lustre/obdfilter/filter.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/fs/filter/filter.c
   5  *
   6  * Copyright (C) 2001  Cluster File Systems, Inc.
   7  *
   8  * This code is issued under the GNU General Public License.
   9  * See the file COPYING in this distribution
  10  *
  11  * by Peter Braam <braam@clusterfs.com>
  12  */
  13
  14 #define EXPORT_SYMTAB
  15 #define DEBUG_SUBSYSTEM S_FILTER
  16
  17 #include <linux/module.h>
  18 #include <linux/obd_filter.h>
  19 #include <linux/ext3_jbd.h>
  20 #include <linux/quotaops.h>
  21
  22 extern struct obd_device obd_dev[MAX_OBD_DEVICES];
  23 long filter_memory;
  24
  25 #define FILTER_ROOTINO 2
  26
  27 #define S_SHIFT 12
  28 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
  29         [0]                     NULL,
  30         [S_IFREG >> S_SHIFT]    "R",
  31         [S_IFDIR >> S_SHIFT]    "D",
  32         [S_IFCHR >> S_SHIFT]    "C",
  33         [S_IFBLK >> S_SHIFT]    "B",
  34         [S_IFIFO >> S_SHIFT]    "F",
  35         [S_IFSOCK >> S_SHIFT]   "S",
  36         [S_IFLNK >> S_SHIFT]    "L"
  37 };
  38
  39 static inline const char *obd_mode_to_type(int mode)
  40 {
  41         return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
  42 }
  43
  44 /* write the pathname into the string */
  45 static int filter_id(char *buf, obd_id id, obd_mode mode)
  46 {
  47         return sprintf(buf, "O/%s/%Ld", obd_mode_to_type(mode),
  48                        (unsigned long long)id);
  49 }
  50
  51 /* setup the object store with correct subdirectories */
  52 static int filter_prep(struct obd_device *obddev)
  53 {
  54         struct obd_run_ctxt saved;
  55         struct filter_obd *filter = &obddev->u.filter;
  56         struct dentry *dentry;
  57         struct file *file;
  58         struct inode *inode;
  59         loff_t off;
  60         int rc = 0;
  61         char rootid[128];
  62         __u64 lastino = 2;
  63         int mode = 0;
  64
  65         push_ctxt(&saved, &filter->fo_ctxt);
  66         dentry = simple_mkdir(current->fs->pwd, "O", 0700);
  67         CDEBUG(D_INODE, "got/created O: %p\n", dentry);
  68         if (IS_ERR(dentry)) {
  69                 rc = PTR_ERR(dentry);
  70                 CERROR("cannot open/create O: rc = %d\n", rc);
  71                 GOTO(out, rc);
  72         }
  73         filter->fo_dentry_O = dentry;
  74         dentry = simple_mkdir(current->fs->pwd, "P", 0700);
  75         CDEBUG(D_INODE, "got/created P: %p\n", dentry);
  76         if (IS_ERR(dentry)) {
  77                 rc = PTR_ERR(dentry);
  78                 CERROR("cannot open/create P: rc = %d\n", rc);
  79                 GOTO(out_O, rc);
  80         }
  81         CDEBUG(D_INODE, "putting P: %p, count = %d\n", dentry,
  82                atomic_read(&dentry->d_count) - 1);
  83         dput(dentry);
  84         dentry = simple_mkdir(current->fs->pwd, "D", 0700);
  85         CDEBUG(D_INODE, "got/created D: %p\n", dentry);
  86         if (IS_ERR(dentry)) {
  87                 rc = PTR_ERR(dentry);
  88                 CERROR("cannot open/create D: rc = %d\n", rc);
  89                 GOTO(out_O, rc);
  90         }
  91         CDEBUG(D_INODE, "putting D: %p, count = %d\n", dentry,
  92                atomic_read(&dentry->d_count) - 1);
  93         dput(dentry);
  94
  95         /*
  96          * Create directories and/or get dentries for each object type.
  97          * This saves us from having to do multiple lookups for each one.
  98          */
  99         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 100                 char *type = obd_type_by_mode[mode];
 101
 102                 if (!type) {
 103                         filter->fo_dentry_O_mode[mode] = NULL;
 104                         continue;
 105                 }
 106                 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
 107                 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
 108                 if (IS_ERR(dentry)) {
 109                         rc = PTR_ERR(dentry);
 110                         CERROR("cannot create O/%s: rc = %d\n", type, rc);
 111                         GOTO(out_O_mode, rc);
 112                 }
 113                 filter->fo_dentry_O_mode[mode] = dentry;
 114         }
 115
 116         filter_id(rootid, FILTER_ROOTINO, S_IFDIR);
 117         file = filp_open(rootid, O_RDWR | O_CREAT, 0755);
 118         if (IS_ERR(file)) {
 119                 rc = PTR_ERR(file);
 120                 CERROR("OBD filter: cannot open/create root %s: rc = %d\n",
 121                        rootid, rc);
 122                 GOTO(out_O_mode, rc);
 123         }
 124         filp_close(file, 0);
 125
 126         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 127         if ( !file || IS_ERR(file) ) {
 128                 rc = PTR_ERR(file);
 129                 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
 130                        "D/status", rc);
 131                 GOTO(out_O_mode, rc);
 132         }
 133
 134         /* steal operations */
 135         inode = file->f_dentry->d_inode;
 136         filter->fo_fop = file->f_op;
 137         filter->fo_iop = inode->i_op;
 138         filter->fo_aops = inode->i_mapping->a_ops;
 139
 140         off = 0;
 141         if (inode->i_size == 0) {
 142                 ssize_t retval = file->f_op->write(file, (char *)&lastino,
 143                                                    sizeof(lastino), &off);
 144                 if (retval != sizeof(lastino)) {
 145                         CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
 146                         filp_close(file, 0);
 147                         GOTO(out_O_mode, rc = -EIO);
 148                 }
 149         } else {
 150                 ssize_t retval = file->f_op->read(file, (char *)&lastino,
 151                                                   sizeof(lastino), &off);
 152                 if (retval != sizeof(lastino)) {
 153                         CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
 154                         filp_close(file, 0);
 155                         GOTO(out_O_mode, rc = -EIO);
 156                 }
 157         }
 158         filter->fo_lastino = lastino;
 159         filp_close(file, 0);
 160
 161         rc = 0;
 162  out:
 163         pop_ctxt(&saved);
 164
 165         return(rc);
 166
 167 out_O_mode:
 168         while (mode-- > 0) {
 169                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 170                 if (dentry) {
 171                         CDEBUG(D_INODE, "putting O/%s: %p, count = %d\n",
 172                                obd_type_by_mode[mode], dentry,
 173                                atomic_read(&dentry->d_count) - 1);
 174                         dput(dentry);
 175                         filter->fo_dentry_O_mode[mode] = NULL;
 176                 }
 177         }
 178 out_O:
 179         CDEBUG(D_INODE, "putting O: %p, count = %d\n", filter->fo_dentry_O,
 180                atomic_read(&filter->fo_dentry_O->d_count) - 1);
 181         dput(filter->fo_dentry_O);
 182         filter->fo_dentry_O = NULL;
 183         goto out;
 184 }
 185
 186 /* cleanup the filter: write last used object id to status file */
 187 static void filter_post(struct obd_device *obddev)
 188 {
 189         struct obd_run_ctxt saved;
 190         struct filter_obd *filter = &obddev->u.filter;
 191         long rc;
 192         struct file *file;
 193         loff_t off = 0;
 194         int mode;
 195
 196         push_ctxt(&saved, &filter->fo_ctxt);
 197         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 198         if (IS_ERR(file)) {
 199                 CERROR("OBD filter: cannot create status file\n");
 200                 goto out;
 201         }
 202         rc = file->f_op->write(file, (char *)&filter->fo_lastino,
 203                        sizeof(filter->fo_lastino), &off);
 204         if (rc != sizeof(filter->fo_lastino))
 205                 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
 206
 207         rc = filp_close(file, NULL);
 208         if (rc)
 209                 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
 210
 211         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 212                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 213                 if (dentry) {
 214                         CDEBUG(D_INODE, "putting O/%s: %p, count = %d\n",
 215                                obd_type_by_mode[mode], dentry,
 216                                atomic_read(&dentry->d_count) - 1);
 217                         dput(dentry);
 218                         filter->fo_dentry_O_mode[mode] = NULL;
 219                 }
 220         }
 221         CDEBUG(D_INODE, "putting O: %p, count = %d\n", filter->fo_dentry_O,
 222                atomic_read(&filter->fo_dentry_O->d_count) - 1);
 223         dput(filter->fo_dentry_O);
 224 out:
 225         pop_ctxt(&saved);
 226 }
 227
 228
 229 static __u64 filter_next_id(struct obd_device *obddev)
 230 {
 231         __u64 id;
 232         spin_lock(&obddev->u.filter.fo_lock);
 233         obddev->u.filter.fo_lastino++;
 234         id =    obddev->u.filter.fo_lastino;
 235         spin_unlock(&obddev->u.filter.fo_lock);
 236         return id;
 237 }
 238
 239 /* how to get files, dentries, inodes from object id's */
 240 /* parent i_sem is already held if needed for exclusivity */
 241 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
 242                                         struct dentry *dparent,
 243                                         __u64 id, __u32 type)
 244 {
 245         struct super_block *sb = obddev->u.filter.fo_sb;
 246         struct dentry *dchild;
 247         char name[32];
 248         int len;
 249         ENTRY;
 250
 251         if (!sb || !sb->s_dev) {
 252                 CERROR("fatal: device not initialized.\n");
 253                 RETURN(ERR_PTR(-ENXIO));
 254         }
 255
 256         if (id == 0) {
 257                 CERROR("fatal: invalid object #0\n");
 258                 LBUG();
 259                 RETURN(ERR_PTR(-ESTALE));
 260         }
 261
 262         if (!(type & S_IFMT)) {
 263                 CERROR("OBD %s, object %Lu has bad type: %o\n", __FUNCTION__,
 264                        (unsigned long long)id, type);
 265                 RETURN(ERR_PTR(-EINVAL));
 266         }
 267
 268         len = sprintf(name, "%Ld", id);
 269         CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
 270                name);
 271         dchild = lookup_one_len(name, dparent, len);
 272         CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
 273                obd_mode_to_type(type), name, dchild,
 274                atomic_read(&dchild->d_count));
 275
 276         if (IS_ERR(dchild)) {
 277                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 278                 RETURN(dchild);
 279         }
 280
 281         RETURN(dchild);
 282 }
 283
 284 static struct file *filter_obj_open(struct obd_device *obddev,
 285                                     __u64 id, __u32 type)
 286 {
 287         struct super_block *sb = obddev->u.filter.fo_sb;
 288         struct obd_run_ctxt saved;
 289         char name[24];
 290         struct file *file;
 291         ENTRY;
 292
 293         if (!sb || !sb->s_dev) {
 294                 CERROR("fatal: device not initialized.\n");
 295                 RETURN(ERR_PTR(-ENXIO));
 296         }
 297
 298         if (!id) {
 299                 CERROR("fatal: invalid obdo %Lu\n", (unsigned long long)id);
 300                 RETURN(ERR_PTR(-ESTALE));
 301         }
 302
 303         if (!(type & S_IFMT)) {
 304                 CERROR("OBD %s, no type (%Ld), mode %o!\n", __FUNCTION__,
 305                        (unsigned long long)id, type);
 306                 RETURN(ERR_PTR(-EINVAL));
 307         }
 308
 309         filter_id(name, id, type);
 310         push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
 311         file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
 312         pop_ctxt(&saved);
 313
 314         CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
 315
 316         if (IS_ERR(file))
 317                 file = NULL;
 318         RETURN(file);
 319 }
 320
 321 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
 322 {
 323         struct filter_obd *filter = &obddev->u.filter;
 324
 325         return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
 326 }
 327
 328
 329 static struct inode *filter_inode_from_obj(struct obd_device *obddev,
 330                                            __u64 id, __u32 type)
 331 {
 332         struct dentry *dentry;
 333         struct inode *inode;
 334
 335         dentry = filter_fid2dentry(obddev, filter_parent(obddev, type),
 336                                    id, type);
 337         if (IS_ERR(dentry)) {
 338                 CERROR("%s: lookup failed: rc = %ld\n", __FUNCTION__,
 339                        PTR_ERR(dentry));
 340                 RETURN(NULL);
 341         }
 342
 343         lock_kernel();
 344         inode = iget(dentry->d_inode->i_sb, dentry->d_inode->i_ino);
 345         unlock_kernel();
 346         CDEBUG(D_INODE, "put child %p, count = %d\n", dentry,
 347                atomic_read(&dentry->d_count) - 1);
 348         dput(dentry);
 349         CDEBUG(D_INODE, "got inode %p (%ld), count = %d\n", inode, inode->i_ino,
 350                atomic_read(&inode->i_count));
 351         return inode;
 352 }
 353
 354 /* obd methods */
 355 static int filter_connect(struct obd_conn *conn, struct obd_device *obd)
 356 {
 357         int rc;
 358
 359         MOD_INC_USE_COUNT;
 360         rc = gen_connect(conn, obd);
 361
 362         if (rc)
 363                 MOD_DEC_USE_COUNT;
 364
 365         return rc;
 366 }
 367
 368 static int filter_disconnect(struct obd_conn *conn)
 369 {
 370         int rc;
 371
 372         rc = gen_disconnect(conn);
 373         if (!rc)
 374                 MOD_DEC_USE_COUNT;
 375
 376         /* XXX cleanup preallocated inodes */
 377         return rc;
 378 }
 379
 380 /* mount the file system (secretly) */
 381 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
 382 {
 383         struct obd_ioctl_data* data = buf;
 384         struct filter_obd *filter;
 385         struct vfsmount *mnt;
 386         int err = 0;
 387         ENTRY;
 388
 389         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 390                 RETURN(-EINVAL);
 391
 392         MOD_INC_USE_COUNT;
 393         mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
 394         err = PTR_ERR(mnt);
 395         if (IS_ERR(mnt))
 396                 GOTO(err_dec, err);
 397
 398         filter = &obddev->u.filter;;
 399         filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
 400         /* XXX is this even possible if do_kern_mount succeeded? */
 401         if (!filter->fo_sb)
 402                 GOTO(err_put, err = -ENODEV);
 403
 404         filter->fo_vfsmnt = mnt;
 405         filter->fo_fstype = strdup(data->ioc_inlbuf2);
 406
 407         OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
 408         filter->fo_ctxt.pwdmnt = mnt;
 409         filter->fo_ctxt.pwd = mnt->mnt_root;
 410         filter->fo_ctxt.fs = get_ds();
 411
 412         err = filter_prep(obddev);
 413         if (err)
 414                 GOTO(err_kfree, err);
 415         spin_lock_init(&filter->fo_lock);
 416
 417         RETURN(0);
 418
 419 err_kfree:
 420         kfree(filter->fo_fstype);
 421 err_put:
 422         unlock_kernel();
 423         mntput(filter->fo_vfsmnt);
 424         filter->fo_sb = 0;
 425         lock_kernel();
 426
 427 err_dec:
 428         MOD_DEC_USE_COUNT;
 429         return err;
 430 }
 431
 432
 433 static int filter_cleanup(struct obd_device * obddev)
 434 {
 435         struct super_block *sb;
 436         ENTRY;
 437
 438         if (!(obddev->obd_flags & OBD_SET_UP))
 439                 RETURN(0);
 440
 441         if (!list_empty(&obddev->obd_exports)) {
 442                 CERROR("still has clients!\n");
 443                 RETURN(-EBUSY);
 444         }
 445
 446         sb = obddev->u.filter.fo_sb;
 447         if (!obddev->u.filter.fo_sb)
 448                 RETURN(0);
 449
 450         filter_post(obddev);
 451
 452         shrink_dcache_parent(sb->s_root);
 453         unlock_kernel();
 454         mntput(obddev->u.filter.fo_vfsmnt);
 455         obddev->u.filter.fo_sb = 0;
 456         kfree(obddev->u.filter.fo_fstype);
 457
 458         lock_kernel();
 459
 460         MOD_DEC_USE_COUNT;
 461         RETURN(0);
 462 }
 463
 464
 465 static inline void filter_from_inode(struct obdo *oa, struct inode *inode)
 466 {
 467         int type = oa->o_mode & S_IFMT;
 468         ENTRY;
 469
 470         CDEBUG(D_INFO, "src inode %ld, dst obdo %ld valid 0x%08x\n",
 471                inode->i_ino, (long)oa->o_id, oa->o_valid);
 472         obdo_from_inode(oa, inode);
 473         oa->o_mode &= ~S_IFMT;
 474         oa->o_mode |= type;
 475
 476         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
 477                 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
 478                 CDEBUG(D_INODE, "copying device %x from inode to obdo\n",
 479                        rdev);
 480                 *((obd_rdev *)oa->o_inline) = rdev;
 481                 oa->o_obdflags |= OBD_FL_INLINEDATA;
 482                 oa->o_valid |= OBD_MD_FLINLINE;
 483         }
 484
 485 #if 0
 486         else if (filter_has_inline(inode)) {
 487                 CDEBUG(D_INFO, "copying inline from inode to obdo\n");
 488                 memcpy(oa->o_inline, inode->u.ext2_i.i_data,
 489                        MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
 490                 oa->o_obdflags |= OBD_FL_INLINEDATA;
 491                 oa->o_valid |= OBD_MD_FLINLINE;
 492         }
 493
 494         if (filter_has_obdmd(inode)) {
 495                 /* XXX this will change when we don't store the obdmd in data */
 496                 CDEBUG(D_INFO, "copying obdmd from inode to obdo\n");
 497                 memcpy(oa->o_obdmd, inode->u.ext2_i.i_data,
 498                        MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
 499                 oa->o_obdflags |= OBD_FL_OBDMDEXISTS;
 500                 oa->o_valid |= OBD_MD_FLOBDMD;
 501         }
 502 #endif
 503         EXIT;
 504 }
 505
 506 static int filter_getattr(struct obd_conn *conn, struct obdo *oa)
 507 {
 508         struct obd_device *obddev = gen_conn2obd(conn);
 509         struct dentry *dentry;
 510         ENTRY;
 511
 512         if (!gen_client(conn)) {
 513                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 514                 RETURN(-EINVAL);
 515         }
 516
 517         obddev = gen_conn2obd(conn);
 518         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 519                                    oa->o_id, oa->o_mode);
 520         if (IS_ERR(dentry))
 521                 RETURN(PTR_ERR(dentry));
 522
 523         oa->o_valid &= ~OBD_MD_FLID;
 524         filter_from_inode(oa, dentry->d_inode);
 525
 526         dput(dentry);
 527         RETURN(0);
 528 }
 529
 530 static int filter_setattr(struct obd_conn *conn, struct obdo *oa)
 531 {
 532         struct obd_run_ctxt saved;
 533         struct obd_device *obd = gen_conn2obd(conn);
 534         struct dentry *dentry;
 535         struct iattr iattr;
 536         struct inode *inode;
 537         int rc;
 538         ENTRY;
 539
 540         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 541                                    oa->o_id, oa->o_mode);
 542         if (IS_ERR(dentry))
 543                 RETURN(PTR_ERR(dentry));
 544
 545         inode = dentry->d_inode;
 546         iattr_from_obdo(&iattr, oa);
 547         iattr.ia_mode &= ~S_IFMT;
 548         iattr.ia_mode |= S_IFREG;
 549         lock_kernel();
 550         if (iattr.ia_mode & ATTR_SIZE)
 551                 down(&inode->i_sem);
 552         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 553         if (inode->i_op->setattr)
 554                 rc = inode->i_op->setattr(dentry, &iattr);
 555         else
 556                 rc = inode_setattr(inode, &iattr);
 557         pop_ctxt(&saved);
 558         if (iattr.ia_mode & ATTR_SIZE) {
 559                 up(&inode->i_sem);
 560                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 561                 obdo_from_inode(oa, inode);
 562         }
 563         unlock_kernel();
 564
 565         CDEBUG(D_INODE, "put dentry %p, count = %d\n", inode,
 566                atomic_read(&dentry->d_count) - 1);
 567         dput(dentry);
 568         RETURN(rc);
 569 }
 570
 571 static int filter_open(struct obd_conn *conn, struct obdo *oa)
 572 {
 573         struct obd_device *obd;
 574         struct dentry *dentry;
 575         /* ENTRY; */
 576
 577         if (!gen_client(conn)) {
 578                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 579                 RETURN(-EINVAL);
 580         }
 581
 582         obd = gen_conn2obd(conn);
 583         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 584                                    oa->o_id, oa->o_mode);
 585         if (IS_ERR(dentry))
 586                 RETURN(PTR_ERR(dentry));
 587
 588         return 0;
 589 } /* filter_open */
 590
 591 static int filter_close(struct obd_conn *conn, struct obdo *oa)
 592 {
 593         struct obd_device *obd;
 594         struct dentry *dentry;
 595         /* ENTRY; */
 596
 597         if (!gen_client(conn)) {
 598                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 599                 RETURN(-EINVAL);
 600         }
 601
 602         obd = gen_conn2obd(conn);
 603         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 604                                    oa->o_id, oa->o_mode);
 605         if (IS_ERR(dentry))
 606                 RETURN(PTR_ERR(dentry));
 607
 608         CDEBUG(D_INODE, "put dentry %p, count = %d\n", dentry,
 609                atomic_read(&dentry->d_count) - 1);
 610         dput(dentry);  /* for the close */
 611         CDEBUG(D_INODE, "put dentry %p, count = %d\n", dentry,
 612                atomic_read(&dentry->d_count) - 1);
 613         dput(dentry);  /* for this call */
 614         return 0;
 615 } /* filter_close */
 616
 617 static int filter_create(struct obd_conn* conn, struct obdo *oa)
 618 {
 619         char name[64];
 620         struct obd_run_ctxt saved;
 621         struct file *file;
 622         int mode;
 623         struct obd_device *obd = gen_conn2obd(conn);
 624         struct iattr;
 625         ENTRY;
 626
 627         if (!gen_client(conn)) {
 628                 CERROR("invalid client %Lx\n", conn->addr);
 629                 return -EINVAL;
 630         }
 631
 632         oa->o_id = filter_next_id(obd);
 633         if (!(oa->o_mode && S_IFMT)) {
 634                 CERROR("filter obd: no type!\n");
 635                 return -ENOENT;
 636         }
 637
 638         filter_id(name, oa->o_id, oa->o_mode);
 639         mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
 640         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 641         file = filp_open(name, O_RDONLY | O_CREAT, mode);
 642         pop_ctxt(&saved);
 643         if (IS_ERR(file)) {
 644                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(file));
 645                 return -ENOENT;
 646         }
 647         filp_close(file, 0);
 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                  OBD_MD_FLUID | OBD_MD_FLGID;
 653
 654         /* XXX Hmm, shouldn't we copy the fields into the obdo here? */
 655         return 0;
 656 }
 657
 658 static int filter_destroy(struct obd_conn *conn, struct obdo *oa)
 659 {
 660         struct obd_run_ctxt saved;
 661         struct obd_device *obd;
 662         struct obd_export *export;
 663         struct inode *inode;
 664         struct dentry *dir_dentry, *object_dentry;
 665         int rc;
 666         ENTRY;
 667
 668         if (!(export = gen_client(conn))) {
 669                 CERROR("invalid client %Lx\n", conn->addr);
 670                 RETURN(-EINVAL);
 671         }
 672
 673         CDEBUG(D_INODE, "destroying object %Ld\n",oa->o_id);
 674         obd = gen_conn2obd(conn);
 675
 676         dir_dentry = filter_parent(obd, oa->o_mode);
 677         down(&dir_dentry->d_inode->i_sem);
 678
 679         object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
 680                                           oa->o_mode);
 681         if (IS_ERR(object_dentry))
 682                 GOTO(out, rc = -ENOENT);
 683
 684         inode = object_dentry->d_inode;
 685         if (inode->i_nlink != 1) {
 686                 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
 687                 inode->i_nlink = 1;
 688         }
 689         inode->i_mode = S_IFREG;
 690
 691         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 692
 693         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 694         pop_ctxt(&saved);
 695         CDEBUG(D_INODE, "put child %p, count = %d\n", object_dentry,
 696                atomic_read(&object_dentry->d_count) - 1);
 697         dput(object_dentry);
 698
 699         EXIT;
 700 out:
 701         up(&dir_dentry->d_inode->i_sem);
 702         return rc;
 703 }
 704
 705 /* NB count and offset are used for punch, but not truncate */
 706 static int filter_truncate(struct obd_conn *conn, struct obdo *oa,
 707                            obd_size count, obd_off offset)
 708 {
 709         int error;
 710         ENTRY;
 711
 712         CDEBUG(D_INODE, "calling truncate for object #%Ld, valid = %x, "
 713                "o_size = %Ld\n", oa->o_id, oa->o_valid, oa->o_size);
 714         error = filter_setattr(conn, oa);
 715
 716         RETURN(error);
 717 }
 718
 719 static int filter_pgcache_brw(int cmd, struct obd_conn *conn, obd_count num_oa,
 720                                struct obdo **oa, obd_count *oa_bufs,
 721                                struct page **pages, obd_size *count,
 722                                obd_off *offset, obd_flag *flags, void *callback)
 723 {
 724         struct obd_run_ctxt      saved;
 725         struct super_block      *sb;
 726         int                      onum;          /* index to oas */
 727         int                      pnum;          /* index to pages (bufs) */
 728         unsigned long            retval;
 729         int                      error;
 730         struct file             *file;
 731         struct obd_device      *obd = gen_conn2obd(conn);
 732         ENTRY;
 733
 734         if (!gen_client(conn)) {
 735                 CDEBUG(D_IOCTL, "invalid client %Lx\n", conn->addr);
 736                 RETURN(-EINVAL);
 737         }
 738
 739         sb = obd->u.filter.fo_sb;
 740         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 741         pnum = 0; /* pnum indexes buf 0..num_pages */
 742         for (onum = 0; onum < num_oa; onum++) {
 743                 int pg;
 744
 745                 file = filter_obj_open(obd, oa[onum]->o_id, oa[onum]->o_mode);
 746                 if (IS_ERR(file))
 747                         GOTO(out, retval = PTR_ERR(file));
 748
 749                 /* count doubles as retval */
 750                 for (pg = 0; pg < oa_bufs[onum]; pg++) {
 751                         CDEBUG(D_INODE, "OP %d obdo no/pno: (%d,%d) (%ld,%ld) "
 752                                "off count (%Ld,%Ld)\n",
 753                                cmd, onum, pnum, file->f_dentry->d_inode->i_ino,
 754                                (unsigned long)offset[pnum] >> PAGE_CACHE_SHIFT,
 755                                (unsigned long long)offset[pnum],
 756                                (unsigned long long)count[pnum]);
 757                         if (cmd & OBD_BRW_WRITE) {
 758                                 loff_t off;
 759                                 char *buffer;
 760                                 off = offset[pnum];
 761                                 buffer = kmap(pages[pnum]);
 762                                 retval = file->f_op->write(file, buffer, count[pnum], &off);
 763                                 kunmap(pages[pnum]);
 764                                 CDEBUG(D_INODE, "retval %ld\n", retval);
 765                         } else {
 766                                 loff_t off = offset[pnum];
 767                                 char *buffer = kmap(pages[pnum]);
 768
 769                                 if (off >= file->f_dentry->d_inode->i_size) {
 770                                         memset(buffer, 0, count[pnum]);
 771                                         retval = count[pnum];
 772                                 } else {
 773                                         retval = file->f_op->read(file, buffer, count[pnum], &off);
 774                                 }
 775                                 kunmap(pages[pnum]);
 776
 777                                 if (retval != count[pnum]) {
 778                                         filp_close(file, 0);
 779                                         GOTO(out, retval = -EIO);
 780                                 }
 781                                 CDEBUG(D_INODE, "retval %ld\n", retval);
 782                         }
 783                         pnum++;
 784                 }
 785                 /* sizes and blocks are set by generic_file_write */
 786                 /* ctimes/mtimes will follow with a setattr call */
 787                 filp_close(file, 0);
 788         }
 789
 790         EXIT;
 791 out:
 792         pop_ctxt(&saved);
 793         error = (retval >= 0) ? 0 : retval;
 794         return error;
 795 }
 796
 797
 798 struct inode *ioobj_to_inode(struct obd_conn *conn, struct obd_ioobj *o)
 799 {
 800         struct obd_device *obd = gen_conn2obd(conn);
 801         struct super_block *sb = obd->u.filter.fo_sb;
 802         struct inode *inode = NULL;
 803         ENTRY;
 804
 805         if (!sb || !sb->s_dev) {
 806                 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
 807                 RETURN(NULL);
 808         }
 809
 810         if ( !o->ioo_id ) {
 811                 CDEBUG(D_INODE, "fatal: invalid obdo %lu\n", (long)o->ioo_id);
 812                 RETURN(NULL);
 813         }
 814
 815         inode = filter_inode_from_obj(obd, o->ioo_id, S_IFREG);
 816         if (!inode || inode->i_nlink == 0 || is_bad_inode(inode)) {
 817                 CERROR("from obdo - fatal: invalid inode %ld (%s).\n",
 818                        (long)o->ioo_id, inode ? inode->i_nlink ? "bad inode" :
 819                        "no links" : "NULL");
 820                 iput(inode);
 821                 RETURN(NULL);
 822         }
 823
 824         RETURN(inode);
 825 }
 826
 827 /*
 828  * Calculate the number of buffer credits needed to write multiple pages in
 829  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
 830  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
 831  *
 832  * See comment above ext3_writepage_trans_blocks for details.  We assume
 833  * no data journaling is being done, but it does allow for all of the pages
 834  * being non-contiguous.  If we are guaranteed contiguous pages we could
 835  * reduce the number of (d)indirect blocks a lot.
 836  *
 837  * With N blocks per page and P pages, for each inode we have at most:
 838  * N*P indirect
 839  * min(N*P, blocksize/4 + 1) dindirect blocks
 840  * 1 tindirect
 841  *
 842  * For the entire filesystem, we have at most:
 843  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 844  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 845  * 1 inode block
 846  * 1 superblock
 847  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 848  */
 849 static int ext3_credits_needed(struct super_block *sb, int objcount,
 850                                struct obd_ioobj *obj)
 851 {
 852         struct obd_ioobj *o = obj;
 853         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 854         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 855         int nbitmaps = 0;
 856         int ngdblocks = 0;
 857         int needed = objcount + 1;
 858         int i;
 859
 860         for (i = 0; i < objcount; i++, o++) {
 861                 int nblocks = o->ioo_bufcnt * blockpp;
 862                 int ndindirect = min(nblocks, addrpp + 1);
 863                 int nindir = nblocks + ndindirect + 1;
 864
 865                 nbitmaps += nindir + nblocks;
 866                 ngdblocks += nindir + nblocks;
 867
 868                 needed += nindir;
 869         }
 870
 871         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 872                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 873         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 874                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 875
 876         needed += nbitmaps + ngdblocks;
 877
 878 #ifdef CONFIG_QUOTA
 879         /* We assume that there will be 1 bit set in s_dquot.flags for each
 880          * quota file that is active.  This is at least true for now.
 881          */
 882         needed += hweight32(sb_any_quota_enabled(sb)) *
 883                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 884 #endif
 885
 886         return needed;
 887 }
 888
 889 /* We have to start a huge journal transaction here to hold all of the
 890  * metadata for the pages being written here.  This is necessitated by
 891  * the fact that we do lots of prepare_write operations before we do
 892  * any of the matching commit_write operations, so even if we split
 893  * up to use "smaller" transactions none of them could complete until
 894  * all of them were opened.  By having a single journal transaction,
 895  * we eliminate duplicate reservations for common blocks like the
 896  * superblock and group descriptors or bitmaps.
 897  *
 898  * We will start the transaction here, but each prepare_write will
 899  * add a refcount to the transaction, and each commit_write will
 900  * remove a refcount.  The transaction will be closed when all of
 901  * the pages have been written.
 902  */
 903 static void *ext3_filter_journal_start(struct filter_obd *filter,
 904                                        int objcount, struct obd_ioobj *obj,
 905                                        int niocount, struct niobuf_remote *nb)
 906 {
 907         journal_t *journal = NULL;
 908         handle_t *handle = NULL;
 909         int needed;
 910
 911         /* Assumes ext3 and extN have same sb_info layout, but avoids issues
 912          * with having extN built properly before filterobd for now.
 913          */
 914         journal = EXT3_SB(filter->fo_sb)->s_journal;
 915         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
 916
 917         /* The number of blocks we could _possibly_ dirty can very large.
 918          * We reduce our request if it is absurd (and we couldn't get that
 919          * many credits for a single handle anyways).
 920          *
 921          * At some point we have to limit the size of I/Os sent at one time,
 922          * increase the size of the journal, or we have to calculate the
 923          * actual journal requirements more carefully by checking all of
 924          * the blocks instead of being maximally pessimistic.  It remains to
 925          * be seen if this is a real problem or not.
 926          */
 927         if (needed > journal->j_max_transaction_buffers) {
 928                 CERROR("want too many journal credits (%d) using %d instead\n",
 929                        needed, journal->j_max_transaction_buffers);
 930                 needed = journal->j_max_transaction_buffers;
 931         }
 932
 933         handle = journal_start(journal, needed);
 934         if (IS_ERR(handle))
 935                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 936                        PTR_ERR(handle));
 937
 938         return(handle);
 939 }
 940
 941 static void *filter_journal_start(void **journal_save,
 942                                   struct filter_obd *filter,
 943                                   int objcount, struct obd_ioobj *obj,
 944                                   int niocount, struct niobuf_remote *nb)
 945 {
 946         void *handle = NULL;
 947
 948         /* This may not be necessary - we probably never have a
 949          * transaction started when we enter here, so we can
 950          * remove the saving of the journal state entirely.
 951          * For now leave it in just to see if it ever happens.
 952          */
 953         *journal_save = current->journal_info;
 954         if (*journal_save) {
 955                 CERROR("Already have handle %p???\n", *journal_save);
 956                 LBUG();
 957                 current->journal_info = NULL;
 958         }
 959
 960         if (!strcmp(filter->fo_fstype, "ext3") ||
 961             !strcmp(filter->fo_fstype, "extN"))
 962                 handle = ext3_filter_journal_start(filter, objcount, obj,
 963                                                    niocount, nb);
 964         return handle;
 965 }
 966
 967 static int ext3_filter_journal_stop(void *handle)
 968 {
 969         int rc;
 970
 971         /* We got a refcount on the handle for each call to prepare_write,
 972          * so we can drop the "parent" handle here to avoid the need for
 973          * osc to call back into filterobd to close the handle.  The
 974          * remaining references will be dropped in commit_write.
 975          */
 976         rc = journal_stop((handle_t *)handle);
 977
 978         return rc;
 979 }
 980
 981 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
 982                                void *handle)
 983 {
 984         int rc = 0;
 985
 986         if (!strcmp(filter->fo_fstype, "ext3") ||
 987             !strcmp(filter->fo_fstype, "extN"))
 988                 rc = ext3_filter_journal_stop(handle);
 989
 990         if (rc)
 991                 CERROR("error on journal stop: rc = %d\n", rc);
 992
 993         current->journal_info = journal_save;
 994
 995         return rc;
 996 }
 997
 998 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
 999                                    struct niobuf_local *lnb)
1000 {
1001         struct address_space *mapping = inode->i_mapping;
1002         struct page *page;
1003         int rc;
1004
1005         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1006         page = grab_cache_page_nowait(mapping, index); /* locked page */
1007
1008         /* This page is currently locked, so get a temporary page instead */
1009         if (!page) {
1010                 unsigned long addr;
1011                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1012                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1013                 if (!addr) {
1014                         CERROR("no memory for a temp page\n");
1015                         LBUG();
1016                         GOTO(err, rc = -ENOMEM);
1017                 }
1018                 page = virt_to_page(addr);
1019                 kmap(page);
1020                 page->index = index;
1021                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1022         } else if (!IS_ERR(page)) {
1023                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1024                  * a no-op for most filesystems, because we write the whole
1025                  * page.  For partial-page I/O this will read in the page.
1026                  */
1027                 kmap(page);
1028                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1029                 if (rc) {
1030                         CERROR("page index %lu, rc = %d\n", index, rc);
1031                         if (rc != -ENOSPC)
1032                                 LBUG();
1033                         GOTO(err_unlock, rc);
1034                 }
1035                 /* XXX not sure if we need this if we are overwriting page */
1036                 if (PageError(page)) {
1037                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1038                         LBUG();
1039                         GOTO(err_unlock, rc = -EIO);
1040                 }
1041         }
1042         return page;
1043
1044 err_unlock:
1045         unlock_page(page);
1046         lustre_put_page(page);
1047 err:
1048         return ERR_PTR(rc);
1049 }
1050
1051 /*
1052  * We need to balance prepare_write() calls with commit_write() calls.
1053  * If the page has been prepared, but we have no data for it, we don't
1054  * want to overwrite valid data on disk, but we still need to zero out
1055  * data for space which was newly allocated.  Like part of what happens
1056  * in __block_prepare_write() for newly allocated blocks.
1057  *
1058  * XXX currently __block_prepare_write() creates buffers for all the
1059  *     pages, and the filesystems mark these buffers as BH_New if they
1060  *     were newly allocated from disk. We use the BH_New flag similarly.
1061  */
1062 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1063                                int err)
1064 {
1065         if (err) {
1066                 unsigned block_start, block_end;
1067                 struct buffer_head *bh, *head = page->buffers;
1068                 unsigned blocksize = head->b_size;
1069                 void *addr = page_address(page);
1070
1071                 /* Currently one buffer per page, but in the future... */
1072                 for (bh = head, block_start = 0; bh != head || !block_start;
1073                      block_start = block_end, bh = bh->b_this_page) {
1074                         block_end = block_start + blocksize;
1075                         if (buffer_new(bh))
1076                                 memset(addr + block_start, 0, blocksize);
1077                 }
1078         }
1079
1080         return lustre_commit_write(page, from, to);
1081 }
1082
1083 static int filter_preprw(int cmd, struct obd_conn *conn,
1084                          int objcount, struct obd_ioobj *obj,
1085                          int niocount, struct niobuf_remote *nb,
1086                          struct niobuf_local *res, void **desc_private)
1087 {
1088         struct obd_run_ctxt saved;
1089         struct obd_device *obd;
1090         struct obd_ioobj *o = obj;
1091         struct niobuf_remote *b = nb;
1092         struct niobuf_local *r = res;
1093         void *journal_save = NULL;
1094         int rc = 0;
1095         int i;
1096         ENTRY;
1097
1098         memset(res, 0, sizeof(*res) * niocount);
1099         obd = gen_conn2obd(conn);
1100
1101         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1102
1103         if (cmd & OBD_BRW_WRITE) {
1104                 *desc_private = filter_journal_start(&journal_save,
1105                                                      &obd->u.filter,
1106                                                      objcount, obj, niocount,
1107                                                      nb);
1108                 if (IS_ERR(*desc_private))
1109                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1110         }
1111
1112         for (i = 0; i < objcount; i++, o++) {
1113                 struct dentry *dentry;
1114                 struct inode *inode;
1115                 int j;
1116
1117                 dentry = filter_fid2dentry(obd,
1118                                            filter_parent(obd, S_IFREG),
1119                                            o->ioo_id, S_IFREG);
1120                 if (IS_ERR(dentry))
1121                         GOTO(out_clean, rc = PTR_ERR(dentry));
1122                 inode = dentry->d_inode;
1123                 if (!inode) {
1124                         CERROR("trying to BRW to non-existent file %Ld\n",
1125                                (unsigned long long)o->ioo_id);
1126                         dput(dentry);
1127                         GOTO(out_clean, rc = -ENOENT);
1128                 }
1129
1130                 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1131                         unsigned long index = b->offset >> PAGE_SHIFT;
1132                         struct page *page;
1133
1134                         if (j == 0)
1135                                 r->dentry = dentry;
1136                         else
1137                                 r->dentry = dget(dentry);
1138
1139                         if (cmd & OBD_BRW_WRITE)
1140                                 page = filter_get_page_write(inode, index, r);
1141                         else
1142                                 page = lustre_get_page_read(inode, index);
1143
1144                         if (IS_ERR(page)) {
1145                                 dput(dentry);
1146                                 GOTO(out_clean, rc = PTR_ERR(page));
1147                         }
1148
1149                         r->addr = page_address(page);
1150                         r->offset = b->offset;
1151                         r->page = page;
1152                         r->len = PAGE_SIZE;
1153                 }
1154         }
1155
1156 out_stop:
1157         if (cmd & OBD_BRW_WRITE) {
1158                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1159                                               *desc_private);
1160                 if (!rc)
1161                         rc = err;
1162         }
1163 out_ctxt:
1164         pop_ctxt(&saved);
1165         RETURN(rc);
1166 out_clean:
1167         while (r-- > res) {
1168                 dput(r->dentry);
1169                 if (cmd & OBD_BRW_WRITE)
1170                         filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1171                 else
1172                         lustre_put_page(r->page);
1173         }
1174         goto out_stop;
1175 }
1176
1177 static int filter_write_locked_page(struct niobuf_local *lnb)
1178 {
1179         struct page *lpage;
1180         int rc;
1181
1182         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1183         if (IS_ERR(lpage)) {
1184                 /* It is highly unlikely that we would ever get an error here.
1185                  * The page we want to get was previously locked, so it had to
1186                  * have already allocated the space, and we were just writing
1187                  * over the same data, so there would be no hole in the file.
1188                  *
1189                  * XXX: possibility of a race with truncate could exist, need
1190                  *      to check that.  There are no guarantees w.r.t.
1191                  *      write order even on a local filesystem, although the
1192                  *      normal response would be to return the number of bytes
1193                  *      successfully written and leave the rest to the app.
1194                  */
1195                 rc = PTR_ERR(lpage);
1196                 CERROR("error getting locked page index %ld: rc = %d\n",
1197                        lnb->page->index, rc);
1198                 GOTO(out, rc);
1199         }
1200
1201         memcpy(page_address(lpage), kmap(lnb->page), PAGE_SIZE);
1202         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1203         if (rc)
1204                 CERROR("error committing locked page %ld: rc = %d\n",
1205                        lnb->page->index, rc);
1206 out:
1207         kunmap(lnb->page);
1208         __free_pages(lnb->page, 0);
1209         dput(lnb->dentry);
1210
1211         return rc;
1212 }
1213
1214 static int filter_commitrw(int cmd, struct obd_conn *conn,
1215                            int objcount, struct obd_ioobj *obj,
1216                            int niocount, struct niobuf_local *res,
1217                            void *private)
1218 {
1219         struct obd_run_ctxt saved;
1220         struct obd_ioobj *o = obj;
1221         struct niobuf_local *r = res;
1222         struct obd_device *obd = gen_conn2obd(conn);
1223         void *journal_save;
1224         int found_locked = 0;
1225         int rc = 0;
1226         int i;
1227         ENTRY;
1228
1229         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1230         journal_save = current->journal_info;
1231         if (journal_save)
1232                 CERROR("Existing handle %p???\n", journal_save);
1233         current->journal_info = private;
1234         for (i = 0; i < objcount; i++, obj++) {
1235                 int j;
1236                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1237                         struct page *page = r->page;
1238
1239                         if (!page)
1240                                 LBUG();
1241
1242                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1243                                 found_locked = 1;
1244                                 continue;
1245                         }
1246
1247                         if (cmd & OBD_BRW_WRITE) {
1248                                 int err = filter_commit_write(page, 0,
1249                                                               PAGE_SIZE, 0);
1250
1251                                 if (!rc)
1252                                         rc = err;
1253                         } else
1254                                 lustre_put_page(page);
1255
1256                         CDEBUG(D_INODE,
1257                                "put inode %p (%ld), count = %d, nlink = %d\n",
1258                                page->mapping->host,
1259                                page->mapping->host->i_ino,
1260                                atomic_read(&page->mapping->host->i_count) - 1,
1261                                page->mapping->host->i_nlink);
1262                         dput(r->dentry);
1263                 }
1264         }
1265         if (!found_locked)
1266                 goto out_ctxt;
1267
1268         for (i = 0; i < objcount; i++, obj++) {
1269                 int j;
1270                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1271                         int err;
1272                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1273                                 continue;
1274
1275                         err = filter_write_locked_page(r);
1276                         if (!rc)
1277                                 rc = err;
1278                 }
1279         }
1280
1281 out_ctxt:
1282         current->journal_info = journal_save;
1283         pop_ctxt(&saved);
1284         RETURN(0);
1285 }
1286
1287 static int filter_statfs(struct obd_conn *conn, struct statfs * statfs)
1288 {
1289         struct super_block *sb;
1290         int err;
1291         ENTRY;
1292
1293         if (!gen_client(conn)) {
1294                 CDEBUG(D_IOCTL, "invalid client %Lx\n", conn->addr);
1295                 RETURN(-EINVAL);
1296         }
1297
1298         sb = gen_conn2obd(conn)->u.filter.fo_sb;
1299
1300         err = sb->s_op->statfs(sb, statfs);
1301         RETURN(err);
1302 } /* filter_statfs */
1303
1304
1305 static int filter_get_info(struct obd_conn *conn, obd_count keylen,
1306                            void *key, obd_count *vallen, void **val)
1307 {
1308         struct obd_device *obd;
1309         struct obd_export * export;
1310         ENTRY;
1311
1312         if (!(export = gen_client(conn))) {
1313                 CDEBUG(D_IOCTL, "invalid client %Lx\n", conn->addr);
1314                 RETURN(-EINVAL);
1315         }
1316
1317         obd = gen_conn2obd(conn);
1318
1319         if ( keylen == strlen("blocksize") &&
1320              memcmp(key, "blocksize", keylen) == 0 ) {
1321                 *vallen = sizeof(long);
1322                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1323                 RETURN(0);
1324         }
1325
1326         if ( keylen == strlen("blocksize_bits") &&
1327              memcmp(key, "blocksize_bits", keylen) == 0 ){
1328                 *vallen = sizeof(long);
1329                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1330                 RETURN(0);
1331         }
1332
1333         if ( keylen == strlen("root_ino") &&
1334              memcmp(key, "root_ino", keylen) == 0 ){
1335                 *vallen = sizeof(long);
1336                 *val = (void *)(long)FILTER_ROOTINO;
1337                 RETURN(0);
1338         }
1339
1340         CDEBUG(D_IOCTL, "invalid key\n");
1341         RETURN(-EINVAL);
1342 }
1343
1344
1345 static struct obd_ops filter_obd_ops = {
1346         o_get_info:    filter_get_info,
1347         o_setup:       filter_setup,
1348         o_cleanup:     filter_cleanup,
1349         o_connect:     filter_connect,
1350         o_disconnect:  filter_disconnect,
1351         o_statfs:      filter_statfs,
1352         o_getattr:     filter_getattr,
1353         o_create:      filter_create,
1354         o_setattr:     filter_setattr,
1355         o_destroy:     filter_destroy,
1356         o_open:        filter_open,
1357         o_close:       filter_close,
1358         o_brw:         filter_pgcache_brw,
1359         o_punch:       filter_truncate,
1360         o_preprw:      filter_preprw,
1361         o_commitrw:    filter_commitrw
1362 #if 0
1363         o_preallocate: filter_preallocate_inodes,
1364         o_migrate:     filter_migrate,
1365         o_copy:        gen_copy_data,
1366         o_iterate:     filter_iterate
1367 #endif
1368 };
1369
1370
1371 static int __init obdfilter_init(void)
1372 {
1373         printk(KERN_INFO "Filtering OBD driver  v0.001, braam@clusterfs.com\n");
1374         return obd_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1375 }
1376
1377 static void __exit obdfilter_exit(void)
1378 {
1379         obd_unregister_type(OBD_FILTER_DEVICENAME);
1380 }
1381
1382 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
1383 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1384 MODULE_LICENSE("GPL");
1385
1386 module_init(obdfilter_init);
1387 module_exit(obdfilter_exit);