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 Lustre, http://www.lustre.org.
  11  *
  12  *   Lustre is free software; you can redistribute it and/or
  13  *   modify it under the terms of version 2 of the GNU General Public
  14  *   License as published by the Free Software Foundation.
  15  *
  16  *   Lustre is distributed in the hope that it will be useful,
  17  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  *   GNU General Public License for more details.
  20  *
  21  *   You should have received a copy of the GNU General Public License
  22  *   along with Lustre; if not, write to the Free Software
  23  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  24  */
  25
  26 #ifndef EXPORT_SYMTAB
  27 # define EXPORT_SYMTAB
  28 #endif
  29 #define DEBUG_SUBSYSTEM S_MDS
  30
  31 #include <linux/module.h>
  32 #include <linux/kmod.h>
  33 #include <linux/version.h>
  34 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  35 #include <linux/mount.h>
  36 #endif
  37 #include <linux/lustre_mds.h>
  38 #include <linux/obd_class.h>
  39 #include <linux/obd_support.h>
  40 #include <linux/lustre_lib.h>
  41 #include <linux/lustre_fsfilt.h>
  42 #include <libcfs/list.h>
  43
  44 #include <linux/lustre_smfs.h>
  45 #include "mds_internal.h"
  46
  47 /* This limit is arbitrary, but for now we fit it in 1 page (32k clients) */
  48 #define MDS_MAX_CLIENTS (PAGE_SIZE * 8)
  49
  50 #define LAST_RCVD "last_rcvd"
  51 #define LOV_OBJID "lov_objid"
  52 #define LAST_FID  "last_fid"
  53 #define VIRT_FID  "virt_fid"
  54
  55 /* Add client data to the MDS.  We use a bitmap to locate a free space
  56  * in the last_rcvd file if cl_off is -1 (i.e. a new client).
  57  * Otherwise, we have just read the data from the last_rcvd file and
  58  * we know its offset.
  59  */
  60 int mds_client_add(struct obd_device *obd, struct mds_obd *mds,
  61                    struct mds_export_data *med, int cl_idx)
  62 {
  63         unsigned long *bitmap = mds->mds_client_bitmap;
  64         int new_client = (cl_idx == -1);
  65         ENTRY;
  66
  67         LASSERT(bitmap != NULL);
  68
  69         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
  70         if (!strcmp((char *)med->med_mcd->mcd_uuid, (char *)obd->obd_uuid.uuid))
  71                 RETURN(0);
  72
  73         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
  74          * there's no need for extra complication here
  75          */
  76         if (new_client) {
  77                 cl_idx = find_first_zero_bit(bitmap, MDS_MAX_CLIENTS);
  78         repeat:
  79                 if (cl_idx >= MDS_MAX_CLIENTS) {
  80                         CERROR("no room for clients - fix MDS_MAX_CLIENTS\n");
  81                         return -ENOMEM;
  82                 }
  83                 if (test_and_set_bit(cl_idx, bitmap)) {
  84                         cl_idx = find_next_zero_bit(bitmap, MDS_MAX_CLIENTS,
  85                                                     cl_idx);
  86                         goto repeat;
  87                 }
  88         } else {
  89                 if (test_and_set_bit(cl_idx, bitmap)) {
  90                         CERROR("MDS client %d: bit already set in bitmap!!\n",
  91                                cl_idx);
  92                         LBUG();
  93                 }
  94         }
  95
  96         CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
  97                cl_idx, med->med_mcd->mcd_uuid);
  98
  99         med->med_idx = cl_idx;
 100         med->med_off = le32_to_cpu(mds->mds_server_data->msd_client_start) +
 101                 (cl_idx * le16_to_cpu(mds->mds_server_data->msd_client_size));
 102
 103         if (new_client) {
 104                 struct file *file = mds->mds_rcvd_filp;
 105                 struct lvfs_run_ctxt saved;
 106                 loff_t off = med->med_off;
 107                 int rc;
 108
 109                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 110                 rc = fsfilt_write_record(obd, file, med->med_mcd,
 111                                          sizeof(*med->med_mcd), &off, 1);
 112                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 113
 114                 if (rc)
 115                         return rc;
 116                 CDEBUG(D_INFO, "wrote client mcd at idx %u off %llu (len %u)\n",
 117                        med->med_idx, med->med_off,
 118                        (unsigned int)sizeof(*med->med_mcd));
 119         }
 120         return 0;
 121 }
 122
 123 int mds_client_free(struct obd_export *exp, int clear_client)
 124 {
 125         struct mds_export_data *med = &exp->exp_mds_data;
 126         struct mds_obd *mds = &exp->exp_obd->u.mds;
 127         unsigned long *bitmap = mds->mds_client_bitmap;
 128         struct obd_device *obd = exp->exp_obd;
 129         struct mds_client_data zero_mcd;
 130         struct lvfs_run_ctxt saved;
 131         int rc;
 132
 133         if (!med->med_mcd)
 134                 RETURN(0);
 135
 136         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 137         if (!strcmp((char *)med->med_mcd->mcd_uuid, (char *)obd->obd_uuid.uuid))
 138                 GOTO(free_and_out, 0);
 139
 140         CDEBUG(D_INFO, "freeing client at idx %u (%lld)with UUID '%s'\n",
 141                med->med_idx, med->med_off, med->med_mcd->mcd_uuid);
 142
 143         LASSERT(bitmap);
 144
 145         /* Clear the bit _after_ zeroing out the client so we don't
 146            race with mds_client_add and zero out new clients.*/
 147         if (!test_bit(med->med_idx, bitmap)) {
 148                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 149                        med->med_idx);
 150                 LBUG();
 151         }
 152
 153         if (clear_client) {
 154                 memset(&zero_mcd, 0, sizeof zero_mcd);
 155                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 156                 rc = fsfilt_write_record(obd, mds->mds_rcvd_filp, &zero_mcd,
 157                                          sizeof(zero_mcd), &med->med_off, 1);
 158                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 159
 160                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 161                        "zeroing out client %s idx %u in %s rc %d\n",
 162                        med->med_mcd->mcd_uuid, med->med_idx, LAST_RCVD, rc);
 163         }
 164
 165         if (!test_and_clear_bit(med->med_idx, bitmap)) {
 166                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 167                        med->med_idx);
 168                 LBUG();
 169         }
 170
 171
 172         /* Make sure the server's last_transno is up to date. Do this
 173          * after the client is freed so we know all the client's
 174          * transactions have been committed. */
 175         mds_update_server_data(exp->exp_obd, 1);
 176
 177 free_and_out:
 178         OBD_FREE(med->med_mcd, sizeof(*med->med_mcd));
 179         med->med_mcd = NULL;
 180         return 0;
 181 }
 182
 183 static int mds_server_free_data(struct mds_obd *mds)
 184 {
 185         OBD_FREE(mds->mds_client_bitmap, MDS_MAX_CLIENTS / 8);
 186         OBD_FREE(mds->mds_server_data, sizeof(*mds->mds_server_data));
 187         mds->mds_server_data = NULL;
 188
 189         return 0;
 190 }
 191
 192 static int mds_read_last_fid(struct obd_device *obd, struct file *file)
 193 {
 194         int rc = 0;
 195         loff_t off = 0;
 196         struct mds_obd *mds = &obd->u.mds;
 197         unsigned long last_fid_size = file->f_dentry->d_inode->i_size;
 198         ENTRY;
 199
 200         if (last_fid_size == 0) {
 201                 CWARN("%s: initializing new %s\n", obd->obd_name,
 202                       file->f_dentry->d_name.name);
 203
 204                 /*
 205                  * as fid is used for forming res_id for locking, it should not
 206                  * be zero. This will keep us out of lots possible problems,
 207                  * asserts, etc.
 208                  */
 209                 mds_set_last_fid(obd, 0);
 210         } else {
 211                 __u64 lastfid;
 212
 213                 rc = fsfilt_read_record(obd, file, &lastfid,
 214                                         sizeof(lastfid), &off);
 215                 if (rc) {
 216                         CERROR("error reading MDS %s: rc = %d\n",
 217                                file->f_dentry->d_name.name, rc);
 218                         RETURN(rc);
 219                 }
 220
 221                 /*
 222                  * make sure, that fid is up-to-date.
 223                  */
 224                 mds_set_last_fid(obd, lastfid);
 225         }
 226
 227         CDEBUG(D_INODE, "%s: server last_fid: "LPU64"\n",
 228                obd->obd_name, mds->mds_last_fid);
 229
 230         rc = mds_update_last_fid(obd, NULL, 1);
 231         RETURN(rc);
 232 }
 233
 234 static int mds_read_last_rcvd(struct obd_device *obd, struct file *file)
 235 {
 236         unsigned long last_rcvd_size = file->f_dentry->d_inode->i_size;
 237         struct mds_obd *mds = &obd->u.mds;
 238         struct mds_server_data *msd = NULL;
 239         struct mds_client_data *mcd = NULL;
 240         loff_t off = 0;
 241         __u64 mount_count;
 242         int cl_idx, rc = 0;
 243         ENTRY;
 244
 245         /* ensure padding in the struct is the correct size */
 246         LASSERT(offsetof(struct mds_server_data, msd_padding) +
 247                 sizeof(msd->msd_padding) == MDS_LR_SERVER_SIZE);
 248         LASSERT(offsetof(struct mds_client_data, mcd_padding) +
 249                 sizeof(mcd->mcd_padding) == MDS_LR_CLIENT_SIZE);
 250
 251         OBD_ALLOC_WAIT(msd, sizeof(*msd));
 252         if (!msd)
 253                 RETURN(-ENOMEM);
 254
 255         OBD_ALLOC_WAIT(mds->mds_client_bitmap, MDS_MAX_CLIENTS / 8);
 256         if (!mds->mds_client_bitmap) {
 257                 OBD_FREE(msd, sizeof(*msd));
 258                 RETURN(-ENOMEM);
 259         }
 260
 261         mds->mds_server_data = msd;
 262
 263         if (last_rcvd_size == 0) {
 264                 CWARN("%s: initializing new %s\n", obd->obd_name,
 265                       file->f_dentry->d_name.name);
 266
 267                 memcpy(msd->msd_uuid, obd->obd_uuid.uuid,sizeof(msd->msd_uuid));
 268                 msd->msd_last_transno = 0;
 269                 mount_count = msd->msd_mount_count = 0;
 270                 msd->msd_server_size = cpu_to_le32(MDS_LR_SERVER_SIZE);
 271                 msd->msd_client_start = cpu_to_le32(MDS_LR_CLIENT_START);
 272                 msd->msd_client_size = cpu_to_le16(MDS_LR_CLIENT_SIZE);
 273                 msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
 274         } else {
 275                 rc = fsfilt_read_record(obd, file, msd, sizeof(*msd), &off);
 276                 if (rc) {
 277                         CERROR("error reading MDS %s: rc = %d\n",
 278                                file->f_dentry->d_name.name, rc);
 279                         GOTO(err_msd, rc);
 280                 }
 281                 if (strcmp((char *)msd->msd_uuid, (char *)obd->obd_uuid.uuid)) {
 282                         CERROR("OBD UUID %s does not match last_rcvd UUID %s\n",
 283                                obd->obd_uuid.uuid, msd->msd_uuid);
 284                         GOTO(err_msd, rc = -EINVAL);
 285                 }
 286                 mount_count = le64_to_cpu(msd->msd_mount_count);
 287         }
 288         if (msd->msd_feature_incompat & ~cpu_to_le32(MDS_INCOMPAT_SUPP)) {
 289                 CERROR("unsupported incompat feature %x\n",
 290                        le32_to_cpu(msd->msd_feature_incompat) &
 291                        ~MDS_INCOMPAT_SUPP);
 292                 GOTO(err_msd, rc = -EINVAL);
 293         }
 294         /* XXX updating existing b_devel fs only, can be removed in future */
 295         msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
 296         if (msd->msd_feature_rocompat & ~cpu_to_le32(MDS_ROCOMPAT_SUPP)) {
 297                 CERROR("unsupported read-only feature %x\n",
 298                        le32_to_cpu(msd->msd_feature_rocompat) &
 299                        ~MDS_ROCOMPAT_SUPP);
 300                 /* Do something like remount filesystem read-only */
 301                 GOTO(err_msd, rc = -EINVAL);
 302         }
 303
 304         mds->mds_last_transno = le64_to_cpu(msd->msd_last_transno);
 305
 306         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 307                obd->obd_name, mds->mds_last_transno);
 308         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 309                obd->obd_name, mount_count + 1);
 310         CDEBUG(D_INODE, "%s: server data size: %u\n",
 311                obd->obd_name, le32_to_cpu(msd->msd_server_size));
 312         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 313                obd->obd_name, le32_to_cpu(msd->msd_client_start));
 314         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 315                obd->obd_name, le32_to_cpu(msd->msd_client_size));
 316         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 317                obd->obd_name, last_rcvd_size);
 318         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 319                last_rcvd_size <= le32_to_cpu(msd->msd_client_start) ? 0 :
 320                (last_rcvd_size - le32_to_cpu(msd->msd_client_start)) /
 321                 le16_to_cpu(msd->msd_client_size));
 322
 323         /* When we do a clean MDS shutdown, we save the last_transno into
 324          * the header.  If we find clients with higher last_transno values
 325          * then those clients may need recovery done. */
 326         for (cl_idx = 0, off = le32_to_cpu(msd->msd_client_start);
 327              off < last_rcvd_size; cl_idx++) {
 328                 __u64 last_transno;
 329                 struct obd_export *exp;
 330                 struct mds_export_data *med;
 331
 332                 if (!mcd) {
 333                         OBD_ALLOC_WAIT(mcd, sizeof(*mcd));
 334                         if (!mcd)
 335                                 GOTO(err_client, rc = -ENOMEM);
 336                 }
 337
 338                 /* Don't assume off is incremented properly by
 339                  * fsfilt_read_record(), in case sizeof(*mcd)
 340                  * isn't the same as msd->msd_client_size.  */
 341                 off = le32_to_cpu(msd->msd_client_start) +
 342                         cl_idx * le16_to_cpu(msd->msd_client_size);
 343                 rc = fsfilt_read_record(obd, file, mcd, sizeof(*mcd), &off);
 344                 if (rc) {
 345                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 346                                file->f_dentry->d_name.name, cl_idx, off, rc);
 347                         break; /* read error shouldn't cause startup to fail */
 348                 }
 349
 350                 if (mcd->mcd_uuid[0] == '\0') {
 351                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 352                                cl_idx);
 353                         continue;
 354                 }
 355
 356                 last_transno = le64_to_cpu(mcd->mcd_last_transno) >
 357                                le64_to_cpu(mcd->mcd_last_close_transno) ?
 358                                le64_to_cpu(mcd->mcd_last_transno) :
 359                                le64_to_cpu(mcd->mcd_last_close_transno);
 360
 361                 /* These exports are cleaned up by mds_disconnect(), so they
 362                  * need to be set up like real exports as mds_connect() does.
 363                  */
 364                 CDEBUG(D_HA|D_WARNING,"RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 365                        " srv lr: "LPU64" lx: "LPU64"\n", mcd->mcd_uuid, cl_idx,
 366                        last_transno, le64_to_cpu(msd->msd_last_transno),
 367                        mcd->mcd_last_xid);
 368
 369                 exp = class_new_export(obd);
 370                 if (exp == NULL)
 371                         GOTO(err_client, rc = -ENOMEM);
 372
 373                 memcpy(&exp->exp_client_uuid.uuid, mcd->mcd_uuid,
 374                        sizeof exp->exp_client_uuid.uuid);
 375                 med = &exp->exp_mds_data;
 376                 med->med_mcd = mcd;
 377                 mds_client_add(obd, mds, med, cl_idx);
 378                 /* create helper if export init gets more complex */
 379                 INIT_LIST_HEAD(&med->med_open_head);
 380                 spin_lock_init(&med->med_open_lock);
 381
 382                 mcd = NULL;
 383                 exp->exp_connected = 0;
 384                 exp->exp_req_replay_needed = 1;
 385                 obd->obd_recoverable_clients++;
 386                 obd->obd_max_recoverable_clients++;
 387
 388                 /* track clients to separate req replay
 389                  * from lock replay. bug 6063 */
 390                 atomic_inc(&obd->obd_req_replay_clients);
 391                 exp->exp_req_replay_needed = 1;
 392                 atomic_inc(&obd->obd_lock_replay_clients);
 393                 exp->exp_lock_replay_needed = 1;
 394
 395                 class_export_put(exp);
 396
 397                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
 398                        cl_idx, last_transno);
 399
 400                 if (last_transno > mds->mds_last_transno)
 401                        mds->mds_last_transno = last_transno;
 402         }
 403         if (mcd)
 404                 OBD_FREE(mcd, sizeof(*mcd));
 405         obd->obd_last_committed = mds->mds_last_transno;
 406         if (obd->obd_recoverable_clients) {
 407                 CWARN("RECOVERY: service %s, %d recoverable clients, "
 408                       "last_transno "LPU64"\n", obd->obd_name,
 409                       obd->obd_recoverable_clients, mds->mds_last_transno);
 410                 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
 411                 target_start_recovery_thread(obd, mds_handle);
 412                 obd->obd_recovery_start = LTIME_S(CURRENT_TIME);
 413         }
 414
 415         mds->mds_mount_count = mount_count + 1;
 416         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
 417
 418         /* save it, so mount count and last_transno is current */
 419         rc = mds_update_server_data(obd, 1);
 420         if (rc)
 421                 GOTO(err_client, rc);
 422
 423         RETURN(0);
 424
 425 err_client:
 426         class_disconnect_exports(obd, 0);
 427 err_msd:
 428         mds_server_free_data(mds);
 429         RETURN(rc);
 430 }
 431
 432 /*
 433  * sets up root inode lustre_id. It tries to read it first from root inode and
 434  * if it is not there, new rootid is allocated and saved there.
 435  */
 436 int mds_fs_setup_rootid(struct obd_device *obd)
 437 {
 438         int rc = 0;
 439         void *handle;
 440         struct inode *inode;
 441         struct dentry *dentry;
 442         struct mds_obd *mds = &obd->u.mds;
 443         __u64 fid;
 444         ENTRY;
 445
 446         /* getting root directory and setup its fid. */
 447         dentry = mds_id2dentry(obd, &mds->mds_rootid, NULL);
 448         if (IS_ERR(dentry)) {
 449                 CERROR("Can't find ROOT by "DLID4", err = %d\n",
 450                        OLID4(&mds->mds_rootid), (int)PTR_ERR(dentry));
 451                 RETURN(PTR_ERR(dentry));
 452         }
 453
 454         inode = dentry->d_inode;
 455         LASSERT(dentry->d_inode);
 456
 457         rc = mds_pack_inode2id(obd, &mds->mds_rootid, inode, 1);
 458         if (rc < 0) {
 459                 if (rc != -ENODATA)
 460                         GOTO(out_dentry, rc);
 461         } else {
 462                 /*
 463                  * rootid is filled by mds_read_inode_sid(), so we do not need
 464                  * to allocate it and update. The only thing we need to check is
 465                  * mds_num.
 466                  */
 467                 LASSERT(id_group(&mds->mds_rootid) == mds->mds_num);
 468                 mds_set_last_fid(obd, id_fid(&mds->mds_rootid));
 469                 GOTO(out_dentry, rc);
 470         }
 471
 472         /* allocating new one, as it is not found in root inode. */
 473         handle = fsfilt_start(obd, inode,
 474                               FSFILT_OP_SETATTR, NULL);
 475
 476         if (IS_ERR(handle)) {
 477                 rc = PTR_ERR(handle);
 478                 CERROR("fsfilt_start() failed, rc = %d\n", rc);
 479                 GOTO(out_dentry, rc);
 480         }
 481
 482         fid = mds_alloc_fid(obd);
 483         down(&inode->i_sem);
 484         rc = mds_set_inode_sid(obd, inode, handle, &mds->mds_rootid, fid);
 485         up(&inode->i_sem);
 486
 487         if (rc) {
 488                 CERROR("mds_set_inode_sid() failed, rc = %d\n",
 489                        rc);
 490                 GOTO(out_dentry, rc);
 491         }
 492
 493         rc = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
 494         if (rc)
 495                 CERROR("fsfilt_commit() failed, rc = %d\n", rc);
 496
 497         EXIT;
 498 out_dentry:
 499         l_dput(dentry);
 500         if (rc == 0)
 501                 CWARN("%s: rootid: "DLID4"\n", obd->obd_name,
 502                       OLID4(&mds->mds_rootid));
 503         return rc;
 504 }
 505
 506 static int mds_update_virtid_fid(struct obd_device *obd,
 507                                  void *handle, int force_sync)
 508 {
 509         struct mds_obd *mds = &obd->u.mds;
 510         struct file *filp = mds->mds_virtid_filp;
 511         struct lvfs_run_ctxt saved;
 512         loff_t off = 0;
 513         int rc = 0;
 514         ENTRY;
 515
 516         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 517         rc = fsfilt_write_record(obd, filp, &mds->mds_virtid_fid,
 518                                  sizeof(mds->mds_virtid_fid),
 519                                  &off, force_sync);
 520         if (rc) {
 521                 CERROR("error writing MDS virtid_fid #"LPU64
 522                        ", err = %d\n", mds->mds_virtid_fid, rc);
 523         }
 524
 525         CDEBUG(D_SUPER, "wrote virtid fid #"LPU64" at idx "
 526                "%llu: err = %d\n", mds->mds_virtid_fid,
 527                off, rc);
 528         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 529
 530         RETURN(rc);
 531 }
 532
 533 static int mds_read_virtid_fid(struct obd_device *obd,
 534                                struct file *file)
 535 {
 536         int rc = 0;
 537         loff_t off = 0;
 538         struct mds_obd *mds = &obd->u.mds;
 539         unsigned long virtid_fid_size = file->f_dentry->d_inode->i_size;
 540         ENTRY;
 541
 542         if (virtid_fid_size == 0) {
 543                 mds->mds_virtid_fid = mds_alloc_fid(obd);
 544         } else {
 545                 rc = fsfilt_read_record(obd, file, &mds->mds_virtid_fid,
 546                                         sizeof(mds->mds_virtid_fid), &off);
 547                 if (rc) {
 548                         CERROR("error reading MDS %s: rc = %d\n",
 549                                file->f_dentry->d_name.name, rc);
 550                         RETURN(rc);
 551                 }
 552         }
 553         rc = mds_update_virtid_fid(obd, NULL, 1);
 554
 555         RETURN(rc);
 556 }
 557
 558 /*
 559  * initializes lustre_id for virtual id directory, it is needed sometimes, as it
 560  * is possible that it will be the parent for object an operations is going to
 561  * be performed on.
 562  */
 563 int mds_fs_setup_virtid(struct obd_device *obd)
 564 {
 565         int rc = 0;
 566         void *handle;
 567         struct lustre_id sid;
 568         struct mds_obd *mds = &obd->u.mds;
 569         struct inode *inode = mds->mds_id_dir->d_inode;
 570         ENTRY;
 571
 572         handle = fsfilt_start(obd, inode,
 573                               FSFILT_OP_SETATTR, NULL);
 574
 575         if (IS_ERR(handle)) {
 576                 rc = PTR_ERR(handle);
 577                 CERROR("fsfilt_start() failed, rc = %d\n", rc);
 578                 RETURN(rc);
 579         }
 580
 581         id_group(&sid) = mds->mds_num;
 582         id_fid(&sid) = mds->mds_virtid_fid;
 583
 584         id_ino(&sid) = inode->i_ino;
 585         id_gen(&sid) = inode->i_generation;
 586         id_type(&sid) = (S_IFMT & inode->i_mode);
 587
 588         down(&inode->i_sem);
 589         rc = mds_update_inode_sid(obd, inode, handle, &sid);
 590         up(&inode->i_sem);
 591
 592         if (rc) {
 593                 CERROR("mds_update_inode_sid() failed, rc = %d\n",
 594                        rc);
 595                 RETURN(rc);
 596         }
 597
 598         rc = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
 599         if (rc) {
 600                 CERROR("fsfilt_commit() failed, rc = %d\n", rc);
 601                 RETURN(rc);
 602         }
 603
 604         RETURN(rc);
 605 }
 606
 607 int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
 608 {
 609         struct mds_obd *mds = &obd->u.mds;
 610         struct lvfs_run_ctxt saved;
 611         struct dentry *dentry;
 612         struct file *file;
 613         int rc;
 614         ENTRY;
 615
 616         rc = cleanup_group_info();
 617         if (rc)
 618                 RETURN(rc);
 619
 620         mds->mds_vfsmnt = mnt;
 621         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 622
 623         fsfilt_setup(obd, mds->mds_sb);
 624
 625         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 626         obd->obd_lvfs_ctxt.pwdmnt = mnt;
 627         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
 628         obd->obd_lvfs_ctxt.fs = get_ds();
 629         obd->obd_lvfs_ctxt.cb_ops = mds_lvfs_ops;
 630
 631         /* setup the directory tree */
 632         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 633         dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755, 0);
 634         if (IS_ERR(dentry)) {
 635                 rc = PTR_ERR(dentry);
 636                 CERROR("cannot create ROOT directory: rc = %d\n", rc);
 637                 GOTO(err_pop, rc);
 638         }
 639
 640         mdc_pack_id(&mds->mds_rootid, dentry->d_inode->i_ino,
 641                     dentry->d_inode->i_generation, S_IFDIR, 0, 0);
 642
 643         dput(dentry);
 644
 645         dentry = lookup_one_len("__iopen__", current->fs->pwd,
 646                                 strlen("__iopen__"));
 647         if (IS_ERR(dentry)) {
 648                 rc = PTR_ERR(dentry);
 649                 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
 650                 GOTO(err_pop, rc);
 651         }
 652         mds->mds_id_de = dentry;
 653         if (!dentry->d_inode || is_bad_inode(dentry->d_inode)) {
 654                 rc = -ENOENT;
 655                 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
 656                 GOTO(err_id_de, rc);
 657         }
 658
 659         dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777, 1);
 660         if (IS_ERR(dentry)) {
 661                 rc = PTR_ERR(dentry);
 662                 CERROR("cannot create PENDING directory: rc = %d\n", rc);
 663                 GOTO(err_id_de, rc);
 664         }
 665         mds->mds_pending_dir = dentry;
 666
 667         dentry = simple_mkdir(current->fs->pwd, "LOGS", 0777, 1);
 668         if (IS_ERR(dentry)) {
 669                 rc = PTR_ERR(dentry);
 670                 CERROR("cannot create LOGS directory: rc = %d\n", rc);
 671                 GOTO(err_pending, rc);
 672         }
 673         mds->mds_logs_dir = dentry;
 674
 675         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
 676         if (IS_ERR(dentry)) {
 677                 rc = PTR_ERR(dentry);
 678                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
 679                 GOTO(err_logs, rc);
 680         }
 681         mds->mds_objects_dir = dentry;
 682
 683         dentry = simple_mkdir(current->fs->pwd, "FIDS", 0777, 1);
 684         if (IS_ERR(dentry)) {
 685                 rc = PTR_ERR(dentry);
 686                 CERROR("cannot create FIDS directory: rc = %d\n", rc);
 687                 GOTO(err_objects, rc);
 688         }
 689         mds->mds_id_dir = dentry;
 690
 691         dentry = simple_mkdir(current->fs->pwd, "UNNAMED", 0777, 1);
 692         if (IS_ERR(dentry)) {
 693                 rc = PTR_ERR(dentry);
 694                 CERROR("cannot create UNNAMED directory: rc = %d\n", rc);
 695                 GOTO(err_unnamed, rc);
 696         }
 697         mds->mds_unnamed_dir = dentry;
 698
 699         /* open and test the last rcvd file */
 700         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
 701         if (IS_ERR(file)) {
 702                 rc = PTR_ERR(file);
 703                 CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
 704                 GOTO(err_id_dir, rc = PTR_ERR(file));
 705         }
 706         mds->mds_rcvd_filp = file;
 707
 708         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 709                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 710                        file->f_dentry->d_inode->i_mode);
 711                 GOTO(err_last_rcvd, rc = -ENOENT);
 712         }
 713
 714         rc = mds_read_last_rcvd(obd, file);
 715         if (rc) {
 716                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
 717                 GOTO(err_last_rcvd, rc);
 718         }
 719
 720         /* open and test last fid file */
 721         file = filp_open(LAST_FID, O_RDWR | O_CREAT, 0644);
 722         if (IS_ERR(file)) {
 723                 rc = PTR_ERR(file);
 724                 CERROR("cannot open/create %s file: rc = %d\n",
 725                        LAST_FID, rc);
 726                 GOTO(err_client, rc = PTR_ERR(file));
 727         }
 728         mds->mds_fid_filp = file;
 729         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 730                 CERROR("%s is not a regular file!: mode = %o\n",
 731                        LAST_FID, file->f_dentry->d_inode->i_mode);
 732                 GOTO(err_last_fid, rc = -ENOENT);
 733         }
 734
 735         rc = mds_read_last_fid(obd, file);
 736         if (rc) {
 737                 CERROR("cannot read %s: rc = %d\n", LAST_FID, rc);
 738                 GOTO(err_last_fid, rc);
 739         }
 740
 741         /* open and test virtid fid file */
 742         file = filp_open(VIRT_FID, O_RDWR | O_CREAT, 0644);
 743         if (IS_ERR(file)) {
 744                 rc = PTR_ERR(file);
 745                 CERROR("cannot open/create %s file: rc = %d\n",
 746                        VIRT_FID, rc);
 747                 GOTO(err_last_fid, rc = PTR_ERR(file));
 748         }
 749         mds->mds_virtid_filp = file;
 750         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 751                 CERROR("%s is not a regular file!: mode = %o\n",
 752                        VIRT_FID, file->f_dentry->d_inode->i_mode);
 753                 GOTO(err_virtid_fid, rc = -ENOENT);
 754         }
 755
 756         rc = mds_read_virtid_fid(obd, file);
 757         if (rc) {
 758                 CERROR("cannot read %s: rc = %d\n", VIRT_FID, rc);
 759                 GOTO(err_virtid_fid, rc);
 760         }
 761
 762         /* open and test the lov objid file */
 763         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
 764         if (IS_ERR(file)) {
 765                 rc = PTR_ERR(file);
 766                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
 767                 GOTO(err_last_fid, rc = PTR_ERR(file));
 768         }
 769         mds->mds_dt_objid_filp = file;
 770         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 771                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
 772                        file->f_dentry->d_inode->i_mode);
 773                 GOTO(err_lov_objid, rc = -ENOENT);
 774         }
 775 err_pop:
 776         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 777         return rc;
 778
 779 err_lov_objid:
 780         if (mds->mds_dt_objid_filp && filp_close(mds->mds_dt_objid_filp, 0))
 781                 CERROR("can't close %s after error\n", LOV_OBJID);
 782 err_virtid_fid:
 783         if (mds->mds_virtid_filp && filp_close(mds->mds_virtid_filp, 0))
 784                 CERROR("can't close %s after error\n", VIRT_FID);
 785 err_last_fid:
 786         if (mds->mds_fid_filp && filp_close(mds->mds_fid_filp, 0))
 787                 CERROR("can't close %s after error\n", LAST_FID);
 788 err_client:
 789         class_disconnect_exports(obd, 0);
 790 err_last_rcvd:
 791         if (mds->mds_rcvd_filp && filp_close(mds->mds_rcvd_filp, 0))
 792                 CERROR("can't close %s after error\n", LAST_RCVD);
 793 err_unnamed:
 794         dput(mds->mds_unnamed_dir);
 795 err_id_dir:
 796         dput(mds->mds_id_dir);
 797 err_objects:
 798         dput(mds->mds_objects_dir);
 799 err_logs:
 800         dput(mds->mds_logs_dir);
 801 err_pending:
 802         dput(mds->mds_pending_dir);
 803 err_id_de:
 804         dput(mds->mds_id_de);
 805         goto err_pop;
 806 }
 807
 808 static int  mds_fs_post_cleanup(struct obd_device *obd)
 809 {
 810         int    rc = 0;
 811         rc = fsfilt_post_cleanup(obd);
 812         return rc;
 813 }
 814
 815 int mds_fs_cleanup(struct obd_device *obd, int flags)
 816 {
 817         struct mds_obd *mds = &obd->u.mds;
 818         struct lvfs_run_ctxt saved;
 819         int rc = 0;
 820
 821         if (flags & OBD_OPT_FAILOVER)
 822                 CERROR("%s: shutting down for failover; client state will"
 823                        " be preserved.\n", obd->obd_name);
 824
 825         class_disconnect_exports(obd, flags); /* cleans up client info too */
 826         target_cleanup_recovery(obd);
 827         mds_server_free_data(mds);
 828
 829         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 830         if (mds->mds_virtid_filp) {
 831                 rc = filp_close(mds->mds_virtid_filp, 0);
 832                 mds->mds_virtid_filp = NULL;
 833                 if (rc)
 834                         CERROR("%s file won't close, rc = %d\n", VIRT_FID, rc);
 835         }
 836         if (mds->mds_fid_filp) {
 837                 rc = filp_close(mds->mds_fid_filp, 0);
 838                 mds->mds_fid_filp = NULL;
 839                 if (rc)
 840                         CERROR("%s file won't close, rc = %d\n", LAST_FID, rc);
 841         }
 842         if (mds->mds_rcvd_filp) {
 843                 rc = filp_close(mds->mds_rcvd_filp, 0);
 844                 mds->mds_rcvd_filp = NULL;
 845                 if (rc)
 846                         CERROR("%s file won't close, rc = %d\n", LAST_RCVD, rc);
 847         }
 848         if (mds->mds_dt_objid_filp) {
 849                 rc = filp_close(mds->mds_dt_objid_filp, 0);
 850                 mds->mds_dt_objid_filp = NULL;
 851                 if (rc)
 852                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
 853         }
 854         if (mds->mds_unnamed_dir != NULL) {
 855                 l_dput(mds->mds_unnamed_dir);
 856                 mds->mds_unnamed_dir = NULL;
 857         }
 858         if (mds->mds_id_dir != NULL) {
 859                 l_dput(mds->mds_id_dir);
 860                 mds->mds_id_dir = NULL;
 861         }
 862         if (mds->mds_objects_dir != NULL) {
 863                 l_dput(mds->mds_objects_dir);
 864                 mds->mds_objects_dir = NULL;
 865         }
 866         if (mds->mds_logs_dir) {
 867                 l_dput(mds->mds_logs_dir);
 868                 mds->mds_logs_dir = NULL;
 869         }
 870         if (mds->mds_pending_dir) {
 871                 l_dput(mds->mds_pending_dir);
 872                 mds->mds_pending_dir = NULL;
 873         }
 874         rc = mds_fs_post_cleanup(obd);
 875
 876         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 877         shrink_dcache_parent(mds->mds_id_de);
 878         dput(mds->mds_id_de);
 879
 880         return rc;
 881 }
 882
 883 /* Creates an object with the same name as its id.  Because this is not at all
 884  * performance sensitive, it is accomplished by creating a file, checking the
 885  * id, and renaming it. */
 886 int mds_obd_create(struct obd_export *exp, struct obdo *oa,
 887                    void *acl, int acl_size,
 888                    struct lov_stripe_md **ea, struct obd_trans_info *oti)
 889 {
 890         struct mds_obd *mds = &exp->exp_obd->u.mds;
 891         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 892         struct file *filp;
 893         struct dentry *dchild;
 894         struct lvfs_run_ctxt saved;
 895         char idname[LL_ID_NAMELEN];
 896         int rc = 0, err, idlen;
 897         void *handle;
 898         ENTRY;
 899
 900         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 901         down(&parent_inode->i_sem);
 902         if (oa->o_id) {
 903                 idlen = ll_id2str(idname, oa->o_id, oa->o_generation);
 904                 dchild = lookup_one_len(idname, mds->mds_objects_dir, idlen);
 905                 if (IS_ERR(dchild))
 906                         GOTO(out_pop, rc = PTR_ERR(dchild));
 907
 908                 if (dchild->d_inode == NULL) {
 909                         struct dentry_params dp;
 910                         struct inode *inode;
 911
 912                         CWARN("creating log with ID "LPU64"\n", oa->o_id);
 913
 914                         dchild->d_fsdata = (void *) &dp;
 915                         dp.p_ptr = NULL;
 916                         dp.p_inum = oa->o_id;
 917                         rc = ll_vfs_create(parent_inode, dchild, S_IFREG, NULL);
 918                         if (dchild->d_fsdata == (void *)(unsigned long)oa->o_id)
 919                                 dchild->d_fsdata = NULL;
 920                         if (rc) {
 921                                 CDEBUG(D_INODE, "err during create: %d\n", rc);
 922                                 dput(dchild);
 923                                 GOTO(out_pop, rc);
 924                         }
 925                         inode = dchild->d_inode;
 926                         LASSERT(inode->i_ino == oa->o_id);
 927                         inode->i_generation = oa->o_generation;
 928                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
 929                                inode->i_ino, inode->i_generation);
 930                         mark_inode_dirty(inode);
 931                 } else {
 932                         CWARN("it should be here!\n");
 933                 }
 934                 GOTO(out_pop, rc);
 935         }
 936
 937         sprintf(idname, "OBJECTS/%u.%u", ll_insecure_random_int(), current->pid);
 938         filp = filp_open(idname, O_CREAT | O_EXCL, 0644);
 939         if (IS_ERR(filp)) {
 940                 rc = PTR_ERR(filp);
 941                 if (rc == -EEXIST) {
 942                         CERROR("impossible object name collision %s\n",
 943                                idname);
 944                         LBUG();
 945                 }
 946                 CERROR("error creating tmp object %s: rc %d\n",
 947                        idname, rc);
 948                 GOTO(out_pop, rc);
 949         }
 950
 951         LASSERT(mds->mds_objects_dir == filp->f_dentry->d_parent);
 952
 953         oa->o_id = filp->f_dentry->d_inode->i_ino;
 954         oa->o_generation = filp->f_dentry->d_inode->i_generation;
 955         idlen = ll_id2str(idname, oa->o_id, oa->o_generation);
 956
 957         CWARN("created log anonymous "LPU64"/%u\n",
 958               oa->o_id, oa->o_generation);
 959
 960         dchild = lookup_one_len(idname, mds->mds_objects_dir, idlen);
 961         if (IS_ERR(dchild)) {
 962                 CERROR("getting neg dentry for obj rename: %d\n", rc);
 963                 GOTO(out_close, rc = PTR_ERR(dchild));
 964         }
 965         if (dchild->d_inode != NULL) {
 966                 CERROR("impossible non-negative obj dentry " LPU64":%u!\n",
 967                        oa->o_id, oa->o_generation);
 968                 LBUG();
 969         }
 970
 971         handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
 972                               FSFILT_OP_RENAME, NULL);
 973         if (IS_ERR(handle))
 974                 GOTO(out_dput, rc = PTR_ERR(handle));
 975
 976         lock_kernel();
 977         rc = vfs_rename(mds->mds_objects_dir->d_inode, filp->f_dentry,
 978                         mds->mds_objects_dir->d_inode, dchild);
 979         unlock_kernel();
 980         if (rc)
 981                 CERROR("error renaming new object "LPU64":%u: rc %d\n",
 982                        oa->o_id, oa->o_generation, rc);
 983
 984         err = fsfilt_commit(exp->exp_obd, mds->mds_sb,
 985                             mds->mds_objects_dir->d_inode, handle, 0);
 986         if (!err) {
 987                 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
 988                 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLGROUP;
 989         } else if (!rc)
 990                 rc = err;
 991 out_dput:
 992         dput(dchild);
 993 out_close:
 994         err = filp_close(filp, 0);
 995         if (err) {
 996                 CERROR("closing tmpfile %s: rc %d\n", idname, rc);
 997                 if (!rc)
 998                         rc = err;
 999         }
1000 out_pop:
1001         up(&parent_inode->i_sem);
1002         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
1003         RETURN(rc);
1004 }
1005
1006 int mds_obd_destroy(struct obd_export *exp, struct obdo *oa,
1007                     struct lov_stripe_md *ea, struct obd_trans_info *oti)
1008 {
1009         struct mds_obd *mds = &exp->exp_obd->u.mds;
1010         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
1011         struct obd_device *obd = exp->exp_obd;
1012         struct lvfs_run_ctxt saved;
1013         char idname[LL_ID_NAMELEN];
1014         struct dentry *de;
1015         void *handle;
1016         int err, idlen, rc = 0;
1017         ENTRY;
1018
1019         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1020
1021         idlen = ll_id2str(idname, oa->o_id, oa->o_generation);
1022
1023         down(&parent_inode->i_sem);
1024         de = lookup_one_len(idname, mds->mds_objects_dir, idlen);
1025         if (IS_ERR(de) || de->d_inode == NULL) {
1026                 rc = IS_ERR(de) ? PTR_ERR(de) : -ENOENT;
1027                 CERROR("destroying non-existent object "LPU64" %s: rc %d\n",
1028                        oa->o_id, idname, rc);
1029                 GOTO(out_dput, rc);
1030         }
1031         /* Stripe count is 1 here since this is some MDS specific stuff
1032            that is unlinked, not spanned across multiple OSTs */
1033         handle = fsfilt_start_log(obd, mds->mds_objects_dir->d_inode,
1034                                   FSFILT_OP_UNLINK, oti, 1);
1035
1036         if (IS_ERR(handle))
1037                 GOTO(out_dput, rc = PTR_ERR(handle));
1038
1039         rc = vfs_unlink(mds->mds_objects_dir->d_inode, de);
1040         if (rc)
1041                 CERROR("error destroying object "LPU64":%u: rc %d\n",
1042                        oa->o_id, oa->o_generation, rc);
1043
1044         err = fsfilt_commit(obd, mds->mds_sb, mds->mds_objects_dir->d_inode,
1045                             handle, exp->exp_sync);
1046         if (err && !rc)
1047                 rc = err;
1048 out_dput:
1049         if (de != NULL)
1050                 l_dput(de);
1051         up(&parent_inode->i_sem);
1052         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1053         RETURN(rc);
1054 }