lustre/mds/mds_fs.c

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