lustre/obdfilter/filter.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/fs/filter/filter.c
   5  *
   6  * Copyright (C) 2001  Cluster File Systems, Inc.
   7  *
   8  * This code is issued under the GNU General Public License.
   9  * See the file COPYING in this distribution
  10  *
  11  * by Peter Braam <braam@clusterfs.com>
  12  */
  13
  14 #define EXPORT_SYMTAB
  15 #define DEBUG_SUBSYSTEM S_FILTER
  16
  17 #include <linux/module.h>
  18 #include <linux/lustre_dlm.h>
  19 #include <linux/obd_filter.h>
  20 #include <linux/ext3_jbd.h>
  21 #include <linux/quotaops.h>
  22 #include <linux/init.h>
  23
  24 extern struct obd_device obd_dev[MAX_OBD_DEVICES];
  25 long filter_memory;
  26
  27 #define FILTER_ROOTINO 2
  28
  29 #define S_SHIFT 12
  30 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
  31         [0]                     NULL,
  32         [S_IFREG >> S_SHIFT]    "R",
  33         [S_IFDIR >> S_SHIFT]    "D",
  34         [S_IFCHR >> S_SHIFT]    "C",
  35         [S_IFBLK >> S_SHIFT]    "B",
  36         [S_IFIFO >> S_SHIFT]    "F",
  37         [S_IFSOCK >> S_SHIFT]   "S",
  38         [S_IFLNK >> S_SHIFT]    "L"
  39 };
  40
  41 static inline const char *obd_mode_to_type(int mode)
  42 {
  43         return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
  44 }
  45
  46 /* write the pathname into the string */
  47 static int filter_id(char *buf, obd_id id, obd_mode mode)
  48 {
  49         return sprintf(buf, "O/%s/%Ld", obd_mode_to_type(mode),
  50                        (unsigned long long)id);
  51 }
  52
  53 /* setup the object store with correct subdirectories */
  54 static int filter_prep(struct obd_device *obddev)
  55 {
  56         struct obd_run_ctxt saved;
  57         struct filter_obd *filter = &obddev->u.filter;
  58         struct dentry *dentry;
  59         struct file *file;
  60         struct inode *inode;
  61         loff_t off;
  62         int rc = 0;
  63         char rootid[128];
  64         __u64 lastino = 2;
  65         int mode = 0;
  66
  67         push_ctxt(&saved, &filter->fo_ctxt);
  68         dentry = simple_mkdir(current->fs->pwd, "O", 0700);
  69         CDEBUG(D_INODE, "got/created O: %p\n", dentry);
  70         if (IS_ERR(dentry)) {
  71                 rc = PTR_ERR(dentry);
  72                 CERROR("cannot open/create O: rc = %d\n", rc);
  73                 GOTO(out, rc);
  74         }
  75         filter->fo_dentry_O = dentry;
  76         dentry = simple_mkdir(current->fs->pwd, "P", 0700);
  77         CDEBUG(D_INODE, "got/created P: %p\n", dentry);
  78         if (IS_ERR(dentry)) {
  79                 rc = PTR_ERR(dentry);
  80                 CERROR("cannot open/create P: rc = %d\n", rc);
  81                 GOTO(out_O, rc);
  82         }
  83         CDEBUG(D_INODE, "putting P: %p, count = %d\n", dentry,
  84                atomic_read(&dentry->d_count) - 1);
  85         dput(dentry);
  86         dentry = simple_mkdir(current->fs->pwd, "D", 0700);
  87         CDEBUG(D_INODE, "got/created D: %p\n", dentry);
  88         if (IS_ERR(dentry)) {
  89                 rc = PTR_ERR(dentry);
  90                 CERROR("cannot open/create D: rc = %d\n", rc);
  91                 GOTO(out_O, rc);
  92         }
  93         CDEBUG(D_INODE, "putting D: %p, count = %d\n", dentry,
  94                atomic_read(&dentry->d_count) - 1);
  95         dput(dentry);
  96
  97         /*
  98          * Create directories and/or get dentries for each object type.
  99          * This saves us from having to do multiple lookups for each one.
 100          */
 101         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 102                 char *type = obd_type_by_mode[mode];
 103
 104                 if (!type) {
 105                         filter->fo_dentry_O_mode[mode] = NULL;
 106                         continue;
 107                 }
 108                 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
 109                 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
 110                 if (IS_ERR(dentry)) {
 111                         rc = PTR_ERR(dentry);
 112                         CERROR("cannot create O/%s: rc = %d\n", type, rc);
 113                         GOTO(out_O_mode, rc);
 114                 }
 115                 filter->fo_dentry_O_mode[mode] = dentry;
 116         }
 117
 118         filter_id(rootid, FILTER_ROOTINO, S_IFDIR);
 119         file = filp_open(rootid, O_RDWR | O_CREAT, 0755);
 120         if (IS_ERR(file)) {
 121                 rc = PTR_ERR(file);
 122                 CERROR("OBD filter: cannot open/create root %s: rc = %d\n",
 123                        rootid, rc);
 124                 GOTO(out_O_mode, rc);
 125         }
 126         filp_close(file, 0);
 127
 128         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 129         if ( !file || IS_ERR(file) ) {
 130                 rc = PTR_ERR(file);
 131                 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
 132                        "D/status", rc);
 133                 GOTO(out_O_mode, rc);
 134         }
 135
 136         /* steal operations */
 137         inode = file->f_dentry->d_inode;
 138         filter->fo_fop = file->f_op;
 139         filter->fo_iop = inode->i_op;
 140         filter->fo_aops = inode->i_mapping->a_ops;
 141
 142         off = 0;
 143         if (inode->i_size == 0) {
 144                 ssize_t retval = file->f_op->write(file, (char *)&lastino,
 145                                                    sizeof(lastino), &off);
 146                 if (retval != sizeof(lastino)) {
 147                         CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
 148                         filp_close(file, 0);
 149                         GOTO(out_O_mode, rc = -EIO);
 150                 }
 151         } else {
 152                 ssize_t retval = file->f_op->read(file, (char *)&lastino,
 153                                                   sizeof(lastino), &off);
 154                 if (retval != sizeof(lastino)) {
 155                         CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
 156                         filp_close(file, 0);
 157                         GOTO(out_O_mode, rc = -EIO);
 158                 }
 159         }
 160         filter->fo_lastino = lastino;
 161         filp_close(file, 0);
 162
 163         rc = 0;
 164  out:
 165         pop_ctxt(&saved);
 166
 167         return(rc);
 168
 169 out_O_mode:
 170         while (mode-- > 0) {
 171                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 172                 if (dentry) {
 173                         CDEBUG(D_INODE, "putting O/%s: %p, count = %d\n",
 174                                obd_type_by_mode[mode], dentry,
 175                                atomic_read(&dentry->d_count) - 1);
 176                         dput(dentry);
 177                         filter->fo_dentry_O_mode[mode] = NULL;
 178                 }
 179         }
 180 out_O:
 181         CDEBUG(D_INODE, "putting O: %p, count = %d\n", filter->fo_dentry_O,
 182                atomic_read(&filter->fo_dentry_O->d_count) - 1);
 183         dput(filter->fo_dentry_O);
 184         filter->fo_dentry_O = NULL;
 185         goto out;
 186 }
 187
 188 /* cleanup the filter: write last used object id to status file */
 189 static void filter_post(struct obd_device *obddev)
 190 {
 191         struct obd_run_ctxt saved;
 192         struct filter_obd *filter = &obddev->u.filter;
 193         long rc;
 194         struct file *file;
 195         loff_t off = 0;
 196         int mode;
 197
 198         push_ctxt(&saved, &filter->fo_ctxt);
 199         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 200         if (IS_ERR(file)) {
 201                 CERROR("OBD filter: cannot create status file\n");
 202                 goto out;
 203         }
 204         rc = file->f_op->write(file, (char *)&filter->fo_lastino,
 205                        sizeof(filter->fo_lastino), &off);
 206         if (rc != sizeof(filter->fo_lastino))
 207                 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
 208
 209         rc = filp_close(file, NULL);
 210         if (rc)
 211                 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
 212
 213         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 214                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 215                 if (dentry) {
 216                         CDEBUG(D_INODE, "putting O/%s: %p, count = %d\n",
 217                                obd_type_by_mode[mode], dentry,
 218                                atomic_read(&dentry->d_count) - 1);
 219                         dput(dentry);
 220                         filter->fo_dentry_O_mode[mode] = NULL;
 221                 }
 222         }
 223         CDEBUG(D_INODE, "putting O: %p, count = %d\n", filter->fo_dentry_O,
 224                atomic_read(&filter->fo_dentry_O->d_count) - 1);
 225         dput(filter->fo_dentry_O);
 226 out:
 227         pop_ctxt(&saved);
 228 }
 229
 230
 231 static __u64 filter_next_id(struct obd_device *obddev)
 232 {
 233         __u64 id;
 234         spin_lock(&obddev->u.filter.fo_lock);
 235         obddev->u.filter.fo_lastino++;
 236         id =    obddev->u.filter.fo_lastino;
 237         spin_unlock(&obddev->u.filter.fo_lock);
 238         return id;
 239 }
 240
 241 /* how to get files, dentries, inodes from object id's */
 242 /* parent i_sem is already held if needed for exclusivity */
 243 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
 244                                         struct dentry *dparent,
 245                                         __u64 id, __u32 type)
 246 {
 247         struct super_block *sb = obddev->u.filter.fo_sb;
 248         struct dentry *dchild;
 249         char name[32];
 250         int len;
 251         ENTRY;
 252
 253         if (!sb || !sb->s_dev) {
 254                 CERROR("fatal: device not initialized.\n");
 255                 RETURN(ERR_PTR(-ENXIO));
 256         }
 257
 258         if (id == 0) {
 259                 CERROR("fatal: invalid object #0\n");
 260                 LBUG();
 261                 RETURN(ERR_PTR(-ESTALE));
 262         }
 263
 264         if (!(type & S_IFMT)) {
 265                 CERROR("OBD %s, object %Lu has bad type: %o\n", __FUNCTION__,
 266                        (unsigned long long)id, type);
 267                 RETURN(ERR_PTR(-EINVAL));
 268         }
 269
 270         len = sprintf(name, "%Ld", id);
 271         CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
 272                name);
 273         dchild = lookup_one_len(name, dparent, len);
 274         CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
 275                obd_mode_to_type(type), name, dchild,
 276                atomic_read(&dchild->d_count));
 277
 278         if (IS_ERR(dchild)) {
 279                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 280                 RETURN(dchild);
 281         }
 282
 283         RETURN(dchild);
 284 }
 285
 286 static struct file *filter_obj_open(struct obd_device *obddev,
 287                                     __u64 id, __u32 type)
 288 {
 289         struct super_block *sb = obddev->u.filter.fo_sb;
 290         struct obd_run_ctxt saved;
 291         char name[24];
 292         struct file *file;
 293         ENTRY;
 294
 295         if (!sb || !sb->s_dev) {
 296                 CERROR("fatal: device not initialized.\n");
 297                 RETURN(ERR_PTR(-ENXIO));
 298         }
 299
 300         if (!id) {
 301                 CERROR("fatal: invalid obdo %Lu\n", (unsigned long long)id);
 302                 RETURN(ERR_PTR(-ESTALE));
 303         }
 304
 305         if (!(type & S_IFMT)) {
 306                 CERROR("OBD %s, no type (%Ld), mode %o!\n", __FUNCTION__,
 307                        (unsigned long long)id, type);
 308                 RETURN(ERR_PTR(-EINVAL));
 309         }
 310
 311         filter_id(name, id, type);
 312         push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
 313         file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
 314         pop_ctxt(&saved);
 315
 316         CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
 317
 318         if (IS_ERR(file))
 319                 file = NULL;
 320         RETURN(file);
 321 }
 322
 323 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
 324 {
 325         struct filter_obd *filter = &obddev->u.filter;
 326
 327         return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
 328 }
 329
 330
 331 static struct inode *filter_inode_from_obj(struct obd_device *obddev,
 332                                            __u64 id, __u32 type)
 333 {
 334         struct dentry *dentry;
 335         struct inode *inode;
 336
 337         dentry = filter_fid2dentry(obddev, filter_parent(obddev, type),
 338                                    id, type);
 339         if (IS_ERR(dentry)) {
 340                 CERROR("%s: lookup failed: rc = %ld\n", __FUNCTION__,
 341                        PTR_ERR(dentry));
 342                 RETURN(NULL);
 343         }
 344
 345         lock_kernel();
 346         inode = iget(dentry->d_inode->i_sb, dentry->d_inode->i_ino);
 347         unlock_kernel();
 348         CDEBUG(D_INODE, "put child %p, count = %d\n", dentry,
 349                atomic_read(&dentry->d_count) - 1);
 350         dput(dentry);
 351         CDEBUG(D_INODE, "got inode %p (%ld), count = %d\n", inode, inode->i_ino,
 352                atomic_read(&inode->i_count));
 353         return inode;
 354 }
 355
 356 /* obd methods */
 357 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
 358                           char *cluuid)
 359 {
 360         int rc;
 361         ENTRY;
 362         MOD_INC_USE_COUNT;
 363         rc = class_connect(conn, obd, cluuid);
 364         if (rc)
 365                 MOD_DEC_USE_COUNT;
 366         RETURN(rc);
 367 }
 368
 369 static int filter_disconnect(struct lustre_handle *conn)
 370 {
 371         int rc;
 372         ENTRY;
 373
 374         rc = class_disconnect(conn);
 375         if (!rc)
 376                 MOD_DEC_USE_COUNT;
 377
 378         /* XXX cleanup preallocated inodes */
 379         RETURN(rc);
 380 }
 381
 382 /* mount the file system (secretly) */
 383 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
 384 {
 385         struct obd_ioctl_data* data = buf;
 386         struct filter_obd *filter;
 387         struct vfsmount *mnt;
 388         int err = 0;
 389         ENTRY;
 390
 391         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 392                 RETURN(-EINVAL);
 393
 394         MOD_INC_USE_COUNT;
 395         mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
 396         err = PTR_ERR(mnt);
 397         if (IS_ERR(mnt))
 398                 GOTO(err_dec, err);
 399
 400         filter = &obddev->u.filter;;
 401         filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
 402         /* XXX is this even possible if do_kern_mount succeeded? */
 403         if (!filter->fo_sb)
 404                 GOTO(err_put, err = -ENODEV);
 405
 406         filter->fo_vfsmnt = mnt;
 407         filter->fo_fstype = strdup(data->ioc_inlbuf2);
 408
 409         OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
 410         filter->fo_ctxt.pwdmnt = mnt;
 411         filter->fo_ctxt.pwd = mnt->mnt_root;
 412         filter->fo_ctxt.fs = get_ds();
 413
 414         err = filter_prep(obddev);
 415         if (err)
 416                 GOTO(err_kfree, err);
 417         spin_lock_init(&filter->fo_lock);
 418
 419         obddev->obd_namespace =
 420                 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
 421         if (obddev->obd_namespace == NULL)
 422                 LBUG();
 423
 424         RETURN(0);
 425
 426 err_kfree:
 427         kfree(filter->fo_fstype);
 428 err_put:
 429         unlock_kernel();
 430         mntput(filter->fo_vfsmnt);
 431         filter->fo_sb = 0;
 432         lock_kernel();
 433
 434 err_dec:
 435         MOD_DEC_USE_COUNT;
 436         return err;
 437 }
 438
 439
 440 static int filter_cleanup(struct obd_device * obddev)
 441 {
 442         struct super_block *sb;
 443         ENTRY;
 444
 445         if (!(obddev->obd_flags & OBD_SET_UP))
 446                 RETURN(0);
 447
 448         if (!list_empty(&obddev->obd_exports)) {
 449                 CERROR("still has clients!\n");
 450                 RETURN(-EBUSY);
 451         }
 452
 453         ldlm_namespace_free(obddev->obd_namespace);
 454
 455         sb = obddev->u.filter.fo_sb;
 456         if (!obddev->u.filter.fo_sb)
 457                 RETURN(0);
 458
 459         filter_post(obddev);
 460
 461         shrink_dcache_parent(sb->s_root);
 462         unlock_kernel();
 463         mntput(obddev->u.filter.fo_vfsmnt);
 464         obddev->u.filter.fo_sb = 0;
 465         kfree(obddev->u.filter.fo_fstype);
 466
 467         lock_kernel();
 468
 469         MOD_DEC_USE_COUNT;
 470         RETURN(0);
 471 }
 472
 473
 474 static inline void filter_from_inode(struct obdo *oa, struct inode *inode)
 475 {
 476         int type = oa->o_mode & S_IFMT;
 477         ENTRY;
 478
 479         CDEBUG(D_INFO, "src inode %ld, dst obdo %ld valid 0x%08x\n",
 480                inode->i_ino, (long)oa->o_id, oa->o_valid);
 481         obdo_from_inode(oa, inode);
 482         oa->o_mode &= ~S_IFMT;
 483         oa->o_mode |= type;
 484
 485         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
 486                 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
 487                 oa->o_rdev = rdev;
 488                 oa->o_valid |= OBD_MD_FLRDEV;
 489         }
 490
 491         EXIT;
 492 }
 493
 494 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
 495                           struct lov_stripe_md *md)
 496 {
 497         struct obd_device *obddev = class_conn2obd(conn);
 498         struct dentry *dentry;
 499         ENTRY;
 500
 501         if (!class_conn2export(conn)) {
 502                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 503                 RETURN(-EINVAL);
 504         }
 505
 506         obddev = class_conn2obd(conn);
 507         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 508                                    oa->o_id, oa->o_mode);
 509         if (IS_ERR(dentry))
 510                 RETURN(PTR_ERR(dentry));
 511
 512         oa->o_valid &= ~OBD_MD_FLID;
 513         filter_from_inode(oa, dentry->d_inode);
 514
 515         dput(dentry);
 516         RETURN(0);
 517 }
 518
 519 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
 520                           struct lov_stripe_md *md)
 521 {
 522         struct obd_run_ctxt saved;
 523         struct obd_device *obd = class_conn2obd(conn);
 524         struct dentry *dentry;
 525         struct iattr iattr;
 526         struct inode *inode;
 527         int rc;
 528         ENTRY;
 529
 530         iattr_from_obdo(&iattr, oa);
 531         iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
 532         dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
 533                                    oa->o_id, iattr.ia_mode);
 534         if (IS_ERR(dentry))
 535                 RETURN(PTR_ERR(dentry));
 536
 537         inode = dentry->d_inode;
 538         lock_kernel();
 539         if (iattr.ia_mode & ATTR_SIZE)
 540                 down(&inode->i_sem);
 541         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 542         if (inode->i_op->setattr)
 543                 rc = inode->i_op->setattr(dentry, &iattr);
 544         else
 545                 rc = inode_setattr(inode, &iattr);
 546         pop_ctxt(&saved);
 547         if (iattr.ia_mode & ATTR_SIZE) {
 548                 up(&inode->i_sem);
 549                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 550                 obdo_from_inode(oa, inode);
 551         }
 552         unlock_kernel();
 553
 554         CDEBUG(D_INODE, "put dentry %p, count = %d\n", inode,
 555                atomic_read(&dentry->d_count) - 1);
 556         dput(dentry);
 557         RETURN(rc);
 558 }
 559
 560 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
 561                           struct lov_stripe_md *ea)
 562 {
 563         struct obd_device *obd;
 564         struct dentry *dentry;
 565         /* ENTRY; */
 566
 567         if (!class_conn2export(conn)) {
 568                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 569                 RETURN(-EINVAL);
 570         }
 571
 572         obd = class_conn2obd(conn);
 573         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 574                                    oa->o_id, oa->o_mode);
 575         if (IS_ERR(dentry))
 576                 RETURN(PTR_ERR(dentry));
 577
 578         return 0;
 579 } /* filter_open */
 580
 581 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
 582                           struct lov_stripe_md *ea)
 583 {
 584         struct obd_device *obd;
 585         struct dentry *dentry;
 586         /* ENTRY; */
 587
 588         if (!class_conn2export(conn)) {
 589                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 590                 RETURN(-EINVAL);
 591         }
 592
 593         obd = class_conn2obd(conn);
 594         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 595                                    oa->o_id, oa->o_mode);
 596         if (IS_ERR(dentry))
 597                 RETURN(PTR_ERR(dentry));
 598
 599         CDEBUG(D_INODE, "put dentry %p, count = %d\n", dentry,
 600                atomic_read(&dentry->d_count) - 1);
 601         dput(dentry);  /* for the close */
 602         CDEBUG(D_INODE, "put dentry %p, count = %d\n", dentry,
 603                atomic_read(&dentry->d_count) - 1);
 604         dput(dentry);  /* for this call */
 605         return 0;
 606 } /* filter_close */
 607
 608 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
 609                          struct lov_stripe_md **ea)
 610 {
 611         char name[64];
 612         struct obd_run_ctxt saved;
 613         struct file *file;
 614         int mode;
 615         struct obd_device *obd = class_conn2obd(conn);
 616         struct iattr;
 617         ENTRY;
 618
 619         if (!class_conn2export(conn)) {
 620                 CERROR("invalid client %Lx\n", conn->addr);
 621                 return -EINVAL;
 622         }
 623
 624         oa->o_id = filter_next_id(obd);
 625         if (!(oa->o_mode && S_IFMT)) {
 626                 CERROR("filter obd: no type!\n");
 627                 return -ENOENT;
 628         }
 629
 630         filter_id(name, oa->o_id, oa->o_mode);
 631         mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
 632         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 633         file = filp_open(name, O_RDONLY | O_CREAT, mode);
 634         pop_ctxt(&saved);
 635         if (IS_ERR(file)) {
 636                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(file));
 637                 return -ENOENT;
 638         }
 639         filp_close(file, 0);
 640
 641         /* Set flags for fields we have set in the inode struct */
 642         oa->o_valid |= OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 643                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME |
 644                  OBD_MD_FLUID | OBD_MD_FLGID;
 645
 646         /* XXX Hmm, shouldn't we copy the fields into the obdo here? */
 647         return 0;
 648 }
 649
 650 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
 651                           struct lov_stripe_md *ea)
 652 {
 653         struct obd_device *obd;
 654         struct filter_obd *filter;
 655         struct obd_run_ctxt saved;
 656         struct obd_export *export;
 657         struct inode *inode;
 658         struct dentry *dir_dentry, *object_dentry;
 659         int rc;
 660         ENTRY;
 661
 662         if (!(export = class_conn2export(conn))) {
 663                 CERROR("invalid client %Lx\n", conn->addr);
 664                 RETURN(-EINVAL);
 665         }
 666
 667         CDEBUG(D_INODE, "destroying object %Ld\n", oa->o_id);
 668         obd = class_conn2obd(conn);
 669
 670         dir_dentry = filter_parent(obd, oa->o_mode);
 671         down(&dir_dentry->d_inode->i_sem);
 672
 673         object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
 674                                           oa->o_mode);
 675         if (IS_ERR(object_dentry))
 676                 GOTO(out, rc = -ENOENT);
 677
 678         inode = object_dentry->d_inode;
 679         if (inode == NULL) {
 680                 CERROR("trying to destroy negative inode %Ld!\n", oa->o_id);
 681                 GOTO(out, rc = -ENOENT);
 682         }
 683
 684         if (inode->i_nlink != 1) {
 685                 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
 686                 LBUG();
 687                 inode->i_nlink = 1;
 688         }
 689         inode->i_mode = S_IFREG;
 690
 691         filter = &obd->u.filter;
 692         push_ctxt(&saved, &filter->fo_ctxt);
 693
 694         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 695         pop_ctxt(&saved);
 696         CDEBUG(D_INODE, "put child %p, count = %d\n", object_dentry,
 697                atomic_read(&object_dentry->d_count) - 1);
 698         dput(object_dentry);
 699
 700         EXIT;
 701 out:
 702         up(&dir_dentry->d_inode->i_sem);
 703         return rc;
 704 }
 705
 706 /* NB count and offset are used for punch, but not truncate */
 707 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
 708                            struct lov_stripe_md *md,
 709                            obd_off start, obd_off end)
 710 {
 711         int error;
 712         ENTRY;
 713
 714         if ( end != 0xffffffffffffffff ) {
 715                 CERROR("PUNCH not supported, only truncate works\n");
 716         }
 717
 718         CDEBUG(D_INODE, "calling truncate for object #%Ld, valid = %x, "
 719                "o_size = %Ld\n", oa->o_id, oa->o_valid, start);
 720         oa->o_size = start;
 721         error = filter_setattr(conn, oa, NULL);
 722         RETURN(error);
 723 }
 724
 725 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
 726                                struct lov_stripe_md *md, obd_count oa_bufs,
 727                                struct page **pages, obd_size *count,
 728                                obd_off *offset, obd_flag *flags,
 729                                brw_callback_t callback, void *data)
 730 {
 731         struct obd_run_ctxt      saved;
 732         struct super_block      *sb;
 733         int                      pnum;          /* index to pages (bufs) */
 734         unsigned long            retval;
 735         int                      error;
 736         struct file             *file;
 737         struct obd_device      *obd = class_conn2obd(conn);
 738         int pg;
 739         ENTRY;
 740
 741         if (!class_conn2export(conn)) {
 742                 CDEBUG(D_IOCTL, "invalid client %Lx\n", conn->addr);
 743                 RETURN(-EINVAL);
 744         }
 745
 746         sb = obd->u.filter.fo_sb;
 747         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 748         pnum = 0; /* pnum indexes buf 0..num_pages */
 749
 750         file = filter_obj_open(obd, md->lmd_object_id, S_IFREG);
 751         if (IS_ERR(file))
 752                 GOTO(out, retval = PTR_ERR(file));
 753
 754         /* count doubles as retval */
 755         for (pg = 0; pg < oa_bufs; pg++) {
 756                 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,%ld) "
 757                        "off count (%Ld,%Ld)\n",
 758                        cmd, pnum, file->f_dentry->d_inode->i_ino,
 759                        (unsigned long)offset[pnum] >> PAGE_CACHE_SHIFT,
 760                        (unsigned long long)offset[pnum],
 761                        (unsigned long long)count[pnum]);
 762                 if (cmd & OBD_BRW_WRITE) {
 763                         loff_t off;
 764                         char *buffer;
 765                         off = offset[pnum];
 766                         buffer = kmap(pages[pnum]);
 767                         retval = file->f_op->write(file, buffer, count[pnum],
 768                                                    &off);
 769                         kunmap(pages[pnum]);
 770                         CDEBUG(D_INODE, "retval %ld\n", retval);
 771                 } else {
 772                         loff_t off = offset[pnum];
 773                         char *buffer = kmap(pages[pnum]);
 774
 775                         if (off >= file->f_dentry->d_inode->i_size) {
 776                                 memset(buffer, 0, count[pnum]);
 777                                 retval = count[pnum];
 778                         } else {
 779                                 retval = file->f_op->read(file, buffer,
 780                                                           count[pnum], &off);
 781                         }
 782                         kunmap(pages[pnum]);
 783
 784                         if (retval != count[pnum]) {
 785                                 filp_close(file, 0);
 786                                 GOTO(out, retval = -EIO);
 787                         }
 788                         CDEBUG(D_INODE, "retval %ld\n", retval);
 789                 }
 790                 pnum++;
 791         }
 792         /* sizes and blocks are set by generic_file_write */
 793         /* ctimes/mtimes will follow with a setattr call */
 794         filp_close(file, 0);
 795
 796         /* XXX: do something with callback if it is set? */
 797
 798         EXIT;
 799 out:
 800         pop_ctxt(&saved);
 801         error = (retval >= 0) ? 0 : retval;
 802         return error;
 803 }
 804
 805
 806 struct inode *ioobj_to_inode(struct lustre_handle *conn, struct obd_ioobj *o)
 807 {
 808         struct obd_device *obd = class_conn2obd(conn);
 809         struct super_block *sb = obd->u.filter.fo_sb;
 810         struct inode *inode = NULL;
 811         ENTRY;
 812
 813         if (!sb || !sb->s_dev) {
 814                 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
 815                 RETURN(NULL);
 816         }
 817
 818         if ( !o->ioo_id ) {
 819                 CDEBUG(D_INODE, "fatal: invalid obdo %lu\n", (long)o->ioo_id);
 820                 RETURN(NULL);
 821         }
 822
 823         inode = filter_inode_from_obj(obd, o->ioo_id, S_IFREG);
 824         if (!inode || inode->i_nlink == 0 || is_bad_inode(inode)) {
 825                 CERROR("from obdo - fatal: invalid inode %ld (%s).\n",
 826                        (long)o->ioo_id, inode ? inode->i_nlink ? "bad inode" :
 827                        "no links" : "NULL");
 828                 iput(inode);
 829                 RETURN(NULL);
 830         }
 831
 832         RETURN(inode);
 833 }
 834
 835 /*
 836  * Calculate the number of buffer credits needed to write multiple pages in
 837  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
 838  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
 839  *
 840  * See comment above ext3_writepage_trans_blocks for details.  We assume
 841  * no data journaling is being done, but it does allow for all of the pages
 842  * being non-contiguous.  If we are guaranteed contiguous pages we could
 843  * reduce the number of (d)indirect blocks a lot.
 844  *
 845  * With N blocks per page and P pages, for each inode we have at most:
 846  * N*P indirect
 847  * min(N*P, blocksize/4 + 1) dindirect blocks
 848  * 1 tindirect
 849  *
 850  * For the entire filesystem, we have at most:
 851  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 852  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 853  * 1 inode block
 854  * 1 superblock
 855  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 856  */
 857 static int ext3_credits_needed(struct super_block *sb, int objcount,
 858                                struct obd_ioobj *obj)
 859 {
 860         struct obd_ioobj *o = obj;
 861         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 862         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 863         int nbitmaps = 0;
 864         int ngdblocks = 0;
 865         int needed = objcount + 1;
 866         int i;
 867
 868         for (i = 0; i < objcount; i++, o++) {
 869                 int nblocks = o->ioo_bufcnt * blockpp;
 870                 int ndindirect = min(nblocks, addrpp + 1);
 871                 int nindir = nblocks + ndindirect + 1;
 872
 873                 nbitmaps += nindir + nblocks;
 874                 ngdblocks += nindir + nblocks;
 875
 876                 needed += nindir;
 877         }
 878
 879         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 880                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 881         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 882                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 883
 884         needed += nbitmaps + ngdblocks;
 885
 886 #ifdef CONFIG_QUOTA
 887         /* We assume that there will be 1 bit set in s_dquot.flags for each
 888          * quota file that is active.  This is at least true for now.
 889          */
 890         needed += hweight32(sb_any_quota_enabled(sb)) *
 891                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 892 #endif
 893
 894         return needed;
 895 }
 896
 897 /* We have to start a huge journal transaction here to hold all of the
 898  * metadata for the pages being written here.  This is necessitated by
 899  * the fact that we do lots of prepare_write operations before we do
 900  * any of the matching commit_write operations, so even if we split
 901  * up to use "smaller" transactions none of them could complete until
 902  * all of them were opened.  By having a single journal transaction,
 903  * we eliminate duplicate reservations for common blocks like the
 904  * superblock and group descriptors or bitmaps.
 905  *
 906  * We will start the transaction here, but each prepare_write will
 907  * add a refcount to the transaction, and each commit_write will
 908  * remove a refcount.  The transaction will be closed when all of
 909  * the pages have been written.
 910  */
 911 static void *ext3_filter_journal_start(struct filter_obd *filter,
 912                                        int objcount, struct obd_ioobj *obj,
 913                                        int niocount, struct niobuf_remote *nb)
 914 {
 915         journal_t *journal = NULL;
 916         handle_t *handle = NULL;
 917         int needed;
 918
 919         /* Assumes ext3 and extN have same sb_info layout, but avoids issues
 920          * with having extN built properly before filterobd for now.
 921          */
 922         journal = EXT3_SB(filter->fo_sb)->s_journal;
 923         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
 924
 925         /* The number of blocks we could _possibly_ dirty can very large.
 926          * We reduce our request if it is absurd (and we couldn't get that
 927          * many credits for a single handle anyways).
 928          *
 929          * At some point we have to limit the size of I/Os sent at one time,
 930          * increase the size of the journal, or we have to calculate the
 931          * actual journal requirements more carefully by checking all of
 932          * the blocks instead of being maximally pessimistic.  It remains to
 933          * be seen if this is a real problem or not.
 934          */
 935         if (needed > journal->j_max_transaction_buffers) {
 936                 CERROR("want too many journal credits (%d) using %d instead\n",
 937                        needed, journal->j_max_transaction_buffers);
 938                 needed = journal->j_max_transaction_buffers;
 939         }
 940
 941         handle = journal_start(journal, needed);
 942         if (IS_ERR(handle))
 943                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 944                        PTR_ERR(handle));
 945
 946         return(handle);
 947 }
 948
 949 static void *filter_journal_start(void **journal_save,
 950                                   struct filter_obd *filter,
 951                                   int objcount, struct obd_ioobj *obj,
 952                                   int niocount, struct niobuf_remote *nb)
 953 {
 954         void *handle = NULL;
 955
 956         /* This may not be necessary - we probably never have a
 957          * transaction started when we enter here, so we can
 958          * remove the saving of the journal state entirely.
 959          * For now leave it in just to see if it ever happens.
 960          */
 961         *journal_save = current->journal_info;
 962         if (*journal_save) {
 963                 CERROR("Already have handle %p???\n", *journal_save);
 964                 LBUG();
 965                 current->journal_info = NULL;
 966         }
 967
 968         if (!strcmp(filter->fo_fstype, "ext3") ||
 969             !strcmp(filter->fo_fstype, "extN"))
 970                 handle = ext3_filter_journal_start(filter, objcount, obj,
 971                                                    niocount, nb);
 972         return handle;
 973 }
 974
 975 static int ext3_filter_journal_stop(void *handle)
 976 {
 977         int rc;
 978
 979         /* We got a refcount on the handle for each call to prepare_write,
 980          * so we can drop the "parent" handle here to avoid the need for
 981          * osc to call back into filterobd to close the handle.  The
 982          * remaining references will be dropped in commit_write.
 983          */
 984         rc = journal_stop((handle_t *)handle);
 985
 986         return rc;
 987 }
 988
 989 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
 990                                void *handle)
 991 {
 992         int rc = 0;
 993
 994         if (!strcmp(filter->fo_fstype, "ext3") ||
 995             !strcmp(filter->fo_fstype, "extN"))
 996                 rc = ext3_filter_journal_stop(handle);
 997
 998         if (rc)
 999                 CERROR("error on journal stop: rc = %d\n", rc);
1000
1001         current->journal_info = journal_save;
1002
1003         return rc;
1004 }
1005
1006 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1007                                    struct niobuf_local *lnb)
1008 {
1009         struct address_space *mapping = inode->i_mapping;
1010         struct page *page;
1011         int rc;
1012
1013         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1014         page = grab_cache_page_nowait(mapping, index); /* locked page */
1015
1016         /* This page is currently locked, so get a temporary page instead */
1017         if (!page) {
1018                 unsigned long addr;
1019                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1020                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1021                 if (!addr) {
1022                         CERROR("no memory for a temp page\n");
1023                         LBUG();
1024                         GOTO(err, rc = -ENOMEM);
1025                 }
1026                 page = virt_to_page(addr);
1027                 kmap(page);
1028                 page->index = index;
1029                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1030         } else if (!IS_ERR(page)) {
1031                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1032                  * a no-op for most filesystems, because we write the whole
1033                  * page.  For partial-page I/O this will read in the page.
1034                  */
1035                 kmap(page);
1036                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1037                 if (rc) {
1038                         CERROR("page index %lu, rc = %d\n", index, rc);
1039                         if (rc != -ENOSPC)
1040                                 LBUG();
1041                         GOTO(err_unlock, rc);
1042                 }
1043                 /* XXX not sure if we need this if we are overwriting page */
1044                 if (PageError(page)) {
1045                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1046                         LBUG();
1047                         GOTO(err_unlock, rc = -EIO);
1048                 }
1049         }
1050         return page;
1051
1052 err_unlock:
1053         unlock_page(page);
1054         lustre_put_page(page);
1055 err:
1056         return ERR_PTR(rc);
1057 }
1058
1059 /*
1060  * We need to balance prepare_write() calls with commit_write() calls.
1061  * If the page has been prepared, but we have no data for it, we don't
1062  * want to overwrite valid data on disk, but we still need to zero out
1063  * data for space which was newly allocated.  Like part of what happens
1064  * in __block_prepare_write() for newly allocated blocks.
1065  *
1066  * XXX currently __block_prepare_write() creates buffers for all the
1067  *     pages, and the filesystems mark these buffers as BH_New if they
1068  *     were newly allocated from disk. We use the BH_New flag similarly.
1069  */
1070 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1071                                int err)
1072 {
1073         if (err) {
1074                 unsigned block_start, block_end;
1075                 struct buffer_head *bh, *head = page->buffers;
1076                 unsigned blocksize = head->b_size;
1077                 void *addr = page_address(page);
1078
1079                 /* Currently one buffer per page, but in the future... */
1080                 for (bh = head, block_start = 0; bh != head || !block_start;
1081                      block_start = block_end, bh = bh->b_this_page) {
1082                         block_end = block_start + blocksize;
1083                         if (buffer_new(bh))
1084                                 memset(addr + block_start, 0, blocksize);
1085                 }
1086         }
1087
1088         return lustre_commit_write(page, from, to);
1089 }
1090
1091 static int filter_preprw(int cmd, struct lustre_handle *conn,
1092                          int objcount, struct obd_ioobj *obj,
1093                          int niocount, struct niobuf_remote *nb,
1094                          struct niobuf_local *res, void **desc_private)
1095 {
1096         struct obd_run_ctxt saved;
1097         struct obd_device *obd;
1098         struct obd_ioobj *o = obj;
1099         struct niobuf_remote *b = nb;
1100         struct niobuf_local *r = res;
1101         void *journal_save = NULL;
1102         int rc = 0;
1103         int i;
1104         ENTRY;
1105
1106         memset(res, 0, sizeof(*res) * niocount);
1107         obd = class_conn2obd(conn);
1108
1109         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1110
1111         if (cmd & OBD_BRW_WRITE) {
1112                 *desc_private = filter_journal_start(&journal_save,
1113                                                      &obd->u.filter,
1114                                                      objcount, obj, niocount,
1115                                                      nb);
1116                 if (IS_ERR(*desc_private))
1117                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1118         }
1119
1120         for (i = 0; i < objcount; i++, o++) {
1121                 struct dentry *dentry;
1122                 struct inode *inode;
1123                 int j;
1124
1125                 dentry = filter_fid2dentry(obd,
1126                                            filter_parent(obd, S_IFREG),
1127                                            o->ioo_id, S_IFREG);
1128                 if (IS_ERR(dentry))
1129                         GOTO(out_clean, rc = PTR_ERR(dentry));
1130                 inode = dentry->d_inode;
1131                 if (!inode) {
1132                         CERROR("trying to BRW to non-existent file %Ld\n",
1133                                (unsigned long long)o->ioo_id);
1134                         dput(dentry);
1135                         GOTO(out_clean, rc = -ENOENT);
1136                 }
1137
1138                 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1139                         unsigned long index = b->offset >> PAGE_SHIFT;
1140                         struct page *page;
1141
1142                         if (j == 0)
1143                                 r->dentry = dentry;
1144                         else
1145                                 r->dentry = dget(dentry);
1146
1147                         if (cmd & OBD_BRW_WRITE)
1148                                 page = filter_get_page_write(inode, index, r);
1149                         else
1150                                 page = lustre_get_page_read(inode, index);
1151
1152                         if (IS_ERR(page)) {
1153                                 dput(dentry);
1154                                 GOTO(out_clean, rc = PTR_ERR(page));
1155                         }
1156
1157                         r->addr = page_address(page);
1158                         r->offset = b->offset;
1159                         r->page = page;
1160                         r->len = PAGE_SIZE;
1161                 }
1162         }
1163
1164 out_stop:
1165         if (cmd & OBD_BRW_WRITE) {
1166                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1167                                               *desc_private);
1168                 if (!rc)
1169                         rc = err;
1170         }
1171 out_ctxt:
1172         pop_ctxt(&saved);
1173         RETURN(rc);
1174 out_clean:
1175         while (r-- > res) {
1176                 dput(r->dentry);
1177                 if (cmd & OBD_BRW_WRITE)
1178                         filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1179                 else
1180                         lustre_put_page(r->page);
1181         }
1182         goto out_stop;
1183 }
1184
1185 static int filter_write_locked_page(struct niobuf_local *lnb)
1186 {
1187         struct page *lpage;
1188         int rc;
1189
1190         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1191         if (IS_ERR(lpage)) {
1192                 /* It is highly unlikely that we would ever get an error here.
1193                  * The page we want to get was previously locked, so it had to
1194                  * have already allocated the space, and we were just writing
1195                  * over the same data, so there would be no hole in the file.
1196                  *
1197                  * XXX: possibility of a race with truncate could exist, need
1198                  *      to check that.  There are no guarantees w.r.t.
1199                  *      write order even on a local filesystem, although the
1200                  *      normal response would be to return the number of bytes
1201                  *      successfully written and leave the rest to the app.
1202                  */
1203                 rc = PTR_ERR(lpage);
1204                 CERROR("error getting locked page index %ld: rc = %d\n",
1205                        lnb->page->index, rc);
1206                 GOTO(out, rc);
1207         }
1208
1209         memcpy(page_address(lpage), kmap(lnb->page), PAGE_SIZE);
1210         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1211         if (rc)
1212                 CERROR("error committing locked page %ld: rc = %d\n",
1213                        lnb->page->index, rc);
1214 out:
1215         kunmap(lnb->page);
1216         __free_pages(lnb->page, 0);
1217         dput(lnb->dentry);
1218
1219         return rc;
1220 }
1221
1222 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1223                            int objcount, struct obd_ioobj *obj,
1224                            int niocount, struct niobuf_local *res,
1225                            void *private)
1226 {
1227         struct obd_run_ctxt saved;
1228         struct obd_ioobj *o = obj;
1229         struct niobuf_local *r = res;
1230         struct obd_device *obd = class_conn2obd(conn);
1231         void *journal_save;
1232         int found_locked = 0;
1233         int rc = 0;
1234         int i;
1235         ENTRY;
1236
1237         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1238         journal_save = current->journal_info;
1239         if (journal_save)
1240                 CERROR("Existing handle %p???\n", journal_save);
1241         current->journal_info = private;
1242         for (i = 0; i < objcount; i++, obj++) {
1243                 int j;
1244                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1245                         struct page *page = r->page;
1246
1247                         if (!page)
1248                                 LBUG();
1249
1250                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1251                                 found_locked = 1;
1252                                 continue;
1253                         }
1254
1255                         if (cmd & OBD_BRW_WRITE) {
1256                                 int err = filter_commit_write(page, 0,
1257                                                               PAGE_SIZE, 0);
1258
1259                                 if (!rc)
1260                                         rc = err;
1261                         } else
1262                                 lustre_put_page(page);
1263
1264                         CDEBUG(D_INODE,
1265                                "put inode %p (%ld), count = %d, nlink = %d\n",
1266                                page->mapping->host,
1267                                page->mapping->host->i_ino,
1268                                atomic_read(&page->mapping->host->i_count) - 1,
1269                                page->mapping->host->i_nlink);
1270                         dput(r->dentry);
1271                 }
1272         }
1273         if (!found_locked)
1274                 goto out_ctxt;
1275
1276         for (i = 0; i < objcount; i++, obj++) {
1277                 int j;
1278                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1279                         int err;
1280                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1281                                 continue;
1282
1283                         err = filter_write_locked_page(r);
1284                         if (!rc)
1285                                 rc = err;
1286                 }
1287         }
1288
1289 out_ctxt:
1290         current->journal_info = journal_save;
1291         pop_ctxt(&saved);
1292         RETURN(0);
1293 }
1294
1295 static int filter_statfs(struct lustre_handle *conn, struct statfs *statfs)
1296 {
1297         struct obd_device *obd = class_conn2obd(conn);
1298
1299         ENTRY;
1300         RETURN(vfs_statfs(obd->u.filter.fo_sb, statfs));
1301 }
1302
1303 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1304                            void *key, obd_count *vallen, void **val)
1305 {
1306         struct obd_device *obd;
1307         struct obd_export * export;
1308         ENTRY;
1309
1310         if (!(export = class_conn2export(conn))) {
1311                 CDEBUG(D_IOCTL, "invalid client %Lx\n", conn->addr);
1312                 RETURN(-EINVAL);
1313         }
1314
1315         obd = class_conn2obd(conn);
1316
1317         if ( keylen == strlen("blocksize") &&
1318              memcmp(key, "blocksize", keylen) == 0 ) {
1319                 *vallen = sizeof(long);
1320                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1321                 RETURN(0);
1322         }
1323
1324         if ( keylen == strlen("blocksize_bits") &&
1325              memcmp(key, "blocksize_bits", keylen) == 0 ){
1326                 *vallen = sizeof(long);
1327                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1328                 RETURN(0);
1329         }
1330
1331         if ( keylen == strlen("root_ino") &&
1332              memcmp(key, "root_ino", keylen) == 0 ){
1333                 *vallen = sizeof(obd_id);
1334                 *val = (void *)(obd_id)FILTER_ROOTINO;
1335                 RETURN(0);
1336         }
1337
1338         CDEBUG(D_IOCTL, "invalid key\n");
1339         RETURN(-EINVAL);
1340 }
1341
1342 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1343                   struct lustre_handle *src_conn, struct obdo *src,
1344                   obd_size count, obd_off offset)
1345 {
1346         struct page *page;
1347         struct lov_stripe_md srcmd, dstmd;
1348         unsigned long index = 0;
1349         int err = 0;
1350
1351         memset(&srcmd, 0, sizeof(srcmd));
1352         memset(&dstmd, 0, sizeof(dstmd));
1353         srcmd.lmd_object_id = src->o_id;
1354         dstmd.lmd_object_id = dst->o_id;
1355
1356         ENTRY;
1357         CDEBUG(D_INFO, "src: ino %Ld blocks %Ld, size %Ld, dst: ino %Ld\n",
1358                (unsigned long long)src->o_id, (unsigned long long)src->o_blocks,
1359                (unsigned long long)src->o_size, (unsigned long long)dst->o_id);
1360         page = alloc_page(GFP_USER);
1361         if (page == NULL)
1362                 RETURN(-ENOMEM);
1363
1364         while (TryLockPage(page))
1365                 ___wait_on_page(page);
1366
1367         /* XXX with brw vector I/O, we could batch up reads and writes here,
1368          *     all we need to do is allocate multiple pages to handle the I/Os
1369          *     and arrays to handle the request parameters.
1370          */
1371         while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1372                 obd_size         brw_count = PAGE_SIZE;
1373                 obd_off          brw_offset = (page->index) << PAGE_SHIFT;
1374                 obd_flag         flagr = 0;
1375                 obd_flag         flagw = OBD_BRW_CREATE;
1376
1377                 page->index = index;
1378                 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &page,
1379                               &brw_count, &brw_offset, &flagr, NULL, NULL);
1380
1381                 if ( err ) {
1382                         EXIT;
1383                         break;
1384                 }
1385                 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1386
1387                 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &page,
1388                               &brw_count, &brw_offset, &flagw, NULL, NULL);
1389
1390                 /* XXX should handle dst->o_size, dst->o_blocks here */
1391                 if ( err ) {
1392                         EXIT;
1393                         break;
1394                 }
1395
1396                 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1397
1398                 index++;
1399         }
1400         dst->o_size = src->o_size;
1401         dst->o_blocks = src->o_blocks;
1402         dst->o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
1403         UnlockPage(page);
1404         __free_page(page);
1405
1406         RETURN(err);
1407 }
1408
1409
1410 static struct obd_ops filter_obd_ops = {
1411         o_get_info:    filter_get_info,
1412         o_setup:       filter_setup,
1413         o_cleanup:     filter_cleanup,
1414         o_connect:     filter_connect,
1415         o_disconnect:  filter_disconnect,
1416         o_statfs:      filter_statfs,
1417         o_getattr:     filter_getattr,
1418         o_create:      filter_create,
1419         o_setattr:     filter_setattr,
1420         o_destroy:     filter_destroy,
1421         o_open:        filter_open,
1422         o_close:       filter_close,
1423         o_brw:         filter_pgcache_brw,
1424         o_punch:       filter_truncate,
1425         o_preprw:      filter_preprw,
1426         o_commitrw:    filter_commitrw
1427 #if 0
1428         o_preallocate: filter_preallocate_inodes,
1429         o_migrate:     filter_migrate,
1430         o_copy:        filter_copy_data,
1431         o_iterate:     filter_iterate
1432 #endif
1433 };
1434
1435
1436 static int __init obdfilter_init(void)
1437 {
1438         printk(KERN_INFO "Filtering OBD driver  v0.001, info@clusterfs.com\n");
1439         return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1440 }
1441
1442 static void __exit obdfilter_exit(void)
1443 {
1444         class_unregister_type(OBD_FILTER_DEVICENAME);
1445 }
1446
1447 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1448 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1449 MODULE_LICENSE("GPL");
1450
1451 module_init(obdfilter_init);
1452 module_exit(obdfilter_exit);