lustre/obdfilter/filter.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/fs/obdfilter/filter.c
   5  *
   6  * Copyright (C) 2001, 2002 Cluster File Systems, Inc.
   7  *
   8  * This code is issued under the GNU General Public License.
   9  * See the file COPYING in this distribution
  10  *
  11  * by Peter Braam <braam@clusterfs.com>
  12  * and Andreas Dilger <adilger@clusterfs.com>
  13  */
  14
  15 #define EXPORT_SYMTAB
  16 #define DEBUG_SUBSYSTEM S_FILTER
  17
  18 #include <linux/module.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/obd_class.h>
  21 #include <linux/lustre_dlm.h>
  22 #include <linux/obd_filter.h>
  23 #include <linux/ext3_jbd.h>
  24 #include <linux/quotaops.h>
  25 #include <linux/init.h>
  26
  27 long filter_memory;
  28
  29 #define FILTER_ROOTINO 2
  30 #define FILTER_ROOTINO_STR __stringify(FILTER_ROOTINO)
  31
  32 #define S_SHIFT 12
  33 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
  34         [0]                     NULL,
  35         [S_IFREG >> S_SHIFT]    "R",
  36         [S_IFDIR >> S_SHIFT]    "D",
  37         [S_IFCHR >> S_SHIFT]    "C",
  38         [S_IFBLK >> S_SHIFT]    "B",
  39         [S_IFIFO >> S_SHIFT]    "F",
  40         [S_IFSOCK >> S_SHIFT]   "S",
  41         [S_IFLNK >> S_SHIFT]    "L"
  42 };
  43
  44 static inline const char *obd_mode_to_type(int mode)
  45 {
  46         return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
  47 }
  48
  49 /* write the pathname into the string */
  50 static int filter_id(char *buf, obd_id id, obd_mode mode)
  51 {
  52         return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
  53 }
  54
  55 static inline void f_dput(struct dentry *dentry)
  56 {
  57         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
  58                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
  59         LASSERT(atomic_read(&dentry->d_count) > 0);
  60         dput(dentry);
  61 }
  62
  63 /* setup the object store with correct subdirectories */
  64 static int filter_prep(struct obd_device *obddev)
  65 {
  66         struct obd_run_ctxt saved;
  67         struct filter_obd *filter = &obddev->u.filter;
  68         struct dentry *dentry;
  69         struct dentry *root;
  70         struct file *file;
  71         struct inode *inode;
  72         int rc = 0;
  73         __u64 lastino = 2;
  74         int mode = 0;
  75
  76         push_ctxt(&saved, &filter->fo_ctxt, NULL);
  77         dentry = simple_mkdir(current->fs->pwd, "O", 0700);
  78         CDEBUG(D_INODE, "got/created O: %p\n", dentry);
  79         if (IS_ERR(dentry)) {
  80                 rc = PTR_ERR(dentry);
  81                 CERROR("cannot open/create O: rc = %d\n", rc);
  82                 GOTO(out, rc);
  83         }
  84         filter->fo_dentry_O = dentry;
  85         dentry = simple_mkdir(current->fs->pwd, "P", 0700);
  86         CDEBUG(D_INODE, "got/created P: %p\n", dentry);
  87         if (IS_ERR(dentry)) {
  88                 rc = PTR_ERR(dentry);
  89                 CERROR("cannot open/create P: rc = %d\n", rc);
  90                 GOTO(out_O, rc);
  91         }
  92         f_dput(dentry);
  93         dentry = simple_mkdir(current->fs->pwd, "D", 0700);
  94         CDEBUG(D_INODE, "got/created D: %p\n", dentry);
  95         if (IS_ERR(dentry)) {
  96                 rc = PTR_ERR(dentry);
  97                 CERROR("cannot open/create D: rc = %d\n", rc);
  98                 GOTO(out_O, rc);
  99         }
 100
 101         root = simple_mknod(dentry, FILTER_ROOTINO_STR, S_IFREG | 0755);
 102         f_dput(dentry);
 103         if (IS_ERR(root)) {
 104                 rc = PTR_ERR(root);
 105                 CERROR("OBD filter: cannot open/create root %d: rc = %d\n",
 106                        FILTER_ROOTINO, rc);
 107                 GOTO(out_O, rc);
 108         }
 109         f_dput(root);
 110
 111         /*
 112          * Create directories and/or get dentries for each object type.
 113          * This saves us from having to do multiple lookups for each one.
 114          */
 115         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 116                 char *type = obd_type_by_mode[mode];
 117
 118                 if (!type) {
 119                         filter->fo_dentry_O_mode[mode] = NULL;
 120                         continue;
 121                 }
 122                 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
 123                 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
 124                 if (IS_ERR(dentry)) {
 125                         rc = PTR_ERR(dentry);
 126                         CERROR("cannot create O/%s: rc = %d\n", type, rc);
 127                         GOTO(out_O_mode, rc);
 128                 }
 129                 filter->fo_dentry_O_mode[mode] = dentry;
 130         }
 131
 132         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 133         if ( !file || IS_ERR(file) ) {
 134                 rc = PTR_ERR(file);
 135                 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
 136                        "D/status", rc);
 137                 GOTO(out_O_mode, rc);
 138         }
 139
 140         /* steal operations */
 141         inode = file->f_dentry->d_inode;
 142         filter->fo_fop = file->f_op;
 143         filter->fo_iop = inode->i_op;
 144         filter->fo_aops = inode->i_mapping->a_ops;
 145
 146         if (inode->i_size == 0) {
 147                 __u64 disk_lastino = cpu_to_le64(lastino);
 148                 ssize_t retval = file->f_op->write(file, (char *)&disk_lastino,
 149                                                    sizeof(disk_lastino),
 150                                                    &file->f_pos);
 151                 if (retval != sizeof(disk_lastino)) {
 152                         CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
 153                         filp_close(file, 0);
 154                         GOTO(out_O_mode, rc = -EIO);
 155                 }
 156         } else {
 157                 __u64 disk_lastino;
 158                 ssize_t retval = file->f_op->read(file, (char *)&disk_lastino,
 159                                                   sizeof(disk_lastino),
 160                                                   &file->f_pos);
 161                 if (retval != sizeof(disk_lastino)) {
 162                         CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
 163                         filp_close(file, 0);
 164                         GOTO(out_O_mode, rc = -EIO);
 165                 }
 166                 lastino = le64_to_cpu(disk_lastino);
 167         }
 168         filter->fo_lastino = lastino;
 169         filp_close(file, 0);
 170
 171         rc = 0;
 172  out:
 173         pop_ctxt(&saved);
 174
 175         return(rc);
 176
 177 out_O_mode:
 178         while (mode-- > 0) {
 179                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 180                 if (dentry) {
 181                         f_dput(dentry);
 182                         filter->fo_dentry_O_mode[mode] = NULL;
 183                 }
 184         }
 185 out_O:
 186         f_dput(filter->fo_dentry_O);
 187         filter->fo_dentry_O = NULL;
 188         goto out;
 189 }
 190
 191 /* cleanup the filter: write last used object id to status file */
 192 static void filter_post(struct obd_device *obddev)
 193 {
 194         struct obd_run_ctxt saved;
 195         struct filter_obd *filter = &obddev->u.filter;
 196         __u64 disk_lastino;
 197         long rc;
 198         struct file *file;
 199         int mode;
 200
 201         push_ctxt(&saved, &filter->fo_ctxt, NULL);
 202         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 203         if (IS_ERR(file)) {
 204                 CERROR("OBD filter: cannot create status file\n");
 205                 goto out;
 206         }
 207
 208         file->f_pos = 0;
 209         disk_lastino = cpu_to_le64(filter->fo_lastino);
 210         rc = file->f_op->write(file, (char *)&disk_lastino,
 211                        sizeof(disk_lastino), &file->f_pos);
 212         if (rc != sizeof(disk_lastino))
 213                 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
 214
 215         rc = filp_close(file, NULL);
 216         if (rc)
 217                 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
 218
 219         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 220                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 221                 if (dentry) {
 222                         f_dput(dentry);
 223                         filter->fo_dentry_O_mode[mode] = NULL;
 224                 }
 225         }
 226         f_dput(filter->fo_dentry_O);
 227 out:
 228         pop_ctxt(&saved);
 229 }
 230
 231
 232 static __u64 filter_next_id(struct obd_device *obddev)
 233 {
 234         obd_id id;
 235
 236         spin_lock(&obddev->u.filter.fo_lock);
 237         id = ++obddev->u.filter.fo_lastino;
 238         spin_unlock(&obddev->u.filter.fo_lock);
 239
 240         /* FIXME: write the lastino to disk here */
 241         return id;
 242 }
 243
 244 /* how to get files, dentries, inodes from object id's */
 245 /* parent i_sem is already held if needed for exclusivity */
 246 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
 247                                         struct dentry *dparent,
 248                                         __u64 id, __u32 type)
 249 {
 250         struct super_block *sb = obddev->u.filter.fo_sb;
 251         struct dentry *dchild;
 252         char name[32];
 253         int len;
 254         ENTRY;
 255
 256         if (!sb || !sb->s_dev) {
 257                 CERROR("fatal: device not initialized.\n");
 258                 RETURN(ERR_PTR(-ENXIO));
 259         }
 260
 261         if (id == 0) {
 262                 CERROR("fatal: invalid object #0\n");
 263                 LBUG();
 264                 RETURN(ERR_PTR(-ESTALE));
 265         }
 266
 267         if (!(type & S_IFMT)) {
 268                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 269                        __FUNCTION__, id, type);
 270                 RETURN(ERR_PTR(-EINVAL));
 271         }
 272
 273         len = sprintf(name, LPU64, id);
 274         CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
 275                name);
 276         dchild = lookup_one_len(name, dparent, len);
 277         if (IS_ERR(dchild)) {
 278                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 279                 RETURN(dchild);
 280         }
 281
 282         CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
 283                obd_mode_to_type(type), name, dchild,
 284                atomic_read(&dchild->d_count));
 285
 286         LASSERT(atomic_read(&dchild->d_count) > 0);
 287
 288         RETURN(dchild);
 289 }
 290
 291 static struct file *filter_obj_open(struct obd_device *obddev,
 292                                     __u64 id, __u32 type)
 293 {
 294         struct super_block *sb = obddev->u.filter.fo_sb;
 295         struct obd_run_ctxt saved;
 296         char name[24];
 297         struct file *file;
 298         ENTRY;
 299
 300         if (!sb || !sb->s_dev) {
 301                 CERROR("fatal: device not initialized.\n");
 302                 RETURN(ERR_PTR(-ENXIO));
 303         }
 304
 305         if (!id) {
 306                 CERROR("fatal: invalid obdo "LPU64"\n", id);
 307                 RETURN(ERR_PTR(-ESTALE));
 308         }
 309
 310         if (!(type & S_IFMT)) {
 311                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 312                        __FUNCTION__, id, type);
 313                 RETURN(ERR_PTR(-EINVAL));
 314         }
 315
 316         filter_id(name, id, type);
 317         push_ctxt(&saved, &obddev->u.filter.fo_ctxt, NULL);
 318         file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
 319         pop_ctxt(&saved);
 320
 321         CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
 322
 323         if (IS_ERR(file))
 324                 file = NULL;
 325         RETURN(file);
 326 }
 327
 328 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
 329 {
 330         struct filter_obd *filter = &obddev->u.filter;
 331
 332         return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
 333 }
 334
 335
 336 /* obd methods */
 337 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
 338                           obd_uuid_t cluuid)
 339 {
 340         int rc;
 341         ENTRY;
 342         MOD_INC_USE_COUNT;
 343         rc = class_connect(conn, obd, cluuid);
 344         if (rc)
 345                 MOD_DEC_USE_COUNT;
 346         RETURN(rc);
 347 }
 348
 349 static int filter_disconnect(struct lustre_handle *conn)
 350 {
 351         int rc;
 352         ENTRY;
 353
 354         rc = class_disconnect(conn);
 355         if (!rc)
 356                 MOD_DEC_USE_COUNT;
 357
 358         /* XXX cleanup preallocated inodes */
 359         RETURN(rc);
 360 }
 361
 362 /* mount the file system (secretly) */
 363 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
 364 {
 365         struct obd_ioctl_data* data = buf;
 366         struct filter_obd *filter;
 367         struct vfsmount *mnt;
 368         int err = 0;
 369         ENTRY;
 370
 371         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 372                 RETURN(-EINVAL);
 373
 374         MOD_INC_USE_COUNT;
 375         mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
 376         err = PTR_ERR(mnt);
 377         if (IS_ERR(mnt))
 378                 GOTO(err_dec, err);
 379
 380         filter = &obddev->u.filter;;
 381         filter->fo_vfsmnt = mnt;
 382         filter->fo_fstype = strdup(data->ioc_inlbuf2);
 383         filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
 384         /* XXX is this even possible if do_kern_mount succeeded? */
 385         if (!filter->fo_sb)
 386                 GOTO(err_kfree, err = -ENODEV);
 387
 388         OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
 389         filter->fo_ctxt.pwdmnt = mnt;
 390         filter->fo_ctxt.pwd = mnt->mnt_root;
 391         filter->fo_ctxt.fs = get_ds();
 392
 393         err = filter_prep(obddev);
 394         if (err)
 395                 GOTO(err_kfree, err);
 396         spin_lock_init(&filter->fo_lock);
 397
 398         obddev->obd_namespace =
 399                 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
 400         if (obddev->obd_namespace == NULL)
 401                 LBUG();
 402
 403         ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
 404                            "filter_ldlm_client", &obddev->obd_ldlm_client);
 405
 406         RETURN(0);
 407
 408 err_kfree:
 409         kfree(filter->fo_fstype);
 410         unlock_kernel();
 411         mntput(filter->fo_vfsmnt);
 412         filter->fo_sb = 0;
 413         lock_kernel();
 414
 415 err_dec:
 416         MOD_DEC_USE_COUNT;
 417         return err;
 418 }
 419
 420
 421 static int filter_cleanup(struct obd_device * obddev)
 422 {
 423         struct super_block *sb;
 424         ENTRY;
 425
 426         if (!list_empty(&obddev->obd_exports)) {
 427                 CERROR("still has clients!\n");
 428                 class_disconnect_all(obddev);
 429                 if (!list_empty(&obddev->obd_exports)) {
 430                         CERROR("still has exports after forced cleanup?\n");
 431                         RETURN(-EBUSY);
 432                 }
 433         }
 434
 435         ldlm_namespace_free(obddev->obd_namespace);
 436
 437         sb = obddev->u.filter.fo_sb;
 438         if (!obddev->u.filter.fo_sb)
 439                 RETURN(0);
 440
 441         filter_post(obddev);
 442
 443         shrink_dcache_parent(sb->s_root);
 444         unlock_kernel();
 445         mntput(obddev->u.filter.fo_vfsmnt);
 446         obddev->u.filter.fo_sb = 0;
 447         kfree(obddev->u.filter.fo_fstype);
 448
 449         lock_kernel();
 450
 451         MOD_DEC_USE_COUNT;
 452         RETURN(0);
 453 }
 454
 455
 456 static inline void filter_from_inode(struct obdo *oa, struct inode *inode,
 457                                      int valid)
 458 {
 459         int type = oa->o_mode & S_IFMT;
 460         ENTRY;
 461
 462         CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
 463                inode->i_ino, inode, (long)oa->o_id, valid);
 464         /* Don't copy the inode number in place of the object ID */
 465         obdo_from_inode(oa, inode, valid);
 466         oa->o_mode &= ~S_IFMT;
 467         oa->o_mode |= type;
 468
 469         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
 470                 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
 471                 oa->o_rdev = rdev;
 472                 oa->o_valid |= OBD_MD_FLRDEV;
 473         }
 474
 475         EXIT;
 476 }
 477
 478 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
 479                           struct lov_stripe_md *md)
 480 {
 481         struct obd_device *obddev = class_conn2obd(conn);
 482         struct dentry *dentry;
 483         int rc = 0;
 484         ENTRY;
 485
 486         if (!class_conn2export(conn)) {
 487                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 488                 RETURN(-EINVAL);
 489         }
 490
 491         obddev = class_conn2obd(conn);
 492         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 493                                    oa->o_id, oa->o_mode);
 494         if (IS_ERR(dentry))
 495                 RETURN(PTR_ERR(dentry));
 496
 497         if (!dentry->d_inode) {
 498                 CERROR("getattr on non-existent object: "LPU64"\n", oa->o_id);
 499                 GOTO(out_getattr, rc = -ENOENT);
 500         }
 501
 502         filter_from_inode(oa, dentry->d_inode, oa->o_valid);
 503
 504 out_getattr:
 505         f_dput(dentry);
 506         RETURN(rc);
 507 }
 508
 509 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
 510                           struct lov_stripe_md *md)
 511 {
 512         struct obd_run_ctxt saved;
 513         struct obd_device *obd = class_conn2obd(conn);
 514         struct dentry *dentry;
 515         struct iattr iattr;
 516         struct inode *inode;
 517         int rc;
 518         ENTRY;
 519
 520         iattr_from_obdo(&iattr, oa, oa->o_valid);
 521         iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
 522         dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
 523                                    oa->o_id, iattr.ia_mode);
 524         if (IS_ERR(dentry))
 525                 RETURN(PTR_ERR(dentry));
 526
 527         inode = dentry->d_inode;
 528         if (!inode) {
 529                 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
 530                 GOTO(out_setattr, rc = -ENOENT);
 531         }
 532
 533         lock_kernel();
 534         if (iattr.ia_valid & ATTR_SIZE)
 535                 down(&inode->i_sem);
 536         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 537         if (inode->i_op->setattr)
 538                 rc = inode->i_op->setattr(dentry, &iattr);
 539         else
 540                 rc = inode_setattr(inode, &iattr);
 541         pop_ctxt(&saved);
 542         if (iattr.ia_valid & ATTR_SIZE) {
 543                 up(&inode->i_sem);
 544                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 545                 obdo_from_inode(oa, inode, oa->o_valid);
 546         }
 547         unlock_kernel();
 548
 549 out_setattr:
 550         f_dput(dentry);
 551         RETURN(rc);
 552 }
 553
 554 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
 555                        struct lov_stripe_md *ea)
 556 {
 557         struct obd_export *export;
 558         struct obd_device *obd;
 559         struct dentry *dentry;
 560         /* ENTRY; */
 561
 562         export = class_conn2export(conn);
 563         if (!export) {
 564                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 565                 RETURN(-EINVAL);
 566         }
 567
 568         obd = export->exp_obd;
 569         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 570                                    oa->o_id, oa->o_mode);
 571         if (IS_ERR(dentry))
 572                 RETURN(PTR_ERR(dentry));
 573
 574         if (!dentry->d_inode) {
 575                 CERROR("opening non-existent objid "LPX64"\n", oa->o_id);
 576                 RETURN(-ENOENT);
 577         }
 578
 579         filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 580                           OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 581
 582         return 0;
 583 } /* filter_open */
 584
 585 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
 586                         struct lov_stripe_md *ea)
 587 {
 588         struct obd_device *obd;
 589         struct dentry *dentry;
 590         /* ENTRY; */
 591
 592         obd = class_conn2obd(conn);
 593         if (!obd) {
 594                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 595                 RETURN(-EINVAL);
 596         }
 597
 598         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 599                                    oa->o_id, oa->o_mode);
 600         if (IS_ERR(dentry))
 601                 RETURN(PTR_ERR(dentry));
 602         LASSERT(atomic_read(&dentry->d_count) > 1);
 603
 604         f_dput(dentry);  /* for the close */
 605         f_dput(dentry);  /* for this call */
 606         return 0;
 607 } /* filter_close */
 608
 609 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
 610                          struct lov_stripe_md **ea)
 611 {
 612         char name[64];
 613         struct obd_run_ctxt saved;
 614         struct dentry *new;
 615         int mode;
 616         struct obd_device *obd = class_conn2obd(conn);
 617         struct filter_obd *filter = &obd->u.filter;
 618         struct iattr;
 619         ENTRY;
 620
 621         if (!obd) {
 622                 CERROR("invalid client "LPX64"\n", conn->addr);
 623                 return -EINVAL;
 624         }
 625
 626         if (!(oa->o_mode && S_IFMT)) {
 627                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 628                        __FUNCTION__, oa->o_id, oa->o_mode);
 629                 return -ENOENT;
 630         }
 631
 632         oa->o_id = filter_next_id(obd);
 633
 634         //filter_id(name, oa->o_id, oa->o_mode);
 635         sprintf(name, LPU64, oa->o_id);
 636         mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
 637         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 638         new = simple_mknod(filter->fo_dentry_O_mode[S_IFREG >> S_SHIFT], name,
 639                            mode);
 640         pop_ctxt(&saved);
 641         if (IS_ERR(new)) {
 642                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
 643                 return -ENOENT;
 644         }
 645
 646         /* Set flags for fields we have set in the inode struct */
 647         oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 648                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
 649         filter_from_inode(oa, new->d_inode, oa->o_valid);
 650         dput(new);
 651
 652         return 0;
 653 }
 654
 655 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
 656                           struct lov_stripe_md *ea)
 657 {
 658         struct obd_device *obd;
 659         struct filter_obd *filter;
 660         struct obd_run_ctxt saved;
 661         struct inode *inode;
 662         struct dentry *dir_dentry, *object_dentry;
 663         int rc;
 664         ENTRY;
 665
 666         obd = class_conn2obd(conn);
 667         if (!obd) {
 668                 CERROR("invalid client "LPX64"\n", conn->addr);
 669                 RETURN(-EINVAL);
 670         }
 671
 672         CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
 673
 674         dir_dentry = filter_parent(obd, oa->o_mode);
 675         down(&dir_dentry->d_inode->i_sem);
 676
 677         object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
 678                                           oa->o_mode);
 679         if (IS_ERR(object_dentry))
 680                 GOTO(out, rc = -ENOENT);
 681
 682         inode = object_dentry->d_inode;
 683         if (!inode) {
 684                 CERROR("trying to destroy negative inode "LPX64"!\n", oa->o_id);
 685                 GOTO(out, rc = -ENOENT);
 686         }
 687
 688         if (inode->i_nlink != 1) {
 689                 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
 690                 LBUG();
 691                 inode->i_nlink = 1;
 692         }
 693         inode->i_mode = S_IFREG;
 694
 695         if (atomic_read(&inode->i_count) > 1) {
 696 #warning FIXME: need to handle open-unlinked case and move to hold dir
 697                 CERROR("inode has count %d\n", atomic_read(&inode->i_count));
 698         }
 699
 700         filter = &obd->u.filter;
 701         push_ctxt(&saved, &filter->fo_ctxt, NULL);
 702
 703         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 704         pop_ctxt(&saved);
 705         f_dput(object_dentry);
 706
 707         EXIT;
 708 out:
 709         up(&dir_dentry->d_inode->i_sem);
 710         return rc;
 711 }
 712
 713 /* NB count and offset are used for punch, but not truncate */
 714 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
 715                            struct lov_stripe_md *md,
 716                            obd_off start, obd_off end)
 717 {
 718         int error;
 719         ENTRY;
 720
 721         if (end != OBD_PUNCH_EOF)
 722                 CERROR("PUNCH not supported, only truncate works\n");
 723
 724         CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
 725                "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
 726         oa->o_size = start;
 727         error = filter_setattr(conn, oa, NULL);
 728         RETURN(error);
 729 }
 730
 731 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
 732                               struct lov_stripe_md *lsm, obd_count oa_bufs,
 733                               struct brw_page *pga, brw_callback_t callback,
 734                               struct io_cb_data *data)
 735 {
 736         struct obd_run_ctxt      saved;
 737         struct super_block      *sb;
 738         int                      pnum;          /* index to pages (bufs) */
 739         unsigned long            retval;
 740         int                      error;
 741         struct file             *file;
 742         struct obd_device       *obd = class_conn2obd(conn);
 743         int pg;
 744         ENTRY;
 745
 746         if (!obd) {
 747                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
 748                 RETURN(-EINVAL);
 749         }
 750
 751         sb = obd->u.filter.fo_sb;
 752         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 753         pnum = 0; /* pnum indexes buf 0..num_pages */
 754
 755         file = filter_obj_open(obd, lsm->lsm_object_id, S_IFREG);
 756         if (IS_ERR(file))
 757                 GOTO(out, retval = PTR_ERR(file));
 758
 759         /* count doubles as retval */
 760         for (pg = 0; pg < oa_bufs; pg++) {
 761                 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
 762                        ") off count ("LPU64",%d)\n",
 763                        cmd, pnum, file->f_dentry->d_inode->i_ino,
 764                        pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
 765                        (int)pga[pnum].count);
 766                 if (cmd & OBD_BRW_WRITE) {
 767                         loff_t off;
 768                         char *buffer;
 769                         off = pga[pnum].off;
 770                         buffer = kmap(pga[pnum].pg);
 771                         retval = file->f_op->write(file, buffer,
 772                                                    pga[pnum].count,
 773                                                    &off);
 774                         kunmap(pga[pnum].pg);
 775                         CDEBUG(D_INODE, "retval %ld\n", retval);
 776                 } else {
 777                         loff_t off = pga[pnum].off;
 778                         char *buffer = kmap(pga[pnum].pg);
 779
 780                         if (off >= file->f_dentry->d_inode->i_size) {
 781                                 memset(buffer, 0, pga[pnum].count);
 782                                 retval = pga[pnum].count;
 783                         } else {
 784                                 retval = file->f_op->read(file, buffer,
 785                                                           pga[pnum].count, &off);
 786                         }
 787                         kunmap(pga[pnum].pg);
 788
 789                         if (retval != pga[pnum].count) {
 790                                 filp_close(file, 0);
 791                                 GOTO(out, retval = -EIO);
 792                         }
 793                         CDEBUG(D_INODE, "retval %ld\n", retval);
 794                 }
 795                 pnum++;
 796         }
 797         /* sizes and blocks are set by generic_file_write */
 798         /* ctimes/mtimes will follow with a setattr call */
 799         filp_close(file, 0);
 800
 801         /* XXX: do something with callback if it is set? */
 802
 803         EXIT;
 804 out:
 805         pop_ctxt(&saved);
 806         error = (retval >= 0) ? 0 : retval;
 807         return error;
 808 }
 809
 810 /*
 811  * Calculate the number of buffer credits needed to write multiple pages in
 812  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
 813  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
 814  *
 815  * See comment above ext3_writepage_trans_blocks for details.  We assume
 816  * no data journaling is being done, but it does allow for all of the pages
 817  * being non-contiguous.  If we are guaranteed contiguous pages we could
 818  * reduce the number of (d)indirect blocks a lot.
 819  *
 820  * With N blocks per page and P pages, for each inode we have at most:
 821  * N*P indirect
 822  * min(N*P, blocksize/4 + 1) dindirect blocks
 823  * 1 tindirect
 824  *
 825  * For the entire filesystem, we have at most:
 826  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 827  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 828  * 1 inode block
 829  * 1 superblock
 830  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 831  */
 832 static int ext3_credits_needed(struct super_block *sb, int objcount,
 833                                struct obd_ioobj *obj)
 834 {
 835         struct obd_ioobj *o = obj;
 836         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 837         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 838         int nbitmaps = 0;
 839         int ngdblocks = 0;
 840         int needed = objcount + 1;
 841         int i;
 842
 843         for (i = 0; i < objcount; i++, o++) {
 844                 int nblocks = o->ioo_bufcnt * blockpp;
 845                 int ndindirect = min(nblocks, addrpp + 1);
 846                 int nindir = nblocks + ndindirect + 1;
 847
 848                 nbitmaps += nindir + nblocks;
 849                 ngdblocks += nindir + nblocks;
 850
 851                 needed += nindir;
 852         }
 853
 854         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 855                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 856         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 857                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 858
 859         needed += nbitmaps + ngdblocks;
 860
 861 #ifdef CONFIG_QUOTA
 862         /* We assume that there will be 1 bit set in s_dquot.flags for each
 863          * quota file that is active.  This is at least true for now.
 864          */
 865         needed += hweight32(sb_any_quota_enabled(sb)) *
 866                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 867 #endif
 868
 869         return needed;
 870 }
 871
 872 /* We have to start a huge journal transaction here to hold all of the
 873  * metadata for the pages being written here.  This is necessitated by
 874  * the fact that we do lots of prepare_write operations before we do
 875  * any of the matching commit_write operations, so even if we split
 876  * up to use "smaller" transactions none of them could complete until
 877  * all of them were opened.  By having a single journal transaction,
 878  * we eliminate duplicate reservations for common blocks like the
 879  * superblock and group descriptors or bitmaps.
 880  *
 881  * We will start the transaction here, but each prepare_write will
 882  * add a refcount to the transaction, and each commit_write will
 883  * remove a refcount.  The transaction will be closed when all of
 884  * the pages have been written.
 885  */
 886 static void *ext3_filter_journal_start(struct filter_obd *filter,
 887                                        int objcount, struct obd_ioobj *obj,
 888                                        int niocount, struct niobuf_remote *nb)
 889 {
 890         journal_t *journal = NULL;
 891         handle_t *handle = NULL;
 892         int needed;
 893
 894         /* Assumes ext3 and extN have same sb_info layout, but avoids issues
 895          * with having extN built properly before filterobd for now.
 896          */
 897         journal = EXT3_SB(filter->fo_sb)->s_journal;
 898         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
 899
 900         /* The number of blocks we could _possibly_ dirty can very large.
 901          * We reduce our request if it is absurd (and we couldn't get that
 902          * many credits for a single handle anyways).
 903          *
 904          * At some point we have to limit the size of I/Os sent at one time,
 905          * increase the size of the journal, or we have to calculate the
 906          * actual journal requirements more carefully by checking all of
 907          * the blocks instead of being maximally pessimistic.  It remains to
 908          * be seen if this is a real problem or not.
 909          */
 910         if (needed > journal->j_max_transaction_buffers) {
 911                 CERROR("want too many journal credits (%d) using %d instead\n",
 912                        needed, journal->j_max_transaction_buffers);
 913                 needed = journal->j_max_transaction_buffers;
 914         }
 915
 916         handle = journal_start(journal, needed);
 917         if (IS_ERR(handle))
 918                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 919                        PTR_ERR(handle));
 920
 921         return(handle);
 922 }
 923
 924 static void *filter_journal_start(void **journal_save,
 925                                   struct filter_obd *filter,
 926                                   int objcount, struct obd_ioobj *obj,
 927                                   int niocount, struct niobuf_remote *nb)
 928 {
 929         void *handle = NULL;
 930
 931         /* This may not be necessary - we probably never have a
 932          * transaction started when we enter here, so we can
 933          * remove the saving of the journal state entirely.
 934          * For now leave it in just to see if it ever happens.
 935          */
 936         *journal_save = current->journal_info;
 937         if (*journal_save) {
 938                 CERROR("Already have handle %p???\n", *journal_save);
 939                 LBUG();
 940                 current->journal_info = NULL;
 941         }
 942
 943         if (!strcmp(filter->fo_fstype, "ext3") ||
 944             !strcmp(filter->fo_fstype, "extN"))
 945                 handle = ext3_filter_journal_start(filter, objcount, obj,
 946                                                    niocount, nb);
 947         return handle;
 948 }
 949
 950 static int ext3_filter_journal_stop(void *handle)
 951 {
 952         int rc;
 953
 954         /* We got a refcount on the handle for each call to prepare_write,
 955          * so we can drop the "parent" handle here to avoid the need for
 956          * osc to call back into filterobd to close the handle.  The
 957          * remaining references will be dropped in commit_write.
 958          */
 959         rc = journal_stop((handle_t *)handle);
 960
 961         return rc;
 962 }
 963
 964 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
 965                                void *handle)
 966 {
 967         int rc = 0;
 968
 969         if (!strcmp(filter->fo_fstype, "ext3") ||
 970             !strcmp(filter->fo_fstype, "extN"))
 971                 rc = ext3_filter_journal_stop(handle);
 972
 973         if (rc)
 974                 CERROR("error on journal stop: rc = %d\n", rc);
 975
 976         current->journal_info = journal_save;
 977
 978         return rc;
 979 }
 980
 981 static inline void lustre_put_page(struct page *page)
 982 {
 983         kunmap(page);
 984         page_cache_release(page);
 985 }
 986
 987 static struct page *
 988 lustre_get_page_read(struct inode *inode, unsigned long index)
 989 {
 990         struct address_space *mapping = inode->i_mapping;
 991         struct page *page;
 992         int rc;
 993
 994         page = read_cache_page(mapping, index,
 995                                (filler_t*)mapping->a_ops->readpage, NULL);
 996         if (!IS_ERR(page)) {
 997                 wait_on_page(page);
 998                 kmap(page);
 999                 if (!Page_Uptodate(page)) {
1000                         CERROR("page index %lu not uptodate\n", index);
1001                         GOTO(err_page, rc = -EIO);
1002                 }
1003                 if (PageError(page)) {
1004                         CERROR("page index %lu has error\n", index);
1005                         GOTO(err_page, rc = -EIO);
1006                 }
1007         }
1008         return page;
1009
1010 err_page:
1011         lustre_put_page(page);
1012         return ERR_PTR(rc);
1013 }
1014
1015 static struct page *
1016 lustre_get_page_write(struct inode *inode, unsigned long index)
1017 {
1018         struct address_space *mapping = inode->i_mapping;
1019         struct page *page;
1020         int rc;
1021
1022         page = grab_cache_page(mapping, index); /* locked page */
1023
1024         if (!IS_ERR(page)) {
1025                 kmap(page);
1026                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1027                  * a no-op for most filesystems, because we write the whole
1028                  * page.  For partial-page I/O this will read in the page.
1029                  */
1030                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1031                 if (rc) {
1032                         CERROR("page index %lu, rc = %d\n", index, rc);
1033                         if (rc != -ENOSPC)
1034                                 LBUG();
1035                         GOTO(err_unlock, rc);
1036                 }
1037                 /* XXX not sure if we need this if we are overwriting page */
1038                 if (PageError(page)) {
1039                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1040                         LBUG();
1041                         GOTO(err_unlock, rc = -EIO);
1042                 }
1043         }
1044         return page;
1045
1046 err_unlock:
1047         unlock_page(page);
1048         lustre_put_page(page);
1049         return ERR_PTR(rc);
1050 }
1051
1052 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1053 {
1054         struct inode *inode = page->mapping->host;
1055         int err;
1056
1057         err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1058         if (!err && IS_SYNC(inode))
1059                 err = waitfor_one_page(page);
1060
1061         //SetPageUptodate(page); // the client commit_write will do this
1062
1063         SetPageReferenced(page);
1064         unlock_page(page);
1065         lustre_put_page(page);
1066         return err;
1067 }
1068
1069 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1070                                    struct niobuf_local *lnb, int *pglocked)
1071 {
1072         struct address_space *mapping = inode->i_mapping;
1073         struct page *page;
1074         int rc;
1075
1076         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1077         if (*pglocked)
1078                 page = grab_cache_page_nowait(mapping, index); /* locked page */
1079         else
1080                 page = grab_cache_page(mapping, index); /* locked page */
1081
1082
1083         /* This page is currently locked, so get a temporary page instead. */
1084         /* XXX I believe this is a very dangerous thing to do - consider if
1085          *     we had multiple writers for the same file (definitely the case
1086          *     if we are using this codepath).  If writer A locks the page,
1087          *     writer B writes to a copy (as here), writer A drops the page
1088          *     lock, and writer C grabs the lock before B does, then B will
1089          *     later overwrite the data from C, even if C had LDLM locked
1090          *     and initiated the write after B did.
1091          */
1092         if (!page) {
1093                 unsigned long addr;
1094                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1095                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1096                 if (!addr) {
1097                         CERROR("no memory for a temp page\n");
1098                         LBUG();
1099                         GOTO(err, rc = -ENOMEM);
1100                 }
1101                 page = virt_to_page(addr);
1102                 kmap(page);
1103                 page->index = index;
1104                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1105         } else if (!IS_ERR(page)) {
1106                 (*pglocked)++;
1107                 kmap(page);
1108
1109                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1110                  * a no-op for most filesystems, because we write the whole
1111                  * page.  For partial-page I/O this will read in the page.
1112                  */
1113                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1114                 if (rc) {
1115                         CERROR("page index %lu, rc = %d\n", index, rc);
1116                         if (rc != -ENOSPC)
1117                                 LBUG();
1118                         GOTO(err_unlock, rc);
1119                 }
1120                 /* XXX not sure if we need this if we are overwriting page */
1121                 if (PageError(page)) {
1122                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1123                         LBUG();
1124                         GOTO(err_unlock, rc = -EIO);
1125                 }
1126         }
1127         return page;
1128
1129 err_unlock:
1130         unlock_page(page);
1131         lustre_put_page(page);
1132 err:
1133         return ERR_PTR(rc);
1134 }
1135
1136 /*
1137  * We need to balance prepare_write() calls with commit_write() calls.
1138  * If the page has been prepared, but we have no data for it, we don't
1139  * want to overwrite valid data on disk, but we still need to zero out
1140  * data for space which was newly allocated.  Like part of what happens
1141  * in __block_prepare_write() for newly allocated blocks.
1142  *
1143  * XXX currently __block_prepare_write() creates buffers for all the
1144  *     pages, and the filesystems mark these buffers as BH_New if they
1145  *     were newly allocated from disk. We use the BH_New flag similarly.
1146  */
1147 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1148                                int err)
1149 {
1150         if (err) {
1151                 unsigned block_start, block_end;
1152                 struct buffer_head *bh, *head = page->buffers;
1153                 unsigned blocksize = head->b_size;
1154                 void *addr = page_address(page);
1155
1156                 /* debugging: just seeing if this ever happens */
1157                 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1158                        page->index, page->mapping->host->i_ino, err);
1159
1160                 /* Currently one buffer per page, but in the future... */
1161                 for (bh = head, block_start = 0; bh != head || !block_start;
1162                      block_start = block_end, bh = bh->b_this_page) {
1163                         block_end = block_start + blocksize;
1164                         if (buffer_new(bh))
1165                                 memset(addr + block_start, 0, blocksize);
1166                 }
1167         }
1168
1169         return lustre_commit_write(page, from, to);
1170 }
1171
1172 static int filter_preprw(int cmd, struct lustre_handle *conn,
1173                          int objcount, struct obd_ioobj *obj,
1174                          int niocount, struct niobuf_remote *nb,
1175                          struct niobuf_local *res, void **desc_private)
1176 {
1177         struct obd_run_ctxt saved;
1178         struct obd_device *obd;
1179         struct obd_ioobj *o = obj;
1180         struct niobuf_remote *b = nb;
1181         struct niobuf_local *r = res;
1182         void *journal_save = NULL;
1183         int pglocked = 0;
1184         int rc = 0;
1185         int i;
1186         ENTRY;
1187
1188         obd = class_conn2obd(conn);
1189         if (!obd) {
1190                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1191                 RETURN(-EINVAL);
1192         }
1193         memset(res, 0, sizeof(*res) * niocount);
1194
1195         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1196
1197         if (cmd & OBD_BRW_WRITE) {
1198                 *desc_private = filter_journal_start(&journal_save,
1199                                                      &obd->u.filter,
1200                                                      objcount, obj, niocount,
1201                                                      nb);
1202                 if (IS_ERR(*desc_private))
1203                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1204         }
1205
1206         for (i = 0; i < objcount; i++, o++) {
1207                 struct dentry *dentry;
1208                 struct inode *inode;
1209                 int j;
1210
1211                 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1212                                            o->ioo_id, S_IFREG);
1213                 if (IS_ERR(dentry))
1214                         GOTO(out_clean, rc = PTR_ERR(dentry));
1215                 inode = dentry->d_inode;
1216                 if (!inode) {
1217                         CERROR("trying to BRW to non-existent file "LPU64"\n",
1218                                o->ioo_id);
1219                         f_dput(dentry);
1220                         GOTO(out_clean, rc = -ENOENT);
1221                 }
1222
1223                 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1224                         unsigned long index = b->offset >> PAGE_SHIFT;
1225                         struct page *page;
1226
1227                         if (j == 0)
1228                                 r->dentry = dentry;
1229                         else
1230                                 r->dentry = dget(dentry);
1231
1232                         if (cmd & OBD_BRW_WRITE)
1233                                 page = filter_get_page_write(inode, index, r,
1234                                                              &pglocked);
1235                         else
1236                                 page = lustre_get_page_read(inode, index);
1237
1238                         if (IS_ERR(page)) {
1239                                 f_dput(dentry);
1240                                 GOTO(out_clean, rc = PTR_ERR(page));
1241                         }
1242
1243                         r->addr = page_address(page);
1244                         r->offset = b->offset;
1245                         r->page = page;
1246                         r->len = b->len;
1247                 }
1248         }
1249
1250 out_stop:
1251         if (cmd & OBD_BRW_WRITE) {
1252                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1253                                               *desc_private);
1254                 if (!rc)
1255                         rc = err;
1256         }
1257 out_ctxt:
1258         pop_ctxt(&saved);
1259         RETURN(rc);
1260 out_clean:
1261         while (r-- > res) {
1262                 CERROR("error cleanup on brw\n");
1263                 f_dput(r->dentry);
1264                 if (cmd & OBD_BRW_WRITE)
1265                         filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1266                 else
1267                         lustre_put_page(r->page);
1268         }
1269         goto out_stop;
1270 }
1271
1272 static int filter_write_locked_page(struct niobuf_local *lnb)
1273 {
1274         struct page *lpage;
1275         int rc;
1276
1277         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1278         if (IS_ERR(lpage)) {
1279                 /* It is highly unlikely that we would ever get an error here.
1280                  * The page we want to get was previously locked, so it had to
1281                  * have already allocated the space, and we were just writing
1282                  * over the same data, so there would be no hole in the file.
1283                  *
1284                  * XXX: possibility of a race with truncate could exist, need
1285                  *      to check that.  There are no guarantees w.r.t.
1286                  *      write order even on a local filesystem, although the
1287                  *      normal response would be to return the number of bytes
1288                  *      successfully written and leave the rest to the app.
1289                  */
1290                 rc = PTR_ERR(lpage);
1291                 CERROR("error getting locked page index %ld: rc = %d\n",
1292                        lnb->page->index, rc);
1293                 GOTO(out, rc);
1294         }
1295
1296         /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1297          * lustre_commit_write() below, lnb->page was kmapped previously in
1298          * filter_get_page_write() and kunmapped in lustre_put_page() below.
1299          */
1300         memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1301         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1302         if (rc)
1303                 CERROR("error committing locked page %ld: rc = %d\n",
1304                        lnb->page->index, rc);
1305 out:
1306         lustre_put_page(lnb->page);
1307
1308         return rc;
1309 }
1310
1311 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1312                            int objcount, struct obd_ioobj *obj,
1313                            int niocount, struct niobuf_local *res,
1314                            void *private)
1315 {
1316         struct obd_run_ctxt saved;
1317         struct obd_ioobj *o;
1318         struct niobuf_local *r;
1319         struct obd_device *obd = class_conn2obd(conn);
1320         void *journal_save;
1321         int found_locked = 0;
1322         int rc = 0;
1323         int i;
1324         ENTRY;
1325
1326         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1327         journal_save = current->journal_info;
1328         LASSERT(!journal_save);
1329
1330         current->journal_info = private;
1331         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1332                 int j;
1333                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1334                         struct page *page = r->page;
1335
1336                         if (!page)
1337                                 LBUG();
1338
1339                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1340                                 found_locked++;
1341                                 continue;
1342                         }
1343
1344                         if (cmd & OBD_BRW_WRITE) {
1345                                 int err = filter_commit_write(page, 0,
1346                                                               r->len, 0);
1347
1348                                 if (!rc)
1349                                         rc = err;
1350                         } else
1351                                 lustre_put_page(page);
1352
1353                         f_dput(r->dentry);
1354                 }
1355         }
1356         current->journal_info = journal_save;
1357
1358         if (!found_locked)
1359                 goto out_ctxt;
1360
1361         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1362                 int j;
1363                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1364                         int err;
1365                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1366                                 continue;
1367
1368                         err = filter_write_locked_page(r);
1369                         if (!rc)
1370                                 rc = err;
1371                         f_dput(r->dentry);
1372                 }
1373         }
1374
1375 out_ctxt:
1376         pop_ctxt(&saved);
1377         RETURN(rc);
1378 }
1379
1380 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1381 {
1382         struct obd_device *obd = class_conn2obd(conn);
1383         struct statfs sfs;
1384         int rc;
1385
1386         ENTRY;
1387         rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1388         if (!rc)
1389                 statfs_pack(osfs, &sfs);
1390
1391         return rc;
1392 }
1393
1394 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1395                            void *key, obd_count *vallen, void **val)
1396 {
1397         struct obd_device *obd;
1398         ENTRY;
1399
1400         obd = class_conn2obd(conn);
1401         if (!obd) {
1402                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1403                 RETURN(-EINVAL);
1404         }
1405
1406         if ( keylen == strlen("blocksize") &&
1407              memcmp(key, "blocksize", keylen) == 0 ) {
1408                 *vallen = sizeof(long);
1409                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1410                 RETURN(0);
1411         }
1412
1413         if ( keylen == strlen("blocksize_bits") &&
1414              memcmp(key, "blocksize_bits", keylen) == 0 ){
1415                 *vallen = sizeof(long);
1416                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1417                 RETURN(0);
1418         }
1419
1420         if ( keylen == strlen("root_ino") &&
1421              memcmp(key, "root_ino", keylen) == 0 ){
1422                 *vallen = sizeof(obd_id);
1423                 *val = (void *)(obd_id)FILTER_ROOTINO;
1424                 RETURN(0);
1425         }
1426
1427         CDEBUG(D_IOCTL, "invalid key\n");
1428         RETURN(-EINVAL);
1429 }
1430
1431 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1432                   struct lustre_handle *src_conn, struct obdo *src,
1433                   obd_size count, obd_off offset)
1434 {
1435         struct page *page;
1436         struct lov_stripe_md srcmd, dstmd;
1437         unsigned long index = 0;
1438         int err = 0;
1439
1440         memset(&srcmd, 0, sizeof(srcmd));
1441         memset(&dstmd, 0, sizeof(dstmd));
1442         srcmd.lsm_object_id = src->o_id;
1443         dstmd.lsm_object_id = dst->o_id;
1444
1445         ENTRY;
1446         CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1447                ", dst: ino "LPU64"\n",
1448                src->o_id, src->o_blocks, src->o_size, dst->o_id);
1449         page = alloc_page(GFP_USER);
1450         if (page == NULL)
1451                 RETURN(-ENOMEM);
1452
1453         while (TryLockPage(page))
1454                 ___wait_on_page(page);
1455
1456         /* XXX with brw vector I/O, we could batch up reads and writes here,
1457          *     all we need to do is allocate multiple pages to handle the I/Os
1458          *     and arrays to handle the request parameters.
1459          */
1460         while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1461                 struct brw_page pg;
1462                 struct io_cb_data *cbd = ll_init_cb();
1463
1464                 if (!cbd) {
1465                         err = -ENOMEM;
1466                         EXIT;
1467                         break;
1468                 }
1469
1470                 pg.pg = page;
1471                 pg.count = PAGE_SIZE;
1472                 pg.off = (page->index) << PAGE_SHIFT;
1473                 pg.flag = 0;
1474
1475                 page->index = index;
1476                 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1477                               ll_sync_io_cb, cbd);
1478
1479                 if ( err ) {
1480                         EXIT;
1481                         break;
1482                 }
1483
1484                 cbd = ll_init_cb();
1485                 if (!cbd) {
1486                         err = -ENOMEM;
1487                         EXIT;
1488                         break;
1489                 }
1490                 pg.flag = OBD_BRW_CREATE;
1491                 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1492
1493                 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1494                               ll_sync_io_cb, cbd);
1495
1496                 /* XXX should handle dst->o_size, dst->o_blocks here */
1497                 if ( err ) {
1498                         EXIT;
1499                         break;
1500                 }
1501
1502                 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1503
1504                 index++;
1505         }
1506         dst->o_size = src->o_size;
1507         dst->o_blocks = src->o_blocks;
1508         dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1509         unlock_page(page);
1510         __free_page(page);
1511
1512         RETURN(err);
1513 }
1514
1515
1516 static struct obd_ops filter_obd_ops = {
1517         o_get_info:    filter_get_info,
1518         o_setup:       filter_setup,
1519         o_cleanup:     filter_cleanup,
1520         o_connect:     filter_connect,
1521         o_disconnect:  filter_disconnect,
1522         o_statfs:      filter_statfs,
1523         o_getattr:     filter_getattr,
1524         o_create:      filter_create,
1525         o_setattr:     filter_setattr,
1526         o_destroy:     filter_destroy,
1527         o_open:        filter_open,
1528         o_close:       filter_close,
1529         o_brw:         filter_pgcache_brw,
1530         o_punch:       filter_truncate,
1531         o_preprw:      filter_preprw,
1532         o_commitrw:    filter_commitrw
1533 #if 0
1534         o_preallocate: filter_preallocate_inodes,
1535         o_migrate:     filter_migrate,
1536         o_copy:        filter_copy_data,
1537         o_iterate:     filter_iterate
1538 #endif
1539 };
1540
1541
1542 static int __init obdfilter_init(void)
1543 {
1544         printk(KERN_INFO "Filtering OBD driver  v0.001, info@clusterfs.com\n");
1545         return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1546 }
1547
1548 static void __exit obdfilter_exit(void)
1549 {
1550         class_unregister_type(OBD_FILTER_DEVICENAME);
1551 }
1552
1553 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1554 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1555 MODULE_LICENSE("GPL");
1556
1557 module_init(obdfilter_init);
1558 module_exit(obdfilter_exit);