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 <portals/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 #define MDS_MAX_CLIENT_WORDS (MDS_MAX_CLIENTS / sizeof(unsigned long))
  50
  51 #define LAST_RCVD "last_rcvd"
  52 #define LOV_OBJID "lov_objid"
  53 #define LAST_FID  "last_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         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(med->med_mcd->mcd_uuid, 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         return 0;
 180 }
 181
 182 static int mds_server_free_data(struct mds_obd *mds)
 183 {
 184         OBD_FREE(mds->mds_client_bitmap,
 185                  MDS_MAX_CLIENT_WORDS * sizeof(unsigned long));
 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,
 256                   MDS_MAX_CLIENT_WORDS * sizeof(unsigned long));
 257         if (!mds->mds_client_bitmap) {
 258                 OBD_FREE(msd, sizeof(*msd));
 259                 RETURN(-ENOMEM);
 260         }
 261
 262         mds->mds_server_data = msd;
 263
 264         if (last_rcvd_size == 0) {
 265                 CWARN("%s: initializing new %s\n", obd->obd_name,
 266                       file->f_dentry->d_name.name);
 267
 268                 memcpy(msd->msd_uuid, obd->obd_uuid.uuid,sizeof(msd->msd_uuid));
 269                 msd->msd_last_transno = 0;
 270                 mount_count = msd->msd_mount_count = 0;
 271                 msd->msd_server_size = cpu_to_le32(MDS_LR_SERVER_SIZE);
 272                 msd->msd_client_start = cpu_to_le32(MDS_LR_CLIENT_START);
 273                 msd->msd_client_size = cpu_to_le16(MDS_LR_CLIENT_SIZE);
 274                 msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
 275         } else {
 276                 rc = fsfilt_read_record(obd, file, msd, sizeof(*msd), &off);
 277                 if (rc) {
 278                         CERROR("error reading MDS %s: rc = %d\n",
 279                                file->f_dentry->d_name.name, rc);
 280                         GOTO(err_msd, rc);
 281                 }
 282                 if (strcmp(msd->msd_uuid, obd->obd_uuid.uuid) != 0) {
 283                         CERROR("OBD UUID %s does not match last_rcvd UUID %s\n",
 284                                obd->obd_uuid.uuid, msd->msd_uuid);
 285                         GOTO(err_msd, rc = -EINVAL);
 286                 }
 287                 mount_count = le64_to_cpu(msd->msd_mount_count);
 288         }
 289         if (msd->msd_feature_incompat & ~cpu_to_le32(MDS_INCOMPAT_SUPP)) {
 290                 CERROR("unsupported incompat feature %x\n",
 291                        le32_to_cpu(msd->msd_feature_incompat) &
 292                        ~MDS_INCOMPAT_SUPP);
 293                 GOTO(err_msd, rc = -EINVAL);
 294         }
 295         /* XXX updating existing b_devel fs only, can be removed in future */
 296         msd->msd_feature_rocompat = cpu_to_le32(MDS_ROCOMPAT_LOVOBJID);
 297         if (msd->msd_feature_rocompat & ~cpu_to_le32(MDS_ROCOMPAT_SUPP)) {
 298                 CERROR("unsupported read-only feature %x\n",
 299                        le32_to_cpu(msd->msd_feature_rocompat) &
 300                        ~MDS_ROCOMPAT_SUPP);
 301                 /* Do something like remount filesystem read-only */
 302                 GOTO(err_msd, rc = -EINVAL);
 303         }
 304
 305         mds->mds_last_transno = le64_to_cpu(msd->msd_last_transno);
 306
 307         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 308                obd->obd_name, mds->mds_last_transno);
 309         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 310                obd->obd_name, mount_count + 1);
 311         CDEBUG(D_INODE, "%s: server data size: %u\n",
 312                obd->obd_name, le32_to_cpu(msd->msd_server_size));
 313         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 314                obd->obd_name, le32_to_cpu(msd->msd_client_start));
 315         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 316                obd->obd_name, le32_to_cpu(msd->msd_client_size));
 317         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 318                obd->obd_name, last_rcvd_size);
 319         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 320                last_rcvd_size <= le32_to_cpu(msd->msd_client_start) ? 0 :
 321                (last_rcvd_size - le32_to_cpu(msd->msd_client_start)) /
 322                 le16_to_cpu(msd->msd_client_size));
 323
 324         /* When we do a clean MDS shutdown, we save the last_transno into
 325          * the header.  If we find clients with higher last_transno values
 326          * then those clients may need recovery done. */
 327         for (cl_idx = 0, off = le32_to_cpu(msd->msd_client_start);
 328              off < last_rcvd_size; cl_idx++) {
 329                 __u64 last_transno;
 330                 struct obd_export *exp;
 331                 struct mds_export_data *med;
 332
 333                 if (!mcd) {
 334                         OBD_ALLOC_WAIT(mcd, sizeof(*mcd));
 335                         if (!mcd)
 336                                 GOTO(err_client, rc = -ENOMEM);
 337                 }
 338
 339                 /* Don't assume off is incremented properly by
 340                  * fsfilt_read_record(), in case sizeof(*mcd)
 341                  * isn't the same as msd->msd_client_size.  */
 342                 off = le32_to_cpu(msd->msd_client_start) +
 343                         cl_idx * le16_to_cpu(msd->msd_client_size);
 344                 rc = fsfilt_read_record(obd, file, mcd, sizeof(*mcd), &off);
 345                 if (rc) {
 346                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 347                                file->f_dentry->d_name.name, cl_idx, off, rc);
 348                         break; /* read error shouldn't cause startup to fail */
 349                 }
 350
 351                 if (mcd->mcd_uuid[0] == '\0') {
 352                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 353                                cl_idx);
 354                         continue;
 355                 }
 356
 357                 last_transno = le64_to_cpu(mcd->mcd_last_transno);
 358
 359                 /* These exports are cleaned up by mds_disconnect(), so they
 360                  * need to be set up like real exports as mds_connect() does.
 361                  */
 362                 CDEBUG(D_HA|D_WARNING,"RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 363                        " srv lr: "LPU64" lx: "LPU64"\n", mcd->mcd_uuid, cl_idx,
 364                        last_transno, le64_to_cpu(msd->msd_last_transno),
 365                        mcd->mcd_last_xid);
 366
 367                 exp = class_new_export(obd);
 368                 if (exp == NULL)
 369                         GOTO(err_client, rc = -ENOMEM);
 370
 371                 memcpy(&exp->exp_client_uuid.uuid, mcd->mcd_uuid,
 372                        sizeof exp->exp_client_uuid.uuid);
 373                 med = &exp->exp_mds_data;
 374                 med->med_mcd = mcd;
 375                 mds_client_add(obd, mds, med, cl_idx);
 376                 /* create helper if export init gets more complex */
 377                 INIT_LIST_HEAD(&med->med_open_head);
 378                 spin_lock_init(&med->med_open_lock);
 379
 380                 mcd = NULL;
 381                 exp->exp_replay_needed = 1;
 382                 obd->obd_recoverable_clients++;
 383                 obd->obd_max_recoverable_clients++;
 384                 class_export_put(exp);
 385
 386                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
 387                        cl_idx, last_transno);
 388
 389                 if (last_transno > mds->mds_last_transno)
 390                        mds->mds_last_transno = last_transno;
 391         }
 392         if (mcd)
 393                 OBD_FREE(mcd, sizeof(*mcd));
 394         obd->obd_last_committed = mds->mds_last_transno;
 395         if (obd->obd_recoverable_clients) {
 396                 CWARN("RECOVERY: service %s, %d recoverable clients, "
 397                       "last_transno "LPU64"\n", obd->obd_name,
 398                       obd->obd_recoverable_clients, mds->mds_last_transno);
 399                 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
 400                 target_start_recovery_thread(obd, mds_handle);
 401                 obd->obd_recovery_start = LTIME_S(CURRENT_TIME);
 402         }
 403
 404         mds->mds_mount_count = mount_count + 1;
 405         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
 406
 407         /* save it, so mount count and last_transno is current */
 408         rc = mds_update_server_data(obd, 1);
 409         if (rc)
 410                 GOTO(err_client, rc);
 411
 412         RETURN(0);
 413
 414 err_client:
 415         class_disconnect_exports(obd, 0);
 416 err_msd:
 417         mds_server_free_data(mds);
 418         RETURN(rc);
 419 }
 420
 421 static int mds_fs_post_setup(struct obd_device *obd)
 422 {
 423         struct mds_obd *mds = &obd->u.mds;
 424         struct dentry *dentry;
 425         int rc = 0;
 426         ENTRY;
 427
 428         dentry = mds_id2dentry(obd, &mds->mds_rootid, NULL);
 429         if (IS_ERR(dentry)) {
 430                 CERROR("Can't find ROOT, err = %d\n",
 431                        (int)PTR_ERR(dentry));
 432                 RETURN(PTR_ERR(dentry));
 433         }
 434
 435         rc = fsfilt_post_setup(obd, dentry);
 436         if (rc)
 437                 goto out_dentry;
 438
 439         LASSERT(dentry->d_inode != NULL);
 440
 441         fsfilt_set_fs_flags(obd, dentry->d_inode,
 442                             SM_DO_REC | SM_DO_COW);
 443
 444         fsfilt_set_fs_flags(obd, mds->mds_pending_dir->d_inode,
 445                             SM_DO_REC | SM_DO_COW);
 446
 447         fsfilt_set_mds_flags(obd, mds->mds_sb);
 448
 449 out_dentry:
 450         l_dput(dentry);
 451         RETURN(rc);
 452 }
 453
 454 /*
 455  * sets up root inode lustre_id. It tries to read it first from root inode and
 456  * if it is not there, new rootid is allocated and saved there.
 457  */
 458 int mds_fs_setup_rootid(struct obd_device *obd)
 459 {
 460         int rc = 0;
 461         void *handle;
 462         struct inode *inode;
 463         struct dentry *dentry;
 464         struct mds_obd *mds = &obd->u.mds;
 465         ENTRY;
 466
 467         /* getting root directory and setup its fid. */
 468         dentry = mds_id2dentry(obd, &mds->mds_rootid, NULL);
 469         if (IS_ERR(dentry)) {
 470                 CERROR("Can't find ROOT by "DLID4", err = %d\n",
 471                        OLID4(&mds->mds_rootid), (int)PTR_ERR(dentry));
 472                 RETURN(PTR_ERR(dentry));
 473         }
 474
 475         inode = dentry->d_inode;
 476         LASSERT(dentry->d_inode);
 477
 478         rc = mds_pack_inode2id(obd, &mds->mds_rootid, inode, 1);
 479         if (rc < 0) {
 480                 if (rc != -ENODATA)
 481                         GOTO(out_dentry, rc);
 482         } else {
 483                 /*
 484                  * rootid is filled by mds_read_inode_sid(), so we do not need
 485                  * to allocate it and update. The only thing we need to check is
 486                  * mds_num.
 487                  */
 488                 LASSERT(id_group(&mds->mds_rootid) == mds->mds_num);
 489                 mds_set_last_fid(obd, id_fid(&mds->mds_rootid));
 490                 GOTO(out_dentry, rc);
 491         }
 492
 493         /* allocating new one, as it is not found in root inode. */
 494         handle = fsfilt_start(obd, inode,
 495                               FSFILT_OP_SETATTR, NULL);
 496
 497         if (IS_ERR(handle)) {
 498                 rc = PTR_ERR(handle);
 499                 CERROR("fsfilt_start() failed, rc = %d\n", rc);
 500                 GOTO(out_dentry, rc);
 501         }
 502
 503         down(&inode->i_sem);
 504         rc = mds_alloc_inode_sid(obd, inode, handle, &mds->mds_rootid);
 505         up(&inode->i_sem);
 506
 507         if (rc) {
 508                 CERROR("mds_alloc_inode_sid() failed, rc = %d\n",
 509                        rc);
 510                 GOTO(out_dentry, rc);
 511         }
 512
 513         rc = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
 514         if (rc)
 515                 CERROR("fsfilt_commit() failed, rc = %d\n", rc);
 516
 517         EXIT;
 518 out_dentry:
 519         l_dput(dentry);
 520         if (rc == 0)
 521                 CWARN("%s: rootid: "DLID4"\n", obd->obd_name,
 522                       OLID4(&mds->mds_rootid));
 523         return rc;
 524 }
 525
 526 /*
 527  * initializes lustre_id for virtual id directory, it is needed sometimes, as it
 528  * is possible that it will be the parent for object an operations is going to
 529  * be performed on.
 530  */
 531 int mds_fs_setup_virtid(struct obd_device *obd)
 532 {
 533         int rc = 0;
 534         void *handle;
 535         struct lustre_id id;
 536         struct mds_obd *mds = &obd->u.mds;
 537         struct inode *inode = mds->mds_id_dir->d_inode;
 538         ENTRY;
 539
 540         handle = fsfilt_start(obd, inode,
 541                               FSFILT_OP_SETATTR, NULL);
 542
 543         if (IS_ERR(handle)) {
 544                 rc = PTR_ERR(handle);
 545                 CERROR("fsfilt_start() failed, rc = %d\n", rc);
 546                 RETURN(rc);
 547         }
 548
 549         down(&inode->i_sem);
 550         rc = mds_alloc_inode_sid(obd, inode, handle, &id);
 551         up(&inode->i_sem);
 552
 553         if (rc) {
 554                 CERROR("mds_alloc_inode_sid() failed, rc = %d\n",
 555                        rc);
 556                 RETURN(rc);
 557         }
 558
 559         rc = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
 560         if (rc) {
 561                 CERROR("fsfilt_commit() failed, rc = %d\n", rc);
 562                 RETURN(rc);
 563         }
 564
 565         RETURN(rc);
 566 }
 567
 568 int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
 569 {
 570         struct mds_obd *mds = &obd->u.mds;
 571         struct lvfs_run_ctxt saved;
 572         struct dentry *dentry;
 573         struct file *file;
 574         int rc;
 575         ENTRY;
 576
 577         rc = cleanup_group_info();
 578         if (rc)
 579                 RETURN(rc);
 580
 581         mds->mds_vfsmnt = mnt;
 582         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 583
 584         fsfilt_setup(obd, mds->mds_sb);
 585
 586         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 587         obd->obd_lvfs_ctxt.pwdmnt = mnt;
 588         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
 589         obd->obd_lvfs_ctxt.fs = get_ds();
 590         obd->obd_lvfs_ctxt.cb_ops = mds_lvfs_ops;
 591
 592         /* setup the directory tree */
 593         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 594         dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755, 0);
 595         if (IS_ERR(dentry)) {
 596                 rc = PTR_ERR(dentry);
 597                 CERROR("cannot create ROOT directory: rc = %d\n", rc);
 598                 GOTO(err_pop, rc);
 599         }
 600
 601         mdc_pack_id(&mds->mds_rootid, dentry->d_inode->i_ino,
 602                     dentry->d_inode->i_generation, S_IFDIR, 0, 0);
 603
 604         dput(dentry);
 605
 606         dentry = lookup_one_len("__iopen__", current->fs->pwd,
 607                                 strlen("__iopen__"));
 608         if (IS_ERR(dentry)) {
 609                 rc = PTR_ERR(dentry);
 610                 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
 611                 GOTO(err_pop, rc);
 612         }
 613         mds->mds_id_de = dentry;
 614
 615         if (!dentry->d_inode) {
 616                 rc = -ENOENT;
 617                 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
 618                 GOTO(err_id_de, rc);
 619         }
 620
 621         dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777, 1);
 622         if (IS_ERR(dentry)) {
 623                 rc = PTR_ERR(dentry);
 624                 CERROR("cannot create PENDING directory: rc = %d\n", rc);
 625                 GOTO(err_id_de, rc);
 626         }
 627         mds->mds_pending_dir = dentry;
 628
 629         dentry = simple_mkdir(current->fs->pwd, "LOGS", 0777, 1);
 630         if (IS_ERR(dentry)) {
 631                 rc = PTR_ERR(dentry);
 632                 CERROR("cannot create LOGS directory: rc = %d\n", rc);
 633                 GOTO(err_pending, rc);
 634         }
 635         mds->mds_logs_dir = dentry;
 636
 637         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
 638         if (IS_ERR(dentry)) {
 639                 rc = PTR_ERR(dentry);
 640                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
 641                 GOTO(err_logs, rc);
 642         }
 643         mds->mds_objects_dir = dentry;
 644
 645         dentry = simple_mkdir(current->fs->pwd, "FIDS", 0777, 1);
 646         if (IS_ERR(dentry)) {
 647                 rc = PTR_ERR(dentry);
 648                 CERROR("cannot create FIDS directory: rc = %d\n", rc);
 649                 GOTO(err_objects, rc);
 650         }
 651         mds->mds_id_dir = dentry;
 652
 653         dentry = simple_mkdir(current->fs->pwd, "UNNAMED", 0777, 1);
 654         if (IS_ERR(dentry)) {
 655                 rc = PTR_ERR(dentry);
 656                 CERROR("cannot create UNNAMED directory: rc = %d\n", rc);
 657                 GOTO(err_unnamed, rc);
 658         }
 659         mds->mds_unnamed_dir = dentry;
 660
 661         /* open and test the last rcvd file */
 662         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
 663         if (IS_ERR(file)) {
 664                 rc = PTR_ERR(file);
 665                 CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
 666                 GOTO(err_id_dir, rc = PTR_ERR(file));
 667         }
 668         mds->mds_rcvd_filp = file;
 669
 670         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 671                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 672                        file->f_dentry->d_inode->i_mode);
 673                 GOTO(err_last_rcvd, rc = -ENOENT);
 674         }
 675
 676         rc = mds_read_last_rcvd(obd, file);
 677         if (rc) {
 678                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
 679                 GOTO(err_last_rcvd, rc);
 680         }
 681
 682         /* open and test the last fid file */
 683         file = filp_open(LAST_FID, O_RDWR | O_CREAT, 0644);
 684         if (IS_ERR(file)) {
 685                 rc = PTR_ERR(file);
 686                 CERROR("cannot open/create %s file: rc = %d\n",
 687                        LAST_FID, rc);
 688                 GOTO(err_client, rc = PTR_ERR(file));
 689         }
 690         mds->mds_fid_filp = file;
 691         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 692                 CERROR("%s is not a regular file!: mode = %o\n",
 693                        LAST_FID, file->f_dentry->d_inode->i_mode);
 694                 GOTO(err_last_fid, rc = -ENOENT);
 695         }
 696
 697         rc = mds_read_last_fid(obd, file);
 698         if (rc) {
 699                 CERROR("cannot read %s: rc = %d\n", LAST_FID, rc);
 700                 GOTO(err_last_fid, rc);
 701         }
 702
 703         /* open and test the lov objid file */
 704         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
 705         if (IS_ERR(file)) {
 706                 rc = PTR_ERR(file);
 707                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
 708                 GOTO(err_last_fid, rc = PTR_ERR(file));
 709         }
 710         mds->mds_lov_objid_filp = file;
 711         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 712                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
 713                        file->f_dentry->d_inode->i_mode);
 714                 GOTO(err_lov_objid, rc = -ENOENT);
 715         }
 716 err_pop:
 717         if (!rc) {
 718                 rc = mds_fs_post_setup(obd);
 719                 if (rc)
 720                         CERROR("can not post setup fsfilt\n");
 721         }
 722         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 723         return rc;
 724
 725 err_lov_objid:
 726         if (mds->mds_lov_objid_filp && filp_close(mds->mds_lov_objid_filp, 0))
 727                 CERROR("can't close %s after error\n", LOV_OBJID);
 728 err_last_fid:
 729         if (mds->mds_fid_filp && filp_close(mds->mds_fid_filp, 0))
 730                 CERROR("can't close %s after error\n", LAST_FID);
 731 err_client:
 732         class_disconnect_exports(obd, 0);
 733 err_last_rcvd:
 734         if (mds->mds_rcvd_filp && filp_close(mds->mds_rcvd_filp, 0))
 735                 CERROR("can't close %s after error\n", LAST_RCVD);
 736 err_unnamed:
 737         dput(mds->mds_unnamed_dir);
 738 err_id_dir:
 739         dput(mds->mds_id_dir);
 740 err_objects:
 741         dput(mds->mds_objects_dir);
 742 err_logs:
 743         dput(mds->mds_logs_dir);
 744 err_pending:
 745         dput(mds->mds_pending_dir);
 746 err_id_de:
 747         dput(mds->mds_id_de);
 748         goto err_pop;
 749 }
 750
 751 static int  mds_fs_post_cleanup(struct obd_device *obd)
 752 {
 753         int    rc = 0;
 754         rc = fsfilt_post_cleanup(obd);
 755         return rc;
 756 }
 757
 758 int mds_fs_cleanup(struct obd_device *obd, int flags)
 759 {
 760         struct mds_obd *mds = &obd->u.mds;
 761         struct lvfs_run_ctxt saved;
 762         int rc = 0;
 763
 764         if (flags & OBD_OPT_FAILOVER)
 765                 CERROR("%s: shutting down for failover; client state will"
 766                        " be preserved.\n", obd->obd_name);
 767
 768         class_disconnect_exports(obd, flags); /* cleans up client info too */
 769         target_cleanup_recovery(obd);
 770         mds_server_free_data(mds);
 771
 772         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 773         if (mds->mds_fid_filp) {
 774                 rc = filp_close(mds->mds_fid_filp, 0);
 775                 mds->mds_fid_filp = NULL;
 776                 if (rc)
 777                         CERROR("%s file won't close, rc = %d\n", LAST_FID, rc);
 778         }
 779         if (mds->mds_rcvd_filp) {
 780                 rc = filp_close(mds->mds_rcvd_filp, 0);
 781                 mds->mds_rcvd_filp = NULL;
 782                 if (rc)
 783                         CERROR("%s file won't close, rc = %d\n", LAST_RCVD, rc);
 784         }
 785         if (mds->mds_lov_objid_filp) {
 786                 rc = filp_close(mds->mds_lov_objid_filp, 0);
 787                 mds->mds_lov_objid_filp = NULL;
 788                 if (rc)
 789                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
 790         }
 791         if (mds->mds_unnamed_dir != NULL) {
 792                 l_dput(mds->mds_unnamed_dir);
 793                 mds->mds_unnamed_dir = NULL;
 794         }
 795         if (mds->mds_id_dir != NULL) {
 796                 l_dput(mds->mds_id_dir);
 797                 mds->mds_id_dir = NULL;
 798         }
 799         if (mds->mds_objects_dir != NULL) {
 800                 l_dput(mds->mds_objects_dir);
 801                 mds->mds_objects_dir = NULL;
 802         }
 803         if (mds->mds_logs_dir) {
 804                 l_dput(mds->mds_logs_dir);
 805                 mds->mds_logs_dir = NULL;
 806         }
 807         if (mds->mds_pending_dir) {
 808                 l_dput(mds->mds_pending_dir);
 809                 mds->mds_pending_dir = NULL;
 810         }
 811         rc = mds_fs_post_cleanup(obd);
 812
 813         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 814         shrink_dcache_parent(mds->mds_id_de);
 815         dput(mds->mds_id_de);
 816
 817         return rc;
 818 }
 819
 820 /* Creates an object with the same name as its id.  Because this is not at all
 821  * performance sensitive, it is accomplished by creating a file, checking the
 822  * id, and renaming it. */
 823 int mds_obd_create(struct obd_export *exp, struct obdo *oa,
 824                    struct lov_stripe_md **ea, struct obd_trans_info *oti)
 825 {
 826         struct mds_obd *mds = &exp->exp_obd->u.mds;
 827         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 828         struct file *filp;
 829         struct dentry *dchild;
 830         struct lvfs_run_ctxt saved;
 831         char idname[LL_ID_NAMELEN];
 832         int rc = 0, err, idlen;
 833         void *handle;
 834         ENTRY;
 835
 836         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 837         down(&parent_inode->i_sem);
 838         if (oa->o_id) {
 839                 idlen = ll_id2str(idname, oa->o_id, oa->o_generation);
 840                 dchild = lookup_one_len(idname, mds->mds_objects_dir, idlen);
 841                 if (IS_ERR(dchild))
 842                         GOTO(out_pop, rc = PTR_ERR(dchild));
 843
 844                 if (dchild->d_inode == NULL) {
 845                         struct dentry_params dp;
 846                         struct inode *inode;
 847
 848                         CWARN("creating log with ID "LPU64"\n", oa->o_id);
 849
 850                         dchild->d_fsdata = (void *) &dp;
 851                         dp.p_ptr = NULL;
 852                         dp.p_inum = oa->o_id;
 853                         rc = ll_vfs_create(parent_inode, dchild, S_IFREG, NULL);
 854                         if (dchild->d_fsdata == (void *)(unsigned long)oa->o_id)
 855                                 dchild->d_fsdata = NULL;
 856                         if (rc) {
 857                                 CDEBUG(D_INODE, "err during create: %d\n", rc);
 858                                 dput(dchild);
 859                                 GOTO(out_pop, rc);
 860                         }
 861                         inode = dchild->d_inode;
 862                         LASSERT(inode->i_ino == oa->o_id);
 863                         inode->i_generation = oa->o_generation;
 864                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
 865                                inode->i_ino, inode->i_generation);
 866                         mark_inode_dirty(inode);
 867                 } else {
 868                         CWARN("it should be here!\n");
 869                 }
 870                 GOTO(out_pop, rc);
 871         }
 872
 873         sprintf(idname, "OBJECTS/%u.%u", ll_insecure_random_int(), current->pid);
 874         filp = filp_open(idname, O_CREAT | O_EXCL, 0644);
 875         if (IS_ERR(filp)) {
 876                 rc = PTR_ERR(filp);
 877                 if (rc == -EEXIST) {
 878                         CERROR("impossible object name collision %s\n",
 879                                idname);
 880                         LBUG();
 881                 }
 882                 CERROR("error creating tmp object %s: rc %d\n",
 883                        idname, rc);
 884                 GOTO(out_pop, rc);
 885         }
 886
 887         LASSERT(mds->mds_objects_dir == filp->f_dentry->d_parent);
 888
 889         oa->o_id = filp->f_dentry->d_inode->i_ino;
 890         oa->o_generation = filp->f_dentry->d_inode->i_generation;
 891         idlen = ll_id2str(idname, oa->o_id, oa->o_generation);
 892
 893         CWARN("created log anonymous "LPU64"/%u\n",
 894               oa->o_id, oa->o_generation);
 895
 896         dchild = lookup_one_len(idname, mds->mds_objects_dir, idlen);
 897         if (IS_ERR(dchild)) {
 898                 CERROR("getting neg dentry for obj rename: %d\n", rc);
 899                 GOTO(out_close, rc = PTR_ERR(dchild));
 900         }
 901         if (dchild->d_inode != NULL) {
 902                 CERROR("impossible non-negative obj dentry " LPU64":%u!\n",
 903                        oa->o_id, oa->o_generation);
 904                 LBUG();
 905         }
 906
 907         handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
 908                               FSFILT_OP_RENAME, NULL);
 909         if (IS_ERR(handle))
 910                 GOTO(out_dput, rc = PTR_ERR(handle));
 911
 912         lock_kernel();
 913         rc = vfs_rename(mds->mds_objects_dir->d_inode, filp->f_dentry,
 914                         mds->mds_objects_dir->d_inode, dchild);
 915         unlock_kernel();
 916         if (rc)
 917                 CERROR("error renaming new object "LPU64":%u: rc %d\n",
 918                        oa->o_id, oa->o_generation, rc);
 919
 920         err = fsfilt_commit(exp->exp_obd, mds->mds_sb,
 921                             mds->mds_objects_dir->d_inode, handle, 0);
 922         if (!err) {
 923                 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
 924                 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLGROUP;
 925         } else if (!rc)
 926                 rc = err;
 927 out_dput:
 928         dput(dchild);
 929 out_close:
 930         err = filp_close(filp, 0);
 931         if (err) {
 932                 CERROR("closing tmpfile %s: rc %d\n", idname, rc);
 933                 if (!rc)
 934                         rc = err;
 935         }
 936 out_pop:
 937         up(&parent_inode->i_sem);
 938         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 939         RETURN(rc);
 940 }
 941
 942 int mds_obd_destroy(struct obd_export *exp, struct obdo *oa,
 943                     struct lov_stripe_md *ea, struct obd_trans_info *oti)
 944 {
 945         struct mds_obd *mds = &exp->exp_obd->u.mds;
 946         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 947         struct obd_device *obd = exp->exp_obd;
 948         struct lvfs_run_ctxt saved;
 949         char idname[LL_ID_NAMELEN];
 950         struct dentry *de;
 951         void *handle;
 952         int err, idlen, rc = 0;
 953         ENTRY;
 954
 955         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 956
 957         idlen = ll_id2str(idname, oa->o_id, oa->o_generation);
 958
 959         down(&parent_inode->i_sem);
 960         de = lookup_one_len(idname, mds->mds_objects_dir, idlen);
 961         if (IS_ERR(de) || de->d_inode == NULL) {
 962                 rc = IS_ERR(de) ? PTR_ERR(de) : -ENOENT;
 963                 CERROR("destroying non-existent object "LPU64" %s: rc %d\n",
 964                        oa->o_id, idname, rc);
 965                 GOTO(out_dput, rc);
 966         }
 967         /* Stripe count is 1 here since this is some MDS specific stuff
 968            that is unlinked, not spanned across multiple OSTs */
 969         handle = fsfilt_start_log(obd, mds->mds_objects_dir->d_inode,
 970                                   FSFILT_OP_UNLINK, oti, 1);
 971
 972         if (IS_ERR(handle))
 973                 GOTO(out_dput, rc = PTR_ERR(handle));
 974
 975         rc = vfs_unlink(mds->mds_objects_dir->d_inode, de);
 976         if (rc)
 977                 CERROR("error destroying object "LPU64":%u: rc %d\n",
 978                        oa->o_id, oa->o_generation, rc);
 979
 980         err = fsfilt_commit(obd, mds->mds_sb, mds->mds_objects_dir->d_inode,
 981                             handle, 0);
 982         if (err && !rc)
 983                 rc = err;
 984 out_dput:
 985         if (de != NULL)
 986                 l_dput(de);
 987         up(&parent_inode->i_sem);
 988         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 989         RETURN(rc);
 990 }