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 {
 359         int rc;
 360         ENTRY;
 361         MOD_INC_USE_COUNT;
 362         rc = class_connect(conn, obd);
 363         if (rc)
 364                 MOD_DEC_USE_COUNT;
 365         RETURN(rc);
 366 }
 367
 368 static int filter_disconnect(struct lustre_handle *conn)
 369 {
 370         int rc;
 371         ENTRY;
 372
 373         rc = class_disconnect(conn);
 374         if (!rc)
 375                 MOD_DEC_USE_COUNT;
 376
 377         /* XXX cleanup preallocated inodes */
 378         RETURN(rc);
 379 }
 380
 381 /* mount the file system (secretly) */
 382 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
 383 {
 384         struct obd_ioctl_data* data = buf;
 385         struct filter_obd *filter;
 386         struct vfsmount *mnt;
 387         int err = 0;
 388         ENTRY;
 389
 390         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 391                 RETURN(-EINVAL);
 392
 393         MOD_INC_USE_COUNT;
 394         mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
 395         err = PTR_ERR(mnt);
 396         if (IS_ERR(mnt))
 397                 GOTO(err_dec, err);
 398
 399         filter = &obddev->u.filter;;
 400         filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
 401         /* XXX is this even possible if do_kern_mount succeeded? */
 402         if (!filter->fo_sb)
 403                 GOTO(err_put, err = -ENODEV);
 404
 405         filter->fo_vfsmnt = mnt;
 406         filter->fo_fstype = strdup(data->ioc_inlbuf2);
 407
 408         OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
 409         filter->fo_ctxt.pwdmnt = mnt;
 410         filter->fo_ctxt.pwd = mnt->mnt_root;
 411         filter->fo_ctxt.fs = get_ds();
 412
 413         err = filter_prep(obddev);
 414         if (err)
 415                 GOTO(err_kfree, err);
 416         spin_lock_init(&filter->fo_lock);
 417
 418         obddev->obd_namespace =
 419                 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
 420         if (obddev->obd_namespace == NULL)
 421                 LBUG();
 422
 423         RETURN(0);
 424
 425 err_kfree:
 426         kfree(filter->fo_fstype);
 427 err_put:
 428         unlock_kernel();
 429         mntput(filter->fo_vfsmnt);
 430         filter->fo_sb = 0;
 431         lock_kernel();
 432
 433 err_dec:
 434         MOD_DEC_USE_COUNT;
 435         return err;
 436 }
 437
 438
 439 static int filter_cleanup(struct obd_device * obddev)
 440 {
 441         struct super_block *sb;
 442         ENTRY;
 443
 444         if (!(obddev->obd_flags & OBD_SET_UP))
 445                 RETURN(0);
 446
 447         if (!list_empty(&obddev->obd_exports)) {
 448                 CERROR("still has clients!\n");
 449                 RETURN(-EBUSY);
 450         }
 451
 452         ldlm_namespace_free(obddev->obd_namespace);
 453
 454         sb = obddev->u.filter.fo_sb;
 455         if (!obddev->u.filter.fo_sb)
 456                 RETURN(0);
 457
 458         filter_post(obddev);
 459
 460         shrink_dcache_parent(sb->s_root);
 461         unlock_kernel();
 462         mntput(obddev->u.filter.fo_vfsmnt);
 463         obddev->u.filter.fo_sb = 0;
 464         kfree(obddev->u.filter.fo_fstype);
 465
 466         lock_kernel();
 467
 468         MOD_DEC_USE_COUNT;
 469         RETURN(0);
 470 }
 471
 472
 473 static inline void filter_from_inode(struct obdo *oa, struct inode *inode)
 474 {
 475         int type = oa->o_mode & S_IFMT;
 476         ENTRY;
 477
 478         CDEBUG(D_INFO, "src inode %ld, dst obdo %ld valid 0x%08x\n",
 479                inode->i_ino, (long)oa->o_id, oa->o_valid);
 480         obdo_from_inode(oa, inode);
 481         oa->o_mode &= ~S_IFMT;
 482         oa->o_mode |= type;
 483
 484         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
 485                 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
 486                 oa->o_rdev = rdev;
 487                 oa->o_valid |= OBD_MD_FLRDEV;
 488         }
 489
 490         EXIT;
 491 }
 492
 493 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
 494                           struct lov_stripe_md *md)
 495 {
 496         struct obd_device *obddev = class_conn2obd(conn);
 497         struct dentry *dentry;
 498         ENTRY;
 499
 500         if (!class_conn2export(conn)) {
 501                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 502                 RETURN(-EINVAL);
 503         }
 504
 505         obddev = class_conn2obd(conn);
 506         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 507                                    oa->o_id, oa->o_mode);
 508         if (IS_ERR(dentry))
 509                 RETURN(PTR_ERR(dentry));
 510
 511         oa->o_valid &= ~OBD_MD_FLID;
 512         filter_from_inode(oa, dentry->d_inode);
 513
 514         dput(dentry);
 515         RETURN(0);
 516 }
 517
 518 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
 519                           struct lov_stripe_md *md)
 520 {
 521         struct obd_run_ctxt saved;
 522         struct obd_device *obd = class_conn2obd(conn);
 523         struct dentry *dentry;
 524         struct iattr iattr;
 525         struct inode *inode;
 526         int rc;
 527         ENTRY;
 528
 529         iattr_from_obdo(&iattr, oa);
 530         iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
 531         dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
 532                                    oa->o_id, iattr.ia_mode);
 533         if (IS_ERR(dentry))
 534                 RETURN(PTR_ERR(dentry));
 535
 536         inode = dentry->d_inode;
 537         lock_kernel();
 538         if (iattr.ia_mode & ATTR_SIZE)
 539                 down(&inode->i_sem);
 540         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 541         if (inode->i_op->setattr)
 542                 rc = inode->i_op->setattr(dentry, &iattr);
 543         else
 544                 rc = inode_setattr(inode, &iattr);
 545         pop_ctxt(&saved);
 546         if (iattr.ia_mode & ATTR_SIZE) {
 547                 up(&inode->i_sem);
 548                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 549                 obdo_from_inode(oa, inode);
 550         }
 551         unlock_kernel();
 552
 553         CDEBUG(D_INODE, "put dentry %p, count = %d\n", inode,
 554                atomic_read(&dentry->d_count) - 1);
 555         dput(dentry);
 556         RETURN(rc);
 557 }
 558
 559 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
 560                           struct lov_stripe_md *ea)
 561 {
 562         struct obd_device *obd;
 563         struct dentry *dentry;
 564         /* ENTRY; */
 565
 566         if (!class_conn2export(conn)) {
 567                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 568                 RETURN(-EINVAL);
 569         }
 570
 571         obd = class_conn2obd(conn);
 572         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 573                                    oa->o_id, oa->o_mode);
 574         if (IS_ERR(dentry))
 575                 RETURN(PTR_ERR(dentry));
 576
 577         return 0;
 578 } /* filter_open */
 579
 580 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
 581                           struct lov_stripe_md *ea)
 582 {
 583         struct obd_device *obd;
 584         struct dentry *dentry;
 585         /* ENTRY; */
 586
 587         if (!class_conn2export(conn)) {
 588                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 589                 RETURN(-EINVAL);
 590         }
 591
 592         obd = class_conn2obd(conn);
 593         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 594                                    oa->o_id, oa->o_mode);
 595         if (IS_ERR(dentry))
 596                 RETURN(PTR_ERR(dentry));
 597
 598         CDEBUG(D_INODE, "put dentry %p, count = %d\n", dentry,
 599                atomic_read(&dentry->d_count) - 1);
 600         dput(dentry);  /* for the close */
 601         CDEBUG(D_INODE, "put dentry %p, count = %d\n", dentry,
 602                atomic_read(&dentry->d_count) - 1);
 603         dput(dentry);  /* for this call */
 604         return 0;
 605 } /* filter_close */
 606
 607 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
 608                          struct lov_stripe_md **ea)
 609 {
 610         char name[64];
 611         struct obd_run_ctxt saved;
 612         struct file *file;
 613         int mode;
 614         struct obd_device *obd = class_conn2obd(conn);
 615         struct iattr;
 616         ENTRY;
 617
 618         if (!class_conn2export(conn)) {
 619                 CERROR("invalid client %Lx\n", conn->addr);
 620                 return -EINVAL;
 621         }
 622
 623         oa->o_id = filter_next_id(obd);
 624         if (!(oa->o_mode && S_IFMT)) {
 625                 CERROR("filter obd: no type!\n");
 626                 return -ENOENT;
 627         }
 628
 629         filter_id(name, oa->o_id, oa->o_mode);
 630         mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
 631         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 632         file = filp_open(name, O_RDONLY | O_CREAT, mode);
 633         pop_ctxt(&saved);
 634         if (IS_ERR(file)) {
 635                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(file));
 636                 return -ENOENT;
 637         }
 638         filp_close(file, 0);
 639
 640         /* Set flags for fields we have set in the inode struct */
 641         oa->o_valid |= OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 642                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME |
 643                  OBD_MD_FLUID | OBD_MD_FLGID;
 644
 645         /* XXX Hmm, shouldn't we copy the fields into the obdo here? */
 646         return 0;
 647 }
 648
 649 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
 650                           struct lov_stripe_md *ea)
 651 {
 652         struct obd_device *obd;
 653         struct filter_obd *filter;
 654         struct obd_run_ctxt saved;
 655         struct obd_export *export;
 656         struct inode *inode;
 657         struct dentry *dir_dentry, *object_dentry;
 658         int rc;
 659         ENTRY;
 660
 661         if (!(export = class_conn2export(conn))) {
 662                 CERROR("invalid client %Lx\n", conn->addr);
 663                 RETURN(-EINVAL);
 664         }
 665
 666         CDEBUG(D_INODE, "destroying object %Ld\n", oa->o_id);
 667         obd = class_conn2obd(conn);
 668
 669         dir_dentry = filter_parent(obd, oa->o_mode);
 670         down(&dir_dentry->d_inode->i_sem);
 671
 672         object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
 673                                           oa->o_mode);
 674         if (IS_ERR(object_dentry))
 675                 GOTO(out, rc = -ENOENT);
 676
 677         inode = object_dentry->d_inode;
 678         if (inode->i_nlink != 1) {
 679                 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
 680                 LBUG();
 681                 inode->i_nlink = 1;
 682         }
 683         inode->i_mode = S_IFREG;
 684
 685         filter = &obd->u.filter;
 686         push_ctxt(&saved, &filter->fo_ctxt);
 687
 688         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 689         pop_ctxt(&saved);
 690         CDEBUG(D_INODE, "put child %p, count = %d\n", object_dentry,
 691                atomic_read(&object_dentry->d_count) - 1);
 692         dput(object_dentry);
 693
 694         EXIT;
 695 out:
 696         up(&dir_dentry->d_inode->i_sem);
 697         return rc;
 698 }
 699
 700 /* NB count and offset are used for punch, but not truncate */
 701 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
 702                            struct lov_stripe_md *md,
 703                            obd_off start, obd_off end)
 704 {
 705         int error;
 706         ENTRY;
 707
 708         if ( end != 0xffffffffffffffff ) {
 709                 CERROR("PUNCH not supported, only truncate works\n");
 710         }
 711
 712         CDEBUG(D_INODE, "calling truncate for object #%Ld, valid = %x, "
 713                "o_size = %Ld\n", oa->o_id, oa->o_valid, start);
 714         oa->o_size = start;
 715         error = filter_setattr(conn, oa, NULL);
 716         RETURN(error);
 717 }
 718
 719 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
 720                                struct lov_stripe_md *md, obd_count oa_bufs,
 721                                struct page **pages, obd_size *count,
 722                                obd_off *offset, obd_flag *flags,
 723                                brw_callback_t callback, void *data)
 724 {
 725         struct obd_run_ctxt      saved;
 726         struct super_block      *sb;
 727         int                      pnum;          /* index to pages (bufs) */
 728         unsigned long            retval;
 729         int                      error;
 730         struct file             *file;
 731         struct obd_device      *obd = class_conn2obd(conn);
 732         int pg;
 733         ENTRY;
 734
 735         if (!class_conn2export(conn)) {
 736                 CDEBUG(D_IOCTL, "invalid client %Lx\n", conn->addr);
 737                 RETURN(-EINVAL);
 738         }
 739
 740         sb = obd->u.filter.fo_sb;
 741         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 742         pnum = 0; /* pnum indexes buf 0..num_pages */
 743
 744         file = filter_obj_open(obd, md->lmd_object_id, S_IFREG);
 745         if (IS_ERR(file))
 746                 GOTO(out, retval = PTR_ERR(file));
 747
 748         /* count doubles as retval */
 749         for (pg = 0; pg < oa_bufs; pg++) {
 750                 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,%ld) "
 751                        "off count (%Ld,%Ld)\n",
 752                        cmd, pnum, file->f_dentry->d_inode->i_ino,
 753                        (unsigned long)offset[pnum] >> PAGE_CACHE_SHIFT,
 754                        (unsigned long long)offset[pnum],
 755                        (unsigned long long)count[pnum]);
 756                 if (cmd & OBD_BRW_WRITE) {
 757                         loff_t off;
 758                         char *buffer;
 759                         off = offset[pnum];
 760                         buffer = kmap(pages[pnum]);
 761                         retval = file->f_op->write(file, buffer, count[pnum],
 762                                                    &off);
 763                         kunmap(pages[pnum]);
 764                         CDEBUG(D_INODE, "retval %ld\n", retval);
 765                 } else {
 766                         loff_t off = offset[pnum];
 767                         char *buffer = kmap(pages[pnum]);
 768
 769                         if (off >= file->f_dentry->d_inode->i_size) {
 770                                 memset(buffer, 0, count[pnum]);
 771                                 retval = count[pnum];
 772                         } else {
 773                                 retval = file->f_op->read(file, buffer,
 774                                                           count[pnum], &off);
 775                         }
 776                         kunmap(pages[pnum]);
 777
 778                         if (retval != count[pnum]) {
 779                                 filp_close(file, 0);
 780                                 GOTO(out, retval = -EIO);
 781                         }
 782                         CDEBUG(D_INODE, "retval %ld\n", retval);
 783                 }
 784                 pnum++;
 785         }
 786         /* sizes and blocks are set by generic_file_write */
 787         /* ctimes/mtimes will follow with a setattr call */
 788         filp_close(file, 0);
 789
 790         /* XXX: do something with callback if it is set? */
 791
 792         EXIT;
 793 out:
 794         pop_ctxt(&saved);
 795         error = (retval >= 0) ? 0 : retval;
 796         return error;
 797 }
 798
 799
 800 struct inode *ioobj_to_inode(struct lustre_handle *conn, struct obd_ioobj *o)
 801 {
 802         struct obd_device *obd = class_conn2obd(conn);
 803         struct super_block *sb = obd->u.filter.fo_sb;
 804         struct inode *inode = NULL;
 805         ENTRY;
 806
 807         if (!sb || !sb->s_dev) {
 808                 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
 809                 RETURN(NULL);
 810         }
 811
 812         if ( !o->ioo_id ) {
 813                 CDEBUG(D_INODE, "fatal: invalid obdo %lu\n", (long)o->ioo_id);
 814                 RETURN(NULL);
 815         }
 816
 817         inode = filter_inode_from_obj(obd, o->ioo_id, S_IFREG);
 818         if (!inode || inode->i_nlink == 0 || is_bad_inode(inode)) {
 819                 CERROR("from obdo - fatal: invalid inode %ld (%s).\n",
 820                        (long)o->ioo_id, inode ? inode->i_nlink ? "bad inode" :
 821                        "no links" : "NULL");
 822                 iput(inode);
 823                 RETURN(NULL);
 824         }
 825
 826         RETURN(inode);
 827 }
 828
 829 /*
 830  * Calculate the number of buffer credits needed to write multiple pages in
 831  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
 832  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
 833  *
 834  * See comment above ext3_writepage_trans_blocks for details.  We assume
 835  * no data journaling is being done, but it does allow for all of the pages
 836  * being non-contiguous.  If we are guaranteed contiguous pages we could
 837  * reduce the number of (d)indirect blocks a lot.
 838  *
 839  * With N blocks per page and P pages, for each inode we have at most:
 840  * N*P indirect
 841  * min(N*P, blocksize/4 + 1) dindirect blocks
 842  * 1 tindirect
 843  *
 844  * For the entire filesystem, we have at most:
 845  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 846  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 847  * 1 inode block
 848  * 1 superblock
 849  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 850  */
 851 static int ext3_credits_needed(struct super_block *sb, int objcount,
 852                                struct obd_ioobj *obj)
 853 {
 854         struct obd_ioobj *o = obj;
 855         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 856         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 857         int nbitmaps = 0;
 858         int ngdblocks = 0;
 859         int needed = objcount + 1;
 860         int i;
 861
 862         for (i = 0; i < objcount; i++, o++) {
 863                 int nblocks = o->ioo_bufcnt * blockpp;
 864                 int ndindirect = min(nblocks, addrpp + 1);
 865                 int nindir = nblocks + ndindirect + 1;
 866
 867                 nbitmaps += nindir + nblocks;
 868                 ngdblocks += nindir + nblocks;
 869
 870                 needed += nindir;
 871         }
 872
 873         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 874                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 875         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 876                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 877
 878         needed += nbitmaps + ngdblocks;
 879
 880 #ifdef CONFIG_QUOTA
 881         /* We assume that there will be 1 bit set in s_dquot.flags for each
 882          * quota file that is active.  This is at least true for now.
 883          */
 884         needed += hweight32(sb_any_quota_enabled(sb)) *
 885                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 886 #endif
 887
 888         return needed;
 889 }
 890
 891 /* We have to start a huge journal transaction here to hold all of the
 892  * metadata for the pages being written here.  This is necessitated by
 893  * the fact that we do lots of prepare_write operations before we do
 894  * any of the matching commit_write operations, so even if we split
 895  * up to use "smaller" transactions none of them could complete until
 896  * all of them were opened.  By having a single journal transaction,
 897  * we eliminate duplicate reservations for common blocks like the
 898  * superblock and group descriptors or bitmaps.
 899  *
 900  * We will start the transaction here, but each prepare_write will
 901  * add a refcount to the transaction, and each commit_write will
 902  * remove a refcount.  The transaction will be closed when all of
 903  * the pages have been written.
 904  */
 905 static void *ext3_filter_journal_start(struct filter_obd *filter,
 906                                        int objcount, struct obd_ioobj *obj,
 907                                        int niocount, struct niobuf_remote *nb)
 908 {
 909         journal_t *journal = NULL;
 910         handle_t *handle = NULL;
 911         int needed;
 912
 913         /* Assumes ext3 and extN have same sb_info layout, but avoids issues
 914          * with having extN built properly before filterobd for now.
 915          */
 916         journal = EXT3_SB(filter->fo_sb)->s_journal;
 917         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
 918
 919         /* The number of blocks we could _possibly_ dirty can very large.
 920          * We reduce our request if it is absurd (and we couldn't get that
 921          * many credits for a single handle anyways).
 922          *
 923          * At some point we have to limit the size of I/Os sent at one time,
 924          * increase the size of the journal, or we have to calculate the
 925          * actual journal requirements more carefully by checking all of
 926          * the blocks instead of being maximally pessimistic.  It remains to
 927          * be seen if this is a real problem or not.
 928          */
 929         if (needed > journal->j_max_transaction_buffers) {
 930                 CERROR("want too many journal credits (%d) using %d instead\n",
 931                        needed, journal->j_max_transaction_buffers);
 932                 needed = journal->j_max_transaction_buffers;
 933         }
 934
 935         handle = journal_start(journal, needed);
 936         if (IS_ERR(handle))
 937                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 938                        PTR_ERR(handle));
 939
 940         return(handle);
 941 }
 942
 943 static void *filter_journal_start(void **journal_save,
 944                                   struct filter_obd *filter,
 945                                   int objcount, struct obd_ioobj *obj,
 946                                   int niocount, struct niobuf_remote *nb)
 947 {
 948         void *handle = NULL;
 949
 950         /* This may not be necessary - we probably never have a
 951          * transaction started when we enter here, so we can
 952          * remove the saving of the journal state entirely.
 953          * For now leave it in just to see if it ever happens.
 954          */
 955         *journal_save = current->journal_info;
 956         if (*journal_save) {
 957                 CERROR("Already have handle %p???\n", *journal_save);
 958                 LBUG();
 959                 current->journal_info = NULL;
 960         }
 961
 962         if (!strcmp(filter->fo_fstype, "ext3") ||
 963             !strcmp(filter->fo_fstype, "extN"))
 964                 handle = ext3_filter_journal_start(filter, objcount, obj,
 965                                                    niocount, nb);
 966         return handle;
 967 }
 968
 969 static int ext3_filter_journal_stop(void *handle)
 970 {
 971         int rc;
 972
 973         /* We got a refcount on the handle for each call to prepare_write,
 974          * so we can drop the "parent" handle here to avoid the need for
 975          * osc to call back into filterobd to close the handle.  The
 976          * remaining references will be dropped in commit_write.
 977          */
 978         rc = journal_stop((handle_t *)handle);
 979
 980         return rc;
 981 }
 982
 983 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
 984                                void *handle)
 985 {
 986         int rc = 0;
 987
 988         if (!strcmp(filter->fo_fstype, "ext3") ||
 989             !strcmp(filter->fo_fstype, "extN"))
 990                 rc = ext3_filter_journal_stop(handle);
 991
 992         if (rc)
 993                 CERROR("error on journal stop: rc = %d\n", rc);
 994
 995         current->journal_info = journal_save;
 996
 997         return rc;
 998 }
 999
1000 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1001                                    struct niobuf_local *lnb)
1002 {
1003         struct address_space *mapping = inode->i_mapping;
1004         struct page *page;
1005         int rc;
1006
1007         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1008         page = grab_cache_page_nowait(mapping, index); /* locked page */
1009
1010         /* This page is currently locked, so get a temporary page instead */
1011         if (!page) {
1012                 unsigned long addr;
1013                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1014                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1015                 if (!addr) {
1016                         CERROR("no memory for a temp page\n");
1017                         LBUG();
1018                         GOTO(err, rc = -ENOMEM);
1019                 }
1020                 page = virt_to_page(addr);
1021                 kmap(page);
1022                 page->index = index;
1023                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1024         } else if (!IS_ERR(page)) {
1025                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1026                  * a no-op for most filesystems, because we write the whole
1027                  * page.  For partial-page I/O this will read in the page.
1028                  */
1029                 kmap(page);
1030                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1031                 if (rc) {
1032                         CERROR("page index %lu, rc = %d\n", index, rc);
1033                         if (rc != -ENOSPC)
1034                                 LBUG();
1035                         GOTO(err_unlock, rc);
1036                 }
1037                 /* XXX not sure if we need this if we are overwriting page */
1038                 if (PageError(page)) {
1039                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1040                         LBUG();
1041                         GOTO(err_unlock, rc = -EIO);
1042                 }
1043         }
1044         return page;
1045
1046 err_unlock:
1047         unlock_page(page);
1048         lustre_put_page(page);
1049 err:
1050         return ERR_PTR(rc);
1051 }
1052
1053 /*
1054  * We need to balance prepare_write() calls with commit_write() calls.
1055  * If the page has been prepared, but we have no data for it, we don't
1056  * want to overwrite valid data on disk, but we still need to zero out
1057  * data for space which was newly allocated.  Like part of what happens
1058  * in __block_prepare_write() for newly allocated blocks.
1059  *
1060  * XXX currently __block_prepare_write() creates buffers for all the
1061  *     pages, and the filesystems mark these buffers as BH_New if they
1062  *     were newly allocated from disk. We use the BH_New flag similarly.
1063  */
1064 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1065                                int err)
1066 {
1067         if (err) {
1068                 unsigned block_start, block_end;
1069                 struct buffer_head *bh, *head = page->buffers;
1070                 unsigned blocksize = head->b_size;
1071                 void *addr = page_address(page);
1072
1073                 /* Currently one buffer per page, but in the future... */
1074                 for (bh = head, block_start = 0; bh != head || !block_start;
1075                      block_start = block_end, bh = bh->b_this_page) {
1076                         block_end = block_start + blocksize;
1077                         if (buffer_new(bh))
1078                                 memset(addr + block_start, 0, blocksize);
1079                 }
1080         }
1081
1082         return lustre_commit_write(page, from, to);
1083 }
1084
1085 static int filter_preprw(int cmd, struct lustre_handle *conn,
1086                          int objcount, struct obd_ioobj *obj,
1087                          int niocount, struct niobuf_remote *nb,
1088                          struct niobuf_local *res, void **desc_private)
1089 {
1090         struct obd_run_ctxt saved;
1091         struct obd_device *obd;
1092         struct obd_ioobj *o = obj;
1093         struct niobuf_remote *b = nb;
1094         struct niobuf_local *r = res;
1095         void *journal_save = NULL;
1096         int rc = 0;
1097         int i;
1098         ENTRY;
1099
1100         memset(res, 0, sizeof(*res) * niocount);
1101         obd = class_conn2obd(conn);
1102
1103         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1104
1105         if (cmd & OBD_BRW_WRITE) {
1106                 *desc_private = filter_journal_start(&journal_save,
1107                                                      &obd->u.filter,
1108                                                      objcount, obj, niocount,
1109                                                      nb);
1110                 if (IS_ERR(*desc_private))
1111                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1112         }
1113
1114         for (i = 0; i < objcount; i++, o++) {
1115                 struct dentry *dentry;
1116                 struct inode *inode;
1117                 int j;
1118
1119                 dentry = filter_fid2dentry(obd,
1120                                            filter_parent(obd, S_IFREG),
1121                                            o->ioo_id, S_IFREG);
1122                 if (IS_ERR(dentry))
1123                         GOTO(out_clean, rc = PTR_ERR(dentry));
1124                 inode = dentry->d_inode;
1125                 if (!inode) {
1126                         CERROR("trying to BRW to non-existent file %Ld\n",
1127                                (unsigned long long)o->ioo_id);
1128                         dput(dentry);
1129                         GOTO(out_clean, rc = -ENOENT);
1130                 }
1131
1132                 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1133                         unsigned long index = b->offset >> PAGE_SHIFT;
1134                         struct page *page;
1135
1136                         if (j == 0)
1137                                 r->dentry = dentry;
1138                         else
1139                                 r->dentry = dget(dentry);
1140
1141                         if (cmd & OBD_BRW_WRITE)
1142                                 page = filter_get_page_write(inode, index, r);
1143                         else
1144                                 page = lustre_get_page_read(inode, index);
1145
1146                         if (IS_ERR(page)) {
1147                                 dput(dentry);
1148                                 GOTO(out_clean, rc = PTR_ERR(page));
1149                         }
1150
1151                         r->addr = page_address(page);
1152                         r->offset = b->offset;
1153                         r->page = page;
1154                         r->len = PAGE_SIZE;
1155                 }
1156         }
1157
1158 out_stop:
1159         if (cmd & OBD_BRW_WRITE) {
1160                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1161                                               *desc_private);
1162                 if (!rc)
1163                         rc = err;
1164         }
1165 out_ctxt:
1166         pop_ctxt(&saved);
1167         RETURN(rc);
1168 out_clean:
1169         while (r-- > res) {
1170                 dput(r->dentry);
1171                 if (cmd & OBD_BRW_WRITE)
1172                         filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1173                 else
1174                         lustre_put_page(r->page);
1175         }
1176         goto out_stop;
1177 }
1178
1179 static int filter_write_locked_page(struct niobuf_local *lnb)
1180 {
1181         struct page *lpage;
1182         int rc;
1183
1184         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1185         if (IS_ERR(lpage)) {
1186                 /* It is highly unlikely that we would ever get an error here.
1187                  * The page we want to get was previously locked, so it had to
1188                  * have already allocated the space, and we were just writing
1189                  * over the same data, so there would be no hole in the file.
1190                  *
1191                  * XXX: possibility of a race with truncate could exist, need
1192                  *      to check that.  There are no guarantees w.r.t.
1193                  *      write order even on a local filesystem, although the
1194                  *      normal response would be to return the number of bytes
1195                  *      successfully written and leave the rest to the app.
1196                  */
1197                 rc = PTR_ERR(lpage);
1198                 CERROR("error getting locked page index %ld: rc = %d\n",
1199                        lnb->page->index, rc);
1200                 GOTO(out, rc);
1201         }
1202
1203         memcpy(page_address(lpage), kmap(lnb->page), PAGE_SIZE);
1204         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1205         if (rc)
1206                 CERROR("error committing locked page %ld: rc = %d\n",
1207                        lnb->page->index, rc);
1208 out:
1209         kunmap(lnb->page);
1210         __free_pages(lnb->page, 0);
1211         dput(lnb->dentry);
1212
1213         return rc;
1214 }
1215
1216 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1217                            int objcount, struct obd_ioobj *obj,
1218                            int niocount, struct niobuf_local *res,
1219                            void *private)
1220 {
1221         struct obd_run_ctxt saved;
1222         struct obd_ioobj *o = obj;
1223         struct niobuf_local *r = res;
1224         struct obd_device *obd = class_conn2obd(conn);
1225         void *journal_save;
1226         int found_locked = 0;
1227         int rc = 0;
1228         int i;
1229         ENTRY;
1230
1231         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1232         journal_save = current->journal_info;
1233         if (journal_save)
1234                 CERROR("Existing handle %p???\n", journal_save);
1235         current->journal_info = private;
1236         for (i = 0; i < objcount; i++, obj++) {
1237                 int j;
1238                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1239                         struct page *page = r->page;
1240
1241                         if (!page)
1242                                 LBUG();
1243
1244                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1245                                 found_locked = 1;
1246                                 continue;
1247                         }
1248
1249                         if (cmd & OBD_BRW_WRITE) {
1250                                 int err = filter_commit_write(page, 0,
1251                                                               PAGE_SIZE, 0);
1252
1253                                 if (!rc)
1254                                         rc = err;
1255                         } else
1256                                 lustre_put_page(page);
1257
1258                         CDEBUG(D_INODE,
1259                                "put inode %p (%ld), count = %d, nlink = %d\n",
1260                                page->mapping->host,
1261                                page->mapping->host->i_ino,
1262                                atomic_read(&page->mapping->host->i_count) - 1,
1263                                page->mapping->host->i_nlink);
1264                         dput(r->dentry);
1265                 }
1266         }
1267         if (!found_locked)
1268                 goto out_ctxt;
1269
1270         for (i = 0; i < objcount; i++, obj++) {
1271                 int j;
1272                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1273                         int err;
1274                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1275                                 continue;
1276
1277                         err = filter_write_locked_page(r);
1278                         if (!rc)
1279                                 rc = err;
1280                 }
1281         }
1282
1283 out_ctxt:
1284         current->journal_info = journal_save;
1285         pop_ctxt(&saved);
1286         RETURN(0);
1287 }
1288
1289 static int filter_statfs(struct lustre_handle *conn, struct statfs *statfs)
1290 {
1291         struct obd_device *obd = class_conn2obd(conn);
1292
1293         ENTRY;
1294         RETURN(vfs_statfs(obd->u.filter.fo_sb, statfs));
1295 }
1296
1297 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1298                            void *key, obd_count *vallen, void **val)
1299 {
1300         struct obd_device *obd;
1301         struct obd_export * export;
1302         ENTRY;
1303
1304         if (!(export = class_conn2export(conn))) {
1305                 CDEBUG(D_IOCTL, "invalid client %Lx\n", conn->addr);
1306                 RETURN(-EINVAL);
1307         }
1308
1309         obd = class_conn2obd(conn);
1310
1311         if ( keylen == strlen("blocksize") &&
1312              memcmp(key, "blocksize", keylen) == 0 ) {
1313                 *vallen = sizeof(long);
1314                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1315                 RETURN(0);
1316         }
1317
1318         if ( keylen == strlen("blocksize_bits") &&
1319              memcmp(key, "blocksize_bits", keylen) == 0 ){
1320                 *vallen = sizeof(long);
1321                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1322                 RETURN(0);
1323         }
1324
1325         if ( keylen == strlen("root_ino") &&
1326              memcmp(key, "root_ino", keylen) == 0 ){
1327                 *vallen = sizeof(obd_id);
1328                 *val = (void *)(obd_id)FILTER_ROOTINO;
1329                 RETURN(0);
1330         }
1331
1332         CDEBUG(D_IOCTL, "invalid key\n");
1333         RETURN(-EINVAL);
1334 }
1335
1336 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1337                   struct lustre_handle *src_conn, struct obdo *src,
1338                   obd_size count, obd_off offset)
1339 {
1340         struct page *page;
1341         struct lov_stripe_md srcmd, dstmd;
1342         unsigned long index = 0;
1343         int err = 0;
1344
1345         memset(&srcmd, 0, sizeof(srcmd));
1346         memset(&dstmd, 0, sizeof(dstmd));
1347         srcmd.lmd_object_id = src->o_id;
1348         dstmd.lmd_object_id = dst->o_id;
1349
1350         ENTRY;
1351         CDEBUG(D_INFO, "src: ino %Ld blocks %Ld, size %Ld, dst: ino %Ld\n",
1352                (unsigned long long)src->o_id, (unsigned long long)src->o_blocks,
1353                (unsigned long long)src->o_size, (unsigned long long)dst->o_id);
1354         page = alloc_page(GFP_USER);
1355         if (page == NULL)
1356                 RETURN(-ENOMEM);
1357
1358         while (TryLockPage(page))
1359                 ___wait_on_page(page);
1360
1361         /* XXX with brw vector I/O, we could batch up reads and writes here,
1362          *     all we need to do is allocate multiple pages to handle the I/Os
1363          *     and arrays to handle the request parameters.
1364          */
1365         while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1366                 obd_size         brw_count = PAGE_SIZE;
1367                 obd_off          brw_offset = (page->index) << PAGE_SHIFT;
1368                 obd_flag         flagr = 0;
1369                 obd_flag         flagw = OBD_BRW_CREATE;
1370
1371                 page->index = index;
1372                 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &page,
1373                               &brw_count, &brw_offset, &flagr, NULL, NULL);
1374
1375                 if ( err ) {
1376                         EXIT;
1377                         break;
1378                 }
1379                 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1380
1381                 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &page,
1382                               &brw_count, &brw_offset, &flagw, NULL, NULL);
1383
1384                 /* XXX should handle dst->o_size, dst->o_blocks here */
1385                 if ( err ) {
1386                         EXIT;
1387                         break;
1388                 }
1389
1390                 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1391
1392                 index++;
1393         }
1394         dst->o_size = src->o_size;
1395         dst->o_blocks = src->o_blocks;
1396         dst->o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
1397         UnlockPage(page);
1398         __free_page(page);
1399
1400         RETURN(err);
1401 }
1402
1403
1404 static struct obd_ops filter_obd_ops = {
1405         o_get_info:    filter_get_info,
1406         o_setup:       filter_setup,
1407         o_cleanup:     filter_cleanup,
1408         o_connect:     filter_connect,
1409         o_disconnect:  filter_disconnect,
1410         o_statfs:      filter_statfs,
1411         o_getattr:     filter_getattr,
1412         o_create:      filter_create,
1413         o_setattr:     filter_setattr,
1414         o_destroy:     filter_destroy,
1415         o_open:        filter_open,
1416         o_close:       filter_close,
1417         o_brw:         filter_pgcache_brw,
1418         o_punch:       filter_truncate,
1419         o_preprw:      filter_preprw,
1420         o_commitrw:    filter_commitrw
1421 #if 0
1422         o_preallocate: filter_preallocate_inodes,
1423         o_migrate:     filter_migrate,
1424         o_copy:        filter_copy_data,
1425         o_iterate:     filter_iterate
1426 #endif
1427 };
1428
1429
1430 static int __init obdfilter_init(void)
1431 {
1432         printk(KERN_INFO "Filtering OBD driver  v0.001, info@clusterfs.com\n");
1433         return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1434 }
1435
1436 static void __exit obdfilter_exit(void)
1437 {
1438         class_unregister_type(OBD_FILTER_DEVICENAME);
1439 }
1440
1441 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1442 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1443 MODULE_LICENSE("GPL");
1444
1445 module_init(obdfilter_init);
1446 module_exit(obdfilter_exit);