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 static inline void f_dput(struct dentry *dentry)
  54 {
  55         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
  56                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
  57         LASSERT(atomic_read(&dentry->d_count) > 0);
  58         dput(dentry);
  59 }
  60
  61 /* setup the object store with correct subdirectories */
  62 static int filter_prep(struct obd_device *obddev)
  63 {
  64         struct obd_run_ctxt saved;
  65         struct filter_obd *filter = &obddev->u.filter;
  66         struct dentry *dentry;
  67         struct file *file;
  68         struct inode *inode;
  69         int rc = 0;
  70         char rootid[128];
  71         __u64 lastino = 2;
  72         int mode = 0;
  73
  74         push_ctxt(&saved, &filter->fo_ctxt);
  75         dentry = simple_mkdir(current->fs->pwd, "O", 0700);
  76         CDEBUG(D_INODE, "got/created O: %p\n", dentry);
  77         if (IS_ERR(dentry)) {
  78                 rc = PTR_ERR(dentry);
  79                 CERROR("cannot open/create O: rc = %d\n", rc);
  80                 GOTO(out, rc);
  81         }
  82         filter->fo_dentry_O = dentry;
  83         dentry = simple_mkdir(current->fs->pwd, "P", 0700);
  84         CDEBUG(D_INODE, "got/created P: %p\n", dentry);
  85         if (IS_ERR(dentry)) {
  86                 rc = PTR_ERR(dentry);
  87                 CERROR("cannot open/create P: rc = %d\n", rc);
  88                 GOTO(out_O, rc);
  89         }
  90         CDEBUG(D_INODE, "putting P: %p, count = %d\n", dentry,
  91                atomic_read(&dentry->d_count) - 1);
  92         f_dput(dentry);
  93         dentry = simple_mkdir(current->fs->pwd, "D", 0700);
  94         CDEBUG(D_INODE, "got/created D: %p\n", dentry);
  95         if (IS_ERR(dentry)) {
  96                 rc = PTR_ERR(dentry);
  97                 CERROR("cannot open/create D: rc = %d\n", rc);
  98                 GOTO(out_O, rc);
  99         }
 100         CDEBUG(D_INODE, "putting D: %p, count = %d\n", dentry,
 101                atomic_read(&dentry->d_count) - 1);
 102         f_dput(dentry);
 103
 104         /*
 105          * Create directories and/or get dentries for each object type.
 106          * This saves us from having to do multiple lookups for each one.
 107          */
 108         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 109                 char *type = obd_type_by_mode[mode];
 110
 111                 if (!type) {
 112                         filter->fo_dentry_O_mode[mode] = NULL;
 113                         continue;
 114                 }
 115                 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
 116                 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
 117                 if (IS_ERR(dentry)) {
 118                         rc = PTR_ERR(dentry);
 119                         CERROR("cannot create O/%s: rc = %d\n", type, rc);
 120                         GOTO(out_O_mode, rc);
 121                 }
 122                 filter->fo_dentry_O_mode[mode] = dentry;
 123         }
 124
 125         filter_id(rootid, FILTER_ROOTINO, S_IFDIR);
 126         file = filp_open(rootid, O_RDWR | O_CREAT, 0755);
 127         if (IS_ERR(file)) {
 128                 rc = PTR_ERR(file);
 129                 CERROR("OBD filter: cannot open/create root %s: rc = %d\n",
 130                        rootid, rc);
 131                 GOTO(out_O_mode, rc);
 132         }
 133         filp_close(file, 0);
 134
 135         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 136         if ( !file || IS_ERR(file) ) {
 137                 rc = PTR_ERR(file);
 138                 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
 139                        "D/status", rc);
 140                 GOTO(out_O_mode, rc);
 141         }
 142
 143         /* steal operations */
 144         inode = file->f_dentry->d_inode;
 145         filter->fo_fop = file->f_op;
 146         filter->fo_iop = inode->i_op;
 147         filter->fo_aops = inode->i_mapping->a_ops;
 148
 149         if (inode->i_size == 0) {
 150                 __u64 disk_lastino = cpu_to_le64(lastino);
 151                 ssize_t retval = file->f_op->write(file, (char *)&disk_lastino,
 152                                                    sizeof(disk_lastino),
 153                                                    &file->f_pos);
 154                 if (retval != sizeof(disk_lastino)) {
 155                         CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
 156                         filp_close(file, 0);
 157                         GOTO(out_O_mode, rc = -EIO);
 158                 }
 159         } else {
 160                 __u64 disk_lastino;
 161                 ssize_t retval = file->f_op->read(file, (char *)&disk_lastino,
 162                                                   sizeof(disk_lastino),
 163                                                   &file->f_pos);
 164                 if (retval != sizeof(disk_lastino)) {
 165                         CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
 166                         filp_close(file, 0);
 167                         GOTO(out_O_mode, rc = -EIO);
 168                 }
 169                 lastino = le64_to_cpu(disk_lastino);
 170         }
 171         filter->fo_lastino = lastino;
 172         filp_close(file, 0);
 173
 174         rc = 0;
 175  out:
 176         pop_ctxt(&saved);
 177
 178         return(rc);
 179
 180 out_O_mode:
 181         while (mode-- > 0) {
 182                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 183                 if (dentry) {
 184                         f_dput(dentry);
 185                         filter->fo_dentry_O_mode[mode] = NULL;
 186                 }
 187         }
 188 out_O:
 189         f_dput(filter->fo_dentry_O);
 190         filter->fo_dentry_O = NULL;
 191         goto out;
 192 }
 193
 194 /* cleanup the filter: write last used object id to status file */
 195 static void filter_post(struct obd_device *obddev)
 196 {
 197         struct obd_run_ctxt saved;
 198         struct filter_obd *filter = &obddev->u.filter;
 199         __u64 disk_lastino;
 200         long rc;
 201         struct file *file;
 202         int mode;
 203
 204         push_ctxt(&saved, &filter->fo_ctxt);
 205         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 206         if (IS_ERR(file)) {
 207                 CERROR("OBD filter: cannot create status file\n");
 208                 goto out;
 209         }
 210
 211         file->f_pos = 0;
 212         disk_lastino = cpu_to_le64(filter->fo_lastino);
 213         rc = file->f_op->write(file, (char *)&disk_lastino,
 214                        sizeof(disk_lastino), &file->f_pos);
 215         if (rc != sizeof(disk_lastino))
 216                 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
 217
 218         rc = filp_close(file, NULL);
 219         if (rc)
 220                 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
 221
 222         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 223                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 224                 if (dentry) {
 225                         f_dput(dentry);
 226                         filter->fo_dentry_O_mode[mode] = NULL;
 227                 }
 228         }
 229         f_dput(filter->fo_dentry_O);
 230 out:
 231         pop_ctxt(&saved);
 232 }
 233
 234
 235 static __u64 filter_next_id(struct obd_device *obddev)
 236 {
 237         obd_id id;
 238
 239         spin_lock(&obddev->u.filter.fo_lock);
 240         id = ++obddev->u.filter.fo_lastino;
 241         spin_unlock(&obddev->u.filter.fo_lock);
 242
 243         /* FIXME: write the lastino to disk here */
 244         return id;
 245 }
 246
 247 /* how to get files, dentries, inodes from object id's */
 248 /* parent i_sem is already held if needed for exclusivity */
 249 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
 250                                         struct dentry *dparent,
 251                                         __u64 id, __u32 type)
 252 {
 253         struct super_block *sb = obddev->u.filter.fo_sb;
 254         struct dentry *dchild;
 255         char name[32];
 256         int len;
 257         ENTRY;
 258
 259         if (!sb || !sb->s_dev) {
 260                 CERROR("fatal: device not initialized.\n");
 261                 RETURN(ERR_PTR(-ENXIO));
 262         }
 263
 264         if (id == 0) {
 265                 CERROR("fatal: invalid object #0\n");
 266                 LBUG();
 267                 RETURN(ERR_PTR(-ESTALE));
 268         }
 269
 270         if (!(type & S_IFMT)) {
 271                 CERROR("OBD %s, object %Lu has bad type: %o\n", __FUNCTION__,
 272                        (unsigned long long)id, type);
 273                 RETURN(ERR_PTR(-EINVAL));
 274         }
 275
 276         len = sprintf(name, "%Ld", id);
 277         CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
 278                name);
 279         dchild = lookup_one_len(name, dparent, len);
 280         if (IS_ERR(dchild)) {
 281                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 282                 RETURN(dchild);
 283         }
 284
 285         CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
 286                obd_mode_to_type(type), name, dchild,
 287                atomic_read(&dchild->d_count));
 288
 289         LASSERT(atomic_read(&dchild->d_count) > 0);
 290
 291         RETURN(dchild);
 292 }
 293
 294 static struct file *filter_obj_open(struct obd_device *obddev,
 295                                     __u64 id, __u32 type)
 296 {
 297         struct super_block *sb = obddev->u.filter.fo_sb;
 298         struct obd_run_ctxt saved;
 299         char name[24];
 300         struct file *file;
 301         ENTRY;
 302
 303         if (!sb || !sb->s_dev) {
 304                 CERROR("fatal: device not initialized.\n");
 305                 RETURN(ERR_PTR(-ENXIO));
 306         }
 307
 308         if (!id) {
 309                 CERROR("fatal: invalid obdo %Lu\n", (unsigned long long)id);
 310                 RETURN(ERR_PTR(-ESTALE));
 311         }
 312
 313         if (!(type & S_IFMT)) {
 314                 CERROR("OBD %s, no type (%Ld), mode %o!\n", __FUNCTION__,
 315                        (unsigned long long)id, type);
 316                 RETURN(ERR_PTR(-EINVAL));
 317         }
 318
 319         filter_id(name, id, type);
 320         push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
 321         file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
 322         pop_ctxt(&saved);
 323
 324         CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
 325
 326         if (IS_ERR(file))
 327                 file = NULL;
 328         RETURN(file);
 329 }
 330
 331 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
 332 {
 333         struct filter_obd *filter = &obddev->u.filter;
 334
 335         return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
 336 }
 337
 338
 339 /* obd methods */
 340 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
 341                           char *cluuid)
 342 {
 343         int rc;
 344         ENTRY;
 345         MOD_INC_USE_COUNT;
 346         rc = class_connect(conn, obd, cluuid);
 347         if (rc)
 348                 MOD_DEC_USE_COUNT;
 349         RETURN(rc);
 350 }
 351
 352 static int filter_disconnect(struct lustre_handle *conn)
 353 {
 354         int rc;
 355         ENTRY;
 356
 357         rc = class_disconnect(conn);
 358         if (!rc)
 359                 MOD_DEC_USE_COUNT;
 360
 361         /* XXX cleanup preallocated inodes */
 362         RETURN(rc);
 363 }
 364
 365 /* mount the file system (secretly) */
 366 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
 367 {
 368         struct obd_ioctl_data* data = buf;
 369         struct filter_obd *filter;
 370         struct vfsmount *mnt;
 371         int err = 0;
 372         ENTRY;
 373
 374         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 375                 RETURN(-EINVAL);
 376
 377         MOD_INC_USE_COUNT;
 378         mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
 379         err = PTR_ERR(mnt);
 380         if (IS_ERR(mnt))
 381                 GOTO(err_dec, err);
 382
 383         filter = &obddev->u.filter;;
 384         filter->fo_vfsmnt = mnt;
 385         filter->fo_fstype = strdup(data->ioc_inlbuf2);
 386         filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
 387         /* XXX is this even possible if do_kern_mount succeeded? */
 388         if (!filter->fo_sb)
 389                 GOTO(err_kfree, err = -ENODEV);
 390
 391         OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
 392         filter->fo_ctxt.pwdmnt = mnt;
 393         filter->fo_ctxt.pwd = mnt->mnt_root;
 394         filter->fo_ctxt.fs = get_ds();
 395
 396         err = filter_prep(obddev);
 397         if (err)
 398                 GOTO(err_kfree, err);
 399         spin_lock_init(&filter->fo_lock);
 400
 401         obddev->obd_namespace =
 402                 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
 403         if (obddev->obd_namespace == NULL)
 404                 LBUG();
 405
 406         RETURN(0);
 407
 408 err_kfree:
 409         kfree(filter->fo_fstype);
 410         unlock_kernel();
 411         mntput(filter->fo_vfsmnt);
 412         filter->fo_sb = 0;
 413         lock_kernel();
 414
 415 err_dec:
 416         MOD_DEC_USE_COUNT;
 417         return err;
 418 }
 419
 420
 421 static int filter_cleanup(struct obd_device * obddev)
 422 {
 423         struct super_block *sb;
 424         ENTRY;
 425
 426         if (!list_empty(&obddev->obd_exports)) {
 427                 CERROR("still has clients!\n");
 428                 class_disconnect_all(obddev);
 429                 if (!list_empty(&obddev->obd_exports)) {
 430                         CERROR("still has exports after forced cleanup?\n");
 431                         RETURN(-EBUSY);
 432                 }
 433         }
 434
 435         ldlm_namespace_free(obddev->obd_namespace);
 436
 437         sb = obddev->u.filter.fo_sb;
 438         if (!obddev->u.filter.fo_sb)
 439                 RETURN(0);
 440
 441         filter_post(obddev);
 442
 443         shrink_dcache_parent(sb->s_root);
 444         unlock_kernel();
 445         mntput(obddev->u.filter.fo_vfsmnt);
 446         obddev->u.filter.fo_sb = 0;
 447         kfree(obddev->u.filter.fo_fstype);
 448
 449         lock_kernel();
 450
 451         MOD_DEC_USE_COUNT;
 452         RETURN(0);
 453 }
 454
 455
 456 static inline void filter_from_inode(struct obdo *oa, struct inode *inode,
 457                                      int valid)
 458 {
 459         int type = oa->o_mode & S_IFMT;
 460         ENTRY;
 461
 462         CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
 463                inode->i_ino, inode, (long)oa->o_id, valid);
 464         /* Don't copy the inode number in place of the object ID */
 465         obdo_from_inode(oa, inode, valid);
 466         oa->o_mode &= ~S_IFMT;
 467         oa->o_mode |= type;
 468
 469         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
 470                 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
 471                 oa->o_rdev = rdev;
 472                 oa->o_valid |= OBD_MD_FLRDEV;
 473         }
 474
 475         EXIT;
 476 }
 477
 478 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
 479                           struct lov_stripe_md *md)
 480 {
 481         struct obd_device *obddev = class_conn2obd(conn);
 482         struct dentry *dentry;
 483         ENTRY;
 484
 485         if (!class_conn2export(conn)) {
 486                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 487                 RETURN(-EINVAL);
 488         }
 489
 490         obddev = class_conn2obd(conn);
 491         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 492                                    oa->o_id, oa->o_mode);
 493         if (IS_ERR(dentry))
 494                 RETURN(PTR_ERR(dentry));
 495
 496         filter_from_inode(oa, dentry->d_inode, oa->o_valid);
 497
 498         f_dput(dentry);
 499         RETURN(0);
 500 }
 501
 502 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
 503                           struct lov_stripe_md *md)
 504 {
 505         struct obd_run_ctxt saved;
 506         struct obd_device *obd = class_conn2obd(conn);
 507         struct dentry *dentry;
 508         struct iattr iattr;
 509         struct inode *inode;
 510         int rc;
 511         ENTRY;
 512
 513         iattr_from_obdo(&iattr, oa, oa->o_valid);
 514         iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
 515         dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
 516                                    oa->o_id, iattr.ia_mode);
 517         if (IS_ERR(dentry))
 518                 RETURN(PTR_ERR(dentry));
 519
 520         inode = dentry->d_inode;
 521         if (!inode) {
 522                 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
 523                 GOTO(out_setattr, rc = -ENOENT);
 524         }
 525
 526         lock_kernel();
 527         if (iattr.ia_valid & ATTR_SIZE)
 528                 down(&inode->i_sem);
 529         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 530         if (inode->i_op->setattr)
 531                 rc = inode->i_op->setattr(dentry, &iattr);
 532         else
 533                 rc = inode_setattr(inode, &iattr);
 534         pop_ctxt(&saved);
 535         if (iattr.ia_valid & ATTR_SIZE) {
 536                 up(&inode->i_sem);
 537                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 538                 obdo_from_inode(oa, inode, oa->o_valid);
 539         }
 540         unlock_kernel();
 541
 542 out_setattr:
 543         f_dput(dentry);
 544         RETURN(rc);
 545 }
 546
 547 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
 548                           struct lov_stripe_md *ea)
 549 {
 550         struct obd_export *export;
 551         struct obd_device *obd;
 552         struct dentry *dentry;
 553         /* ENTRY; */
 554
 555         export = class_conn2export(conn);
 556         if (!export) {
 557                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 558                 RETURN(-EINVAL);
 559         }
 560
 561         obd = export->exp_obd;
 562         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 563                                    oa->o_id, oa->o_mode);
 564         if (IS_ERR(dentry))
 565                 RETURN(PTR_ERR(dentry));
 566
 567         filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 568                           OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 569
 570         return 0;
 571 } /* filter_open */
 572
 573 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
 574                         struct lov_stripe_md *ea)
 575 {
 576         struct obd_device *obd;
 577         struct dentry *dentry;
 578         /* ENTRY; */
 579
 580         obd = class_conn2obd(conn);
 581         if (!obd) {
 582                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 583                 RETURN(-EINVAL);
 584         }
 585
 586         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 587                                    oa->o_id, oa->o_mode);
 588         if (IS_ERR(dentry))
 589                 RETURN(PTR_ERR(dentry));
 590         LASSERT(atomic_read(&dentry->d_count) > 1);
 591
 592         f_dput(dentry);  /* for the close */
 593         f_dput(dentry);  /* for this call */
 594         return 0;
 595 } /* filter_close */
 596
 597 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
 598                          struct lov_stripe_md **ea)
 599 {
 600         char name[64];
 601         struct obd_run_ctxt saved;
 602         struct file *file;
 603         int mode;
 604         struct obd_device *obd = class_conn2obd(conn);
 605         struct iattr;
 606         ENTRY;
 607
 608         if (!obd) {
 609                 CERROR("invalid client %Lx\n", conn->addr);
 610                 return -EINVAL;
 611         }
 612
 613         if (!(oa->o_mode && S_IFMT)) {
 614                 CERROR("filter obd: no type!\n");
 615                 return -ENOENT;
 616         }
 617
 618         oa->o_id = filter_next_id(obd);
 619
 620         filter_id(name, oa->o_id, oa->o_mode);
 621         mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
 622         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 623         file = filp_open(name, O_RDONLY | O_CREAT, mode);
 624         pop_ctxt(&saved);
 625         if (IS_ERR(file)) {
 626                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(file));
 627                 return -ENOENT;
 628         }
 629
 630         /* Set flags for fields we have set in the inode struct */
 631         oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 632                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
 633         filter_from_inode(oa, file->f_dentry->d_inode, oa->o_valid);
 634         filp_close(file, 0);
 635
 636         return 0;
 637 }
 638
 639 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
 640                           struct lov_stripe_md *ea)
 641 {
 642         struct obd_device *obd;
 643         struct filter_obd *filter;
 644         struct obd_run_ctxt saved;
 645         struct inode *inode;
 646         struct dentry *dir_dentry, *object_dentry;
 647         int rc;
 648         ENTRY;
 649
 650         obd = class_conn2obd(conn);
 651         if (!obd) {
 652                 CERROR("invalid client %Lx\n", conn->addr);
 653                 RETURN(-EINVAL);
 654         }
 655
 656         CDEBUG(D_INODE, "destroying object %Ld\n", oa->o_id);
 657
 658         dir_dentry = filter_parent(obd, oa->o_mode);
 659         down(&dir_dentry->d_inode->i_sem);
 660
 661         object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
 662                                           oa->o_mode);
 663         if (IS_ERR(object_dentry))
 664                 GOTO(out, rc = -ENOENT);
 665
 666         inode = object_dentry->d_inode;
 667         if (inode == NULL) {
 668                 CERROR("trying to destroy negative inode %Ld!\n", oa->o_id);
 669                 GOTO(out, rc = -ENOENT);
 670         }
 671
 672         if (inode->i_nlink != 1) {
 673                 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
 674                 LBUG();
 675                 inode->i_nlink = 1;
 676         }
 677         inode->i_mode = S_IFREG;
 678
 679         filter = &obd->u.filter;
 680         push_ctxt(&saved, &filter->fo_ctxt);
 681
 682         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 683         pop_ctxt(&saved);
 684         f_dput(object_dentry);
 685
 686         EXIT;
 687 out:
 688         up(&dir_dentry->d_inode->i_sem);
 689         return rc;
 690 }
 691
 692 /* NB count and offset are used for punch, but not truncate */
 693 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
 694                            struct lov_stripe_md *md,
 695                            obd_off start, obd_off end)
 696 {
 697         int error;
 698         ENTRY;
 699
 700         if (end != 0xffffffffffffffff)
 701                 CERROR("PUNCH not supported, only truncate works\n");
 702
 703         CDEBUG(D_INODE, "calling truncate for object #%Ld, valid = %x, "
 704                "o_size = %Ld\n", oa->o_id, oa->o_valid, start);
 705         oa->o_size = start;
 706         error = filter_setattr(conn, oa, NULL);
 707         RETURN(error);
 708 }
 709
 710 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
 711                               struct lov_stripe_md *md, obd_count oa_bufs,
 712                               struct brw_page *pga, brw_callback_t callback,
 713                               struct io_cb_data *data)
 714 {
 715         struct obd_run_ctxt      saved;
 716         struct super_block      *sb;
 717         int                      pnum;          /* index to pages (bufs) */
 718         unsigned long            retval;
 719         int                      error;
 720         struct file             *file;
 721         struct obd_device       *obd = class_conn2obd(conn);
 722         int pg;
 723         ENTRY;
 724
 725         if (!obd) {
 726                 CDEBUG(D_IOCTL, "invalid client %Lx\n", conn->addr);
 727                 RETURN(-EINVAL);
 728         }
 729
 730         sb = obd->u.filter.fo_sb;
 731         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 732         pnum = 0; /* pnum indexes buf 0..num_pages */
 733
 734         file = filter_obj_open(obd, md->lmd_object_id, S_IFREG);
 735         if (IS_ERR(file))
 736                 GOTO(out, retval = PTR_ERR(file));
 737
 738         /* count doubles as retval */
 739         for (pg = 0; pg < oa_bufs; pg++) {
 740                 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
 741                        ") off count ("LPU64",%d)\n",
 742                        cmd, pnum, file->f_dentry->d_inode->i_ino,
 743                        pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
 744                        (int)pga[pnum].count);
 745                 if (cmd & OBD_BRW_WRITE) {
 746                         loff_t off;
 747                         char *buffer;
 748                         off = pga[pnum].off;
 749                         buffer = kmap(pga[pnum].pg);
 750                         retval = file->f_op->write(file, buffer,
 751                                                    pga[pnum].count,
 752                                                    &off);
 753                         kunmap(pga[pnum].pg);
 754                         CDEBUG(D_INODE, "retval %ld\n", retval);
 755                 } else {
 756                         loff_t off = pga[pnum].off;
 757                         char *buffer = kmap(pga[pnum].pg);
 758
 759                         if (off >= file->f_dentry->d_inode->i_size) {
 760                                 memset(buffer, 0, pga[pnum].count);
 761                                 retval = pga[pnum].count;
 762                         } else {
 763                                 retval = file->f_op->read(file, buffer,
 764                                                           pga[pnum].count, &off);
 765                         }
 766                         kunmap(pga[pnum].pg);
 767
 768                         if (retval != pga[pnum].count) {
 769                                 filp_close(file, 0);
 770                                 GOTO(out, retval = -EIO);
 771                         }
 772                         CDEBUG(D_INODE, "retval %ld\n", retval);
 773                 }
 774                 pnum++;
 775         }
 776         /* sizes and blocks are set by generic_file_write */
 777         /* ctimes/mtimes will follow with a setattr call */
 778         filp_close(file, 0);
 779
 780         /* XXX: do something with callback if it is set? */
 781
 782         EXIT;
 783 out:
 784         pop_ctxt(&saved);
 785         error = (retval >= 0) ? 0 : retval;
 786         return error;
 787 }
 788
 789 /*
 790  * Calculate the number of buffer credits needed to write multiple pages in
 791  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
 792  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
 793  *
 794  * See comment above ext3_writepage_trans_blocks for details.  We assume
 795  * no data journaling is being done, but it does allow for all of the pages
 796  * being non-contiguous.  If we are guaranteed contiguous pages we could
 797  * reduce the number of (d)indirect blocks a lot.
 798  *
 799  * With N blocks per page and P pages, for each inode we have at most:
 800  * N*P indirect
 801  * min(N*P, blocksize/4 + 1) dindirect blocks
 802  * 1 tindirect
 803  *
 804  * For the entire filesystem, we have at most:
 805  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 806  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 807  * 1 inode block
 808  * 1 superblock
 809  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 810  */
 811 static int ext3_credits_needed(struct super_block *sb, int objcount,
 812                                struct obd_ioobj *obj)
 813 {
 814         struct obd_ioobj *o = obj;
 815         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 816         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 817         int nbitmaps = 0;
 818         int ngdblocks = 0;
 819         int needed = objcount + 1;
 820         int i;
 821
 822         for (i = 0; i < objcount; i++, o++) {
 823                 int nblocks = o->ioo_bufcnt * blockpp;
 824                 int ndindirect = min(nblocks, addrpp + 1);
 825                 int nindir = nblocks + ndindirect + 1;
 826
 827                 nbitmaps += nindir + nblocks;
 828                 ngdblocks += nindir + nblocks;
 829
 830                 needed += nindir;
 831         }
 832
 833         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 834                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 835         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 836                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 837
 838         needed += nbitmaps + ngdblocks;
 839
 840 #ifdef CONFIG_QUOTA
 841         /* We assume that there will be 1 bit set in s_dquot.flags for each
 842          * quota file that is active.  This is at least true for now.
 843          */
 844         needed += hweight32(sb_any_quota_enabled(sb)) *
 845                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 846 #endif
 847
 848         return needed;
 849 }
 850
 851 /* We have to start a huge journal transaction here to hold all of the
 852  * metadata for the pages being written here.  This is necessitated by
 853  * the fact that we do lots of prepare_write operations before we do
 854  * any of the matching commit_write operations, so even if we split
 855  * up to use "smaller" transactions none of them could complete until
 856  * all of them were opened.  By having a single journal transaction,
 857  * we eliminate duplicate reservations for common blocks like the
 858  * superblock and group descriptors or bitmaps.
 859  *
 860  * We will start the transaction here, but each prepare_write will
 861  * add a refcount to the transaction, and each commit_write will
 862  * remove a refcount.  The transaction will be closed when all of
 863  * the pages have been written.
 864  */
 865 static void *ext3_filter_journal_start(struct filter_obd *filter,
 866                                        int objcount, struct obd_ioobj *obj,
 867                                        int niocount, struct niobuf_remote *nb)
 868 {
 869         journal_t *journal = NULL;
 870         handle_t *handle = NULL;
 871         int needed;
 872
 873         /* Assumes ext3 and extN have same sb_info layout, but avoids issues
 874          * with having extN built properly before filterobd for now.
 875          */
 876         journal = EXT3_SB(filter->fo_sb)->s_journal;
 877         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
 878
 879         /* The number of blocks we could _possibly_ dirty can very large.
 880          * We reduce our request if it is absurd (and we couldn't get that
 881          * many credits for a single handle anyways).
 882          *
 883          * At some point we have to limit the size of I/Os sent at one time,
 884          * increase the size of the journal, or we have to calculate the
 885          * actual journal requirements more carefully by checking all of
 886          * the blocks instead of being maximally pessimistic.  It remains to
 887          * be seen if this is a real problem or not.
 888          */
 889         if (needed > journal->j_max_transaction_buffers) {
 890                 CERROR("want too many journal credits (%d) using %d instead\n",
 891                        needed, journal->j_max_transaction_buffers);
 892                 needed = journal->j_max_transaction_buffers;
 893         }
 894
 895         handle = journal_start(journal, needed);
 896         if (IS_ERR(handle))
 897                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 898                        PTR_ERR(handle));
 899
 900         return(handle);
 901 }
 902
 903 static void *filter_journal_start(void **journal_save,
 904                                   struct filter_obd *filter,
 905                                   int objcount, struct obd_ioobj *obj,
 906                                   int niocount, struct niobuf_remote *nb)
 907 {
 908         void *handle = NULL;
 909
 910         /* This may not be necessary - we probably never have a
 911          * transaction started when we enter here, so we can
 912          * remove the saving of the journal state entirely.
 913          * For now leave it in just to see if it ever happens.
 914          */
 915         *journal_save = current->journal_info;
 916         if (*journal_save) {
 917                 CERROR("Already have handle %p???\n", *journal_save);
 918                 LBUG();
 919                 current->journal_info = NULL;
 920         }
 921
 922         if (!strcmp(filter->fo_fstype, "ext3") ||
 923             !strcmp(filter->fo_fstype, "extN"))
 924                 handle = ext3_filter_journal_start(filter, objcount, obj,
 925                                                    niocount, nb);
 926         return handle;
 927 }
 928
 929 static int ext3_filter_journal_stop(void *handle)
 930 {
 931         int rc;
 932
 933         /* We got a refcount on the handle for each call to prepare_write,
 934          * so we can drop the "parent" handle here to avoid the need for
 935          * osc to call back into filterobd to close the handle.  The
 936          * remaining references will be dropped in commit_write.
 937          */
 938         rc = journal_stop((handle_t *)handle);
 939
 940         return rc;
 941 }
 942
 943 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
 944                                void *handle)
 945 {
 946         int rc = 0;
 947
 948         if (!strcmp(filter->fo_fstype, "ext3") ||
 949             !strcmp(filter->fo_fstype, "extN"))
 950                 rc = ext3_filter_journal_stop(handle);
 951
 952         if (rc)
 953                 CERROR("error on journal stop: rc = %d\n", rc);
 954
 955         current->journal_info = journal_save;
 956
 957         return rc;
 958 }
 959
 960 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
 961                                    struct niobuf_local *lnb, int *pglocked)
 962 {
 963         struct address_space *mapping = inode->i_mapping;
 964         struct page *page;
 965         int rc;
 966
 967         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
 968         if (*pglocked)
 969                 page = grab_cache_page_nowait(mapping, index); /* locked page */
 970         else
 971                 page = grab_cache_page(mapping, index); /* locked page */
 972
 973
 974         /* This page is currently locked, so get a temporary page instead. */
 975         /* XXX I believe this is a very dangerous thing to do - consider if
 976          *     we had multiple writers for the same file (definitely the case
 977          *     if we are using this codepath).  If writer A locks the page,
 978          *     writer B writes to a copy (as here), writer A drops the page
 979          *     lock, and writer C grabs the lock before B does, then B will
 980          *     later overwrite the data from C, even if C had LDLM locked
 981          *     and initiated the write after B did.
 982          */
 983         if (!page) {
 984                 unsigned long addr;
 985                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
 986                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
 987                 if (!addr) {
 988                         CERROR("no memory for a temp page\n");
 989                         LBUG();
 990                         GOTO(err, rc = -ENOMEM);
 991                 }
 992                 page = virt_to_page(addr);
 993                 kmap(page);
 994                 page->index = index;
 995                 lnb->flags |= N_LOCAL_TEMP_PAGE;
 996         } else if (!IS_ERR(page)) {
 997                 (*pglocked)++;
 998                 kmap(page);
 999
1000                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1001                  * a no-op for most filesystems, because we write the whole
1002                  * page.  For partial-page I/O this will read in the page.
1003                  */
1004                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1005                 if (rc) {
1006                         CERROR("page index %lu, rc = %d\n", index, rc);
1007                         if (rc != -ENOSPC)
1008                                 LBUG();
1009                         GOTO(err_unlock, rc);
1010                 }
1011                 /* XXX not sure if we need this if we are overwriting page */
1012                 if (PageError(page)) {
1013                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1014                         LBUG();
1015                         GOTO(err_unlock, rc = -EIO);
1016                 }
1017         }
1018         return page;
1019
1020 err_unlock:
1021         unlock_page(page);
1022         lustre_put_page(page);
1023 err:
1024         return ERR_PTR(rc);
1025 }
1026
1027 /*
1028  * We need to balance prepare_write() calls with commit_write() calls.
1029  * If the page has been prepared, but we have no data for it, we don't
1030  * want to overwrite valid data on disk, but we still need to zero out
1031  * data for space which was newly allocated.  Like part of what happens
1032  * in __block_prepare_write() for newly allocated blocks.
1033  *
1034  * XXX currently __block_prepare_write() creates buffers for all the
1035  *     pages, and the filesystems mark these buffers as BH_New if they
1036  *     were newly allocated from disk. We use the BH_New flag similarly.
1037  */
1038 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1039                                int err)
1040 {
1041         if (err) {
1042                 unsigned block_start, block_end;
1043                 struct buffer_head *bh, *head = page->buffers;
1044                 unsigned blocksize = head->b_size;
1045                 void *addr = page_address(page);
1046
1047                 /* debugging: just seeing if this ever happens */
1048                 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1049                        page->index, page->mapping->host->i_ino, err);
1050
1051                 /* Currently one buffer per page, but in the future... */
1052                 for (bh = head, block_start = 0; bh != head || !block_start;
1053                      block_start = block_end, bh = bh->b_this_page) {
1054                         block_end = block_start + blocksize;
1055                         if (buffer_new(bh))
1056                                 memset(addr + block_start, 0, blocksize);
1057                 }
1058         }
1059
1060         return lustre_commit_write(page, from, to);
1061 }
1062
1063 static int filter_preprw(int cmd, struct lustre_handle *conn,
1064                          int objcount, struct obd_ioobj *obj,
1065                          int niocount, struct niobuf_remote *nb,
1066                          struct niobuf_local *res, void **desc_private)
1067 {
1068         struct obd_run_ctxt saved;
1069         struct obd_device *obd;
1070         struct obd_ioobj *o = obj;
1071         struct niobuf_remote *b = nb;
1072         struct niobuf_local *r = res;
1073         void *journal_save = NULL;
1074         int pglocked = 0;
1075         int rc = 0;
1076         int i;
1077         ENTRY;
1078
1079         obd = class_conn2obd(conn);
1080         if (!obd) {
1081                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1082                 RETURN(-EINVAL);
1083         }
1084         memset(res, 0, sizeof(*res) * niocount);
1085
1086         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1087
1088         if (cmd & OBD_BRW_WRITE) {
1089                 *desc_private = filter_journal_start(&journal_save,
1090                                                      &obd->u.filter,
1091                                                      objcount, obj, niocount,
1092                                                      nb);
1093                 if (IS_ERR(*desc_private))
1094                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1095         }
1096
1097         for (i = 0; i < objcount; i++, o++) {
1098                 struct dentry *dentry;
1099                 struct inode *inode;
1100                 int j;
1101
1102                 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1103                                            o->ioo_id, S_IFREG);
1104                 if (IS_ERR(dentry))
1105                         GOTO(out_clean, rc = PTR_ERR(dentry));
1106                 inode = dentry->d_inode;
1107                 if (!inode) {
1108                         CERROR("trying to BRW to non-existent file %Ld\n",
1109                                (unsigned long long)o->ioo_id);
1110                         f_dput(dentry);
1111                         GOTO(out_clean, rc = -ENOENT);
1112                 }
1113
1114                 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1115                         unsigned long index = b->offset >> PAGE_SHIFT;
1116                         struct page *page;
1117
1118                         if (j == 0)
1119                                 r->dentry = dentry;
1120                         else
1121                                 r->dentry = dget(dentry);
1122
1123                         if (cmd & OBD_BRW_WRITE)
1124                                 page = filter_get_page_write(inode, index, r,
1125                                                              &pglocked);
1126                         else
1127                                 page = lustre_get_page_read(inode, index);
1128
1129                         if (IS_ERR(page)) {
1130                                 f_dput(dentry);
1131                                 GOTO(out_clean, rc = PTR_ERR(page));
1132                         }
1133
1134                         r->addr = page_address(page);
1135                         r->offset = b->offset;
1136                         r->page = page;
1137                         r->len = b->len;
1138                 }
1139         }
1140
1141 out_stop:
1142         if (cmd & OBD_BRW_WRITE) {
1143                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1144                                               *desc_private);
1145                 if (!rc)
1146                         rc = err;
1147         }
1148 out_ctxt:
1149         pop_ctxt(&saved);
1150         RETURN(rc);
1151 out_clean:
1152         while (r-- > res) {
1153                 CERROR("error cleanup on brw\n");
1154                 f_dput(r->dentry);
1155                 if (cmd & OBD_BRW_WRITE)
1156                         filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1157                 else
1158                         lustre_put_page(r->page);
1159         }
1160         goto out_stop;
1161 }
1162
1163 static int filter_write_locked_page(struct niobuf_local *lnb)
1164 {
1165         struct page *lpage;
1166         int rc;
1167
1168         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1169         if (IS_ERR(lpage)) {
1170                 /* It is highly unlikely that we would ever get an error here.
1171                  * The page we want to get was previously locked, so it had to
1172                  * have already allocated the space, and we were just writing
1173                  * over the same data, so there would be no hole in the file.
1174                  *
1175                  * XXX: possibility of a race with truncate could exist, need
1176                  *      to check that.  There are no guarantees w.r.t.
1177                  *      write order even on a local filesystem, although the
1178                  *      normal response would be to return the number of bytes
1179                  *      successfully written and leave the rest to the app.
1180                  */
1181                 rc = PTR_ERR(lpage);
1182                 CERROR("error getting locked page index %ld: rc = %d\n",
1183                        lnb->page->index, rc);
1184                 GOTO(out, rc);
1185         }
1186
1187         /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1188          * lustre_commit_write() below, lnb->page was kmapped previously in
1189          * filter_get_page_write() and kunmapped in lustre_put_page() below.
1190          */
1191         memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1192         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1193         if (rc)
1194                 CERROR("error committing locked page %ld: rc = %d\n",
1195                        lnb->page->index, rc);
1196 out:
1197         lustre_put_page(lnb->page);
1198
1199         return rc;
1200 }
1201
1202 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1203                            int objcount, struct obd_ioobj *obj,
1204                            int niocount, struct niobuf_local *res,
1205                            void *private)
1206 {
1207         struct obd_run_ctxt saved;
1208         struct obd_ioobj *o;
1209         struct niobuf_local *r;
1210         struct obd_device *obd = class_conn2obd(conn);
1211         void *journal_save;
1212         int found_locked = 0;
1213         int rc = 0;
1214         int i;
1215         ENTRY;
1216
1217         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1218         journal_save = current->journal_info;
1219         LASSERT(!journal_save);
1220
1221         current->journal_info = private;
1222         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1223                 int j;
1224                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1225                         struct page *page = r->page;
1226
1227                         if (!page)
1228                                 LBUG();
1229
1230                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1231                                 found_locked++;
1232                                 continue;
1233                         }
1234
1235                         if (cmd & OBD_BRW_WRITE) {
1236                                 int err = filter_commit_write(page, 0,
1237                                                               r->len, 0);
1238
1239                                 if (!rc)
1240                                         rc = err;
1241                         } else
1242                                 lustre_put_page(page);
1243
1244                         f_dput(r->dentry);
1245                 }
1246         }
1247         current->journal_info = journal_save;
1248
1249         if (!found_locked)
1250                 goto out_ctxt;
1251
1252         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1253                 int j;
1254                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1255                         int err;
1256                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1257                                 continue;
1258
1259                         err = filter_write_locked_page(r);
1260                         if (!rc)
1261                                 rc = err;
1262                         f_dput(r->dentry);
1263                 }
1264         }
1265
1266 out_ctxt:
1267         pop_ctxt(&saved);
1268         RETURN(rc);
1269 }
1270
1271 static int filter_statfs(struct lustre_handle *conn, struct statfs *statfs)
1272 {
1273         struct obd_device *obd = class_conn2obd(conn);
1274
1275         ENTRY;
1276         RETURN(vfs_statfs(obd->u.filter.fo_sb, statfs));
1277 }
1278
1279 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1280                            void *key, obd_count *vallen, void **val)
1281 {
1282         struct obd_device *obd;
1283         struct obd_export * export;
1284         ENTRY;
1285
1286         if (!(export = class_conn2export(conn))) {
1287                 CDEBUG(D_IOCTL, "invalid client %Lx\n", conn->addr);
1288                 RETURN(-EINVAL);
1289         }
1290
1291         obd = class_conn2obd(conn);
1292
1293         if ( keylen == strlen("blocksize") &&
1294              memcmp(key, "blocksize", keylen) == 0 ) {
1295                 *vallen = sizeof(long);
1296                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1297                 RETURN(0);
1298         }
1299
1300         if ( keylen == strlen("blocksize_bits") &&
1301              memcmp(key, "blocksize_bits", keylen) == 0 ){
1302                 *vallen = sizeof(long);
1303                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1304                 RETURN(0);
1305         }
1306
1307         if ( keylen == strlen("root_ino") &&
1308              memcmp(key, "root_ino", keylen) == 0 ){
1309                 *vallen = sizeof(obd_id);
1310                 *val = (void *)(obd_id)FILTER_ROOTINO;
1311                 RETURN(0);
1312         }
1313
1314         CDEBUG(D_IOCTL, "invalid key\n");
1315         RETURN(-EINVAL);
1316 }
1317
1318 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1319                   struct lustre_handle *src_conn, struct obdo *src,
1320                   obd_size count, obd_off offset)
1321 {
1322         struct page *page;
1323         struct lov_stripe_md srcmd, dstmd;
1324         unsigned long index = 0;
1325         int err = 0;
1326
1327         memset(&srcmd, 0, sizeof(srcmd));
1328         memset(&dstmd, 0, sizeof(dstmd));
1329         srcmd.lmd_object_id = src->o_id;
1330         dstmd.lmd_object_id = dst->o_id;
1331
1332         ENTRY;
1333         CDEBUG(D_INFO, "src: ino %Ld blocks %Ld, size %Ld, dst: ino %Ld\n",
1334                (unsigned long long)src->o_id, (unsigned long long)src->o_blocks,
1335                (unsigned long long)src->o_size, (unsigned long long)dst->o_id);
1336         page = alloc_page(GFP_USER);
1337         if (page == NULL)
1338                 RETURN(-ENOMEM);
1339
1340         while (TryLockPage(page))
1341                 ___wait_on_page(page);
1342
1343         /* XXX with brw vector I/O, we could batch up reads and writes here,
1344          *     all we need to do is allocate multiple pages to handle the I/Os
1345          *     and arrays to handle the request parameters.
1346          */
1347         while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1348                 struct brw_page pg;
1349                 struct io_cb_data *cbd = ll_init_cb();
1350
1351                 if (!cbd) {
1352                         err = -ENOMEM;
1353                         EXIT;
1354                         break;
1355                 }
1356
1357                 pg.pg = page;
1358                 pg.count = PAGE_SIZE;
1359                 pg.off = (page->index) << PAGE_SHIFT;
1360                 pg.flag = 0;
1361
1362                 page->index = index;
1363                 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1364                               ll_sync_io_cb, cbd);
1365
1366                 if ( err ) {
1367                         EXIT;
1368                         break;
1369                 }
1370
1371                 cbd = ll_init_cb();
1372                 if (!cbd) {
1373                         err = -ENOMEM;
1374                         EXIT;
1375                         break;
1376                 }
1377                 pg.flag = OBD_BRW_CREATE;
1378                 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1379
1380                 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1381                               ll_sync_io_cb, cbd);
1382
1383                 /* XXX should handle dst->o_size, dst->o_blocks here */
1384                 if ( err ) {
1385                         EXIT;
1386                         break;
1387                 }
1388
1389                 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1390
1391                 index++;
1392         }
1393         dst->o_size = src->o_size;
1394         dst->o_blocks = src->o_blocks;
1395         dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1396         unlock_page(page);
1397         __free_page(page);
1398
1399         RETURN(err);
1400 }
1401
1402
1403 static struct obd_ops filter_obd_ops = {
1404         o_get_info:    filter_get_info,
1405         o_setup:       filter_setup,
1406         o_cleanup:     filter_cleanup,
1407         o_connect:     filter_connect,
1408         o_disconnect:  filter_disconnect,
1409         o_statfs:      filter_statfs,
1410         o_getattr:     filter_getattr,
1411         o_create:      filter_create,
1412         o_setattr:     filter_setattr,
1413         o_destroy:     filter_destroy,
1414         o_open:        filter_open,
1415         o_close:       filter_close,
1416         o_brw:         filter_pgcache_brw,
1417         o_punch:       filter_truncate,
1418         o_preprw:      filter_preprw,
1419         o_commitrw:    filter_commitrw
1420 #if 0
1421         o_preallocate: filter_preallocate_inodes,
1422         o_migrate:     filter_migrate,
1423         o_copy:        filter_copy_data,
1424         o_iterate:     filter_iterate
1425 #endif
1426 };
1427
1428
1429 static int __init obdfilter_init(void)
1430 {
1431         printk(KERN_INFO "Filtering OBD driver  v0.001, info@clusterfs.com\n");
1432         return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1433 }
1434
1435 static void __exit obdfilter_exit(void)
1436 {
1437         class_unregister_type(OBD_FILTER_DEVICENAME);
1438 }
1439
1440 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1441 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1442 MODULE_LICENSE("GPL");
1443
1444 module_init(obdfilter_init);
1445 module_exit(obdfilter_exit);