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