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/obdfilter/filter.c
   5  *
   6  * Copyright (C) 2001, 2002 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  * and Andreas Dilger <adilger@clusterfs.com>
  13  */
  14
  15 #define EXPORT_SYMTAB
  16 #define DEBUG_SUBSYSTEM S_FILTER
  17
  18 #include <linux/module.h>
  19 #include <linux/lustre_dlm.h>
  20 #include <linux/obd_filter.h>
  21 #include <linux/ext3_jbd.h>
  22 #include <linux/quotaops.h>
  23 #include <linux/init.h>
  24
  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                           obd_uuid_t 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         ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
 407                            "filter_ldlm_client", &obddev->obd_ldlm_client);
 408
 409         RETURN(0);
 410
 411 err_kfree:
 412         kfree(filter->fo_fstype);
 413         unlock_kernel();
 414         mntput(filter->fo_vfsmnt);
 415         filter->fo_sb = 0;
 416         lock_kernel();
 417
 418 err_dec:
 419         MOD_DEC_USE_COUNT;
 420         return err;
 421 }
 422
 423
 424 static int filter_cleanup(struct obd_device * obddev)
 425 {
 426         struct super_block *sb;
 427         ENTRY;
 428
 429         if (!list_empty(&obddev->obd_exports)) {
 430                 CERROR("still has clients!\n");
 431                 class_disconnect_all(obddev);
 432                 if (!list_empty(&obddev->obd_exports)) {
 433                         CERROR("still has exports after forced cleanup?\n");
 434                         RETURN(-EBUSY);
 435                 }
 436         }
 437
 438         ldlm_namespace_free(obddev->obd_namespace);
 439
 440         sb = obddev->u.filter.fo_sb;
 441         if (!obddev->u.filter.fo_sb)
 442                 RETURN(0);
 443
 444         filter_post(obddev);
 445
 446         shrink_dcache_parent(sb->s_root);
 447         unlock_kernel();
 448         mntput(obddev->u.filter.fo_vfsmnt);
 449         obddev->u.filter.fo_sb = 0;
 450         kfree(obddev->u.filter.fo_fstype);
 451
 452         lock_kernel();
 453
 454         MOD_DEC_USE_COUNT;
 455         RETURN(0);
 456 }
 457
 458
 459 static inline void filter_from_inode(struct obdo *oa, struct inode *inode,
 460                                      int valid)
 461 {
 462         int type = oa->o_mode & S_IFMT;
 463         ENTRY;
 464
 465         CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
 466                inode->i_ino, inode, (long)oa->o_id, valid);
 467         /* Don't copy the inode number in place of the object ID */
 468         obdo_from_inode(oa, inode, valid);
 469         oa->o_mode &= ~S_IFMT;
 470         oa->o_mode |= type;
 471
 472         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
 473                 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
 474                 oa->o_rdev = rdev;
 475                 oa->o_valid |= OBD_MD_FLRDEV;
 476         }
 477
 478         EXIT;
 479 }
 480
 481 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
 482                           struct lov_stripe_md *md)
 483 {
 484         struct obd_device *obddev = class_conn2obd(conn);
 485         struct dentry *dentry;
 486         ENTRY;
 487
 488         if (!class_conn2export(conn)) {
 489                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 490                 RETURN(-EINVAL);
 491         }
 492
 493         obddev = class_conn2obd(conn);
 494         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 495                                    oa->o_id, oa->o_mode);
 496         if (IS_ERR(dentry))
 497                 RETURN(PTR_ERR(dentry));
 498
 499         filter_from_inode(oa, dentry->d_inode, oa->o_valid);
 500
 501         f_dput(dentry);
 502         RETURN(0);
 503 }
 504
 505 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
 506                           struct lov_stripe_md *md)
 507 {
 508         struct obd_run_ctxt saved;
 509         struct obd_device *obd = class_conn2obd(conn);
 510         struct dentry *dentry;
 511         struct iattr iattr;
 512         struct inode *inode;
 513         int rc;
 514         ENTRY;
 515
 516         iattr_from_obdo(&iattr, oa, oa->o_valid);
 517         iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
 518         dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
 519                                    oa->o_id, iattr.ia_mode);
 520         if (IS_ERR(dentry))
 521                 RETURN(PTR_ERR(dentry));
 522
 523         inode = dentry->d_inode;
 524         if (!inode) {
 525                 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
 526                 GOTO(out_setattr, rc = -ENOENT);
 527         }
 528
 529         lock_kernel();
 530         if (iattr.ia_valid & ATTR_SIZE)
 531                 down(&inode->i_sem);
 532         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 533         if (inode->i_op->setattr)
 534                 rc = inode->i_op->setattr(dentry, &iattr);
 535         else
 536                 rc = inode_setattr(inode, &iattr);
 537         pop_ctxt(&saved);
 538         if (iattr.ia_valid & ATTR_SIZE) {
 539                 up(&inode->i_sem);
 540                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 541                 obdo_from_inode(oa, inode, oa->o_valid);
 542         }
 543         unlock_kernel();
 544
 545 out_setattr:
 546         f_dput(dentry);
 547         RETURN(rc);
 548 }
 549
 550 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
 551                           struct lov_stripe_md *ea)
 552 {
 553         struct obd_export *export;
 554         struct obd_device *obd;
 555         struct dentry *dentry;
 556         /* ENTRY; */
 557
 558         export = class_conn2export(conn);
 559         if (!export) {
 560                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 561                 RETURN(-EINVAL);
 562         }
 563
 564         obd = export->exp_obd;
 565         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 566                                    oa->o_id, oa->o_mode);
 567         if (IS_ERR(dentry))
 568                 RETURN(PTR_ERR(dentry));
 569
 570         filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 571                           OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 572
 573         return 0;
 574 } /* filter_open */
 575
 576 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
 577                         struct lov_stripe_md *ea)
 578 {
 579         struct obd_device *obd;
 580         struct dentry *dentry;
 581         /* ENTRY; */
 582
 583         obd = class_conn2obd(conn);
 584         if (!obd) {
 585                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 586                 RETURN(-EINVAL);
 587         }
 588
 589         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 590                                    oa->o_id, oa->o_mode);
 591         if (IS_ERR(dentry))
 592                 RETURN(PTR_ERR(dentry));
 593         LASSERT(atomic_read(&dentry->d_count) > 1);
 594
 595         f_dput(dentry);  /* for the close */
 596         f_dput(dentry);  /* for this call */
 597         return 0;
 598 } /* filter_close */
 599
 600 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
 601                          struct lov_stripe_md **ea)
 602 {
 603         char name[64];
 604         struct obd_run_ctxt saved;
 605         struct file *file;
 606         int mode;
 607         struct obd_device *obd = class_conn2obd(conn);
 608         struct iattr;
 609         ENTRY;
 610
 611         if (!obd) {
 612                 CERROR("invalid client "LPX64"\n", conn->addr);
 613                 return -EINVAL;
 614         }
 615
 616         if (!(oa->o_mode && S_IFMT)) {
 617                 CERROR("filter obd: no type!\n");
 618                 return -ENOENT;
 619         }
 620
 621         oa->o_id = filter_next_id(obd);
 622
 623         filter_id(name, oa->o_id, oa->o_mode);
 624         mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
 625         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 626         file = filp_open(name, O_RDONLY | O_CREAT, mode);
 627         pop_ctxt(&saved);
 628         if (IS_ERR(file)) {
 629                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(file));
 630                 return -ENOENT;
 631         }
 632
 633         /* Set flags for fields we have set in the inode struct */
 634         oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 635                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
 636         filter_from_inode(oa, file->f_dentry->d_inode, oa->o_valid);
 637         filp_close(file, 0);
 638
 639         return 0;
 640 }
 641
 642 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
 643                           struct lov_stripe_md *ea)
 644 {
 645         struct obd_device *obd;
 646         struct filter_obd *filter;
 647         struct obd_run_ctxt saved;
 648         struct inode *inode;
 649         struct dentry *dir_dentry, *object_dentry;
 650         int rc;
 651         ENTRY;
 652
 653         obd = class_conn2obd(conn);
 654         if (!obd) {
 655                 CERROR("invalid client "LPX64"\n", conn->addr);
 656                 RETURN(-EINVAL);
 657         }
 658
 659         CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
 660
 661         dir_dentry = filter_parent(obd, oa->o_mode);
 662         down(&dir_dentry->d_inode->i_sem);
 663
 664         object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
 665                                           oa->o_mode);
 666         if (IS_ERR(object_dentry))
 667                 GOTO(out, rc = -ENOENT);
 668
 669         inode = object_dentry->d_inode;
 670         if (inode == NULL) {
 671                 CERROR("trying to destroy negative inode "LPX64"!\n", oa->o_id);
 672                 GOTO(out, rc = -ENOENT);
 673         }
 674
 675         if (inode->i_nlink != 1) {
 676                 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
 677                 LBUG();
 678                 inode->i_nlink = 1;
 679         }
 680         inode->i_mode = S_IFREG;
 681
 682         filter = &obd->u.filter;
 683         push_ctxt(&saved, &filter->fo_ctxt);
 684
 685         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 686         pop_ctxt(&saved);
 687         f_dput(object_dentry);
 688
 689         EXIT;
 690 out:
 691         up(&dir_dentry->d_inode->i_sem);
 692         return rc;
 693 }
 694
 695 /* NB count and offset are used for punch, but not truncate */
 696 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
 697                            struct lov_stripe_md *md,
 698                            obd_off start, obd_off end)
 699 {
 700         int error;
 701         ENTRY;
 702
 703         if (end != OBD_PUNCH_EOF)
 704                 CERROR("PUNCH not supported, only truncate works\n");
 705
 706         CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
 707                "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
 708         oa->o_size = start;
 709         error = filter_setattr(conn, oa, NULL);
 710         RETURN(error);
 711 }
 712
 713 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
 714                               struct lov_stripe_md *lsm, obd_count oa_bufs,
 715                               struct brw_page *pga, brw_callback_t callback,
 716                               struct io_cb_data *data)
 717 {
 718         struct obd_run_ctxt      saved;
 719         struct super_block      *sb;
 720         int                      pnum;          /* index to pages (bufs) */
 721         unsigned long            retval;
 722         int                      error;
 723         struct file             *file;
 724         struct obd_device       *obd = class_conn2obd(conn);
 725         int pg;
 726         ENTRY;
 727
 728         if (!obd) {
 729                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
 730                 RETURN(-EINVAL);
 731         }
 732
 733         sb = obd->u.filter.fo_sb;
 734         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 735         pnum = 0; /* pnum indexes buf 0..num_pages */
 736
 737         file = filter_obj_open(obd, lsm->lsm_object_id, S_IFREG);
 738         if (IS_ERR(file))
 739                 GOTO(out, retval = PTR_ERR(file));
 740
 741         /* count doubles as retval */
 742         for (pg = 0; pg < oa_bufs; pg++) {
 743                 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
 744                        ") off count ("LPU64",%d)\n",
 745                        cmd, pnum, file->f_dentry->d_inode->i_ino,
 746                        pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
 747                        (int)pga[pnum].count);
 748                 if (cmd & OBD_BRW_WRITE) {
 749                         loff_t off;
 750                         char *buffer;
 751                         off = pga[pnum].off;
 752                         buffer = kmap(pga[pnum].pg);
 753                         retval = file->f_op->write(file, buffer,
 754                                                    pga[pnum].count,
 755                                                    &off);
 756                         kunmap(pga[pnum].pg);
 757                         CDEBUG(D_INODE, "retval %ld\n", retval);
 758                 } else {
 759                         loff_t off = pga[pnum].off;
 760                         char *buffer = kmap(pga[pnum].pg);
 761
 762                         if (off >= file->f_dentry->d_inode->i_size) {
 763                                 memset(buffer, 0, pga[pnum].count);
 764                                 retval = pga[pnum].count;
 765                         } else {
 766                                 retval = file->f_op->read(file, buffer,
 767                                                           pga[pnum].count, &off);
 768                         }
 769                         kunmap(pga[pnum].pg);
 770
 771                         if (retval != pga[pnum].count) {
 772                                 filp_close(file, 0);
 773                                 GOTO(out, retval = -EIO);
 774                         }
 775                         CDEBUG(D_INODE, "retval %ld\n", retval);
 776                 }
 777                 pnum++;
 778         }
 779         /* sizes and blocks are set by generic_file_write */
 780         /* ctimes/mtimes will follow with a setattr call */
 781         filp_close(file, 0);
 782
 783         /* XXX: do something with callback if it is set? */
 784
 785         EXIT;
 786 out:
 787         pop_ctxt(&saved);
 788         error = (retval >= 0) ? 0 : retval;
 789         return error;
 790 }
 791
 792 /*
 793  * Calculate the number of buffer credits needed to write multiple pages in
 794  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
 795  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
 796  *
 797  * See comment above ext3_writepage_trans_blocks for details.  We assume
 798  * no data journaling is being done, but it does allow for all of the pages
 799  * being non-contiguous.  If we are guaranteed contiguous pages we could
 800  * reduce the number of (d)indirect blocks a lot.
 801  *
 802  * With N blocks per page and P pages, for each inode we have at most:
 803  * N*P indirect
 804  * min(N*P, blocksize/4 + 1) dindirect blocks
 805  * 1 tindirect
 806  *
 807  * For the entire filesystem, we have at most:
 808  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 809  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 810  * 1 inode block
 811  * 1 superblock
 812  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 813  */
 814 static int ext3_credits_needed(struct super_block *sb, int objcount,
 815                                struct obd_ioobj *obj)
 816 {
 817         struct obd_ioobj *o = obj;
 818         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 819         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 820         int nbitmaps = 0;
 821         int ngdblocks = 0;
 822         int needed = objcount + 1;
 823         int i;
 824
 825         for (i = 0; i < objcount; i++, o++) {
 826                 int nblocks = o->ioo_bufcnt * blockpp;
 827                 int ndindirect = min(nblocks, addrpp + 1);
 828                 int nindir = nblocks + ndindirect + 1;
 829
 830                 nbitmaps += nindir + nblocks;
 831                 ngdblocks += nindir + nblocks;
 832
 833                 needed += nindir;
 834         }
 835
 836         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 837                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 838         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 839                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 840
 841         needed += nbitmaps + ngdblocks;
 842
 843 #ifdef CONFIG_QUOTA
 844         /* We assume that there will be 1 bit set in s_dquot.flags for each
 845          * quota file that is active.  This is at least true for now.
 846          */
 847         needed += hweight32(sb_any_quota_enabled(sb)) *
 848                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 849 #endif
 850
 851         return needed;
 852 }
 853
 854 /* We have to start a huge journal transaction here to hold all of the
 855  * metadata for the pages being written here.  This is necessitated by
 856  * the fact that we do lots of prepare_write operations before we do
 857  * any of the matching commit_write operations, so even if we split
 858  * up to use "smaller" transactions none of them could complete until
 859  * all of them were opened.  By having a single journal transaction,
 860  * we eliminate duplicate reservations for common blocks like the
 861  * superblock and group descriptors or bitmaps.
 862  *
 863  * We will start the transaction here, but each prepare_write will
 864  * add a refcount to the transaction, and each commit_write will
 865  * remove a refcount.  The transaction will be closed when all of
 866  * the pages have been written.
 867  */
 868 static void *ext3_filter_journal_start(struct filter_obd *filter,
 869                                        int objcount, struct obd_ioobj *obj,
 870                                        int niocount, struct niobuf_remote *nb)
 871 {
 872         journal_t *journal = NULL;
 873         handle_t *handle = NULL;
 874         int needed;
 875
 876         /* Assumes ext3 and extN have same sb_info layout, but avoids issues
 877          * with having extN built properly before filterobd for now.
 878          */
 879         journal = EXT3_SB(filter->fo_sb)->s_journal;
 880         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
 881
 882         /* The number of blocks we could _possibly_ dirty can very large.
 883          * We reduce our request if it is absurd (and we couldn't get that
 884          * many credits for a single handle anyways).
 885          *
 886          * At some point we have to limit the size of I/Os sent at one time,
 887          * increase the size of the journal, or we have to calculate the
 888          * actual journal requirements more carefully by checking all of
 889          * the blocks instead of being maximally pessimistic.  It remains to
 890          * be seen if this is a real problem or not.
 891          */
 892         if (needed > journal->j_max_transaction_buffers) {
 893                 CERROR("want too many journal credits (%d) using %d instead\n",
 894                        needed, journal->j_max_transaction_buffers);
 895                 needed = journal->j_max_transaction_buffers;
 896         }
 897
 898         handle = journal_start(journal, needed);
 899         if (IS_ERR(handle))
 900                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 901                        PTR_ERR(handle));
 902
 903         return(handle);
 904 }
 905
 906 static void *filter_journal_start(void **journal_save,
 907                                   struct filter_obd *filter,
 908                                   int objcount, struct obd_ioobj *obj,
 909                                   int niocount, struct niobuf_remote *nb)
 910 {
 911         void *handle = NULL;
 912
 913         /* This may not be necessary - we probably never have a
 914          * transaction started when we enter here, so we can
 915          * remove the saving of the journal state entirely.
 916          * For now leave it in just to see if it ever happens.
 917          */
 918         *journal_save = current->journal_info;
 919         if (*journal_save) {
 920                 CERROR("Already have handle %p???\n", *journal_save);
 921                 LBUG();
 922                 current->journal_info = NULL;
 923         }
 924
 925         if (!strcmp(filter->fo_fstype, "ext3") ||
 926             !strcmp(filter->fo_fstype, "extN"))
 927                 handle = ext3_filter_journal_start(filter, objcount, obj,
 928                                                    niocount, nb);
 929         return handle;
 930 }
 931
 932 static int ext3_filter_journal_stop(void *handle)
 933 {
 934         int rc;
 935
 936         /* We got a refcount on the handle for each call to prepare_write,
 937          * so we can drop the "parent" handle here to avoid the need for
 938          * osc to call back into filterobd to close the handle.  The
 939          * remaining references will be dropped in commit_write.
 940          */
 941         rc = journal_stop((handle_t *)handle);
 942
 943         return rc;
 944 }
 945
 946 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
 947                                void *handle)
 948 {
 949         int rc = 0;
 950
 951         if (!strcmp(filter->fo_fstype, "ext3") ||
 952             !strcmp(filter->fo_fstype, "extN"))
 953                 rc = ext3_filter_journal_stop(handle);
 954
 955         if (rc)
 956                 CERROR("error on journal stop: rc = %d\n", rc);
 957
 958         current->journal_info = journal_save;
 959
 960         return rc;
 961 }
 962
 963 static inline void lustre_put_page(struct page *page)
 964 {
 965         kunmap(page);
 966         page_cache_release(page);
 967 }
 968
 969 static struct page *
 970 lustre_get_page_read(struct inode *inode, unsigned long index)
 971 {
 972         struct address_space *mapping = inode->i_mapping;
 973         struct page *page;
 974         int rc;
 975
 976         page = read_cache_page(mapping, index,
 977                                (filler_t*)mapping->a_ops->readpage, NULL);
 978         if (!IS_ERR(page)) {
 979                 wait_on_page(page);
 980                 kmap(page);
 981                 if (!Page_Uptodate(page)) {
 982                         CERROR("page index %lu not uptodate\n", index);
 983                         GOTO(err_page, rc = -EIO);
 984                 }
 985                 if (PageError(page)) {
 986                         CERROR("page index %lu has error\n", index);
 987                         GOTO(err_page, rc = -EIO);
 988                 }
 989         }
 990         return page;
 991
 992 err_page:
 993         lustre_put_page(page);
 994         return ERR_PTR(rc);
 995 }
 996
 997 static struct page *
 998 lustre_get_page_write(struct inode *inode, unsigned long index)
 999 {
1000         struct address_space *mapping = inode->i_mapping;
1001         struct page *page;
1002         int rc;
1003
1004         page = grab_cache_page(mapping, index); /* locked page */
1005
1006         if (!IS_ERR(page)) {
1007                 kmap(page);
1008                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1009                  * a no-op for most filesystems, because we write the whole
1010                  * page.  For partial-page I/O this will read in the page.
1011                  */
1012                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1013                 if (rc) {
1014                         CERROR("page index %lu, rc = %d\n", index, rc);
1015                         if (rc != -ENOSPC)
1016                                 LBUG();
1017                         GOTO(err_unlock, rc);
1018                 }
1019                 /* XXX not sure if we need this if we are overwriting page */
1020                 if (PageError(page)) {
1021                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1022                         LBUG();
1023                         GOTO(err_unlock, rc = -EIO);
1024                 }
1025         }
1026         return page;
1027
1028 err_unlock:
1029         unlock_page(page);
1030         lustre_put_page(page);
1031         return ERR_PTR(rc);
1032 }
1033
1034 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1035 {
1036         struct inode *inode = page->mapping->host;
1037         int err;
1038
1039         err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1040         if (!err && IS_SYNC(inode))
1041                 err = waitfor_one_page(page);
1042
1043         //SetPageUptodate(page); // the client commit_write will do this
1044
1045         SetPageReferenced(page);
1046         unlock_page(page);
1047         lustre_put_page(page);
1048         return err;
1049 }
1050
1051 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1052                                    struct niobuf_local *lnb, int *pglocked)
1053 {
1054         struct address_space *mapping = inode->i_mapping;
1055         struct page *page;
1056         int rc;
1057
1058         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1059         if (*pglocked)
1060                 page = grab_cache_page_nowait(mapping, index); /* locked page */
1061         else
1062                 page = grab_cache_page(mapping, index); /* locked page */
1063
1064
1065         /* This page is currently locked, so get a temporary page instead. */
1066         /* XXX I believe this is a very dangerous thing to do - consider if
1067          *     we had multiple writers for the same file (definitely the case
1068          *     if we are using this codepath).  If writer A locks the page,
1069          *     writer B writes to a copy (as here), writer A drops the page
1070          *     lock, and writer C grabs the lock before B does, then B will
1071          *     later overwrite the data from C, even if C had LDLM locked
1072          *     and initiated the write after B did.
1073          */
1074         if (!page) {
1075                 unsigned long addr;
1076                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1077                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1078                 if (!addr) {
1079                         CERROR("no memory for a temp page\n");
1080                         LBUG();
1081                         GOTO(err, rc = -ENOMEM);
1082                 }
1083                 page = virt_to_page(addr);
1084                 kmap(page);
1085                 page->index = index;
1086                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1087         } else if (!IS_ERR(page)) {
1088                 (*pglocked)++;
1089                 kmap(page);
1090
1091                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1092                  * a no-op for most filesystems, because we write the whole
1093                  * page.  For partial-page I/O this will read in the page.
1094                  */
1095                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1096                 if (rc) {
1097                         CERROR("page index %lu, rc = %d\n", index, rc);
1098                         if (rc != -ENOSPC)
1099                                 LBUG();
1100                         GOTO(err_unlock, rc);
1101                 }
1102                 /* XXX not sure if we need this if we are overwriting page */
1103                 if (PageError(page)) {
1104                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1105                         LBUG();
1106                         GOTO(err_unlock, rc = -EIO);
1107                 }
1108         }
1109         return page;
1110
1111 err_unlock:
1112         unlock_page(page);
1113         lustre_put_page(page);
1114 err:
1115         return ERR_PTR(rc);
1116 }
1117
1118 /*
1119  * We need to balance prepare_write() calls with commit_write() calls.
1120  * If the page has been prepared, but we have no data for it, we don't
1121  * want to overwrite valid data on disk, but we still need to zero out
1122  * data for space which was newly allocated.  Like part of what happens
1123  * in __block_prepare_write() for newly allocated blocks.
1124  *
1125  * XXX currently __block_prepare_write() creates buffers for all the
1126  *     pages, and the filesystems mark these buffers as BH_New if they
1127  *     were newly allocated from disk. We use the BH_New flag similarly.
1128  */
1129 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1130                                int err)
1131 {
1132         if (err) {
1133                 unsigned block_start, block_end;
1134                 struct buffer_head *bh, *head = page->buffers;
1135                 unsigned blocksize = head->b_size;
1136                 void *addr = page_address(page);
1137
1138                 /* debugging: just seeing if this ever happens */
1139                 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1140                        page->index, page->mapping->host->i_ino, err);
1141
1142                 /* Currently one buffer per page, but in the future... */
1143                 for (bh = head, block_start = 0; bh != head || !block_start;
1144                      block_start = block_end, bh = bh->b_this_page) {
1145                         block_end = block_start + blocksize;
1146                         if (buffer_new(bh))
1147                                 memset(addr + block_start, 0, blocksize);
1148                 }
1149         }
1150
1151         return lustre_commit_write(page, from, to);
1152 }
1153
1154 static int filter_preprw(int cmd, struct lustre_handle *conn,
1155                          int objcount, struct obd_ioobj *obj,
1156                          int niocount, struct niobuf_remote *nb,
1157                          struct niobuf_local *res, void **desc_private)
1158 {
1159         struct obd_run_ctxt saved;
1160         struct obd_device *obd;
1161         struct obd_ioobj *o = obj;
1162         struct niobuf_remote *b = nb;
1163         struct niobuf_local *r = res;
1164         void *journal_save = NULL;
1165         int pglocked = 0;
1166         int rc = 0;
1167         int i;
1168         ENTRY;
1169
1170         obd = class_conn2obd(conn);
1171         if (!obd) {
1172                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1173                 RETURN(-EINVAL);
1174         }
1175         memset(res, 0, sizeof(*res) * niocount);
1176
1177         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1178
1179         if (cmd & OBD_BRW_WRITE) {
1180                 *desc_private = filter_journal_start(&journal_save,
1181                                                      &obd->u.filter,
1182                                                      objcount, obj, niocount,
1183                                                      nb);
1184                 if (IS_ERR(*desc_private))
1185                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1186         }
1187
1188         for (i = 0; i < objcount; i++, o++) {
1189                 struct dentry *dentry;
1190                 struct inode *inode;
1191                 int j;
1192
1193                 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1194                                            o->ioo_id, S_IFREG);
1195                 if (IS_ERR(dentry))
1196                         GOTO(out_clean, rc = PTR_ERR(dentry));
1197                 inode = dentry->d_inode;
1198                 if (!inode) {
1199                         CERROR("trying to BRW to non-existent file %Ld\n",
1200                                (unsigned long long)o->ioo_id);
1201                         f_dput(dentry);
1202                         GOTO(out_clean, rc = -ENOENT);
1203                 }
1204
1205                 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1206                         unsigned long index = b->offset >> PAGE_SHIFT;
1207                         struct page *page;
1208
1209                         if (j == 0)
1210                                 r->dentry = dentry;
1211                         else
1212                                 r->dentry = dget(dentry);
1213
1214                         if (cmd & OBD_BRW_WRITE)
1215                                 page = filter_get_page_write(inode, index, r,
1216                                                              &pglocked);
1217                         else
1218                                 page = lustre_get_page_read(inode, index);
1219
1220                         if (IS_ERR(page)) {
1221                                 f_dput(dentry);
1222                                 GOTO(out_clean, rc = PTR_ERR(page));
1223                         }
1224
1225                         r->addr = page_address(page);
1226                         r->offset = b->offset;
1227                         r->page = page;
1228                         r->len = b->len;
1229                 }
1230         }
1231
1232 out_stop:
1233         if (cmd & OBD_BRW_WRITE) {
1234                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1235                                               *desc_private);
1236                 if (!rc)
1237                         rc = err;
1238         }
1239 out_ctxt:
1240         pop_ctxt(&saved);
1241         RETURN(rc);
1242 out_clean:
1243         while (r-- > res) {
1244                 CERROR("error cleanup on brw\n");
1245                 f_dput(r->dentry);
1246                 if (cmd & OBD_BRW_WRITE)
1247                         filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1248                 else
1249                         lustre_put_page(r->page);
1250         }
1251         goto out_stop;
1252 }
1253
1254 static int filter_write_locked_page(struct niobuf_local *lnb)
1255 {
1256         struct page *lpage;
1257         int rc;
1258
1259         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1260         if (IS_ERR(lpage)) {
1261                 /* It is highly unlikely that we would ever get an error here.
1262                  * The page we want to get was previously locked, so it had to
1263                  * have already allocated the space, and we were just writing
1264                  * over the same data, so there would be no hole in the file.
1265                  *
1266                  * XXX: possibility of a race with truncate could exist, need
1267                  *      to check that.  There are no guarantees w.r.t.
1268                  *      write order even on a local filesystem, although the
1269                  *      normal response would be to return the number of bytes
1270                  *      successfully written and leave the rest to the app.
1271                  */
1272                 rc = PTR_ERR(lpage);
1273                 CERROR("error getting locked page index %ld: rc = %d\n",
1274                        lnb->page->index, rc);
1275                 GOTO(out, rc);
1276         }
1277
1278         /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1279          * lustre_commit_write() below, lnb->page was kmapped previously in
1280          * filter_get_page_write() and kunmapped in lustre_put_page() below.
1281          */
1282         memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1283         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1284         if (rc)
1285                 CERROR("error committing locked page %ld: rc = %d\n",
1286                        lnb->page->index, rc);
1287 out:
1288         lustre_put_page(lnb->page);
1289
1290         return rc;
1291 }
1292
1293 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1294                            int objcount, struct obd_ioobj *obj,
1295                            int niocount, struct niobuf_local *res,
1296                            void *private)
1297 {
1298         struct obd_run_ctxt saved;
1299         struct obd_ioobj *o;
1300         struct niobuf_local *r;
1301         struct obd_device *obd = class_conn2obd(conn);
1302         void *journal_save;
1303         int found_locked = 0;
1304         int rc = 0;
1305         int i;
1306         ENTRY;
1307
1308         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1309         journal_save = current->journal_info;
1310         LASSERT(!journal_save);
1311
1312         current->journal_info = private;
1313         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1314                 int j;
1315                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1316                         struct page *page = r->page;
1317
1318                         if (!page)
1319                                 LBUG();
1320
1321                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1322                                 found_locked++;
1323                                 continue;
1324                         }
1325
1326                         if (cmd & OBD_BRW_WRITE) {
1327                                 int err = filter_commit_write(page, 0,
1328                                                               r->len, 0);
1329
1330                                 if (!rc)
1331                                         rc = err;
1332                         } else
1333                                 lustre_put_page(page);
1334
1335                         f_dput(r->dentry);
1336                 }
1337         }
1338         current->journal_info = journal_save;
1339
1340         if (!found_locked)
1341                 goto out_ctxt;
1342
1343         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1344                 int j;
1345                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1346                         int err;
1347                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1348                                 continue;
1349
1350                         err = filter_write_locked_page(r);
1351                         if (!rc)
1352                                 rc = err;
1353                         f_dput(r->dentry);
1354                 }
1355         }
1356
1357 out_ctxt:
1358         pop_ctxt(&saved);
1359         RETURN(rc);
1360 }
1361
1362 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1363 {
1364         struct obd_device *obd = class_conn2obd(conn);
1365         struct statfs sfs;
1366         int rc;
1367
1368         ENTRY;
1369         rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1370         if (!rc)
1371                 statfs_pack(osfs, &sfs);
1372
1373         return rc;
1374 }
1375
1376 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1377                            void *key, obd_count *vallen, void **val)
1378 {
1379         struct obd_device *obd;
1380         ENTRY;
1381
1382         obd = class_conn2obd(conn);
1383         if (!obd) {
1384                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1385                 RETURN(-EINVAL);
1386         }
1387
1388         if ( keylen == strlen("blocksize") &&
1389              memcmp(key, "blocksize", keylen) == 0 ) {
1390                 *vallen = sizeof(long);
1391                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1392                 RETURN(0);
1393         }
1394
1395         if ( keylen == strlen("blocksize_bits") &&
1396              memcmp(key, "blocksize_bits", keylen) == 0 ){
1397                 *vallen = sizeof(long);
1398                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1399                 RETURN(0);
1400         }
1401
1402         if ( keylen == strlen("root_ino") &&
1403              memcmp(key, "root_ino", keylen) == 0 ){
1404                 *vallen = sizeof(obd_id);
1405                 *val = (void *)(obd_id)FILTER_ROOTINO;
1406                 RETURN(0);
1407         }
1408
1409         CDEBUG(D_IOCTL, "invalid key\n");
1410         RETURN(-EINVAL);
1411 }
1412
1413 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1414                   struct lustre_handle *src_conn, struct obdo *src,
1415                   obd_size count, obd_off offset)
1416 {
1417         struct page *page;
1418         struct lov_stripe_md srcmd, dstmd;
1419         unsigned long index = 0;
1420         int err = 0;
1421
1422         memset(&srcmd, 0, sizeof(srcmd));
1423         memset(&dstmd, 0, sizeof(dstmd));
1424         srcmd.lsm_object_id = src->o_id;
1425         dstmd.lsm_object_id = dst->o_id;
1426
1427         ENTRY;
1428         CDEBUG(D_INFO, "src: ino %Ld blocks %Ld, size %Ld, dst: ino %Ld\n",
1429                (unsigned long long)src->o_id, (unsigned long long)src->o_blocks,
1430                (unsigned long long)src->o_size, (unsigned long long)dst->o_id);
1431         page = alloc_page(GFP_USER);
1432         if (page == NULL)
1433                 RETURN(-ENOMEM);
1434
1435         while (TryLockPage(page))
1436                 ___wait_on_page(page);
1437
1438         /* XXX with brw vector I/O, we could batch up reads and writes here,
1439          *     all we need to do is allocate multiple pages to handle the I/Os
1440          *     and arrays to handle the request parameters.
1441          */
1442         while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1443                 struct brw_page pg;
1444                 struct io_cb_data *cbd = ll_init_cb();
1445
1446                 if (!cbd) {
1447                         err = -ENOMEM;
1448                         EXIT;
1449                         break;
1450                 }
1451
1452                 pg.pg = page;
1453                 pg.count = PAGE_SIZE;
1454                 pg.off = (page->index) << PAGE_SHIFT;
1455                 pg.flag = 0;
1456
1457                 page->index = index;
1458                 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1459                               ll_sync_io_cb, cbd);
1460
1461                 if ( err ) {
1462                         EXIT;
1463                         break;
1464                 }
1465
1466                 cbd = ll_init_cb();
1467                 if (!cbd) {
1468                         err = -ENOMEM;
1469                         EXIT;
1470                         break;
1471                 }
1472                 pg.flag = OBD_BRW_CREATE;
1473                 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1474
1475                 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1476                               ll_sync_io_cb, cbd);
1477
1478                 /* XXX should handle dst->o_size, dst->o_blocks here */
1479                 if ( err ) {
1480                         EXIT;
1481                         break;
1482                 }
1483
1484                 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1485
1486                 index++;
1487         }
1488         dst->o_size = src->o_size;
1489         dst->o_blocks = src->o_blocks;
1490         dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1491         unlock_page(page);
1492         __free_page(page);
1493
1494         RETURN(err);
1495 }
1496
1497
1498 static struct obd_ops filter_obd_ops = {
1499         o_get_info:    filter_get_info,
1500         o_setup:       filter_setup,
1501         o_cleanup:     filter_cleanup,
1502         o_connect:     filter_connect,
1503         o_disconnect:  filter_disconnect,
1504         o_statfs:      filter_statfs,
1505         o_getattr:     filter_getattr,
1506         o_create:      filter_create,
1507         o_setattr:     filter_setattr,
1508         o_destroy:     filter_destroy,
1509         o_open:        filter_open,
1510         o_close:       filter_close,
1511         o_brw:         filter_pgcache_brw,
1512         o_punch:       filter_truncate,
1513         o_preprw:      filter_preprw,
1514         o_commitrw:    filter_commitrw
1515 #if 0
1516         o_preallocate: filter_preallocate_inodes,
1517         o_migrate:     filter_migrate,
1518         o_copy:        filter_copy_data,
1519         o_iterate:     filter_iterate
1520 #endif
1521 };
1522
1523
1524 static int __init obdfilter_init(void)
1525 {
1526         printk(KERN_INFO "Filtering OBD driver  v0.001, info@clusterfs.com\n");
1527         return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1528 }
1529
1530 static void __exit obdfilter_exit(void)
1531 {
1532         class_unregister_type(OBD_FILTER_DEVICENAME);
1533 }
1534
1535 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1536 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1537 MODULE_LICENSE("GPL");
1538
1539 module_init(obdfilter_init);
1540 module_exit(obdfilter_exit);