lustre/obdfilter/filter.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/fs/filter/filter.c
   5  *
   6  * Copyright (C) 2001  Cluster File Systems, Inc.
   7  *
   8  * This code is issued under the GNU General Public License.
   9  * See the file COPYING in this distribution
  10  *
  11  * by Peter Braam <braam@clusterfs.com>
  12  */
  13
  14 #define EXPORT_SYMTAB
  15 #define DEBUG_SUBSYSTEM S_FILTER
  16
  17 #include <linux/module.h>
  18 #include <linux/lustre_dlm.h>
  19 #include <linux/obd_filter.h>
  20 #include <linux/ext3_jbd.h>
  21 #include <linux/quotaops.h>
  22
  23 extern struct obd_device obd_dev[MAX_OBD_DEVICES];
  24 long filter_memory;
  25
  26 #define FILTER_ROOTINO 2
  27
  28 #define S_SHIFT 12
  29 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
  30         [0]                     NULL,
  31         [S_IFREG >> S_SHIFT]    "R",
  32         [S_IFDIR >> S_SHIFT]    "D",
  33         [S_IFCHR >> S_SHIFT]    "C",
  34         [S_IFBLK >> S_SHIFT]    "B",
  35         [S_IFIFO >> S_SHIFT]    "F",
  36         [S_IFSOCK >> S_SHIFT]   "S",
  37         [S_IFLNK >> S_SHIFT]    "L"
  38 };
  39
  40 static inline const char *obd_mode_to_type(int mode)
  41 {
  42         return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
  43 }
  44
  45 /* write the pathname into the string */
  46 static int filter_id(char *buf, obd_id id, obd_mode mode)
  47 {
  48         return sprintf(buf, "O/%s/%Ld", obd_mode_to_type(mode),
  49                        (unsigned long long)id);
  50 }
  51
  52 /* setup the object store with correct subdirectories */
  53 static int filter_prep(struct obd_device *obddev)
  54 {
  55         struct obd_run_ctxt saved;
  56         struct filter_obd *filter = &obddev->u.filter;
  57         struct dentry *dentry;
  58         struct file *file;
  59         struct inode *inode;
  60         loff_t off;
  61         int rc = 0;
  62         char rootid[128];
  63         __u64 lastino = 2;
  64         int mode = 0;
  65
  66         push_ctxt(&saved, &filter->fo_ctxt);
  67         dentry = simple_mkdir(current->fs->pwd, "O", 0700);
  68         CDEBUG(D_INODE, "got/created O: %p\n", dentry);
  69         if (IS_ERR(dentry)) {
  70                 rc = PTR_ERR(dentry);
  71                 CERROR("cannot open/create O: rc = %d\n", rc);
  72                 GOTO(out, rc);
  73         }
  74         filter->fo_dentry_O = dentry;
  75         dentry = simple_mkdir(current->fs->pwd, "P", 0700);
  76         CDEBUG(D_INODE, "got/created P: %p\n", dentry);
  77         if (IS_ERR(dentry)) {
  78                 rc = PTR_ERR(dentry);
  79                 CERROR("cannot open/create P: rc = %d\n", rc);
  80                 GOTO(out_O, rc);
  81         }
  82         CDEBUG(D_INODE, "putting P: %p, count = %d\n", dentry,
  83                atomic_read(&dentry->d_count) - 1);
  84         dput(dentry);
  85         dentry = simple_mkdir(current->fs->pwd, "D", 0700);
  86         CDEBUG(D_INODE, "got/created D: %p\n", dentry);
  87         if (IS_ERR(dentry)) {
  88                 rc = PTR_ERR(dentry);
  89                 CERROR("cannot open/create D: rc = %d\n", rc);
  90                 GOTO(out_O, rc);
  91         }
  92         CDEBUG(D_INODE, "putting D: %p, count = %d\n", dentry,
  93                atomic_read(&dentry->d_count) - 1);
  94         dput(dentry);
  95
  96         /*
  97          * Create directories and/or get dentries for each object type.
  98          * This saves us from having to do multiple lookups for each one.
  99          */
 100         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 101                 char *type = obd_type_by_mode[mode];
 102
 103                 if (!type) {
 104                         filter->fo_dentry_O_mode[mode] = NULL;
 105                         continue;
 106                 }
 107                 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
 108                 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
 109                 if (IS_ERR(dentry)) {
 110                         rc = PTR_ERR(dentry);
 111                         CERROR("cannot create O/%s: rc = %d\n", type, rc);
 112                         GOTO(out_O_mode, rc);
 113                 }
 114                 filter->fo_dentry_O_mode[mode] = dentry;
 115         }
 116
 117         filter_id(rootid, FILTER_ROOTINO, S_IFDIR);
 118         file = filp_open(rootid, O_RDWR | O_CREAT, 0755);
 119         if (IS_ERR(file)) {
 120                 rc = PTR_ERR(file);
 121                 CERROR("OBD filter: cannot open/create root %s: rc = %d\n",
 122                        rootid, rc);
 123                 GOTO(out_O_mode, rc);
 124         }
 125         filp_close(file, 0);
 126
 127         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 128         if ( !file || IS_ERR(file) ) {
 129                 rc = PTR_ERR(file);
 130                 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
 131                        "D/status", rc);
 132                 GOTO(out_O_mode, rc);
 133         }
 134
 135         /* steal operations */
 136         inode = file->f_dentry->d_inode;
 137         filter->fo_fop = file->f_op;
 138         filter->fo_iop = inode->i_op;
 139         filter->fo_aops = inode->i_mapping->a_ops;
 140
 141         off = 0;
 142         if (inode->i_size == 0) {
 143                 ssize_t retval = file->f_op->write(file, (char *)&lastino,
 144                                                    sizeof(lastino), &off);
 145                 if (retval != sizeof(lastino)) {
 146                         CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
 147                         filp_close(file, 0);
 148                         GOTO(out_O_mode, rc = -EIO);
 149                 }
 150         } else {
 151                 ssize_t retval = file->f_op->read(file, (char *)&lastino,
 152                                                   sizeof(lastino), &off);
 153                 if (retval != sizeof(lastino)) {
 154                         CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
 155                         filp_close(file, 0);
 156                         GOTO(out_O_mode, rc = -EIO);
 157                 }
 158         }
 159         filter->fo_lastino = lastino;
 160         filp_close(file, 0);
 161
 162         rc = 0;
 163  out:
 164         pop_ctxt(&saved);
 165
 166         return(rc);
 167
 168 out_O_mode:
 169         while (mode-- > 0) {
 170                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 171                 if (dentry) {
 172                         CDEBUG(D_INODE, "putting O/%s: %p, count = %d\n",
 173                                obd_type_by_mode[mode], dentry,
 174                                atomic_read(&dentry->d_count) - 1);
 175                         dput(dentry);
 176                         filter->fo_dentry_O_mode[mode] = NULL;
 177                 }
 178         }
 179 out_O:
 180         CDEBUG(D_INODE, "putting O: %p, count = %d\n", filter->fo_dentry_O,
 181                atomic_read(&filter->fo_dentry_O->d_count) - 1);
 182         dput(filter->fo_dentry_O);
 183         filter->fo_dentry_O = NULL;
 184         goto out;
 185 }
 186
 187 /* cleanup the filter: write last used object id to status file */
 188 static void filter_post(struct obd_device *obddev)
 189 {
 190         struct obd_run_ctxt saved;
 191         struct filter_obd *filter = &obddev->u.filter;
 192         long rc;
 193         struct file *file;
 194         loff_t off = 0;
 195         int mode;
 196
 197         push_ctxt(&saved, &filter->fo_ctxt);
 198         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 199         if (IS_ERR(file)) {
 200                 CERROR("OBD filter: cannot create status file\n");
 201                 goto out;
 202         }
 203         rc = file->f_op->write(file, (char *)&filter->fo_lastino,
 204                        sizeof(filter->fo_lastino), &off);
 205         if (rc != sizeof(filter->fo_lastino))
 206                 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
 207
 208         rc = filp_close(file, NULL);
 209         if (rc)
 210                 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
 211
 212         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 213                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 214                 if (dentry) {
 215                         CDEBUG(D_INODE, "putting O/%s: %p, count = %d\n",
 216                                obd_type_by_mode[mode], dentry,
 217                                atomic_read(&dentry->d_count) - 1);
 218                         dput(dentry);
 219                         filter->fo_dentry_O_mode[mode] = NULL;
 220                 }
 221         }
 222         CDEBUG(D_INODE, "putting O: %p, count = %d\n", filter->fo_dentry_O,
 223                atomic_read(&filter->fo_dentry_O->d_count) - 1);
 224         dput(filter->fo_dentry_O);
 225 out:
 226         pop_ctxt(&saved);
 227 }
 228
 229
 230 static __u64 filter_next_id(struct obd_device *obddev)
 231 {
 232         __u64 id;
 233         spin_lock(&obddev->u.filter.fo_lock);
 234         obddev->u.filter.fo_lastino++;
 235         id =    obddev->u.filter.fo_lastino;
 236         spin_unlock(&obddev->u.filter.fo_lock);
 237         return id;
 238 }
 239
 240 /* how to get files, dentries, inodes from object id's */
 241 /* parent i_sem is already held if needed for exclusivity */
 242 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
 243                                         struct dentry *dparent,
 244                                         __u64 id, __u32 type)
 245 {
 246         struct super_block *sb = obddev->u.filter.fo_sb;
 247         struct dentry *dchild;
 248         char name[32];
 249         int len;
 250         ENTRY;
 251
 252         if (!sb || !sb->s_dev) {
 253                 CERROR("fatal: device not initialized.\n");
 254                 RETURN(ERR_PTR(-ENXIO));
 255         }
 256
 257         if (id == 0) {
 258                 CERROR("fatal: invalid object #0\n");
 259                 LBUG();
 260                 RETURN(ERR_PTR(-ESTALE));
 261         }
 262
 263         if (!(type & S_IFMT)) {
 264                 CERROR("OBD %s, object %Lu has bad type: %o\n", __FUNCTION__,
 265                        (unsigned long long)id, type);
 266                 RETURN(ERR_PTR(-EINVAL));
 267         }
 268
 269         len = sprintf(name, "%Ld", id);
 270         CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
 271                name);
 272         dchild = lookup_one_len(name, dparent, len);
 273         CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
 274                obd_mode_to_type(type), name, dchild,
 275                atomic_read(&dchild->d_count));
 276
 277         if (IS_ERR(dchild)) {
 278                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 279                 RETURN(dchild);
 280         }
 281
 282         RETURN(dchild);
 283 }
 284
 285 static struct file *filter_obj_open(struct obd_device *obddev,
 286                                     __u64 id, __u32 type)
 287 {
 288         struct super_block *sb = obddev->u.filter.fo_sb;
 289         struct obd_run_ctxt saved;
 290         char name[24];
 291         struct file *file;
 292         ENTRY;
 293
 294         if (!sb || !sb->s_dev) {
 295                 CERROR("fatal: device not initialized.\n");
 296                 RETURN(ERR_PTR(-ENXIO));
 297         }
 298
 299         if (!id) {
 300                 CERROR("fatal: invalid obdo %Lu\n", (unsigned long long)id);
 301                 RETURN(ERR_PTR(-ESTALE));
 302         }
 303
 304         if (!(type & S_IFMT)) {
 305                 CERROR("OBD %s, no type (%Ld), mode %o!\n", __FUNCTION__,
 306                        (unsigned long long)id, type);
 307                 RETURN(ERR_PTR(-EINVAL));
 308         }
 309
 310         filter_id(name, id, type);
 311         push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
 312         file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
 313         pop_ctxt(&saved);
 314
 315         CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
 316
 317         if (IS_ERR(file))
 318                 file = NULL;
 319         RETURN(file);
 320 }
 321
 322 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
 323 {
 324         struct filter_obd *filter = &obddev->u.filter;
 325
 326         return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
 327 }
 328
 329
 330 static struct inode *filter_inode_from_obj(struct obd_device *obddev,
 331                                            __u64 id, __u32 type)
 332 {
 333         struct dentry *dentry;
 334         struct inode *inode;
 335
 336         dentry = filter_fid2dentry(obddev, filter_parent(obddev, type),
 337                                    id, type);
 338         if (IS_ERR(dentry)) {
 339                 CERROR("%s: lookup failed: rc = %ld\n", __FUNCTION__,
 340                        PTR_ERR(dentry));
 341                 RETURN(NULL);
 342         }
 343
 344         lock_kernel();
 345         inode = iget(dentry->d_inode->i_sb, dentry->d_inode->i_ino);
 346         unlock_kernel();
 347         CDEBUG(D_INODE, "put child %p, count = %d\n", dentry,
 348                atomic_read(&dentry->d_count) - 1);
 349         dput(dentry);
 350         CDEBUG(D_INODE, "got inode %p (%ld), count = %d\n", inode, inode->i_ino,
 351                atomic_read(&inode->i_count));
 352         return inode;
 353 }
 354
 355 /* obd methods */
 356 static int filter_connect(struct obd_conn *conn, struct obd_device *obd)
 357 {
 358         int rc;
 359
 360         MOD_INC_USE_COUNT;
 361         rc = gen_connect(conn, obd);
 362
 363         if (rc)
 364                 MOD_DEC_USE_COUNT;
 365
 366         return rc;
 367 }
 368
 369 static int filter_disconnect(struct obd_conn *conn)
 370 {
 371         int rc;
 372
 373         rc = gen_disconnect(conn);
 374         if (!rc)
 375                 MOD_DEC_USE_COUNT;
 376
 377         /* XXX cleanup preallocated inodes */
 378         return rc;
 379 }
 380
 381 /* mount the file system (secretly) */
 382 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
 383 {
 384         struct obd_ioctl_data* data = buf;
 385         struct filter_obd *filter;
 386         struct vfsmount *mnt;
 387         int err = 0;
 388         ENTRY;
 389
 390         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 391                 RETURN(-EINVAL);
 392
 393         MOD_INC_USE_COUNT;
 394         mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
 395         err = PTR_ERR(mnt);
 396         if (IS_ERR(mnt))
 397                 GOTO(err_dec, err);
 398
 399         filter = &obddev->u.filter;;
 400         filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
 401         /* XXX is this even possible if do_kern_mount succeeded? */
 402         if (!filter->fo_sb)
 403                 GOTO(err_put, err = -ENODEV);
 404
 405         filter->fo_vfsmnt = mnt;
 406         filter->fo_fstype = strdup(data->ioc_inlbuf2);
 407
 408         OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
 409         filter->fo_ctxt.pwdmnt = mnt;
 410         filter->fo_ctxt.pwd = mnt->mnt_root;
 411         filter->fo_ctxt.fs = get_ds();
 412
 413         err = filter_prep(obddev);
 414         if (err)
 415                 GOTO(err_kfree, err);
 416         spin_lock_init(&filter->fo_lock);
 417
 418         obddev->obd_namespace =
 419                 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
 420         if (obddev->obd_namespace == NULL)
 421                 LBUG();
 422
 423         RETURN(0);
 424
 425 err_kfree:
 426         kfree(filter->fo_fstype);
 427 err_put:
 428         unlock_kernel();
 429         mntput(filter->fo_vfsmnt);
 430         filter->fo_sb = 0;
 431         lock_kernel();
 432
 433 err_dec:
 434         MOD_DEC_USE_COUNT;
 435         return err;
 436 }
 437
 438
 439 static int filter_cleanup(struct obd_device * obddev)
 440 {
 441         struct super_block *sb;
 442         ENTRY;
 443
 444         if (!(obddev->obd_flags & OBD_SET_UP))
 445                 RETURN(0);
 446
 447         if (!list_empty(&obddev->obd_exports)) {
 448                 CERROR("still has clients!\n");
 449                 RETURN(-EBUSY);
 450         }
 451
 452         ldlm_namespace_free(obddev->obd_namespace);
 453
 454         sb = obddev->u.filter.fo_sb;
 455         if (!obddev->u.filter.fo_sb)
 456                 RETURN(0);
 457
 458         filter_post(obddev);
 459
 460         shrink_dcache_parent(sb->s_root);
 461         unlock_kernel();
 462         mntput(obddev->u.filter.fo_vfsmnt);
 463         obddev->u.filter.fo_sb = 0;
 464         kfree(obddev->u.filter.fo_fstype);
 465
 466         lock_kernel();
 467
 468         MOD_DEC_USE_COUNT;
 469         RETURN(0);
 470 }
 471
 472
 473 static inline void filter_from_inode(struct obdo *oa, struct inode *inode)
 474 {
 475         int type = oa->o_mode & S_IFMT;
 476         ENTRY;
 477
 478         CDEBUG(D_INFO, "src inode %ld, dst obdo %ld valid 0x%08x\n",
 479                inode->i_ino, (long)oa->o_id, oa->o_valid);
 480         obdo_from_inode(oa, inode);
 481         oa->o_mode &= ~S_IFMT;
 482         oa->o_mode |= type;
 483
 484         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
 485                 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
 486                 CDEBUG(D_INODE, "copying device %x from inode to obdo\n",
 487                        rdev);
 488                 *((obd_rdev *)oa->o_inline) = rdev;
 489                 oa->o_obdflags |= OBD_FL_INLINEDATA;
 490                 oa->o_valid |= OBD_MD_FLINLINE;
 491         }
 492
 493 #if 0
 494         else if (filter_has_inline(inode)) {
 495                 CDEBUG(D_INFO, "copying inline from inode to obdo\n");
 496                 memcpy(oa->o_inline, inode->u.ext2_i.i_data,
 497                        MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
 498                 oa->o_obdflags |= OBD_FL_INLINEDATA;
 499                 oa->o_valid |= OBD_MD_FLINLINE;
 500         }
 501
 502         if (filter_has_obdmd(inode)) {
 503                 /* XXX this will change when we don't store the obdmd in data */
 504                 CDEBUG(D_INFO, "copying obdmd from inode to obdo\n");
 505                 memcpy(oa->o_obdmd, inode->u.ext2_i.i_data,
 506                        MIN(sizeof(inode->u.ext2_i.i_data),OBD_INLINESZ));
 507                 oa->o_obdflags |= OBD_FL_OBDMDEXISTS;
 508                 oa->o_valid |= OBD_MD_FLOBDMD;
 509         }
 510 #endif
 511         EXIT;
 512 }
 513
 514 static int filter_getattr(struct obd_conn *conn, struct obdo *oa)
 515 {
 516         struct obd_device *obddev = gen_conn2obd(conn);
 517         struct dentry *dentry;
 518         ENTRY;
 519
 520         if (!gen_client(conn)) {
 521                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 522                 RETURN(-EINVAL);
 523         }
 524
 525         obddev = gen_conn2obd(conn);
 526         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 527                                    oa->o_id, oa->o_mode);
 528         if (IS_ERR(dentry))
 529                 RETURN(PTR_ERR(dentry));
 530
 531         oa->o_valid &= ~OBD_MD_FLID;
 532         filter_from_inode(oa, dentry->d_inode);
 533
 534         dput(dentry);
 535         RETURN(0);
 536 }
 537
 538 static int filter_setattr(struct obd_conn *conn, struct obdo *oa)
 539 {
 540         struct obd_run_ctxt saved;
 541         struct obd_device *obd = gen_conn2obd(conn);
 542         struct dentry *dentry;
 543         struct iattr iattr;
 544         struct inode *inode;
 545         int rc;
 546         ENTRY;
 547
 548         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 549                                    oa->o_id, oa->o_mode);
 550         if (IS_ERR(dentry))
 551                 RETURN(PTR_ERR(dentry));
 552
 553         inode = dentry->d_inode;
 554         iattr_from_obdo(&iattr, oa);
 555         iattr.ia_mode &= ~S_IFMT;
 556         iattr.ia_mode |= S_IFREG;
 557         lock_kernel();
 558         if (iattr.ia_mode & ATTR_SIZE)
 559                 down(&inode->i_sem);
 560         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 561         if (inode->i_op->setattr)
 562                 rc = inode->i_op->setattr(dentry, &iattr);
 563         else
 564                 rc = inode_setattr(inode, &iattr);
 565         pop_ctxt(&saved);
 566         if (iattr.ia_mode & ATTR_SIZE) {
 567                 up(&inode->i_sem);
 568                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 569                 obdo_from_inode(oa, inode);
 570         }
 571         unlock_kernel();
 572
 573         CDEBUG(D_INODE, "put dentry %p, count = %d\n", inode,
 574                atomic_read(&dentry->d_count) - 1);
 575         dput(dentry);
 576         RETURN(rc);
 577 }
 578
 579 static int filter_open(struct obd_conn *conn, struct obdo *oa)
 580 {
 581         struct obd_device *obd;
 582         struct dentry *dentry;
 583         /* ENTRY; */
 584
 585         if (!gen_client(conn)) {
 586                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 587                 RETURN(-EINVAL);
 588         }
 589
 590         obd = gen_conn2obd(conn);
 591         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 592                                    oa->o_id, oa->o_mode);
 593         if (IS_ERR(dentry))
 594                 RETURN(PTR_ERR(dentry));
 595
 596         return 0;
 597 } /* filter_open */
 598
 599 static int filter_close(struct obd_conn *conn, struct obdo *oa)
 600 {
 601         struct obd_device *obd;
 602         struct dentry *dentry;
 603         /* ENTRY; */
 604
 605         if (!gen_client(conn)) {
 606                 CDEBUG(D_IOCTL, "fatal: invalid client %Lx\n", conn->addr);
 607                 RETURN(-EINVAL);
 608         }
 609
 610         obd = gen_conn2obd(conn);
 611         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 612                                    oa->o_id, oa->o_mode);
 613         if (IS_ERR(dentry))
 614                 RETURN(PTR_ERR(dentry));
 615
 616         CDEBUG(D_INODE, "put dentry %p, count = %d\n", dentry,
 617                atomic_read(&dentry->d_count) - 1);
 618         dput(dentry);  /* for the close */
 619         CDEBUG(D_INODE, "put dentry %p, count = %d\n", dentry,
 620                atomic_read(&dentry->d_count) - 1);
 621         dput(dentry);  /* for this call */
 622         return 0;
 623 } /* filter_close */
 624
 625 static int filter_create(struct obd_conn* conn, struct obdo *oa)
 626 {
 627         char name[64];
 628         struct obd_run_ctxt saved;
 629         struct file *file;
 630         int mode;
 631         struct obd_device *obd = gen_conn2obd(conn);
 632         struct iattr;
 633         ENTRY;
 634
 635         if (!gen_client(conn)) {
 636                 CERROR("invalid client %Lx\n", conn->addr);
 637                 return -EINVAL;
 638         }
 639
 640         oa->o_id = filter_next_id(obd);
 641         if (!(oa->o_mode && S_IFMT)) {
 642                 CERROR("filter obd: no type!\n");
 643                 return -ENOENT;
 644         }
 645
 646         filter_id(name, oa->o_id, oa->o_mode);
 647         mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
 648         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 649         file = filp_open(name, O_RDONLY | O_CREAT, mode);
 650         pop_ctxt(&saved);
 651         if (IS_ERR(file)) {
 652                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(file));
 653                 return -ENOENT;
 654         }
 655         filp_close(file, 0);
 656
 657         /* Set flags for fields we have set in the inode struct */
 658         oa->o_valid |= OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 659                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME |
 660                  OBD_MD_FLUID | OBD_MD_FLGID;
 661
 662         /* XXX Hmm, shouldn't we copy the fields into the obdo here? */
 663         return 0;
 664 }
 665
 666 static int filter_destroy(struct obd_conn *conn, struct obdo *oa)
 667 {
 668         struct obd_run_ctxt saved;
 669         struct obd_device *obd;
 670         struct obd_export *export;
 671         struct inode *inode;
 672         struct dentry *dir_dentry, *object_dentry;
 673         int rc;
 674         ENTRY;
 675
 676         if (!(export = gen_client(conn))) {
 677                 CERROR("invalid client %Lx\n", conn->addr);
 678                 RETURN(-EINVAL);
 679         }
 680
 681         CDEBUG(D_INODE, "destroying object %Ld\n",oa->o_id);
 682         obd = gen_conn2obd(conn);
 683
 684         dir_dentry = filter_parent(obd, oa->o_mode);
 685         down(&dir_dentry->d_inode->i_sem);
 686
 687         object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
 688                                           oa->o_mode);
 689         if (IS_ERR(object_dentry))
 690                 GOTO(out, rc = -ENOENT);
 691
 692         inode = object_dentry->d_inode;
 693         if (inode->i_nlink != 1) {
 694                 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
 695                 inode->i_nlink = 1;
 696         }
 697         inode->i_mode = S_IFREG;
 698
 699         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 700
 701         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 702         pop_ctxt(&saved);
 703         CDEBUG(D_INODE, "put child %p, count = %d\n", object_dentry,
 704                atomic_read(&object_dentry->d_count) - 1);
 705         dput(object_dentry);
 706
 707         EXIT;
 708 out:
 709         up(&dir_dentry->d_inode->i_sem);
 710         return rc;
 711 }
 712
 713 /* NB count and offset are used for punch, but not truncate */
 714 static int filter_truncate(struct obd_conn *conn, struct obdo *oa,
 715                            obd_size count, obd_off offset)
 716 {
 717         int error;
 718         ENTRY;
 719
 720         CDEBUG(D_INODE, "calling truncate for object #%Ld, valid = %x, "
 721                "o_size = %Ld\n", oa->o_id, oa->o_valid, oa->o_size);
 722         error = filter_setattr(conn, oa);
 723
 724         RETURN(error);
 725 }
 726
 727 static int filter_pgcache_brw(int cmd, struct obd_conn *conn, obd_count num_oa,
 728                                struct obdo **oa, obd_count *oa_bufs,
 729                                struct page **pages, obd_size *count,
 730                                obd_off *offset, obd_flag *flags, void *callback)
 731 {
 732         struct obd_run_ctxt      saved;
 733         struct super_block      *sb;
 734         int                      onum;          /* index to oas */
 735         int                      pnum;          /* index to pages (bufs) */
 736         unsigned long            retval;
 737         int                      error;
 738         struct file             *file;
 739         struct obd_device      *obd = gen_conn2obd(conn);
 740         ENTRY;
 741
 742         if (!gen_client(conn)) {
 743                 CDEBUG(D_IOCTL, "invalid client %Lx\n", conn->addr);
 744                 RETURN(-EINVAL);
 745         }
 746
 747         sb = obd->u.filter.fo_sb;
 748         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 749         pnum = 0; /* pnum indexes buf 0..num_pages */
 750         for (onum = 0; onum < num_oa; onum++) {
 751                 int pg;
 752
 753                 file = filter_obj_open(obd, oa[onum]->o_id, oa[onum]->o_mode);
 754                 if (IS_ERR(file))
 755                         GOTO(out, retval = PTR_ERR(file));
 756
 757                 /* count doubles as retval */
 758                 for (pg = 0; pg < oa_bufs[onum]; pg++) {
 759                         CDEBUG(D_INODE, "OP %d obdo no/pno: (%d,%d) (%ld,%ld) "
 760                                "off count (%Ld,%Ld)\n",
 761                                cmd, onum, pnum, file->f_dentry->d_inode->i_ino,
 762                                (unsigned long)offset[pnum] >> PAGE_CACHE_SHIFT,
 763                                (unsigned long long)offset[pnum],
 764                                (unsigned long long)count[pnum]);
 765                         if (cmd & OBD_BRW_WRITE) {
 766                                 loff_t off;
 767                                 char *buffer;
 768                                 off = offset[pnum];
 769                                 buffer = kmap(pages[pnum]);
 770                                 retval = file->f_op->write(file, buffer, count[pnum], &off);
 771                                 kunmap(pages[pnum]);
 772                                 CDEBUG(D_INODE, "retval %ld\n", retval);
 773                         } else {
 774                                 loff_t off = offset[pnum];
 775                                 char *buffer = kmap(pages[pnum]);
 776
 777                                 if (off >= file->f_dentry->d_inode->i_size) {
 778                                         memset(buffer, 0, count[pnum]);
 779                                         retval = count[pnum];
 780                                 } else {
 781                                         retval = file->f_op->read(file, buffer, count[pnum], &off);
 782                                 }
 783                                 kunmap(pages[pnum]);
 784
 785                                 if (retval != count[pnum]) {
 786                                         filp_close(file, 0);
 787                                         GOTO(out, retval = -EIO);
 788                                 }
 789                                 CDEBUG(D_INODE, "retval %ld\n", retval);
 790                         }
 791                         pnum++;
 792                 }
 793                 /* sizes and blocks are set by generic_file_write */
 794                 /* ctimes/mtimes will follow with a setattr call */
 795                 filp_close(file, 0);
 796         }
 797
 798         EXIT;
 799 out:
 800         pop_ctxt(&saved);
 801         error = (retval >= 0) ? 0 : retval;
 802         return error;
 803 }
 804
 805
 806 struct inode *ioobj_to_inode(struct obd_conn *conn, struct obd_ioobj *o)
 807 {
 808         struct obd_device *obd = gen_conn2obd(conn);
 809         struct super_block *sb = obd->u.filter.fo_sb;
 810         struct inode *inode = NULL;
 811         ENTRY;
 812
 813         if (!sb || !sb->s_dev) {
 814                 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
 815                 RETURN(NULL);
 816         }
 817
 818         if ( !o->ioo_id ) {
 819                 CDEBUG(D_INODE, "fatal: invalid obdo %lu\n", (long)o->ioo_id);
 820                 RETURN(NULL);
 821         }
 822
 823         inode = filter_inode_from_obj(obd, o->ioo_id, S_IFREG);
 824         if (!inode || inode->i_nlink == 0 || is_bad_inode(inode)) {
 825                 CERROR("from obdo - fatal: invalid inode %ld (%s).\n",
 826                        (long)o->ioo_id, inode ? inode->i_nlink ? "bad inode" :
 827                        "no links" : "NULL");
 828                 iput(inode);
 829                 RETURN(NULL);
 830         }
 831
 832         RETURN(inode);
 833 }
 834
 835 /*
 836  * Calculate the number of buffer credits needed to write multiple pages in
 837  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
 838  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
 839  *
 840  * See comment above ext3_writepage_trans_blocks for details.  We assume
 841  * no data journaling is being done, but it does allow for all of the pages
 842  * being non-contiguous.  If we are guaranteed contiguous pages we could
 843  * reduce the number of (d)indirect blocks a lot.
 844  *
 845  * With N blocks per page and P pages, for each inode we have at most:
 846  * N*P indirect
 847  * min(N*P, blocksize/4 + 1) dindirect blocks
 848  * 1 tindirect
 849  *
 850  * For the entire filesystem, we have at most:
 851  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 852  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 853  * 1 inode block
 854  * 1 superblock
 855  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 856  */
 857 static int ext3_credits_needed(struct super_block *sb, int objcount,
 858                                struct obd_ioobj *obj)
 859 {
 860         struct obd_ioobj *o = obj;
 861         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 862         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 863         int nbitmaps = 0;
 864         int ngdblocks = 0;
 865         int needed = objcount + 1;
 866         int i;
 867
 868         for (i = 0; i < objcount; i++, o++) {
 869                 int nblocks = o->ioo_bufcnt * blockpp;
 870                 int ndindirect = min(nblocks, addrpp + 1);
 871                 int nindir = nblocks + ndindirect + 1;
 872
 873                 nbitmaps += nindir + nblocks;
 874                 ngdblocks += nindir + nblocks;
 875
 876                 needed += nindir;
 877         }
 878
 879         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 880                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 881         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 882                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 883
 884         needed += nbitmaps + ngdblocks;
 885
 886 #ifdef CONFIG_QUOTA
 887         /* We assume that there will be 1 bit set in s_dquot.flags for each
 888          * quota file that is active.  This is at least true for now.
 889          */
 890         needed += hweight32(sb_any_quota_enabled(sb)) *
 891                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 892 #endif
 893
 894         return needed;
 895 }
 896
 897 /* We have to start a huge journal transaction here to hold all of the
 898  * metadata for the pages being written here.  This is necessitated by
 899  * the fact that we do lots of prepare_write operations before we do
 900  * any of the matching commit_write operations, so even if we split
 901  * up to use "smaller" transactions none of them could complete until
 902  * all of them were opened.  By having a single journal transaction,
 903  * we eliminate duplicate reservations for common blocks like the
 904  * superblock and group descriptors or bitmaps.
 905  *
 906  * We will start the transaction here, but each prepare_write will
 907  * add a refcount to the transaction, and each commit_write will
 908  * remove a refcount.  The transaction will be closed when all of
 909  * the pages have been written.
 910  */
 911 static void *ext3_filter_journal_start(struct filter_obd *filter,
 912                                        int objcount, struct obd_ioobj *obj,
 913                                        int niocount, struct niobuf_remote *nb)
 914 {
 915         journal_t *journal = NULL;
 916         handle_t *handle = NULL;
 917         int needed;
 918
 919         /* Assumes ext3 and extN have same sb_info layout, but avoids issues
 920          * with having extN built properly before filterobd for now.
 921          */
 922         journal = EXT3_SB(filter->fo_sb)->s_journal;
 923         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
 924
 925         /* The number of blocks we could _possibly_ dirty can very large.
 926          * We reduce our request if it is absurd (and we couldn't get that
 927          * many credits for a single handle anyways).
 928          *
 929          * At some point we have to limit the size of I/Os sent at one time,
 930          * increase the size of the journal, or we have to calculate the
 931          * actual journal requirements more carefully by checking all of
 932          * the blocks instead of being maximally pessimistic.  It remains to
 933          * be seen if this is a real problem or not.
 934          */
 935         if (needed > journal->j_max_transaction_buffers) {
 936                 CERROR("want too many journal credits (%d) using %d instead\n",
 937                        needed, journal->j_max_transaction_buffers);
 938                 needed = journal->j_max_transaction_buffers;
 939         }
 940
 941         handle = journal_start(journal, needed);
 942         if (IS_ERR(handle))
 943                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 944                        PTR_ERR(handle));
 945
 946         return(handle);
 947 }
 948
 949 static void *filter_journal_start(void **journal_save,
 950                                   struct filter_obd *filter,
 951                                   int objcount, struct obd_ioobj *obj,
 952                                   int niocount, struct niobuf_remote *nb)
 953 {
 954         void *handle = NULL;
 955
 956         /* This may not be necessary - we probably never have a
 957          * transaction started when we enter here, so we can
 958          * remove the saving of the journal state entirely.
 959          * For now leave it in just to see if it ever happens.
 960          */
 961         *journal_save = current->journal_info;
 962         if (*journal_save) {
 963                 CERROR("Already have handle %p???\n", *journal_save);
 964                 LBUG();
 965                 current->journal_info = NULL;
 966         }
 967
 968         if (!strcmp(filter->fo_fstype, "ext3") ||
 969             !strcmp(filter->fo_fstype, "extN"))
 970                 handle = ext3_filter_journal_start(filter, objcount, obj,
 971                                                    niocount, nb);
 972         return handle;
 973 }
 974
 975 static int ext3_filter_journal_stop(void *handle)
 976 {
 977         int rc;
 978
 979         /* We got a refcount on the handle for each call to prepare_write,
 980          * so we can drop the "parent" handle here to avoid the need for
 981          * osc to call back into filterobd to close the handle.  The
 982          * remaining references will be dropped in commit_write.
 983          */
 984         rc = journal_stop((handle_t *)handle);
 985
 986         return rc;
 987 }
 988
 989 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
 990                                void *handle)
 991 {
 992         int rc = 0;
 993
 994         if (!strcmp(filter->fo_fstype, "ext3") ||
 995             !strcmp(filter->fo_fstype, "extN"))
 996                 rc = ext3_filter_journal_stop(handle);
 997
 998         if (rc)
 999                 CERROR("error on journal stop: rc = %d\n", rc);
1000
1001         current->journal_info = journal_save;
1002
1003         return rc;
1004 }
1005
1006 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1007                                    struct niobuf_local *lnb)
1008 {
1009         struct address_space *mapping = inode->i_mapping;
1010         struct page *page;
1011         int rc;
1012
1013         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1014         page = grab_cache_page_nowait(mapping, index); /* locked page */
1015
1016         /* This page is currently locked, so get a temporary page instead */
1017         if (!page) {
1018                 unsigned long addr;
1019                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1020                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1021                 if (!addr) {
1022                         CERROR("no memory for a temp page\n");
1023                         LBUG();
1024                         GOTO(err, rc = -ENOMEM);
1025                 }
1026                 page = virt_to_page(addr);
1027                 kmap(page);
1028                 page->index = index;
1029                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1030         } else if (!IS_ERR(page)) {
1031                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1032                  * a no-op for most filesystems, because we write the whole
1033                  * page.  For partial-page I/O this will read in the page.
1034                  */
1035                 kmap(page);
1036                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1037                 if (rc) {
1038                         CERROR("page index %lu, rc = %d\n", index, rc);
1039                         if (rc != -ENOSPC)
1040                                 LBUG();
1041                         GOTO(err_unlock, rc);
1042                 }
1043                 /* XXX not sure if we need this if we are overwriting page */
1044                 if (PageError(page)) {
1045                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1046                         LBUG();
1047                         GOTO(err_unlock, rc = -EIO);
1048                 }
1049         }
1050         return page;
1051
1052 err_unlock:
1053         unlock_page(page);
1054         lustre_put_page(page);
1055 err:
1056         return ERR_PTR(rc);
1057 }
1058
1059 /*
1060  * We need to balance prepare_write() calls with commit_write() calls.
1061  * If the page has been prepared, but we have no data for it, we don't
1062  * want to overwrite valid data on disk, but we still need to zero out
1063  * data for space which was newly allocated.  Like part of what happens
1064  * in __block_prepare_write() for newly allocated blocks.
1065  *
1066  * XXX currently __block_prepare_write() creates buffers for all the
1067  *     pages, and the filesystems mark these buffers as BH_New if they
1068  *     were newly allocated from disk. We use the BH_New flag similarly.
1069  */
1070 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1071                                int err)
1072 {
1073         if (err) {
1074                 unsigned block_start, block_end;
1075                 struct buffer_head *bh, *head = page->buffers;
1076                 unsigned blocksize = head->b_size;
1077                 void *addr = page_address(page);
1078
1079                 /* Currently one buffer per page, but in the future... */
1080                 for (bh = head, block_start = 0; bh != head || !block_start;
1081                      block_start = block_end, bh = bh->b_this_page) {
1082                         block_end = block_start + blocksize;
1083                         if (buffer_new(bh))
1084                                 memset(addr + block_start, 0, blocksize);
1085                 }
1086         }
1087
1088         return lustre_commit_write(page, from, to);
1089 }
1090
1091 static int filter_preprw(int cmd, struct obd_conn *conn,
1092                          int objcount, struct obd_ioobj *obj,
1093                          int niocount, struct niobuf_remote *nb,
1094                          struct niobuf_local *res, void **desc_private)
1095 {
1096         struct obd_run_ctxt saved;
1097         struct obd_device *obd;
1098         struct obd_ioobj *o = obj;
1099         struct niobuf_remote *b = nb;
1100         struct niobuf_local *r = res;
1101         void *journal_save = NULL;
1102         int rc = 0;
1103         int i;
1104         ENTRY;
1105
1106         memset(res, 0, sizeof(*res) * niocount);
1107         obd = gen_conn2obd(conn);
1108
1109         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1110
1111         if (cmd & OBD_BRW_WRITE) {
1112                 *desc_private = filter_journal_start(&journal_save,
1113                                                      &obd->u.filter,
1114                                                      objcount, obj, niocount,
1115                                                      nb);
1116                 if (IS_ERR(*desc_private))
1117                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1118         }
1119
1120         for (i = 0; i < objcount; i++, o++) {
1121                 struct dentry *dentry;
1122                 struct inode *inode;
1123                 int j;
1124
1125                 dentry = filter_fid2dentry(obd,
1126                                            filter_parent(obd, S_IFREG),
1127                                            o->ioo_id, S_IFREG);
1128                 if (IS_ERR(dentry))
1129                         GOTO(out_clean, rc = PTR_ERR(dentry));
1130                 inode = dentry->d_inode;
1131                 if (!inode) {
1132                         CERROR("trying to BRW to non-existent file %Ld\n",
1133                                (unsigned long long)o->ioo_id);
1134                         dput(dentry);
1135                         GOTO(out_clean, rc = -ENOENT);
1136                 }
1137
1138                 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1139                         unsigned long index = b->offset >> PAGE_SHIFT;
1140                         struct page *page;
1141
1142                         if (j == 0)
1143                                 r->dentry = dentry;
1144                         else
1145                                 r->dentry = dget(dentry);
1146
1147                         if (cmd & OBD_BRW_WRITE)
1148                                 page = filter_get_page_write(inode, index, r);
1149                         else
1150                                 page = lustre_get_page_read(inode, index);
1151
1152                         if (IS_ERR(page)) {
1153                                 dput(dentry);
1154                                 GOTO(out_clean, rc = PTR_ERR(page));
1155                         }
1156
1157                         r->addr = page_address(page);
1158                         r->offset = b->offset;
1159                         r->page = page;
1160                         r->len = PAGE_SIZE;
1161                 }
1162         }
1163
1164 out_stop:
1165         if (cmd & OBD_BRW_WRITE) {
1166                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1167                                               *desc_private);
1168                 if (!rc)
1169                         rc = err;
1170         }
1171 out_ctxt:
1172         pop_ctxt(&saved);
1173         RETURN(rc);
1174 out_clean:
1175         while (r-- > res) {
1176                 dput(r->dentry);
1177                 if (cmd & OBD_BRW_WRITE)
1178                         filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1179                 else
1180                         lustre_put_page(r->page);
1181         }
1182         goto out_stop;
1183 }
1184
1185 static int filter_write_locked_page(struct niobuf_local *lnb)
1186 {
1187         struct page *lpage;
1188         int rc;
1189
1190         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1191         if (IS_ERR(lpage)) {
1192                 /* It is highly unlikely that we would ever get an error here.
1193                  * The page we want to get was previously locked, so it had to
1194                  * have already allocated the space, and we were just writing
1195                  * over the same data, so there would be no hole in the file.
1196                  *
1197                  * XXX: possibility of a race with truncate could exist, need
1198                  *      to check that.  There are no guarantees w.r.t.
1199                  *      write order even on a local filesystem, although the
1200                  *      normal response would be to return the number of bytes
1201                  *      successfully written and leave the rest to the app.
1202                  */
1203                 rc = PTR_ERR(lpage);
1204                 CERROR("error getting locked page index %ld: rc = %d\n",
1205                        lnb->page->index, rc);
1206                 GOTO(out, rc);
1207         }
1208
1209         memcpy(page_address(lpage), kmap(lnb->page), PAGE_SIZE);
1210         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1211         if (rc)
1212                 CERROR("error committing locked page %ld: rc = %d\n",
1213                        lnb->page->index, rc);
1214 out:
1215         kunmap(lnb->page);
1216         __free_pages(lnb->page, 0);
1217         dput(lnb->dentry);
1218
1219         return rc;
1220 }
1221
1222 static int filter_commitrw(int cmd, struct obd_conn *conn,
1223                            int objcount, struct obd_ioobj *obj,
1224                            int niocount, struct niobuf_local *res,
1225                            void *private)
1226 {
1227         struct obd_run_ctxt saved;
1228         struct obd_ioobj *o = obj;
1229         struct niobuf_local *r = res;
1230         struct obd_device *obd = gen_conn2obd(conn);
1231         void *journal_save;
1232         int found_locked = 0;
1233         int rc = 0;
1234         int i;
1235         ENTRY;
1236
1237         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1238         journal_save = current->journal_info;
1239         if (journal_save)
1240                 CERROR("Existing handle %p???\n", journal_save);
1241         current->journal_info = private;
1242         for (i = 0; i < objcount; i++, obj++) {
1243                 int j;
1244                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1245                         struct page *page = r->page;
1246
1247                         if (!page)
1248                                 LBUG();
1249
1250                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1251                                 found_locked = 1;
1252                                 continue;
1253                         }
1254
1255                         if (cmd & OBD_BRW_WRITE) {
1256                                 int err = filter_commit_write(page, 0,
1257                                                               PAGE_SIZE, 0);
1258
1259                                 if (!rc)
1260                                         rc = err;
1261                         } else
1262                                 lustre_put_page(page);
1263
1264                         CDEBUG(D_INODE,
1265                                "put inode %p (%ld), count = %d, nlink = %d\n",
1266                                page->mapping->host,
1267                                page->mapping->host->i_ino,
1268                                atomic_read(&page->mapping->host->i_count) - 1,
1269                                page->mapping->host->i_nlink);
1270                         dput(r->dentry);
1271                 }
1272         }
1273         if (!found_locked)
1274                 goto out_ctxt;
1275
1276         for (i = 0; i < objcount; i++, obj++) {
1277                 int j;
1278                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1279                         int err;
1280                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1281                                 continue;
1282
1283                         err = filter_write_locked_page(r);
1284                         if (!rc)
1285                                 rc = err;
1286                 }
1287         }
1288
1289 out_ctxt:
1290         current->journal_info = journal_save;
1291         pop_ctxt(&saved);
1292         RETURN(0);
1293 }
1294
1295 static int filter_statfs(struct obd_conn *conn, struct statfs * statfs)
1296 {
1297         struct super_block *sb;
1298         int err;
1299         ENTRY;
1300
1301         if (!gen_client(conn)) {
1302                 CDEBUG(D_IOCTL, "invalid client %Lx\n", conn->addr);
1303                 RETURN(-EINVAL);
1304         }
1305
1306         sb = gen_conn2obd(conn)->u.filter.fo_sb;
1307
1308         err = sb->s_op->statfs(sb, statfs);
1309         RETURN(err);
1310 } /* filter_statfs */
1311
1312
1313 static int filter_get_info(struct obd_conn *conn, obd_count keylen,
1314                            void *key, obd_count *vallen, void **val)
1315 {
1316         struct obd_device *obd;
1317         struct obd_export * export;
1318         ENTRY;
1319
1320         if (!(export = gen_client(conn))) {
1321                 CDEBUG(D_IOCTL, "invalid client %Lx\n", conn->addr);
1322                 RETURN(-EINVAL);
1323         }
1324
1325         obd = gen_conn2obd(conn);
1326
1327         if ( keylen == strlen("blocksize") &&
1328              memcmp(key, "blocksize", keylen) == 0 ) {
1329                 *vallen = sizeof(long);
1330                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1331                 RETURN(0);
1332         }
1333
1334         if ( keylen == strlen("blocksize_bits") &&
1335              memcmp(key, "blocksize_bits", keylen) == 0 ){
1336                 *vallen = sizeof(long);
1337                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1338                 RETURN(0);
1339         }
1340
1341         if ( keylen == strlen("root_ino") &&
1342              memcmp(key, "root_ino", keylen) == 0 ){
1343                 *vallen = sizeof(long);
1344                 *val = (void *)(long)FILTER_ROOTINO;
1345                 RETURN(0);
1346         }
1347
1348         CDEBUG(D_IOCTL, "invalid key\n");
1349         RETURN(-EINVAL);
1350 }
1351
1352
1353 static struct obd_ops filter_obd_ops = {
1354         o_get_info:    filter_get_info,
1355         o_setup:       filter_setup,
1356         o_cleanup:     filter_cleanup,
1357         o_connect:     filter_connect,
1358         o_disconnect:  filter_disconnect,
1359         o_statfs:      filter_statfs,
1360         o_getattr:     filter_getattr,
1361         o_create:      filter_create,
1362         o_setattr:     filter_setattr,
1363         o_destroy:     filter_destroy,
1364         o_open:        filter_open,
1365         o_close:       filter_close,
1366         o_brw:         filter_pgcache_brw,
1367         o_punch:       filter_truncate,
1368         o_preprw:      filter_preprw,
1369         o_commitrw:    filter_commitrw
1370 #if 0
1371         o_preallocate: filter_preallocate_inodes,
1372         o_migrate:     filter_migrate,
1373         o_copy:        gen_copy_data,
1374         o_iterate:     filter_iterate
1375 #endif
1376 };
1377
1378
1379 static int __init obdfilter_init(void)
1380 {
1381         printk(KERN_INFO "Filtering OBD driver  v0.001, braam@clusterfs.com\n");
1382         return obd_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1383 }
1384
1385 static void __exit obdfilter_exit(void)
1386 {
1387         obd_unregister_type(OBD_FILTER_DEVICENAME);
1388 }
1389
1390 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
1391 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1392 MODULE_LICENSE("GPL");
1393
1394 module_init(obdfilter_init);
1395 module_exit(obdfilter_exit);