lustre/obdfilter/filter.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/fs/obdfilter/filter.c
   5  *
   6  * Copyright (C) 2001, 2002 Cluster File Systems, Inc.
   7  *
   8  * This code is issued under the GNU General Public License.
   9  * See the file COPYING in this distribution
  10  *
  11  * by Peter Braam <braam@clusterfs.com>
  12  * and Andreas Dilger <adilger@clusterfs.com>
  13  */
  14
  15 #define EXPORT_SYMTAB
  16 #define DEBUG_SUBSYSTEM S_FILTER
  17
  18 #include <linux/module.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/lustre_dlm.h>
  21 #include <linux/obd_filter.h>
  22 #include <linux/ext3_jbd.h>
  23 #include <linux/quotaops.h>
  24 #include <linux/init.h>
  25
  26 long filter_memory;
  27
  28 #define FILTER_ROOTINO 2
  29 #define FILTER_ROOTINO_STR __stringify(FILTER_ROOTINO)
  30
  31 #define S_SHIFT 12
  32 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
  33         [0]                     NULL,
  34         [S_IFREG >> S_SHIFT]    "R",
  35         [S_IFDIR >> S_SHIFT]    "D",
  36         [S_IFCHR >> S_SHIFT]    "C",
  37         [S_IFBLK >> S_SHIFT]    "B",
  38         [S_IFIFO >> S_SHIFT]    "F",
  39         [S_IFSOCK >> S_SHIFT]   "S",
  40         [S_IFLNK >> S_SHIFT]    "L"
  41 };
  42
  43 static inline const char *obd_mode_to_type(int mode)
  44 {
  45         return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
  46 }
  47
  48 /* write the pathname into the string */
  49 static int filter_id(char *buf, obd_id id, obd_mode mode)
  50 {
  51         return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
  52 }
  53
  54 static inline void f_dput(struct dentry *dentry)
  55 {
  56         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
  57                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
  58         LASSERT(atomic_read(&dentry->d_count) > 0);
  59         dput(dentry);
  60 }
  61
  62 /* setup the object store with correct subdirectories */
  63 static int filter_prep(struct obd_device *obddev)
  64 {
  65         struct obd_run_ctxt saved;
  66         struct filter_obd *filter = &obddev->u.filter;
  67         struct dentry *dentry;
  68         struct dentry *root;
  69         struct file *file;
  70         struct inode *inode;
  71         int rc = 0;
  72         __u64 lastino = 2;
  73         int mode = 0;
  74
  75         push_ctxt(&saved, &filter->fo_ctxt);
  76         dentry = simple_mkdir(current->fs->pwd, "O", 0700);
  77         CDEBUG(D_INODE, "got/created O: %p\n", dentry);
  78         if (IS_ERR(dentry)) {
  79                 rc = PTR_ERR(dentry);
  80                 CERROR("cannot open/create O: rc = %d\n", rc);
  81                 GOTO(out, rc);
  82         }
  83         filter->fo_dentry_O = dentry;
  84         dentry = simple_mkdir(current->fs->pwd, "P", 0700);
  85         CDEBUG(D_INODE, "got/created P: %p\n", dentry);
  86         if (IS_ERR(dentry)) {
  87                 rc = PTR_ERR(dentry);
  88                 CERROR("cannot open/create P: rc = %d\n", rc);
  89                 GOTO(out_O, rc);
  90         }
  91         f_dput(dentry);
  92         dentry = simple_mkdir(current->fs->pwd, "D", 0700);
  93         CDEBUG(D_INODE, "got/created D: %p\n", dentry);
  94         if (IS_ERR(dentry)) {
  95                 rc = PTR_ERR(dentry);
  96                 CERROR("cannot open/create D: rc = %d\n", rc);
  97                 GOTO(out_O, rc);
  98         }
  99
 100         root = simple_mknod(dentry, FILTER_ROOTINO_STR, S_IFREG | 0755);
 101         f_dput(dentry);
 102         if (IS_ERR(root)) {
 103                 rc = PTR_ERR(root);
 104                 CERROR("OBD filter: cannot open/create root %d: rc = %d\n",
 105                        FILTER_ROOTINO, rc);
 106                 GOTO(out_O, rc);
 107         }
 108         f_dput(root);
 109
 110         /*
 111          * Create directories and/or get dentries for each object type.
 112          * This saves us from having to do multiple lookups for each one.
 113          */
 114         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 115                 char *type = obd_type_by_mode[mode];
 116
 117                 if (!type) {
 118                         filter->fo_dentry_O_mode[mode] = NULL;
 119                         continue;
 120                 }
 121                 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
 122                 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
 123                 if (IS_ERR(dentry)) {
 124                         rc = PTR_ERR(dentry);
 125                         CERROR("cannot create O/%s: rc = %d\n", type, rc);
 126                         GOTO(out_O_mode, rc);
 127                 }
 128                 filter->fo_dentry_O_mode[mode] = dentry;
 129         }
 130
 131         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 132         if ( !file || IS_ERR(file) ) {
 133                 rc = PTR_ERR(file);
 134                 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
 135                        "D/status", rc);
 136                 GOTO(out_O_mode, rc);
 137         }
 138
 139         /* steal operations */
 140         inode = file->f_dentry->d_inode;
 141         filter->fo_fop = file->f_op;
 142         filter->fo_iop = inode->i_op;
 143         filter->fo_aops = inode->i_mapping->a_ops;
 144
 145         if (inode->i_size == 0) {
 146                 __u64 disk_lastino = cpu_to_le64(lastino);
 147                 ssize_t retval = file->f_op->write(file, (char *)&disk_lastino,
 148                                                    sizeof(disk_lastino),
 149                                                    &file->f_pos);
 150                 if (retval != sizeof(disk_lastino)) {
 151                         CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
 152                         filp_close(file, 0);
 153                         GOTO(out_O_mode, rc = -EIO);
 154                 }
 155         } else {
 156                 __u64 disk_lastino;
 157                 ssize_t retval = file->f_op->read(file, (char *)&disk_lastino,
 158                                                   sizeof(disk_lastino),
 159                                                   &file->f_pos);
 160                 if (retval != sizeof(disk_lastino)) {
 161                         CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
 162                         filp_close(file, 0);
 163                         GOTO(out_O_mode, rc = -EIO);
 164                 }
 165                 lastino = le64_to_cpu(disk_lastino);
 166         }
 167         filter->fo_lastino = lastino;
 168         filp_close(file, 0);
 169
 170         rc = 0;
 171  out:
 172         pop_ctxt(&saved);
 173
 174         return(rc);
 175
 176 out_O_mode:
 177         while (mode-- > 0) {
 178                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 179                 if (dentry) {
 180                         f_dput(dentry);
 181                         filter->fo_dentry_O_mode[mode] = NULL;
 182                 }
 183         }
 184 out_O:
 185         f_dput(filter->fo_dentry_O);
 186         filter->fo_dentry_O = NULL;
 187         goto out;
 188 }
 189
 190 /* cleanup the filter: write last used object id to status file */
 191 static void filter_post(struct obd_device *obddev)
 192 {
 193         struct obd_run_ctxt saved;
 194         struct filter_obd *filter = &obddev->u.filter;
 195         __u64 disk_lastino;
 196         long rc;
 197         struct file *file;
 198         int mode;
 199
 200         push_ctxt(&saved, &filter->fo_ctxt);
 201         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 202         if (IS_ERR(file)) {
 203                 CERROR("OBD filter: cannot create status file\n");
 204                 goto out;
 205         }
 206
 207         file->f_pos = 0;
 208         disk_lastino = cpu_to_le64(filter->fo_lastino);
 209         rc = file->f_op->write(file, (char *)&disk_lastino,
 210                        sizeof(disk_lastino), &file->f_pos);
 211         if (rc != sizeof(disk_lastino))
 212                 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
 213
 214         rc = filp_close(file, NULL);
 215         if (rc)
 216                 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
 217
 218         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 219                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 220                 if (dentry) {
 221                         f_dput(dentry);
 222                         filter->fo_dentry_O_mode[mode] = NULL;
 223                 }
 224         }
 225         f_dput(filter->fo_dentry_O);
 226 out:
 227         pop_ctxt(&saved);
 228 }
 229
 230
 231 static __u64 filter_next_id(struct obd_device *obddev)
 232 {
 233         obd_id id;
 234
 235         spin_lock(&obddev->u.filter.fo_lock);
 236         id = ++obddev->u.filter.fo_lastino;
 237         spin_unlock(&obddev->u.filter.fo_lock);
 238
 239         /* FIXME: write the lastino to disk here */
 240         return id;
 241 }
 242
 243 /* how to get files, dentries, inodes from object id's */
 244 /* parent i_sem is already held if needed for exclusivity */
 245 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
 246                                         struct dentry *dparent,
 247                                         __u64 id, __u32 type)
 248 {
 249         struct super_block *sb = obddev->u.filter.fo_sb;
 250         struct dentry *dchild;
 251         char name[32];
 252         int len;
 253         ENTRY;
 254
 255         if (!sb || !sb->s_dev) {
 256                 CERROR("fatal: device not initialized.\n");
 257                 RETURN(ERR_PTR(-ENXIO));
 258         }
 259
 260         if (id == 0) {
 261                 CERROR("fatal: invalid object #0\n");
 262                 LBUG();
 263                 RETURN(ERR_PTR(-ESTALE));
 264         }
 265
 266         if (!(type & S_IFMT)) {
 267                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 268                        __FUNCTION__, id, type);
 269                 RETURN(ERR_PTR(-EINVAL));
 270         }
 271
 272         len = sprintf(name, LPU64, id);
 273         CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
 274                name);
 275         dchild = lookup_one_len(name, dparent, len);
 276         if (IS_ERR(dchild)) {
 277                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 278                 RETURN(dchild);
 279         }
 280
 281         CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
 282                obd_mode_to_type(type), name, dchild,
 283                atomic_read(&dchild->d_count));
 284
 285         LASSERT(atomic_read(&dchild->d_count) > 0);
 286
 287         RETURN(dchild);
 288 }
 289
 290 static struct file *filter_obj_open(struct obd_device *obddev,
 291                                     __u64 id, __u32 type)
 292 {
 293         struct super_block *sb = obddev->u.filter.fo_sb;
 294         struct obd_run_ctxt saved;
 295         char name[24];
 296         struct file *file;
 297         ENTRY;
 298
 299         if (!sb || !sb->s_dev) {
 300                 CERROR("fatal: device not initialized.\n");
 301                 RETURN(ERR_PTR(-ENXIO));
 302         }
 303
 304         if (!id) {
 305                 CERROR("fatal: invalid obdo "LPU64"\n", id);
 306                 RETURN(ERR_PTR(-ESTALE));
 307         }
 308
 309         if (!(type & S_IFMT)) {
 310                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 311                        __FUNCTION__, id, type);
 312                 RETURN(ERR_PTR(-EINVAL));
 313         }
 314
 315         filter_id(name, id, type);
 316         push_ctxt(&saved, &obddev->u.filter.fo_ctxt);
 317         file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
 318         pop_ctxt(&saved);
 319
 320         CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
 321
 322         if (IS_ERR(file))
 323                 file = NULL;
 324         RETURN(file);
 325 }
 326
 327 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
 328 {
 329         struct filter_obd *filter = &obddev->u.filter;
 330
 331         return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
 332 }
 333
 334
 335 /* obd methods */
 336 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
 337                           obd_uuid_t cluuid)
 338 {
 339         int rc;
 340         ENTRY;
 341         MOD_INC_USE_COUNT;
 342         rc = class_connect(conn, obd, cluuid);
 343         if (rc)
 344                 MOD_DEC_USE_COUNT;
 345         RETURN(rc);
 346 }
 347
 348 static int filter_disconnect(struct lustre_handle *conn)
 349 {
 350         int rc;
 351         ENTRY;
 352
 353         rc = class_disconnect(conn);
 354         if (!rc)
 355                 MOD_DEC_USE_COUNT;
 356
 357         /* XXX cleanup preallocated inodes */
 358         RETURN(rc);
 359 }
 360
 361 /* mount the file system (secretly) */
 362 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
 363 {
 364         struct obd_ioctl_data* data = buf;
 365         struct filter_obd *filter;
 366         struct vfsmount *mnt;
 367         int err = 0;
 368         ENTRY;
 369
 370         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 371                 RETURN(-EINVAL);
 372
 373         MOD_INC_USE_COUNT;
 374         mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
 375         err = PTR_ERR(mnt);
 376         if (IS_ERR(mnt))
 377                 GOTO(err_dec, err);
 378
 379         filter = &obddev->u.filter;;
 380         filter->fo_vfsmnt = mnt;
 381         filter->fo_fstype = strdup(data->ioc_inlbuf2);
 382         filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
 383         /* XXX is this even possible if do_kern_mount succeeded? */
 384         if (!filter->fo_sb)
 385                 GOTO(err_kfree, err = -ENODEV);
 386
 387         OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
 388         filter->fo_ctxt.pwdmnt = mnt;
 389         filter->fo_ctxt.pwd = mnt->mnt_root;
 390         filter->fo_ctxt.fs = get_ds();
 391
 392         err = filter_prep(obddev);
 393         if (err)
 394                 GOTO(err_kfree, err);
 395         spin_lock_init(&filter->fo_lock);
 396
 397         obddev->obd_namespace =
 398                 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
 399         if (obddev->obd_namespace == NULL)
 400                 LBUG();
 401
 402         ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
 403                            "filter_ldlm_client", &obddev->obd_ldlm_client);
 404
 405         RETURN(0);
 406
 407 err_kfree:
 408         kfree(filter->fo_fstype);
 409         unlock_kernel();
 410         mntput(filter->fo_vfsmnt);
 411         filter->fo_sb = 0;
 412         lock_kernel();
 413
 414 err_dec:
 415         MOD_DEC_USE_COUNT;
 416         return err;
 417 }
 418
 419
 420 static int filter_cleanup(struct obd_device * obddev)
 421 {
 422         struct super_block *sb;
 423         ENTRY;
 424
 425         if (!list_empty(&obddev->obd_exports)) {
 426                 CERROR("still has clients!\n");
 427                 class_disconnect_all(obddev);
 428                 if (!list_empty(&obddev->obd_exports)) {
 429                         CERROR("still has exports after forced cleanup?\n");
 430                         RETURN(-EBUSY);
 431                 }
 432         }
 433
 434         ldlm_namespace_free(obddev->obd_namespace);
 435
 436         sb = obddev->u.filter.fo_sb;
 437         if (!obddev->u.filter.fo_sb)
 438                 RETURN(0);
 439
 440         filter_post(obddev);
 441
 442         shrink_dcache_parent(sb->s_root);
 443         unlock_kernel();
 444         mntput(obddev->u.filter.fo_vfsmnt);
 445         obddev->u.filter.fo_sb = 0;
 446         kfree(obddev->u.filter.fo_fstype);
 447
 448         lock_kernel();
 449
 450         MOD_DEC_USE_COUNT;
 451         RETURN(0);
 452 }
 453
 454
 455 static inline void filter_from_inode(struct obdo *oa, struct inode *inode,
 456                                      int valid)
 457 {
 458         int type = oa->o_mode & S_IFMT;
 459         ENTRY;
 460
 461         CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
 462                inode->i_ino, inode, (long)oa->o_id, valid);
 463         /* Don't copy the inode number in place of the object ID */
 464         obdo_from_inode(oa, inode, valid);
 465         oa->o_mode &= ~S_IFMT;
 466         oa->o_mode |= type;
 467
 468         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
 469                 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
 470                 oa->o_rdev = rdev;
 471                 oa->o_valid |= OBD_MD_FLRDEV;
 472         }
 473
 474         EXIT;
 475 }
 476
 477 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
 478                           struct lov_stripe_md *md)
 479 {
 480         struct obd_device *obddev = class_conn2obd(conn);
 481         struct dentry *dentry;
 482         int rc = 0;
 483         ENTRY;
 484
 485         if (!class_conn2export(conn)) {
 486                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 487                 RETURN(-EINVAL);
 488         }
 489
 490         obddev = class_conn2obd(conn);
 491         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 492                                    oa->o_id, oa->o_mode);
 493         if (IS_ERR(dentry))
 494                 RETURN(PTR_ERR(dentry));
 495
 496         if (!dentry->d_inode) {
 497                 CERROR("getattr on non-existent object: "LPU64"\n", oa->o_id);
 498                 GOTO(out_getattr, rc = -ENOENT);
 499         }
 500
 501         filter_from_inode(oa, dentry->d_inode, oa->o_valid);
 502
 503 out_getattr:
 504         f_dput(dentry);
 505         RETURN(rc);
 506 }
 507
 508 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
 509                           struct lov_stripe_md *md)
 510 {
 511         struct obd_run_ctxt saved;
 512         struct obd_device *obd = class_conn2obd(conn);
 513         struct dentry *dentry;
 514         struct iattr iattr;
 515         struct inode *inode;
 516         int rc;
 517         ENTRY;
 518
 519         iattr_from_obdo(&iattr, oa, oa->o_valid);
 520         iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
 521         dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
 522                                    oa->o_id, iattr.ia_mode);
 523         if (IS_ERR(dentry))
 524                 RETURN(PTR_ERR(dentry));
 525
 526         inode = dentry->d_inode;
 527         if (!inode) {
 528                 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
 529                 GOTO(out_setattr, rc = -ENOENT);
 530         }
 531
 532         lock_kernel();
 533         if (iattr.ia_valid & ATTR_SIZE)
 534                 down(&inode->i_sem);
 535         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 536         if (inode->i_op->setattr)
 537                 rc = inode->i_op->setattr(dentry, &iattr);
 538         else
 539                 rc = inode_setattr(inode, &iattr);
 540         pop_ctxt(&saved);
 541         if (iattr.ia_valid & ATTR_SIZE) {
 542                 up(&inode->i_sem);
 543                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 544                 obdo_from_inode(oa, inode, oa->o_valid);
 545         }
 546         unlock_kernel();
 547
 548 out_setattr:
 549         f_dput(dentry);
 550         RETURN(rc);
 551 }
 552
 553 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
 554                        struct lov_stripe_md *ea)
 555 {
 556         struct obd_export *export;
 557         struct obd_device *obd;
 558         struct dentry *dentry;
 559         /* ENTRY; */
 560
 561         export = class_conn2export(conn);
 562         if (!export) {
 563                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 564                 RETURN(-EINVAL);
 565         }
 566
 567         obd = export->exp_obd;
 568         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 569                                    oa->o_id, oa->o_mode);
 570         if (IS_ERR(dentry))
 571                 RETURN(PTR_ERR(dentry));
 572
 573         if (!dentry->d_inode) {
 574                 CERROR("opening non-existent objid "LPX64"\n", oa->o_id);
 575                 RETURN(-ENOENT);
 576         }
 577
 578         filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 579                           OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 580
 581         return 0;
 582 } /* filter_open */
 583
 584 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
 585                         struct lov_stripe_md *ea)
 586 {
 587         struct obd_device *obd;
 588         struct dentry *dentry;
 589         /* ENTRY; */
 590
 591         obd = class_conn2obd(conn);
 592         if (!obd) {
 593                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 594                 RETURN(-EINVAL);
 595         }
 596
 597         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 598                                    oa->o_id, oa->o_mode);
 599         if (IS_ERR(dentry))
 600                 RETURN(PTR_ERR(dentry));
 601         LASSERT(atomic_read(&dentry->d_count) > 1);
 602
 603         f_dput(dentry);  /* for the close */
 604         f_dput(dentry);  /* for this call */
 605         return 0;
 606 } /* filter_close */
 607
 608 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
 609                          struct lov_stripe_md **ea)
 610 {
 611         char name[64];
 612         struct obd_run_ctxt saved;
 613         struct dentry *new;
 614         int mode;
 615         struct obd_device *obd = class_conn2obd(conn);
 616         struct filter_obd *filter = &obd->u.filter;
 617         struct iattr;
 618         ENTRY;
 619
 620         if (!obd) {
 621                 CERROR("invalid client "LPX64"\n", conn->addr);
 622                 return -EINVAL;
 623         }
 624
 625         if (!(oa->o_mode && S_IFMT)) {
 626                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 627                        __FUNCTION__, oa->o_id, oa->o_mode);
 628                 return -ENOENT;
 629         }
 630
 631         oa->o_id = filter_next_id(obd);
 632
 633         //filter_id(name, oa->o_id, oa->o_mode);
 634         sprintf(name, LPU64, oa->o_id);
 635         mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
 636         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 637         new = simple_mknod(filter->fo_dentry_O_mode[S_IFREG >> S_SHIFT], name,
 638                            mode);
 639         pop_ctxt(&saved);
 640         if (IS_ERR(new)) {
 641                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
 642                 return -ENOENT;
 643         }
 644
 645         /* Set flags for fields we have set in the inode struct */
 646         oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 647                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
 648         filter_from_inode(oa, new->d_inode, oa->o_valid);
 649         dput(new);
 650
 651         return 0;
 652 }
 653
 654 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
 655                           struct lov_stripe_md *ea)
 656 {
 657         struct obd_device *obd;
 658         struct filter_obd *filter;
 659         struct obd_run_ctxt saved;
 660         struct inode *inode;
 661         struct dentry *dir_dentry, *object_dentry;
 662         int rc;
 663         ENTRY;
 664
 665         obd = class_conn2obd(conn);
 666         if (!obd) {
 667                 CERROR("invalid client "LPX64"\n", conn->addr);
 668                 RETURN(-EINVAL);
 669         }
 670
 671         CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
 672
 673         dir_dentry = filter_parent(obd, oa->o_mode);
 674         down(&dir_dentry->d_inode->i_sem);
 675
 676         object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
 677                                           oa->o_mode);
 678         if (IS_ERR(object_dentry))
 679                 GOTO(out, rc = -ENOENT);
 680
 681         inode = object_dentry->d_inode;
 682         if (!inode) {
 683                 CERROR("trying to destroy negative inode "LPX64"!\n", oa->o_id);
 684                 GOTO(out, rc = -ENOENT);
 685         }
 686
 687         if (inode->i_nlink != 1) {
 688                 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
 689                 LBUG();
 690                 inode->i_nlink = 1;
 691         }
 692         inode->i_mode = S_IFREG;
 693
 694         if (atomic_read(&inode->i_count) > 1) {
 695 #warning FIXME: need to handle open-unlinked case and move to hold dir
 696                 CERROR("inode has count %d\n", atomic_read(&inode->i_count));
 697         }
 698
 699         filter = &obd->u.filter;
 700         push_ctxt(&saved, &filter->fo_ctxt);
 701
 702         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 703         pop_ctxt(&saved);
 704         f_dput(object_dentry);
 705
 706         EXIT;
 707 out:
 708         up(&dir_dentry->d_inode->i_sem);
 709         return rc;
 710 }
 711
 712 /* NB count and offset are used for punch, but not truncate */
 713 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
 714                            struct lov_stripe_md *md,
 715                            obd_off start, obd_off end)
 716 {
 717         int error;
 718         ENTRY;
 719
 720         if (end != OBD_PUNCH_EOF)
 721                 CERROR("PUNCH not supported, only truncate works\n");
 722
 723         CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
 724                "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
 725         oa->o_size = start;
 726         error = filter_setattr(conn, oa, NULL);
 727         RETURN(error);
 728 }
 729
 730 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
 731                               struct lov_stripe_md *lsm, obd_count oa_bufs,
 732                               struct brw_page *pga, brw_callback_t callback,
 733                               struct io_cb_data *data)
 734 {
 735         struct obd_run_ctxt      saved;
 736         struct super_block      *sb;
 737         int                      pnum;          /* index to pages (bufs) */
 738         unsigned long            retval;
 739         int                      error;
 740         struct file             *file;
 741         struct obd_device       *obd = class_conn2obd(conn);
 742         int pg;
 743         ENTRY;
 744
 745         if (!obd) {
 746                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
 747                 RETURN(-EINVAL);
 748         }
 749
 750         sb = obd->u.filter.fo_sb;
 751         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
 752         pnum = 0; /* pnum indexes buf 0..num_pages */
 753
 754         file = filter_obj_open(obd, lsm->lsm_object_id, S_IFREG);
 755         if (IS_ERR(file))
 756                 GOTO(out, retval = PTR_ERR(file));
 757
 758         /* count doubles as retval */
 759         for (pg = 0; pg < oa_bufs; pg++) {
 760                 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
 761                        ") off count ("LPU64",%d)\n",
 762                        cmd, pnum, file->f_dentry->d_inode->i_ino,
 763                        pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
 764                        (int)pga[pnum].count);
 765                 if (cmd & OBD_BRW_WRITE) {
 766                         loff_t off;
 767                         char *buffer;
 768                         off = pga[pnum].off;
 769                         buffer = kmap(pga[pnum].pg);
 770                         retval = file->f_op->write(file, buffer,
 771                                                    pga[pnum].count,
 772                                                    &off);
 773                         kunmap(pga[pnum].pg);
 774                         CDEBUG(D_INODE, "retval %ld\n", retval);
 775                 } else {
 776                         loff_t off = pga[pnum].off;
 777                         char *buffer = kmap(pga[pnum].pg);
 778
 779                         if (off >= file->f_dentry->d_inode->i_size) {
 780                                 memset(buffer, 0, pga[pnum].count);
 781                                 retval = pga[pnum].count;
 782                         } else {
 783                                 retval = file->f_op->read(file, buffer,
 784                                                           pga[pnum].count, &off);
 785                         }
 786                         kunmap(pga[pnum].pg);
 787
 788                         if (retval != pga[pnum].count) {
 789                                 filp_close(file, 0);
 790                                 GOTO(out, retval = -EIO);
 791                         }
 792                         CDEBUG(D_INODE, "retval %ld\n", retval);
 793                 }
 794                 pnum++;
 795         }
 796         /* sizes and blocks are set by generic_file_write */
 797         /* ctimes/mtimes will follow with a setattr call */
 798         filp_close(file, 0);
 799
 800         /* XXX: do something with callback if it is set? */
 801
 802         EXIT;
 803 out:
 804         pop_ctxt(&saved);
 805         error = (retval >= 0) ? 0 : retval;
 806         return error;
 807 }
 808
 809 /*
 810  * Calculate the number of buffer credits needed to write multiple pages in
 811  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
 812  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
 813  *
 814  * See comment above ext3_writepage_trans_blocks for details.  We assume
 815  * no data journaling is being done, but it does allow for all of the pages
 816  * being non-contiguous.  If we are guaranteed contiguous pages we could
 817  * reduce the number of (d)indirect blocks a lot.
 818  *
 819  * With N blocks per page and P pages, for each inode we have at most:
 820  * N*P indirect
 821  * min(N*P, blocksize/4 + 1) dindirect blocks
 822  * 1 tindirect
 823  *
 824  * For the entire filesystem, we have at most:
 825  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 826  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 827  * 1 inode block
 828  * 1 superblock
 829  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 830  */
 831 static int ext3_credits_needed(struct super_block *sb, int objcount,
 832                                struct obd_ioobj *obj)
 833 {
 834         struct obd_ioobj *o = obj;
 835         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 836         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 837         int nbitmaps = 0;
 838         int ngdblocks = 0;
 839         int needed = objcount + 1;
 840         int i;
 841
 842         for (i = 0; i < objcount; i++, o++) {
 843                 int nblocks = o->ioo_bufcnt * blockpp;
 844                 int ndindirect = min(nblocks, addrpp + 1);
 845                 int nindir = nblocks + ndindirect + 1;
 846
 847                 nbitmaps += nindir + nblocks;
 848                 ngdblocks += nindir + nblocks;
 849
 850                 needed += nindir;
 851         }
 852
 853         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 854                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 855         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 856                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 857
 858         needed += nbitmaps + ngdblocks;
 859
 860 #ifdef CONFIG_QUOTA
 861         /* We assume that there will be 1 bit set in s_dquot.flags for each
 862          * quota file that is active.  This is at least true for now.
 863          */
 864         needed += hweight32(sb_any_quota_enabled(sb)) *
 865                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 866 #endif
 867
 868         return needed;
 869 }
 870
 871 /* We have to start a huge journal transaction here to hold all of the
 872  * metadata for the pages being written here.  This is necessitated by
 873  * the fact that we do lots of prepare_write operations before we do
 874  * any of the matching commit_write operations, so even if we split
 875  * up to use "smaller" transactions none of them could complete until
 876  * all of them were opened.  By having a single journal transaction,
 877  * we eliminate duplicate reservations for common blocks like the
 878  * superblock and group descriptors or bitmaps.
 879  *
 880  * We will start the transaction here, but each prepare_write will
 881  * add a refcount to the transaction, and each commit_write will
 882  * remove a refcount.  The transaction will be closed when all of
 883  * the pages have been written.
 884  */
 885 static void *ext3_filter_journal_start(struct filter_obd *filter,
 886                                        int objcount, struct obd_ioobj *obj,
 887                                        int niocount, struct niobuf_remote *nb)
 888 {
 889         journal_t *journal = NULL;
 890         handle_t *handle = NULL;
 891         int needed;
 892
 893         /* Assumes ext3 and extN have same sb_info layout, but avoids issues
 894          * with having extN built properly before filterobd for now.
 895          */
 896         journal = EXT3_SB(filter->fo_sb)->s_journal;
 897         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
 898
 899         /* The number of blocks we could _possibly_ dirty can very large.
 900          * We reduce our request if it is absurd (and we couldn't get that
 901          * many credits for a single handle anyways).
 902          *
 903          * At some point we have to limit the size of I/Os sent at one time,
 904          * increase the size of the journal, or we have to calculate the
 905          * actual journal requirements more carefully by checking all of
 906          * the blocks instead of being maximally pessimistic.  It remains to
 907          * be seen if this is a real problem or not.
 908          */
 909         if (needed > journal->j_max_transaction_buffers) {
 910                 CERROR("want too many journal credits (%d) using %d instead\n",
 911                        needed, journal->j_max_transaction_buffers);
 912                 needed = journal->j_max_transaction_buffers;
 913         }
 914
 915         handle = journal_start(journal, needed);
 916         if (IS_ERR(handle))
 917                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 918                        PTR_ERR(handle));
 919
 920         return(handle);
 921 }
 922
 923 static void *filter_journal_start(void **journal_save,
 924                                   struct filter_obd *filter,
 925                                   int objcount, struct obd_ioobj *obj,
 926                                   int niocount, struct niobuf_remote *nb)
 927 {
 928         void *handle = NULL;
 929
 930         /* This may not be necessary - we probably never have a
 931          * transaction started when we enter here, so we can
 932          * remove the saving of the journal state entirely.
 933          * For now leave it in just to see if it ever happens.
 934          */
 935         *journal_save = current->journal_info;
 936         if (*journal_save) {
 937                 CERROR("Already have handle %p???\n", *journal_save);
 938                 LBUG();
 939                 current->journal_info = NULL;
 940         }
 941
 942         if (!strcmp(filter->fo_fstype, "ext3") ||
 943             !strcmp(filter->fo_fstype, "extN"))
 944                 handle = ext3_filter_journal_start(filter, objcount, obj,
 945                                                    niocount, nb);
 946         return handle;
 947 }
 948
 949 static int ext3_filter_journal_stop(void *handle)
 950 {
 951         int rc;
 952
 953         /* We got a refcount on the handle for each call to prepare_write,
 954          * so we can drop the "parent" handle here to avoid the need for
 955          * osc to call back into filterobd to close the handle.  The
 956          * remaining references will be dropped in commit_write.
 957          */
 958         rc = journal_stop((handle_t *)handle);
 959
 960         return rc;
 961 }
 962
 963 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
 964                                void *handle)
 965 {
 966         int rc = 0;
 967
 968         if (!strcmp(filter->fo_fstype, "ext3") ||
 969             !strcmp(filter->fo_fstype, "extN"))
 970                 rc = ext3_filter_journal_stop(handle);
 971
 972         if (rc)
 973                 CERROR("error on journal stop: rc = %d\n", rc);
 974
 975         current->journal_info = journal_save;
 976
 977         return rc;
 978 }
 979
 980 static inline void lustre_put_page(struct page *page)
 981 {
 982         kunmap(page);
 983         page_cache_release(page);
 984 }
 985
 986 static struct page *
 987 lustre_get_page_read(struct inode *inode, unsigned long index)
 988 {
 989         struct address_space *mapping = inode->i_mapping;
 990         struct page *page;
 991         int rc;
 992
 993         page = read_cache_page(mapping, index,
 994                                (filler_t*)mapping->a_ops->readpage, NULL);
 995         if (!IS_ERR(page)) {
 996                 wait_on_page(page);
 997                 kmap(page);
 998                 if (!Page_Uptodate(page)) {
 999                         CERROR("page index %lu not uptodate\n", index);
1000                         GOTO(err_page, rc = -EIO);
1001                 }
1002                 if (PageError(page)) {
1003                         CERROR("page index %lu has error\n", index);
1004                         GOTO(err_page, rc = -EIO);
1005                 }
1006         }
1007         return page;
1008
1009 err_page:
1010         lustre_put_page(page);
1011         return ERR_PTR(rc);
1012 }
1013
1014 static struct page *
1015 lustre_get_page_write(struct inode *inode, unsigned long index)
1016 {
1017         struct address_space *mapping = inode->i_mapping;
1018         struct page *page;
1019         int rc;
1020
1021         page = grab_cache_page(mapping, index); /* locked page */
1022
1023         if (!IS_ERR(page)) {
1024                 kmap(page);
1025                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1026                  * a no-op for most filesystems, because we write the whole
1027                  * page.  For partial-page I/O this will read in the page.
1028                  */
1029                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1030                 if (rc) {
1031                         CERROR("page index %lu, rc = %d\n", index, rc);
1032                         if (rc != -ENOSPC)
1033                                 LBUG();
1034                         GOTO(err_unlock, rc);
1035                 }
1036                 /* XXX not sure if we need this if we are overwriting page */
1037                 if (PageError(page)) {
1038                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1039                         LBUG();
1040                         GOTO(err_unlock, rc = -EIO);
1041                 }
1042         }
1043         return page;
1044
1045 err_unlock:
1046         unlock_page(page);
1047         lustre_put_page(page);
1048         return ERR_PTR(rc);
1049 }
1050
1051 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1052 {
1053         struct inode *inode = page->mapping->host;
1054         int err;
1055
1056         err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1057         if (!err && IS_SYNC(inode))
1058                 err = waitfor_one_page(page);
1059
1060         //SetPageUptodate(page); // the client commit_write will do this
1061
1062         SetPageReferenced(page);
1063         unlock_page(page);
1064         lustre_put_page(page);
1065         return err;
1066 }
1067
1068 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1069                                    struct niobuf_local *lnb, int *pglocked)
1070 {
1071         struct address_space *mapping = inode->i_mapping;
1072         struct page *page;
1073         int rc;
1074
1075         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1076         if (*pglocked)
1077                 page = grab_cache_page_nowait(mapping, index); /* locked page */
1078         else
1079                 page = grab_cache_page(mapping, index); /* locked page */
1080
1081
1082         /* This page is currently locked, so get a temporary page instead. */
1083         /* XXX I believe this is a very dangerous thing to do - consider if
1084          *     we had multiple writers for the same file (definitely the case
1085          *     if we are using this codepath).  If writer A locks the page,
1086          *     writer B writes to a copy (as here), writer A drops the page
1087          *     lock, and writer C grabs the lock before B does, then B will
1088          *     later overwrite the data from C, even if C had LDLM locked
1089          *     and initiated the write after B did.
1090          */
1091         if (!page) {
1092                 unsigned long addr;
1093                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1094                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1095                 if (!addr) {
1096                         CERROR("no memory for a temp page\n");
1097                         LBUG();
1098                         GOTO(err, rc = -ENOMEM);
1099                 }
1100                 page = virt_to_page(addr);
1101                 kmap(page);
1102                 page->index = index;
1103                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1104         } else if (!IS_ERR(page)) {
1105                 (*pglocked)++;
1106                 kmap(page);
1107
1108                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1109                  * a no-op for most filesystems, because we write the whole
1110                  * page.  For partial-page I/O this will read in the page.
1111                  */
1112                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1113                 if (rc) {
1114                         CERROR("page index %lu, rc = %d\n", index, rc);
1115                         if (rc != -ENOSPC)
1116                                 LBUG();
1117                         GOTO(err_unlock, rc);
1118                 }
1119                 /* XXX not sure if we need this if we are overwriting page */
1120                 if (PageError(page)) {
1121                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1122                         LBUG();
1123                         GOTO(err_unlock, rc = -EIO);
1124                 }
1125         }
1126         return page;
1127
1128 err_unlock:
1129         unlock_page(page);
1130         lustre_put_page(page);
1131 err:
1132         return ERR_PTR(rc);
1133 }
1134
1135 /*
1136  * We need to balance prepare_write() calls with commit_write() calls.
1137  * If the page has been prepared, but we have no data for it, we don't
1138  * want to overwrite valid data on disk, but we still need to zero out
1139  * data for space which was newly allocated.  Like part of what happens
1140  * in __block_prepare_write() for newly allocated blocks.
1141  *
1142  * XXX currently __block_prepare_write() creates buffers for all the
1143  *     pages, and the filesystems mark these buffers as BH_New if they
1144  *     were newly allocated from disk. We use the BH_New flag similarly.
1145  */
1146 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1147                                int err)
1148 {
1149         if (err) {
1150                 unsigned block_start, block_end;
1151                 struct buffer_head *bh, *head = page->buffers;
1152                 unsigned blocksize = head->b_size;
1153                 void *addr = page_address(page);
1154
1155                 /* debugging: just seeing if this ever happens */
1156                 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1157                        page->index, page->mapping->host->i_ino, err);
1158
1159                 /* Currently one buffer per page, but in the future... */
1160                 for (bh = head, block_start = 0; bh != head || !block_start;
1161                      block_start = block_end, bh = bh->b_this_page) {
1162                         block_end = block_start + blocksize;
1163                         if (buffer_new(bh))
1164                                 memset(addr + block_start, 0, blocksize);
1165                 }
1166         }
1167
1168         return lustre_commit_write(page, from, to);
1169 }
1170
1171 static int filter_preprw(int cmd, struct lustre_handle *conn,
1172                          int objcount, struct obd_ioobj *obj,
1173                          int niocount, struct niobuf_remote *nb,
1174                          struct niobuf_local *res, void **desc_private)
1175 {
1176         struct obd_run_ctxt saved;
1177         struct obd_device *obd;
1178         struct obd_ioobj *o = obj;
1179         struct niobuf_remote *b = nb;
1180         struct niobuf_local *r = res;
1181         void *journal_save = NULL;
1182         int pglocked = 0;
1183         int rc = 0;
1184         int i;
1185         ENTRY;
1186
1187         obd = class_conn2obd(conn);
1188         if (!obd) {
1189                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1190                 RETURN(-EINVAL);
1191         }
1192         memset(res, 0, sizeof(*res) * niocount);
1193
1194         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1195
1196         if (cmd & OBD_BRW_WRITE) {
1197                 *desc_private = filter_journal_start(&journal_save,
1198                                                      &obd->u.filter,
1199                                                      objcount, obj, niocount,
1200                                                      nb);
1201                 if (IS_ERR(*desc_private))
1202                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1203         }
1204
1205         for (i = 0; i < objcount; i++, o++) {
1206                 struct dentry *dentry;
1207                 struct inode *inode;
1208                 int j;
1209
1210                 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1211                                            o->ioo_id, S_IFREG);
1212                 if (IS_ERR(dentry))
1213                         GOTO(out_clean, rc = PTR_ERR(dentry));
1214                 inode = dentry->d_inode;
1215                 if (!inode) {
1216                         CERROR("trying to BRW to non-existent file "LPU64"\n",
1217                                o->ioo_id);
1218                         f_dput(dentry);
1219                         GOTO(out_clean, rc = -ENOENT);
1220                 }
1221
1222                 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1223                         unsigned long index = b->offset >> PAGE_SHIFT;
1224                         struct page *page;
1225
1226                         if (j == 0)
1227                                 r->dentry = dentry;
1228                         else
1229                                 r->dentry = dget(dentry);
1230
1231                         if (cmd & OBD_BRW_WRITE)
1232                                 page = filter_get_page_write(inode, index, r,
1233                                                              &pglocked);
1234                         else
1235                                 page = lustre_get_page_read(inode, index);
1236
1237                         if (IS_ERR(page)) {
1238                                 f_dput(dentry);
1239                                 GOTO(out_clean, rc = PTR_ERR(page));
1240                         }
1241
1242                         r->addr = page_address(page);
1243                         r->offset = b->offset;
1244                         r->page = page;
1245                         r->len = b->len;
1246                 }
1247         }
1248
1249 out_stop:
1250         if (cmd & OBD_BRW_WRITE) {
1251                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1252                                               *desc_private);
1253                 if (!rc)
1254                         rc = err;
1255         }
1256 out_ctxt:
1257         pop_ctxt(&saved);
1258         RETURN(rc);
1259 out_clean:
1260         while (r-- > res) {
1261                 CERROR("error cleanup on brw\n");
1262                 f_dput(r->dentry);
1263                 if (cmd & OBD_BRW_WRITE)
1264                         filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1265                 else
1266                         lustre_put_page(r->page);
1267         }
1268         goto out_stop;
1269 }
1270
1271 static int filter_write_locked_page(struct niobuf_local *lnb)
1272 {
1273         struct page *lpage;
1274         int rc;
1275
1276         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1277         if (IS_ERR(lpage)) {
1278                 /* It is highly unlikely that we would ever get an error here.
1279                  * The page we want to get was previously locked, so it had to
1280                  * have already allocated the space, and we were just writing
1281                  * over the same data, so there would be no hole in the file.
1282                  *
1283                  * XXX: possibility of a race with truncate could exist, need
1284                  *      to check that.  There are no guarantees w.r.t.
1285                  *      write order even on a local filesystem, although the
1286                  *      normal response would be to return the number of bytes
1287                  *      successfully written and leave the rest to the app.
1288                  */
1289                 rc = PTR_ERR(lpage);
1290                 CERROR("error getting locked page index %ld: rc = %d\n",
1291                        lnb->page->index, rc);
1292                 GOTO(out, rc);
1293         }
1294
1295         /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1296          * lustre_commit_write() below, lnb->page was kmapped previously in
1297          * filter_get_page_write() and kunmapped in lustre_put_page() below.
1298          */
1299         memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1300         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1301         if (rc)
1302                 CERROR("error committing locked page %ld: rc = %d\n",
1303                        lnb->page->index, rc);
1304 out:
1305         lustre_put_page(lnb->page);
1306
1307         return rc;
1308 }
1309
1310 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1311                            int objcount, struct obd_ioobj *obj,
1312                            int niocount, struct niobuf_local *res,
1313                            void *private)
1314 {
1315         struct obd_run_ctxt saved;
1316         struct obd_ioobj *o;
1317         struct niobuf_local *r;
1318         struct obd_device *obd = class_conn2obd(conn);
1319         void *journal_save;
1320         int found_locked = 0;
1321         int rc = 0;
1322         int i;
1323         ENTRY;
1324
1325         push_ctxt(&saved, &obd->u.filter.fo_ctxt);
1326         journal_save = current->journal_info;
1327         LASSERT(!journal_save);
1328
1329         current->journal_info = private;
1330         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1331                 int j;
1332                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1333                         struct page *page = r->page;
1334
1335                         if (!page)
1336                                 LBUG();
1337
1338                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1339                                 found_locked++;
1340                                 continue;
1341                         }
1342
1343                         if (cmd & OBD_BRW_WRITE) {
1344                                 int err = filter_commit_write(page, 0,
1345                                                               r->len, 0);
1346
1347                                 if (!rc)
1348                                         rc = err;
1349                         } else
1350                                 lustre_put_page(page);
1351
1352                         f_dput(r->dentry);
1353                 }
1354         }
1355         current->journal_info = journal_save;
1356
1357         if (!found_locked)
1358                 goto out_ctxt;
1359
1360         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1361                 int j;
1362                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1363                         int err;
1364                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1365                                 continue;
1366
1367                         err = filter_write_locked_page(r);
1368                         if (!rc)
1369                                 rc = err;
1370                         f_dput(r->dentry);
1371                 }
1372         }
1373
1374 out_ctxt:
1375         pop_ctxt(&saved);
1376         RETURN(rc);
1377 }
1378
1379 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1380 {
1381         struct obd_device *obd = class_conn2obd(conn);
1382         struct statfs sfs;
1383         int rc;
1384
1385         ENTRY;
1386         rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1387         if (!rc)
1388                 statfs_pack(osfs, &sfs);
1389
1390         return rc;
1391 }
1392
1393 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1394                            void *key, obd_count *vallen, void **val)
1395 {
1396         struct obd_device *obd;
1397         ENTRY;
1398
1399         obd = class_conn2obd(conn);
1400         if (!obd) {
1401                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1402                 RETURN(-EINVAL);
1403         }
1404
1405         if ( keylen == strlen("blocksize") &&
1406              memcmp(key, "blocksize", keylen) == 0 ) {
1407                 *vallen = sizeof(long);
1408                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1409                 RETURN(0);
1410         }
1411
1412         if ( keylen == strlen("blocksize_bits") &&
1413              memcmp(key, "blocksize_bits", keylen) == 0 ){
1414                 *vallen = sizeof(long);
1415                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1416                 RETURN(0);
1417         }
1418
1419         if ( keylen == strlen("root_ino") &&
1420              memcmp(key, "root_ino", keylen) == 0 ){
1421                 *vallen = sizeof(obd_id);
1422                 *val = (void *)(obd_id)FILTER_ROOTINO;
1423                 RETURN(0);
1424         }
1425
1426         CDEBUG(D_IOCTL, "invalid key\n");
1427         RETURN(-EINVAL);
1428 }
1429
1430 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1431                   struct lustre_handle *src_conn, struct obdo *src,
1432                   obd_size count, obd_off offset)
1433 {
1434         struct page *page;
1435         struct lov_stripe_md srcmd, dstmd;
1436         unsigned long index = 0;
1437         int err = 0;
1438
1439         memset(&srcmd, 0, sizeof(srcmd));
1440         memset(&dstmd, 0, sizeof(dstmd));
1441         srcmd.lsm_object_id = src->o_id;
1442         dstmd.lsm_object_id = dst->o_id;
1443
1444         ENTRY;
1445         CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1446                ", dst: ino "LPU64"\n",
1447                src->o_id, src->o_blocks, src->o_size, dst->o_id);
1448         page = alloc_page(GFP_USER);
1449         if (page == NULL)
1450                 RETURN(-ENOMEM);
1451
1452         while (TryLockPage(page))
1453                 ___wait_on_page(page);
1454
1455         /* XXX with brw vector I/O, we could batch up reads and writes here,
1456          *     all we need to do is allocate multiple pages to handle the I/Os
1457          *     and arrays to handle the request parameters.
1458          */
1459         while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1460                 struct brw_page pg;
1461                 struct io_cb_data *cbd = ll_init_cb();
1462
1463                 if (!cbd) {
1464                         err = -ENOMEM;
1465                         EXIT;
1466                         break;
1467                 }
1468
1469                 pg.pg = page;
1470                 pg.count = PAGE_SIZE;
1471                 pg.off = (page->index) << PAGE_SHIFT;
1472                 pg.flag = 0;
1473
1474                 page->index = index;
1475                 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1476                               ll_sync_io_cb, cbd);
1477
1478                 if ( err ) {
1479                         EXIT;
1480                         break;
1481                 }
1482
1483                 cbd = ll_init_cb();
1484                 if (!cbd) {
1485                         err = -ENOMEM;
1486                         EXIT;
1487                         break;
1488                 }
1489                 pg.flag = OBD_BRW_CREATE;
1490                 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1491
1492                 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1493                               ll_sync_io_cb, cbd);
1494
1495                 /* XXX should handle dst->o_size, dst->o_blocks here */
1496                 if ( err ) {
1497                         EXIT;
1498                         break;
1499                 }
1500
1501                 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1502
1503                 index++;
1504         }
1505         dst->o_size = src->o_size;
1506         dst->o_blocks = src->o_blocks;
1507         dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1508         unlock_page(page);
1509         __free_page(page);
1510
1511         RETURN(err);
1512 }
1513
1514
1515 static struct obd_ops filter_obd_ops = {
1516         o_get_info:    filter_get_info,
1517         o_setup:       filter_setup,
1518         o_cleanup:     filter_cleanup,
1519         o_connect:     filter_connect,
1520         o_disconnect:  filter_disconnect,
1521         o_statfs:      filter_statfs,
1522         o_getattr:     filter_getattr,
1523         o_create:      filter_create,
1524         o_setattr:     filter_setattr,
1525         o_destroy:     filter_destroy,
1526         o_open:        filter_open,
1527         o_close:       filter_close,
1528         o_brw:         filter_pgcache_brw,
1529         o_punch:       filter_truncate,
1530         o_preprw:      filter_preprw,
1531         o_commitrw:    filter_commitrw
1532 #if 0
1533         o_preallocate: filter_preallocate_inodes,
1534         o_migrate:     filter_migrate,
1535         o_copy:        filter_copy_data,
1536         o_iterate:     filter_iterate
1537 #endif
1538 };
1539
1540
1541 static int __init obdfilter_init(void)
1542 {
1543         printk(KERN_INFO "Filtering OBD driver  v0.001, info@clusterfs.com\n");
1544         return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1545 }
1546
1547 static void __exit obdfilter_exit(void)
1548 {
1549         class_unregister_type(OBD_FILTER_DEVICENAME);
1550 }
1551
1552 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1553 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1554 MODULE_LICENSE("GPL");
1555
1556 module_init(obdfilter_init);
1557 module_exit(obdfilter_exit);