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