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(med->med_mcd->mcd_uuid, 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         mds_idmap_cleanup(med);
 134
 135         if (!med->med_mcd)
 136                 RETURN(0);
 137
 138         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 139         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
 140                 GOTO(free_and_out, 0);
 141
 142         CDEBUG(D_INFO, "freeing client at idx %u (%lld)with UUID '%s'\n",
 143                med->med_idx, med->med_off, med->med_mcd->mcd_uuid);
 144
 145         LASSERT(bitmap);
 146
 147         /* Clear the bit _after_ zeroing out the client so we don't
 148            race with mds_client_add and zero out new clients.*/
 149         if (!test_bit(med->med_idx, bitmap)) {
 150                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 151                        med->med_idx);
 152                 LBUG();
 153         }
 154
 155         if (clear_client) {
 156                 memset(&zero_mcd, 0, sizeof zero_mcd);
 157                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 158                 rc = fsfilt_write_record(obd, mds->mds_rcvd_filp, &zero_mcd,
 159                                          sizeof(zero_mcd), &med->med_off, 1);
 160                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 161
 162                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 163                        "zeroing out client %s idx %u in %s rc %d\n",
 164                        med->med_mcd->mcd_uuid, med->med_idx, LAST_RCVD, rc);
 165         }
 166
 167         if (!test_and_clear_bit(med->med_idx, bitmap)) {
 168                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 169                        med->med_idx);
 170                 LBUG();
 171         }
 172
 173
 174         /* Make sure the server's last_transno is up to date. Do this
 175          * after the client is freed so we know all the client's
 176          * transactions have been committed. */
 177         mds_update_server_data(exp->exp_obd, 1);
 178
 179 free_and_out:
 180         OBD_FREE(med->med_mcd, sizeof(*med->med_mcd));
 181         med->med_mcd = NULL;
 182         return 0;
 183 }
 184
 185 static int mds_server_free_data(struct mds_obd *mds)
 186 {
 187         OBD_FREE(mds->mds_client_bitmap, MDS_MAX_CLIENTS / 8);
 188         OBD_FREE(mds->mds_server_data, sizeof(*mds->mds_server_data));
 189         mds->mds_server_data = NULL;
 190
 191         return 0;
 192 }
 193
 194 static int mds_read_last_fid(struct obd_device *obd, struct file *file)
 195 {
 196         int rc = 0;
 197         loff_t off = 0;
 198         struct mds_obd *mds = &obd->u.mds;
 199         unsigned long last_fid_size = file->f_dentry->d_inode->i_size;
 200         ENTRY;
 201
 202         if (last_fid_size == 0) {
 203                 CWARN("%s: initializing new %s\n", obd->obd_name,
 204                       file->f_dentry->d_name.name);
 205
 206                 /*
 207                  * as fid is used for forming res_id for locking, it should not
 208                  * be zero. This will keep us out of lots possible problems,
 209                  * asserts, etc.
 210                  */
 211                 mds_set_last_fid(obd, 0);
 212         } else {
 213                 __u64 lastfid;
 214
 215                 rc = fsfilt_read_record(obd, file, &lastfid,
 216                                         sizeof(lastfid), &off);
 217                 if (rc) {
 218                         CERROR("error reading MDS %s: rc = %d\n",
 219                                file->f_dentry->d_name.name, rc);
 220                         RETURN(rc);
 221                 }
 222
 223                 /*
 224                  * make sure, that fid is up-to-date.
 225                  */
 226                 mds_set_last_fid(obd, lastfid);
 227         }
 228
 229         CDEBUG(D_INODE, "%s: server last_fid: "LPU64"\n",
 230                obd->obd_name, mds->mds_last_fid);
 231
 232         rc = mds_update_last_fid(obd, NULL, 1);
 233         RETURN(rc);
 234 }
 235
 236 static int mds_read_last_rcvd(struct obd_device *obd, struct file *file)
 237 {
 238         unsigned long last_rcvd_size = file->f_dentry->d_inode->i_size;
 239         struct mds_obd *mds = &obd->u.mds;
 240         struct mds_server_data *msd = NULL;
 241         struct mds_client_data *mcd = NULL;
 242         loff_t off = 0;
 243         __u64 mount_count;
 244         int cl_idx, rc = 0;
 245         ENTRY;
 246
 247         /* ensure padding in the struct is the correct size */
 248         LASSERT(offsetof(struct mds_server_data, msd_padding) +
 249                 sizeof(msd->msd_padding) == MDS_LR_SERVER_SIZE);
 250         LASSERT(offsetof(struct mds_client_data, mcd_padding) +
 251                 sizeof(mcd->mcd_padding) == MDS_LR_CLIENT_SIZE);
 252
 253         OBD_ALLOC_WAIT(msd, sizeof(*msd));
 254         if (!msd)
 255                 RETURN(-ENOMEM);
 256
 257         OBD_ALLOC_WAIT(mds->mds_client_bitmap, MDS_MAX_CLIENTS / 8);
 258         if (!mds->mds_client_bitmap) {
 259                 OBD_FREE(msd, sizeof(*msd));
 260                 RETURN(-ENOMEM);
 261         }
 262
 263         mds->mds_server_data = msd;
 264
 265         if (last_rcvd_size == 0) {
 266                 CWARN("%s: initializing new %s\n", obd->obd_name,
 267                       file->f_dentry->d_name.name);
 268
 269                 memcpy(msd->msd_uuid, obd->obd_uuid.uuid,sizeof(msd->msd_uuid));
 270                 msd->msd_last_transno = 0;
 271                 mount_count = msd->msd_mount_count = 0;
 272                 msd->msd_server_size = cpu_to_le32(MDS_LR_SERVER_SIZE);
 273                 msd->msd_client_start = cpu_to_le32(MDS_LR_CLIENT_START);
 274                 msd->msd_client_size = cpu_to_le16(MDS_LR_CLIENT_SIZE);
 275                 msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
 276         } else {
 277                 rc = fsfilt_read_record(obd, file, msd, sizeof(*msd), &off);
 278                 if (rc) {
 279                         CERROR("error reading MDS %s: rc = %d\n",
 280                                file->f_dentry->d_name.name, rc);
 281                         GOTO(err_msd, rc);
 282                 }
 283                 if (strcmp(msd->msd_uuid, obd->obd_uuid.uuid) != 0) {
 284                         CERROR("OBD UUID %s does not match last_rcvd UUID %s\n",
 285                                obd->obd_uuid.uuid, msd->msd_uuid);
 286                         GOTO(err_msd, rc = -EINVAL);
 287                 }
 288                 mount_count = le64_to_cpu(msd->msd_mount_count);
 289         }
 290         if (msd->msd_feature_incompat & ~cpu_to_le32(MDS_INCOMPAT_SUPP)) {
 291                 CERROR("unsupported incompat feature %x\n",
 292                        le32_to_cpu(msd->msd_feature_incompat) &
 293                        ~MDS_INCOMPAT_SUPP);
 294                 GOTO(err_msd, rc = -EINVAL);
 295         }
 296         /* XXX updating existing b_devel fs only, can be removed in future */
 297         msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
 298         if (msd->msd_feature_rocompat & ~cpu_to_le32(MDS_ROCOMPAT_SUPP)) {
 299                 CERROR("unsupported read-only feature %x\n",
 300                        le32_to_cpu(msd->msd_feature_rocompat) &
 301                        ~MDS_ROCOMPAT_SUPP);
 302                 /* Do something like remount filesystem read-only */
 303                 GOTO(err_msd, rc = -EINVAL);
 304         }
 305
 306         mds->mds_last_transno = le64_to_cpu(msd->msd_last_transno);
 307
 308         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 309                obd->obd_name, mds->mds_last_transno);
 310         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 311                obd->obd_name, mount_count + 1);
 312         CDEBUG(D_INODE, "%s: server data size: %u\n",
 313                obd->obd_name, le32_to_cpu(msd->msd_server_size));
 314         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 315                obd->obd_name, le32_to_cpu(msd->msd_client_start));
 316         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 317                obd->obd_name, le32_to_cpu(msd->msd_client_size));
 318         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 319                obd->obd_name, last_rcvd_size);
 320         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 321                last_rcvd_size <= le32_to_cpu(msd->msd_client_start) ? 0 :
 322                (last_rcvd_size - le32_to_cpu(msd->msd_client_start)) /
 323                 le16_to_cpu(msd->msd_client_size));
 324
 325         /* When we do a clean MDS shutdown, we save the last_transno into
 326          * the header.  If we find clients with higher last_transno values
 327          * then those clients may need recovery done. */
 328         for (cl_idx = 0, off = le32_to_cpu(msd->msd_client_start);
 329              off < last_rcvd_size; cl_idx++) {
 330                 __u64 last_transno;
 331                 struct obd_export *exp;
 332                 struct mds_export_data *med;
 333
 334                 if (!mcd) {
 335                         OBD_ALLOC_WAIT(mcd, sizeof(*mcd));
 336                         if (!mcd)
 337                                 GOTO(err_client, rc = -ENOMEM);
 338                 }
 339
 340                 /* Don't assume off is incremented properly by
 341                  * fsfilt_read_record(), in case sizeof(*mcd)
 342                  * isn't the same as msd->msd_client_size.  */
 343                 off = le32_to_cpu(msd->msd_client_start) +
 344                         cl_idx * le16_to_cpu(msd->msd_client_size);
 345                 rc = fsfilt_read_record(obd, file, mcd, sizeof(*mcd), &off);
 346                 if (rc) {
 347                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 348                                file->f_dentry->d_name.name, cl_idx, off, rc);
 349                         break; /* read error shouldn't cause startup to fail */
 350                 }
 351
 352                 if (mcd->mcd_uuid[0] == '\0') {
 353                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 354                                cl_idx);
 355                         continue;
 356                 }
 357
 358                 last_transno = le64_to_cpu(mcd->mcd_last_transno) >
 359                                le64_to_cpu(mcd->mcd_last_close_transno) ?
 360                                le64_to_cpu(mcd->mcd_last_transno) :
 361                                le64_to_cpu(mcd->mcd_last_close_transno);
 362
 363                 /* These exports are cleaned up by mds_disconnect(), so they
 364                  * need to be set up like real exports as mds_connect() does.
 365                  */
 366                 CDEBUG(D_HA|D_WARNING,"RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 367                        " srv lr: "LPU64" lx: "LPU64"\n", mcd->mcd_uuid, cl_idx,
 368                        last_transno, le64_to_cpu(msd->msd_last_transno),
 369                        mcd->mcd_last_xid);
 370
 371                 exp = class_new_export(obd);
 372                 if (exp == NULL)
 373                         GOTO(err_client, rc = -ENOMEM);
 374
 375                 memcpy(&exp->exp_client_uuid.uuid, mcd->mcd_uuid,
 376                        sizeof exp->exp_client_uuid.uuid);
 377                 med = &exp->exp_mds_data;
 378                 med->med_mcd = mcd;
 379                 mds_client_add(obd, mds, med, cl_idx);
 380                 /* create helper if export init gets more complex */
 381                 INIT_LIST_HEAD(&med->med_open_head);
 382                 spin_lock_init(&med->med_open_lock);
 383
 384                 mcd = NULL;
 385                 exp->exp_replay_needed = 1;
 386                 obd->obd_recoverable_clients++;
 387                 obd->obd_max_recoverable_clients++;
 388                 class_export_put(exp);
 389
 390                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
 391                        cl_idx, last_transno);
 392
 393                 if (last_transno > mds->mds_last_transno)
 394                        mds->mds_last_transno = last_transno;
 395         }
 396         if (mcd)
 397                 OBD_FREE(mcd, sizeof(*mcd));
 398         obd->obd_last_committed = mds->mds_last_transno;
 399         if (obd->obd_recoverable_clients) {
 400                 CWARN("RECOVERY: service %s, %d recoverable clients, "
 401                       "last_transno "LPU64"\n", obd->obd_name,
 402                       obd->obd_recoverable_clients, mds->mds_last_transno);
 403                 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
 404                 target_start_recovery_thread(obd, mds_handle);
 405                 obd->obd_recovery_start = LTIME_S(CURRENT_TIME);
 406         }
 407
 408         mds->mds_mount_count = mount_count + 1;
 409         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
 410
 411         /* save it, so mount count and last_transno is current */
 412         rc = mds_update_server_data(obd, 1);
 413         if (rc)
 414                 GOTO(err_client, rc);
 415
 416         RETURN(0);
 417
 418 err_client:
 419         class_disconnect_exports(obd, 0);
 420 err_msd:
 421         mds_server_free_data(mds);
 422         RETURN(rc);
 423 }
 424
 425 static int mds_fs_post_setup(struct obd_device *obd)
 426 {
 427         struct mds_obd *mds = &obd->u.mds;
 428         struct dentry *dentry;
 429         int rc = 0;
 430         ENTRY;
 431
 432         dentry = mds_id2dentry(obd, &mds->mds_rootid, NULL);
 433         if (IS_ERR(dentry)) {
 434                 CERROR("Can't find ROOT, err = %d\n",
 435                        (int)PTR_ERR(dentry));
 436                 RETURN(PTR_ERR(dentry));
 437         }
 438
 439         rc = fsfilt_post_setup(obd, dentry);
 440         if (rc)
 441                 goto out_dentry;
 442
 443         LASSERT(dentry->d_inode != NULL);
 444
 445         fsfilt_set_fs_flags(obd, dentry->d_inode,
 446                             SM_DO_REC | SM_DO_COW);
 447
 448         fsfilt_set_fs_flags(obd, mds->mds_pending_dir->d_inode,
 449                             SM_DO_REC | SM_DO_COW);
 450
 451         fsfilt_set_mds_flags(obd, mds->mds_sb);
 452
 453 out_dentry:
 454         l_dput(dentry);
 455         RETURN(rc);
 456 }
 457
 458 /*
 459  * sets up root inode lustre_id. It tries to read it first from root inode and
 460  * if it is not there, new rootid is allocated and saved there.
 461  */
 462 int mds_fs_setup_rootid(struct obd_device *obd)
 463 {
 464         int rc = 0;
 465         void *handle;
 466         struct inode *inode;
 467         struct dentry *dentry;
 468         struct mds_obd *mds = &obd->u.mds;
 469         ENTRY;
 470
 471         /* getting root directory and setup its fid. */
 472         dentry = mds_id2dentry(obd, &mds->mds_rootid, NULL);
 473         if (IS_ERR(dentry)) {
 474                 CERROR("Can't find ROOT by "DLID4", err = %d\n",
 475                        OLID4(&mds->mds_rootid), (int)PTR_ERR(dentry));
 476                 RETURN(PTR_ERR(dentry));
 477         }
 478
 479         inode = dentry->d_inode;
 480         LASSERT(dentry->d_inode);
 481
 482         rc = mds_pack_inode2id(obd, &mds->mds_rootid, inode, 1);
 483         if (rc < 0) {
 484                 if (rc != -ENODATA)
 485                         GOTO(out_dentry, rc);
 486         } else {
 487                 /*
 488                  * rootid is filled by mds_read_inode_sid(), so we do not need
 489                  * to allocate it and update. The only thing we need to check is
 490                  * mds_num.
 491                  */
 492                 LASSERT(id_group(&mds->mds_rootid) == mds->mds_num);
 493                 mds_set_last_fid(obd, id_fid(&mds->mds_rootid));
 494                 GOTO(out_dentry, rc);
 495         }
 496
 497         /* allocating new one, as it is not found in root inode. */
 498         handle = fsfilt_start(obd, inode,
 499                               FSFILT_OP_SETATTR, NULL);
 500
 501         if (IS_ERR(handle)) {
 502                 rc = PTR_ERR(handle);
 503                 CERROR("fsfilt_start() failed, rc = %d\n", rc);
 504                 GOTO(out_dentry, rc);
 505         }
 506
 507         down(&inode->i_sem);
 508         rc = mds_alloc_inode_sid(obd, inode, handle, &mds->mds_rootid);
 509         up(&inode->i_sem);
 510
 511         if (rc) {
 512                 CERROR("mds_alloc_inode_sid() failed, rc = %d\n",
 513                        rc);
 514                 GOTO(out_dentry, rc);
 515         }
 516
 517         rc = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
 518         if (rc)
 519                 CERROR("fsfilt_commit() failed, rc = %d\n", rc);
 520
 521         EXIT;
 522 out_dentry:
 523         l_dput(dentry);
 524         if (rc == 0)
 525                 CWARN("%s: rootid: "DLID4"\n", obd->obd_name,
 526                       OLID4(&mds->mds_rootid));
 527         return rc;
 528 }
 529
 530 static int mds_update_virtid_fid(struct obd_device *obd,
 531                                  void *handle, int force_sync)
 532 {
 533         struct mds_obd *mds = &obd->u.mds;
 534         struct file *filp = mds->mds_virtid_filp;
 535         struct lvfs_run_ctxt saved;
 536         loff_t off = 0;
 537         int rc = 0;
 538         ENTRY;
 539
 540         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 541         rc = fsfilt_write_record(obd, filp, &mds->mds_virtid_fid,
 542                                  sizeof(mds->mds_virtid_fid),
 543                                  &off, force_sync);
 544         if (rc) {
 545                 CERROR("error writing MDS virtid_fid #"LPU64
 546                        ", err = %d\n", mds->mds_virtid_fid, rc);
 547         }
 548
 549         CDEBUG(D_SUPER, "wrote virtid fid #"LPU64" at idx "
 550                "%llu: err = %d\n", mds->mds_virtid_fid,
 551                off, rc);
 552         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 553
 554         RETURN(rc);
 555 }
 556
 557 static int mds_read_virtid_fid(struct obd_device *obd,
 558                                struct file *file)
 559 {
 560         int rc = 0;
 561         loff_t off = 0;
 562         struct mds_obd *mds = &obd->u.mds;
 563         unsigned long virtid_fid_size = file->f_dentry->d_inode->i_size;
 564         ENTRY;
 565
 566         if (virtid_fid_size == 0) {
 567                 mds->mds_virtid_fid = mds_alloc_fid(obd);
 568         } else {
 569                 rc = fsfilt_read_record(obd, file, &mds->mds_virtid_fid,
 570                                         sizeof(mds->mds_virtid_fid), &off);
 571                 if (rc) {
 572                         CERROR("error reading MDS %s: rc = %d\n",
 573                                file->f_dentry->d_name.name, rc);
 574                         RETURN(rc);
 575                 }
 576         }
 577         rc = mds_update_virtid_fid(obd, NULL, 1);
 578
 579         RETURN(rc);
 580 }
 581
 582 /*
 583  * initializes lustre_id for virtual id directory, it is needed sometimes, as it
 584  * is possible that it will be the parent for object an operations is going to
 585  * be performed on.
 586  */
 587 int mds_fs_setup_virtid(struct obd_device *obd)
 588 {
 589         int rc = 0;
 590         void *handle;
 591         struct lustre_id sid;
 592         struct mds_obd *mds = &obd->u.mds;
 593         struct inode *inode = mds->mds_id_dir->d_inode;
 594         ENTRY;
 595
 596         handle = fsfilt_start(obd, inode,
 597                               FSFILT_OP_SETATTR, NULL);
 598
 599         if (IS_ERR(handle)) {
 600                 rc = PTR_ERR(handle);
 601                 CERROR("fsfilt_start() failed, rc = %d\n", rc);
 602                 RETURN(rc);
 603         }
 604
 605         id_group(&sid) = mds->mds_num;
 606         id_fid(&sid) = mds->mds_virtid_fid;
 607
 608         id_ino(&sid) = inode->i_ino;
 609         id_gen(&sid) = inode->i_generation;
 610         id_type(&sid) = (S_IFMT & inode->i_mode);
 611
 612         down(&inode->i_sem);
 613         rc = mds_update_inode_sid(obd, inode, handle, &sid);
 614         up(&inode->i_sem);
 615
 616         if (rc) {
 617                 CERROR("mds_update_inode_sid() failed, rc = %d\n",
 618                        rc);
 619                 RETURN(rc);
 620         }
 621
 622         rc = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
 623         if (rc) {
 624                 CERROR("fsfilt_commit() failed, rc = %d\n", rc);
 625                 RETURN(rc);
 626         }
 627
 628         RETURN(rc);
 629 }
 630
 631 int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
 632 {
 633         struct mds_obd *mds = &obd->u.mds;
 634         struct lvfs_run_ctxt saved;
 635         struct dentry *dentry;
 636         struct file *file;
 637         int rc;
 638         ENTRY;
 639
 640         rc = cleanup_group_info();
 641         if (rc)
 642                 RETURN(rc);
 643
 644         mds->mds_vfsmnt = mnt;
 645         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 646
 647         fsfilt_setup(obd, mds->mds_sb);
 648
 649         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 650         obd->obd_lvfs_ctxt.pwdmnt = mnt;
 651         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
 652         obd->obd_lvfs_ctxt.fs = get_ds();
 653         obd->obd_lvfs_ctxt.cb_ops = mds_lvfs_ops;
 654
 655         /* setup the directory tree */
 656         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 657         dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755, 0);
 658         if (IS_ERR(dentry)) {
 659                 rc = PTR_ERR(dentry);
 660                 CERROR("cannot create ROOT directory: rc = %d\n", rc);
 661                 GOTO(err_pop, rc);
 662         }
 663
 664         mdc_pack_id(&mds->mds_rootid, dentry->d_inode->i_ino,
 665                     dentry->d_inode->i_generation, S_IFDIR, 0, 0);
 666
 667         dput(dentry);
 668
 669         dentry = lookup_one_len("__iopen__", current->fs->pwd,
 670                                 strlen("__iopen__"));
 671         if (IS_ERR(dentry)) {
 672                 rc = PTR_ERR(dentry);
 673                 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
 674                 GOTO(err_pop, rc);
 675         }
 676         mds->mds_id_de = dentry;
 677         if (!dentry->d_inode || is_bad_inode(dentry->d_inode)) {
 678                 rc = -ENOENT;
 679                 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
 680                 GOTO(err_id_de, rc);
 681         }
 682
 683         dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777, 1);
 684         if (IS_ERR(dentry)) {
 685                 rc = PTR_ERR(dentry);
 686                 CERROR("cannot create PENDING directory: rc = %d\n", rc);
 687                 GOTO(err_id_de, rc);
 688         }
 689         mds->mds_pending_dir = dentry;
 690
 691         dentry = simple_mkdir(current->fs->pwd, "LOGS", 0777, 1);
 692         if (IS_ERR(dentry)) {
 693                 rc = PTR_ERR(dentry);
 694                 CERROR("cannot create LOGS directory: rc = %d\n", rc);
 695                 GOTO(err_pending, rc);
 696         }
 697         mds->mds_logs_dir = dentry;
 698
 699         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
 700         if (IS_ERR(dentry)) {
 701                 rc = PTR_ERR(dentry);
 702                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
 703                 GOTO(err_logs, rc);
 704         }
 705         mds->mds_objects_dir = dentry;
 706
 707         dentry = simple_mkdir(current->fs->pwd, "FIDS", 0777, 1);
 708         if (IS_ERR(dentry)) {
 709                 rc = PTR_ERR(dentry);
 710                 CERROR("cannot create FIDS directory: rc = %d\n", rc);
 711                 GOTO(err_objects, rc);
 712         }
 713         mds->mds_id_dir = dentry;
 714
 715         dentry = simple_mkdir(current->fs->pwd, "UNNAMED", 0777, 1);
 716         if (IS_ERR(dentry)) {
 717                 rc = PTR_ERR(dentry);
 718                 CERROR("cannot create UNNAMED directory: rc = %d\n", rc);
 719                 GOTO(err_unnamed, rc);
 720         }
 721         mds->mds_unnamed_dir = dentry;
 722
 723         /* open and test the last rcvd file */
 724         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
 725         if (IS_ERR(file)) {
 726                 rc = PTR_ERR(file);
 727                 CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
 728                 GOTO(err_id_dir, rc = PTR_ERR(file));
 729         }
 730         mds->mds_rcvd_filp = file;
 731
 732         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 733                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 734                        file->f_dentry->d_inode->i_mode);
 735                 GOTO(err_last_rcvd, rc = -ENOENT);
 736         }
 737
 738         rc = mds_read_last_rcvd(obd, file);
 739         if (rc) {
 740                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
 741                 GOTO(err_last_rcvd, rc);
 742         }
 743
 744         /* open and test last fid file */
 745         file = filp_open(LAST_FID, O_RDWR | O_CREAT, 0644);
 746         if (IS_ERR(file)) {
 747                 rc = PTR_ERR(file);
 748                 CERROR("cannot open/create %s file: rc = %d\n",
 749                        LAST_FID, rc);
 750                 GOTO(err_client, rc = PTR_ERR(file));
 751         }
 752         mds->mds_fid_filp = file;
 753         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 754                 CERROR("%s is not a regular file!: mode = %o\n",
 755                        LAST_FID, file->f_dentry->d_inode->i_mode);
 756                 GOTO(err_last_fid, rc = -ENOENT);
 757         }
 758
 759         rc = mds_read_last_fid(obd, file);
 760         if (rc) {
 761                 CERROR("cannot read %s: rc = %d\n", LAST_FID, rc);
 762                 GOTO(err_last_fid, rc);
 763         }
 764
 765         /* open and test virtid fid file */
 766         file = filp_open(VIRT_FID, O_RDWR | O_CREAT, 0644);
 767         if (IS_ERR(file)) {
 768                 rc = PTR_ERR(file);
 769                 CERROR("cannot open/create %s file: rc = %d\n",
 770                        VIRT_FID, rc);
 771                 GOTO(err_last_fid, rc = PTR_ERR(file));
 772         }
 773         mds->mds_virtid_filp = file;
 774         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 775                 CERROR("%s is not a regular file!: mode = %o\n",
 776                        VIRT_FID, file->f_dentry->d_inode->i_mode);
 777                 GOTO(err_virtid_fid, rc = -ENOENT);
 778         }
 779
 780         rc = mds_read_virtid_fid(obd, file);
 781         if (rc) {
 782                 CERROR("cannot read %s: rc = %d\n", VIRT_FID, rc);
 783                 GOTO(err_virtid_fid, rc);
 784         }
 785
 786         /* open and test the lov objid file */
 787         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
 788         if (IS_ERR(file)) {
 789                 rc = PTR_ERR(file);
 790                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
 791                 GOTO(err_last_fid, rc = PTR_ERR(file));
 792         }
 793         mds->mds_dt_objid_filp = file;
 794         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 795                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
 796                        file->f_dentry->d_inode->i_mode);
 797                 GOTO(err_lov_objid, rc = -ENOENT);
 798         }
 799 err_pop:
 800         if (!rc) {
 801                 rc = mds_fs_post_setup(obd);
 802                 if (rc)
 803                         CERROR("can not post setup fsfilt\n");
 804         }
 805         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 806         return rc;
 807
 808 err_lov_objid:
 809         if (mds->mds_dt_objid_filp && filp_close(mds->mds_dt_objid_filp, 0))
 810                 CERROR("can't close %s after error\n", LOV_OBJID);
 811 err_virtid_fid:
 812         if (mds->mds_virtid_filp && filp_close(mds->mds_virtid_filp, 0))
 813                 CERROR("can't close %s after error\n", VIRT_FID);
 814 err_last_fid:
 815         if (mds->mds_fid_filp && filp_close(mds->mds_fid_filp, 0))
 816                 CERROR("can't close %s after error\n", LAST_FID);
 817 err_client:
 818         class_disconnect_exports(obd, 0);
 819 err_last_rcvd:
 820         if (mds->mds_rcvd_filp && filp_close(mds->mds_rcvd_filp, 0))
 821                 CERROR("can't close %s after error\n", LAST_RCVD);
 822 err_unnamed:
 823         dput(mds->mds_unnamed_dir);
 824 err_id_dir:
 825         dput(mds->mds_id_dir);
 826 err_objects:
 827         dput(mds->mds_objects_dir);
 828 err_logs:
 829         dput(mds->mds_logs_dir);
 830 err_pending:
 831         dput(mds->mds_pending_dir);
 832 err_id_de:
 833         dput(mds->mds_id_de);
 834         goto err_pop;
 835 }
 836
 837 static int  mds_fs_post_cleanup(struct obd_device *obd)
 838 {
 839         int    rc = 0;
 840         rc = fsfilt_post_cleanup(obd);
 841         return rc;
 842 }
 843
 844 int mds_fs_cleanup(struct obd_device *obd, int flags)
 845 {
 846         struct mds_obd *mds = &obd->u.mds;
 847         struct lvfs_run_ctxt saved;
 848         int rc = 0;
 849
 850         if (flags & OBD_OPT_FAILOVER)
 851                 CERROR("%s: shutting down for failover; client state will"
 852                        " be preserved.\n", obd->obd_name);
 853
 854         class_disconnect_exports(obd, flags); /* cleans up client info too */
 855         target_cleanup_recovery(obd);
 856         mds_server_free_data(mds);
 857
 858         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 859         if (mds->mds_virtid_filp) {
 860                 rc = filp_close(mds->mds_virtid_filp, 0);
 861                 mds->mds_virtid_filp = NULL;
 862                 if (rc)
 863                         CERROR("%s file won't close, rc = %d\n", VIRT_FID, rc);
 864         }
 865         if (mds->mds_fid_filp) {
 866                 rc = filp_close(mds->mds_fid_filp, 0);
 867                 mds->mds_fid_filp = NULL;
 868                 if (rc)
 869                         CERROR("%s file won't close, rc = %d\n", LAST_FID, rc);
 870         }
 871         if (mds->mds_rcvd_filp) {
 872                 rc = filp_close(mds->mds_rcvd_filp, 0);
 873                 mds->mds_rcvd_filp = NULL;
 874                 if (rc)
 875                         CERROR("%s file won't close, rc = %d\n", LAST_RCVD, rc);
 876         }
 877         if (mds->mds_dt_objid_filp) {
 878                 rc = filp_close(mds->mds_dt_objid_filp, 0);
 879                 mds->mds_dt_objid_filp = NULL;
 880                 if (rc)
 881                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
 882         }
 883         if (mds->mds_unnamed_dir != NULL) {
 884                 l_dput(mds->mds_unnamed_dir);
 885                 mds->mds_unnamed_dir = NULL;
 886         }
 887         if (mds->mds_id_dir != NULL) {
 888                 l_dput(mds->mds_id_dir);
 889                 mds->mds_id_dir = NULL;
 890         }
 891         if (mds->mds_objects_dir != NULL) {
 892                 l_dput(mds->mds_objects_dir);
 893                 mds->mds_objects_dir = NULL;
 894         }
 895         if (mds->mds_logs_dir) {
 896                 l_dput(mds->mds_logs_dir);
 897                 mds->mds_logs_dir = NULL;
 898         }
 899         if (mds->mds_pending_dir) {
 900                 l_dput(mds->mds_pending_dir);
 901                 mds->mds_pending_dir = NULL;
 902         }
 903         rc = mds_fs_post_cleanup(obd);
 904
 905         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 906         shrink_dcache_parent(mds->mds_id_de);
 907         dput(mds->mds_id_de);
 908
 909         return rc;
 910 }
 911
 912 /* Creates an object with the same name as its id.  Because this is not at all
 913  * performance sensitive, it is accomplished by creating a file, checking the
 914  * id, and renaming it. */
 915 int mds_obd_create(struct obd_export *exp, struct obdo *oa,
 916                    struct lov_stripe_md **ea, struct obd_trans_info *oti)
 917 {
 918         struct mds_obd *mds = &exp->exp_obd->u.mds;
 919         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 920         struct file *filp;
 921         struct dentry *dchild;
 922         struct lvfs_run_ctxt saved;
 923         char idname[LL_ID_NAMELEN];
 924         int rc = 0, err, idlen;
 925         void *handle;
 926         ENTRY;
 927
 928         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 929         down(&parent_inode->i_sem);
 930         if (oa->o_id) {
 931                 idlen = ll_id2str(idname, oa->o_id, oa->o_generation);
 932                 dchild = lookup_one_len(idname, mds->mds_objects_dir, idlen);
 933                 if (IS_ERR(dchild))
 934                         GOTO(out_pop, rc = PTR_ERR(dchild));
 935
 936                 if (dchild->d_inode == NULL) {
 937                         struct dentry_params dp;
 938                         struct inode *inode;
 939
 940                         CWARN("creating log with ID "LPU64"\n", oa->o_id);
 941
 942                         dchild->d_fsdata = (void *) &dp;
 943                         dp.p_ptr = NULL;
 944                         dp.p_inum = oa->o_id;
 945                         rc = ll_vfs_create(parent_inode, dchild, S_IFREG, NULL);
 946                         if (dchild->d_fsdata == (void *)(unsigned long)oa->o_id)
 947                                 dchild->d_fsdata = NULL;
 948                         if (rc) {
 949                                 CDEBUG(D_INODE, "err during create: %d\n", rc);
 950                                 dput(dchild);
 951                                 GOTO(out_pop, rc);
 952                         }
 953                         inode = dchild->d_inode;
 954                         LASSERT(inode->i_ino == oa->o_id);
 955                         inode->i_generation = oa->o_generation;
 956                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
 957                                inode->i_ino, inode->i_generation);
 958                         mark_inode_dirty(inode);
 959                 } else {
 960                         CWARN("it should be here!\n");
 961                 }
 962                 GOTO(out_pop, rc);
 963         }
 964
 965         sprintf(idname, "OBJECTS/%u.%u", ll_insecure_random_int(), current->pid);
 966         filp = filp_open(idname, O_CREAT | O_EXCL, 0644);
 967         if (IS_ERR(filp)) {
 968                 rc = PTR_ERR(filp);
 969                 if (rc == -EEXIST) {
 970                         CERROR("impossible object name collision %s\n",
 971                                idname);
 972                         LBUG();
 973                 }
 974                 CERROR("error creating tmp object %s: rc %d\n",
 975                        idname, rc);
 976                 GOTO(out_pop, rc);
 977         }
 978
 979         LASSERT(mds->mds_objects_dir == filp->f_dentry->d_parent);
 980
 981         oa->o_id = filp->f_dentry->d_inode->i_ino;
 982         oa->o_generation = filp->f_dentry->d_inode->i_generation;
 983         idlen = ll_id2str(idname, oa->o_id, oa->o_generation);
 984
 985         CWARN("created log anonymous "LPU64"/%u\n",
 986               oa->o_id, oa->o_generation);
 987
 988         dchild = lookup_one_len(idname, mds->mds_objects_dir, idlen);
 989         if (IS_ERR(dchild)) {
 990                 CERROR("getting neg dentry for obj rename: %d\n", rc);
 991                 GOTO(out_close, rc = PTR_ERR(dchild));
 992         }
 993         if (dchild->d_inode != NULL) {
 994                 CERROR("impossible non-negative obj dentry " LPU64":%u!\n",
 995                        oa->o_id, oa->o_generation);
 996                 LBUG();
 997         }
 998
 999         handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
1000                               FSFILT_OP_RENAME, NULL);
1001         if (IS_ERR(handle))
1002                 GOTO(out_dput, rc = PTR_ERR(handle));
1003
1004         lock_kernel();
1005         rc = vfs_rename(mds->mds_objects_dir->d_inode, filp->f_dentry,
1006                         mds->mds_objects_dir->d_inode, dchild);
1007         unlock_kernel();
1008         if (rc)
1009                 CERROR("error renaming new object "LPU64":%u: rc %d\n",
1010                        oa->o_id, oa->o_generation, rc);
1011
1012         err = fsfilt_commit(exp->exp_obd, mds->mds_sb,
1013                             mds->mds_objects_dir->d_inode, handle, 0);
1014         if (!err) {
1015                 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
1016                 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLGROUP;
1017         } else if (!rc)
1018                 rc = err;
1019 out_dput:
1020         dput(dchild);
1021 out_close:
1022         err = filp_close(filp, 0);
1023         if (err) {
1024                 CERROR("closing tmpfile %s: rc %d\n", idname, rc);
1025                 if (!rc)
1026                         rc = err;
1027         }
1028 out_pop:
1029         up(&parent_inode->i_sem);
1030         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
1031         RETURN(rc);
1032 }
1033
1034 int mds_obd_destroy(struct obd_export *exp, struct obdo *oa,
1035                     struct lov_stripe_md *ea, struct obd_trans_info *oti)
1036 {
1037         struct mds_obd *mds = &exp->exp_obd->u.mds;
1038         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
1039         struct obd_device *obd = exp->exp_obd;
1040         struct lvfs_run_ctxt saved;
1041         char idname[LL_ID_NAMELEN];
1042         struct dentry *de;
1043         void *handle;
1044         int err, idlen, rc = 0;
1045         ENTRY;
1046
1047         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1048
1049         idlen = ll_id2str(idname, oa->o_id, oa->o_generation);
1050
1051         down(&parent_inode->i_sem);
1052         de = lookup_one_len(idname, mds->mds_objects_dir, idlen);
1053         if (IS_ERR(de) || de->d_inode == NULL) {
1054                 rc = IS_ERR(de) ? PTR_ERR(de) : -ENOENT;
1055                 CERROR("destroying non-existent object "LPU64" %s: rc %d\n",
1056                        oa->o_id, idname, rc);
1057                 GOTO(out_dput, rc);
1058         }
1059         /* Stripe count is 1 here since this is some MDS specific stuff
1060            that is unlinked, not spanned across multiple OSTs */
1061         handle = fsfilt_start_log(obd, mds->mds_objects_dir->d_inode,
1062                                   FSFILT_OP_UNLINK, oti, 1);
1063
1064         if (IS_ERR(handle))
1065                 GOTO(out_dput, rc = PTR_ERR(handle));
1066
1067         rc = vfs_unlink(mds->mds_objects_dir->d_inode, de);
1068         if (rc)
1069                 CERROR("error destroying object "LPU64":%u: rc %d\n",
1070                        oa->o_id, oa->o_generation, rc);
1071
1072         err = fsfilt_commit(obd, mds->mds_sb, mds->mds_objects_dir->d_inode,
1073                             handle, exp->exp_sync);
1074         if (err && !rc)
1075                 rc = err;
1076 out_dput:
1077         if (de != NULL)
1078                 l_dput(de);
1079         up(&parent_inode->i_sem);
1080         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1081         RETURN(rc);
1082 }