lustre/mds/mds_fs.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  mds/mds_fs.c
   5  *  Lustre Metadata Server (MDS) filesystem interface code
   6  *
   7  *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   8  *   Author: Andreas Dilger <adilger@clusterfs.com>
   9  *
  10  *   This file is part of the Lustre file system, http://www.lustre.org
  11  *   Lustre is a trademark of Cluster File Systems, Inc.
  12  *
  13  *   You may have signed or agreed to another license before downloading
  14  *   this software.  If so, you are bound by the terms and conditions
  15  *   of that agreement, and the following does not apply to you.  See the
  16  *   LICENSE file included with this distribution for more information.
  17  *
  18  *   If you did not agree to a different license, then this copy of Lustre
  19  *   is open source software; you can redistribute it and/or modify it
  20  *   under the terms of version 2 of the GNU General Public License as
  21  *   published by the Free Software Foundation.
  22  *
  23  *   In either case, Lustre is distributed in the hope that it will be
  24  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  25  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  26  *   license text for more details.
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_MDS
  33
  34 #include <linux/module.h>
  35 #include <linux/kmod.h>
  36 #include <linux/version.h>
  37 #include <linux/sched.h>
  38 #include <lustre_quota.h>
  39 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  40 #include <linux/mount.h>
  41 #endif
  42 #include <lustre_mds.h>
  43 #include <obd_class.h>
  44 #include <obd_support.h>
  45 #include <lustre_lib.h>
  46 #include <lustre_fsfilt.h>
  47 #include <lustre_disk.h>
  48 #include <libcfs/list.h>
  49
  50 #include "mds_internal.h"
  51
  52
  53 static int mds_export_stats_init(struct obd_device *obd, struct obd_export *exp)
  54 {
  55         int rc, num_stats;
  56
  57         rc = lprocfs_exp_setup(exp);
  58         if (rc)
  59                 return rc;
  60         num_stats = (sizeof(*obd->obd_type->typ_ops) / sizeof(void *)) +
  61                      LPROC_MDS_LAST - 1;
  62         exp->exp_ops_stats = lprocfs_alloc_stats(num_stats,
  63                                                  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         int rc;
  88         ENTRY;
  89
  90         LASSERT(bitmap != NULL);
  91         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
  92
  93         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
  94         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
  95                 RETURN(0);
  96
  97         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
  98          * there's no need for extra complication here
  99          */
 100         if (new_client) {
 101                 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
 102         repeat:
 103                 if (cl_idx >= LR_MAX_CLIENTS ||
 104                     OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_CLIENT_ADD)) {
 105                         CERROR("no room for %u clients - fix LR_MAX_CLIENTS\n",
 106                                cl_idx);
 107                         return -EOVERFLOW;
 108                 }
 109                 if (test_and_set_bit(cl_idx, bitmap)) {
 110                         cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
 111                                                     cl_idx);
 112                         goto repeat;
 113                 }
 114         } else {
 115                 if (test_and_set_bit(cl_idx, bitmap)) {
 116                         CERROR("MDS client %d: bit already set in bitmap!!\n",
 117                                cl_idx);
 118                         LBUG();
 119                 }
 120         }
 121
 122         CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
 123                cl_idx, med->med_mcd->mcd_uuid);
 124
 125         med->med_lr_idx = cl_idx;
 126         med->med_lr_off = le32_to_cpu(mds->mds_server_data->lsd_client_start) +
 127                 (cl_idx * le16_to_cpu(mds->mds_server_data->lsd_client_size));
 128         LASSERTF(med->med_lr_off > 0, "med_lr_off = %llu\n", med->med_lr_off);
 129         mds_export_stats_init(obd, exp);
 130
 131         if (new_client) {
 132                 struct lvfs_run_ctxt saved;
 133                 loff_t off = med->med_lr_off;
 134                 struct file *file = mds->mds_rcvd_filp;
 135                 void *handle;
 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 using"
 302                                            " the wrong disk %s. Were the /dev/ "
 303                                            "assignments rearranged?\n",
 304                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
 305                         GOTO(err_msd, rc = -EINVAL);
 306                 }
 307                 /* COMPAT_146 */
 308                 /* Assume old last_rcvd format unless I_C_LR is set */
 309                 if (!(lsd->lsd_feature_incompat &
 310                       cpu_to_le32(OBD_INCOMPAT_COMMON_LR)))
 311                         lsd->lsd_mount_count = lsd->lsd_compat14;
 312                 /* end COMPAT_146 */
 313                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
 314         }
 315
 316         if (lsd->lsd_feature_incompat & ~cpu_to_le32(MDT_INCOMPAT_SUPP)) {
 317                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
 318                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
 319                        ~MDT_INCOMPAT_SUPP);
 320                 GOTO(err_msd, rc = -EINVAL);
 321         }
 322         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(MDT_ROCOMPAT_SUPP)) {
 323                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
 324                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
 325                        ~MDT_ROCOMPAT_SUPP);
 326                 /* Do something like remount filesystem read-only */
 327                 GOTO(err_msd, rc = -EINVAL);
 328         }
 329
 330         lsd->lsd_feature_compat = cpu_to_le32(OBD_COMPAT_MDT);
 331
 332         mds->mds_last_transno = le64_to_cpu(lsd->lsd_last_transno);
 333
 334         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 335                obd->obd_name, mds->mds_last_transno);
 336         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 337                obd->obd_name, mount_count + 1);
 338         CDEBUG(D_INODE, "%s: server data size: %u\n",
 339                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
 340         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 341                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
 342         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 343                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
 344         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 345                obd->obd_name, last_rcvd_size);
 346         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 347                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
 348                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
 349                 le16_to_cpu(lsd->lsd_client_size));
 350
 351         if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
 352             !lsd->lsd_client_size) {
 353                 CERROR("Bad last_rcvd contents!\n");
 354                 GOTO(err_msd, rc = -EINVAL);
 355         }
 356
 357         /* When we do a clean MDS shutdown, we save the last_transno into
 358          * the header.  If we find clients with higher last_transno values
 359          * then those clients may need recovery done. */
 360         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
 361              off < last_rcvd_size; cl_idx++) {
 362                 __u64 last_transno;
 363                 struct obd_export *exp;
 364                 struct mds_export_data *med;
 365
 366                 if (!mcd) {
 367                         OBD_ALLOC_WAIT(mcd, sizeof(*mcd));
 368                         if (!mcd)
 369                                 GOTO(err_client, rc = -ENOMEM);
 370                 }
 371
 372                 /* Don't assume off is incremented properly by
 373                  * fsfilt_read_record(), in case sizeof(*mcd)
 374                  * isn't the same as lsd->lsd_client_size.  */
 375                 off = le32_to_cpu(lsd->lsd_client_start) +
 376                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
 377                 rc = fsfilt_read_record(obd, file, mcd, sizeof(*mcd), &off);
 378                 if (rc) {
 379                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 380                                LAST_RCVD, cl_idx, off, rc);
 381                         break; /* read error shouldn't cause startup to fail */
 382                 }
 383
 384                 if (mcd->mcd_uuid[0] == '\0') {
 385                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 386                                cl_idx);
 387                         continue;
 388                 }
 389
 390                 last_transno = le64_to_cpu(mcd->mcd_last_transno) >
 391                                le64_to_cpu(mcd->mcd_last_close_transno) ?
 392                                le64_to_cpu(mcd->mcd_last_transno) :
 393                                le64_to_cpu(mcd->mcd_last_close_transno);
 394
 395                 /* These exports are cleaned up by mds_disconnect(), so they
 396                  * need to be set up like real exports as mds_connect() does.
 397                  */
 398                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 399                        " srv lr: "LPU64" lx: "LPU64"\n", mcd->mcd_uuid, cl_idx,
 400                        last_transno, le64_to_cpu(lsd->lsd_last_transno),
 401                        le64_to_cpu(mcd->mcd_last_xid));
 402
 403                 exp = class_new_export(obd, (struct obd_uuid *)mcd->mcd_uuid);
 404                 if (IS_ERR(exp)) {
 405                         if (PTR_ERR(exp) == -EALREADY) {
 406                                 /* export already exists, zero out this one */
 407                                 mcd->mcd_uuid[0] = '\0';
 408                         } else {
 409                                 GOTO(err_client, rc = PTR_ERR(exp));
 410                         }
 411                 } else {
 412                         med = &exp->exp_mds_data;
 413                         med->med_mcd = mcd;
 414                         rc = mds_client_add(obd, exp, cl_idx);
 415                         /* can't fail for existing client */
 416                         LASSERTF(rc == 0, "rc = %d\n", rc);
 417
 418                         mcd = NULL;
 419
 420                         spin_lock(&exp->exp_lock);
 421                         exp->exp_replay_needed = 1;
 422                         exp->exp_connecting = 0;
 423                         spin_unlock(&exp->exp_lock);
 424
 425                         obd->obd_recoverable_clients++;
 426                         obd->obd_max_recoverable_clients++;
 427                         class_export_put(exp);
 428                 }
 429
 430                 /* Need to check last_rcvd even for duplicated exports. */
 431                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
 432                        cl_idx, last_transno);
 433
 434                 if (last_transno > mds->mds_last_transno)
 435                         mds->mds_last_transno = last_transno;
 436         }
 437
 438         if (mcd)
 439                 OBD_FREE(mcd, sizeof(*mcd));
 440
 441         obd->obd_last_committed = mds->mds_last_transno;
 442
 443         if (obd->obd_recoverable_clients) {
 444                 CWARN("RECOVERY: service %s, %d recoverable clients, "
 445                       "last_transno "LPU64"\n", obd->obd_name,
 446                       obd->obd_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 = 0;
 450                 obd->obd_recovery_end = 0;
 451         }
 452
 453         mds->mds_mount_count = mount_count + 1;
 454         lsd->lsd_mount_count = lsd->lsd_compat14 =
 455                 cpu_to_le64(mds->mds_mount_count);
 456
 457         /* save it, so mount count and last_transno is current */
 458         rc = mds_update_server_data(obd, 1);
 459         if (rc)
 460                 GOTO(err_client, rc);
 461
 462         RETURN(0);
 463
 464 err_client:
 465         class_disconnect_exports(obd);
 466 err_msd:
 467         mds_server_free_data(mds);
 468         RETURN(rc);
 469 }
 470
 471 int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
 472 {
 473         struct mds_obd *mds = &obd->u.mds;
 474         struct lvfs_run_ctxt saved;
 475         struct dentry *dentry;
 476         struct file *file;
 477         int rc;
 478         ENTRY;
 479
 480         OBD_FAIL_RETURN(OBD_FAIL_MDS_FS_SETUP, -ENOENT);
 481
 482         rc = cleanup_group_info();
 483         if (rc)
 484                 RETURN(rc);
 485
 486         mds->mds_vfsmnt = mnt;
 487         /* why not mnt->mnt_sb instead of mnt->mnt_root->d_inode->i_sb? */
 488         obd->u.obt.obt_sb = mnt->mnt_root->d_inode->i_sb;
 489
 490         fsfilt_setup(obd, obd->u.obt.obt_sb);
 491
 492         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 493         obd->obd_lvfs_ctxt.pwdmnt = mnt;
 494         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
 495         obd->obd_lvfs_ctxt.fs = get_ds();
 496         obd->obd_lvfs_ctxt.cb_ops = mds_lvfs_ops;
 497
 498         /* setup the directory tree */
 499         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 500         dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755, 0);
 501         if (IS_ERR(dentry)) {
 502                 rc = PTR_ERR(dentry);
 503                 CERROR("cannot create ROOT directory: rc = %d\n", rc);
 504                 GOTO(err_pop, rc);
 505         }
 506
 507         mds->mds_rootfid.id = dentry->d_inode->i_ino;
 508         mds->mds_rootfid.generation = dentry->d_inode->i_generation;
 509         mds->mds_rootfid.f_type = S_IFDIR;
 510
 511         dput(dentry);
 512
 513         dentry = lookup_one_len("__iopen__", current->fs->pwd,
 514                                 strlen("__iopen__"));
 515         if (IS_ERR(dentry)) {
 516                 rc = PTR_ERR(dentry);
 517                 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
 518                 GOTO(err_pop, rc);
 519         }
 520
 521         mds->mds_fid_de = dentry;
 522         if (!dentry->d_inode || is_bad_inode(dentry->d_inode)) {
 523                 rc = -ENOENT;
 524                 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
 525                 GOTO(err_fid, rc);
 526         }
 527
 528         dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777, 1);
 529         if (IS_ERR(dentry)) {
 530                 rc = PTR_ERR(dentry);
 531                 CERROR("cannot create PENDING directory: rc = %d\n", rc);
 532                 GOTO(err_fid, rc);
 533         }
 534         mds->mds_pending_dir = dentry;
 535
 536         /* COMPAT_146 */
 537         dentry = simple_mkdir(current->fs->pwd, MDT_LOGS_DIR, 0777, 1);
 538         if (IS_ERR(dentry)) {
 539                 rc = PTR_ERR(dentry);
 540                 CERROR("cannot create %s directory: rc = %d\n",
 541                        MDT_LOGS_DIR, rc);
 542                 GOTO(err_pending, rc);
 543         }
 544         mds->mds_logs_dir = dentry;
 545         /* end COMPAT_146 */
 546
 547         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
 548         if (IS_ERR(dentry)) {
 549                 rc = PTR_ERR(dentry);
 550                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
 551                 GOTO(err_logs, rc);
 552         }
 553         mds->mds_objects_dir = dentry;
 554
 555         /* open and test the last rcvd file */
 556         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
 557         if (IS_ERR(file)) {
 558                 rc = PTR_ERR(file);
 559                 CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
 560                 GOTO(err_objects, rc = PTR_ERR(file));
 561         }
 562         mds->mds_rcvd_filp = file;
 563         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 564                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 565                        file->f_dentry->d_inode->i_mode);
 566                 GOTO(err_last_rcvd, rc = -ENOENT);
 567         }
 568
 569         rc = mds_init_server_data(obd, file);
 570         if (rc) {
 571                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
 572                 GOTO(err_last_rcvd, rc);
 573         }
 574
 575         /* open and test the lov objd file */
 576         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
 577         if (IS_ERR(file)) {
 578                 rc = PTR_ERR(file);
 579                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
 580                 GOTO(err_client, rc = PTR_ERR(file));
 581         }
 582         mds->mds_lov_objid_filp = file;
 583         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 584                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
 585                        file->f_dentry->d_inode->i_mode);
 586                 GOTO(err_lov_objid, rc = -ENOENT);
 587         }
 588
 589         /* open and test the check io file junk */
 590         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
 591         if (IS_ERR(file)) {
 592                 rc = PTR_ERR(file);
 593                 CERROR("cannot open/create %s file: rc = %d\n",HEALTH_CHECK,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 && filp_close(mds->mds_lov_objid_filp, 0))
 616                 CERROR("can't close %s after error\n", LOV_OBJID);
 617 err_client:
 618         class_disconnect_exports(obd);
 619 err_last_rcvd:
 620         if (mds->mds_rcvd_filp && filp_close(mds->mds_rcvd_filp, 0))
 621                 CERROR("can't close %s after error\n", LAST_RCVD);
 622 err_objects:
 623         dput(mds->mds_objects_dir);
 624 err_logs:
 625         dput(mds->mds_logs_dir);
 626 err_pending:
 627         dput(mds->mds_pending_dir);
 628 err_fid:
 629         dput(mds->mds_fid_de);
 630         goto err_pop;
 631 }
 632
 633
 634 int mds_fs_cleanup(struct obd_device *obd)
 635 {
 636         struct mds_obd *mds = &obd->u.mds;
 637         struct lvfs_run_ctxt saved;
 638         int rc = 0;
 639
 640         if (obd->obd_fail)
 641                 LCONSOLE_WARN("%s: shutting down for failover; client state "
 642                               "will be preserved.\n", obd->obd_name);
 643
 644         class_disconnect_exports(obd); /* cleans up client info too */
 645         mds_server_free_data(mds);
 646
 647         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 648         if (mds->mds_rcvd_filp) {
 649                 rc = filp_close(mds->mds_rcvd_filp, 0);
 650                 mds->mds_rcvd_filp = NULL;
 651                 if (rc)
 652                         CERROR("%s file won't close, rc=%d\n", LAST_RCVD, rc);
 653         }
 654         if (mds->mds_lov_objid_filp) {
 655                 rc = filp_close(mds->mds_lov_objid_filp, 0);
 656                 mds->mds_lov_objid_filp = NULL;
 657                 if (rc)
 658                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
 659         }
 660         if (mds->mds_health_check_filp) {
 661                 rc = filp_close(mds->mds_health_check_filp, 0);
 662                 mds->mds_health_check_filp = NULL;
 663                 if (rc)
 664                         CERROR("%s file won't close, rc=%d\n", HEALTH_CHECK,rc);
 665         }
 666         if (mds->mds_objects_dir != NULL) {
 667                 l_dput(mds->mds_objects_dir);
 668                 mds->mds_objects_dir = NULL;
 669         }
 670         if (mds->mds_logs_dir) {
 671                 l_dput(mds->mds_logs_dir);
 672                 mds->mds_logs_dir = NULL;
 673         }
 674         if (mds->mds_pending_dir) {
 675                 l_dput(mds->mds_pending_dir);
 676                 mds->mds_pending_dir = NULL;
 677         }
 678
 679         lquota_fs_cleanup(mds_quota_interface_ref, obd);
 680
 681         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 682         shrink_dcache_parent(mds->mds_fid_de);
 683         dput(mds->mds_fid_de);
 684         LL_DQUOT_OFF(obd->u.obt.obt_sb);
 685
 686         return rc;
 687 }
 688
 689 /* Creates an object with the same name as its fid.  Because this is not at all
 690  * performance sensitive, it is accomplished by creating a file, checking the
 691  * fid, and renaming it. */
 692 int mds_obd_create(struct obd_export *exp, struct obdo *oa,
 693                    struct lov_stripe_md **ea, struct obd_trans_info *oti)
 694 {
 695         struct mds_obd *mds = &exp->exp_obd->u.mds;
 696         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 697         unsigned int tmpname = ll_rand();
 698         struct file *filp;
 699         struct dentry *new_child;
 700         struct lvfs_run_ctxt saved;
 701         char fidname[LL_FID_NAMELEN];
 702         void *handle;
 703         struct lvfs_ucred ucred = { 0 };
 704         int rc = 0, err, namelen;
 705         ENTRY;
 706
 707         /* the owner of object file should always be root */
 708         cap_raise(ucred.luc_cap, CAP_SYS_RESOURCE);
 709
 710         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, &ucred);
 711
 712         sprintf(fidname, "OBJECTS/%u.%u", tmpname, current->pid);
 713         filp = filp_open(fidname, O_CREAT | O_EXCL, 0666);
 714         if (IS_ERR(filp)) {
 715                 rc = PTR_ERR(filp);
 716                 if (rc == -EEXIST) {
 717                         CERROR("impossible object name collision %u\n",
 718                                tmpname);
 719                         LBUG();
 720                 }
 721                 CERROR("error creating tmp object %u: rc %d\n", tmpname, rc);
 722                 GOTO(out_pop, rc);
 723         }
 724
 725         LASSERT(mds->mds_objects_dir == filp->f_dentry->d_parent);
 726
 727         oa->o_id = filp->f_dentry->d_inode->i_ino;
 728         oa->o_generation = filp->f_dentry->d_inode->i_generation;
 729         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 730
 731         LOCK_INODE_MUTEX(parent_inode);
 732         new_child = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 733
 734         if (IS_ERR(new_child)) {
 735                 CERROR("getting neg dentry for obj rename: %d\n", rc);
 736                 GOTO(out_close, rc = PTR_ERR(new_child));
 737         }
 738         if (new_child->d_inode != NULL) {
 739                 CERROR("impossible non-negative obj dentry " LPU64":%u!\n",
 740                        oa->o_id, oa->o_generation);
 741                 LBUG();
 742         }
 743
 744         handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
 745                               FSFILT_OP_RENAME, NULL);
 746         if (IS_ERR(handle))
 747                 GOTO(out_dput, rc = PTR_ERR(handle));
 748
 749         lock_kernel();
 750         rc = vfs_rename(mds->mds_objects_dir->d_inode, filp->f_dentry,
 751                         mds->mds_objects_dir->d_inode, new_child);
 752         unlock_kernel();
 753         if (rc)
 754                 CERROR("error renaming new object "LPU64":%u: rc %d\n",
 755                        oa->o_id, oa->o_generation, rc);
 756
 757         err = fsfilt_commit(exp->exp_obd, mds->mds_objects_dir->d_inode,
 758                             handle, 0);
 759         if (!err)
 760                 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
 761         else if (!rc)
 762                 rc = err;
 763 out_dput:
 764         dput(new_child);
 765 out_close:
 766         UNLOCK_INODE_MUTEX(parent_inode);
 767         err = filp_close(filp, 0);
 768         if (err) {
 769                 CERROR("closing tmpfile %u: rc %d\n", tmpname, rc);
 770                 if (!rc)
 771                         rc = err;
 772         }
 773 out_pop:
 774         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, &ucred);
 775         RETURN(rc);
 776 }
 777
 778 int mds_obd_destroy(struct obd_export *exp, struct obdo *oa,
 779                     struct lov_stripe_md *ea, struct obd_trans_info *oti,
 780                     struct obd_export *md_exp)
 781 {
 782         struct mds_obd *mds = &exp->exp_obd->u.mds;
 783         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 784         struct obd_device *obd = exp->exp_obd;
 785         struct lvfs_run_ctxt saved;
 786         struct lvfs_ucred ucred = { 0 };
 787         char fidname[LL_FID_NAMELEN];
 788         struct inode *inode = NULL;
 789         struct dentry *de;
 790         void *handle;
 791         int err, namelen, rc = 0;
 792         ENTRY;
 793
 794         cap_raise(ucred.luc_cap, CAP_SYS_RESOURCE);
 795         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &ucred);
 796
 797         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 798
 799         LOCK_INODE_MUTEX(parent_inode);
 800         de = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 801         if (IS_ERR(de)) {
 802                 rc = IS_ERR(de);
 803                 de = NULL;
 804                 CERROR("error looking up object "LPU64" %s: rc %d\n",
 805                        oa->o_id, fidname, rc);
 806                 GOTO(out_dput, rc);
 807         }
 808         if (de->d_inode == NULL) {
 809                 CERROR("destroying non-existent object "LPU64" %s: rc %d\n",
 810                        oa->o_id, fidname, rc);
 811                 GOTO(out_dput, rc = -ENOENT);
 812         }
 813
 814         /* Stripe count is 1 here since this is some MDS specific stuff
 815            that is unlinked, not spanned across multiple OSTs */
 816         handle = fsfilt_start_log(obd, mds->mds_objects_dir->d_inode,
 817                                   FSFILT_OP_UNLINK, oti, 1);
 818
 819         if (IS_ERR(handle))
 820                 GOTO(out_dput, rc = PTR_ERR(handle));
 821
 822         /* take a reference to protect inode from truncation within
 823            vfs_unlink() context. bug 10409 */
 824         inode = de->d_inode;
 825         atomic_inc(&inode->i_count);
 826         rc = vfs_unlink(mds->mds_objects_dir->d_inode, de);
 827         if (rc)
 828                 CERROR("error destroying object "LPU64":%u: rc %d\n",
 829                        oa->o_id, oa->o_generation, rc);
 830
 831         err = fsfilt_commit(obd, mds->mds_objects_dir->d_inode, handle, 0);
 832         if (err && !rc)
 833                 rc = err;
 834 out_dput:
 835         if (de != NULL)
 836                 l_dput(de);
 837         UNLOCK_INODE_MUTEX(parent_inode);
 838
 839         if (inode)
 840                 iput(inode);
 841
 842         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &ucred);
 843         RETURN(rc);
 844 }