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