd_size == 0) {
 223                 CWARN("%s: initializing new %s\n", obd->obd_name, LAST_RCVD);
 224
 225                 memcpy(msd->msd_uuid, obd->obd_uuid.uuid,sizeof(msd->msd_uuid));
 226                 msd->msd_last_transno = 0;
 227                 mount_count = msd->msd_mount_count = 0;
 228                 msd->msd_server_size = cpu_to_le32(MDS_LR_SERVER_SIZE);
 229                 msd->msd_client_start = cpu_to_le32(MDS_LR_CLIENT_START);
 230                 msd->msd_client_size = cpu_to_le16(MDS_LR_CLIENT_SIZE);
 231                 msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
 232         } else {
 233                 rc = fsfilt_read_record(obd, file, msd, sizeof(*msd), &off);
 234                 if (rc) {
 235                         CERROR("error reading MDS %s: rc = %d\n", LAST_RCVD, rc);
 236                         GOTO(err_msd, rc);
 237                 }
 238                 if (strcmp(msd->msd_uuid, obd->obd_uuid.uuid) != 0) {
 239                         CERROR("OBD UUID %s does not match last_rcvd UUID %s\n",
 240                                obd->obd_uuid.uuid, msd->msd_uuid);
 241                         GOTO(err_msd, rc = -EINVAL);
 242                 }
 243                 mount_count = le64_to_cpu(msd->msd_mount_count);
 244         }
 245         if (msd->msd_feature_incompat & ~cpu_to_le32(MDS_INCOMPAT_SUPP)) {
 246                 CERROR("unsupported incompat feature %x\n",
 247                        le32_to_cpu(msd->msd_feature_incompat) &
 248                        ~MDS_INCOMPAT_SUPP);
 249                 GOTO(err_msd, rc = -EINVAL);
 250         }
 251         /* XXX updating existing b_devel fs only, can be removed in future */
 252         msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
 253         if (msd->msd_feature_rocompat & ~cpu_to_le32(MDS_ROCOMPAT_SUPP)) {
 254                 CERROR("unsupported read-only feature %x\n",
 255                        le32_to_cpu(msd->msd_feature_rocompat) &
 256                        ~MDS_ROCOMPAT_SUPP);
 257                 /* Do something like remount filesystem read-only */
 258                 GOTO(err_msd, rc = -EINVAL);
 259         }
 260
 261         mds->mds_last_transno = le64_to_cpu(msd->msd_last_transno);
 262
 263         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 264                obd->obd_name, mds->mds_last_transno);
 265         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 266                obd->obd_name, mount_count + 1);
 267         CDEBUG(D_INODE, "%s: server data size: %u\n",
 268                obd->obd_name, le32_to_cpu(msd->msd_server_size));
 269         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 270                obd->obd_name, le32_to_cpu(msd->msd_client_start));
 271         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 272                obd->obd_name, le32_to_cpu(msd->msd_client_size));
 273         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 274                obd->obd_name, last_rcvd_size);
 275         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 276                last_rcvd_size <= le32_to_cpu(msd->msd_client_start) ? 0 :
 277                (last_rcvd_size - le32_to_cpu(msd->msd_client_start)) /
 278                 le16_to_cpu(msd->msd_client_size));
 279
 280         /* When we do a clean MDS shutdown, we save the last_transno into
 281          * the header.  If we find clients with higher last_transno values
 282          * then those clients may need recovery done. */
 283         for (cl_idx = 0, off = le32_to_cpu(msd->msd_client_start);
 284              off < last_rcvd_size; cl_idx++) {
 285                 __u64 last_transno;
 286                 struct obd_export *exp;
 287                 struct mds_export_data *med;
 288
 289                 if (!mcd) {
 290                         OBD_ALLOC_WAIT(mcd, sizeof(*mcd));
 291                         if (!mcd)
 292                                 GOTO(err_client, rc = -ENOMEM);
 293                 }
 294
 295                 /* Don't assume off is incremented properly by
 296                  * fsfilt_read_record(), in case sizeof(*mcd)
 297                  * isn't the same as msd->msd_client_size.  */
 298                 off = le32_to_cpu(msd->msd_client_start) +
 299                         cl_idx * le16_to_cpu(msd->msd_client_size);
 300                 rc = fsfilt_read_record(obd, file, mcd, sizeof(*mcd), &off);
 301                 if (rc) {
 302                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 303                                LAST_RCVD, cl_idx, off, rc);
 304                         break; /* read error shouldn't cause startup to fail */
 305                 }
 306
 307                 if (mcd->mcd_uuid[0] == '\0') {
 308                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 309                                cl_idx);
 310                         continue;
 311                 }
 312
 313                 last_transno = le64_to_cpu(mcd->mcd_last_transno);
 314
 315                 /* These exports are cleaned up by mds_disconnect(), so they
 316                  * need to be set up like real exports as mds_connect() does.
 317                  */
 318                 CDEBUG(D_HA|D_WARNING,"RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 319                        " srv lr: "LPU64" lx: "LPU64"\n", mcd->mcd_uuid, cl_idx,
 320                        last_transno, le64_to_cpu(msd->msd_last_transno),
 321                        mcd->mcd_last_xid);
 322
 323                 exp = class_new_export(obd);
 324                 if (exp == NULL)
 325                         GOTO(err_client, rc = -ENOMEM);
 326
 327                 memcpy(&exp->exp_client_uuid.uuid, mcd->mcd_uuid,
 328                        sizeof exp->exp_client_uuid.uuid);
 329                 med = &exp->exp_mds_data;
 330                 med->med_mcd = mcd;
 331                 mds_client_add(obd, mds, med, cl_idx);
 332                 /* create helper if export init gets more complex */
 333                 INIT_LIST_HEAD(&med->med_open_head);
 334                 spin_lock_init(&med->med_open_lock);
 335
 336                 mcd = NULL;
 337                 exp->exp_replay_needed = 1;
 338                 obd->obd_recoverable_clients++;
 339                 obd->obd_max_recoverable_clients++;
 340                 class_export_put(exp);
 341
 342                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
 343                        cl_idx, last_transno);
 344
 345                 if (last_transno > mds->mds_last_transno)
 346                        mds->mds_last_transno = last_transno;
 347         }
 348         if (mcd)
 349                 OBD_FREE(mcd, sizeof(*mcd));
 350         obd->obd_last_committed = mds->mds_last_transno;
 351         if (obd->obd_recoverable_clients) {
 352                 CWARN("RECOVERY: service %s, %d recoverable clients, "
 353                       "last_transno "LPU64"\n", obd->obd_name,
 354                       obd->obd_recoverable_clients, mds->mds_last_transno);
 355                 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
 356                 target_start_recovery_thread(obd, mds_handle);
 357                 obd->obd_recovery_start = LTIME_S(CURRENT_TIME);
 358         }
 359
 360         mds->mds_mount_count = mount_count + 1;
 361         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
 362
 363         /* save it, so mount count and last_transno is current */
 364         rc = mds_update_server_data(obd, 1);
 365         if (rc)
 366                 GOTO(err_client, rc);
 367
 368         RETURN(0);
 369
 370 err_client:
 371         class_disconnect_exports(obd, 0);
 372 err_msd:
 373         mds_server_free_data(mds);
 374         RETURN(rc);
 375 }
 376
 377 static int  mds_fs_post_setup(struct obd_device *obd)
 378 {
 379         struct mds_obd *mds = &obd->u.mds;
 380         struct dentry *de = mds_fid2dentry(mds, &mds->mds_rootfid, NULL);
 381         int    rc = 0;
 382
 383         rc = fsfilt_post_setup(obd, de);
 384         if (rc)
 385                 GOTO(out, rc);
 386
 387         fsfilt_set_fs_flags(obd, de->d_inode,
 388                               SM_DO_REC | SM_DO_COW);
 389         fsfilt_set_fs_flags(obd, mds->mds_pending_dir->d_inode,
 390                               SM_DO_REC | SM_DO_COW);
 391         fsfilt_set_mds_flags(obd, mds->mds_sb);
 392 out:
 393         l_dput(de);
 394         return rc;
 395 }
 396
 397 int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
 398 {
 399         struct mds_obd *mds = &obd->u.mds;
 400         struct lvfs_run_ctxt saved;
 401         struct dentry *dentry;
 402         struct file *file;
 403         int rc;
 404         ENTRY;
 405
 406         rc = cleanup_group_info();
 407         if (rc)
 408                 RETURN(rc);
 409
 410         mds->mds_vfsmnt = mnt;
 411         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 412
 413         fsfilt_setup(obd, mds->mds_sb);
 414
 415         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 416         obd->obd_lvfs_ctxt.pwdmnt = mnt;
 417         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
 418         obd->obd_lvfs_ctxt.fs = get_ds();
 419         obd->obd_lvfs_ctxt.cb_ops = mds_lvfs_ops;
 420
 421         /* setup the directory tree */
 422         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 423         dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755, 0);
 424         if (IS_ERR(dentry)) {
 425                 rc = PTR_ERR(dentry);
 426                 CERROR("cannot create ROOT directory: rc = %d\n", rc);
 427                 GOTO(err_pop, rc);
 428         }
 429
 430         mds->mds_rootfid.id = dentry->d_inode->i_ino;
 431         mds->mds_rootfid.generation = dentry->d_inode->i_generation;
 432         mds->mds_rootfid.f_type = S_IFDIR;
 433
 434         dput(dentry);
 435
 436         dentry = lookup_one_len("__iopen__", current->fs->pwd,
 437                                 strlen("__iopen__"));
 438         if (IS_ERR(dentry)) {
 439                 rc = PTR_ERR(dentry);
 440                 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
 441                 GOTO(err_pop, rc);
 442         }
 443         if (!dentry->d_inode) {
 444                 rc = -ENOENT;
 445                 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
 446                 GOTO(err_fid, rc);
 447         }
 448         mds->mds_fid_de = dentry;
 449
 450         dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777, 1);
 451         if (IS_ERR(dentry)) {
 452                 rc = PTR_ERR(dentry);
 453                 CERROR("cannot create PENDING directory: rc = %d\n", rc);
 454                 GOTO(err_fid, rc);
 455         }
 456         mds->mds_pending_dir = dentry;
 457
 458         dentry = simple_mkdir(current->fs->pwd, "LOGS", 0777, 1);
 459         if (IS_ERR(dentry)) {
 460                 rc = PTR_ERR(dentry);
 461                 CERROR("cannot create LOGS directory: rc = %d\n", rc);
 462                 GOTO(err_pending, rc);
 463         }
 464         mds->mds_logs_dir = dentry;
 465
 466         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
 467         if (IS_ERR(dentry)) {
 468                 rc = PTR_ERR(dentry);
 469                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
 470                 GOTO(err_logs, rc);
 471         }
 472         mds->mds_objects_dir = dentry;
 473
 474         dentry = simple_mkdir(current->fs->pwd, "FIDS", 0777, 1);
 475         if (IS_ERR(dentry)) {
 476                 rc = PTR_ERR(dentry);
 477                 CERROR("cannot create FIDS directory: rc = %d\n", rc);
 478                 GOTO(err_fids, rc);
 479         }
 480         mds->mds_fids_dir = dentry;
 481
 482         dentry = simple_mkdir(current->fs->pwd, "UNNAMED", 0777, 1);
 483         if (IS_ERR(dentry)) {
 484                 rc = PTR_ERR(dentry);
 485                 CERROR("cannot create UNNAMED directory: rc = %d\n", rc);
 486                 GOTO(err_unnamed, rc);
 487         }
 488         mds->mds_unnamed_dir = dentry;
 489
 490         /* open and test the last rcvd file */
 491         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
 492         if (IS_ERR(file)) {
 493                 rc = PTR_ERR(file);
 494                 CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
 495                 GOTO(err_objects, rc = PTR_ERR(file));
 496         }
 497         mds->mds_rcvd_filp = file;
 498         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 499                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 500                        file->f_dentry->d_inode->i_mode);
 501                 GOTO(err_last_rcvd, rc = -ENOENT);
 502         }
 503
 504         rc = mds_read_last_rcvd(obd, file);
 505         if (rc) {
 506                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
 507                 GOTO(err_last_rcvd, rc);
 508         }
 509
 510         /* open and test the lov objd file */
 511         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
 512         if (IS_ERR(file)) {
 513                 rc = PTR_ERR(file);
 514                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
 515                 GOTO(err_client, rc = PTR_ERR(file));
 516         }
 517         mds->mds_lov_objid_filp = file;
 518         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 519                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
 520                        file->f_dentry->d_inode->i_mode);
 521                 GOTO(err_lov_objid, rc = -ENOENT);
 522         }
 523 err_pop:
 524         if (!rc) {
 525                 rc = mds_fs_post_setup(obd);
 526                 if (rc)
 527                         CERROR("can not post setup fsfilt\n");
 528         }
 529         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 530         return rc;
 531
 532 err_lov_objid:
 533         if (mds->mds_lov_objid_filp && filp_close(mds->mds_lov_objid_filp, 0))
 534                 CERROR("can't close %s after error\n", LOV_OBJID);
 535 err_client:
 536         class_disconnect_exports(obd, 0);
 537 err_last_rcvd:
 538         if (mds->mds_rcvd_filp && filp_close(mds->mds_rcvd_filp, 0))
 539                 CERROR("can't close %s after error\n", LAST_RCVD);
 540 err_unnamed:
 541         dput(mds->mds_unnamed_dir);
 542 err_fids:
 543         dput(mds->mds_fids_dir);
 544 err_objects:
 545         dput(mds->mds_objects_dir);
 546 err_logs:
 547         dput(mds->mds_logs_dir);
 548 err_pending:
 549         dput(mds->mds_pending_dir);
 550 err_fid:
 551         dput(mds->mds_fid_de);
 552         goto err_pop;
 553 }
 554
 555 static int  mds_fs_post_cleanup(struct obd_device *obd)
 556 {
 557         int    rc = 0;
 558         rc = fsfilt_post_cleanup(obd);
 559         return rc;
 560 }
 561
 562 int mds_fs_cleanup(struct obd_device *obd, int flags)
 563 {
 564         struct mds_obd *mds = &obd->u.mds;
 565         struct lvfs_run_ctxt saved;
 566         int rc = 0;
 567
 568         if (flags & OBD_OPT_FAILOVER)
 569                 CERROR("%s: shutting down for failover; client state will"
 570                        " be preserved.\n", obd->obd_name);
 571
 572         class_disconnect_exports(obd, flags); /* cleans up client info too */
 573         target_cleanup_recovery(obd);
 574         mds_server_free_data(mds);
 575
 576         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 577         if (mds->mds_rcvd_filp) {
 578                 rc = filp_close(mds->mds_rcvd_filp, 0);
 579                 mds->mds_rcvd_filp = NULL;
 580                 if (rc)
 581                         CERROR("%s file won't close, rc=%d\n", LAST_RCVD, rc);
 582         }
 583         if (mds->mds_lov_objid_filp) {
 584                 rc = filp_close(mds->mds_lov_objid_filp, 0);
 585                 mds->mds_lov_objid_filp = NULL;
 586                 if (rc)
 587                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
 588         }
 589         if (mds->mds_unnamed_dir != NULL) {
 590                 l_dput(mds->mds_unnamed_dir);
 591                 mds->mds_unnamed_dir = NULL;
 592         }
 593         if (mds->mds_fids_dir != NULL) {
 594                 l_dput(mds->mds_fids_dir);
 595                 mds->mds_fids_dir = NULL;
 596         }
 597         if (mds->mds_objects_dir != NULL) {
 598                 l_dput(mds->mds_objects_dir);
 599                 mds->mds_objects_dir = NULL;
 600         }
 601         if (mds->mds_logs_dir) {
 602                 l_dput(mds->mds_logs_dir);
 603                 mds->mds_logs_dir = NULL;
 604         }
 605         if (mds->mds_pending_dir) {
 606                 l_dput(mds->mds_pending_dir);
 607                 mds->mds_pending_dir = NULL;
 608         }
 609         rc = mds_fs_post_cleanup(obd);
 610
 611         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 612         shrink_dcache_parent(mds->mds_fid_de);
 613         dput(mds->mds_fid_de);
 614
 615         return rc;
 616 }
 617
 618 /* Creates an object with the same name as its fid.  Because this is not at all
 619  * performance sensitive, it is accomplished by creating a file, checking the
 620  * fid, and renaming it. */
 621 int mds_obd_create(struct obd_export *exp, struct obdo *oa,
 622                    struct lov_stripe_md **ea, struct obd_trans_info *oti)
 623 {
 624         struct mds_obd *mds = &exp->exp_obd->u.mds;
 625         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 626         struct file *filp;
 627         struct dentry *dchild;
 628         struct lvfs_run_ctxt saved;
 629         char fidname[LL_FID_NAMELEN];
 630         void *handle;
 631         int rc = 0, err, namelen;
 632         ENTRY;
 633
 634         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 635         down(&parent_inode->i_sem);
 636         if (oa->o_id) {
 637                 namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 638
 639                 dchild = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 640                 if (IS_ERR(dchild))
 641                         GOTO(out_pop, rc = PTR_ERR(dchild));
 642
 643                 if (dchild->d_inode == NULL) {
 644                         struct dentry_params dp;
 645                         struct inode *inode;
 646
 647                         CWARN("creating log with ID "LPU64"\n", oa->o_id);
 648
 649                         dchild->d_fsdata = (void *) &dp;
 650                         dp.p_ptr = NULL;
 651                         dp.p_inum = oa->o_id;
 652                         rc = ll_vfs_create(parent_inode, dchild, S_IFREG, NULL);
 653                         if (dchild->d_fsdata == (void *)(unsigned long)oa->o_id)
 654                                 dchild->d_fsdata = NULL;
 655                         if (rc) {
 656                                 CDEBUG(D_INODE, "err during create: %d\n", rc);
 657                                 dput(dchild);
 658                                 GOTO(out_pop, rc);
 659                         }
 660                         inode = dchild->d_inode;
 661                         LASSERT(inode->i_ino == oa->o_id);
 662                         inode->i_generation = oa->o_generation;
 663                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
 664                                inode->i_ino, inode->i_generation);
 665                         mark_inode_dirty(inode);
 666                 } else {
 667                         CWARN("it should be here!\n");
 668                 }
 669                 GOTO(out_pop, rc);
 670         }
 671
 672         sprintf(fidname, "OBJECTS/%u.%u",ll_insecure_random_int(),current->pid);
 673         filp = filp_open(fidname, O_CREAT | O_EXCL, 0644);
 674         if (IS_ERR(filp)) {
 675                 rc = PTR_ERR(filp);
 676                 if (rc == -EEXIST) {
 677                         CERROR("impossible object name collision %s\n",
 678                                fidname);
 679                         LBUG();
 680                 }
 681                 CERROR("error creating tmp object %s: rc %d\n", fidname, rc);
 682                 GOTO(out_pop, rc);
 683         }
 684
 685         LASSERT(mds->mds_objects_dir == filp->f_dentry->d_parent);
 686
 687         oa->o_id = filp->f_dentry->d_inode->i_ino;
 688         oa->o_generation = filp->f_dentry->d_inode->i_generation;
 689         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 690         CWARN("created log anonymous "LPU64"/%u\n", oa->o_id, oa->o_generation);
 691
 692         dchild = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 693
 694         if (IS_ERR(dchild)) {
 695                 CERROR("getting neg dentry for obj rename: %d\n", rc);
 696                 GOTO(out_close, rc = PTR_ERR(dchild));
 697         }
 698         if (dchild->d_inode != NULL) {
 699                 CERROR("impossible non-negative obj dentry " LPU64":%u!\n",
 700                        oa->o_id, oa->o_generation);
 701                 LBUG();
 702         }
 703
 704         handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
 705                               FSFILT_OP_RENAME, NULL);
 706         if (IS_ERR(handle))
 707                 GOTO(out_dput, rc = PTR_ERR(handle));
 708
 709         lock_kernel();
 710         rc = vfs_rename(mds->mds_objects_dir->d_inode, filp->f_dentry,
 711                         mds->mds_objects_dir->d_inode, dchild);
 712         unlock_kernel();
 713         if (rc)
 714                 CERROR("error renaming new object "LPU64":%u: rc %d\n",
 715                        oa->o_id, oa->o_generation, rc);
 716
 717         err = fsfilt_commit(exp->exp_obd, mds->mds_sb,
 718                             mds->mds_objects_dir->d_inode, handle, 0);
 719         if (!err) {
 720                 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
 721                 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLGROUP;
 722         } else if (!rc)
 723                 rc = err;
 724 out_dput:
 725         dput(dchild);
 726 out_close:
 727         err = filp_close(filp, 0);
 728         if (err) {
 729                 CERROR("closing tmpfile %s: rc %d\n", fidname, rc);
 730                 if (!rc)
 731                         rc = err;
 732         }
 733 out_pop:
 734         up(&parent_inode->i_sem);
 735         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 736         RETURN(rc);
 737 }
 738
 739 int mds_obd_destroy(struct obd_export *exp, struct obdo *oa,
 740                     struct lov_stripe_md *ea, struct obd_trans_info *oti)
 741 {
 742         struct mds_obd *mds = &exp->exp_obd->u.mds;
 743         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 744         struct obd_device *obd = exp->exp_obd;
 745         struct lvfs_run_ctxt saved;
 746         char fidname[LL_FID_NAMELEN];
 747         struct dentry *de;
 748         void *handle;
 749         int err, namelen, rc = 0;
 750         ENTRY;
 751
 752         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 753
 754         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 755
 756         down(&parent_inode->i_sem);
 757         de = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 758         if (IS_ERR(de) || de->d_inode == NULL) {
 759                 rc = IS_ERR(de) ? PTR_ERR(de) : -ENOENT;
 760                 CERROR("destroying non-existent object "LPU64" %s: rc %d\n",
 761                        oa->o_id, fidname, rc);
 762                 GOTO(out_dput, rc);
 763         }
 764         /* Stripe count is 1 here since this is some MDS specific stuff
 765            that is unlinked, not spanned across multiple OSTs */
 766         handle = fsfilt_start_log(obd, mds->mds_objects_dir->d_inode,
 767                                   FSFILT_OP_UNLINK, oti, 1);
 768
 769         if (IS_ERR(handle))
 770                 GOTO(out_dput, rc = PTR_ERR(handle));
 771
 772         rc = vfs_unlink(mds->mds_objects_dir->d_inode, de);
 773         if (rc)
 774                 CERROR("error destroying object "LPU64":%u: rc %d\n",
 775                        oa->o_id, oa->o_generation, rc);
 776
 777         err = fsfilt_commit(obd, mds->mds_sb, mds->mds_objects_dir->d_inode,
 778                             handle, 0);
 779         if (err && !rc)
 780                 rc = err;
 781 out_dput:
 782         if (de != NULL)
 783                 l_dput(de);
 784         up(&parent_inode->i_sem);
 785         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 786         RETURN(rc);
 787 }