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/pagemap.h>
  20 #include <linux/obd_class.h>
  21 #include <linux/lustre_dlm.h>
  22 #include <linux/obd_filter.h>
  23 #include <linux/ext3_jbd.h>
  24 #include <linux/extN_jbd.h>
  25 #include <linux/quotaops.h>
  26 #include <linux/init.h>
  27 #include <linux/stringify.h>
  28
  29 long filter_memory;
  30
  31 #define FILTER_ROOTINO 2
  32 #define FILTER_ROOTINO_STR __stringify(FILTER_ROOTINO)
  33
  34 #define S_SHIFT 12
  35 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
  36         [0]                     NULL,
  37         [S_IFREG >> S_SHIFT]    "R",
  38         [S_IFDIR >> S_SHIFT]    "D",
  39         [S_IFCHR >> S_SHIFT]    "C",
  40         [S_IFBLK >> S_SHIFT]    "B",
  41         [S_IFIFO >> S_SHIFT]    "F",
  42         [S_IFSOCK >> S_SHIFT]   "S",
  43         [S_IFLNK >> S_SHIFT]    "L"
  44 };
  45
  46 static inline const char *obd_mode_to_type(int mode)
  47 {
  48         return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
  49 }
  50
  51 /* write the pathname into the string */
  52 static int filter_id(char *buf, obd_id id, obd_mode mode)
  53 {
  54         return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
  55 }
  56
  57 static inline void f_dput(struct dentry *dentry)
  58 {
  59         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
  60                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
  61         LASSERT(atomic_read(&dentry->d_count) > 0);
  62         dput(dentry);
  63 }
  64
  65 /* setup the object store with correct subdirectories */
  66 static int filter_prep(struct obd_device *obddev)
  67 {
  68         struct obd_run_ctxt saved;
  69         struct filter_obd *filter = &obddev->u.filter;
  70         struct dentry *dentry;
  71         struct dentry *root;
  72         struct file *file;
  73         struct inode *inode;
  74         int rc = 0;
  75         __u64 lastino = 2;
  76         int mode = 0;
  77
  78         push_ctxt(&saved, &filter->fo_ctxt, NULL);
  79         dentry = simple_mkdir(current->fs->pwd, "O", 0700);
  80         CDEBUG(D_INODE, "got/created O: %p\n", dentry);
  81         if (IS_ERR(dentry)) {
  82                 rc = PTR_ERR(dentry);
  83                 CERROR("cannot open/create O: rc = %d\n", rc);
  84                 GOTO(out, rc);
  85         }
  86         filter->fo_dentry_O = dentry;
  87         dentry = simple_mkdir(current->fs->pwd, "P", 0700);
  88         CDEBUG(D_INODE, "got/created P: %p\n", dentry);
  89         if (IS_ERR(dentry)) {
  90                 rc = PTR_ERR(dentry);
  91                 CERROR("cannot open/create P: rc = %d\n", rc);
  92                 GOTO(out_O, rc);
  93         }
  94         f_dput(dentry);
  95         dentry = simple_mkdir(current->fs->pwd, "D", 0700);
  96         CDEBUG(D_INODE, "got/created D: %p\n", dentry);
  97         if (IS_ERR(dentry)) {
  98                 rc = PTR_ERR(dentry);
  99                 CERROR("cannot open/create D: rc = %d\n", rc);
 100                 GOTO(out_O, rc);
 101         }
 102
 103         root = simple_mknod(dentry, FILTER_ROOTINO_STR, S_IFREG | 0755);
 104         f_dput(dentry);
 105         if (IS_ERR(root)) {
 106                 rc = PTR_ERR(root);
 107                 CERROR("OBD filter: cannot open/create root %d: rc = %d\n",
 108                        FILTER_ROOTINO, rc);
 109                 GOTO(out_O, rc);
 110         }
 111         f_dput(root);
 112
 113         /*
 114          * Create directories and/or get dentries for each object type.
 115          * This saves us from having to do multiple lookups for each one.
 116          */
 117         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 118                 char *type = obd_type_by_mode[mode];
 119
 120                 if (!type) {
 121                         filter->fo_dentry_O_mode[mode] = NULL;
 122                         continue;
 123                 }
 124                 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
 125                 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
 126                 if (IS_ERR(dentry)) {
 127                         rc = PTR_ERR(dentry);
 128                         CERROR("cannot create O/%s: rc = %d\n", type, rc);
 129                         GOTO(out_O_mode, rc);
 130                 }
 131                 filter->fo_dentry_O_mode[mode] = dentry;
 132         }
 133
 134         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 135         if ( !file || IS_ERR(file) ) {
 136                 rc = PTR_ERR(file);
 137                 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
 138                        "D/status", rc);
 139                 GOTO(out_O_mode, rc);
 140         }
 141
 142         /* steal operations */
 143         inode = file->f_dentry->d_inode;
 144         filter->fo_fop = file->f_op;
 145         filter->fo_iop = inode->i_op;
 146         filter->fo_aops = inode->i_mapping->a_ops;
 147
 148         if (inode->i_size == 0) {
 149                 __u64 disk_lastino = cpu_to_le64(lastino);
 150                 ssize_t retval = file->f_op->write(file, (char *)&disk_lastino,
 151                                                    sizeof(disk_lastino),
 152                                                    &file->f_pos);
 153                 if (retval != sizeof(disk_lastino)) {
 154                         CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
 155                         filp_close(file, 0);
 156                         GOTO(out_O_mode, rc = -EIO);
 157                 }
 158         } else {
 159                 __u64 disk_lastino;
 160                 ssize_t retval = file->f_op->read(file, (char *)&disk_lastino,
 161                                                   sizeof(disk_lastino),
 162                                                   &file->f_pos);
 163                 if (retval != sizeof(disk_lastino)) {
 164                         CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
 165                         filp_close(file, 0);
 166                         GOTO(out_O_mode, rc = -EIO);
 167                 }
 168                 lastino = le64_to_cpu(disk_lastino);
 169         }
 170         filter->fo_lastino = lastino;
 171         filp_close(file, 0);
 172
 173         rc = 0;
 174  out:
 175         pop_ctxt(&saved);
 176
 177         return(rc);
 178
 179 out_O_mode:
 180         while (mode-- > 0) {
 181                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 182                 if (dentry) {
 183                         f_dput(dentry);
 184                         filter->fo_dentry_O_mode[mode] = NULL;
 185                 }
 186         }
 187 out_O:
 188         f_dput(filter->fo_dentry_O);
 189         filter->fo_dentry_O = NULL;
 190         goto out;
 191 }
 192
 193 /* cleanup the filter: write last used object id to status file */
 194 static void filter_post(struct obd_device *obddev)
 195 {
 196         struct obd_run_ctxt saved;
 197         struct filter_obd *filter = &obddev->u.filter;
 198         __u64 disk_lastino;
 199         long rc;
 200         struct file *file;
 201         int mode;
 202
 203         push_ctxt(&saved, &filter->fo_ctxt, NULL);
 204         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 205         if (IS_ERR(file)) {
 206                 CERROR("OBD filter: cannot create status file\n");
 207                 goto out;
 208         }
 209
 210         file->f_pos = 0;
 211         disk_lastino = cpu_to_le64(filter->fo_lastino);
 212         rc = file->f_op->write(file, (char *)&disk_lastino,
 213                        sizeof(disk_lastino), &file->f_pos);
 214         if (rc != sizeof(disk_lastino))
 215                 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
 216
 217         rc = filp_close(file, NULL);
 218         if (rc)
 219                 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
 220
 221         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 222                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 223                 if (dentry) {
 224                         f_dput(dentry);
 225                         filter->fo_dentry_O_mode[mode] = NULL;
 226                 }
 227         }
 228         f_dput(filter->fo_dentry_O);
 229 out:
 230         pop_ctxt(&saved);
 231 }
 232
 233
 234 static __u64 filter_next_id(struct obd_device *obddev)
 235 {
 236         obd_id id;
 237
 238         spin_lock(&obddev->u.filter.fo_lock);
 239         id = ++obddev->u.filter.fo_lastino;
 240         spin_unlock(&obddev->u.filter.fo_lock);
 241
 242         /* FIXME: write the lastino to disk here */
 243         return id;
 244 }
 245
 246 /* how to get files, dentries, inodes from object id's */
 247 /* parent i_sem is already held if needed for exclusivity */
 248 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
 249                                         struct dentry *dparent,
 250                                         __u64 id, __u32 type)
 251 {
 252         struct super_block *sb = obddev->u.filter.fo_sb;
 253         struct dentry *dchild;
 254         char name[32];
 255         int len;
 256         ENTRY;
 257
 258         if (!sb || !sb->s_dev) {
 259                 CERROR("fatal: device not initialized.\n");
 260                 RETURN(ERR_PTR(-ENXIO));
 261         }
 262
 263         if (id == 0) {
 264                 CERROR("fatal: invalid object #0\n");
 265                 LBUG();
 266                 RETURN(ERR_PTR(-ESTALE));
 267         }
 268
 269         if (!(type & S_IFMT)) {
 270                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 271                        __FUNCTION__, id, type);
 272                 RETURN(ERR_PTR(-EINVAL));
 273         }
 274
 275         len = sprintf(name, LPU64, id);
 276         CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
 277                name);
 278         dchild = lookup_one_len(name, dparent, len);
 279         if (IS_ERR(dchild)) {
 280                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 281                 RETURN(dchild);
 282         }
 283
 284         CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
 285                obd_mode_to_type(type), name, dchild,
 286                atomic_read(&dchild->d_count));
 287
 288         LASSERT(atomic_read(&dchild->d_count) > 0);
 289
 290         RETURN(dchild);
 291 }
 292
 293 static struct file *filter_obj_open(struct obd_device *obddev,
 294                                     __u64 id, __u32 type)
 295 {
 296         struct super_block *sb = obddev->u.filter.fo_sb;
 297         struct obd_run_ctxt saved;
 298         char name[24];
 299         struct file *file;
 300         ENTRY;
 301
 302         if (!sb || !sb->s_dev) {
 303                 CERROR("fatal: device not initialized.\n");
 304                 RETURN(ERR_PTR(-ENXIO));
 305         }
 306
 307         if (!id) {
 308                 CERROR("fatal: invalid obdo "LPU64"\n", id);
 309                 RETURN(ERR_PTR(-ESTALE));
 310         }
 311
 312         if (!(type & S_IFMT)) {
 313                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 314                        __FUNCTION__, id, type);
 315                 RETURN(ERR_PTR(-EINVAL));
 316         }
 317
 318         filter_id(name, id, type);
 319         push_ctxt(&saved, &obddev->u.filter.fo_ctxt, NULL);
 320         file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
 321         pop_ctxt(&saved);
 322
 323         CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
 324
 325         if (IS_ERR(file))
 326                 file = NULL;
 327         RETURN(file);
 328 }
 329
 330 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
 331 {
 332         struct filter_obd *filter = &obddev->u.filter;
 333
 334         return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
 335 }
 336
 337
 338 /* obd methods */
 339 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
 340                           obd_uuid_t cluuid, struct recovd_obd *recovd,
 341                           ptlrpc_recovery_cb_t recover)
 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         int rc = 0;
 487         ENTRY;
 488
 489         if (!class_conn2export(conn)) {
 490                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 491                 RETURN(-EINVAL);
 492         }
 493
 494         obddev = class_conn2obd(conn);
 495         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 496                                    oa->o_id, oa->o_mode);
 497         if (IS_ERR(dentry))
 498                 RETURN(PTR_ERR(dentry));
 499
 500         if (!dentry->d_inode) {
 501                 CERROR("getattr on non-existent object: "LPU64"\n", oa->o_id);
 502                 GOTO(out_getattr, rc = -ENOENT);
 503         }
 504
 505         filter_from_inode(oa, dentry->d_inode, oa->o_valid);
 506
 507 out_getattr:
 508         f_dput(dentry);
 509         RETURN(rc);
 510 }
 511
 512 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
 513                           struct lov_stripe_md *md)
 514 {
 515         struct obd_run_ctxt saved;
 516         struct obd_device *obd = class_conn2obd(conn);
 517         struct dentry *dentry;
 518         struct iattr iattr;
 519         struct inode *inode;
 520         int rc;
 521         ENTRY;
 522
 523         iattr_from_obdo(&iattr, oa, oa->o_valid);
 524         iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
 525         dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
 526                                    oa->o_id, iattr.ia_mode);
 527         if (IS_ERR(dentry))
 528                 RETURN(PTR_ERR(dentry));
 529
 530         inode = dentry->d_inode;
 531         if (!inode) {
 532                 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
 533                 GOTO(out_setattr, rc = -ENOENT);
 534         }
 535
 536         lock_kernel();
 537         if (iattr.ia_valid & ATTR_SIZE)
 538                 down(&inode->i_sem);
 539         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 540         if (inode->i_op->setattr)
 541                 rc = inode->i_op->setattr(dentry, &iattr);
 542         else
 543                 rc = inode_setattr(inode, &iattr);
 544         pop_ctxt(&saved);
 545         if (iattr.ia_valid & ATTR_SIZE) {
 546                 up(&inode->i_sem);
 547                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 548                 obdo_from_inode(oa, inode, oa->o_valid);
 549         }
 550         unlock_kernel();
 551
 552 out_setattr:
 553         f_dput(dentry);
 554         RETURN(rc);
 555 }
 556
 557 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
 558                        struct lov_stripe_md *ea)
 559 {
 560         struct obd_export *export;
 561         struct obd_device *obd;
 562         struct dentry *dentry;
 563         /* ENTRY; */
 564
 565         export = class_conn2export(conn);
 566         if (!export) {
 567                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 568                 RETURN(-EINVAL);
 569         }
 570
 571         obd = export->exp_obd;
 572         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 573                                    oa->o_id, oa->o_mode);
 574         if (IS_ERR(dentry))
 575                 RETURN(PTR_ERR(dentry));
 576
 577         if (!dentry->d_inode) {
 578                 CERROR("opening non-existent objid "LPX64"\n", oa->o_id);
 579                 RETURN(-ENOENT);
 580         }
 581
 582         filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 583                           OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 584
 585         return 0;
 586 } /* filter_open */
 587
 588 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
 589                         struct lov_stripe_md *ea)
 590 {
 591         struct obd_device *obd;
 592         struct dentry *dentry;
 593         /* ENTRY; */
 594
 595         obd = class_conn2obd(conn);
 596         if (!obd) {
 597                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 598                 RETURN(-EINVAL);
 599         }
 600
 601         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 602                                    oa->o_id, oa->o_mode);
 603         if (IS_ERR(dentry))
 604                 RETURN(PTR_ERR(dentry));
 605         LASSERT(atomic_read(&dentry->d_count) > 1);
 606
 607         f_dput(dentry);  /* for the close */
 608         f_dput(dentry);  /* for this call */
 609         return 0;
 610 } /* filter_close */
 611
 612 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
 613                          struct lov_stripe_md **ea)
 614 {
 615         char name[64];
 616         struct obd_run_ctxt saved;
 617         struct dentry *new;
 618         int mode;
 619         struct obd_device *obd = class_conn2obd(conn);
 620         struct filter_obd *filter = &obd->u.filter;
 621         struct iattr;
 622         ENTRY;
 623
 624         if (!obd) {
 625                 CERROR("invalid client "LPX64"\n", conn->addr);
 626                 return -EINVAL;
 627         }
 628
 629         if (!(oa->o_mode && S_IFMT)) {
 630                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 631                        __FUNCTION__, oa->o_id, oa->o_mode);
 632                 return -ENOENT;
 633         }
 634
 635         oa->o_id = filter_next_id(obd);
 636
 637         //filter_id(name, oa->o_id, oa->o_mode);
 638         sprintf(name, LPU64, oa->o_id);
 639         mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
 640         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 641         new = simple_mknod(filter->fo_dentry_O_mode[S_IFREG >> S_SHIFT], name,
 642                            mode);
 643         pop_ctxt(&saved);
 644         if (IS_ERR(new)) {
 645                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
 646                 return -ENOENT;
 647         }
 648
 649         /* Set flags for fields we have set in the inode struct */
 650         oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 651                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
 652         filter_from_inode(oa, new->d_inode, oa->o_valid);
 653         dput(new);
 654
 655         return 0;
 656 }
 657
 658 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
 659                           struct lov_stripe_md *ea)
 660 {
 661         struct obd_device *obd;
 662         struct filter_obd *filter;
 663         struct obd_run_ctxt saved;
 664         struct inode *inode;
 665         struct dentry *dir_dentry, *object_dentry;
 666         int rc;
 667         ENTRY;
 668
 669         obd = class_conn2obd(conn);
 670         if (!obd) {
 671                 CERROR("invalid client "LPX64"\n", conn->addr);
 672                 RETURN(-EINVAL);
 673         }
 674
 675         CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
 676
 677         dir_dentry = filter_parent(obd, oa->o_mode);
 678         down(&dir_dentry->d_inode->i_sem);
 679
 680         object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
 681                                           oa->o_mode);
 682         if (IS_ERR(object_dentry))
 683                 GOTO(out, rc = -ENOENT);
 684
 685         inode = object_dentry->d_inode;
 686         if (!inode) {
 687                 CERROR("trying to destroy negative inode "LPX64"!\n", oa->o_id);
 688                 GOTO(out, rc = -ENOENT);
 689         }
 690
 691         if (inode->i_nlink != 1) {
 692                 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
 693                 LBUG();
 694                 inode->i_nlink = 1;
 695         }
 696         inode->i_mode = S_IFREG;
 697
 698         if (atomic_read(&inode->i_count) > 1) {
 699 #warning FIXME: need to handle open-unlinked case and move to hold dir
 700                 CERROR("inode has count %d\n", atomic_read(&inode->i_count));
 701         }
 702
 703         filter = &obd->u.filter;
 704         push_ctxt(&saved, &filter->fo_ctxt, NULL);
 705
 706         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 707         pop_ctxt(&saved);
 708         f_dput(object_dentry);
 709
 710         EXIT;
 711 out:
 712         up(&dir_dentry->d_inode->i_sem);
 713         return rc;
 714 }
 715
 716 /* NB count and offset are used for punch, but not truncate */
 717 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
 718                            struct lov_stripe_md *md,
 719                            obd_off start, obd_off end)
 720 {
 721         int error;
 722         ENTRY;
 723
 724         if (end != OBD_OBJECT_EOF)
 725                 CERROR("PUNCH not supported, only truncate works\n");
 726
 727         CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
 728                "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
 729         oa->o_size = start;
 730         error = filter_setattr(conn, oa, NULL);
 731         RETURN(error);
 732 }
 733
 734 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
 735                               struct lov_stripe_md *lsm, obd_count oa_bufs,
 736                               struct brw_page *pga, brw_callback_t callback,
 737                               struct io_cb_data *data)
 738 {
 739         struct obd_run_ctxt      saved;
 740         struct super_block      *sb;
 741         int                      pnum;          /* index to pages (bufs) */
 742         unsigned long            retval;
 743         int                      error;
 744         struct file             *file;
 745         struct obd_device       *obd = class_conn2obd(conn);
 746         int pg;
 747         ENTRY;
 748
 749         if (!obd) {
 750                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
 751                 RETURN(-EINVAL);
 752         }
 753
 754         sb = obd->u.filter.fo_sb;
 755         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 756         pnum = 0; /* pnum indexes buf 0..num_pages */
 757
 758         file = filter_obj_open(obd, lsm->lsm_object_id, S_IFREG);
 759         if (IS_ERR(file))
 760                 GOTO(out, retval = PTR_ERR(file));
 761
 762         /* count doubles as retval */
 763         for (pg = 0; pg < oa_bufs; pg++) {
 764                 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
 765                        ") off count ("LPU64",%d)\n",
 766                        cmd, pnum, file->f_dentry->d_inode->i_ino,
 767                        pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
 768                        (int)pga[pnum].count);
 769                 if (cmd & OBD_BRW_WRITE) {
 770                         loff_t off;
 771                         char *buffer;
 772                         off = pga[pnum].off;
 773                         buffer = kmap(pga[pnum].pg);
 774                         retval = file->f_op->write(file, buffer,
 775                                                    pga[pnum].count,
 776                                                    &off);
 777                         kunmap(pga[pnum].pg);
 778                         CDEBUG(D_INODE, "retval %ld\n", retval);
 779                 } else {
 780                         loff_t off = pga[pnum].off;
 781                         char *buffer = kmap(pga[pnum].pg);
 782
 783                         if (off >= file->f_dentry->d_inode->i_size) {
 784                                 memset(buffer, 0, pga[pnum].count);
 785                                 retval = pga[pnum].count;
 786                         } else {
 787                                 retval = file->f_op->read(file, buffer,
 788                                                           pga[pnum].count, &off);
 789                         }
 790                         kunmap(pga[pnum].pg);
 791
 792                         if (retval != pga[pnum].count) {
 793                                 filp_close(file, 0);
 794                                 GOTO(out, retval = -EIO);
 795                         }
 796                         CDEBUG(D_INODE, "retval %ld\n", retval);
 797                 }
 798                 pnum++;
 799         }
 800         /* sizes and blocks are set by generic_file_write */
 801         /* ctimes/mtimes will follow with a setattr call */
 802         filp_close(file, 0);
 803
 804         /* XXX: do something with callback if it is set? */
 805
 806         EXIT;
 807 out:
 808         pop_ctxt(&saved);
 809         error = (retval >= 0) ? 0 : retval;
 810         return error;
 811 }
 812
 813 /*
 814  * Calculate the number of buffer credits needed to write multiple pages in
 815  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
 816  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
 817  *
 818  * See comment above ext3_writepage_trans_blocks for details.  We assume
 819  * no data journaling is being done, but it does allow for all of the pages
 820  * being non-contiguous.  If we are guaranteed contiguous pages we could
 821  * reduce the number of (d)indirect blocks a lot.
 822  *
 823  * With N blocks per page and P pages, for each inode we have at most:
 824  * N*P indirect
 825  * min(N*P, blocksize/4 + 1) dindirect blocks
 826  * 1 tindirect
 827  *
 828  * For the entire filesystem, we have at most:
 829  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 830  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 831  * 1 inode block
 832  * 1 superblock
 833  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 834  */
 835 static int ext3_credits_needed(struct super_block *sb, int objcount,
 836                                struct obd_ioobj *obj)
 837 {
 838         struct obd_ioobj *o = obj;
 839         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 840         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 841         int nbitmaps = 0;
 842         int ngdblocks = 0;
 843         int needed = objcount + 1;
 844         int i;
 845
 846         for (i = 0; i < objcount; i++, o++) {
 847                 int nblocks = o->ioo_bufcnt * blockpp;
 848                 int ndindirect = min(nblocks, addrpp + 1);
 849                 int nindir = nblocks + ndindirect + 1;
 850
 851                 nbitmaps += nindir + nblocks;
 852                 ngdblocks += nindir + nblocks;
 853
 854                 needed += nindir;
 855         }
 856
 857         /* Assumes ext3 and extN have same sb_info layout at the start. */
 858         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 859                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 860         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 861                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 862
 863         needed += nbitmaps + ngdblocks;
 864
 865 #ifdef CONFIG_QUOTA
 866         /* We assume that there will be 1 bit set in s_dquot.flags for each
 867          * quota file that is active.  This is at least true for now.
 868          */
 869         needed += hweight32(sb_any_quota_enabled(sb)) *
 870                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 871 #endif
 872
 873         return needed;
 874 }
 875
 876 /* We have to start a huge journal transaction here to hold all of the
 877  * metadata for the pages being written here.  This is necessitated by
 878  * the fact that we do lots of prepare_write operations before we do
 879  * any of the matching commit_write operations, so even if we split
 880  * up to use "smaller" transactions none of them could complete until
 881  * all of them were opened.  By having a single journal transaction,
 882  * we eliminate duplicate reservations for common blocks like the
 883  * superblock and group descriptors or bitmaps.
 884  *
 885  * We will start the transaction here, but each prepare_write will
 886  * add a refcount to the transaction, and each commit_write will
 887  * remove a refcount.  The transaction will be closed when all of
 888  * the pages have been written.
 889  */
 890 static void *ext3_filter_journal_start(struct filter_obd *filter,
 891                                        int objcount, struct obd_ioobj *obj,
 892                                        int niocount, struct niobuf_remote *nb)
 893 {
 894         journal_t *journal = NULL;
 895         handle_t *handle = NULL;
 896         int needed;
 897
 898         /* It appears that some kernels have different values for
 899          * EXT*_MAX_GROUP_LOADED (either 8 or 32), so we cannot
 900          * assume anything after s_inode_bitmap_number is the same.
 901          */
 902         if (!strcmp(filter->fo_fstype, "ext3"))
 903                 journal = EXT3_SB(filter->fo_sb)->s_journal;
 904         else if (!strcmp(filter->fo_fstype, "extN"))
 905                 journal = EXTN_SB(filter->fo_sb)->s_journal;
 906         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
 907
 908         /* The number of blocks we could _possibly_ dirty can very large.
 909          * We reduce our request if it is absurd (and we couldn't get that
 910          * many credits for a single handle anyways).
 911          *
 912          * At some point we have to limit the size of I/Os sent at one time,
 913          * increase the size of the journal, or we have to calculate the
 914          * actual journal requirements more carefully by checking all of
 915          * the blocks instead of being maximally pessimistic.  It remains to
 916          * be seen if this is a real problem or not.
 917          */
 918         if (needed > journal->j_max_transaction_buffers) {
 919                 CERROR("want too many journal credits (%d) using %d instead\n",
 920                        needed, journal->j_max_transaction_buffers);
 921                 needed = journal->j_max_transaction_buffers;
 922         }
 923
 924         lock_kernel();
 925         handle = journal_start(journal, needed);
 926         unlock_kernel();
 927         if (IS_ERR(handle))
 928                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 929                        PTR_ERR(handle));
 930
 931         return(handle);
 932 }
 933
 934 static void *filter_journal_start(void **journal_save,
 935                                   struct filter_obd *filter,
 936                                   int objcount, struct obd_ioobj *obj,
 937                                   int niocount, struct niobuf_remote *nb)
 938 {
 939         void *handle = NULL;
 940
 941         /* This may not be necessary - we probably never have a
 942          * transaction started when we enter here, so we can
 943          * remove the saving of the journal state entirely.
 944          * For now leave it in just to see if it ever happens.
 945          */
 946         *journal_save = current->journal_info;
 947         if (*journal_save) {
 948                 CERROR("Already have handle %p???\n", *journal_save);
 949                 LBUG();
 950                 current->journal_info = NULL;
 951         }
 952
 953         if (!strcmp(filter->fo_fstype, "ext3") ||
 954             !strcmp(filter->fo_fstype, "extN"))
 955                 handle = ext3_filter_journal_start(filter, objcount, obj,
 956                                                    niocount, nb);
 957         return handle;
 958 }
 959
 960 static int ext3_filter_journal_stop(void *handle)
 961 {
 962         int rc;
 963
 964         /* We got a refcount on the handle for each call to prepare_write,
 965          * so we can drop the "parent" handle here to avoid the need for
 966          * osc to call back into filterobd to close the handle.  The
 967          * remaining references will be dropped in commit_write.
 968          */
 969         lock_kernel();
 970         rc = journal_stop((handle_t *)handle);
 971         unlock_kernel();
 972
 973         return rc;
 974 }
 975
 976 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
 977                                void *handle)
 978 {
 979         int rc = 0;
 980
 981         if (!strcmp(filter->fo_fstype, "ext3") ||
 982             !strcmp(filter->fo_fstype, "extN"))
 983                 rc = ext3_filter_journal_stop(handle);
 984
 985         if (rc)
 986                 CERROR("error on journal stop: rc = %d\n", rc);
 987
 988         current->journal_info = journal_save;
 989
 990         return rc;
 991 }
 992
 993 static inline void lustre_put_page(struct page *page)
 994 {
 995         kunmap(page);
 996         page_cache_release(page);
 997 }
 998
 999 static struct page *
1000 lustre_get_page_read(struct inode *inode, unsigned long index)
1001 {
1002         struct address_space *mapping = inode->i_mapping;
1003         struct page *page;
1004         int rc;
1005
1006         page = read_cache_page(mapping, index,
1007                                (filler_t*)mapping->a_ops->readpage, NULL);
1008         if (!IS_ERR(page)) {
1009                 wait_on_page(page);
1010                 kmap(page);
1011                 if (!Page_Uptodate(page)) {
1012                         CERROR("page index %lu not uptodate\n", index);
1013                         GOTO(err_page, rc = -EIO);
1014                 }
1015                 if (PageError(page)) {
1016                         CERROR("page index %lu has error\n", index);
1017                         GOTO(err_page, rc = -EIO);
1018                 }
1019         }
1020         return page;
1021
1022 err_page:
1023         lustre_put_page(page);
1024         return ERR_PTR(rc);
1025 }
1026
1027 static struct page *
1028 lustre_get_page_write(struct inode *inode, unsigned long index)
1029 {
1030         struct address_space *mapping = inode->i_mapping;
1031         struct page *page;
1032         int rc;
1033
1034         page = grab_cache_page(mapping, index); /* locked page */
1035
1036         if (!IS_ERR(page)) {
1037                 kmap(page);
1038                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1039                  * a no-op for most filesystems, because we write the whole
1040                  * page.  For partial-page I/O this will read in the page.
1041                  */
1042                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1043                 if (rc) {
1044                         CERROR("page index %lu, rc = %d\n", index, rc);
1045                         if (rc != -ENOSPC)
1046                                 LBUG();
1047                         GOTO(err_unlock, rc);
1048                 }
1049                 /* XXX not sure if we need this if we are overwriting page */
1050                 if (PageError(page)) {
1051                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1052                         LBUG();
1053                         GOTO(err_unlock, rc = -EIO);
1054                 }
1055         }
1056         return page;
1057
1058 err_unlock:
1059         unlock_page(page);
1060         lustre_put_page(page);
1061         return ERR_PTR(rc);
1062 }
1063
1064 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1065 {
1066         struct inode *inode = page->mapping->host;
1067         int err;
1068
1069         err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1070         if (!err && IS_SYNC(inode))
1071                 err = waitfor_one_page(page);
1072
1073         //SetPageUptodate(page); // the client commit_write will do this
1074
1075         SetPageReferenced(page);
1076         unlock_page(page);
1077         lustre_put_page(page);
1078         return err;
1079 }
1080
1081 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1082                                    struct niobuf_local *lnb, int *pglocked)
1083 {
1084         struct address_space *mapping = inode->i_mapping;
1085         struct page *page;
1086         int rc;
1087
1088         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1089         if (*pglocked)
1090                 page = grab_cache_page_nowait(mapping, index); /* locked page */
1091         else
1092                 page = grab_cache_page(mapping, index); /* locked page */
1093
1094
1095         /* This page is currently locked, so get a temporary page instead. */
1096         /* XXX I believe this is a very dangerous thing to do - consider if
1097          *     we had multiple writers for the same file (definitely the case
1098          *     if we are using this codepath).  If writer A locks the page,
1099          *     writer B writes to a copy (as here), writer A drops the page
1100          *     lock, and writer C grabs the lock before B does, then B will
1101          *     later overwrite the data from C, even if C had LDLM locked
1102          *     and initiated the write after B did.
1103          */
1104         if (!page) {
1105                 unsigned long addr;
1106                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1107                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1108                 if (!addr) {
1109                         CERROR("no memory for a temp page\n");
1110                         LBUG();
1111                         GOTO(err, rc = -ENOMEM);
1112                 }
1113                 page = virt_to_page(addr);
1114                 kmap(page);
1115                 page->index = index;
1116                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1117         } else if (!IS_ERR(page)) {
1118                 (*pglocked)++;
1119                 kmap(page);
1120
1121                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1122                  * a no-op for most filesystems, because we write the whole
1123                  * page.  For partial-page I/O this will read in the page.
1124                  */
1125                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1126                 if (rc) {
1127                         CERROR("page index %lu, rc = %d\n", index, rc);
1128                         if (rc != -ENOSPC)
1129                                 LBUG();
1130                         GOTO(err_unlock, rc);
1131                 }
1132                 /* XXX not sure if we need this if we are overwriting page */
1133                 if (PageError(page)) {
1134                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1135                         LBUG();
1136                         GOTO(err_unlock, rc = -EIO);
1137                 }
1138         }
1139         return page;
1140
1141 err_unlock:
1142         unlock_page(page);
1143         lustre_put_page(page);
1144 err:
1145         return ERR_PTR(rc);
1146 }
1147
1148 /*
1149  * We need to balance prepare_write() calls with commit_write() calls.
1150  * If the page has been prepared, but we have no data for it, we don't
1151  * want to overwrite valid data on disk, but we still need to zero out
1152  * data for space which was newly allocated.  Like part of what happens
1153  * in __block_prepare_write() for newly allocated blocks.
1154  *
1155  * XXX currently __block_prepare_write() creates buffers for all the
1156  *     pages, and the filesystems mark these buffers as BH_New if they
1157  *     were newly allocated from disk. We use the BH_New flag similarly.
1158  */
1159 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1160                                int err)
1161 {
1162         if (err) {
1163                 unsigned block_start, block_end;
1164                 struct buffer_head *bh, *head = page->buffers;
1165                 unsigned blocksize = head->b_size;
1166                 void *addr = page_address(page);
1167
1168                 /* debugging: just seeing if this ever happens */
1169                 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1170                        page->index, page->mapping->host->i_ino, err);
1171
1172                 /* Currently one buffer per page, but in the future... */
1173                 for (bh = head, block_start = 0; bh != head || !block_start;
1174                      block_start = block_end, bh = bh->b_this_page) {
1175                         block_end = block_start + blocksize;
1176                         if (buffer_new(bh))
1177                                 memset(addr + block_start, 0, blocksize);
1178                 }
1179         }
1180
1181         return lustre_commit_write(page, from, to);
1182 }
1183
1184 static int filter_preprw(int cmd, struct lustre_handle *conn,
1185                          int objcount, struct obd_ioobj *obj,
1186                          int niocount, struct niobuf_remote *nb,
1187                          struct niobuf_local *res, void **desc_private)
1188 {
1189         struct obd_run_ctxt saved;
1190         struct obd_device *obd;
1191         struct obd_ioobj *o = obj;
1192         struct niobuf_remote *b = nb;
1193         struct niobuf_local *r = res;
1194         void *journal_save = NULL;
1195         int pglocked = 0;
1196         int rc = 0;
1197         int i;
1198         ENTRY;
1199
1200         obd = class_conn2obd(conn);
1201         if (!obd) {
1202                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1203                 RETURN(-EINVAL);
1204         }
1205         memset(res, 0, sizeof(*res) * niocount);
1206
1207         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1208
1209         if (cmd & OBD_BRW_WRITE) {
1210                 *desc_private = filter_journal_start(&journal_save,
1211                                                      &obd->u.filter,
1212                                                      objcount, obj, niocount,
1213                                                      nb);
1214                 if (IS_ERR(*desc_private))
1215                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1216         }
1217
1218         for (i = 0; i < objcount; i++, o++) {
1219                 struct dentry *dentry;
1220                 struct inode *inode;
1221                 int j;
1222
1223                 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1224                                            o->ioo_id, S_IFREG);
1225                 if (IS_ERR(dentry))
1226                         GOTO(out_clean, rc = PTR_ERR(dentry));
1227                 inode = dentry->d_inode;
1228                 if (!inode) {
1229                         CERROR("trying to BRW to non-existent file "LPU64"\n",
1230                                o->ioo_id);
1231                         f_dput(dentry);
1232                         GOTO(out_clean, rc = -ENOENT);
1233                 }
1234
1235                 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1236                         unsigned long index = b->offset >> PAGE_SHIFT;
1237                         struct page *page;
1238
1239                         if (j == 0)
1240                                 r->dentry = dentry;
1241                         else
1242                                 r->dentry = dget(dentry);
1243
1244                         if (cmd & OBD_BRW_WRITE)
1245                                 page = filter_get_page_write(inode, index, r,
1246                                                              &pglocked);
1247                         else
1248                                 page = lustre_get_page_read(inode, index);
1249
1250                         if (IS_ERR(page)) {
1251                                 f_dput(dentry);
1252                                 GOTO(out_clean, rc = PTR_ERR(page));
1253                         }
1254
1255                         r->addr = page_address(page);
1256                         r->offset = b->offset;
1257                         r->page = page;
1258                         r->len = b->len;
1259                 }
1260         }
1261
1262 out_stop:
1263         if (cmd & OBD_BRW_WRITE) {
1264                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1265                                               *desc_private);
1266                 if (!rc)
1267                         rc = err;
1268         }
1269 out_ctxt:
1270         pop_ctxt(&saved);
1271         RETURN(rc);
1272 out_clean:
1273         while (r-- > res) {
1274                 CERROR("error cleanup on brw\n");
1275                 f_dput(r->dentry);
1276                 if (cmd & OBD_BRW_WRITE)
1277                         filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1278                 else
1279                         lustre_put_page(r->page);
1280         }
1281         goto out_stop;
1282 }
1283
1284 static int filter_write_locked_page(struct niobuf_local *lnb)
1285 {
1286         struct page *lpage;
1287         int rc;
1288
1289         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1290         if (IS_ERR(lpage)) {
1291                 /* It is highly unlikely that we would ever get an error here.
1292                  * The page we want to get was previously locked, so it had to
1293                  * have already allocated the space, and we were just writing
1294                  * over the same data, so there would be no hole in the file.
1295                  *
1296                  * XXX: possibility of a race with truncate could exist, need
1297                  *      to check that.  There are no guarantees w.r.t.
1298                  *      write order even on a local filesystem, although the
1299                  *      normal response would be to return the number of bytes
1300                  *      successfully written and leave the rest to the app.
1301                  */
1302                 rc = PTR_ERR(lpage);
1303                 CERROR("error getting locked page index %ld: rc = %d\n",
1304                        lnb->page->index, rc);
1305                 GOTO(out, rc);
1306         }
1307
1308         /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1309          * lustre_commit_write() below, lnb->page was kmapped previously in
1310          * filter_get_page_write() and kunmapped in lustre_put_page() below.
1311          */
1312         memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1313         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1314         if (rc)
1315                 CERROR("error committing locked page %ld: rc = %d\n",
1316                        lnb->page->index, rc);
1317 out:
1318         lustre_put_page(lnb->page);
1319
1320         return rc;
1321 }
1322
1323 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1324                            int objcount, struct obd_ioobj *obj,
1325                            int niocount, struct niobuf_local *res,
1326                            void *private)
1327 {
1328         struct obd_run_ctxt saved;
1329         struct obd_ioobj *o;
1330         struct niobuf_local *r;
1331         struct obd_device *obd = class_conn2obd(conn);
1332         void *journal_save;
1333         int found_locked = 0;
1334         int rc = 0;
1335         int i;
1336         ENTRY;
1337
1338         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1339         journal_save = current->journal_info;
1340         LASSERT(!journal_save);
1341
1342         current->journal_info = private;
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                         struct page *page = r->page;
1347
1348                         if (!page)
1349                                 LBUG();
1350
1351                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1352                                 found_locked++;
1353                                 continue;
1354                         }
1355
1356                         if (cmd & OBD_BRW_WRITE) {
1357                                 int err = filter_commit_write(page, 0,
1358                                                               r->len, 0);
1359
1360                                 if (!rc)
1361                                         rc = err;
1362                         } else
1363                                 lustre_put_page(page);
1364
1365                         f_dput(r->dentry);
1366                 }
1367         }
1368         current->journal_info = journal_save;
1369
1370         if (!found_locked)
1371                 goto out_ctxt;
1372
1373         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1374                 int j;
1375                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1376                         int err;
1377                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1378                                 continue;
1379
1380                         err = filter_write_locked_page(r);
1381                         if (!rc)
1382                                 rc = err;
1383                         f_dput(r->dentry);
1384                 }
1385         }
1386
1387 out_ctxt:
1388         pop_ctxt(&saved);
1389         RETURN(rc);
1390 }
1391
1392 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1393 {
1394         struct obd_device *obd = class_conn2obd(conn);
1395         struct statfs sfs;
1396         int rc;
1397
1398         ENTRY;
1399         rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1400         if (!rc)
1401                 statfs_pack(osfs, &sfs);
1402
1403         return rc;
1404 }
1405
1406 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1407                            void *key, obd_count *vallen, void **val)
1408 {
1409         struct obd_device *obd;
1410         ENTRY;
1411
1412         obd = class_conn2obd(conn);
1413         if (!obd) {
1414                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1415                 RETURN(-EINVAL);
1416         }
1417
1418         if ( keylen == strlen("blocksize") &&
1419              memcmp(key, "blocksize", keylen) == 0 ) {
1420                 *vallen = sizeof(long);
1421                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1422                 RETURN(0);
1423         }
1424
1425         if ( keylen == strlen("blocksize_bits") &&
1426              memcmp(key, "blocksize_bits", keylen) == 0 ){
1427                 *vallen = sizeof(long);
1428                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1429                 RETURN(0);
1430         }
1431
1432         if ( keylen == strlen("root_ino") &&
1433              memcmp(key, "root_ino", keylen) == 0 ){
1434                 *vallen = sizeof(obd_id);
1435                 *val = (void *)(obd_id)FILTER_ROOTINO;
1436                 RETURN(0);
1437         }
1438
1439         CDEBUG(D_IOCTL, "invalid key\n");
1440         RETURN(-EINVAL);
1441 }
1442
1443 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1444                   struct lustre_handle *src_conn, struct obdo *src,
1445                   obd_size count, obd_off offset)
1446 {
1447         struct page *page;
1448         struct lov_stripe_md srcmd, dstmd;
1449         unsigned long index = 0;
1450         int err = 0;
1451
1452         memset(&srcmd, 0, sizeof(srcmd));
1453         memset(&dstmd, 0, sizeof(dstmd));
1454         srcmd.lsm_object_id = src->o_id;
1455         dstmd.lsm_object_id = dst->o_id;
1456
1457         ENTRY;
1458         CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1459                ", dst: ino "LPU64"\n",
1460                src->o_id, src->o_blocks, src->o_size, dst->o_id);
1461         page = alloc_page(GFP_USER);
1462         if (page == NULL)
1463                 RETURN(-ENOMEM);
1464
1465         while (TryLockPage(page))
1466                 ___wait_on_page(page);
1467
1468         /* XXX with brw vector I/O, we could batch up reads and writes here,
1469          *     all we need to do is allocate multiple pages to handle the I/Os
1470          *     and arrays to handle the request parameters.
1471          */
1472         while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1473                 struct brw_page pg;
1474                 struct io_cb_data *cbd = ll_init_cb();
1475
1476                 if (!cbd) {
1477                         err = -ENOMEM;
1478                         EXIT;
1479                         break;
1480                 }
1481
1482                 pg.pg = page;
1483                 pg.count = PAGE_SIZE;
1484                 pg.off = (page->index) << PAGE_SHIFT;
1485                 pg.flag = 0;
1486
1487                 page->index = index;
1488                 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1489                               ll_sync_io_cb, cbd);
1490
1491                 if ( err ) {
1492                         EXIT;
1493                         break;
1494                 }
1495
1496                 cbd = ll_init_cb();
1497                 if (!cbd) {
1498                         err = -ENOMEM;
1499                         EXIT;
1500                         break;
1501                 }
1502                 pg.flag = OBD_BRW_CREATE;
1503                 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1504
1505                 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1506                               ll_sync_io_cb, cbd);
1507
1508                 /* XXX should handle dst->o_size, dst->o_blocks here */
1509                 if ( err ) {
1510                         EXIT;
1511                         break;
1512                 }
1513
1514                 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1515
1516                 index++;
1517         }
1518         dst->o_size = src->o_size;
1519         dst->o_blocks = src->o_blocks;
1520         dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1521         unlock_page(page);
1522         __free_page(page);
1523
1524         RETURN(err);
1525 }
1526
1527
1528 static struct obd_ops filter_obd_ops = {
1529         o_get_info:    filter_get_info,
1530         o_setup:       filter_setup,
1531         o_cleanup:     filter_cleanup,
1532         o_connect:     filter_connect,
1533         o_disconnect:  filter_disconnect,
1534         o_statfs:      filter_statfs,
1535         o_getattr:     filter_getattr,
1536         o_create:      filter_create,
1537         o_setattr:     filter_setattr,
1538         o_destroy:     filter_destroy,
1539         o_open:        filter_open,
1540         o_close:       filter_close,
1541         o_brw:         filter_pgcache_brw,
1542         o_punch:       filter_truncate,
1543         o_preprw:      filter_preprw,
1544         o_commitrw:    filter_commitrw
1545 #if 0
1546         o_preallocate: filter_preallocate_inodes,
1547         o_migrate:     filter_migrate,
1548         o_copy:        filter_copy_data,
1549         o_iterate:     filter_iterate
1550 #endif
1551 };
1552
1553
1554 static int __init obdfilter_init(void)
1555 {
1556         printk(KERN_INFO "Filtering OBD driver  v0.001, info@clusterfs.com\n");
1557         return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1558 }
1559
1560 static void __exit obdfilter_exit(void)
1561 {
1562         class_unregister_type(OBD_FILTER_DEVICENAME);
1563 }
1564
1565 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1566 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1567 MODULE_LICENSE("GPL");
1568
1569 module_init(obdfilter_init);
1570 module_exit(obdfilter_exit);