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)
 356 {
 357         int rc;
 358
 359         MOD_INC_USE_COUNT;
 360         rc = gen_connect(conn);
 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_gen_clients)) {
 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;
 509         struct dentry *dentry;
 510         ENTRY;
 511
 512         if (!gen_client(conn)) {
 513                 CDEBUG(D_IOCTL, "fatal: invalid client %u\n", conn->oc_id);
 514                 RETURN(-EINVAL);
 515         }
 516
 517         obddev = conn->oc_dev;
 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 *obddev;
 534         struct dentry *dentry;
 535         struct iattr iattr;
 536         struct inode *inode;
 537         int rc;
 538         ENTRY;
 539
 540         if (!gen_client(conn)) {
 541                 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
 542                 RETURN(-EINVAL);
 543         }
 544
 545         obddev = conn->oc_dev;
 546         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 547                                    oa->o_id, oa->o_mode);
 548         if (IS_ERR(dentry))
 549                 RETURN(PTR_ERR(dentry));
 550
 551         inode = dentry->d_inode;
 552         iattr_from_obdo(&iattr, oa);
 553         iattr.ia_mode &= ~S_IFMT;
 554         iattr.ia_mode |= S_IFREG;
 555         lock_kernel();
 556         if (iattr.ia_mode & ATTR_SIZE)
 557                 down(&inode->i_sem);
 558         push_ctxt(&saved, &conn->oc_dev->u.filter.fo_ctxt);
 559         if (inode->i_op->setattr)
 560                 rc = inode->i_op->setattr(dentry, &iattr);
 561         else
 562                 rc = inode_setattr(inode, &iattr);
 563         pop_ctxt(&saved);
 564         if (iattr.ia_mode & ATTR_SIZE) {
 565                 up(&inode->i_sem);
 566                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 567                 obdo_from_inode(oa, inode);
 568         }
 569         unlock_kernel();
 570
 571         CDEBUG(D_INODE, "put dentry %p, count = %d\n", inode,
 572                atomic_read(&dentry->d_count) - 1);
 573         dput(dentry);
 574         RETURN(rc);
 575 }
 576
 577 static int filter_open(struct obd_conn *conn, struct obdo *oa)
 578 {
 579         struct obd_device *obddev;
 580         struct dentry *dentry;
 581         /* ENTRY; */
 582
 583         if (!gen_client(conn)) {
 584                 CDEBUG(D_IOCTL, "fatal: invalid client %u\n", conn->oc_id);
 585                 RETURN(-EINVAL);
 586         }
 587
 588         obddev = conn->oc_dev;
 589         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 590                                    oa->o_id, oa->o_mode);
 591         if (IS_ERR(dentry))
 592                 RETURN(PTR_ERR(dentry));
 593
 594         return 0;
 595 } /* filter_open */
 596
 597 static int filter_close(struct obd_conn *conn, struct obdo *oa)
 598 {
 599         struct obd_device *obddev;
 600         struct dentry *dentry;
 601         /* ENTRY; */
 602
 603         if (!gen_client(conn)) {
 604                 CDEBUG(D_IOCTL, "fatal: invalid client %u\n", conn->oc_id);
 605                 RETURN(-EINVAL);
 606         }
 607
 608         obddev = conn->oc_dev;
 609         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 610                                    oa->o_id, oa->o_mode);
 611         if (IS_ERR(dentry))
 612                 RETURN(PTR_ERR(dentry));
 613
 614         CDEBUG(D_INODE, "put dentry %p, count = %d\n", dentry,
 615                atomic_read(&dentry->d_count) - 1);
 616         dput(dentry);  /* for the close */
 617         CDEBUG(D_INODE, "put dentry %p, count = %d\n", dentry,
 618                atomic_read(&dentry->d_count) - 1);
 619         dput(dentry);  /* for this call */
 620         return 0;
 621 } /* filter_close */
 622
 623 static int filter_create(struct obd_conn* conn, struct obdo *oa)
 624 {
 625         char name[64];
 626         struct obd_run_ctxt saved;
 627         struct file *file;
 628         int mode;
 629         struct obd_device *obddev = conn->oc_dev;
 630         struct iattr;
 631         ENTRY;
 632
 633         if (!gen_client(conn)) {
 634                 CERROR("invalid client %u\n", conn->oc_id);
 635                 return -EINVAL;
 636         }
 637
 638         oa->o_id = filter_next_id(conn->oc_dev);
 639         if (!(oa->o_mode && S_IFMT)) {
 640                 CERROR("filter obd: no type!\n");
 641                 return -ENOENT;
 642         }
 643
 644         filter_id(name, oa->o_id, oa->o_mode);
 645         mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
 646         push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
 647         file = filp_open(name, O_RDONLY | O_CREAT, mode);
 648         pop_ctxt(&saved);
 649         if (IS_ERR(file)) {
 650                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(file));
 651                 return -ENOENT;
 652         }
 653         filp_close(file, 0);
 654
 655         /* Set flags for fields we have set in the inode struct */
 656         oa->o_valid |= OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 657                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME |
 658                  OBD_MD_FLUID | OBD_MD_FLGID;
 659
 660         /* XXX Hmm, shouldn't we copy the fields into the obdo here? */
 661         return 0;
 662 }
 663
 664 static int filter_destroy(struct obd_conn *conn, struct obdo *oa)
 665 {
 666         struct obd_run_ctxt saved;
 667         struct obd_device *obddev;
 668         struct obd_client *cli;
 669         struct inode *inode;
 670         struct dentry *dir_dentry, *object_dentry;
 671         int rc;
 672         ENTRY;
 673
 674         if (!(cli = gen_client(conn))) {
 675                 CERROR("invalid client %u\n", conn->oc_id);
 676                 RETURN(-EINVAL);
 677         }
 678
 679         CDEBUG(D_INODE, "destroying object %Ld\n",oa->o_id);
 680         obddev = conn->oc_dev;
 681
 682         dir_dentry = filter_parent(obddev, oa->o_mode);
 683         down(&dir_dentry->d_inode->i_sem);
 684
 685         object_dentry = filter_fid2dentry(obddev, dir_dentry, oa->o_id,
 686                                           oa->o_mode);
 687         if (IS_ERR(object_dentry))
 688                 GOTO(out, rc = -ENOENT);
 689
 690         inode = object_dentry->d_inode;
 691         if (inode->i_nlink != 1) {
 692                 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
 693                 inode->i_nlink = 1;
 694         }
 695         inode->i_mode = S_IFREG;
 696
 697         push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
 698
 699         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 700         pop_ctxt(&saved);
 701         CDEBUG(D_INODE, "put child %p, count = %d\n", object_dentry,
 702                atomic_read(&object_dentry->d_count) - 1);
 703         dput(object_dentry);
 704
 705         EXIT;
 706 out:
 707         up(&dir_dentry->d_inode->i_sem);
 708         return rc;
 709 }
 710
 711 /* NB count and offset are used for punch, but not truncate */
 712 static int filter_truncate(struct obd_conn *conn, struct obdo *oa,
 713                            obd_size count, obd_off offset)
 714 {
 715         int error;
 716         ENTRY;
 717
 718         CDEBUG(D_INODE, "calling truncate for object #%Ld, valid = %x, "
 719                "o_size = %Ld\n", oa->o_id, oa->o_valid, oa->o_size);
 720         error = filter_setattr(conn, oa);
 721
 722         RETURN(error);
 723 }
 724
 725 /* buffer must lie in user memory here */
 726 static int filter_read(struct obd_conn *conn, struct obdo *oa, char *buf,
 727                         obd_size *count, obd_off offset)
 728 {
 729         struct file * file;
 730         unsigned long retval;
 731         int err;
 732         ENTRY;
 733
 734         if (!gen_client(conn)) {
 735                 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
 736                 RETURN(-EINVAL);
 737         }
 738
 739         file = filter_obj_open(conn->oc_dev, oa->o_id, oa->o_mode);
 740         if (IS_ERR(file))
 741                 RETURN(PTR_ERR(file));
 742
 743         /* count doubles as retval */
 744         retval = file->f_op->read(file, buf, *count, (loff_t *)&offset);
 745         filp_close(file, 0);
 746
 747         if ( retval >= 0 ) {
 748                 err = 0;
 749                 *count = retval;
 750         } else {
 751                 err = retval;
 752                 *count = 0;
 753         }
 754
 755         return err;
 756 }
 757
 758
 759 /* buffer must lie in user memory here */
 760 static int filter_write(struct obd_conn *conn, struct obdo *oa, char *buf,
 761                          obd_size *count, obd_off offset)
 762 {
 763         struct obd_run_ctxt saved;
 764         int err;
 765         struct file * file;
 766         unsigned long retval;
 767         ENTRY;
 768
 769         if (!gen_client(conn)) {
 770                 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
 771                 RETURN(-EINVAL);
 772         }
 773
 774         file = filter_obj_open(conn->oc_dev, oa->o_id, oa->o_mode);
 775         if (IS_ERR(file))
 776                 RETURN(PTR_ERR(file));
 777
 778         /* count doubles as retval */
 779         push_ctxt(&saved, &conn->oc_dev->u.filter.fo_ctxt);
 780         retval = file->f_op->write(file, buf, *count, (loff_t *)&offset);
 781         pop_ctxt(&saved);
 782         filp_close(file, 0);
 783
 784         if ( retval >= 0 ) {
 785                 err = 0;
 786                 *count = retval;
 787                 EXIT;
 788         } else {
 789                 err = retval;
 790                 *count = 0;
 791                 EXIT;
 792         }
 793
 794         return err;
 795 } /* filter_write */
 796
 797 static int filter_pgcache_brw(int cmd, struct obd_conn *conn, obd_count num_oa,
 798                                struct obdo **oa, obd_count *oa_bufs,
 799                                struct page **pages, obd_size *count,
 800                                obd_off *offset, obd_flag *flags, void *callback)
 801 {
 802         struct obd_run_ctxt      saved;
 803         struct super_block      *sb;
 804         int                      onum;          /* index to oas */
 805         int                      pnum;          /* index to pages (bufs) */
 806         unsigned long            retval;
 807         int                      error;
 808         struct file             *file;
 809         ENTRY;
 810
 811         if (!gen_client(conn)) {
 812                 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
 813                 RETURN(-EINVAL);
 814         }
 815
 816         sb = conn->oc_dev->u.filter.fo_sb;
 817         push_ctxt(&saved, &conn->oc_dev->u.filter.fo_ctxt);
 818         pnum = 0; /* pnum indexes buf 0..num_pages */
 819         for (onum = 0; onum < num_oa; onum++) {
 820                 int pg;
 821
 822                 file = filter_obj_open(conn->oc_dev, oa[onum]->o_id,
 823                                        oa[onum]->o_mode);
 824                 if (IS_ERR(file))
 825                         GOTO(out, retval = PTR_ERR(file));
 826
 827                 /* count doubles as retval */
 828                 for (pg = 0; pg < oa_bufs[onum]; pg++) {
 829                         CDEBUG(D_INODE, "OP %d obdo no/pno: (%d,%d) (%ld,%ld) "
 830                                "off count (%Ld,%Ld)\n",
 831                                cmd, onum, pnum, file->f_dentry->d_inode->i_ino,
 832                                (unsigned long)offset[pnum] >> PAGE_CACHE_SHIFT,
 833                                (unsigned long long)offset[pnum],
 834                                (unsigned long long)count[pnum]);
 835                         if (cmd & OBD_BRW_WRITE) {
 836                                 loff_t off;
 837                                 char *buffer;
 838                                 off = offset[pnum];
 839                                 buffer = kmap(pages[pnum]);
 840                                 retval = file->f_op->write(file, buffer, count[pnum], &off);
 841                                 kunmap(pages[pnum]);
 842                                 CDEBUG(D_INODE, "retval %ld\n", retval);
 843                         } else {
 844                                 loff_t off = offset[pnum];
 845                                 char *buffer = kmap(pages[pnum]);
 846
 847                                 if (off >= file->f_dentry->d_inode->i_size) {
 848                                         memset(buffer, 0, count[pnum]);
 849                                         retval = count[pnum];
 850                                 } else {
 851                                         retval = file->f_op->read(file, buffer, count[pnum], &off);
 852                                 }
 853                                 kunmap(pages[pnum]);
 854
 855                                 if (retval != count[pnum]) {
 856                                         filp_close(file, 0);
 857                                         GOTO(out, retval = -EIO);
 858                                 }
 859                                 CDEBUG(D_INODE, "retval %ld\n", retval);
 860                         }
 861                         pnum++;
 862                 }
 863                 /* sizes and blocks are set by generic_file_write */
 864                 /* ctimes/mtimes will follow with a setattr call */
 865                 filp_close(file, 0);
 866         }
 867
 868         EXIT;
 869 out:
 870         pop_ctxt(&saved);
 871         error = (retval >= 0) ? 0 : retval;
 872         return error;
 873 }
 874
 875
 876 struct inode *ioobj_to_inode(struct obd_conn *conn, struct obd_ioobj *o)
 877 {
 878         struct super_block *sb = conn->oc_dev->u.filter.fo_sb;
 879         struct inode *inode = NULL;
 880         ENTRY;
 881
 882         if (!sb || !sb->s_dev) {
 883                 CDEBUG(D_SUPER, "fatal: device not initialized.\n");
 884                 RETURN(NULL);
 885         }
 886
 887         if ( !o->ioo_id ) {
 888                 CDEBUG(D_INODE, "fatal: invalid obdo %lu\n", (long)o->ioo_id);
 889                 RETURN(NULL);
 890         }
 891
 892         inode = filter_inode_from_obj(conn->oc_dev, o->ioo_id, S_IFREG);
 893         if (!inode || inode->i_nlink == 0 || is_bad_inode(inode)) {
 894                 CERROR("from obdo - fatal: invalid inode %ld (%s).\n",
 895                        (long)o->ioo_id, inode ? inode->i_nlink ? "bad inode" :
 896                        "no links" : "NULL");
 897                 iput(inode);
 898                 RETURN(NULL);
 899         }
 900
 901         RETURN(inode);
 902 }
 903
 904 /*
 905  * Calculate the number of buffer credits needed to write multiple pages in
 906  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
 907  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
 908  *
 909  * See comment above ext3_writepage_trans_blocks for details.  We assume
 910  * no data journaling is being done, but it does allow for all of the pages
 911  * being non-contiguous.  If we are guaranteed contiguous pages we could
 912  * reduce the number of (d)indirect blocks a lot.
 913  *
 914  * With N blocks per page and P pages, for each inode we have at most:
 915  * N*P indirect
 916  * min(N*P, blocksize/4 + 1) dindirect blocks
 917  * 1 tindirect
 918  *
 919  * For the entire filesystem, we have at most:
 920  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 921  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 922  * 1 inode block
 923  * 1 superblock
 924  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 925  */
 926 static int ext3_credits_needed(struct super_block *sb, int objcount,
 927                                struct obd_ioobj *obj)
 928 {
 929         struct obd_ioobj *o = obj;
 930         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 931         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 932         int nbitmaps = 0;
 933         int ngdblocks = 0;
 934         int needed = objcount + 1;
 935         int i;
 936
 937         for (i = 0; i < objcount; i++, o++) {
 938                 int nblocks = o->ioo_bufcnt * blockpp;
 939                 int ndindirect = min(nblocks, addrpp + 1);
 940                 int nindir = nblocks + ndindirect + 1;
 941
 942                 nbitmaps += nindir + nblocks;
 943                 ngdblocks += nindir + nblocks;
 944
 945                 needed += nindir;
 946         }
 947
 948         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 949                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 950         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 951                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 952
 953         needed += nbitmaps + ngdblocks;
 954
 955 #ifdef CONFIG_QUOTA
 956         /* We assume that there will be 1 bit set in s_dquot.flags for each
 957          * quota file that is active.  This is at least true for now.
 958          */
 959         needed += hweight32(sb_any_quota_enabled(sb)) *
 960                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 961 #endif
 962
 963         return needed;
 964 }
 965
 966 /* We have to start a huge journal transaction here to hold all of the
 967  * metadata for the pages being written here.  This is necessitated by
 968  * the fact that we do lots of prepare_write operations before we do
 969  * any of the matching commit_write operations, so even if we split
 970  * up to use "smaller" transactions none of them could complete until
 971  * all of them were opened.  By having a single journal transaction,
 972  * we eliminate duplicate reservations for common blocks like the
 973  * superblock and group descriptors or bitmaps.
 974  *
 975  * We will start the transaction here, but each prepare_write will
 976  * add a refcount to the transaction, and each commit_write will
 977  * remove a refcount.  The transaction will be closed when all of
 978  * the pages have been written.
 979  */
 980 static void *ext3_filter_journal_start(struct filter_obd *filter,
 981                                        int objcount, struct obd_ioobj *obj,
 982                                        int niocount, struct niobuf_remote *nb)
 983 {
 984         journal_t *journal = NULL;
 985         handle_t *handle = NULL;
 986         int needed;
 987
 988         /* Assumes ext3 and extN have same sb_info layout, but avoids issues
 989          * with having extN built properly before filterobd for now.
 990          */
 991         journal = EXT3_SB(filter->fo_sb)->s_journal;
 992         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
 993
 994         /* The number of blocks we could _possibly_ dirty can very large.
 995          * We reduce our request if it is absurd (and we couldn't get that
 996          * many credits for a single handle anyways).
 997          *
 998          * At some point we have to limit the size of I/Os sent at one time,
 999          * increase the size of the journal, or we have to calculate the
1000          * actual journal requirements more carefully by checking all of
1001          * the blocks instead of being maximally pessimistic.  It remains to
1002          * be seen if this is a real problem or not.
1003          */
1004         if (needed > journal->j_max_transaction_buffers) {
1005                 CERROR("want too many journal credits (%d) using %d instead\n",
1006                        needed, journal->j_max_transaction_buffers);
1007                 needed = journal->j_max_transaction_buffers;
1008         }
1009
1010         handle = journal_start(journal, needed);
1011         if (IS_ERR(handle))
1012                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
1013                        PTR_ERR(handle));
1014
1015         return(handle);
1016 }
1017
1018 static void *filter_journal_start(void **journal_save,
1019                                   struct filter_obd *filter,
1020                                   int objcount, struct obd_ioobj *obj,
1021                                   int niocount, struct niobuf_remote *nb)
1022 {
1023         void *handle = NULL;
1024
1025         /* This may not be necessary - we probably never have a
1026          * transaction started when we enter here, so we can
1027          * remove the saving of the journal state entirely.
1028          * For now leave it in just to see if it ever happens.
1029          */
1030         *journal_save = current->journal_info;
1031         if (*journal_save) {
1032                 CERROR("Already have handle %p???\n", *journal_save);
1033                 LBUG();
1034                 current->journal_info = NULL;
1035         }
1036
1037         if (!strcmp(filter->fo_fstype, "ext3") ||
1038             !strcmp(filter->fo_fstype, "extN"))
1039                 handle = ext3_filter_journal_start(filter, objcount, obj,
1040                                                    niocount, nb);
1041         return handle;
1042 }
1043
1044 static int ext3_filter_journal_stop(void *handle)
1045 {
1046         int rc;
1047
1048         /* We got a refcount on the handle for each call to prepare_write,
1049          * so we can drop the "parent" handle here to avoid the need for
1050          * osc to call back into filterobd to close the handle.  The
1051          * remaining references will be dropped in commit_write.
1052          */
1053         rc = journal_stop((handle_t *)handle);
1054
1055         return rc;
1056 }
1057
1058 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
1059                                void *handle)
1060 {
1061         int rc = 0;
1062
1063         if (!strcmp(filter->fo_fstype, "ext3") ||
1064             !strcmp(filter->fo_fstype, "extN"))
1065                 rc = ext3_filter_journal_stop(handle);
1066
1067         if (rc)
1068                 CERROR("error on journal stop: rc = %d\n", rc);
1069
1070         current->journal_info = journal_save;
1071
1072         return rc;
1073 }
1074
1075 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1076                                    struct niobuf_local *lnb)
1077 {
1078         struct address_space *mapping = inode->i_mapping;
1079         struct page *page;
1080         int rc;
1081
1082         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1083         page = grab_cache_page_nowait(mapping, index); /* locked page */
1084
1085         /* This page is currently locked, so get a temporary page instead */
1086         if (!page) {
1087                 unsigned long addr;
1088                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1089                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1090                 if (!addr) {
1091                         CERROR("no memory for a temp page\n");
1092                         LBUG();
1093                         GOTO(err, rc = -ENOMEM);
1094                 }
1095                 page = virt_to_page(addr);
1096                 kmap(page);
1097                 page->index = index;
1098                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1099         } else if (!IS_ERR(page)) {
1100                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1101                  * a no-op for most filesystems, because we write the whole
1102                  * page.  For partial-page I/O this will read in the page.
1103                  */
1104                 kmap(page);
1105                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1106                 if (rc) {
1107                         CERROR("page index %lu, rc = %d\n", index, rc);
1108                         if (rc != -ENOSPC)
1109                                 LBUG();
1110                         GOTO(err_unlock, rc);
1111                 }
1112                 /* XXX not sure if we need this if we are overwriting page */
1113                 if (PageError(page)) {
1114                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1115                         LBUG();
1116                         GOTO(err_unlock, rc = -EIO);
1117                 }
1118         }
1119         return page;
1120
1121 err_unlock:
1122         unlock_page(page);
1123         lustre_put_page(page);
1124 err:
1125         return ERR_PTR(rc);
1126 }
1127
1128 /*
1129  * We need to balance prepare_write() calls with commit_write() calls.
1130  * If the page has been prepared, but we have no data for it, we don't
1131  * want to overwrite valid data on disk, but we still need to zero out
1132  * data for space which was newly allocated.  Like part of what happens
1133  * in __block_prepare_write() for newly allocated blocks.
1134  *
1135  * XXX currently __block_prepare_write() creates buffers for all the
1136  *     pages, and the filesystems mark these buffers as BH_New if they
1137  *     were newly allocated from disk. We use the BH_New flag similarly.
1138  */
1139 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1140                                int err)
1141 {
1142         if (err) {
1143                 unsigned block_start, block_end;
1144                 struct buffer_head *bh, *head = page->buffers;
1145                 unsigned blocksize = head->b_size;
1146                 void *addr = page_address(page);
1147
1148                 /* Currently one buffer per page, but in the future... */
1149                 for (bh = head, block_start = 0; bh != head || !block_start;
1150                      block_start = block_end, bh = bh->b_this_page) {
1151                         block_end = block_start + blocksize;
1152                         if (buffer_new(bh))
1153                                 memset(addr + block_start, 0, blocksize);
1154                 }
1155         }
1156
1157         return lustre_commit_write(page, from, to);
1158 }
1159
1160 static int filter_preprw(int cmd, struct obd_conn *conn,
1161                          int objcount, struct obd_ioobj *obj,
1162                          int niocount, struct niobuf_remote *nb,
1163                          struct niobuf_local *res, void **desc_private)
1164 {
1165         struct obd_run_ctxt saved;
1166         struct obd_device *obddev;
1167         struct obd_ioobj *o = obj;
1168         struct niobuf_remote *b = nb;
1169         struct niobuf_local *r = res;
1170         void *journal_save = NULL;
1171         int rc = 0;
1172         int i;
1173         ENTRY;
1174
1175         memset(res, 0, sizeof(*res) * niocount);
1176         obddev = conn->oc_dev;
1177
1178         push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
1179
1180         if (cmd & OBD_BRW_WRITE) {
1181                 *desc_private = filter_journal_start(&journal_save,
1182                                                      &obddev->u.filter,
1183                                                      objcount, obj, niocount,
1184                                                      nb);
1185                 if (IS_ERR(*desc_private))
1186                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1187         }
1188
1189         for (i = 0; i < objcount; i++, o++) {
1190                 struct dentry *dentry;
1191                 struct inode *inode;
1192                 int j;
1193
1194                 dentry = filter_fid2dentry(obddev,
1195                                            filter_parent(obddev, S_IFREG),
1196                                            o->ioo_id, S_IFREG);
1197                 if (IS_ERR(dentry))
1198                         GOTO(out_clean, rc = PTR_ERR(dentry));
1199                 inode = dentry->d_inode;
1200                 if (!inode) {
1201                         CERROR("trying to BRW to non-existent file %Ld\n",
1202                                (unsigned long long)o->ioo_id);
1203                         dput(dentry);
1204                         GOTO(out_clean, rc = -ENOENT);
1205                 }
1206
1207                 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1208                         unsigned long index = b->offset >> PAGE_SHIFT;
1209                         struct page *page;
1210
1211                         if (j == 0)
1212                                 r->dentry = dentry;
1213                         else
1214                                 r->dentry = dget(dentry);
1215
1216                         if (cmd & OBD_BRW_WRITE)
1217                                 page = filter_get_page_write(inode, index, r);
1218                         else
1219                                 page = lustre_get_page_read(inode, index);
1220
1221                         if (IS_ERR(page)) {
1222                                 dput(dentry);
1223                                 GOTO(out_clean, rc = PTR_ERR(page));
1224                         }
1225
1226                         r->addr = page_address(page);
1227                         r->offset = b->offset;
1228                         r->page = page;
1229                         r->len = PAGE_SIZE;
1230                 }
1231         }
1232
1233 out_stop:
1234         if (cmd & OBD_BRW_WRITE) {
1235                 int err = filter_journal_stop(journal_save, &obddev->u.filter,
1236                                               *desc_private);
1237                 if (!rc)
1238                         rc = err;
1239         }
1240 out_ctxt:
1241         pop_ctxt(&saved);
1242         RETURN(rc);
1243 out_clean:
1244         while (r-- > res) {
1245                 dput(r->dentry);
1246                 if (cmd & OBD_BRW_WRITE)
1247                         filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1248                 else
1249                         lustre_put_page(r->page);
1250         }
1251         goto out_stop;
1252 }
1253
1254 static int filter_write_locked_page(struct niobuf_local *lnb)
1255 {
1256         struct page *lpage;
1257         int rc;
1258
1259         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1260         if (IS_ERR(lpage)) {
1261                 /* It is highly unlikely that we would ever get an error here.
1262                  * The page we want to get was previously locked, so it had to
1263                  * have already allocated the space, and we were just writing
1264                  * over the same data, so there would be no hole in the file.
1265                  *
1266                  * XXX: possibility of a race with truncate could exist, need
1267                  *      to check that.  There are no guarantees w.r.t.
1268                  *      write order even on a local filesystem, although the
1269                  *      normal response would be to return the number of bytes
1270                  *      successfully written and leave the rest to the app.
1271                  */
1272                 rc = PTR_ERR(lpage);
1273                 CERROR("error getting locked page index %ld: rc = %d\n",
1274                        lnb->page->index, rc);
1275                 GOTO(out, rc);
1276         }
1277
1278         memcpy(page_address(lpage), kmap(lnb->page), PAGE_SIZE);
1279         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1280         if (rc)
1281                 CERROR("error committing locked page %ld: rc = %d\n",
1282                        lnb->page->index, rc);
1283 out:
1284         kunmap(lnb->page);
1285         __free_pages(lnb->page, 0);
1286         dput(lnb->dentry);
1287
1288         return rc;
1289 }
1290
1291 static int filter_commitrw(int cmd, struct obd_conn *conn,
1292                            int objcount, struct obd_ioobj *obj,
1293                            int niocount, struct niobuf_local *res,
1294                            void *private)
1295 {
1296         struct obd_run_ctxt saved;
1297         struct obd_ioobj *o = obj;
1298         struct niobuf_local *r = res;
1299         void *journal_save;
1300         int found_locked = 0;
1301         int rc = 0;
1302         int i;
1303         ENTRY;
1304
1305         push_ctxt(&saved, &conn->oc_dev->u.filter.fo_ctxt);
1306         journal_save = current->journal_info;
1307         if (journal_save)
1308                 CERROR("Existing handle %p???\n", journal_save);
1309         current->journal_info = private;
1310         for (i = 0; i < objcount; i++, obj++) {
1311                 int j;
1312                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1313                         struct page *page = r->page;
1314
1315                         if (!page)
1316                                 LBUG();
1317
1318                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1319                                 found_locked = 1;
1320                                 continue;
1321                         }
1322
1323                         if (cmd & OBD_BRW_WRITE) {
1324                                 int err = filter_commit_write(page, 0,
1325                                                               PAGE_SIZE, 0);
1326
1327                                 if (!rc)
1328                                         rc = err;
1329                         } else
1330                                 lustre_put_page(page);
1331
1332                         CDEBUG(D_INODE,
1333                                "put inode %p (%ld), count = %d, nlink = %d\n",
1334                                page->mapping->host,
1335                                page->mapping->host->i_ino,
1336                                atomic_read(&page->mapping->host->i_count) - 1,
1337                                page->mapping->host->i_nlink);
1338                         dput(r->dentry);
1339                 }
1340         }
1341         if (!found_locked)
1342                 goto out_ctxt;
1343
1344         for (i = 0; i < objcount; i++, obj++) {
1345                 int j;
1346                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1347                         int err;
1348                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1349                                 continue;
1350
1351                         err = filter_write_locked_page(r);
1352                         if (!rc)
1353                                 rc = err;
1354                 }
1355         }
1356
1357 out_ctxt:
1358         current->journal_info = journal_save;
1359         pop_ctxt(&saved);
1360         RETURN(0);
1361 }
1362
1363 static int filter_statfs(struct obd_conn *conn, struct statfs * statfs)
1364 {
1365         struct super_block *sb;
1366         int err;
1367         ENTRY;
1368
1369         if (!gen_client(conn)) {
1370                 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
1371                 RETURN(-EINVAL);
1372         }
1373
1374         sb = conn->oc_dev->u.filter.fo_sb;
1375
1376         err = sb->s_op->statfs(sb, statfs);
1377         RETURN(err);
1378 } /* filter_statfs */
1379
1380
1381 static int filter_get_info(struct obd_conn *conn, obd_count keylen,
1382                            void *key, obd_count *vallen, void **val)
1383 {
1384         struct obd_device *obddev;
1385         struct obd_client * cli;
1386         ENTRY;
1387
1388         if (!(cli = gen_client(conn))) {
1389                 CDEBUG(D_IOCTL, "invalid client %u\n", conn->oc_id);
1390                 RETURN(-EINVAL);
1391         }
1392
1393         obddev = conn->oc_dev;
1394
1395         if ( keylen == strlen("blocksize") &&
1396              memcmp(key, "blocksize", keylen) == 0 ) {
1397                 *vallen = sizeof(long);
1398                 *val = (void *)(long)obddev->u.filter.fo_sb->s_blocksize;
1399                 RETURN(0);
1400         }
1401
1402         if ( keylen == strlen("blocksize_bits") &&
1403              memcmp(key, "blocksize_bits", keylen) == 0 ){
1404                 *vallen = sizeof(long);
1405                 *val = (void *)(long)obddev->u.filter.fo_sb->s_blocksize_bits;
1406                 RETURN(0);
1407         }
1408
1409         if ( keylen == strlen("root_ino") &&
1410              memcmp(key, "root_ino", keylen) == 0 ){
1411                 *vallen = sizeof(long);
1412                 *val = (void *)(long)FILTER_ROOTINO;
1413                 RETURN(0);
1414         }
1415
1416         CDEBUG(D_IOCTL, "invalid key\n");
1417         RETURN(-EINVAL);
1418 }
1419
1420
1421 static struct obd_ops filter_obd_ops = {
1422         o_get_info:    filter_get_info,
1423         o_setup:       filter_setup,
1424         o_cleanup:     filter_cleanup,
1425         o_connect:     filter_connect,
1426         o_disconnect:  filter_disconnect,
1427         o_statfs:      filter_statfs,
1428         o_getattr:     filter_getattr,
1429         o_create:      filter_create,
1430         o_setattr:     filter_setattr,
1431         o_destroy:     filter_destroy,
1432         o_open:        filter_open,
1433         o_close:       filter_close,
1434         o_read:        filter_read,
1435         o_write:       filter_write,
1436         o_brw:         filter_pgcache_brw,
1437         o_punch:       filter_truncate,
1438         o_preprw:      filter_preprw,
1439         o_commitrw:    filter_commitrw
1440 #if 0
1441         o_preallocate: filter_preallocate_inodes,
1442         o_migrate:     filter_migrate,
1443         o_copy:        gen_copy_data,
1444         o_iterate:     filter_iterate
1445 #endif
1446 };
1447
1448
1449 static int __init obdfilter_init(void)
1450 {
1451         printk(KERN_INFO "Filtering OBD driver  v0.001, braam@clusterfs.com\n");
1452         return obd_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1453 }
1454
1455 static void __exit obdfilter_exit(void)
1456 {
1457         obd_unregister_type(OBD_FILTER_DEVICENAME);
1458 }
1459
1460 MODULE_AUTHOR("Peter J. Braam <braam@clusterfs.com>");
1461 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1462 MODULE_LICENSE("GPL");
1463
1464 module_init(obdfilter_init);
1465 module_exit(obdfilter_exit);