lustre/mds/mds_fs.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  mds/mds_fs.c
   5  *  Lustre Metadata Server (MDS) filesystem interface code
   6  *
   7  *  Copyright (C) 2002, 2003 Cluster File Systems, Inc.
   8  *   Author: Andreas Dilger <adilger@clusterfs.com>
   9  *
  10  *   This file is part of the Lustre file system, http://www.lustre.org
  11  *   Lustre is a trademark of Cluster File Systems, Inc.
  12  *
  13  *   You may have signed or agreed to another license before downloading
  14  *   this software.  If so, you are bound by the terms and conditions
  15  *   of that agreement, and the following does not apply to you.  See the
  16  *   LICENSE file included with this distribution for more information.
  17  *
  18  *   If you did not agree to a different license, then this copy of Lustre
  19  *   is open source software; you can redistribute it and/or modify it
  20  *   under the terms of version 2 of the GNU General Public License as
  21  *   published by the Free Software Foundation.
  22  *
  23  *   In either case, Lustre is distributed in the hope that it will be
  24  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  25  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  26  *   license text for more details.
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_MDS
  33
  34 #include <linux/module.h>
  35 #include <linux/kmod.h>
  36 #include <linux/version.h>
  37 #include <linux/sched.h>
  38 #include <lustre_quota.h>
  39 #include <linux/mount.h>
  40 #include <lustre_mds.h>
  41 #include <obd_class.h>
  42 #include <obd_support.h>
  43 #include <lustre_lib.h>
  44 #include <lustre_fsfilt.h>
  45 #include <lustre_disk.h>
  46 #include <libcfs/list.h>
  47
  48 #include "mds_internal.h"
  49
  50
  51 static int mds_export_stats_init(struct obd_device *obd,
  52                                  struct obd_export *exp,
  53                                  lnet_nid_t client_nid)
  54 {
  55         int rc, num_stats, newnid;
  56         rc = lprocfs_exp_setup(exp, client_nid, &newnid);
  57         if (rc)
  58                 return rc;
  59
  60         if (client_nid && newnid) {
  61                 struct nid_stat *tmp = exp->exp_nid_stats;
  62                 LASSERT(tmp != NULL);
  63                 num_stats = (sizeof(*obd->obd_type->typ_dt_ops) / sizeof(void *)) +
  64                              LPROC_MDS_LAST - 1;
  65                 tmp->nid_stats = lprocfs_alloc_stats(num_stats,
  66                                                      LPROCFS_STATS_FLAG_NOPERCPU);
  67                 if (tmp->nid_stats == NULL)
  68                         return -ENOMEM;
  69
  70                 lprocfs_init_ops_stats(LPROC_MDS_LAST, tmp->nid_stats);
  71                 rc = lprocfs_register_stats(tmp->nid_proc, "stats",
  72                                             tmp->nid_stats);
  73                 if (rc)
  74                         return rc;
  75                 mds_stats_counter_init(tmp->nid_stats);
  76         }
  77         return 0;
  78 }
  79
  80 /* Add client data to the MDS.  We use a bitmap to locate a free space
  81  * in the last_rcvd file if cl_off is -1 (i.e. a new client).
  82  * Otherwise, we have just read the data from the last_rcvd file and
  83  * we know its offset.
  84  *
  85  * It should not be possible to fail adding an existing client - otherwise
  86  * mds_init_server_data() callsite needs to be fixed.
  87  */
  88 int mds_client_add(struct obd_device *obd, struct obd_export *exp,
  89                    int cl_idx, lnet_nid_t client_nid)
  90 {
  91         struct mds_obd *mds = &obd->u.mds;
  92         struct mds_export_data *med = &exp->exp_mds_data;
  93         unsigned long *bitmap = mds->mds_client_bitmap;
  94         int new_client = (cl_idx == -1);
  95         ENTRY;
  96
  97         LASSERT(bitmap != NULL);
  98         LASSERTF(cl_idx > -2, "%d\n", cl_idx);
  99
 100         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 101         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
 102                 RETURN(0);
 103
 104         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
 105          * there's no need for extra complication here
 106          */
 107         if (new_client) {
 108                 cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
 109         repeat:
 110                 if (cl_idx >= LR_MAX_CLIENTS ||
 111                     OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
 112                         CERROR("no room for %u client - fix LR_MAX_CLIENTS\n",
 113                                cl_idx);
 114                         return -EOVERFLOW;
 115                 }
 116                 if (test_and_set_bit(cl_idx, bitmap)) {
 117                         cl_idx = find_next_zero_bit(bitmap, LR_MAX_CLIENTS,
 118                                                     cl_idx);
 119                         goto repeat;
 120                 }
 121         } else {
 122                 if (test_and_set_bit(cl_idx, bitmap)) {
 123                         CERROR("MDS client %d: bit already set in bitmap!!\n",
 124                                cl_idx);
 125                         LBUG();
 126                 }
 127         }
 128
 129         CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
 130                cl_idx, med->med_mcd->mcd_uuid);
 131
 132         med->med_lr_idx = cl_idx;
 133         med->med_lr_off = le32_to_cpu(mds->mds_server_data->lsd_client_start) +
 134                 (cl_idx * le16_to_cpu(mds->mds_server_data->lsd_client_size));
 135         LASSERTF(med->med_lr_off > 0, "med_lr_off = %llu\n", med->med_lr_off);
 136         mds_export_stats_init(obd, exp, client_nid);
 137
 138         if (new_client) {
 139                 struct lvfs_run_ctxt saved;
 140                 loff_t off = med->med_lr_off;
 141                 struct file *file = mds->mds_rcvd_filp;
 142                 void *handle;
 143                 int rc;
 144
 145                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 146                 handle = fsfilt_start(obd, file->f_dentry->d_inode,
 147                                       FSFILT_OP_SETATTR, NULL);
 148                 if (IS_ERR(handle)) {
 149                         rc = PTR_ERR(handle);
 150                         CERROR("unable to start transaction: rc %d\n", rc);
 151                 } else {
 152                         rc = fsfilt_add_journal_cb(obd, 0, handle,
 153                                                    target_client_add_cb, exp);
 154                         if (rc == 0) {
 155                                 spin_lock(&exp->exp_lock);
 156                                 exp->exp_need_sync = 1;
 157                                 spin_unlock(&exp->exp_lock);
 158                         }
 159                         rc = fsfilt_write_record(obd, file, med->med_mcd,
 160                                                  sizeof(*med->med_mcd),
 161                                                  &off, rc /* sync if no cb */);
 162                         fsfilt_commit(obd, file->f_dentry->d_inode, handle, 0);
 163                 }
 164
 165                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 166
 167                 if (rc)
 168                         return rc;
 169                 CDEBUG(D_INFO, "wrote client mcd at idx %u off %llu (len %u)\n",
 170                        med->med_lr_idx, med->med_lr_off,
 171                        (unsigned int)sizeof(*med->med_mcd));
 172         }
 173         return 0;
 174 }
 175
 176 int mds_client_free(struct obd_export *exp)
 177 {
 178         struct mds_export_data *med = &exp->exp_mds_data;
 179         struct mds_obd *mds = &exp->exp_obd->u.mds;
 180         struct obd_device *obd = exp->exp_obd;
 181         struct mds_client_data zero_mcd;
 182         struct lvfs_run_ctxt saved;
 183         int rc;
 184         loff_t off;
 185         ENTRY;
 186
 187         if (!med->med_mcd)
 188                 RETURN(0);
 189
 190         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 191         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
 192                 GOTO(free, 0);
 193
 194         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld with UUID '%s'\n",
 195                med->med_lr_idx, med->med_lr_off, med->med_mcd->mcd_uuid);
 196
 197         LASSERT(mds->mds_client_bitmap != NULL);
 198
 199         lprocfs_exp_cleanup(exp);
 200
 201         off = med->med_lr_off;
 202
 203         /* Don't clear med_lr_idx here as it is likely also unset.  At worst
 204          * we leak a client slot that will be cleaned on the next recovery. */
 205         if (off <= 0) {
 206                 CERROR("%s: client idx %d has offset %lld\n",
 207                         obd->obd_name, med->med_lr_idx, off);
 208                 GOTO(free, rc = -EINVAL);
 209         }
 210
 211         /* Clear the bit _after_ zeroing out the client so we don't
 212            race with mds_client_add and zero out new clients.*/
 213         if (!test_bit(med->med_lr_idx, mds->mds_client_bitmap)) {
 214                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 215                        med->med_lr_idx);
 216                 LBUG();
 217         }
 218
 219         if (!(exp->exp_flags & OBD_OPT_FAILOVER)) {
 220                 memset(&zero_mcd, 0, sizeof zero_mcd);
 221                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 222                 rc = fsfilt_write_record(obd, mds->mds_rcvd_filp, &zero_mcd,
 223                                          sizeof(zero_mcd), &off,
 224                                          (!exp->exp_libclient ||
 225                                           exp->exp_need_sync));
 226                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 227
 228                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 229                        "zeroing out client %s idx %u in %s rc %d\n",
 230                        med->med_mcd->mcd_uuid, med->med_lr_idx, LAST_RCVD, rc);
 231         }
 232
 233         if (!test_and_clear_bit(med->med_lr_idx, mds->mds_client_bitmap)) {
 234                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 235                        med->med_lr_idx);
 236                 LBUG();
 237         }
 238
 239
 240         /* Make sure the server's last_transno is up to date. Do this
 241          * after the client is freed so we know all the client's
 242          * transactions have been committed. */
 243         mds_update_server_data(exp->exp_obd, 0);
 244
 245         EXIT;
 246  free:
 247         OBD_FREE(med->med_mcd, sizeof(*med->med_mcd));
 248         med->med_mcd = NULL;
 249
 250         return 0;
 251 }
 252
 253 static int mds_server_free_data(struct mds_obd *mds)
 254 {
 255         OBD_FREE(mds->mds_client_bitmap, LR_MAX_CLIENTS / 8);
 256         OBD_FREE(mds->mds_server_data, sizeof(*mds->mds_server_data));
 257         mds->mds_server_data = NULL;
 258
 259         return 0;
 260 }
 261
 262 static int mds_init_server_data(struct obd_device *obd, struct file *file)
 263 {
 264         struct mds_obd *mds = &obd->u.mds;
 265         struct lr_server_data *lsd;
 266         struct mds_client_data *mcd = NULL;
 267         loff_t off = 0;
 268         unsigned long last_rcvd_size = i_size_read(file->f_dentry->d_inode);
 269         __u64 mount_count;
 270         int cl_idx, rc = 0;
 271         ENTRY;
 272
 273         /* ensure padding in the struct is the correct size */
 274         LASSERT(offsetof(struct lr_server_data, lsd_padding) +
 275                 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
 276         LASSERT(offsetof(struct mds_client_data, mcd_padding) +
 277                 sizeof(mcd->mcd_padding) == LR_CLIENT_SIZE);
 278
 279         OBD_ALLOC_WAIT(lsd, sizeof(*lsd));
 280         if (!lsd)
 281                 RETURN(-ENOMEM);
 282
 283         OBD_ALLOC_WAIT(mds->mds_client_bitmap, LR_MAX_CLIENTS / 8);
 284         if (!mds->mds_client_bitmap) {
 285                 OBD_FREE(lsd, sizeof(*lsd));
 286                 RETURN(-ENOMEM);
 287         }
 288
 289         mds->mds_server_data = lsd;
 290
 291         if (last_rcvd_size == 0) {
 292                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
 293
 294                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,sizeof(lsd->lsd_uuid));
 295                 lsd->lsd_last_transno = 0;
 296                 mount_count = lsd->lsd_mount_count = 0;
 297                 lsd->lsd_server_size = cpu_to_le32(LR_SERVER_SIZE);
 298                 lsd->lsd_client_start = cpu_to_le32(LR_CLIENT_START);
 299                 lsd->lsd_client_size = cpu_to_le16(LR_CLIENT_SIZE);
 300                 lsd->lsd_feature_rocompat = cpu_to_le32(OBD_ROCOMPAT_LOVOBJID);
 301                 lsd->lsd_feature_incompat = cpu_to_le32(OBD_INCOMPAT_MDT);
 302         } else {
 303                 rc = fsfilt_read_record(obd, file, lsd, sizeof(*lsd), &off);
 304                 if (rc) {
 305                         CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
 306                         GOTO(err_msd, rc);
 307                 }
 308                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
 309                         LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s "
 310                                            "using the wrong disk %s. Were the "
 311                                            "/dev/ assignments rearranged?\n",
 312                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
 313                         GOTO(err_msd, rc = -EINVAL);
 314                 }
 315                 mount_count = le64_to_cpu(lsd->lsd_mount_count);
 316         }
 317
 318         if (lsd->lsd_feature_incompat & ~cpu_to_le32(MDT_INCOMPAT_SUPP)) {
 319                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
 320                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_incompat) &
 321                        ~MDT_INCOMPAT_SUPP);
 322                 GOTO(err_msd, rc = -EINVAL);
 323         }
 324         if (lsd->lsd_feature_rocompat & ~cpu_to_le32(MDT_ROCOMPAT_SUPP)) {
 325                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
 326                        obd->obd_name, le32_to_cpu(lsd->lsd_feature_rocompat) &
 327                        ~MDT_ROCOMPAT_SUPP);
 328                 /* Do something like remount filesystem read-only */
 329                 GOTO(err_msd, rc = -EINVAL);
 330         }
 331
 332         lsd->lsd_feature_compat = cpu_to_le32(OBD_COMPAT_MDT);
 333
 334         mds->mds_last_transno = le64_to_cpu(lsd->lsd_last_transno);
 335
 336         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 337                obd->obd_name, mds->mds_last_transno);
 338         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 339                obd->obd_name, mount_count + 1);
 340         CDEBUG(D_INODE, "%s: server data size: %u\n",
 341                obd->obd_name, le32_to_cpu(lsd->lsd_server_size));
 342         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 343                obd->obd_name, le32_to_cpu(lsd->lsd_client_start));
 344         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 345                obd->obd_name, le32_to_cpu(lsd->lsd_client_size));
 346         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 347                obd->obd_name, last_rcvd_size);
 348         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 349                last_rcvd_size <= le32_to_cpu(lsd->lsd_client_start) ? 0 :
 350                (last_rcvd_size - le32_to_cpu(lsd->lsd_client_start)) /
 351                 le16_to_cpu(lsd->lsd_client_size));
 352
 353         if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
 354             !lsd->lsd_client_size) {
 355                 CERROR("Bad last_rcvd contents!\n");
 356                 GOTO(err_msd, rc = -EINVAL);
 357         }
 358
 359         /* When we do a clean MDS shutdown, we save the last_transno into
 360          * the header.  If we find clients with higher last_transno values
 361          * then those clients may need recovery done. */
 362         for (cl_idx = 0, off = le32_to_cpu(lsd->lsd_client_start);
 363              off < last_rcvd_size; cl_idx++) {
 364                 __u64 last_transno;
 365                 struct obd_export *exp;
 366                 struct mds_export_data *med;
 367
 368                 if (!mcd) {
 369                         OBD_ALLOC_WAIT(mcd, sizeof(*mcd));
 370                         if (!mcd)
 371                                 GOTO(err_client, rc = -ENOMEM);
 372                 }
 373
 374                 /* Don't assume off is incremented properly by
 375                  * fsfilt_read_record(), in case sizeof(*mcd)
 376                  * isn't the same as lsd->lsd_client_size.  */
 377                 off = le32_to_cpu(lsd->lsd_client_start) +
 378                         cl_idx * le16_to_cpu(lsd->lsd_client_size);
 379                 rc = fsfilt_read_record(obd, file, mcd, sizeof(*mcd), &off);
 380                 if (rc) {
 381                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 382                                LAST_RCVD, cl_idx, off, rc);
 383                         break; /* read error shouldn't cause startup to fail */
 384                 }
 385
 386                 if (mcd->mcd_uuid[0] == '\0') {
 387                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 388                                cl_idx);
 389                         continue;
 390                 }
 391
 392                 last_transno = le64_to_cpu(mcd->mcd_last_transno) >
 393                                le64_to_cpu(mcd->mcd_last_close_transno) ?
 394                                le64_to_cpu(mcd->mcd_last_transno) :
 395                                le64_to_cpu(mcd->mcd_last_close_transno);
 396
 397                 /* These exports are cleaned up by mds_disconnect(), so they
 398                  * need to be set up like real exports as mds_connect() does.
 399                  */
 400                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 401                        " srv lr: "LPU64" lx: "LPU64"\n", mcd->mcd_uuid, cl_idx,
 402                        last_transno, le64_to_cpu(lsd->lsd_last_transno),
 403                        le64_to_cpu(mcd->mcd_last_xid));
 404
 405                 exp = class_new_export(obd, (struct obd_uuid *)mcd->mcd_uuid);
 406                 if (IS_ERR(exp)) {
 407                         if (PTR_ERR(exp) == -EALREADY) {
 408                                 /* export already exists, zero out this one */
 409                                 mcd->mcd_uuid[0] = '\0';
 410                         } else {
 411                                 GOTO(err_client, rc = PTR_ERR(exp));
 412                         }
 413                 } else {
 414                         med = &exp->exp_mds_data;
 415                         med->med_mcd = mcd;
 416                         rc = mds_client_add(obd, exp, cl_idx, 0);
 417                         LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
 418
 419                         mcd = NULL;
 420                         spin_lock(&exp->exp_lock);
 421                         exp->exp_req_replay_needed = 1;
 422                         exp->exp_connecting = 0;
 423                         spin_unlock(&exp->exp_lock);
 424                         obd->obd_max_recoverable_clients++;
 425                         class_export_put(exp);
 426                 }
 427                 /* Need to check last_rcvd even for duplicated exports. */
 428                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
 429                        cl_idx, last_transno);
 430
 431                 if (last_transno > mds->mds_last_transno)
 432                         mds->mds_last_transno = last_transno;
 433         }
 434
 435         if (mcd)
 436                 OBD_FREE(mcd, sizeof(*mcd));
 437
 438         obd->obd_last_committed = mds->mds_last_transno;
 439
 440         if (obd->obd_max_recoverable_clients) {
 441                 /* shouldn't happen in b_new_cmd */
 442                 LBUG();
 443                 CWARN("RECOVERY: service %s, %d recoverable clients, "
 444                       "last_transno "LPU64"\n", obd->obd_name,
 445                       obd->obd_max_recoverable_clients, mds->mds_last_transno);
 446                 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
 447                 obd->obd_recovering = 1;
 448                 obd->obd_recovery_start = CURRENT_SECONDS;
 449                 /* Only used for lprocfs_status */
 450                 obd->obd_recovery_end = obd->obd_recovery_start +
 451                         OBD_RECOVERY_TIMEOUT;
 452         }
 453
 454         mds->mds_mount_count = mount_count + 1;
 455         lsd->lsd_mount_count = lsd->lsd_compat14 =
 456                 cpu_to_le64(mds->mds_mount_count);
 457
 458         /* save it, so mount count and last_transno is current */
 459         rc = mds_update_server_data(obd, 1);
 460         if (rc)
 461                 GOTO(err_client, rc);
 462
 463         RETURN(0);
 464
 465 err_client:
 466         class_disconnect_exports(obd);
 467 err_msd:
 468         mds_server_free_data(mds);
 469         RETURN(rc);
 470 }
 471
 472 void mds_init_ctxt(struct obd_device *obd, struct vfsmount *mnt)
 473 {
 474         struct mds_obd *mds = &obd->u.mds;
 475
 476         mds->mds_vfsmnt = mnt;
 477         /* why not mnt->mnt_sb instead of mnt->mnt_root->d_inode->i_sb? */
 478         obd->u.obt.obt_sb = mnt->mnt_root->d_inode->i_sb;
 479
 480         fsfilt_setup(obd, obd->u.obt.obt_sb);
 481
 482         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 483         obd->obd_lvfs_ctxt.pwdmnt = mnt;
 484         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
 485         obd->obd_lvfs_ctxt.fs = get_ds();
 486         obd->obd_lvfs_ctxt.cb_ops = mds_lvfs_ops;
 487         return;
 488 }
 489
 490 int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
 491 {
 492         struct mds_obd *mds = &obd->u.mds;
 493         struct lvfs_run_ctxt saved;
 494         struct dentry *dentry;
 495         struct file *file;
 496         int rc;
 497         ENTRY;
 498
 499         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
 500                 RETURN(-ENOENT);
 501
 502         rc = cleanup_group_info();
 503         if (rc)
 504                 RETURN(rc);
 505
 506         mds_init_ctxt(obd, mnt);
 507
 508         /* setup the directory tree */
 509         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 510         dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755, 0);
 511         if (IS_ERR(dentry)) {
 512                 rc = PTR_ERR(dentry);
 513                 CERROR("cannot create ROOT directory: rc = %d\n", rc);
 514                 GOTO(err_pop, rc);
 515         }
 516
 517         mds->mds_rootfid.id = dentry->d_inode->i_ino;
 518         mds->mds_rootfid.generation = dentry->d_inode->i_generation;
 519         mds->mds_rootfid.f_type = S_IFDIR;
 520
 521         dput(dentry);
 522
 523         dentry = lookup_one_len("__iopen__", current->fs->pwd,
 524                                 strlen("__iopen__"));
 525         if (IS_ERR(dentry)) {
 526                 rc = PTR_ERR(dentry);
 527                 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
 528                 GOTO(err_pop, rc);
 529         }
 530
 531         mds->mds_fid_de = dentry;
 532         if (!dentry->d_inode || is_bad_inode(dentry->d_inode)) {
 533                 rc = -ENOENT;
 534                 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
 535                 GOTO(err_fid, rc);
 536         }
 537
 538         dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777, 1);
 539         if (IS_ERR(dentry)) {
 540                 rc = PTR_ERR(dentry);
 541                 CERROR("cannot create PENDING directory: rc = %d\n", rc);
 542                 GOTO(err_fid, rc);
 543         }
 544         mds->mds_pending_dir = dentry;
 545
 546         /* COMPAT_146 */
 547         dentry = simple_mkdir(current->fs->pwd, MDT_LOGS_DIR, 0777, 1);
 548         if (IS_ERR(dentry)) {
 549                 rc = PTR_ERR(dentry);
 550                 CERROR("cannot create %s directory: rc = %d\n",
 551                        MDT_LOGS_DIR, rc);
 552                 GOTO(err_pending, rc);
 553         }
 554         mds->mds_logs_dir = dentry;
 555         /* end COMPAT_146 */
 556
 557         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
 558         if (IS_ERR(dentry)) {
 559                 rc = PTR_ERR(dentry);
 560                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
 561                 GOTO(err_logs, rc);
 562         }
 563         mds->mds_objects_dir = dentry;
 564
 565         /* open and test the last rcvd file */
 566         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
 567         if (IS_ERR(file)) {
 568                 rc = PTR_ERR(file);
 569                 CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
 570                 GOTO(err_objects, rc = PTR_ERR(file));
 571         }
 572         mds->mds_rcvd_filp = file;
 573         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 574                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 575                        file->f_dentry->d_inode->i_mode);
 576                 GOTO(err_last_rcvd, rc = -ENOENT);
 577         }
 578
 579         rc = mds_init_server_data(obd, file);
 580         if (rc) {
 581                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
 582                 GOTO(err_last_rcvd, rc);
 583         }
 584
 585         /* open and test the lov objd file */
 586         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
 587         if (IS_ERR(file)) {
 588                 rc = PTR_ERR(file);
 589                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
 590                 GOTO(err_client, rc = PTR_ERR(file));
 591         }
 592         mds->mds_lov_objid_filp = file;
 593         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 594                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
 595                        file->f_dentry->d_inode->i_mode);
 596                 GOTO(err_lov_objid, rc = -ENOENT);
 597         }
 598
 599         /* open and test the check io file junk */
 600         file = filp_open(HEALTH_CHECK, O_RDWR | O_CREAT, 0644);
 601         if (IS_ERR(file)) {
 602                 rc = PTR_ERR(file);
 603                 CERROR("cannot open/create %s file: rc = %d\n", HEALTH_CHECK,
 604                        rc);
 605                 GOTO(err_lov_objid, rc = PTR_ERR(file));
 606         }
 607         mds->mds_health_check_filp = file;
 608         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 609                 CERROR("%s is not a regular file!: mode = %o\n", HEALTH_CHECK,
 610                        file->f_dentry->d_inode->i_mode);
 611                 GOTO(err_health_check, rc = -ENOENT);
 612         }
 613         rc = lvfs_check_io_health(obd, file);
 614         if (rc)
 615                 GOTO(err_health_check, rc);
 616 err_pop:
 617         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 618
 619         return rc;
 620
 621 err_health_check:
 622         if (mds->mds_health_check_filp &&
 623             filp_close(mds->mds_health_check_filp, 0))
 624                 CERROR("can't close %s after error\n", HEALTH_CHECK);
 625 err_lov_objid:
 626         if (mds->mds_lov_objid_filp &&
 627                 filp_close((struct file *)mds->mds_lov_objid_filp, 0))
 628                 CERROR("can't close %s after error\n", LOV_OBJID);
 629 err_client:
 630         class_disconnect_exports(obd);
 631 err_last_rcvd:
 632         if (mds->mds_rcvd_filp && filp_close(mds->mds_rcvd_filp, 0))
 633                 CERROR("can't close %s after error\n", LAST_RCVD);
 634 err_objects:
 635         dput(mds->mds_objects_dir);
 636 err_logs:
 637         dput(mds->mds_logs_dir);
 638 err_pending:
 639         dput(mds->mds_pending_dir);
 640 err_fid:
 641         dput(mds->mds_fid_de);
 642         goto err_pop;
 643 }
 644
 645 int mds_fs_cleanup(struct obd_device *obd)
 646 {
 647         struct mds_obd *mds = &obd->u.mds;
 648         struct lvfs_run_ctxt saved;
 649         int rc = 0;
 650
 651         if (obd->obd_fail)
 652                 LCONSOLE_WARN("%s: shutting down for failover; client state "
 653                               "will be preserved.\n", obd->obd_name);
 654
 655         class_disconnect_exports(obd); /* cleans up client info too */
 656         mds_server_free_data(mds);
 657
 658         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 659         if (mds->mds_rcvd_filp) {
 660                 rc = filp_close(mds->mds_rcvd_filp, 0);
 661                 mds->mds_rcvd_filp = NULL;
 662                 if (rc)
 663                         CERROR("%s file won't close, rc=%d\n", LAST_RCVD, rc);
 664         }
 665         if (mds->mds_lov_objid_filp) {
 666                 rc = filp_close((struct file *)mds->mds_lov_objid_filp, 0);
 667                 mds->mds_lov_objid_filp = NULL;
 668                 if (rc)
 669                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
 670         }
 671         if (mds->mds_health_check_filp) {
 672                 rc = filp_close(mds->mds_health_check_filp, 0);
 673                 mds->mds_health_check_filp = NULL;
 674                 if (rc)
 675                         CERROR("%s file won't close, rc=%d\n", HEALTH_CHECK,
 676                                rc);
 677         }
 678         if (mds->mds_objects_dir != NULL) {
 679                 l_dput(mds->mds_objects_dir);
 680                 mds->mds_objects_dir = NULL;
 681         }
 682         if (mds->mds_logs_dir) {
 683                 l_dput(mds->mds_logs_dir);
 684                 mds->mds_logs_dir = NULL;
 685         }
 686         if (mds->mds_pending_dir) {
 687                 l_dput(mds->mds_pending_dir);
 688                 mds->mds_pending_dir = NULL;
 689         }
 690
 691         lquota_fs_cleanup(mds_quota_interface_ref, obd);
 692
 693         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 694         shrink_dcache_parent(mds->mds_fid_de);
 695         dput(mds->mds_fid_de);
 696         LL_DQUOT_OFF(obd->u.obt.obt_sb);
 697
 698         return rc;
 699 }
 700
 701 /* Creates an object with the same name as its fid.  Because this is not at all
 702  * performance sensitive, it is accomplished by creating a file, checking the
 703  * fid, and renaming it. */
 704 int mds_obd_create(struct obd_export *exp, struct obdo *oa,
 705                    struct lov_stripe_md **ea, struct obd_trans_info *oti)
 706 {
 707         struct mds_obd *mds = &exp->exp_obd->u.mds;
 708         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 709         unsigned int tmpname = ll_rand();
 710         struct file *filp;
 711         struct dentry *new_child;
 712         struct lvfs_run_ctxt saved;
 713         char fidname[LL_FID_NAMELEN];
 714         void *handle;
 715         struct lvfs_ucred ucred = { 0 };
 716         int rc = 0, err, namelen;
 717         ENTRY;
 718
 719         /* the owner of object file should always be root */
 720         cap_raise(ucred.luc_cap, CAP_SYS_RESOURCE);
 721
 722         if (strncmp(exp->exp_obd->obd_name, MDD_OBD_NAME,
 723                                    strlen(MDD_OBD_NAME))) {
 724                 RETURN(0);
 725         }
 726
 727         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, &ucred);
 728
 729         sprintf(fidname, "OBJECTS/%u.%u", tmpname, current->pid);
 730         filp = filp_open(fidname, O_CREAT | O_EXCL, 0666);
 731         if (IS_ERR(filp)) {
 732                 rc = PTR_ERR(filp);
 733                 if (rc == -EEXIST) {
 734                         CERROR("impossible object name collision %u\n",
 735                                tmpname);
 736                         LBUG();
 737                 }
 738                 CERROR("error creating tmp object %u: rc %d\n", tmpname, rc);
 739                 GOTO(out_pop, rc);
 740         }
 741
 742         LASSERT(mds->mds_objects_dir == filp->f_dentry->d_parent);
 743
 744         oa->o_id = filp->f_dentry->d_inode->i_ino;
 745         oa->o_generation = filp->f_dentry->d_inode->i_generation;
 746         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 747
 748         LOCK_INODE_MUTEX(parent_inode);
 749         new_child = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 750
 751         if (IS_ERR(new_child)) {
 752                 CERROR("getting neg dentry for obj rename: %d\n", rc);
 753                 GOTO(out_close, rc = PTR_ERR(new_child));
 754         }
 755         if (new_child->d_inode != NULL) {
 756                 CERROR("impossible non-negative obj dentry " LPU64":%u!\n",
 757                        oa->o_id, oa->o_generation);
 758                 LBUG();
 759         }
 760
 761         handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
 762                               FSFILT_OP_RENAME, NULL);
 763         if (IS_ERR(handle))
 764                 GOTO(out_dput, rc = PTR_ERR(handle));
 765
 766         lock_kernel();
 767         rc = vfs_rename(mds->mds_objects_dir->d_inode, filp->f_dentry,
 768                         mds->mds_objects_dir->d_inode, new_child);
 769         unlock_kernel();
 770         if (rc)
 771                 CERROR("error renaming new object "LPU64":%u: rc %d\n",
 772                        oa->o_id, oa->o_generation, rc);
 773
 774         err = fsfilt_commit(exp->exp_obd, mds->mds_objects_dir->d_inode,
 775                             handle, 0);
 776         if (!err) {
 777                 oa->o_gr = FILTER_GROUP_MDS0 + mds->mds_id;
 778                 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLGROUP;
 779         } else if (!rc)
 780                 rc = err;
 781 out_dput:
 782         dput(new_child);
 783 out_close:
 784         UNLOCK_INODE_MUTEX(parent_inode);
 785         err = filp_close(filp, 0);
 786         if (err) {
 787                 CERROR("closing tmpfile %u: rc %d\n", tmpname, rc);
 788                 if (!rc)
 789                         rc = err;
 790         }
 791 out_pop:
 792         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, &ucred);
 793         RETURN(rc);
 794 }
 795
 796 int mds_obd_destroy(struct obd_export *exp, struct obdo *oa,
 797                     struct lov_stripe_md *ea, struct obd_trans_info *oti,
 798                     struct obd_export *md_exp)
 799 {
 800         struct mds_obd *mds = &exp->exp_obd->u.mds;
 801         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 802         struct obd_device *obd = exp->exp_obd;
 803         struct lvfs_run_ctxt saved;
 804         struct lvfs_ucred ucred = { 0 };
 805         char fidname[LL_FID_NAMELEN];
 806         struct inode *inode = NULL;
 807         struct dentry *de;
 808         void *handle;
 809         int err, namelen, rc = 0;
 810         ENTRY;
 811
 812         cap_raise(ucred.luc_cap, CAP_SYS_RESOURCE);
 813         push_ctxt(&saved, &obd->obd_lvfs_ctxt, &ucred);
 814
 815         namelen = ll_fid2str(fidname, oa->o_id, oa->o_generation);
 816
 817         LOCK_INODE_MUTEX(parent_inode);
 818         de = lookup_one_len(fidname, mds->mds_objects_dir, namelen);
 819         if (IS_ERR(de)) {
 820                 rc = IS_ERR(de);
 821                 de = NULL;
 822                 CERROR("error looking up object "LPU64" %s: rc %d\n",
 823                        oa->o_id, fidname, rc);
 824                 GOTO(out_dput, rc);
 825         }
 826         if (de->d_inode == NULL) {
 827                 CERROR("destroying non-existent object "LPU64" %s: rc %d\n",
 828                        oa->o_id, fidname, rc);
 829                 GOTO(out_dput, rc = -ENOENT);
 830         }
 831
 832         /* Stripe count is 1 here since this is some MDS specific stuff
 833            that is unlinked, not spanned across multiple OSTs */
 834         handle = fsfilt_start_log(obd, mds->mds_objects_dir->d_inode,
 835                                   FSFILT_OP_UNLINK, oti, 1);
 836
 837         if (IS_ERR(handle))
 838                 GOTO(out_dput, rc = PTR_ERR(handle));
 839
 840         /* take a reference to protect inode from truncation within
 841            vfs_unlink() context. bug 10409 */
 842         inode = de->d_inode;
 843         atomic_inc(&inode->i_count);
 844         rc = vfs_unlink(mds->mds_objects_dir->d_inode, de);
 845         if (rc)
 846                 CERROR("error destroying object "LPU64":%u: rc %d\n",
 847                        oa->o_id, oa->o_generation, rc);
 848
 849         err = fsfilt_commit(obd, mds->mds_objects_dir->d_inode, handle, 0);
 850         if (err && !rc)
 851                 rc = err;
 852 out_dput:
 853         if (de != NULL)
 854                 l_dput(de);
 855         UNLOCK_INODE_MUTEX(parent_inode);
 856
 857         if (inode)
 858                 iput(inode);
 859
 860         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &ucred);
 861         RETURN(rc);
 862 }