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 #include <linux/mount.h>
  40 #include <lustre_mds.h>
  41 #include <obd_class.h>
  42 #include <obd_support.h>
  43 #include <lustre_lib.h>
  44 #include <lustre_fsfilt.h>
  45 #include <lustre_disk.h>
  46 #include <libcfs/list.h>
  47
  48 #include "mds_internal.h"
  49
  50
  51 static int mds_export_stats_init(struct obd_device *obd,
  52                                  struct obd_export *exp,
  53                                  void *localdata)
  54 {
  55         int rc, num_stats, newnid;
  56
  57         rc = lprocfs_exp_setup(exp, localdata, &newnid);
  58         if (rc)
  59                 return rc;
  60
  61         if (newnid) {
  62                 struct nid_stat *tmp = exp->exp_nid_stats;
  63                 LASSERT(tmp != NULL);
  64                 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
  65                              LPROC_MDS_LAST - 1;
  66                 tmp->nid_stats = lprocfs_alloc_stats(num_stats,
  67                                                      LPROCFS_STATS_FLAG_NOPERCPU);
  68                 if (tmp->nid_stats == NULL)
  69                         return -ENOMEM;
  70
  71                 lprocfs_init_ops_stats(LPROC_MDS_LAST, tmp->nid_stats);
  72                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
  73                                             tmp->nid_stats);
  74                 if (rc)
  75                         return rc;
  76                 mds_stats_counter_init(tmp->nid_stats);
  77         }
  78         return 0;
  79 }
  80
  81 /* Add client data to the MDS.  We use a bitmap to locate a free space
  82  * in the last_rcvd file if cl_off is -1 (i.e. a new client).
  83  * Otherwise, we have just read the data from the last_rcvd file and
  84  * we know its offset.
  85  *
  86  * It should not be possible to fail adding an existing client - otherwise
  87  * mds_init_server_data() callsite needs to be fixed.
  88  */
  89 int mds_client_add(struct obd_device *obd, struct obd_export *exp,
  90                    int cl_idx, void *localdata)
  91 {
  92         struct mds_obd *mds = &obd->u.mds;
  93         struct mds_export_data *med = &exp->exp_mds_data;
  94         unsigned long *bitmap = mds->mds_client_bitmap;
  95         int new_client = (cl_idx == -1);
  96         ENTRY;
  97
  98         LASSERT(bitmap != NULL);
  99         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
 100
 101         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 102         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
 103                 RETURN(0);
 104
 105         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
 106          * there's no need for extra complication here
 107          */
 108         if (new_client) {
 109                 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
 110         repeat:
 111                 if (cl_idx >= LR_MAX_CLIENTS ||
 112                     OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
 113                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
 114                                cl_idx);
 115                         return -EOVERFLOW;
 116                 }
 117                 if (test_and_set_bit(cl_idx, bitmap)) {
 118                         cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
 119                                                     cl_idx);
 120                         goto repeat;
 121                 }
 122         } else {
 123                 if (test_and_set_bit(cl_idx, bitmap)) {
 124                         CERROR("MDS client %d: bit already set in bitmap!!\n",
 125                                cl_idx);
 126                         LBUG();
 127                 }
 128         }
 129
 130         CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
 131                cl_idx, med->med_mcd->mcd_uuid);
 132
 133         med->med_lr_idx = cl_idx;
 134         med->med_lr_off = le32_to_cpu(mds->mds_server_data->lsd_client_start) +
 135                 (cl_idx * le16_to_cpu(mds->mds_server_data->lsd_client_size));
 136         LASSERTF(med->med_lr_off > 0, "med_lr_off = %llu\n", med->med_lr_off);
 137         mds_export_stats_init(obd, exp, localdata);
 138
 139         if (new_client) {
 140                 struct lvfs_run_ctxt saved;
 141                 loff_t off = med->med_lr_off;
 142                 struct file *file = mds->mds_rcvd_filp;
 143                 void *handle;
 144                 int rc;
 145
 146                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 147                 handle = fsfilt_start(obd, file->f_dentry->d_inode,
 148                                       FSFILT_OP_SETATTR, NULL);
 149                 if (IS_ERR(handle)) {
 150                         rc = PTR_ERR(handle);
 151                         CERROR("unable to start transaction: rc %d\n", rc);
 152                 } else {
 153                         rc = fsfilt_add_journal_cb(obd, 0, handle,
 154                                                    target_client_add_cb, exp);
 155                         if (rc == 0) {
 156                                 spin_lock(&exp->exp_lock);
 157                                 exp->exp_need_sync = 1;
 158                                 spin_unlock(&exp->exp_lock);
 159                         }
 160                         rc = fsfilt_write_record(obd, file, med->med_mcd,
 161                                                  sizeof(*med->med_mcd),
 162                                                  &off, rc /* sync if no cb */);
 163                         fsfilt_commit(obd, file->f_dentry->d_inode, handle, 0);
 164                 }
 165
 166                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 167
 168                 if (rc)
 169                         return rc;
 170                 CDEBUG(D_INFO, "wrote client mcd at idx %u off %llu (len %u)\n",
 171                        med->med_lr_idx, med->med_lr_off,
 172                        (unsigned int)sizeof(*med->med_mcd));
 173         }
 174         return 0;
 175 }
 176
 177 int mds_client_free(struct obd_export *exp)
 178 {
 179         struct mds_export_data *med = &exp->exp_mds_data;
 180         struct mds_obd *mds = &exp->exp_obd->u.mds;
 181         struct obd_device *obd = exp->exp_obd;
 182         struct mds_client_data zero_mcd;
 183         struct lvfs_run_ctxt saved;
 184         int rc;
 185         loff_t off;
 186         ENTRY;
 187
 188         if (!med->med_mcd)
 189                 RETURN(0);
 190
 191         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 192         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
 193                 GOTO(free, 0);
 194
 195         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
 196                med->med_lr_idx, med->med_lr_off, med->med_mcd->mcd_uuid);
 197
 198         LASSERT(mds->mds_client_bitmap != NULL);
 199
 200         lprocfs_exp_cleanup(exp);
 201
 202         off = med->med_lr_off;
 203
 204         /* Don't clear med_lr_idx here as it is likely also unset.  At worst
 205          * we leak a client slot that will be cleaned on the next recovery. */
 206         if (off <= 0) {
 207                 CERROR("%s: client idx %d has offset %lld\n",
 208                         obd->obd_name, med->med_lr_idx, off);
 209                 GOTO(free, rc = -EINVAL);
 210         }
 211
 212         /* Clear the bit _after_ zeroing out the client so we don't
 213            race with mds_client_add and zero out new clients.*/
 214         if (!test_bit(med->med_lr_idx, mds->mds_client_bitmap)) {
 215                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 216                        med->med_lr_idx);
 217                 LBUG();
 218         }
 219
 220         if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
 221                 memset(&zero_mcd, 0, sizeof zero_mcd);
 222                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 223                 rc = fsfilt_write_record(obd, mds->mds_rcvd_filp, &zero_mcd,
 224                                          sizeof(zero_mcd), &off,
 225                                          (!exp->exp_libclient ||
 226                                           exp->exp_need_sync));
 227                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 228
 229                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 230                        "zeroing out client %s idx %u in %s rc %d\n",
 231                        med->med_mcd->mcd_uuid, med->med_lr_idx, LAST_RCVD, rc);
 232         }
 233
 234         if (!test_and_clear_bit(med->med_lr_idx, mds->mds_client_bitmap)) {
 235                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 236                        med->med_lr_idx);
 237                 LBUG();
 238         }
 239
 240
 241         /* Make sure the server's last_transno is up to date. Do this
 242          * after the client is freed so we know all the client's
 243          * transactions have been committed. */
 244         mds_update_server_data(exp->exp_obd, 0);
 245
 246         EXIT;
 247  free:
 248         OBD_FREE(med->med_mcd, sizeof(*med->med_mcd));
 249         med->med_mcd = NULL;
 250
 251         return 0;
 252 }
 253
 254 static int mds_server_free_data(struct mds_obd *mds)
 255 {
 256         OBD_FREE(mds->mds_client_bitmap, LR_MAX_CLIENTS / 8);
 257         OBD_FREE(mds->mds_server_data, sizeof(*mds->mds_server_data));
 258         mds->mds_server_data = NULL;
 259
 260         return 0;
 261 }
 262
 263 static int mds_init_server_data(struct obd_device *obd, struct file *file)
 264 {
 265         struct mds_obd *mds = &obd->u.mds;
 266         struct lr_server_data *lsd;
 267         struct mds_client_data *mcd = NULL;
 268         loff_t off = 0;
 269         unsigned long last_rcvd_size = i_size_read(file->f_dentry->d_inode);
 270         __u64 mount_count;
 271         int cl_idx, rc = 0;
 272         ENTRY;
 273
 274         /* ensure padding in the struct is the correct size */
 275         LASSERT(offsetof(struct lr_server_data, lsd_padding) +
 276                 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
 277         LASSERT(offsetof(struct mds_client_data, mcd_padding) +
 278                 sizeof(mcd->mcd_padding) == LR_CLIENT_SIZE);
 279
 280         OBD_ALLOC_WAIT(lsd, sizeof(*lsd));
 281         if (!lsd)
 282                 RETURN(-ENOMEM);
 283
 284         OBD_ALLOC_WAIT(mds->mds_client_bitmap, LR_MAX_CLIENTS / 8);
 285         if (!mds->mds_client_bitmap) {
 286                 OBD_FREE(lsd, sizeof(*lsd));
 287                 RETURN(-ENOMEM);
 288         }
 289
 290         mds->mds_server_data = lsd;
 291
 292         if (last_rcvd_size == 0) {
 293                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
 294
 295                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
 296                 lsd->lsd_last_transno = 0;
 297                 mount_count = lsd->lsd_mount_count = 0;
 298                 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
 299                 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
 300                 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
 301                 lsd->lsd_feature_rocompat = cpu_to_le32(OBD_ROCOMPAT_LOVOBJID);
 302                 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_MDT);
 303         } else {
 304                 rc = fsfilt_read_record(obd, file, lsd, sizeof(*lsd), &off);
 305                 if (rc) {
 306                         CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
 307                         GOTO(err_msd, rc);
 308                 }
 309                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
 310                         LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s "
 311                                            "using the wrong disk %s. Were the "
 312                                            "/dev/ assignments rearranged?\n",
 313                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
 314                         GOTO(err_msd, rc = -EINVAL);
 315                 }
 316                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
 317         }
 318
 319         if (lsd->lsd_feature_incompat & ~cpu_to_le32(MDT_INCOMPAT_SUPP)) {
 320                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
 321                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
 322                        ~MDT_INCOMPAT_SUPP);
 323                 GOTO(err_msd, rc = -EINVAL);
 324         }
 325         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(MDT_ROCOMPAT_SUPP)) {
 326                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
 327                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
 328                        ~MDT_ROCOMPAT_SUPP);
 329                 /* Do something like remount filesystem read-only */
 330                 GOTO(err_msd, rc = -EINVAL);
 331         }
 332
 333         lsd->lsd_feature_compat = cpu_to_le32(OBD_COMPAT_MDT);
 334
 335         mds->mds_last_transno = le64_to_cpu(lsd->lsd_last_transno);
 336
 337         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 338                obd->obd_name, mds->mds_last_transno);
 339         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 340                obd->obd_name, mount_count + 1);
 341         CDEBUG(D_INODE, "%s: server data size: %u\n",
 342                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
 343         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 344                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
 345         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 346                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
 347         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 348                obd->obd_name, last_rcvd_size);
 349         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 350                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
 351                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
 352                 le16_to_cpu(lsd->lsd_client_size));
 353
 354         if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
 355             !lsd->lsd_client_size) {
 356                 CERROR("Bad last_rcvd contents!\n");
 357                 GOTO(err_msd, rc = -EINVAL);
 358         }
 359
 360         /* When we do a clean MDS shutdown, we save the last_transno into
 361          * the header.  If we find clients with higher last_transno values
 362          * then those clients may need recovery done. */
 363         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
 364              off < last_rcvd_size; cl_idx++) {
 365                 __u64 last_transno;
 366                 struct obd_export *exp;
 367                 struct mds_export_data *med;
 368
 369                 if (!mcd) {
 370                         OBD_ALLOC_WAIT(mcd, sizeof(*mcd));
 371                         if (!mcd)
 372                                 GOTO(err_client, rc = -ENOMEM);
 373                 }
 374
 375                 /* Don't assume off is incremented properly by
 376                  * fsfilt_read_record(), in case sizeof(*mcd)
 377                  * isn't the same as lsd->lsd_client_size.  */
 378                 off = le32_to_cpu(lsd->lsd_client_start) +
 379                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
 380                 rc = fsfilt_read_record(obd, file, mcd, sizeof(*mcd), &off);
 381                 if (rc) {
 382                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 383                                LAST_RCVD, cl_idx, off, rc);
 384                         break; /* read error shouldn't cause startup to fail */
 385                 }
 386
 387                 if (mcd->mcd_uuid[0] == '\0') {
 388                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 389                                cl_idx);
 390                         continue;
 391                 }
 392
 393                 last_transno = le64_to_cpu(mcd->mcd_last_transno) >
 394                                le64_to_cpu(mcd->mcd_last_close_transno) ?
 395                                le64_to_cpu(mcd->mcd_last_transno) :
 396                                le64_to_cpu(mcd->mcd_last_close_transno);
 397
 398                 /* These exports are cleaned up by mds_disconnect(), so they
 399                  * need to be set up like real exports as mds_connect() does.
 400                  */
 401                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 402                        " srv lr: "LPU64" lx: "LPU64"\n", mcd->mcd_uuid, cl_idx,
 403                        last_transno, le64_to_cpu(lsd->lsd_last_transno),
 404                        le64_to_cpu(mcd->mcd_last_xid));
 405
 406                 exp = class_new_export(obd, (struct obd_uuid *)mcd->mcd_uuid);
 407                 if (IS_ERR(exp)) {
 408                         if (PTR_ERR(exp) == -EALREADY) {
 409                                 /* export already exists, zero out this one */
 410                                 mcd->mcd_uuid[0] = '\0';
 411                         } else {
 412                                 GOTO(err_client, rc = PTR_ERR(exp));
 413                         }
 414                 } else {
 415                         med = &exp->exp_mds_data;
 416                         med->med_mcd = mcd;
 417                         rc = mds_client_add(obd, exp, cl_idx, NULL);
 418                         LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
 419
 420                         mcd = NULL;
 421                         spin_lock(&exp->exp_lock);
 422                         exp->exp_req_replay_needed = 1;
 423                         exp->exp_connecting = 0;
 424                         spin_unlock(&exp->exp_lock);
 425                         obd->obd_max_recoverable_clients++;
 426                         class_export_put(exp);
 427                 }
 428                 /* Need to check last_rcvd even for duplicated exports. */
 429                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
 430                        cl_idx, last_transno);
 431
 432                 if (last_transno > mds->mds_last_transno)
 433                         mds->mds_last_transno = last_transno;
 434         }
 435
 436         if (mcd)
 437                 OBD_FREE(mcd, sizeof(*mcd));
 438
 439         obd->obd_last_committed = mds->mds_last_transno;
 440
 441         if (obd->obd_max_recoverable_clients) {
 442                 /* shouldn't happen in b_new_cmd */
 443                 LBUG();
 444                 CWARN("RECOVERY: service %s, %d recoverable clients, "
 445                       "last_transno "LPU64"\n", obd->obd_name,
 446                       obd->obd_max_recoverable_clients, mds->mds_last_transno);
 447                 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
 448                 obd->obd_recovering = 1;
 449                 obd->obd_recovery_start = CURRENT_SECONDS;
 450                 /* Only used for lprocfs_status */
 451                 obd->obd_recovery_end = obd->obd_recovery_start +
 452                         OBD_RECOVERY_TIMEOUT;
 453         }
 454
 455         mds->mds_mount_count = mount_count + 1;
 456         lsd->lsd_mount_count = lsd->lsd_compat14 =
 457                 cpu_to_le64(mds->mds_mount_count);
 458
 459         /* save it, so mount count and last_transno is current */
 460         rc = mds_update_server_data(obd, 1);
 461         if (rc)
 462                 GOTO(err_client, rc);
 463
 464         RETURN(0);
 465
 466 err_client:
 467         class_disconnect_exports(obd);
 468 err_msd:
 469         mds_server_free_data(mds);
 470         RETURN(rc);
 471 }
 472
 473 void mds_init_ctxt(struct obd_device *obd, struct vfsmount *mnt)
 474 {
 475         struct mds_obd *mds = &obd->u.mds;
 476
 477         mds->mds_vfsmnt = mnt;
 478         /* why not mnt->mnt_sb instead of mnt->mnt_root->d_inode->i_sb? */
 479         obd->u.obt.obt_sb = mnt->mnt_root->d_inode->i_sb;
 480
 481         fsfilt_setup(obd, obd->u.obt.obt_sb);
 482
 483         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 484         obd->obd_lvfs_ctxt.pwdmnt = mnt;
 485         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
 486         obd->obd_lvfs_ctxt.fs = get_ds();
 487         obd->obd_lvfs_ctxt.cb_ops = mds_lvfs_ops;
 488         return;
 489 }
 490
 491 int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
 492 {
 493         struct mds_obd *mds = &obd->u.mds;
 494         struct lvfs_run_ctxt saved;
 495         struct dentry *dentry;
 496         struct file *file;
 497         int rc;
 498         ENTRY;
 499
 500         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
 501                 RETURN(-ENOENT);
 502
 503         rc = cleanup_group_info();
 504         if (rc)
 505                 RETURN(rc);
 506
 507         mds_init_ctxt(obd, mnt);
 508
 509         /* setup the directory tree */
 510         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 511         dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755, 0);
 512         if (IS_ERR(dentry)) {
 513                 rc = PTR_ERR(dentry);
 514                 CERROR("cannot create ROOT directory: rc = %d\n", rc);
 515                 GOTO(err_pop, rc);
 516         }
 517
 518         mds->mds_rootfid.id = dentry->d_inode->i_ino;
 519         mds->mds_rootfid.generation = dentry->d_inode->i_generation;
 520         mds->mds_rootfid.f_type = S_IFDIR;
 521
 522         dput(dentry);
 523
 524         dentry = lookup_one_len("__iopen__", current->fs->pwd,
 525                                 strlen("__iopen__"));
 526         if (IS_ERR(dentry)) {
 527                 rc = PTR_ERR(dentry);
 528                 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
 529                 GOTO(err_pop, rc);
 530         }
 531
 532         mds->mds_fid_de = dentry;
 533         if (!dentry->d_inode || is_bad_inode(dentry->d_inode)) {
 534                 rc = -ENOENT;
 535                 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
 536                 GOTO(err_fid, rc);
 537         }
 538
 539         dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777, 1);
 540         if (IS_ERR(dentry)) {
 541                 rc = PTR_ERR(dentry);
 542                 CERROR("cannot create PENDING directory: rc = %d\n", rc);
 543                 GOTO(err_fid, rc);
 544         }
 545         mds->mds_pending_dir = dentry;
 546
 547         /* COMPAT_146 */
 548         dentry = simple_mkdir(current->fs->pwd, MDT_LOGS_DIR, 0777, 1);
 549         if (IS_ERR(dentry)) {
 550                 rc = PTR_ERR(dentry);
 551                 CERROR("cannot create %s directory: rc = %d\n",
 552                        MDT_LOGS_DIR, rc);
 553                 GOTO(err_pending, rc);
 554         }
 555         mds->mds_logs_dir = dentry;
 556         /* end COMPAT_146 */
 557
 558         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
 559         if (IS_ERR(dentry)) {
 560                 rc = PTR_ERR(dentry);
 561                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
 562                 GOTO(err_logs, rc);
 563         }
 564         mds->mds_objects_dir = dentry;
 565
 566         /* open and test the last rcvd file */
 567         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
 568         if (IS_ERR(file)) {
 569                 rc = PTR_ERR(file);
 570                 CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
 571                 GOTO(err_objects, rc = PTR_ERR(file));
 572         }
 573         mds->mds_rcvd_filp = file;
 574         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 575                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 576                        file->f_dentry->d_inode->i_mode);
 577                 GOTO(err_last_rcvd, rc = -ENOENT);
 578         }
 579
 580         rc = mds_init_server_data(obd, file);
 581         if (rc) {
 582                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
 583                 GOTO(err_last_rcvd, rc);
 584         }
 585
 586         /* open and test the lov objd file */
 587         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
 588         if (IS_ERR(file)) {
 589                 rc = PTR_ERR(file);
 590                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
 591                 GOTO(err_client, rc = PTR_ERR(file));
 592         }
 593         mds->mds_lov_objid_filp = file;
 594         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 595                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
 596                        file->f_dentry->d_inode->i_mode);
 597                 GOTO(err_lov_objid, rc = -ENOENT);
 598         }
 599
 600         /* open and test the check io file junk */
 601         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
 602         if (IS_ERR(file)) {
 603                 rc = PTR_ERR(file);
 604                 CERROR("cannot open/create %s file: rc = %d\n", HEALTH_CHECK,
 605                        rc);
 606                 GOTO(err_lov_objid, rc = PTR_ERR(file));
 607         }
 608         mds->mds_health_check_filp = file;
 609         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 610                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
 611                        file->f_dentry->d_inode->i_mode);
 612                 GOTO(err_health_check, rc = -ENOENT);
 613         }
 614         rc = lvfs_check_io_health(obd, file);
 615         if (rc)
 616                 GOTO(err_health_check, rc);
 617 err_pop:
 618         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 619
 620         return rc;
 621
 622 err_health_check:
 623         if (mds->mds_health_check_filp &&
 624             filp_close(mds->mds_health_check_filp, 0))
 625                 CERROR("can't close %s after error\n", HEALTH_CHECK);
 626 err_lov_objid:
 627         if (mds->mds_lov_objid_filp &&
 628                 filp_close((struct file *)mds->mds_lov_objid_filp, 0))
 629                 CERROR("can't close %s after error\n", LOV_OBJID);
 630 err_client:
 631         class_disconnect_exports(obd);
 632 err_last_rcvd:
 633         if (mds->mds_rcvd_filp && filp_close(mds->mds_rcvd_filp, 0))
 634                 CERROR("can't close %s after error\n", LAST_RCVD);
 635 err_objects:
 636         dput(mds->mds_objects_dir);
 637 err_logs:
 638         dput(mds->mds_logs_dir);
 639 err_pending:
 640         dput(mds->mds_pending_dir);
 641 err_fid:
 642         dput(mds->mds_fid_de);
 643         goto err_pop;
 644 }
 645
 646 int mds_fs_cleanup(struct obd_device *obd)
 647 {
 648         struct mds_obd *mds = &obd->u.mds;
 649         struct lvfs_run_ctxt saved;
 650         int rc = 0;
 651
 652         if (obd->obd_fail)
 653                 LCONSOLE_WARN("%s: shutting down for failover; client state "
 654                               "will be preserved.\n", obd->obd_name);
 655
 656         class_disconnect_exports(obd); /* cleans up client info too */
 657         mds_server_free_data(mds);
 658
 659         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 660         if (mds->mds_rcvd_filp) {
 661                 rc = filp_close(mds->mds_rcvd_filp, 0);
 662                 mds->mds_rcvd_filp = NULL;
 663                 if (rc)
 664                         CERROR("%s file won't close, rc=%d\n", LAST_RCVD, rc);
 665         }
 666         if (mds->mds_lov_objid_filp) {
 667                 rc = filp_close((struct file *)mds->mds_lov_objid_filp, 0);
 668                 mds->mds_lov_objid_filp = NULL;
 669                 if (rc)
 670                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
 671         }
 672         if (mds->mds_health_check_filp) {
 673                 rc = filp_close(mds->mds_health_check_filp, 0);
 674                 mds->mds_health_check_filp = NULL;
 675                 if (rc)
 676                         CERROR("%s file won't close, rc=%d\n", HEALTH_CHECK,
 677                                rc);
 678         }
 679         if (mds->mds_objects_dir != NULL) {
 680                 l_dput(mds->mds_objects_dir);
 681                 mds->mds_objects_dir = NULL;
 682         }
 683         if (mds->mds_logs_dir) {
 684                 l_dput(mds->mds_logs_dir);
 685                 mds->mds_logs_dir = NULL;
 686         }
 687         if (mds->mds_pending_dir) {
 688                 l_dput(mds->mds_pending_dir);
 689                 mds->mds_pending_dir = NULL;
 690         }
 691
 692         lquota_fs_cleanup(mds_quota_interface_ref, obd);
 693
 694         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 695         shrink_dcache_parent(mds->mds_fid_de);
 696         dput(mds->mds_fid_de);
 697         LL_DQUOT_OFF(obd->u.obt.obt_sb);
 698
 699         return rc;
 700 }
 701
 702 /* Creates an object with the same name as its fid.  Because this is not at all
 703  * performance sensitive, it is accomplished by creating a file, checking the
 704  * fid, and renaming it. */
 705 int mds_obd_create(struct obd_export *exp, struct obdo *oa,
 706                    struct lov_stripe_md **ea, struct obd_trans_info *oti)
 707 {
 708         struct mds_obd *mds = &exp->exp_obd->u.mds;
 709         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 710         unsigned int tmpname = ll_rand();
 711         struct file *filp;
 712         struct dentry *new_child;
 713         struct lvfs_run_ctxt saved;
 714         char fidname[LL_FID_NAMELEN];
 715         void *handle;
 716         struct lvfs_ucred ucred = { 0 };
 717         int rc = 0, err, namelen;
 718         ENTRY;
 719
 720         /* the owner of object file should always be root */
 721         cap_raise(ucred.luc_cap, CAP_SYS_RESOURCE);
 722
 723         if (strncmp(exp->exp_obd->obd_name, MDD_OBD_NAME,
 724                                    strlen(MDD_OBD_NAME))) {
 725                 RETURN(0);
 726         }
 727
 728         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, &ucred);
 729
 730         sprintf(fidname, "OBJECTS/%u.%u", tmpname, current->pid);
 731         filp = filp_open(fidname, O_CREAT | O_EXCL, 0666);
 732         if (IS_ERR(filp)) {
 733                 rc = PTR_ERR(filp);
 734                 if (rc == -EEXIST) {
 735                         CERROR("impossible object name collision %u\n",
 736                                tmpname);
 737                         LBUG();
 738                 }
 739                 CERROR("error creating tmp object %u: rc %d\n", tmpname, rc);
 740                 GOTO(out_pop, rc);
 741         }
 742
 743         LASSERT(mds->mds_objects_dir == filp->f_dentry->d_parent);
 744
 745         oa->o_id = filp->f_dentry->d_inode->i_ino;
 746         oa->o_generation = filp->f_dentry->d_inode->i_generation;
 747         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 748
 749         LOCK_INODE_MUTEX(parent_inode);
 750         new_child = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 751
 752         if (IS_ERR(new_child)) {
 753                 CERROR("getting neg dentry for obj rename: %d\n", rc);
 754                 GOTO(out_close, rc = PTR_ERR(new_child));
 755         }
 756         if (new_child->d_inode != NULL) {
 757                 CERROR("impossible non-negative obj dentry " LPU64":%u!\n",
 758                        oa->o_id, oa->o_generation);
 759                 LBUG();
 760         }
 761
 762         handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
 763                               FSFILT_OP_RENAME, NULL);
 764         if (IS_ERR(handle))
 765                 GOTO(out_dput, rc = PTR_ERR(handle));
 766
 767         lock_kernel();
 768         rc = vfs_rename(mds->mds_objects_dir->d_inode, filp->f_dentry,
 769                         mds->mds_objects_dir->d_inode, new_child);
 770         unlock_kernel();
 771         if (rc)
 772                 CERROR("error renaming new object "LPU64":%u: rc %d\n",
 773                        oa->o_id, oa->o_generation, rc);
 774
 775         err = fsfilt_commit(exp->exp_obd, mds->mds_objects_dir->d_inode,
 776                             handle, 0);
 777         if (!err) {
 778                 oa->o_gr = FILTER_GROUP_MDS0 + mds->mds_id;
 779                 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLGROUP;
 780         } else if (!rc)
 781                 rc = err;
 782 out_dput:
 783         dput(new_child);
 784 out_close:
 785         UNLOCK_INODE_MUTEX(parent_inode);
 786         err = filp_close(filp, 0);
 787         if (err) {
 788                 CERROR("closing tmpfile %u: rc %d\n", tmpname, rc);
 789                 if (!rc)
 790                         rc = err;
 791         }
 792 out_pop:
 793         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, &ucred);
 794         RETURN(rc);
 795 }
 796
 797 int mds_obd_destroy(struct obd_export *exp, struct obdo *oa,
 798                     struct lov_stripe_md *ea, struct obd_trans_info *oti,
 799                     struct obd_export *md_exp)
 800 {
 801         struct mds_obd *mds = &exp->exp_obd->u.mds;
 802         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 803         struct obd_device *obd = exp->exp_obd;
 804         struct lvfs_run_ctxt saved;
 805         struct lvfs_ucred ucred = { 0 };
 806         char fidname[LL_FID_NAMELEN];
 807         struct inode *inode = NULL;
 808         struct dentry *de;
 809         void *handle;
 810         int err, namelen, rc = 0;
 811         ENTRY;
 812
 813         cap_raise(ucred.luc_cap, CAP_SYS_RESOURCE);
 814         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &ucred);
 815
 816         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 817
 818         LOCK_INODE_MUTEX(parent_inode);
 819         de = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 820         if (IS_ERR(de)) {
 821                 rc = IS_ERR(de);
 822                 de = NULL;
 823                 CERROR("error looking up object "LPU64" %s: rc %d\n",
 824                        oa->o_id, fidname, rc);
 825                 GOTO(out_dput, rc);
 826         }
 827         if (de->d_inode == NULL) {
 828                 CERROR("destroying non-existent object "LPU64" %s: rc %d\n",
 829                        oa->o_id, fidname, rc);
 830                 GOTO(out_dput, rc = -ENOENT);
 831         }
 832
 833         /* Stripe count is 1 here since this is some MDS specific stuff
 834            that is unlinked, not spanned across multiple OSTs */
 835         handle = fsfilt_start_log(obd, mds->mds_objects_dir->d_inode,
 836                                   FSFILT_OP_UNLINK, oti, 1);
 837
 838         if (IS_ERR(handle))
 839                 GOTO(out_dput, rc = PTR_ERR(handle));
 840
 841         /* take a reference to protect inode from truncation within
 842            vfs_unlink() context. bug 10409 */
 843         inode = de->d_inode;
 844         atomic_inc(&inode->i_count);
 845         rc = vfs_unlink(mds->mds_objects_dir->d_inode, de);
 846         if (rc)
 847                 CERROR("error destroying object "LPU64":%u: rc %d\n",
 848                        oa->o_id, oa->o_generation, rc);
 849
 850         err = fsfilt_commit(obd, mds->mds_objects_dir->d_inode, handle, 0);
 851         if (err && !rc)
 852                 rc = err;
 853 out_dput:
 854         if (de != NULL)
 855                 l_dput(de);
 856         UNLOCK_INODE_MUTEX(parent_inode);
 857
 858         if (inode)
 859                 iput(inode);
 860
 861         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &ucred);
 862         RETURN(rc);
 863 }