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