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