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 the Lustre file system, http://www.lustre.org
  11  *   Lustre is a trademark of Cluster File Systems, Inc.
  12  *
  13  *   You may have signed or agreed to another license before downloading
  14  *   this software.  If so, you are bound by the terms and conditions
  15  *   of that agreement, and the following does not apply to you.  See the
  16  *   LICENSE file included with this distribution for more information.
  17  *
  18  *   If you did not agree to a different license, then this copy of Lustre
  19  *   is open source software; you can redistribute it and/or modify it
  20  *   under the terms of version 2 of the GNU General Public License as
  21  *   published by the Free Software Foundation.
  22  *
  23  *   In either case, Lustre is distributed in the hope that it will be
  24  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  25  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  26  *   license text for more details.
  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/kmod.h>
  36 #include <linux/version.h>
  37 #include <linux/sched.h>
  38 #include <lustre_quota.h>
  39 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  40 #include <linux/mount.h>
  41 #endif
  42 #include <lustre_mds.h>
  43 #include <obd_class.h>
  44 #include <obd_support.h>
  45 #include <lustre_lib.h>
  46 #include <lustre_fsfilt.h>
  47 #include <lustre_disk.h>
  48 #include <libcfs/list.h>
  49
  50 #include "mds_internal.h"
  51
  52
  53 static int mds_export_stats_init(struct obd_device *obd, struct obd_export *exp)
  54 {
  55         int rc, num_stats;
  56
  57         rc = lprocfs_exp_setup(exp);
  58         if (rc)
  59                 return rc;
  60         num_stats = (sizeof(*obd->obd_type->typ_ops) / sizeof(void *)) +
  61                      LPROC_MDS_LAST - 1;
  62         exp->exp_ops_stats = lprocfs_alloc_stats(num_stats);
  63         if (exp->exp_ops_stats == NULL)
  64                 return -ENOMEM;
  65         lprocfs_init_ops_stats(LPROC_MDS_LAST, exp->exp_ops_stats);
  66         mds_stats_counter_init(exp->exp_ops_stats);
  67         lprocfs_register_stats(exp->exp_proc, "stats", exp->exp_ops_stats);
  68         return 0;
  69 }
  70
  71 /* Add client data to the MDS.  We use a bitmap to locate a free space
  72  * in the last_rcvd file if cl_off is -1 (i.e. a new client).
  73  * Otherwise, we have just read the data from the last_rcvd file and
  74  * we know its offset.
  75  *
  76  * It should not be possible to fail adding an existing client - otherwise
  77  * mds_init_server_data() callsite needs to be fixed.
  78  */
  79 int mds_client_add(struct obd_device *obd, struct obd_export *exp,
  80                    int cl_idx)
  81 {
  82         struct mds_obd *mds = &obd->u.mds;
  83         struct mds_export_data *med = &exp->exp_mds_data;
  84         unsigned long *bitmap = mds->mds_client_bitmap;
  85         int new_client = (cl_idx == -1);
  86         int rc;
  87         ENTRY;
  88
  89         LASSERT(bitmap != NULL);
  90         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
  91
  92         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
  93         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
  94                 RETURN(0);
  95
  96         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
  97          * there's no need for extra complication here
  98          */
  99         if (new_client) {
 100                 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
 101         repeat:
 102                 if (cl_idx >= LR_MAX_CLIENTS ||
 103                     OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_CLIENT_ADD)) {
 104                         CERROR("no room for %u clients - fix LR_MAX_CLIENTS\n",
 105                                cl_idx);
 106                         return -EOVERFLOW;
 107                 }
 108                 if (test_and_set_bit(cl_idx, bitmap)) {
 109                         cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
 110                                                     cl_idx);
 111                         goto repeat;
 112                 }
 113         } else {
 114                 if (test_and_set_bit(cl_idx, bitmap)) {
 115                         CERROR("MDS client %d: bit already set in bitmap!!\n",
 116                                cl_idx);
 117                         LBUG();
 118                 }
 119         }
 120
 121         CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
 122                cl_idx, med->med_mcd->mcd_uuid);
 123
 124         med->med_lr_idx = cl_idx;
 125         med->med_lr_off = le32_to_cpu(mds->mds_server_data->lsd_client_start) +
 126                 (cl_idx * le16_to_cpu(mds->mds_server_data->lsd_client_size));
 127         LASSERTF(med->med_lr_off > 0, "med_lr_off = %llu\n", med->med_lr_off);
 128         mds_export_stats_init(obd, exp);
 129
 130         if (new_client) {
 131                 struct lvfs_run_ctxt saved;
 132                 loff_t off = med->med_lr_off;
 133                 struct file *file = mds->mds_rcvd_filp;
 134                 void *handle;
 135
 136                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 137                 handle = fsfilt_start(obd, file->f_dentry->d_inode,
 138                                       FSFILT_OP_SETATTR, NULL);
 139                 if (IS_ERR(handle)) {
 140                         rc = PTR_ERR(handle);
 141                         CERROR("unable to start transaction: rc %d\n", rc);
 142                 } else {
 143                         rc = fsfilt_add_journal_cb(obd, 0, handle,
 144                                                    target_client_add_cb, exp);
 145                         if (rc == 0) {
 146                                 spin_lock(&exp->exp_lock);
 147                                 exp->exp_need_sync = 1;
 148                                 spin_unlock(&exp->exp_lock);
 149                         }
 150                         rc = fsfilt_write_record(obd, file, med->med_mcd,
 151                                                  sizeof(*med->med_mcd),
 152                                                  &off, rc /* sync if no cb */);
 153                         fsfilt_commit(obd, file->f_dentry->d_inode, handle, 0);
 154                 }
 155
 156                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 157
 158                 if (rc)
 159                         return rc;
 160                 CDEBUG(D_INFO, "wrote client mcd at idx %u off %llu (len %u)\n",
 161                        med->med_lr_idx, med->med_lr_off,
 162                        (unsigned int)sizeof(*med->med_mcd));
 163         }
 164         return 0;
 165 }
 166
 167 int mds_client_free(struct obd_export *exp)
 168 {
 169         struct mds_export_data *med = &exp->exp_mds_data;
 170         struct mds_obd *mds = &exp->exp_obd->u.mds;
 171         struct obd_device *obd = exp->exp_obd;
 172         struct mds_client_data zero_mcd;
 173         struct lvfs_run_ctxt saved;
 174         int rc;
 175         loff_t off;
 176         ENTRY;
 177
 178         if (!med->med_mcd)
 179                 RETURN(0);
 180
 181         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 182         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
 183                 GOTO(free, 0);
 184
 185         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
 186                med->med_lr_idx, med->med_lr_off, med->med_mcd->mcd_uuid);
 187
 188         LASSERT(mds->mds_client_bitmap != NULL);
 189
 190         lprocfs_exp_cleanup(exp);
 191
 192         off = med->med_lr_off;
 193
 194         /* Don't clear med_lr_idx here as it is likely also unset.  At worst
 195          * we leak a client slot that will be cleaned on the next recovery. */
 196         if (off <= 0) {
 197                 CERROR("%s: client idx %d has offset %lld\n",
 198                         obd->obd_name, med->med_lr_idx, off);
 199                 GOTO(free, rc = -EINVAL);
 200         }
 201
 202         /* Clear the bit _after_ zeroing out the client so we don't
 203            race with mds_client_add and zero out new clients.*/
 204         if (!test_bit(med->med_lr_idx, mds->mds_client_bitmap)) {
 205                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 206                        med->med_lr_idx);
 207                 LBUG();
 208         }
 209
 210         if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
 211                 memset(&zero_mcd, 0, sizeof zero_mcd);
 212                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 213                 rc = fsfilt_write_record(obd, mds->mds_rcvd_filp, &zero_mcd,
 214                                          sizeof(zero_mcd), &off,
 215                                          (!exp->exp_libclient || exp->exp_need_sync));
 216                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 217
 218                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 219                        "zeroing out client %s idx %u in %s rc %d\n",
 220                        med->med_mcd->mcd_uuid, med->med_lr_idx, LAST_RCVD, rc);
 221         }
 222
 223         if (!test_and_clear_bit(med->med_lr_idx, mds->mds_client_bitmap)) {
 224                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 225                        med->med_lr_idx);
 226                 LBUG();
 227         }
 228
 229
 230         /* Make sure the server's last_transno is up to date. Do this
 231          * after the client is freed so we know all the client's
 232          * transactions have been committed. */
 233         mds_update_server_data(exp->exp_obd, 0);
 234
 235         EXIT;
 236  free:
 237         OBD_FREE(med->med_mcd, sizeof(*med->med_mcd));
 238         med->med_mcd = NULL;
 239
 240         return 0;
 241 }
 242
 243 static int mds_server_free_data(struct mds_obd *mds)
 244 {
 245         OBD_FREE(mds->mds_client_bitmap, LR_MAX_CLIENTS / 8);
 246         OBD_FREE(mds->mds_server_data, sizeof(*mds->mds_server_data));
 247         mds->mds_server_data = NULL;
 248
 249         return 0;
 250 }
 251
 252 static int mds_init_server_data(struct obd_device *obd, struct file *file)
 253 {
 254         struct mds_obd *mds = &obd->u.mds;
 255         struct lr_server_data *lsd;
 256         struct mds_client_data *mcd = NULL;
 257         loff_t off = 0;
 258         unsigned long last_rcvd_size = file->f_dentry->d_inode->i_size;
 259         __u64 mount_count;
 260         int cl_idx, rc = 0;
 261         ENTRY;
 262
 263         /* ensure padding in the struct is the correct size */
 264         LASSERT(offsetof(struct lr_server_data, lsd_padding) +
 265                 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
 266         LASSERT(offsetof(struct mds_client_data, mcd_padding) +
 267                 sizeof(mcd->mcd_padding) == LR_CLIENT_SIZE);
 268
 269         OBD_ALLOC_WAIT(lsd, sizeof(*lsd));
 270         if (!lsd)
 271                 RETURN(-ENOMEM);
 272
 273         OBD_ALLOC_WAIT(mds->mds_client_bitmap, LR_MAX_CLIENTS / 8);
 274         if (!mds->mds_client_bitmap) {
 275                 OBD_FREE(lsd, sizeof(*lsd));
 276                 RETURN(-ENOMEM);
 277         }
 278
 279         mds->mds_server_data = lsd;
 280
 281         if (last_rcvd_size == 0) {
 282                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
 283
 284                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
 285                 lsd->lsd_last_transno = 0;
 286                 mount_count = lsd->lsd_mount_count = 0;
 287                 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
 288                 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
 289                 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
 290                 lsd->lsd_feature_rocompat = cpu_to_le32(OBD_ROCOMPAT_LOVOBJID);
 291                 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_MDT);
 292         } else {
 293                 rc = fsfilt_read_record(obd, file, lsd, sizeof(*lsd), &off);
 294                 if (rc) {
 295                         CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
 296                         GOTO(err_msd, rc);
 297                 }
 298                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
 299                         LCONSOLE_ERROR("Trying to start OBD %s using the wrong"
 300                                        " disk %s. Were the /dev/ assignments "
 301                                        "rearranged?\n",
 302                                        obd->obd_uuid.uuid, lsd->lsd_uuid);
 303                         GOTO(err_msd, rc = -EINVAL);
 304                 }
 305                 /* COMPAT_146 */
 306                 /* Assume old last_rcvd format unless I_C_LR is set */
 307                 if (!(lsd->lsd_feature_incompat &
 308                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
 309                         lsd->lsd_mount_count = lsd->lsd_compat14;
 310                 /* end COMPAT_146 */
 311                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
 312         }
 313
 314         if (lsd->lsd_feature_incompat & ~cpu_to_le32(MDT_INCOMPAT_SUPP)) {
 315                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
 316                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
 317                        ~MDT_INCOMPAT_SUPP);
 318                 GOTO(err_msd, rc = -EINVAL);
 319         }
 320         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(MDT_ROCOMPAT_SUPP)) {
 321                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
 322                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
 323                        ~MDT_ROCOMPAT_SUPP);
 324                 /* Do something like remount filesystem read-only */
 325                 GOTO(err_msd, rc = -EINVAL);
 326         }
 327
 328         lsd->lsd_feature_compat = cpu_to_le32(OBD_COMPAT_MDT);
 329
 330         mds->mds_last_transno = le64_to_cpu(lsd->lsd_last_transno);
 331
 332         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 333                obd->obd_name, mds->mds_last_transno);
 334         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 335                obd->obd_name, mount_count + 1);
 336         CDEBUG(D_INODE, "%s: server data size: %u\n",
 337                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
 338         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 339                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
 340         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 341                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
 342         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 343                obd->obd_name, last_rcvd_size);
 344         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 345                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
 346                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
 347                 le16_to_cpu(lsd->lsd_client_size));
 348
 349         if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
 350             !lsd->lsd_client_size) {
 351                 CERROR("Bad last_rcvd contents!\n");
 352                 GOTO(err_msd, rc = -EINVAL);
 353         }
 354
 355         /* When we do a clean MDS shutdown, we save the last_transno into
 356          * the header.  If we find clients with higher last_transno values
 357          * then those clients may need recovery done. */
 358         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
 359              off < last_rcvd_size; cl_idx++) {
 360                 __u64 last_transno;
 361                 struct obd_export *exp;
 362                 struct mds_export_data *med;
 363
 364                 if (!mcd) {
 365                         OBD_ALLOC_WAIT(mcd, sizeof(*mcd));
 366                         if (!mcd)
 367                                 GOTO(err_client, rc = -ENOMEM);
 368                 }
 369
 370                 /* Don't assume off is incremented properly by
 371                  * fsfilt_read_record(), in case sizeof(*mcd)
 372                  * isn't the same as lsd->lsd_client_size.  */
 373                 off = le32_to_cpu(lsd->lsd_client_start) +
 374                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
 375                 rc = fsfilt_read_record(obd, file, mcd, sizeof(*mcd), &off);
 376                 if (rc) {
 377                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 378                                LAST_RCVD, cl_idx, off, rc);
 379                         break; /* read error shouldn't cause startup to fail */
 380                 }
 381
 382                 if (mcd->mcd_uuid[0] == '\0') {
 383                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 384                                cl_idx);
 385                         continue;
 386                 }
 387
 388                 last_transno = le64_to_cpu(mcd->mcd_last_transno) >
 389                                le64_to_cpu(mcd->mcd_last_close_transno) ?
 390                                le64_to_cpu(mcd->mcd_last_transno) :
 391                                le64_to_cpu(mcd->mcd_last_close_transno);
 392
 393                 /* These exports are cleaned up by mds_disconnect(), so they
 394                  * need to be set up like real exports as mds_connect() does.
 395                  */
 396                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 397                        " srv lr: "LPU64" lx: "LPU64"\n", mcd->mcd_uuid, cl_idx,
 398                        last_transno, le64_to_cpu(lsd->lsd_last_transno),
 399                        le64_to_cpu(mcd->mcd_last_xid));
 400
 401                 exp = class_new_export(obd, (struct obd_uuid *)mcd->mcd_uuid);
 402                 if (IS_ERR(exp))
 403                         GOTO(err_client, rc = PTR_ERR(exp));
 404
 405                 med = &exp->exp_mds_data;
 406                 med->med_mcd = mcd;
 407                 rc = mds_client_add(obd, exp, cl_idx);
 408                 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
 409
 410
 411                 mcd = NULL;
 412
 413                 spin_lock(&exp->exp_lock);
 414                 exp->exp_replay_needed = 1;
 415                 exp->exp_connecting = 0;
 416                 spin_unlock(&exp->exp_lock);
 417
 418                 obd->obd_recoverable_clients++;
 419                 obd->obd_max_recoverable_clients++;
 420                 class_export_put(exp);
 421
 422                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
 423                        cl_idx, last_transno);
 424
 425                 if (last_transno > mds->mds_last_transno)
 426                         mds->mds_last_transno = last_transno;
 427         }
 428
 429         if (mcd)
 430                 OBD_FREE(mcd, sizeof(*mcd));
 431
 432         obd->obd_last_committed = mds->mds_last_transno;
 433
 434         if (obd->obd_recoverable_clients) {
 435                 CWARN("RECOVERY: service %s, %d recoverable clients, "
 436                       "last_transno "LPU64"\n", obd->obd_name,
 437                       obd->obd_recoverable_clients, mds->mds_last_transno);
 438                 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
 439                 obd->obd_recovering = 1;
 440                 obd->obd_recovery_start = CURRENT_SECONDS;
 441                 /* Only used for lprocfs_status */
 442                 obd->obd_recovery_end = obd->obd_recovery_start +
 443                         OBD_RECOVERY_TIMEOUT;
 444         }
 445
 446         mds->mds_mount_count = mount_count + 1;
 447         lsd->lsd_mount_count = lsd->lsd_compat14 =
 448                 cpu_to_le64(mds->mds_mount_count);
 449
 450         /* save it, so mount count and last_transno is current */
 451         rc = mds_update_server_data(obd, 1);
 452         if (rc)
 453                 GOTO(err_client, rc);
 454
 455         RETURN(0);
 456
 457 err_client:
 458         class_disconnect_exports(obd);
 459 err_msd:
 460         mds_server_free_data(mds);
 461         RETURN(rc);
 462 }
 463
 464 int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
 465 {
 466         struct mds_obd *mds = &obd->u.mds;
 467         struct lvfs_run_ctxt saved;
 468         struct dentry *dentry;
 469         struct file *file;
 470         int rc;
 471         ENTRY;
 472
 473         OBD_FAIL_RETURN(OBD_FAIL_MDS_FS_SETUP, -ENOENT);
 474
 475         rc = cleanup_group_info();
 476         if (rc)
 477                 RETURN(rc);
 478
 479         mds->mds_vfsmnt = mnt;
 480         /* why not mnt->mnt_sb instead of mnt->mnt_root->d_inode->i_sb? */
 481         obd->u.obt.obt_sb = mnt->mnt_root->d_inode->i_sb;
 482
 483         fsfilt_setup(obd, obd->u.obt.obt_sb);
 484
 485         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 486         obd->obd_lvfs_ctxt.pwdmnt = mnt;
 487         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
 488         obd->obd_lvfs_ctxt.fs = get_ds();
 489         obd->obd_lvfs_ctxt.cb_ops = mds_lvfs_ops;
 490
 491         /* setup the directory tree */
 492         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 493         dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755, 0);
 494         if (IS_ERR(dentry)) {
 495                 rc = PTR_ERR(dentry);
 496                 CERROR("cannot create ROOT directory: rc = %d\n", rc);
 497                 GOTO(err_pop, rc);
 498         }
 499
 500         mds->mds_rootfid.id = dentry->d_inode->i_ino;
 501         mds->mds_rootfid.generation = dentry->d_inode->i_generation;
 502         mds->mds_rootfid.f_type = S_IFDIR;
 503
 504         dput(dentry);
 505
 506         dentry = lookup_one_len("__iopen__", current->fs->pwd,
 507                                 strlen("__iopen__"));
 508         if (IS_ERR(dentry)) {
 509                 rc = PTR_ERR(dentry);
 510                 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
 511                 GOTO(err_pop, rc);
 512         }
 513
 514         mds->mds_fid_de = dentry;
 515         if (!dentry->d_inode || is_bad_inode(dentry->d_inode)) {
 516                 rc = -ENOENT;
 517                 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
 518                 GOTO(err_fid, rc);
 519         }
 520
 521         dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777, 1);
 522         if (IS_ERR(dentry)) {
 523                 rc = PTR_ERR(dentry);
 524                 CERROR("cannot create PENDING directory: rc = %d\n", rc);
 525                 GOTO(err_fid, rc);
 526         }
 527         mds->mds_pending_dir = dentry;
 528
 529         /* COMPAT_146 */
 530         dentry = simple_mkdir(current->fs->pwd, MDT_LOGS_DIR, 0777, 1);
 531         if (IS_ERR(dentry)) {
 532                 rc = PTR_ERR(dentry);
 533                 CERROR("cannot create %s directory: rc = %d\n",
 534                        MDT_LOGS_DIR, rc);
 535                 GOTO(err_pending, rc);
 536         }
 537         mds->mds_logs_dir = dentry;
 538         /* end COMPAT_146 */
 539
 540         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
 541         if (IS_ERR(dentry)) {
 542                 rc = PTR_ERR(dentry);
 543                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
 544                 GOTO(err_logs, rc);
 545         }
 546         mds->mds_objects_dir = dentry;
 547
 548         /* open and test the last rcvd file */
 549         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
 550         if (IS_ERR(file)) {
 551                 rc = PTR_ERR(file);
 552                 CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
 553                 GOTO(err_objects, rc = PTR_ERR(file));
 554         }
 555         mds->mds_rcvd_filp = file;
 556         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 557                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 558                        file->f_dentry->d_inode->i_mode);
 559                 GOTO(err_last_rcvd, rc = -ENOENT);
 560         }
 561
 562         rc = mds_init_server_data(obd, file);
 563         if (rc) {
 564                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
 565                 GOTO(err_last_rcvd, rc);
 566         }
 567
 568         /* open and test the lov objd file */
 569         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
 570         if (IS_ERR(file)) {
 571                 rc = PTR_ERR(file);
 572                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
 573                 GOTO(err_client, rc = PTR_ERR(file));
 574         }
 575         mds->mds_lov_objid_filp = file;
 576         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 577                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
 578                        file->f_dentry->d_inode->i_mode);
 579                 GOTO(err_lov_objid, rc = -ENOENT);
 580         }
 581
 582         /* open and test the check io file junk */
 583         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
 584         if (IS_ERR(file)) {
 585                 rc = PTR_ERR(file);
 586                 CERROR("cannot open/create %s file: rc = %d\n", HEALTH_CHECK, rc);
 587                 GOTO(err_lov_objid, rc = PTR_ERR(file));
 588         }
 589         mds->mds_health_check_filp = file;
 590         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 591                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
 592                        file->f_dentry->d_inode->i_mode);
 593                 GOTO(err_health_check, rc = -ENOENT);
 594         }
 595         rc = lvfs_check_io_health(obd, file);
 596         if (rc)
 597                 GOTO(err_health_check, rc);
 598 err_pop:
 599         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 600
 601         return rc;
 602
 603 err_health_check:
 604         if (mds->mds_health_check_filp &&
 605             filp_close(mds->mds_health_check_filp, 0))
 606                 CERROR("can't close %s after error\n", HEALTH_CHECK);
 607 err_lov_objid:
 608         if (mds->mds_lov_objid_filp && filp_close(mds->mds_lov_objid_filp, 0))
 609                 CERROR("can't close %s after error\n", LOV_OBJID);
 610 err_client:
 611         class_disconnect_exports(obd);
 612 err_last_rcvd:
 613         if (mds->mds_rcvd_filp && filp_close(mds->mds_rcvd_filp, 0))
 614                 CERROR("can't close %s after error\n", LAST_RCVD);
 615 err_objects:
 616         dput(mds->mds_objects_dir);
 617 err_logs:
 618         dput(mds->mds_logs_dir);
 619 err_pending:
 620         dput(mds->mds_pending_dir);
 621 err_fid:
 622         dput(mds->mds_fid_de);
 623         goto err_pop;
 624 }
 625
 626
 627 int mds_fs_cleanup(struct obd_device *obd)
 628 {
 629         struct mds_obd *mds = &obd->u.mds;
 630         struct lvfs_run_ctxt saved;
 631         int rc = 0;
 632
 633         if (obd->obd_fail)
 634                 LCONSOLE_WARN("%s: shutting down for failover; client state "
 635                               "will be preserved.\n", obd->obd_name);
 636
 637         class_disconnect_exports(obd); /* cleans up client info too */
 638         mds_server_free_data(mds);
 639
 640         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 641         if (mds->mds_rcvd_filp) {
 642                 rc = filp_close(mds->mds_rcvd_filp, 0);
 643                 mds->mds_rcvd_filp = NULL;
 644                 if (rc)
 645                         CERROR("%s file won't close, rc=%d\n", LAST_RCVD, rc);
 646         }
 647         if (mds->mds_lov_objid_filp) {
 648                 rc = filp_close(mds->mds_lov_objid_filp, 0);
 649                 mds->mds_lov_objid_filp = NULL;
 650                 if (rc)
 651                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
 652         }
 653         if (mds->mds_health_check_filp) {
 654                 rc = filp_close(mds->mds_health_check_filp, 0);
 655                 mds->mds_health_check_filp = NULL;
 656                 if (rc)
 657                         CERROR("%s file won't close, rc=%d\n", HEALTH_CHECK, rc);
 658         }
 659         if (mds->mds_objects_dir != NULL) {
 660                 l_dput(mds->mds_objects_dir);
 661                 mds->mds_objects_dir = NULL;
 662         }
 663         if (mds->mds_logs_dir) {
 664                 l_dput(mds->mds_logs_dir);
 665                 mds->mds_logs_dir = NULL;
 666         }
 667         if (mds->mds_pending_dir) {
 668                 l_dput(mds->mds_pending_dir);
 669                 mds->mds_pending_dir = NULL;
 670         }
 671
 672         lquota_fs_cleanup(mds_quota_interface_ref, obd);
 673
 674         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 675         shrink_dcache_parent(mds->mds_fid_de);
 676         dput(mds->mds_fid_de);
 677         LL_DQUOT_OFF(obd->u.obt.obt_sb);
 678
 679         return rc;
 680 }
 681
 682 /* Creates an object with the same name as its fid.  Because this is not at all
 683  * performance sensitive, it is accomplished by creating a file, checking the
 684  * fid, and renaming it. */
 685 int mds_obd_create(struct obd_export *exp, struct obdo *oa,
 686                    struct lov_stripe_md **ea, struct obd_trans_info *oti)
 687 {
 688         struct mds_obd *mds = &exp->exp_obd->u.mds;
 689         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 690         unsigned int tmpname = ll_rand();
 691         struct file *filp;
 692         struct dentry *new_child;
 693         struct lvfs_run_ctxt saved;
 694         char fidname[LL_FID_NAMELEN];
 695         void *handle;
 696         struct lvfs_ucred ucred = { 0 };
 697         int rc = 0, err, namelen;
 698         ENTRY;
 699
 700         /* the owner of object file should always be root */
 701         ucred.luc_cap = current->cap_effective | CAP_SYS_RESOURCE;
 702
 703         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, &ucred);
 704
 705         sprintf(fidname, "OBJECTS/%u.%u", tmpname, current->pid);
 706         filp = filp_open(fidname, O_CREAT | O_EXCL, 0666);
 707         if (IS_ERR(filp)) {
 708                 rc = PTR_ERR(filp);
 709                 if (rc == -EEXIST) {
 710                         CERROR("impossible object name collision %u\n",
 711                                tmpname);
 712                         LBUG();
 713                 }
 714                 CERROR("error creating tmp object %u: rc %d\n", tmpname, rc);
 715                 GOTO(out_pop, rc);
 716         }
 717
 718         LASSERT(mds->mds_objects_dir == filp->f_dentry->d_parent);
 719
 720         oa->o_id = filp->f_dentry->d_inode->i_ino;
 721         oa->o_generation = filp->f_dentry->d_inode->i_generation;
 722         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 723
 724         LOCK_INODE_MUTEX(parent_inode);
 725         new_child = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 726
 727         if (IS_ERR(new_child)) {
 728                 CERROR("getting neg dentry for obj rename: %d\n", rc);
 729                 GOTO(out_close, rc = PTR_ERR(new_child));
 730         }
 731         if (new_child->d_inode != NULL) {
 732                 CERROR("impossible non-negative obj dentry " LPU64":%u!\n",
 733                        oa->o_id, oa->o_generation);
 734                 LBUG();
 735         }
 736
 737         handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
 738                               FSFILT_OP_RENAME, NULL);
 739         if (IS_ERR(handle))
 740                 GOTO(out_dput, rc = PTR_ERR(handle));
 741
 742         lock_kernel();
 743         rc = vfs_rename(mds->mds_objects_dir->d_inode, filp->f_dentry,
 744                         mds->mds_objects_dir->d_inode, new_child);
 745         unlock_kernel();
 746         if (rc)
 747                 CERROR("error renaming new object "LPU64":%u: rc %d\n",
 748                        oa->o_id, oa->o_generation, rc);
 749
 750         err = fsfilt_commit(exp->exp_obd, mds->mds_objects_dir->d_inode,
 751                             handle, 0);
 752         if (!err)
 753                 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
 754         else if (!rc)
 755                 rc = err;
 756 out_dput:
 757         dput(new_child);
 758 out_close:
 759         UNLOCK_INODE_MUTEX(parent_inode);
 760         err = filp_close(filp, 0);
 761         if (err) {
 762                 CERROR("closing tmpfile %u: rc %d\n", tmpname, rc);
 763                 if (!rc)
 764                         rc = err;
 765         }
 766 out_pop:
 767         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, &ucred);
 768         RETURN(rc);
 769 }
 770
 771 int mds_obd_destroy(struct obd_export *exp, struct obdo *oa,
 772                     struct lov_stripe_md *ea, struct obd_trans_info *oti,
 773                     struct obd_export *md_exp)
 774 {
 775         struct mds_obd *mds = &exp->exp_obd->u.mds;
 776         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 777         struct obd_device *obd = exp->exp_obd;
 778         struct lvfs_run_ctxt saved;
 779         struct lvfs_ucred ucred = { 0 };
 780         char fidname[LL_FID_NAMELEN];
 781         struct inode *inode = NULL;
 782         struct dentry *de;
 783         void *handle;
 784         int err, namelen, rc = 0;
 785         ENTRY;
 786
 787         ucred.luc_cap = current->cap_effective | CAP_SYS_RESOURCE;
 788         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &ucred);
 789
 790         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 791
 792         LOCK_INODE_MUTEX(parent_inode);
 793         de = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 794         if (IS_ERR(de)) {
 795                 rc = IS_ERR(de);
 796                 de = NULL;
 797                 CERROR("error looking up object "LPU64" %s: rc %d\n",
 798                        oa->o_id, fidname, rc);
 799                 GOTO(out_dput, rc);
 800         }
 801         if (de->d_inode == NULL) {
 802                 CERROR("destroying non-existent object "LPU64" %s: rc %d\n",
 803                        oa->o_id, fidname, rc);
 804                 GOTO(out_dput, rc = -ENOENT);
 805         }
 806
 807         /* Stripe count is 1 here since this is some MDS specific stuff
 808            that is unlinked, not spanned across multiple OSTs */
 809         handle = fsfilt_start_log(obd, mds->mds_objects_dir->d_inode,
 810                                   FSFILT_OP_UNLINK, oti, 1);
 811
 812         if (IS_ERR(handle))
 813                 GOTO(out_dput, rc = PTR_ERR(handle));
 814
 815         /* take a reference to protect inode from truncation within
 816            vfs_unlink() context. bug 10409 */
 817         inode = de->d_inode;
 818         atomic_inc(&inode->i_count);
 819         rc = vfs_unlink(mds->mds_objects_dir->d_inode, de);
 820         if (rc)
 821                 CERROR("error destroying object "LPU64":%u: rc %d\n",
 822                        oa->o_id, oa->o_generation, rc);
 823
 824         err = fsfilt_commit(obd, mds->mds_objects_dir->d_inode, handle, 0);
 825         if (err && !rc)
 826                 rc = err;
 827 out_dput:
 828         if (de != NULL)
 829                 l_dput(de);
 830         UNLOCK_INODE_MUTEX(parent_inode);
 831
 832         if (inode)
 833                 iput(inode);
 834
 835         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &ucred);
 836         RETURN(rc);
 837 }