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 #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 #define VIRT_FID  "virt_fid"
  55
  56 /* Add client data to the MDS.  We use a bitmap to locate a free space
  57  * in the last_rcvd file if cl_off is -1 (i.e. a new client).
  58  * Otherwise, we have just read the data from the last_rcvd file and
  59  * we know its offset.
  60  */
  61 int mds_client_add(struct obd_device *obd, struct mds_obd *mds,
  62                    struct mds_export_data *med, int cl_idx)
  63 {
  64         unsigned long *bitmap = mds->mds_client_bitmap;
  65         int new_client = (cl_idx == -1);
  66         ENTRY;
  67
  68         LASSERT(bitmap != NULL);
  69
  70         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
  71         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
  72                 RETURN(0);
  73
  74         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
  75          * there's no need for extra complication here
  76          */
  77         if (new_client) {
  78                 cl_idx = find_first_zero_bit(bitmap, MDS_MAX_CLIENTS);
  79         repeat:
  80                 if (cl_idx >= MDS_MAX_CLIENTS) {
  81                         CERROR("no room for clients - fix MDS_MAX_CLIENTS\n");
  82                         return -ENOMEM;
  83                 }
  84                 if (test_and_set_bit(cl_idx, bitmap)) {
  85                         cl_idx = find_next_zero_bit(bitmap, MDS_MAX_CLIENTS,
  86                                                     cl_idx);
  87                         goto repeat;
  88                 }
  89         } else {
  90                 if (test_and_set_bit(cl_idx, bitmap)) {
  91                         CERROR("MDS client %d: bit already set in bitmap!!\n",
  92                                cl_idx);
  93                         LBUG();
  94                 }
  95         }
  96
  97         CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
  98                cl_idx, med->med_mcd->mcd_uuid);
  99
 100         med->med_idx = cl_idx;
 101         med->med_off = le32_to_cpu(mds->mds_server_data->msd_client_start) +
 102                 (cl_idx * le16_to_cpu(mds->mds_server_data->msd_client_size));
 103
 104         if (new_client) {
 105                 struct file *file = mds->mds_rcvd_filp;
 106                 struct lvfs_run_ctxt saved;
 107                 loff_t off = med->med_off;
 108                 int rc;
 109
 110                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 111                 rc = fsfilt_write_record(obd, file, med->med_mcd,
 112                                          sizeof(*med->med_mcd), &off, 1);
 113                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 114
 115                 if (rc)
 116                         return rc;
 117                 CDEBUG(D_INFO, "wrote client mcd at idx %u off %llu (len %u)\n",
 118                        med->med_idx, med->med_off,
 119                        (unsigned int)sizeof(*med->med_mcd));
 120         }
 121         return 0;
 122 }
 123
 124 int mds_client_free(struct obd_export *exp, int clear_client)
 125 {
 126         struct mds_export_data *med = &exp->exp_mds_data;
 127         struct mds_obd *mds = &exp->exp_obd->u.mds;
 128         unsigned long *bitmap = mds->mds_client_bitmap;
 129         struct obd_device *obd = exp->exp_obd;
 130         struct mds_client_data zero_mcd;
 131         struct lvfs_run_ctxt saved;
 132         int rc;
 133
 134         if (!med->med_mcd)
 135                 RETURN(0);
 136
 137         /* XXX if mcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 138         if (!strcmp(med->med_mcd->mcd_uuid, obd->obd_uuid.uuid))
 139                 GOTO(free_and_out, 0);
 140
 141         CDEBUG(D_INFO, "freeing client at idx %u (%lld)with UUID '%s'\n",
 142                med->med_idx, med->med_off, med->med_mcd->mcd_uuid);
 143
 144         LASSERT(bitmap);
 145
 146         /* Clear the bit _after_ zeroing out the client so we don't
 147            race with mds_client_add and zero out new clients.*/
 148         if (!test_bit(med->med_idx, bitmap)) {
 149                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 150                        med->med_idx);
 151                 LBUG();
 152         }
 153
 154         if (clear_client) {
 155                 memset(&zero_mcd, 0, sizeof zero_mcd);
 156                 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 157                 rc = fsfilt_write_record(obd, mds->mds_rcvd_filp, &zero_mcd,
 158                                          sizeof(zero_mcd), &med->med_off, 1);
 159                 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 160
 161                 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 162                        "zeroing out client %s idx %u in %s rc %d\n",
 163                        med->med_mcd->mcd_uuid, med->med_idx, LAST_RCVD, rc);
 164         }
 165
 166         if (!test_and_clear_bit(med->med_idx, bitmap)) {
 167                 CERROR("MDS client %u: bit already clear in bitmap!!\n",
 168                        med->med_idx);
 169                 LBUG();
 170         }
 171
 172
 173         /* Make sure the server's last_transno is up to date. Do this
 174          * after the client is freed so we know all the client's
 175          * transactions have been committed. */
 176         mds_update_server_data(exp->exp_obd, 1);
 177
 178 free_and_out:
 179         OBD_FREE(med->med_mcd, sizeof(*med->med_mcd));
 180         return 0;
 181 }
 182
 183 static int mds_server_free_data(struct mds_obd *mds)
 184 {
 185         OBD_FREE(mds->mds_client_bitmap,
 186                  MDS_MAX_CLIENT_WORDS * sizeof(unsigned long));
 187         OBD_FREE(mds->mds_server_data, sizeof(*mds->mds_server_data));
 188         mds->mds_server_data = NULL;
 189
 190         return 0;
 191 }
 192
 193 static int mds_read_last_fid(struct obd_device *obd, struct file *file)
 194 {
 195         int rc = 0;
 196         loff_t off = 0;
 197         struct mds_obd *mds = &obd->u.mds;
 198         unsigned long last_fid_size = file->f_dentry->d_inode->i_size;
 199         ENTRY;
 200
 201         if (last_fid_size == 0) {
 202                 CWARN("%s: initializing new %s\n", obd->obd_name,
 203                       file->f_dentry->d_name.name);
 204
 205                 /*
 206                  * as fid is used for forming res_id for locking, it should not
 207                  * be zero. This will keep us out of lots possible problems,
 208                  * asserts, etc.
 209                  */
 210                 mds_set_last_fid(obd, 0);
 211         } else {
 212                 __u64 lastfid;
 213
 214                 rc = fsfilt_read_record(obd, file, &lastfid,
 215                                         sizeof(lastfid), &off);
 216                 if (rc) {
 217                         CERROR("error reading MDS %s: rc = %d\n",
 218                                file->f_dentry->d_name.name, rc);
 219                         RETURN(rc);
 220                 }
 221
 222                 /*
 223                  * make sure, that fid is up-to-date.
 224                  */
 225                 mds_set_last_fid(obd, lastfid);
 226         }
 227
 228         CDEBUG(D_INODE, "%s: server last_fid: "LPU64"\n",
 229                obd->obd_name, mds->mds_last_fid);
 230
 231         rc = mds_update_last_fid(obd, NULL, 1);
 232         RETURN(rc);
 233 }
 234
 235 static int mds_read_last_rcvd(struct obd_device *obd, struct file *file)
 236 {
 237         unsigned long last_rcvd_size = file->f_dentry->d_inode->i_size;
 238         struct mds_obd *mds = &obd->u.mds;
 239         struct mds_server_data *msd = NULL;
 240         struct mds_client_data *mcd = NULL;
 241         loff_t off = 0;
 242         __u64 mount_count;
 243         int cl_idx, rc = 0;
 244         ENTRY;
 245
 246         /* ensure padding in the struct is the correct size */
 247         LASSERT(offsetof(struct mds_server_data, msd_padding) +
 248                 sizeof(msd->msd_padding) == MDS_LR_SERVER_SIZE);
 249         LASSERT(offsetof(struct mds_client_data, mcd_padding) +
 250                 sizeof(mcd->mcd_padding) == MDS_LR_CLIENT_SIZE);
 251
 252         OBD_ALLOC_WAIT(msd, sizeof(*msd));
 253         if (!msd)
 254                 RETURN(-ENOMEM);
 255
 256         OBD_ALLOC_WAIT(mds->mds_client_bitmap,
 257                   MDS_MAX_CLIENT_WORDS * sizeof(unsigned long));
 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
 360                 /* These exports are cleaned up by mds_disconnect(), so they
 361                  * need to be set up like real exports as mds_connect() does.
 362                  */
 363                 CDEBUG(D_HA|D_WARNING,"RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 364                        " srv lr: "LPU64" lx: "LPU64"\n", mcd->mcd_uuid, cl_idx,
 365                        last_transno, le64_to_cpu(msd->msd_last_transno),
 366                        mcd->mcd_last_xid);
 367
 368                 exp = class_new_export(obd);
 369                 if (exp == NULL)
 370                         GOTO(err_client, rc = -ENOMEM);
 371
 372                 memcpy(&exp->exp_client_uuid.uuid, mcd->mcd_uuid,
 373                        sizeof exp->exp_client_uuid.uuid);
 374                 med = &exp->exp_mds_data;
 375                 med->med_mcd = mcd;
 376                 mds_client_add(obd, mds, med, cl_idx);
 377                 /* create helper if export init gets more complex */
 378                 INIT_LIST_HEAD(&med->med_open_head);
 379                 spin_lock_init(&med->med_open_lock);
 380
 381                 mcd = NULL;
 382                 exp->exp_replay_needed = 1;
 383                 obd->obd_recoverable_clients++;
 384                 obd->obd_max_recoverable_clients++;
 385                 class_export_put(exp);
 386
 387                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
 388                        cl_idx, last_transno);
 389
 390                 if (last_transno > mds->mds_last_transno)
 391                        mds->mds_last_transno = last_transno;
 392         }
 393         if (mcd)
 394                 OBD_FREE(mcd, sizeof(*mcd));
 395         obd->obd_last_committed = mds->mds_last_transno;
 396         if (obd->obd_recoverable_clients) {
 397                 CWARN("RECOVERY: service %s, %d recoverable clients, "
 398                       "last_transno "LPU64"\n", obd->obd_name,
 399                       obd->obd_recoverable_clients, mds->mds_last_transno);
 400                 obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
 401                 target_start_recovery_thread(obd, mds_handle);
 402                 obd->obd_recovery_start = LTIME_S(CURRENT_TIME);
 403         }
 404
 405         mds->mds_mount_count = mount_count + 1;
 406         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
 407
 408         /* save it, so mount count and last_transno is current */
 409         rc = mds_update_server_data(obd, 1);
 410         if (rc)
 411                 GOTO(err_client, rc);
 412
 413         RETURN(0);
 414
 415 err_client:
 416         class_disconnect_exports(obd, 0);
 417 err_msd:
 418         mds_server_free_data(mds);
 419         RETURN(rc);
 420 }
 421
 422 static int mds_fs_post_setup(struct obd_device *obd)
 423 {
 424         struct mds_obd *mds = &obd->u.mds;
 425         struct dentry *dentry;
 426         int rc = 0;
 427         ENTRY;
 428
 429         dentry = mds_id2dentry(obd, &mds->mds_rootid, NULL);
 430         if (IS_ERR(dentry)) {
 431                 CERROR("Can't find ROOT, err = %d\n",
 432                        (int)PTR_ERR(dentry));
 433                 RETURN(PTR_ERR(dentry));
 434         }
 435
 436         rc = fsfilt_post_setup(obd, dentry);
 437         if (rc)
 438                 goto out_dentry;
 439
 440         LASSERT(dentry->d_inode != NULL);
 441
 442         fsfilt_set_fs_flags(obd, dentry->d_inode,
 443                             SM_DO_REC | SM_DO_COW);
 444
 445         fsfilt_set_fs_flags(obd, mds->mds_pending_dir->d_inode,
 446                             SM_DO_REC | SM_DO_COW);
 447
 448         fsfilt_set_mds_flags(obd, mds->mds_sb);
 449
 450 out_dentry:
 451         l_dput(dentry);
 452         RETURN(rc);
 453 }
 454
 455 /*
 456  * sets up root inode lustre_id. It tries to read it first from root inode and
 457  * if it is not there, new rootid is allocated and saved there.
 458  */
 459 int mds_fs_setup_rootid(struct obd_device *obd)
 460 {
 461         int rc = 0;
 462         void *handle;
 463         struct inode *inode;
 464         struct dentry *dentry;
 465         struct mds_obd *mds = &obd->u.mds;
 466         ENTRY;
 467
 468         /* getting root directory and setup its fid. */
 469         dentry = mds_id2dentry(obd, &mds->mds_rootid, NULL);
 470         if (IS_ERR(dentry)) {
 471                 CERROR("Can't find ROOT by "DLID4", err = %d\n",
 472                        OLID4(&mds->mds_rootid), (int)PTR_ERR(dentry));
 473                 RETURN(PTR_ERR(dentry));
 474         }
 475
 476         inode = dentry->d_inode;
 477         LASSERT(dentry->d_inode);
 478
 479         rc = mds_pack_inode2id(obd, &mds->mds_rootid, inode, 1);
 480         if (rc < 0) {
 481                 if (rc != -ENODATA)
 482                         GOTO(out_dentry, rc);
 483         } else {
 484                 /*
 485                  * rootid is filled by mds_read_inode_sid(), so we do not need
 486                  * to allocate it and update. The only thing we need to check is
 487                  * mds_num.
 488                  */
 489                 LASSERT(id_group(&mds->mds_rootid) == mds->mds_num);
 490                 mds_set_last_fid(obd, id_fid(&mds->mds_rootid));
 491                 GOTO(out_dentry, rc);
 492         }
 493
 494         /* allocating new one, as it is not found in root inode. */
 495         handle = fsfilt_start(obd, inode,
 496                               FSFILT_OP_SETATTR, NULL);
 497
 498         if (IS_ERR(handle)) {
 499                 rc = PTR_ERR(handle);
 500                 CERROR("fsfilt_start() failed, rc = %d\n", rc);
 501                 GOTO(out_dentry, rc);
 502         }
 503
 504         down(&inode->i_sem);
 505         rc = mds_alloc_inode_sid(obd, inode, handle, &mds->mds_rootid);
 506         up(&inode->i_sem);
 507
 508         if (rc) {
 509                 CERROR("mds_alloc_inode_sid() failed, rc = %d\n",
 510                        rc);
 511                 GOTO(out_dentry, rc);
 512         }
 513
 514         rc = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
 515         if (rc)
 516                 CERROR("fsfilt_commit() failed, rc = %d\n", rc);
 517
 518         EXIT;
 519 out_dentry:
 520         l_dput(dentry);
 521         if (rc == 0)
 522                 CWARN("%s: rootid: "DLID4"\n", obd->obd_name,
 523                       OLID4(&mds->mds_rootid));
 524         return rc;
 525 }
 526
 527 static int mds_update_virtid_fid(struct obd_device *obd,
 528                                  void *handle, int force_sync)
 529 {
 530         struct mds_obd *mds = &obd->u.mds;
 531         struct file *filp = mds->mds_virtid_filp;
 532         struct lvfs_run_ctxt saved;
 533         loff_t off = 0;
 534         int rc = 0;
 535         ENTRY;
 536
 537         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 538         rc = fsfilt_write_record(obd, filp, &mds->mds_virtid_fid,
 539                                  sizeof(mds->mds_virtid_fid),
 540                                  &off, force_sync);
 541         if (rc) {
 542                 CERROR("error writing MDS virtid_fid #"LPU64
 543                        ", err = %d\n", mds->mds_virtid_fid, rc);
 544         }
 545
 546         CDEBUG(D_SUPER, "wrote virtid fid #"LPU64" at idx "
 547                "%llu: err = %d\n", mds->mds_virtid_fid,
 548                off, rc);
 549         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 550
 551         RETURN(rc);
 552 }
 553
 554 static int mds_read_virtid_fid(struct obd_device *obd,
 555                                struct file *file)
 556 {
 557         int rc = 0;
 558         loff_t off = 0;
 559         struct mds_obd *mds = &obd->u.mds;
 560         unsigned long virtid_fid_size = file->f_dentry->d_inode->i_size;
 561         ENTRY;
 562
 563         if (virtid_fid_size == 0) {
 564                 mds->mds_virtid_fid = mds_alloc_fid(obd);
 565         } else {
 566                 rc = fsfilt_read_record(obd, file, &mds->mds_virtid_fid,
 567                                         sizeof(mds->mds_virtid_fid), &off);
 568                 if (rc) {
 569                         CERROR("error reading MDS %s: rc = %d\n",
 570                                file->f_dentry->d_name.name, rc);
 571                         RETURN(rc);
 572                 }
 573         }
 574         rc = mds_update_virtid_fid(obd, NULL, 1);
 575
 576         RETURN(rc);
 577 }
 578
 579 /*
 580  * initializes lustre_id for virtual id directory, it is needed sometimes, as it
 581  * is possible that it will be the parent for object an operations is going to
 582  * be performed on.
 583  */
 584 int mds_fs_setup_virtid(struct obd_device *obd)
 585 {
 586         int rc = 0;
 587         void *handle;
 588         struct lustre_id sid;
 589         struct mds_obd *mds = &obd->u.mds;
 590         struct inode *inode = mds->mds_id_dir->d_inode;
 591         ENTRY;
 592
 593         handle = fsfilt_start(obd, inode,
 594                               FSFILT_OP_SETATTR, NULL);
 595
 596         if (IS_ERR(handle)) {
 597                 rc = PTR_ERR(handle);
 598                 CERROR("fsfilt_start() failed, rc = %d\n", rc);
 599                 RETURN(rc);
 600         }
 601
 602         id_group(&sid) = mds->mds_num;
 603         id_fid(&sid) = mds->mds_virtid_fid;
 604
 605         id_ino(&sid) = inode->i_ino;
 606         id_gen(&sid) = inode->i_generation;
 607         id_type(&sid) = (S_IFMT & inode->i_mode);
 608
 609         down(&inode->i_sem);
 610         rc = mds_update_inode_sid(obd, inode, handle, &sid);
 611         up(&inode->i_sem);
 612
 613         if (rc) {
 614                 CERROR("mds_update_inode_sid() failed, rc = %d\n",
 615                        rc);
 616                 RETURN(rc);
 617         }
 618
 619         rc = fsfilt_commit(obd, mds->mds_sb, inode, handle, 0);
 620         if (rc) {
 621                 CERROR("fsfilt_commit() failed, rc = %d\n", rc);
 622                 RETURN(rc);
 623         }
 624
 625         RETURN(rc);
 626 }
 627
 628 int mds_fs_setup(struct obd_device *obd, struct vfsmount *mnt)
 629 {
 630         struct mds_obd *mds = &obd->u.mds;
 631         struct lvfs_run_ctxt saved;
 632         struct dentry *dentry;
 633         struct file *file;
 634         int rc;
 635         ENTRY;
 636
 637         rc = cleanup_group_info();
 638         if (rc)
 639                 RETURN(rc);
 640
 641         mds->mds_vfsmnt = mnt;
 642         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 643
 644         fsfilt_setup(obd, mds->mds_sb);
 645
 646         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
 647         obd->obd_lvfs_ctxt.pwdmnt = mnt;
 648         obd->obd_lvfs_ctxt.pwd = mnt->mnt_root;
 649         obd->obd_lvfs_ctxt.fs = get_ds();
 650         obd->obd_lvfs_ctxt.cb_ops = mds_lvfs_ops;
 651
 652         /* setup the directory tree */
 653         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 654         dentry = simple_mkdir(current->fs->pwd, "ROOT", 0755, 0);
 655         if (IS_ERR(dentry)) {
 656                 rc = PTR_ERR(dentry);
 657                 CERROR("cannot create ROOT directory: rc = %d\n", rc);
 658                 GOTO(err_pop, rc);
 659         }
 660
 661         mdc_pack_id(&mds->mds_rootid, dentry->d_inode->i_ino,
 662                     dentry->d_inode->i_generation, S_IFDIR, 0, 0);
 663
 664         dput(dentry);
 665
 666         dentry = lookup_one_len("__iopen__", current->fs->pwd,
 667                                 strlen("__iopen__"));
 668         if (IS_ERR(dentry)) {
 669                 rc = PTR_ERR(dentry);
 670                 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
 671                 GOTO(err_pop, rc);
 672         }
 673         mds->mds_id_de = dentry;
 674
 675         if (!dentry->d_inode) {
 676                 rc = -ENOENT;
 677                 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
 678                 GOTO(err_id_de, rc);
 679         }
 680
 681         dentry = simple_mkdir(current->fs->pwd, "PENDING", 0777, 1);
 682         if (IS_ERR(dentry)) {
 683                 rc = PTR_ERR(dentry);
 684                 CERROR("cannot create PENDING directory: rc = %d\n", rc);
 685                 GOTO(err_id_de, rc);
 686         }
 687         mds->mds_pending_dir = dentry;
 688
 689         dentry = simple_mkdir(current->fs->pwd, "LOGS", 0777, 1);
 690         if (IS_ERR(dentry)) {
 691                 rc = PTR_ERR(dentry);
 692                 CERROR("cannot create LOGS directory: rc = %d\n", rc);
 693                 GOTO(err_pending, rc);
 694         }
 695         mds->mds_logs_dir = dentry;
 696
 697         dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
 698         if (IS_ERR(dentry)) {
 699                 rc = PTR_ERR(dentry);
 700                 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
 701                 GOTO(err_logs, rc);
 702         }
 703         mds->mds_objects_dir = dentry;
 704
 705         dentry = simple_mkdir(current->fs->pwd, "FIDS", 0777, 1);
 706         if (IS_ERR(dentry)) {
 707                 rc = PTR_ERR(dentry);
 708                 CERROR("cannot create FIDS directory: rc = %d\n", rc);
 709                 GOTO(err_objects, rc);
 710         }
 711         mds->mds_id_dir = dentry;
 712
 713         dentry = simple_mkdir(current->fs->pwd, "UNNAMED", 0777, 1);
 714         if (IS_ERR(dentry)) {
 715                 rc = PTR_ERR(dentry);
 716                 CERROR("cannot create UNNAMED directory: rc = %d\n", rc);
 717                 GOTO(err_unnamed, rc);
 718         }
 719         mds->mds_unnamed_dir = dentry;
 720
 721         /* open and test the last rcvd file */
 722         file = filp_open(LAST_RCVD, O_RDWR | O_CREAT, 0644);
 723         if (IS_ERR(file)) {
 724                 rc = PTR_ERR(file);
 725                 CERROR("cannot open/create %s file: rc = %d\n", LAST_RCVD, rc);
 726                 GOTO(err_id_dir, rc = PTR_ERR(file));
 727         }
 728         mds->mds_rcvd_filp = file;
 729
 730         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 731                 CERROR("%s is not a regular file!: mode = %o\n", LAST_RCVD,
 732                        file->f_dentry->d_inode->i_mode);
 733                 GOTO(err_last_rcvd, rc = -ENOENT);
 734         }
 735
 736         rc = mds_read_last_rcvd(obd, file);
 737         if (rc) {
 738                 CERROR("cannot read %s: rc = %d\n", LAST_RCVD, rc);
 739                 GOTO(err_last_rcvd, rc);
 740         }
 741
 742         /* open and test last fid file */
 743         file = filp_open(LAST_FID, O_RDWR | O_CREAT, 0644);
 744         if (IS_ERR(file)) {
 745                 rc = PTR_ERR(file);
 746                 CERROR("cannot open/create %s file: rc = %d\n",
 747                        LAST_FID, rc);
 748                 GOTO(err_client, rc = PTR_ERR(file));
 749         }
 750         mds->mds_fid_filp = file;
 751         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 752                 CERROR("%s is not a regular file!: mode = %o\n",
 753                        LAST_FID, file->f_dentry->d_inode->i_mode);
 754                 GOTO(err_last_fid, rc = -ENOENT);
 755         }
 756
 757         rc = mds_read_last_fid(obd, file);
 758         if (rc) {
 759                 CERROR("cannot read %s: rc = %d\n", LAST_FID, rc);
 760                 GOTO(err_last_fid, rc);
 761         }
 762
 763         /* open and test virtid fid file */
 764         file = filp_open(VIRT_FID, O_RDWR | O_CREAT, 0644);
 765         if (IS_ERR(file)) {
 766                 rc = PTR_ERR(file);
 767                 CERROR("cannot open/create %s file: rc = %d\n",
 768                        VIRT_FID, rc);
 769                 GOTO(err_last_fid, rc = PTR_ERR(file));
 770         }
 771         mds->mds_virtid_filp = file;
 772         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 773                 CERROR("%s is not a regular file!: mode = %o\n",
 774                        VIRT_FID, file->f_dentry->d_inode->i_mode);
 775                 GOTO(err_virtid_fid, rc = -ENOENT);
 776         }
 777
 778         rc = mds_read_virtid_fid(obd, file);
 779         if (rc) {
 780                 CERROR("cannot read %s: rc = %d\n", VIRT_FID, rc);
 781                 GOTO(err_virtid_fid, rc);
 782         }
 783
 784         /* open and test the lov objid file */
 785         file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
 786         if (IS_ERR(file)) {
 787                 rc = PTR_ERR(file);
 788                 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
 789                 GOTO(err_last_fid, rc = PTR_ERR(file));
 790         }
 791         mds->mds_lov_objid_filp = file;
 792         if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
 793                 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
 794                        file->f_dentry->d_inode->i_mode);
 795                 GOTO(err_lov_objid, rc = -ENOENT);
 796         }
 797 err_pop:
 798         if (!rc) {
 799                 rc = mds_fs_post_setup(obd);
 800                 if (rc)
 801                         CERROR("can not post setup fsfilt\n");
 802         }
 803         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 804         return rc;
 805
 806 err_lov_objid:
 807         if (mds->mds_lov_objid_filp && filp_close(mds->mds_lov_objid_filp, 0))
 808                 CERROR("can't close %s after error\n", LOV_OBJID);
 809 err_virtid_fid:
 810         if (mds->mds_virtid_filp && filp_close(mds->mds_virtid_filp, 0))
 811                 CERROR("can't close %s after error\n", VIRT_FID);
 812 err_last_fid:
 813         if (mds->mds_fid_filp && filp_close(mds->mds_fid_filp, 0))
 814                 CERROR("can't close %s after error\n", LAST_FID);
 815 err_client:
 816         class_disconnect_exports(obd, 0);
 817 err_last_rcvd:
 818         if (mds->mds_rcvd_filp && filp_close(mds->mds_rcvd_filp, 0))
 819                 CERROR("can't close %s after error\n", LAST_RCVD);
 820 err_unnamed:
 821         dput(mds->mds_unnamed_dir);
 822 err_id_dir:
 823         dput(mds->mds_id_dir);
 824 err_objects:
 825         dput(mds->mds_objects_dir);
 826 err_logs:
 827         dput(mds->mds_logs_dir);
 828 err_pending:
 829         dput(mds->mds_pending_dir);
 830 err_id_de:
 831         dput(mds->mds_id_de);
 832         goto err_pop;
 833 }
 834
 835 static int  mds_fs_post_cleanup(struct obd_device *obd)
 836 {
 837         int    rc = 0;
 838         rc = fsfilt_post_cleanup(obd);
 839         return rc;
 840 }
 841
 842 int mds_fs_cleanup(struct obd_device *obd, int flags)
 843 {
 844         struct mds_obd *mds = &obd->u.mds;
 845         struct lvfs_run_ctxt saved;
 846         int rc = 0;
 847
 848         if (flags & OBD_OPT_FAILOVER)
 849                 CERROR("%s: shutting down for failover; client state will"
 850                        " be preserved.\n", obd->obd_name);
 851
 852         class_disconnect_exports(obd, flags); /* cleans up client info too */
 853         target_cleanup_recovery(obd);
 854         mds_server_free_data(mds);
 855
 856         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 857         if (mds->mds_virtid_filp) {
 858                 rc = filp_close(mds->mds_virtid_filp, 0);
 859                 mds->mds_virtid_filp = NULL;
 860                 if (rc)
 861                         CERROR("%s file won't close, rc = %d\n", VIRT_FID, rc);
 862         }
 863         if (mds->mds_fid_filp) {
 864                 rc = filp_close(mds->mds_fid_filp, 0);
 865                 mds->mds_fid_filp = NULL;
 866                 if (rc)
 867                         CERROR("%s file won't close, rc = %d\n", LAST_FID, rc);
 868         }
 869         if (mds->mds_rcvd_filp) {
 870                 rc = filp_close(mds->mds_rcvd_filp, 0);
 871                 mds->mds_rcvd_filp = NULL;
 872                 if (rc)
 873                         CERROR("%s file won't close, rc = %d\n", LAST_RCVD, rc);
 874         }
 875         if (mds->mds_lov_objid_filp) {
 876                 rc = filp_close(mds->mds_lov_objid_filp, 0);
 877                 mds->mds_lov_objid_filp = NULL;
 878                 if (rc)
 879                         CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
 880         }
 881         if (mds->mds_unnamed_dir != NULL) {
 882                 l_dput(mds->mds_unnamed_dir);
 883                 mds->mds_unnamed_dir = NULL;
 884         }
 885         if (mds->mds_id_dir != NULL) {
 886                 l_dput(mds->mds_id_dir);
 887                 mds->mds_id_dir = NULL;
 888         }
 889         if (mds->mds_objects_dir != NULL) {
 890                 l_dput(mds->mds_objects_dir);
 891                 mds->mds_objects_dir = NULL;
 892         }
 893         if (mds->mds_logs_dir) {
 894                 l_dput(mds->mds_logs_dir);
 895                 mds->mds_logs_dir = NULL;
 896         }
 897         if (mds->mds_pending_dir) {
 898                 l_dput(mds->mds_pending_dir);
 899                 mds->mds_pending_dir = NULL;
 900         }
 901         rc = mds_fs_post_cleanup(obd);
 902
 903         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 904         shrink_dcache_parent(mds->mds_id_de);
 905         dput(mds->mds_id_de);
 906
 907         return rc;
 908 }
 909
 910 /* Creates an object with the same name as its id.  Because this is not at all
 911  * performance sensitive, it is accomplished by creating a file, checking the
 912  * id, and renaming it. */
 913 int mds_obd_create(struct obd_export *exp, struct obdo *oa,
 914                    struct lov_stripe_md **ea, struct obd_trans_info *oti)
 915 {
 916         struct mds_obd *mds = &exp->exp_obd->u.mds;
 917         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
 918         struct file *filp;
 919         struct dentry *dchild;
 920         struct lvfs_run_ctxt saved;
 921         char idname[LL_ID_NAMELEN];
 922         int rc = 0, err, idlen;
 923         void *handle;
 924         ENTRY;
 925
 926         push_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
 927         down(&parent_inode->i_sem);
 928         if (oa->o_id) {
 929                 idlen = ll_id2str(idname, oa->o_id, oa->o_generation);
 930                 dchild = lookup_one_len(idname, mds->mds_objects_dir, idlen);
 931                 if (IS_ERR(dchild))
 932                         GOTO(out_pop, rc = PTR_ERR(dchild));
 933
 934                 if (dchild->d_inode == NULL) {
 935                         struct dentry_params dp;
 936                         struct inode *inode;
 937
 938                         CWARN("creating log with ID "LPU64"\n", oa->o_id);
 939
 940                         dchild->d_fsdata = (void *) &dp;
 941                         dp.p_ptr = NULL;
 942                         dp.p_inum = oa->o_id;
 943                         rc = ll_vfs_create(parent_inode, dchild, S_IFREG, NULL);
 944                         if (dchild->d_fsdata == (void *)(unsigned long)oa->o_id)
 945                                 dchild->d_fsdata = NULL;
 946                         if (rc) {
 947                                 CDEBUG(D_INODE, "err during create: %d\n", rc);
 948                                 dput(dchild);
 949                                 GOTO(out_pop, rc);
 950                         }
 951                         inode = dchild->d_inode;
 952                         LASSERT(inode->i_ino == oa->o_id);
 953                         inode->i_generation = oa->o_generation;
 954                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
 955                                inode->i_ino, inode->i_generation);
 956                         mark_inode_dirty(inode);
 957                 } else {
 958                         CWARN("it should be here!\n");
 959                 }
 960                 GOTO(out_pop, rc);
 961         }
 962
 963         sprintf(idname, "OBJECTS/%u.%u", ll_insecure_random_int(), current->pid);
 964         filp = filp_open(idname, O_CREAT | O_EXCL, 0644);
 965         if (IS_ERR(filp)) {
 966                 rc = PTR_ERR(filp);
 967                 if (rc == -EEXIST) {
 968                         CERROR("impossible object name collision %s\n",
 969                                idname);
 970                         LBUG();
 971                 }
 972                 CERROR("error creating tmp object %s: rc %d\n",
 973                        idname, rc);
 974                 GOTO(out_pop, rc);
 975         }
 976
 977         LASSERT(mds->mds_objects_dir == filp->f_dentry->d_parent);
 978
 979         oa->o_id = filp->f_dentry->d_inode->i_ino;
 980         oa->o_generation = filp->f_dentry->d_inode->i_generation;
 981         idlen = ll_id2str(idname, oa->o_id, oa->o_generation);
 982
 983         CWARN("created log anonymous "LPU64"/%u\n",
 984               oa->o_id, oa->o_generation);
 985
 986         dchild = lookup_one_len(idname, mds->mds_objects_dir, idlen);
 987         if (IS_ERR(dchild)) {
 988                 CERROR("getting neg dentry for obj rename: %d\n", rc);
 989                 GOTO(out_close, rc = PTR_ERR(dchild));
 990         }
 991         if (dchild->d_inode != NULL) {
 992                 CERROR("impossible non-negative obj dentry " LPU64":%u!\n",
 993                        oa->o_id, oa->o_generation);
 994                 LBUG();
 995         }
 996
 997         handle = fsfilt_start(exp->exp_obd, mds->mds_objects_dir->d_inode,
 998                               FSFILT_OP_RENAME, NULL);
 999         if (IS_ERR(handle))
1000                 GOTO(out_dput, rc = PTR_ERR(handle));
1001
1002         lock_kernel();
1003         rc = vfs_rename(mds->mds_objects_dir->d_inode, filp->f_dentry,
1004                         mds->mds_objects_dir->d_inode, dchild);
1005         unlock_kernel();
1006         if (rc)
1007                 CERROR("error renaming new object "LPU64":%u: rc %d\n",
1008                        oa->o_id, oa->o_generation, rc);
1009
1010         err = fsfilt_commit(exp->exp_obd, mds->mds_sb,
1011                             mds->mds_objects_dir->d_inode, handle, 0);
1012         if (!err) {
1013                 oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
1014                 oa->o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLGROUP;
1015         } else if (!rc)
1016                 rc = err;
1017 out_dput:
1018         dput(dchild);
1019 out_close:
1020         err = filp_close(filp, 0);
1021         if (err) {
1022                 CERROR("closing tmpfile %s: rc %d\n", idname, rc);
1023                 if (!rc)
1024                         rc = err;
1025         }
1026 out_pop:
1027         up(&parent_inode->i_sem);
1028         pop_ctxt(&saved, &exp->exp_obd->obd_lvfs_ctxt, NULL);
1029         RETURN(rc);
1030 }
1031
1032 int mds_obd_destroy(struct obd_export *exp, struct obdo *oa,
1033                     struct lov_stripe_md *ea, struct obd_trans_info *oti)
1034 {
1035         struct mds_obd *mds = &exp->exp_obd->u.mds;
1036         struct inode *parent_inode = mds->mds_objects_dir->d_inode;
1037         struct obd_device *obd = exp->exp_obd;
1038         struct lvfs_run_ctxt saved;
1039         char idname[LL_ID_NAMELEN];
1040         struct dentry *de;
1041         void *handle;
1042         int err, idlen, rc = 0;
1043         ENTRY;
1044
1045         push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1046
1047         idlen = ll_id2str(idname, oa->o_id, oa->o_generation);
1048
1049         down(&parent_inode->i_sem);
1050         de = lookup_one_len(idname, mds->mds_objects_dir, idlen);
1051         if (IS_ERR(de) || de->d_inode == NULL) {
1052                 rc = IS_ERR(de) ? PTR_ERR(de) : -ENOENT;
1053                 CERROR("destroying non-existent object "LPU64" %s: rc %d\n",
1054                        oa->o_id, idname, rc);
1055                 GOTO(out_dput, rc);
1056         }
1057         /* Stripe count is 1 here since this is some MDS specific stuff
1058            that is unlinked, not spanned across multiple OSTs */
1059         handle = fsfilt_start_log(obd, mds->mds_objects_dir->d_inode,
1060                                   FSFILT_OP_UNLINK, oti, 1);
1061
1062         if (IS_ERR(handle))
1063                 GOTO(out_dput, rc = PTR_ERR(handle));
1064
1065         rc = vfs_unlink(mds->mds_objects_dir->d_inode, de);
1066         if (rc)
1067                 CERROR("error destroying object "LPU64":%u: rc %d\n",
1068                        oa->o_id, oa->o_generation, rc);
1069
1070         err = fsfilt_commit(obd, mds->mds_sb, mds->mds_objects_dir->d_inode,
1071                             handle, 0);
1072         if (err && !rc)
1073                 rc = err;
1074 out_dput:
1075         if (de != NULL)
1076                 l_dput(de);
1077         up(&parent_inode->i_sem);
1078         pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1079         RETURN(rc);
1080 }