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
  28 long filter_memory;
  29
  30 #define FILTER_ROOTINO 2
  31 #define FILTER_ROOTINO_STR __stringify(FILTER_ROOTINO)
  32
  33 #define S_SHIFT 12
  34 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
  35         [0]                     NULL,
  36         [S_IFREG >> S_SHIFT]    "R",
  37         [S_IFDIR >> S_SHIFT]    "D",
  38         [S_IFCHR >> S_SHIFT]    "C",
  39         [S_IFBLK >> S_SHIFT]    "B",
  40         [S_IFIFO >> S_SHIFT]    "F",
  41         [S_IFSOCK >> S_SHIFT]   "S",
  42         [S_IFLNK >> S_SHIFT]    "L"
  43 };
  44
  45 static inline const char *obd_mode_to_type(int mode)
  46 {
  47         return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
  48 }
  49
  50 /* write the pathname into the string */
  51 static int filter_id(char *buf, obd_id id, obd_mode mode)
  52 {
  53         return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
  54 }
  55
  56 static inline void f_dput(struct dentry *dentry)
  57 {
  58         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
  59                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
  60         LASSERT(atomic_read(&dentry->d_count) > 0);
  61         dput(dentry);
  62 }
  63
  64 /* setup the object store with correct subdirectories */
  65 static int filter_prep(struct obd_device *obddev)
  66 {
  67         struct obd_run_ctxt saved;
  68         struct filter_obd *filter = &obddev->u.filter;
  69         struct dentry *dentry;
  70         struct dentry *root;
  71         struct file *file;
  72         struct inode *inode;
  73         int rc = 0;
  74         __u64 lastino = 2;
  75         int mode = 0;
  76
  77         push_ctxt(&saved, &filter->fo_ctxt, NULL);
  78         dentry = simple_mkdir(current->fs->pwd, "O", 0700);
  79         CDEBUG(D_INODE, "got/created O: %p\n", dentry);
  80         if (IS_ERR(dentry)) {
  81                 rc = PTR_ERR(dentry);
  82                 CERROR("cannot open/create O: rc = %d\n", rc);
  83                 GOTO(out, rc);
  84         }
  85         filter->fo_dentry_O = dentry;
  86         dentry = simple_mkdir(current->fs->pwd, "P", 0700);
  87         CDEBUG(D_INODE, "got/created P: %p\n", dentry);
  88         if (IS_ERR(dentry)) {
  89                 rc = PTR_ERR(dentry);
  90                 CERROR("cannot open/create P: rc = %d\n", rc);
  91                 GOTO(out_O, rc);
  92         }
  93         f_dput(dentry);
  94         dentry = simple_mkdir(current->fs->pwd, "D", 0700);
  95         CDEBUG(D_INODE, "got/created D: %p\n", dentry);
  96         if (IS_ERR(dentry)) {
  97                 rc = PTR_ERR(dentry);
  98                 CERROR("cannot open/create D: rc = %d\n", rc);
  99                 GOTO(out_O, rc);
 100         }
 101
 102         root = simple_mknod(dentry, FILTER_ROOTINO_STR, S_IFREG | 0755);
 103         f_dput(dentry);
 104         if (IS_ERR(root)) {
 105                 rc = PTR_ERR(root);
 106                 CERROR("OBD filter: cannot open/create root %d: rc = %d\n",
 107                        FILTER_ROOTINO, rc);
 108                 GOTO(out_O, rc);
 109         }
 110         f_dput(root);
 111
 112         /*
 113          * Create directories and/or get dentries for each object type.
 114          * This saves us from having to do multiple lookups for each one.
 115          */
 116         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 117                 char *type = obd_type_by_mode[mode];
 118
 119                 if (!type) {
 120                         filter->fo_dentry_O_mode[mode] = NULL;
 121                         continue;
 122                 }
 123                 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
 124                 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
 125                 if (IS_ERR(dentry)) {
 126                         rc = PTR_ERR(dentry);
 127                         CERROR("cannot create O/%s: rc = %d\n", type, rc);
 128                         GOTO(out_O_mode, rc);
 129                 }
 130                 filter->fo_dentry_O_mode[mode] = dentry;
 131         }
 132
 133         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 134         if ( !file || IS_ERR(file) ) {
 135                 rc = PTR_ERR(file);
 136                 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
 137                        "D/status", rc);
 138                 GOTO(out_O_mode, rc);
 139         }
 140
 141         /* steal operations */
 142         inode = file->f_dentry->d_inode;
 143         filter->fo_fop = file->f_op;
 144         filter->fo_iop = inode->i_op;
 145         filter->fo_aops = inode->i_mapping->a_ops;
 146
 147         if (inode->i_size == 0) {
 148                 __u64 disk_lastino = cpu_to_le64(lastino);
 149                 ssize_t retval = file->f_op->write(file, (char *)&disk_lastino,
 150                                                    sizeof(disk_lastino),
 151                                                    &file->f_pos);
 152                 if (retval != sizeof(disk_lastino)) {
 153                         CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
 154                         filp_close(file, 0);
 155                         GOTO(out_O_mode, rc = -EIO);
 156                 }
 157         } else {
 158                 __u64 disk_lastino;
 159                 ssize_t retval = file->f_op->read(file, (char *)&disk_lastino,
 160                                                   sizeof(disk_lastino),
 161                                                   &file->f_pos);
 162                 if (retval != sizeof(disk_lastino)) {
 163                         CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
 164                         filp_close(file, 0);
 165                         GOTO(out_O_mode, rc = -EIO);
 166                 }
 167                 lastino = le64_to_cpu(disk_lastino);
 168         }
 169         filter->fo_lastino = lastino;
 170         filp_close(file, 0);
 171
 172         rc = 0;
 173  out:
 174         pop_ctxt(&saved);
 175
 176         return(rc);
 177
 178 out_O_mode:
 179         while (mode-- > 0) {
 180                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 181                 if (dentry) {
 182                         f_dput(dentry);
 183                         filter->fo_dentry_O_mode[mode] = NULL;
 184                 }
 185         }
 186 out_O:
 187         f_dput(filter->fo_dentry_O);
 188         filter->fo_dentry_O = NULL;
 189         goto out;
 190 }
 191
 192 /* cleanup the filter: write last used object id to status file */
 193 static void filter_post(struct obd_device *obddev)
 194 {
 195         struct obd_run_ctxt saved;
 196         struct filter_obd *filter = &obddev->u.filter;
 197         __u64 disk_lastino;
 198         long rc;
 199         struct file *file;
 200         int mode;
 201
 202         push_ctxt(&saved, &filter->fo_ctxt, NULL);
 203         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 204         if (IS_ERR(file)) {
 205                 CERROR("OBD filter: cannot create status file\n");
 206                 goto out;
 207         }
 208
 209         file->f_pos = 0;
 210         disk_lastino = cpu_to_le64(filter->fo_lastino);
 211         rc = file->f_op->write(file, (char *)&disk_lastino,
 212                        sizeof(disk_lastino), &file->f_pos);
 213         if (rc != sizeof(disk_lastino))
 214                 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
 215
 216         rc = filp_close(file, NULL);
 217         if (rc)
 218                 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
 219
 220         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 221                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 222                 if (dentry) {
 223                         f_dput(dentry);
 224                         filter->fo_dentry_O_mode[mode] = NULL;
 225                 }
 226         }
 227         f_dput(filter->fo_dentry_O);
 228 out:
 229         pop_ctxt(&saved);
 230 }
 231
 232
 233 static __u64 filter_next_id(struct obd_device *obddev)
 234 {
 235         obd_id id;
 236
 237         spin_lock(&obddev->u.filter.fo_lock);
 238         id = ++obddev->u.filter.fo_lastino;
 239         spin_unlock(&obddev->u.filter.fo_lock);
 240
 241         /* FIXME: write the lastino to disk here */
 242         return id;
 243 }
 244
 245 /* how to get files, dentries, inodes from object id's */
 246 /* parent i_sem is already held if needed for exclusivity */
 247 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
 248                                         struct dentry *dparent,
 249                                         __u64 id, __u32 type)
 250 {
 251         struct super_block *sb = obddev->u.filter.fo_sb;
 252         struct dentry *dchild;
 253         char name[32];
 254         int len;
 255         ENTRY;
 256
 257         if (!sb || !sb->s_dev) {
 258                 CERROR("fatal: device not initialized.\n");
 259                 RETURN(ERR_PTR(-ENXIO));
 260         }
 261
 262         if (id == 0) {
 263                 CERROR("fatal: invalid object #0\n");
 264                 LBUG();
 265                 RETURN(ERR_PTR(-ESTALE));
 266         }
 267
 268         if (!(type & S_IFMT)) {
 269                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 270                        __FUNCTION__, id, type);
 271                 RETURN(ERR_PTR(-EINVAL));
 272         }
 273
 274         len = sprintf(name, LPU64, id);
 275         CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
 276                name);
 277         dchild = lookup_one_len(name, dparent, len);
 278         if (IS_ERR(dchild)) {
 279                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 280                 RETURN(dchild);
 281         }
 282
 283         CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
 284                obd_mode_to_type(type), name, dchild,
 285                atomic_read(&dchild->d_count));
 286
 287         LASSERT(atomic_read(&dchild->d_count) > 0);
 288
 289         RETURN(dchild);
 290 }
 291
 292 static struct file *filter_obj_open(struct obd_device *obddev,
 293                                     __u64 id, __u32 type)
 294 {
 295         struct super_block *sb = obddev->u.filter.fo_sb;
 296         struct obd_run_ctxt saved;
 297         char name[24];
 298         struct file *file;
 299         ENTRY;
 300
 301         if (!sb || !sb->s_dev) {
 302                 CERROR("fatal: device not initialized.\n");
 303                 RETURN(ERR_PTR(-ENXIO));
 304         }
 305
 306         if (!id) {
 307                 CERROR("fatal: invalid obdo "LPU64"\n", id);
 308                 RETURN(ERR_PTR(-ESTALE));
 309         }
 310
 311         if (!(type & S_IFMT)) {
 312                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 313                        __FUNCTION__, id, type);
 314                 RETURN(ERR_PTR(-EINVAL));
 315         }
 316
 317         filter_id(name, id, type);
 318         push_ctxt(&saved, &obddev->u.filter.fo_ctxt, NULL);
 319         file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
 320         pop_ctxt(&saved);
 321
 322         CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
 323
 324         if (IS_ERR(file))
 325                 file = NULL;
 326         RETURN(file);
 327 }
 328
 329 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
 330 {
 331         struct filter_obd *filter = &obddev->u.filter;
 332
 333         return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
 334 }
 335
 336
 337 /* obd methods */
 338 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
 339                           obd_uuid_t cluuid)
 340 {
 341         int rc;
 342         ENTRY;
 343         MOD_INC_USE_COUNT;
 344         rc = class_connect(conn, obd, cluuid);
 345         if (rc)
 346                 MOD_DEC_USE_COUNT;
 347         RETURN(rc);
 348 }
 349
 350 static int filter_disconnect(struct lustre_handle *conn)
 351 {
 352         int rc;
 353         ENTRY;
 354
 355         rc = class_disconnect(conn);
 356         if (!rc)
 357                 MOD_DEC_USE_COUNT;
 358
 359         /* XXX cleanup preallocated inodes */
 360         RETURN(rc);
 361 }
 362
 363 /* mount the file system (secretly) */
 364 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
 365 {
 366         struct obd_ioctl_data* data = buf;
 367         struct filter_obd *filter;
 368         struct vfsmount *mnt;
 369         int err = 0;
 370         ENTRY;
 371
 372         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 373                 RETURN(-EINVAL);
 374
 375         MOD_INC_USE_COUNT;
 376         mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
 377         err = PTR_ERR(mnt);
 378         if (IS_ERR(mnt))
 379                 GOTO(err_dec, err);
 380
 381         filter = &obddev->u.filter;;
 382         filter->fo_vfsmnt = mnt;
 383         filter->fo_fstype = strdup(data->ioc_inlbuf2);
 384         filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
 385         /* XXX is this even possible if do_kern_mount succeeded? */
 386         if (!filter->fo_sb)
 387                 GOTO(err_kfree, err = -ENODEV);
 388
 389         OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
 390         filter->fo_ctxt.pwdmnt = mnt;
 391         filter->fo_ctxt.pwd = mnt->mnt_root;
 392         filter->fo_ctxt.fs = get_ds();
 393
 394         err = filter_prep(obddev);
 395         if (err)
 396                 GOTO(err_kfree, err);
 397         spin_lock_init(&filter->fo_lock);
 398
 399         obddev->obd_namespace =
 400                 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
 401         if (obddev->obd_namespace == NULL)
 402                 LBUG();
 403
 404         ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
 405                            "filter_ldlm_client", &obddev->obd_ldlm_client);
 406
 407         RETURN(0);
 408
 409 err_kfree:
 410         kfree(filter->fo_fstype);
 411         unlock_kernel();
 412         mntput(filter->fo_vfsmnt);
 413         filter->fo_sb = 0;
 414         lock_kernel();
 415
 416 err_dec:
 417         MOD_DEC_USE_COUNT;
 418         return err;
 419 }
 420
 421
 422 static int filter_cleanup(struct obd_device * obddev)
 423 {
 424         struct super_block *sb;
 425         ENTRY;
 426
 427         if (!list_empty(&obddev->obd_exports)) {
 428                 CERROR("still has clients!\n");
 429                 class_disconnect_all(obddev);
 430                 if (!list_empty(&obddev->obd_exports)) {
 431                         CERROR("still has exports after forced cleanup?\n");
 432                         RETURN(-EBUSY);
 433                 }
 434         }
 435
 436         ldlm_namespace_free(obddev->obd_namespace);
 437
 438         sb = obddev->u.filter.fo_sb;
 439         if (!obddev->u.filter.fo_sb)
 440                 RETURN(0);
 441
 442         filter_post(obddev);
 443
 444         shrink_dcache_parent(sb->s_root);
 445         unlock_kernel();
 446         mntput(obddev->u.filter.fo_vfsmnt);
 447         obddev->u.filter.fo_sb = 0;
 448         kfree(obddev->u.filter.fo_fstype);
 449
 450         lock_kernel();
 451
 452         MOD_DEC_USE_COUNT;
 453         RETURN(0);
 454 }
 455
 456
 457 static inline void filter_from_inode(struct obdo *oa, struct inode *inode,
 458                                      int valid)
 459 {
 460         int type = oa->o_mode & S_IFMT;
 461         ENTRY;
 462
 463         CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
 464                inode->i_ino, inode, (long)oa->o_id, valid);
 465         /* Don't copy the inode number in place of the object ID */
 466         obdo_from_inode(oa, inode, valid);
 467         oa->o_mode &= ~S_IFMT;
 468         oa->o_mode |= type;
 469
 470         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
 471                 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
 472                 oa->o_rdev = rdev;
 473                 oa->o_valid |= OBD_MD_FLRDEV;
 474         }
 475
 476         EXIT;
 477 }
 478
 479 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
 480                           struct lov_stripe_md *md)
 481 {
 482         struct obd_device *obddev = class_conn2obd(conn);
 483         struct dentry *dentry;
 484         int rc = 0;
 485         ENTRY;
 486
 487         if (!class_conn2export(conn)) {
 488                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 489                 RETURN(-EINVAL);
 490         }
 491
 492         obddev = class_conn2obd(conn);
 493         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 494                                    oa->o_id, oa->o_mode);
 495         if (IS_ERR(dentry))
 496                 RETURN(PTR_ERR(dentry));
 497
 498         if (!dentry->d_inode) {
 499                 CERROR("getattr on non-existent object: "LPU64"\n", oa->o_id);
 500                 GOTO(out_getattr, rc = -ENOENT);
 501         }
 502
 503         filter_from_inode(oa, dentry->d_inode, oa->o_valid);
 504
 505 out_getattr:
 506         f_dput(dentry);
 507         RETURN(rc);
 508 }
 509
 510 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
 511                           struct lov_stripe_md *md)
 512 {
 513         struct obd_run_ctxt saved;
 514         struct obd_device *obd = class_conn2obd(conn);
 515         struct dentry *dentry;
 516         struct iattr iattr;
 517         struct inode *inode;
 518         int rc;
 519         ENTRY;
 520
 521         iattr_from_obdo(&iattr, oa, oa->o_valid);
 522         iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
 523         dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
 524                                    oa->o_id, iattr.ia_mode);
 525         if (IS_ERR(dentry))
 526                 RETURN(PTR_ERR(dentry));
 527
 528         inode = dentry->d_inode;
 529         if (!inode) {
 530                 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
 531                 GOTO(out_setattr, rc = -ENOENT);
 532         }
 533
 534         lock_kernel();
 535         if (iattr.ia_valid & ATTR_SIZE)
 536                 down(&inode->i_sem);
 537         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 538         if (inode->i_op->setattr)
 539                 rc = inode->i_op->setattr(dentry, &iattr);
 540         else
 541                 rc = inode_setattr(inode, &iattr);
 542         pop_ctxt(&saved);
 543         if (iattr.ia_valid & ATTR_SIZE) {
 544                 up(&inode->i_sem);
 545                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 546                 obdo_from_inode(oa, inode, oa->o_valid);
 547         }
 548         unlock_kernel();
 549
 550 out_setattr:
 551         f_dput(dentry);
 552         RETURN(rc);
 553 }
 554
 555 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
 556                        struct lov_stripe_md *ea)
 557 {
 558         struct obd_export *export;
 559         struct obd_device *obd;
 560         struct dentry *dentry;
 561         /* ENTRY; */
 562
 563         export = class_conn2export(conn);
 564         if (!export) {
 565                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 566                 RETURN(-EINVAL);
 567         }
 568
 569         obd = export->exp_obd;
 570         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 571                                    oa->o_id, oa->o_mode);
 572         if (IS_ERR(dentry))
 573                 RETURN(PTR_ERR(dentry));
 574
 575         if (!dentry->d_inode) {
 576                 CERROR("opening non-existent objid "LPX64"\n", oa->o_id);
 577                 RETURN(-ENOENT);
 578         }
 579
 580         filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 581                           OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 582
 583         return 0;
 584 } /* filter_open */
 585
 586 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
 587                         struct lov_stripe_md *ea)
 588 {
 589         struct obd_device *obd;
 590         struct dentry *dentry;
 591         /* ENTRY; */
 592
 593         obd = class_conn2obd(conn);
 594         if (!obd) {
 595                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 596                 RETURN(-EINVAL);
 597         }
 598
 599         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 600                                    oa->o_id, oa->o_mode);
 601         if (IS_ERR(dentry))
 602                 RETURN(PTR_ERR(dentry));
 603         LASSERT(atomic_read(&dentry->d_count) > 1);
 604
 605         f_dput(dentry);  /* for the close */
 606         f_dput(dentry);  /* for this call */
 607         return 0;
 608 } /* filter_close */
 609
 610 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
 611                          struct lov_stripe_md **ea)
 612 {
 613         char name[64];
 614         struct obd_run_ctxt saved;
 615         struct dentry *new;
 616         int mode;
 617         struct obd_device *obd = class_conn2obd(conn);
 618         struct filter_obd *filter = &obd->u.filter;
 619         struct iattr;
 620         ENTRY;
 621
 622         if (!obd) {
 623                 CERROR("invalid client "LPX64"\n", conn->addr);
 624                 return -EINVAL;
 625         }
 626
 627         if (!(oa->o_mode && S_IFMT)) {
 628                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 629                        __FUNCTION__, oa->o_id, oa->o_mode);
 630                 return -ENOENT;
 631         }
 632
 633         oa->o_id = filter_next_id(obd);
 634
 635         //filter_id(name, oa->o_id, oa->o_mode);
 636         sprintf(name, LPU64, oa->o_id);
 637         mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
 638         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 639         new = simple_mknod(filter->fo_dentry_O_mode[S_IFREG >> S_SHIFT], name,
 640                            mode);
 641         pop_ctxt(&saved);
 642         if (IS_ERR(new)) {
 643                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
 644                 return -ENOENT;
 645         }
 646
 647         /* Set flags for fields we have set in the inode struct */
 648         oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 649                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
 650         filter_from_inode(oa, new->d_inode, oa->o_valid);
 651         dput(new);
 652
 653         return 0;
 654 }
 655
 656 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
 657                           struct lov_stripe_md *ea)
 658 {
 659         struct obd_device *obd;
 660         struct filter_obd *filter;
 661         struct obd_run_ctxt saved;
 662         struct inode *inode;
 663         struct dentry *dir_dentry, *object_dentry;
 664         int rc;
 665         ENTRY;
 666
 667         obd = class_conn2obd(conn);
 668         if (!obd) {
 669                 CERROR("invalid client "LPX64"\n", conn->addr);
 670                 RETURN(-EINVAL);
 671         }
 672
 673         CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
 674
 675         dir_dentry = filter_parent(obd, oa->o_mode);
 676         down(&dir_dentry->d_inode->i_sem);
 677
 678         object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
 679                                           oa->o_mode);
 680         if (IS_ERR(object_dentry))
 681                 GOTO(out, rc = -ENOENT);
 682
 683         inode = object_dentry->d_inode;
 684         if (!inode) {
 685                 CERROR("trying to destroy negative inode "LPX64"!\n", oa->o_id);
 686                 GOTO(out, rc = -ENOENT);
 687         }
 688
 689         if (inode->i_nlink != 1) {
 690                 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
 691                 LBUG();
 692                 inode->i_nlink = 1;
 693         }
 694         inode->i_mode = S_IFREG;
 695
 696         if (atomic_read(&inode->i_count) > 1) {
 697 #warning FIXME: need to handle open-unlinked case and move to hold dir
 698                 CERROR("inode has count %d\n", atomic_read(&inode->i_count));
 699         }
 700
 701         filter = &obd->u.filter;
 702         push_ctxt(&saved, &filter->fo_ctxt, NULL);
 703
 704         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 705         pop_ctxt(&saved);
 706         f_dput(object_dentry);
 707
 708         EXIT;
 709 out:
 710         up(&dir_dentry->d_inode->i_sem);
 711         return rc;
 712 }
 713
 714 /* NB count and offset are used for punch, but not truncate */
 715 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
 716                            struct lov_stripe_md *md,
 717                            obd_off start, obd_off end)
 718 {
 719         int error;
 720         ENTRY;
 721
 722         if (end != OBD_PUNCH_EOF)
 723                 CERROR("PUNCH not supported, only truncate works\n");
 724
 725         CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
 726                "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
 727         oa->o_size = start;
 728         error = filter_setattr(conn, oa, NULL);
 729         RETURN(error);
 730 }
 731
 732 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
 733                               struct lov_stripe_md *lsm, obd_count oa_bufs,
 734                               struct brw_page *pga, brw_callback_t callback,
 735                               struct io_cb_data *data)
 736 {
 737         struct obd_run_ctxt      saved;
 738         struct super_block      *sb;
 739         int                      pnum;          /* index to pages (bufs) */
 740         unsigned long            retval;
 741         int                      error;
 742         struct file             *file;
 743         struct obd_device       *obd = class_conn2obd(conn);
 744         int pg;
 745         ENTRY;
 746
 747         if (!obd) {
 748                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
 749                 RETURN(-EINVAL);
 750         }
 751
 752         sb = obd->u.filter.fo_sb;
 753         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 754         pnum = 0; /* pnum indexes buf 0..num_pages */
 755
 756         file = filter_obj_open(obd, lsm->lsm_object_id, S_IFREG);
 757         if (IS_ERR(file))
 758                 GOTO(out, retval = PTR_ERR(file));
 759
 760         /* count doubles as retval */
 761         for (pg = 0; pg < oa_bufs; pg++) {
 762                 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
 763                        ") off count ("LPU64",%d)\n",
 764                        cmd, pnum, file->f_dentry->d_inode->i_ino,
 765                        pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
 766                        (int)pga[pnum].count);
 767                 if (cmd & OBD_BRW_WRITE) {
 768                         loff_t off;
 769                         char *buffer;
 770                         off = pga[pnum].off;
 771                         buffer = kmap(pga[pnum].pg);
 772                         retval = file->f_op->write(file, buffer,
 773                                                    pga[pnum].count,
 774                                                    &off);
 775                         kunmap(pga[pnum].pg);
 776                         CDEBUG(D_INODE, "retval %ld\n", retval);
 777                 } else {
 778                         loff_t off = pga[pnum].off;
 779                         char *buffer = kmap(pga[pnum].pg);
 780
 781                         if (off >= file->f_dentry->d_inode->i_size) {
 782                                 memset(buffer, 0, pga[pnum].count);
 783                                 retval = pga[pnum].count;
 784                         } else {
 785                                 retval = file->f_op->read(file, buffer,
 786                                                           pga[pnum].count, &off);
 787                         }
 788                         kunmap(pga[pnum].pg);
 789
 790                         if (retval != pga[pnum].count) {
 791                                 filp_close(file, 0);
 792                                 GOTO(out, retval = -EIO);
 793                         }
 794                         CDEBUG(D_INODE, "retval %ld\n", retval);
 795                 }
 796                 pnum++;
 797         }
 798         /* sizes and blocks are set by generic_file_write */
 799         /* ctimes/mtimes will follow with a setattr call */
 800         filp_close(file, 0);
 801
 802         /* XXX: do something with callback if it is set? */
 803
 804         EXIT;
 805 out:
 806         pop_ctxt(&saved);
 807         error = (retval >= 0) ? 0 : retval;
 808         return error;
 809 }
 810
 811 /*
 812  * Calculate the number of buffer credits needed to write multiple pages in
 813  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
 814  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
 815  *
 816  * See comment above ext3_writepage_trans_blocks for details.  We assume
 817  * no data journaling is being done, but it does allow for all of the pages
 818  * being non-contiguous.  If we are guaranteed contiguous pages we could
 819  * reduce the number of (d)indirect blocks a lot.
 820  *
 821  * With N blocks per page and P pages, for each inode we have at most:
 822  * N*P indirect
 823  * min(N*P, blocksize/4 + 1) dindirect blocks
 824  * 1 tindirect
 825  *
 826  * For the entire filesystem, we have at most:
 827  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 828  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 829  * 1 inode block
 830  * 1 superblock
 831  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 832  */
 833 static int ext3_credits_needed(struct super_block *sb, int objcount,
 834                                struct obd_ioobj *obj)
 835 {
 836         struct obd_ioobj *o = obj;
 837         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 838         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 839         int nbitmaps = 0;
 840         int ngdblocks = 0;
 841         int needed = objcount + 1;
 842         int i;
 843
 844         for (i = 0; i < objcount; i++, o++) {
 845                 int nblocks = o->ioo_bufcnt * blockpp;
 846                 int ndindirect = min(nblocks, addrpp + 1);
 847                 int nindir = nblocks + ndindirect + 1;
 848
 849                 nbitmaps += nindir + nblocks;
 850                 ngdblocks += nindir + nblocks;
 851
 852                 needed += nindir;
 853         }
 854
 855         /* Assumes ext3 and extN have same sb_info layout at the start. */
 856         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 857                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 858         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 859                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 860
 861         needed += nbitmaps + ngdblocks;
 862
 863 #ifdef CONFIG_QUOTA
 864         /* We assume that there will be 1 bit set in s_dquot.flags for each
 865          * quota file that is active.  This is at least true for now.
 866          */
 867         needed += hweight32(sb_any_quota_enabled(sb)) *
 868                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 869 #endif
 870
 871         return needed;
 872 }
 873
 874 /* We have to start a huge journal transaction here to hold all of the
 875  * metadata for the pages being written here.  This is necessitated by
 876  * the fact that we do lots of prepare_write operations before we do
 877  * any of the matching commit_write operations, so even if we split
 878  * up to use "smaller" transactions none of them could complete until
 879  * all of them were opened.  By having a single journal transaction,
 880  * we eliminate duplicate reservations for common blocks like the
 881  * superblock and group descriptors or bitmaps.
 882  *
 883  * We will start the transaction here, but each prepare_write will
 884  * add a refcount to the transaction, and each commit_write will
 885  * remove a refcount.  The transaction will be closed when all of
 886  * the pages have been written.
 887  */
 888 static void *ext3_filter_journal_start(struct filter_obd *filter,
 889                                        int objcount, struct obd_ioobj *obj,
 890                                        int niocount, struct niobuf_remote *nb)
 891 {
 892         journal_t *journal = NULL;
 893         handle_t *handle = NULL;
 894         int needed;
 895
 896         /* It appears that some kernels have different values for
 897          * EXT*_MAX_GROUP_LOADED (either 8 or 32), so we cannot
 898          * assume anything after s_inode_bitmap_number is the same.
 899          */
 900         if (!strcmp(filter->fo_fstype, "ext3"))
 901                 journal = EXT3_SB(filter->fo_sb)->s_journal;
 902         else if (!strcmp(filter->fo_fstype, "extN"))
 903                 journal = EXTN_SB(filter->fo_sb)->s_journal;
 904         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
 905
 906         /* The number of blocks we could _possibly_ dirty can very large.
 907          * We reduce our request if it is absurd (and we couldn't get that
 908          * many credits for a single handle anyways).
 909          *
 910          * At some point we have to limit the size of I/Os sent at one time,
 911          * increase the size of the journal, or we have to calculate the
 912          * actual journal requirements more carefully by checking all of
 913          * the blocks instead of being maximally pessimistic.  It remains to
 914          * be seen if this is a real problem or not.
 915          */
 916         if (needed > journal->j_max_transaction_buffers) {
 917                 CERROR("want too many journal credits (%d) using %d instead\n",
 918                        needed, journal->j_max_transaction_buffers);
 919                 needed = journal->j_max_transaction_buffers;
 920         }
 921
 922         lock_kernel();
 923         handle = journal_start(journal, needed);
 924         unlock_kernel();
 925         if (IS_ERR(handle))
 926                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 927                        PTR_ERR(handle));
 928
 929         return(handle);
 930 }
 931
 932 static void *filter_journal_start(void **journal_save,
 933                                   struct filter_obd *filter,
 934                                   int objcount, struct obd_ioobj *obj,
 935                                   int niocount, struct niobuf_remote *nb)
 936 {
 937         void *handle = NULL;
 938
 939         /* This may not be necessary - we probably never have a
 940          * transaction started when we enter here, so we can
 941          * remove the saving of the journal state entirely.
 942          * For now leave it in just to see if it ever happens.
 943          */
 944         *journal_save = current->journal_info;
 945         if (*journal_save) {
 946                 CERROR("Already have handle %p???\n", *journal_save);
 947                 LBUG();
 948                 current->journal_info = NULL;
 949         }
 950
 951         if (!strcmp(filter->fo_fstype, "ext3") ||
 952             !strcmp(filter->fo_fstype, "extN"))
 953                 handle = ext3_filter_journal_start(filter, objcount, obj,
 954                                                    niocount, nb);
 955         return handle;
 956 }
 957
 958 static int ext3_filter_journal_stop(void *handle)
 959 {
 960         int rc;
 961
 962         /* We got a refcount on the handle for each call to prepare_write,
 963          * so we can drop the "parent" handle here to avoid the need for
 964          * osc to call back into filterobd to close the handle.  The
 965          * remaining references will be dropped in commit_write.
 966          */
 967         lock_kernel();
 968         rc = journal_stop((handle_t *)handle);
 969         unlock_kernel();
 970
 971         return rc;
 972 }
 973
 974 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
 975                                void *handle)
 976 {
 977         int rc = 0;
 978
 979         if (!strcmp(filter->fo_fstype, "ext3") ||
 980             !strcmp(filter->fo_fstype, "extN"))
 981                 rc = ext3_filter_journal_stop(handle);
 982
 983         if (rc)
 984                 CERROR("error on journal stop: rc = %d\n", rc);
 985
 986         current->journal_info = journal_save;
 987
 988         return rc;
 989 }
 990
 991 static inline void lustre_put_page(struct page *page)
 992 {
 993         kunmap(page);
 994         page_cache_release(page);
 995 }
 996
 997 static struct page *
 998 lustre_get_page_read(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 = read_cache_page(mapping, index,
1005                                (filler_t*)mapping->a_ops->readpage, NULL);
1006         if (!IS_ERR(page)) {
1007                 wait_on_page(page);
1008                 kmap(page);
1009                 if (!Page_Uptodate(page)) {
1010                         CERROR("page index %lu not uptodate\n", index);
1011                         GOTO(err_page, rc = -EIO);
1012                 }
1013                 if (PageError(page)) {
1014                         CERROR("page index %lu has error\n", index);
1015                         GOTO(err_page, rc = -EIO);
1016                 }
1017         }
1018         return page;
1019
1020 err_page:
1021         lustre_put_page(page);
1022         return ERR_PTR(rc);
1023 }
1024
1025 static struct page *
1026 lustre_get_page_write(struct inode *inode, unsigned long index)
1027 {
1028         struct address_space *mapping = inode->i_mapping;
1029         struct page *page;
1030         int rc;
1031
1032         page = grab_cache_page(mapping, index); /* locked page */
1033
1034         if (!IS_ERR(page)) {
1035                 kmap(page);
1036                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1037                  * a no-op for most filesystems, because we write the whole
1038                  * page.  For partial-page I/O this will read in the page.
1039                  */
1040                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1041                 if (rc) {
1042                         CERROR("page index %lu, rc = %d\n", index, rc);
1043                         if (rc != -ENOSPC)
1044                                 LBUG();
1045                         GOTO(err_unlock, rc);
1046                 }
1047                 /* XXX not sure if we need this if we are overwriting page */
1048                 if (PageError(page)) {
1049                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1050                         LBUG();
1051                         GOTO(err_unlock, rc = -EIO);
1052                 }
1053         }
1054         return page;
1055
1056 err_unlock:
1057         unlock_page(page);
1058         lustre_put_page(page);
1059         return ERR_PTR(rc);
1060 }
1061
1062 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1063 {
1064         struct inode *inode = page->mapping->host;
1065         int err;
1066
1067         err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1068         if (!err && IS_SYNC(inode))
1069                 err = waitfor_one_page(page);
1070
1071         //SetPageUptodate(page); // the client commit_write will do this
1072
1073         SetPageReferenced(page);
1074         unlock_page(page);
1075         lustre_put_page(page);
1076         return err;
1077 }
1078
1079 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1080                                    struct niobuf_local *lnb, int *pglocked)
1081 {
1082         struct address_space *mapping = inode->i_mapping;
1083         struct page *page;
1084         int rc;
1085
1086         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1087         if (*pglocked)
1088                 page = grab_cache_page_nowait(mapping, index); /* locked page */
1089         else
1090                 page = grab_cache_page(mapping, index); /* locked page */
1091
1092
1093         /* This page is currently locked, so get a temporary page instead. */
1094         /* XXX I believe this is a very dangerous thing to do - consider if
1095          *     we had multiple writers for the same file (definitely the case
1096          *     if we are using this codepath).  If writer A locks the page,
1097          *     writer B writes to a copy (as here), writer A drops the page
1098          *     lock, and writer C grabs the lock before B does, then B will
1099          *     later overwrite the data from C, even if C had LDLM locked
1100          *     and initiated the write after B did.
1101          */
1102         if (!page) {
1103                 unsigned long addr;
1104                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1105                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1106                 if (!addr) {
1107                         CERROR("no memory for a temp page\n");
1108                         LBUG();
1109                         GOTO(err, rc = -ENOMEM);
1110                 }
1111                 page = virt_to_page(addr);
1112                 kmap(page);
1113                 page->index = index;
1114                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1115         } else if (!IS_ERR(page)) {
1116                 (*pglocked)++;
1117                 kmap(page);
1118
1119                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1120                  * a no-op for most filesystems, because we write the whole
1121                  * page.  For partial-page I/O this will read in the page.
1122                  */
1123                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1124                 if (rc) {
1125                         CERROR("page index %lu, rc = %d\n", index, rc);
1126                         if (rc != -ENOSPC)
1127                                 LBUG();
1128                         GOTO(err_unlock, rc);
1129                 }
1130                 /* XXX not sure if we need this if we are overwriting page */
1131                 if (PageError(page)) {
1132                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1133                         LBUG();
1134                         GOTO(err_unlock, rc = -EIO);
1135                 }
1136         }
1137         return page;
1138
1139 err_unlock:
1140         unlock_page(page);
1141         lustre_put_page(page);
1142 err:
1143         return ERR_PTR(rc);
1144 }
1145
1146 /*
1147  * We need to balance prepare_write() calls with commit_write() calls.
1148  * If the page has been prepared, but we have no data for it, we don't
1149  * want to overwrite valid data on disk, but we still need to zero out
1150  * data for space which was newly allocated.  Like part of what happens
1151  * in __block_prepare_write() for newly allocated blocks.
1152  *
1153  * XXX currently __block_prepare_write() creates buffers for all the
1154  *     pages, and the filesystems mark these buffers as BH_New if they
1155  *     were newly allocated from disk. We use the BH_New flag similarly.
1156  */
1157 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1158                                int err)
1159 {
1160         if (err) {
1161                 unsigned block_start, block_end;
1162                 struct buffer_head *bh, *head = page->buffers;
1163                 unsigned blocksize = head->b_size;
1164                 void *addr = page_address(page);
1165
1166                 /* debugging: just seeing if this ever happens */
1167                 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1168                        page->index, page->mapping->host->i_ino, err);
1169
1170                 /* Currently one buffer per page, but in the future... */
1171                 for (bh = head, block_start = 0; bh != head || !block_start;
1172                      block_start = block_end, bh = bh->b_this_page) {
1173                         block_end = block_start + blocksize;
1174                         if (buffer_new(bh))
1175                                 memset(addr + block_start, 0, blocksize);
1176                 }
1177         }
1178
1179         return lustre_commit_write(page, from, to);
1180 }
1181
1182 static int filter_preprw(int cmd, struct lustre_handle *conn,
1183                          int objcount, struct obd_ioobj *obj,
1184                          int niocount, struct niobuf_remote *nb,
1185                          struct niobuf_local *res, void **desc_private)
1186 {
1187         struct obd_run_ctxt saved;
1188         struct obd_device *obd;
1189         struct obd_ioobj *o = obj;
1190         struct niobuf_remote *b = nb;
1191         struct niobuf_local *r = res;
1192         void *journal_save = NULL;
1193         int pglocked = 0;
1194         int rc = 0;
1195         int i;
1196         ENTRY;
1197
1198         obd = class_conn2obd(conn);
1199         if (!obd) {
1200                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1201                 RETURN(-EINVAL);
1202         }
1203         memset(res, 0, sizeof(*res) * niocount);
1204
1205         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1206
1207         if (cmd & OBD_BRW_WRITE) {
1208                 *desc_private = filter_journal_start(&journal_save,
1209                                                      &obd->u.filter,
1210                                                      objcount, obj, niocount,
1211                                                      nb);
1212                 if (IS_ERR(*desc_private))
1213                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1214         }
1215
1216         for (i = 0; i < objcount; i++, o++) {
1217                 struct dentry *dentry;
1218                 struct inode *inode;
1219                 int j;
1220
1221                 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1222                                            o->ioo_id, S_IFREG);
1223                 if (IS_ERR(dentry))
1224                         GOTO(out_clean, rc = PTR_ERR(dentry));
1225                 inode = dentry->d_inode;
1226                 if (!inode) {
1227                         CERROR("trying to BRW to non-existent file "LPU64"\n",
1228                                o->ioo_id);
1229                         f_dput(dentry);
1230                         GOTO(out_clean, rc = -ENOENT);
1231                 }
1232
1233                 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1234                         unsigned long index = b->offset >> PAGE_SHIFT;
1235                         struct page *page;
1236
1237                         if (j == 0)
1238                                 r->dentry = dentry;
1239                         else
1240                                 r->dentry = dget(dentry);
1241
1242                         if (cmd & OBD_BRW_WRITE)
1243                                 page = filter_get_page_write(inode, index, r,
1244                                                              &pglocked);
1245                         else
1246                                 page = lustre_get_page_read(inode, index);
1247
1248                         if (IS_ERR(page)) {
1249                                 f_dput(dentry);
1250                                 GOTO(out_clean, rc = PTR_ERR(page));
1251                         }
1252
1253                         r->addr = page_address(page);
1254                         r->offset = b->offset;
1255                         r->page = page;
1256                         r->len = b->len;
1257                 }
1258         }
1259
1260 out_stop:
1261         if (cmd & OBD_BRW_WRITE) {
1262                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1263                                               *desc_private);
1264                 if (!rc)
1265                         rc = err;
1266         }
1267 out_ctxt:
1268         pop_ctxt(&saved);
1269         RETURN(rc);
1270 out_clean:
1271         while (r-- > res) {
1272                 CERROR("error cleanup on brw\n");
1273                 f_dput(r->dentry);
1274                 if (cmd & OBD_BRW_WRITE)
1275                         filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1276                 else
1277                         lustre_put_page(r->page);
1278         }
1279         goto out_stop;
1280 }
1281
1282 static int filter_write_locked_page(struct niobuf_local *lnb)
1283 {
1284         struct page *lpage;
1285         int rc;
1286
1287         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1288         if (IS_ERR(lpage)) {
1289                 /* It is highly unlikely that we would ever get an error here.
1290                  * The page we want to get was previously locked, so it had to
1291                  * have already allocated the space, and we were just writing
1292                  * over the same data, so there would be no hole in the file.
1293                  *
1294                  * XXX: possibility of a race with truncate could exist, need
1295                  *      to check that.  There are no guarantees w.r.t.
1296                  *      write order even on a local filesystem, although the
1297                  *      normal response would be to return the number of bytes
1298                  *      successfully written and leave the rest to the app.
1299                  */
1300                 rc = PTR_ERR(lpage);
1301                 CERROR("error getting locked page index %ld: rc = %d\n",
1302                        lnb->page->index, rc);
1303                 GOTO(out, rc);
1304         }
1305
1306         /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1307          * lustre_commit_write() below, lnb->page was kmapped previously in
1308          * filter_get_page_write() and kunmapped in lustre_put_page() below.
1309          */
1310         memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1311         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1312         if (rc)
1313                 CERROR("error committing locked page %ld: rc = %d\n",
1314                        lnb->page->index, rc);
1315 out:
1316         lustre_put_page(lnb->page);
1317
1318         return rc;
1319 }
1320
1321 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1322                            int objcount, struct obd_ioobj *obj,
1323                            int niocount, struct niobuf_local *res,
1324                            void *private)
1325 {
1326         struct obd_run_ctxt saved;
1327         struct obd_ioobj *o;
1328         struct niobuf_local *r;
1329         struct obd_device *obd = class_conn2obd(conn);
1330         void *journal_save;
1331         int found_locked = 0;
1332         int rc = 0;
1333         int i;
1334         ENTRY;
1335
1336         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1337         journal_save = current->journal_info;
1338         LASSERT(!journal_save);
1339
1340         current->journal_info = private;
1341         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1342                 int j;
1343                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1344                         struct page *page = r->page;
1345
1346                         if (!page)
1347                                 LBUG();
1348
1349                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1350                                 found_locked++;
1351                                 continue;
1352                         }
1353
1354                         if (cmd & OBD_BRW_WRITE) {
1355                                 int err = filter_commit_write(page, 0,
1356                                                               r->len, 0);
1357
1358                                 if (!rc)
1359                                         rc = err;
1360                         } else
1361                                 lustre_put_page(page);
1362
1363                         f_dput(r->dentry);
1364                 }
1365         }
1366         current->journal_info = journal_save;
1367
1368         if (!found_locked)
1369                 goto out_ctxt;
1370
1371         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1372                 int j;
1373                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1374                         int err;
1375                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1376                                 continue;
1377
1378                         err = filter_write_locked_page(r);
1379                         if (!rc)
1380                                 rc = err;
1381                         f_dput(r->dentry);
1382                 }
1383         }
1384
1385 out_ctxt:
1386         pop_ctxt(&saved);
1387         RETURN(rc);
1388 }
1389
1390 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1391 {
1392         struct obd_device *obd = class_conn2obd(conn);
1393         struct statfs sfs;
1394         int rc;
1395
1396         ENTRY;
1397         rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1398         if (!rc)
1399                 statfs_pack(osfs, &sfs);
1400
1401         return rc;
1402 }
1403
1404 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1405                            void *key, obd_count *vallen, void **val)
1406 {
1407         struct obd_device *obd;
1408         ENTRY;
1409
1410         obd = class_conn2obd(conn);
1411         if (!obd) {
1412                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1413                 RETURN(-EINVAL);
1414         }
1415
1416         if ( keylen == strlen("blocksize") &&
1417              memcmp(key, "blocksize", keylen) == 0 ) {
1418                 *vallen = sizeof(long);
1419                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1420                 RETURN(0);
1421         }
1422
1423         if ( keylen == strlen("blocksize_bits") &&
1424              memcmp(key, "blocksize_bits", keylen) == 0 ){
1425                 *vallen = sizeof(long);
1426                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1427                 RETURN(0);
1428         }
1429
1430         if ( keylen == strlen("root_ino") &&
1431              memcmp(key, "root_ino", keylen) == 0 ){
1432                 *vallen = sizeof(obd_id);
1433                 *val = (void *)(obd_id)FILTER_ROOTINO;
1434                 RETURN(0);
1435         }
1436
1437         CDEBUG(D_IOCTL, "invalid key\n");
1438         RETURN(-EINVAL);
1439 }
1440
1441 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1442                   struct lustre_handle *src_conn, struct obdo *src,
1443                   obd_size count, obd_off offset)
1444 {
1445         struct page *page;
1446         struct lov_stripe_md srcmd, dstmd;
1447         unsigned long index = 0;
1448         int err = 0;
1449
1450         memset(&srcmd, 0, sizeof(srcmd));
1451         memset(&dstmd, 0, sizeof(dstmd));
1452         srcmd.lsm_object_id = src->o_id;
1453         dstmd.lsm_object_id = dst->o_id;
1454
1455         ENTRY;
1456         CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1457                ", dst: ino "LPU64"\n",
1458                src->o_id, src->o_blocks, src->o_size, dst->o_id);
1459         page = alloc_page(GFP_USER);
1460         if (page == NULL)
1461                 RETURN(-ENOMEM);
1462
1463         while (TryLockPage(page))
1464                 ___wait_on_page(page);
1465
1466         /* XXX with brw vector I/O, we could batch up reads and writes here,
1467          *     all we need to do is allocate multiple pages to handle the I/Os
1468          *     and arrays to handle the request parameters.
1469          */
1470         while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1471                 struct brw_page pg;
1472                 struct io_cb_data *cbd = ll_init_cb();
1473
1474                 if (!cbd) {
1475                         err = -ENOMEM;
1476                         EXIT;
1477                         break;
1478                 }
1479
1480                 pg.pg = page;
1481                 pg.count = PAGE_SIZE;
1482                 pg.off = (page->index) << PAGE_SHIFT;
1483                 pg.flag = 0;
1484
1485                 page->index = index;
1486                 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1487                               ll_sync_io_cb, cbd);
1488
1489                 if ( err ) {
1490                         EXIT;
1491                         break;
1492                 }
1493
1494                 cbd = ll_init_cb();
1495                 if (!cbd) {
1496                         err = -ENOMEM;
1497                         EXIT;
1498                         break;
1499                 }
1500                 pg.flag = OBD_BRW_CREATE;
1501                 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1502
1503                 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1504                               ll_sync_io_cb, cbd);
1505
1506                 /* XXX should handle dst->o_size, dst->o_blocks here */
1507                 if ( err ) {
1508                         EXIT;
1509                         break;
1510                 }
1511
1512                 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1513
1514                 index++;
1515         }
1516         dst->o_size = src->o_size;
1517         dst->o_blocks = src->o_blocks;
1518         dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1519         unlock_page(page);
1520         __free_page(page);
1521
1522         RETURN(err);
1523 }
1524
1525
1526 static struct obd_ops filter_obd_ops = {
1527         o_get_info:    filter_get_info,
1528         o_setup:       filter_setup,
1529         o_cleanup:     filter_cleanup,
1530         o_connect:     filter_connect,
1531         o_disconnect:  filter_disconnect,
1532         o_statfs:      filter_statfs,
1533         o_getattr:     filter_getattr,
1534         o_create:      filter_create,
1535         o_setattr:     filter_setattr,
1536         o_destroy:     filter_destroy,
1537         o_open:        filter_open,
1538         o_close:       filter_close,
1539         o_brw:         filter_pgcache_brw,
1540         o_punch:       filter_truncate,
1541         o_preprw:      filter_preprw,
1542         o_commitrw:    filter_commitrw
1543 #if 0
1544         o_preallocate: filter_preallocate_inodes,
1545         o_migrate:     filter_migrate,
1546         o_copy:        filter_copy_data,
1547         o_iterate:     filter_iterate
1548 #endif
1549 };
1550
1551
1552 static int __init obdfilter_init(void)
1553 {
1554         printk(KERN_INFO "Filtering OBD driver  v0.001, info@clusterfs.com\n");
1555         return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1556 }
1557
1558 static void __exit obdfilter_exit(void)
1559 {
1560         class_unregister_type(OBD_FILTER_DEVICENAME);
1561 }
1562
1563 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1564 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1565 MODULE_LICENSE("GPL");
1566
1567 module_init(obdfilter_init);
1568 module_exit(obdfilter_exit);