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/fs.h>
  21 #include <linux/dcache.h>
  22 #include <linux/obd_class.h>
  23 #include <linux/lustre_dlm.h>
  24 #include <linux/obd_filter.h>
  25 #include <linux/ext3_jbd.h>
  26 #include <linux/extN_jbd.h>
  27 #include <linux/quotaops.h>
  28 #include <linux/init.h>
  29 #include <linux/stringify.h>
  30
  31 long filter_memory;
  32
  33 #define FILTER_ROOTINO 2
  34 #define FILTER_ROOTINO_STR __stringify(FILTER_ROOTINO)
  35
  36 #define S_SHIFT 12
  37 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
  38         [0]                     NULL,
  39         [S_IFREG >> S_SHIFT]    "R",
  40         [S_IFDIR >> S_SHIFT]    "D",
  41         [S_IFCHR >> S_SHIFT]    "C",
  42         [S_IFBLK >> S_SHIFT]    "B",
  43         [S_IFIFO >> S_SHIFT]    "F",
  44         [S_IFSOCK >> S_SHIFT]   "S",
  45         [S_IFLNK >> S_SHIFT]    "L"
  46 };
  47
  48 static inline const char *obd_mode_to_type(int mode)
  49 {
  50         return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
  51 }
  52
  53 /* write the pathname into the string */
  54 static int filter_id(char *buf, obd_id id, obd_mode mode)
  55 {
  56         return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
  57 }
  58
  59 static inline void f_dput(struct dentry *dentry)
  60 {
  61         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
  62                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
  63         LASSERT(atomic_read(&dentry->d_count) > 0);
  64         dput(dentry);
  65 }
  66
  67 /* setup the object store with correct subdirectories */
  68 static int filter_prep(struct obd_device *obddev)
  69 {
  70         struct obd_run_ctxt saved;
  71         struct filter_obd *filter = &obddev->u.filter;
  72         struct dentry *dentry;
  73         struct dentry *root;
  74         struct file *file;
  75         struct inode *inode;
  76         int rc = 0;
  77         __u64 lastino = 2;
  78         int mode = 0;
  79
  80         push_ctxt(&saved, &filter->fo_ctxt, NULL);
  81         dentry = simple_mkdir(current->fs->pwd, "O", 0700);
  82         CDEBUG(D_INODE, "got/created O: %p\n", dentry);
  83         if (IS_ERR(dentry)) {
  84                 rc = PTR_ERR(dentry);
  85                 CERROR("cannot open/create O: rc = %d\n", rc);
  86                 GOTO(out, rc);
  87         }
  88         filter->fo_dentry_O = dentry;
  89         dentry = simple_mkdir(current->fs->pwd, "P", 0700);
  90         CDEBUG(D_INODE, "got/created P: %p\n", dentry);
  91         if (IS_ERR(dentry)) {
  92                 rc = PTR_ERR(dentry);
  93                 CERROR("cannot open/create P: rc = %d\n", rc);
  94                 GOTO(out_O, rc);
  95         }
  96         f_dput(dentry);
  97         dentry = simple_mkdir(current->fs->pwd, "D", 0700);
  98         CDEBUG(D_INODE, "got/created D: %p\n", dentry);
  99         if (IS_ERR(dentry)) {
 100                 rc = PTR_ERR(dentry);
 101                 CERROR("cannot open/create D: rc = %d\n", rc);
 102                 GOTO(out_O, rc);
 103         }
 104
 105         root = simple_mknod(dentry, FILTER_ROOTINO_STR, S_IFREG | 0755);
 106         f_dput(dentry);
 107         if (IS_ERR(root)) {
 108                 rc = PTR_ERR(root);
 109                 CERROR("OBD filter: cannot open/create root %d: rc = %d\n",
 110                        FILTER_ROOTINO, rc);
 111                 GOTO(out_O, rc);
 112         }
 113         f_dput(root);
 114
 115         /*
 116          * Create directories and/or get dentries for each object type.
 117          * This saves us from having to do multiple lookups for each one.
 118          */
 119         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 120                 char *type = obd_type_by_mode[mode];
 121
 122                 if (!type) {
 123                         filter->fo_dentry_O_mode[mode] = NULL;
 124                         continue;
 125                 }
 126                 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
 127                 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
 128                 if (IS_ERR(dentry)) {
 129                         rc = PTR_ERR(dentry);
 130                         CERROR("cannot create O/%s: rc = %d\n", type, rc);
 131                         GOTO(out_O_mode, rc);
 132                 }
 133                 filter->fo_dentry_O_mode[mode] = dentry;
 134         }
 135
 136         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 137         if ( !file || IS_ERR(file) ) {
 138                 rc = PTR_ERR(file);
 139                 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
 140                        "D/status", rc);
 141                 GOTO(out_O_mode, rc);
 142         }
 143
 144         /* steal operations */
 145         inode = file->f_dentry->d_inode;
 146         filter->fo_fop = file->f_op;
 147         filter->fo_iop = inode->i_op;
 148         filter->fo_aops = inode->i_mapping->a_ops;
 149
 150         if (inode->i_size == 0) {
 151                 __u64 disk_lastino = cpu_to_le64(lastino);
 152                 ssize_t retval = file->f_op->write(file, (char *)&disk_lastino,
 153                                                    sizeof(disk_lastino),
 154                                                    &file->f_pos);
 155                 if (retval != sizeof(disk_lastino)) {
 156                         CDEBUG(D_INODE, "OBD filter: error writing lastino\n");
 157                         filp_close(file, 0);
 158                         GOTO(out_O_mode, rc = -EIO);
 159                 }
 160         } else {
 161                 __u64 disk_lastino;
 162                 ssize_t retval = file->f_op->read(file, (char *)&disk_lastino,
 163                                                   sizeof(disk_lastino),
 164                                                   &file->f_pos);
 165                 if (retval != sizeof(disk_lastino)) {
 166                         CDEBUG(D_INODE, "OBD filter: error reading lastino\n");
 167                         filp_close(file, 0);
 168                         GOTO(out_O_mode, rc = -EIO);
 169                 }
 170                 lastino = le64_to_cpu(disk_lastino);
 171         }
 172         filter->fo_lastino = lastino;
 173         filp_close(file, 0);
 174
 175         rc = 0;
 176  out:
 177         pop_ctxt(&saved);
 178
 179         return(rc);
 180
 181 out_O_mode:
 182         while (mode-- > 0) {
 183                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 184                 if (dentry) {
 185                         f_dput(dentry);
 186                         filter->fo_dentry_O_mode[mode] = NULL;
 187                 }
 188         }
 189 out_O:
 190         f_dput(filter->fo_dentry_O);
 191         filter->fo_dentry_O = NULL;
 192         goto out;
 193 }
 194
 195 /* cleanup the filter: write last used object id to status file */
 196 static void filter_post(struct obd_device *obddev)
 197 {
 198         struct obd_run_ctxt saved;
 199         struct filter_obd *filter = &obddev->u.filter;
 200         __u64 disk_lastino;
 201         long rc;
 202         struct file *file;
 203         int mode;
 204
 205         push_ctxt(&saved, &filter->fo_ctxt, NULL);
 206         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 207         if (IS_ERR(file)) {
 208                 CERROR("OBD filter: cannot create status file\n");
 209                 goto out;
 210         }
 211
 212         file->f_pos = 0;
 213         disk_lastino = cpu_to_le64(filter->fo_lastino);
 214         rc = file->f_op->write(file, (char *)&disk_lastino,
 215                        sizeof(disk_lastino), &file->f_pos);
 216         if (rc != sizeof(disk_lastino))
 217                 CERROR("OBD filter: error writing lastino: rc = %ld\n", rc);
 218
 219         rc = filp_close(file, NULL);
 220         if (rc)
 221                 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
 222
 223         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 224                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 225                 if (dentry) {
 226                         f_dput(dentry);
 227                         filter->fo_dentry_O_mode[mode] = NULL;
 228                 }
 229         }
 230         f_dput(filter->fo_dentry_O);
 231 out:
 232         pop_ctxt(&saved);
 233 }
 234
 235
 236 static __u64 filter_next_id(struct obd_device *obddev)
 237 {
 238         obd_id id;
 239
 240         spin_lock(&obddev->u.filter.fo_lock);
 241         id = ++obddev->u.filter.fo_lastino;
 242         spin_unlock(&obddev->u.filter.fo_lock);
 243
 244         /* FIXME: write the lastino to disk here */
 245         return id;
 246 }
 247
 248 /* how to get files, dentries, inodes from object id's */
 249 /* parent i_sem is already held if needed for exclusivity */
 250 static struct dentry *filter_fid2dentry(struct obd_device *obddev,
 251                                         struct dentry *dparent,
 252                                         __u64 id, __u32 type)
 253 {
 254         struct super_block *sb = obddev->u.filter.fo_sb;
 255         struct dentry *dchild;
 256         char name[32];
 257         int len;
 258         ENTRY;
 259
 260         if (!sb || !sb->s_dev) {
 261                 CERROR("fatal: device not initialized.\n");
 262                 RETURN(ERR_PTR(-ENXIO));
 263         }
 264
 265         if (id == 0) {
 266                 CERROR("fatal: invalid object #0\n");
 267                 LBUG();
 268                 RETURN(ERR_PTR(-ESTALE));
 269         }
 270
 271         if (!(type & S_IFMT)) {
 272                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 273                        __FUNCTION__, id, type);
 274                 RETURN(ERR_PTR(-EINVAL));
 275         }
 276
 277         len = sprintf(name, LPU64, id);
 278         CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
 279                name);
 280         dchild = lookup_one_len(name, dparent, len);
 281         if (IS_ERR(dchild)) {
 282                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 283                 RETURN(dchild);
 284         }
 285
 286         CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
 287                obd_mode_to_type(type), name, dchild,
 288                atomic_read(&dchild->d_count));
 289
 290         LASSERT(atomic_read(&dchild->d_count) > 0);
 291
 292         RETURN(dchild);
 293 }
 294
 295 static struct file *filter_obj_open(struct obd_device *obddev,
 296                                     __u64 id, __u32 type)
 297 {
 298         struct super_block *sb = obddev->u.filter.fo_sb;
 299         struct obd_run_ctxt saved;
 300         char name[24];
 301         struct file *file;
 302         ENTRY;
 303
 304         if (!sb || !sb->s_dev) {
 305                 CERROR("fatal: device not initialized.\n");
 306                 RETURN(ERR_PTR(-ENXIO));
 307         }
 308
 309         if (!id) {
 310                 CERROR("fatal: invalid obdo "LPU64"\n", id);
 311                 RETURN(ERR_PTR(-ESTALE));
 312         }
 313
 314         if (!(type & S_IFMT)) {
 315                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 316                        __FUNCTION__, id, type);
 317                 RETURN(ERR_PTR(-EINVAL));
 318         }
 319
 320         filter_id(name, id, type);
 321         push_ctxt(&saved, &obddev->u.filter.fo_ctxt, NULL);
 322         file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
 323         pop_ctxt(&saved);
 324
 325         CDEBUG(D_INODE, "opening obdo %s: rc = %p\n", name, file);
 326
 327         if (IS_ERR(file))
 328                 file = NULL;
 329         RETURN(file);
 330 }
 331
 332 static struct dentry *filter_parent(struct obd_device *obddev, obd_mode mode)
 333 {
 334         struct filter_obd *filter = &obddev->u.filter;
 335
 336         return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
 337 }
 338
 339
 340 /* obd methods */
 341 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
 342                           obd_uuid_t cluuid, struct recovd_obd *recovd,
 343                           ptlrpc_recovery_cb_t recover)
 344 {
 345         int rc;
 346         ENTRY;
 347         MOD_INC_USE_COUNT;
 348         rc = class_connect(conn, obd, cluuid);
 349         if (rc)
 350                 MOD_DEC_USE_COUNT;
 351         RETURN(rc);
 352 }
 353
 354 static int filter_disconnect(struct lustre_handle *conn)
 355 {
 356         int rc;
 357         ENTRY;
 358
 359         rc = class_disconnect(conn);
 360         if (!rc)
 361                 MOD_DEC_USE_COUNT;
 362
 363         /* XXX cleanup preallocated inodes */
 364         RETURN(rc);
 365 }
 366
 367 /* mount the file system (secretly) */
 368 static int filter_setup(struct obd_device *obddev, obd_count len, void *buf)
 369 {
 370         struct obd_ioctl_data* data = buf;
 371         struct filter_obd *filter;
 372         struct vfsmount *mnt;
 373         int err = 0;
 374         ENTRY;
 375
 376         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 377                 RETURN(-EINVAL);
 378
 379         MOD_INC_USE_COUNT;
 380         mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
 381         err = PTR_ERR(mnt);
 382         if (IS_ERR(mnt))
 383                 GOTO(err_dec, err);
 384
 385         filter = &obddev->u.filter;;
 386         filter->fo_vfsmnt = mnt;
 387         filter->fo_fstype = strdup(data->ioc_inlbuf2);
 388         filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
 389         /* XXX is this even possible if do_kern_mount succeeded? */
 390         if (!filter->fo_sb)
 391                 GOTO(err_kfree, err = -ENODEV);
 392
 393         OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
 394         filter->fo_ctxt.pwdmnt = mnt;
 395         filter->fo_ctxt.pwd = mnt->mnt_root;
 396         filter->fo_ctxt.fs = get_ds();
 397
 398         err = filter_prep(obddev);
 399         if (err)
 400                 GOTO(err_kfree, err);
 401         spin_lock_init(&filter->fo_lock);
 402
 403         obddev->obd_namespace =
 404                 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
 405         if (obddev->obd_namespace == NULL)
 406                 LBUG();
 407
 408         ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
 409                            "filter_ldlm_client", &obddev->obd_ldlm_client);
 410
 411         RETURN(0);
 412
 413 err_kfree:
 414         kfree(filter->fo_fstype);
 415         unlock_kernel();
 416         mntput(filter->fo_vfsmnt);
 417         filter->fo_sb = 0;
 418         lock_kernel();
 419
 420 err_dec:
 421         MOD_DEC_USE_COUNT;
 422         return err;
 423 }
 424
 425
 426 static int filter_cleanup(struct obd_device * obddev)
 427 {
 428         struct super_block *sb;
 429         ENTRY;
 430
 431         if (!list_empty(&obddev->obd_exports)) {
 432                 CERROR("still has clients!\n");
 433                 class_disconnect_all(obddev);
 434                 if (!list_empty(&obddev->obd_exports)) {
 435                         CERROR("still has exports after forced cleanup?\n");
 436                         RETURN(-EBUSY);
 437                 }
 438         }
 439
 440         ldlm_namespace_free(obddev->obd_namespace);
 441
 442         sb = obddev->u.filter.fo_sb;
 443         if (!obddev->u.filter.fo_sb)
 444                 RETURN(0);
 445
 446         filter_post(obddev);
 447
 448         shrink_dcache_parent(sb->s_root);
 449         unlock_kernel();
 450         mntput(obddev->u.filter.fo_vfsmnt);
 451         obddev->u.filter.fo_sb = 0;
 452         kfree(obddev->u.filter.fo_fstype);
 453
 454         lock_kernel();
 455
 456         MOD_DEC_USE_COUNT;
 457         RETURN(0);
 458 }
 459
 460
 461 static inline void filter_from_inode(struct obdo *oa, struct inode *inode,
 462                                      int valid)
 463 {
 464         int type = oa->o_mode & S_IFMT;
 465         ENTRY;
 466
 467         CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
 468                inode->i_ino, inode, (long)oa->o_id, valid);
 469         /* Don't copy the inode number in place of the object ID */
 470         obdo_from_inode(oa, inode, valid);
 471         oa->o_mode &= ~S_IFMT;
 472         oa->o_mode |= type;
 473
 474         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
 475                 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
 476                 oa->o_rdev = rdev;
 477                 oa->o_valid |= OBD_MD_FLRDEV;
 478         }
 479
 480         EXIT;
 481 }
 482
 483 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
 484                           struct lov_stripe_md *md)
 485 {
 486         struct obd_device *obddev = class_conn2obd(conn);
 487         struct dentry *dentry;
 488         int rc = 0;
 489         ENTRY;
 490
 491         if (!class_conn2export(conn)) {
 492                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 493                 RETURN(-EINVAL);
 494         }
 495
 496         obddev = class_conn2obd(conn);
 497         dentry = filter_fid2dentry(obddev, filter_parent(obddev, oa->o_mode),
 498                                    oa->o_id, oa->o_mode);
 499         if (IS_ERR(dentry))
 500                 RETURN(PTR_ERR(dentry));
 501
 502         if (!dentry->d_inode) {
 503                 CERROR("getattr on non-existent object: "LPU64"\n", oa->o_id);
 504                 GOTO(out_getattr, rc = -ENOENT);
 505         }
 506
 507         filter_from_inode(oa, dentry->d_inode, oa->o_valid);
 508
 509 out_getattr:
 510         f_dput(dentry);
 511         RETURN(rc);
 512 }
 513
 514 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
 515                           struct lov_stripe_md *md)
 516 {
 517         struct obd_run_ctxt saved;
 518         struct obd_device *obd = class_conn2obd(conn);
 519         struct dentry *dentry;
 520         struct iattr iattr;
 521         struct inode *inode;
 522         int rc;
 523         ENTRY;
 524
 525         iattr_from_obdo(&iattr, oa, oa->o_valid);
 526         iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
 527         dentry = filter_fid2dentry(obd, filter_parent(obd, iattr.ia_mode),
 528                                    oa->o_id, iattr.ia_mode);
 529         if (IS_ERR(dentry))
 530                 RETURN(PTR_ERR(dentry));
 531
 532         inode = dentry->d_inode;
 533         if (!inode) {
 534                 CERROR("setattr on non-existent object: "LPU64"\n", oa->o_id);
 535                 GOTO(out_setattr, rc = -ENOENT);
 536         }
 537
 538         lock_kernel();
 539         if (iattr.ia_valid & ATTR_SIZE)
 540                 down(&inode->i_sem);
 541         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 542         if (inode->i_op->setattr)
 543                 rc = inode->i_op->setattr(dentry, &iattr);
 544         else
 545                 rc = inode_setattr(inode, &iattr);
 546         pop_ctxt(&saved);
 547         if (iattr.ia_valid & ATTR_SIZE) {
 548                 up(&inode->i_sem);
 549                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 550                 obdo_from_inode(oa, inode, oa->o_valid);
 551         }
 552         unlock_kernel();
 553
 554 out_setattr:
 555         f_dput(dentry);
 556         RETURN(rc);
 557 }
 558
 559 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
 560                        struct lov_stripe_md *ea)
 561 {
 562         struct obd_export *export;
 563         struct obd_device *obd;
 564         struct dentry *dentry;
 565         /* ENTRY; */
 566
 567         export = class_conn2export(conn);
 568         if (!export) {
 569                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 570                 RETURN(-EINVAL);
 571         }
 572
 573         obd = export->exp_obd;
 574         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 575                                    oa->o_id, oa->o_mode);
 576         if (IS_ERR(dentry))
 577                 RETURN(PTR_ERR(dentry));
 578
 579         if (!dentry->d_inode) {
 580                 CERROR("opening non-existent objid "LPX64"\n", oa->o_id);
 581                 RETURN(-ENOENT);
 582         }
 583
 584         filter_from_inode(oa, dentry->d_inode, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 585                           OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 586
 587         return 0;
 588 } /* filter_open */
 589
 590 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
 591                         struct lov_stripe_md *ea)
 592 {
 593         struct obd_device *obd;
 594         struct dentry *dentry;
 595         /* ENTRY; */
 596
 597         obd = class_conn2obd(conn);
 598         if (!obd) {
 599                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 600                 RETURN(-EINVAL);
 601         }
 602
 603         dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 604                                    oa->o_id, oa->o_mode);
 605         if (IS_ERR(dentry))
 606                 RETURN(PTR_ERR(dentry));
 607         LASSERT(atomic_read(&dentry->d_count) > 1);
 608
 609         f_dput(dentry);  /* for the close */
 610         f_dput(dentry);  /* for this call */
 611         return 0;
 612 } /* filter_close */
 613
 614 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
 615                          struct lov_stripe_md **ea)
 616 {
 617         char name[64];
 618         struct obd_run_ctxt saved;
 619         struct dentry *new;
 620         int mode;
 621         struct obd_device *obd = class_conn2obd(conn);
 622         struct filter_obd *filter = &obd->u.filter;
 623         struct iattr;
 624         ENTRY;
 625
 626         if (!obd) {
 627                 CERROR("invalid client "LPX64"\n", conn->addr);
 628                 return -EINVAL;
 629         }
 630
 631         if (!(oa->o_mode && S_IFMT)) {
 632                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 633                        __FUNCTION__, oa->o_id, oa->o_mode);
 634                 return -ENOENT;
 635         }
 636
 637         oa->o_id = filter_next_id(obd);
 638
 639         //filter_id(name, oa->o_id, oa->o_mode);
 640         sprintf(name, LPU64, oa->o_id);
 641         mode = (oa->o_mode & ~S_IFMT) | S_IFREG;
 642         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 643         new = simple_mknod(filter->fo_dentry_O_mode[S_IFREG >> S_SHIFT], name,
 644                            mode);
 645         pop_ctxt(&saved);
 646         if (IS_ERR(new)) {
 647                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
 648                 return -ENOENT;
 649         }
 650
 651         /* Set flags for fields we have set in the inode struct */
 652         oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 653                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
 654         filter_from_inode(oa, new->d_inode, oa->o_valid);
 655         dput(new);
 656
 657         return 0;
 658 }
 659
 660 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
 661                           struct lov_stripe_md *ea)
 662 {
 663         struct obd_device *obd;
 664         struct filter_obd *filter;
 665         struct obd_run_ctxt saved;
 666         struct inode *inode;
 667         struct dentry *dir_dentry, *object_dentry;
 668         int rc;
 669         ENTRY;
 670
 671         obd = class_conn2obd(conn);
 672         if (!obd) {
 673                 CERROR("invalid client "LPX64"\n", conn->addr);
 674                 RETURN(-EINVAL);
 675         }
 676
 677         CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
 678
 679         dir_dentry = filter_parent(obd, oa->o_mode);
 680         down(&dir_dentry->d_inode->i_sem);
 681
 682         object_dentry = filter_fid2dentry(obd, dir_dentry, oa->o_id,
 683                                           oa->o_mode);
 684         if (IS_ERR(object_dentry))
 685                 GOTO(out, rc = -ENOENT);
 686
 687         inode = object_dentry->d_inode;
 688         if (!inode) {
 689                 CERROR("trying to destroy negative inode "LPX64"!\n", oa->o_id);
 690                 GOTO(out, rc = -ENOENT);
 691         }
 692
 693         if (inode->i_nlink != 1) {
 694                 CERROR("destroying inode with nlink = %d\n", inode->i_nlink);
 695                 LBUG();
 696                 inode->i_nlink = 1;
 697         }
 698         inode->i_mode = S_IFREG;
 699
 700         if (atomic_read(&inode->i_count) > 1) {
 701 #warning FIXME: need to handle open-unlinked case and move to hold dir
 702                 CERROR("inode has count %d\n", atomic_read(&inode->i_count));
 703         }
 704
 705         filter = &obd->u.filter;
 706         push_ctxt(&saved, &filter->fo_ctxt, NULL);
 707
 708         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 709         pop_ctxt(&saved);
 710         f_dput(object_dentry);
 711
 712         EXIT;
 713 out:
 714         up(&dir_dentry->d_inode->i_sem);
 715         return rc;
 716 }
 717
 718 /* NB count and offset are used for punch, but not truncate */
 719 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
 720                            struct lov_stripe_md *md,
 721                            obd_off start, obd_off end)
 722 {
 723         int error;
 724         ENTRY;
 725
 726         if (end != OBD_OBJECT_EOF)
 727                 CERROR("PUNCH not supported, only truncate works\n");
 728
 729         CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
 730                "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
 731         oa->o_size = start;
 732         error = filter_setattr(conn, oa, NULL);
 733         RETURN(error);
 734 }
 735
 736 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
 737                               struct lov_stripe_md *lsm, obd_count oa_bufs,
 738                               struct brw_page *pga, brw_callback_t callback,
 739                               struct io_cb_data *data)
 740 {
 741         struct obd_run_ctxt      saved;
 742         struct super_block      *sb;
 743         int                      pnum;          /* index to pages (bufs) */
 744         unsigned long            retval;
 745         int                      error;
 746         struct file             *file;
 747         struct obd_device       *obd = class_conn2obd(conn);
 748         int pg;
 749         ENTRY;
 750
 751         if (!obd) {
 752                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
 753                 RETURN(-EINVAL);
 754         }
 755
 756         sb = obd->u.filter.fo_sb;
 757         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 758         pnum = 0; /* pnum indexes buf 0..num_pages */
 759
 760         file = filter_obj_open(obd, lsm->lsm_object_id, S_IFREG);
 761         if (IS_ERR(file))
 762                 GOTO(out, retval = PTR_ERR(file));
 763
 764         /* count doubles as retval */
 765         for (pg = 0; pg < oa_bufs; pg++) {
 766                 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
 767                        ") off count ("LPU64",%d)\n",
 768                        cmd, pnum, file->f_dentry->d_inode->i_ino,
 769                        pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
 770                        (int)pga[pnum].count);
 771                 if (cmd & OBD_BRW_WRITE) {
 772                         loff_t off;
 773                         char *buffer;
 774                         off = pga[pnum].off;
 775                         buffer = kmap(pga[pnum].pg);
 776                         retval = file->f_op->write(file, buffer,
 777                                                    pga[pnum].count,
 778                                                    &off);
 779                         kunmap(pga[pnum].pg);
 780                         CDEBUG(D_INODE, "retval %ld\n", retval);
 781                 } else {
 782                         loff_t off = pga[pnum].off;
 783                         char *buffer = kmap(pga[pnum].pg);
 784
 785                         if (off >= file->f_dentry->d_inode->i_size) {
 786                                 memset(buffer, 0, pga[pnum].count);
 787                                 retval = pga[pnum].count;
 788                         } else {
 789                                 retval = file->f_op->read(file, buffer,
 790                                                           pga[pnum].count, &off);
 791                         }
 792                         kunmap(pga[pnum].pg);
 793
 794                         if (retval != pga[pnum].count) {
 795                                 filp_close(file, 0);
 796                                 GOTO(out, retval = -EIO);
 797                         }
 798                         CDEBUG(D_INODE, "retval %ld\n", retval);
 799                 }
 800                 pnum++;
 801         }
 802         /* sizes and blocks are set by generic_file_write */
 803         /* ctimes/mtimes will follow with a setattr call */
 804         filp_close(file, 0);
 805
 806         /* XXX: do something with callback if it is set? */
 807
 808         EXIT;
 809 out:
 810         pop_ctxt(&saved);
 811         error = (retval >= 0) ? 0 : retval;
 812         return error;
 813 }
 814
 815 /*
 816  * Calculate the number of buffer credits needed to write multiple pages in
 817  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
 818  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
 819  *
 820  * See comment above ext3_writepage_trans_blocks for details.  We assume
 821  * no data journaling is being done, but it does allow for all of the pages
 822  * being non-contiguous.  If we are guaranteed contiguous pages we could
 823  * reduce the number of (d)indirect blocks a lot.
 824  *
 825  * With N blocks per page and P pages, for each inode we have at most:
 826  * N*P indirect
 827  * min(N*P, blocksize/4 + 1) dindirect blocks
 828  * 1 tindirect
 829  *
 830  * For the entire filesystem, we have at most:
 831  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
 832  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
 833  * 1 inode block
 834  * 1 superblock
 835  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
 836  */
 837 static int ext3_credits_needed(struct super_block *sb, int objcount,
 838                                struct obd_ioobj *obj)
 839 {
 840         struct obd_ioobj *o = obj;
 841         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
 842         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
 843         int nbitmaps = 0;
 844         int ngdblocks = 0;
 845         int needed = objcount + 1;
 846         int i;
 847
 848         for (i = 0; i < objcount; i++, o++) {
 849                 int nblocks = o->ioo_bufcnt * blockpp;
 850                 int ndindirect = min(nblocks, addrpp + 1);
 851                 int nindir = nblocks + ndindirect + 1;
 852
 853                 nbitmaps += nindir + nblocks;
 854                 ngdblocks += nindir + nblocks;
 855
 856                 needed += nindir;
 857         }
 858
 859         /* Assumes ext3 and extN have same sb_info layout at the start. */
 860         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
 861                 nbitmaps = EXT3_SB(sb)->s_groups_count;
 862         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
 863                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
 864
 865         needed += nbitmaps + ngdblocks;
 866
 867 #ifdef CONFIG_QUOTA
 868         /* We assume that there will be 1 bit set in s_dquot.flags for each
 869          * quota file that is active.  This is at least true for now.
 870          */
 871         needed += hweight32(sb_any_quota_enabled(sb)) *
 872                 EXT3_SINGLEDATA_TRANS_BLOCKS;
 873 #endif
 874
 875         return needed;
 876 }
 877
 878 /* We have to start a huge journal transaction here to hold all of the
 879  * metadata for the pages being written here.  This is necessitated by
 880  * the fact that we do lots of prepare_write operations before we do
 881  * any of the matching commit_write operations, so even if we split
 882  * up to use "smaller" transactions none of them could complete until
 883  * all of them were opened.  By having a single journal transaction,
 884  * we eliminate duplicate reservations for common blocks like the
 885  * superblock and group descriptors or bitmaps.
 886  *
 887  * We will start the transaction here, but each prepare_write will
 888  * add a refcount to the transaction, and each commit_write will
 889  * remove a refcount.  The transaction will be closed when all of
 890  * the pages have been written.
 891  */
 892 static void *ext3_filter_journal_start(struct filter_obd *filter,
 893                                        int objcount, struct obd_ioobj *obj,
 894                                        int niocount, struct niobuf_remote *nb)
 895 {
 896         journal_t *journal = NULL;
 897         handle_t *handle = NULL;
 898         int needed;
 899
 900         /* It appears that some kernels have different values for
 901          * EXT*_MAX_GROUP_LOADED (either 8 or 32), so we cannot
 902          * assume anything after s_inode_bitmap_number is the same.
 903          */
 904         if (!strcmp(filter->fo_fstype, "ext3"))
 905                 journal = EXT3_SB(filter->fo_sb)->s_journal;
 906         else if (!strcmp(filter->fo_fstype, "extN"))
 907                 journal = EXTN_SB(filter->fo_sb)->s_journal;
 908         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
 909
 910         /* The number of blocks we could _possibly_ dirty can very large.
 911          * We reduce our request if it is absurd (and we couldn't get that
 912          * many credits for a single handle anyways).
 913          *
 914          * At some point we have to limit the size of I/Os sent at one time,
 915          * increase the size of the journal, or we have to calculate the
 916          * actual journal requirements more carefully by checking all of
 917          * the blocks instead of being maximally pessimistic.  It remains to
 918          * be seen if this is a real problem or not.
 919          */
 920         if (needed > journal->j_max_transaction_buffers) {
 921                 CERROR("want too many journal credits (%d) using %d instead\n",
 922                        needed, journal->j_max_transaction_buffers);
 923                 needed = journal->j_max_transaction_buffers;
 924         }
 925
 926         lock_kernel();
 927         handle = journal_start(journal, needed);
 928         unlock_kernel();
 929         if (IS_ERR(handle))
 930                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
 931                        PTR_ERR(handle));
 932
 933         return(handle);
 934 }
 935
 936 static void *filter_journal_start(void **journal_save,
 937                                   struct filter_obd *filter,
 938                                   int objcount, struct obd_ioobj *obj,
 939                                   int niocount, struct niobuf_remote *nb)
 940 {
 941         void *handle = NULL;
 942
 943         /* This may not be necessary - we probably never have a
 944          * transaction started when we enter here, so we can
 945          * remove the saving of the journal state entirely.
 946          * For now leave it in just to see if it ever happens.
 947          */
 948         *journal_save = current->journal_info;
 949         if (*journal_save) {
 950                 CERROR("Already have handle %p???\n", *journal_save);
 951                 LBUG();
 952                 current->journal_info = NULL;
 953         }
 954
 955         if (!strcmp(filter->fo_fstype, "ext3") ||
 956             !strcmp(filter->fo_fstype, "extN"))
 957                 handle = ext3_filter_journal_start(filter, objcount, obj,
 958                                                    niocount, nb);
 959         return handle;
 960 }
 961
 962 static int ext3_filter_journal_stop(void *handle)
 963 {
 964         int rc;
 965
 966         /* We got a refcount on the handle for each call to prepare_write,
 967          * so we can drop the "parent" handle here to avoid the need for
 968          * osc to call back into filterobd to close the handle.  The
 969          * remaining references will be dropped in commit_write.
 970          */
 971         lock_kernel();
 972         rc = journal_stop((handle_t *)handle);
 973         unlock_kernel();
 974
 975         return rc;
 976 }
 977
 978 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
 979                                void *handle)
 980 {
 981         int rc = 0;
 982
 983         if (!strcmp(filter->fo_fstype, "ext3") ||
 984             !strcmp(filter->fo_fstype, "extN"))
 985                 rc = ext3_filter_journal_stop(handle);
 986
 987         if (rc)
 988                 CERROR("error on journal stop: rc = %d\n", rc);
 989
 990         current->journal_info = journal_save;
 991
 992         return rc;
 993 }
 994
 995 static inline void lustre_put_page(struct page *page)
 996 {
 997         kunmap(page);
 998         page_cache_release(page);
 999 }
1000
1001 #ifndef PageUptodate
1002 #define PageUptodate(page) Page_Uptodate(page)
1003 #endif
1004
1005 static struct page *
1006 lustre_get_page_read(struct inode *inode, unsigned long index)
1007 {
1008         struct address_space *mapping = inode->i_mapping;
1009         struct page *page;
1010         int rc;
1011
1012         page = read_cache_page(mapping, index,
1013                                (filler_t*)mapping->a_ops->readpage, NULL);
1014         if (!IS_ERR(page)) {
1015                 wait_on_page(page);
1016                 kmap(page);
1017                 if (!PageUptodate(page)) {
1018                         CERROR("page index %lu not uptodate\n", index);
1019                         GOTO(err_page, rc = -EIO);
1020                 }
1021                 if (PageError(page)) {
1022                         CERROR("page index %lu has error\n", index);
1023                         GOTO(err_page, rc = -EIO);
1024                 }
1025         }
1026         return page;
1027
1028 err_page:
1029         lustre_put_page(page);
1030         return ERR_PTR(rc);
1031 }
1032
1033 static struct page *
1034 lustre_get_page_write(struct inode *inode, unsigned long index)
1035 {
1036         struct address_space *mapping = inode->i_mapping;
1037         struct page *page;
1038         int rc;
1039
1040         page = grab_cache_page(mapping, index); /* locked page */
1041
1042         if (!IS_ERR(page)) {
1043                 kmap(page);
1044                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1045                  * a no-op for most filesystems, because we write the whole
1046                  * page.  For partial-page I/O this will read in the page.
1047                  */
1048                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1049                 if (rc) {
1050                         CERROR("page index %lu, rc = %d\n", index, rc);
1051                         if (rc != -ENOSPC)
1052                                 LBUG();
1053                         GOTO(err_unlock, rc);
1054                 }
1055                 /* XXX not sure if we need this if we are overwriting page */
1056                 if (PageError(page)) {
1057                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1058                         LBUG();
1059                         GOTO(err_unlock, rc = -EIO);
1060                 }
1061         }
1062         return page;
1063
1064 err_unlock:
1065         unlock_page(page);
1066         lustre_put_page(page);
1067         return ERR_PTR(rc);
1068 }
1069
1070 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1071 {
1072         struct inode *inode = page->mapping->host;
1073         int err;
1074
1075         err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1076         if (!err && IS_SYNC(inode))
1077                 err = waitfor_one_page(page);
1078
1079         //SetPageUptodate(page); // the client commit_write will do this
1080
1081         SetPageReferenced(page);
1082         unlock_page(page);
1083         lustre_put_page(page);
1084         return err;
1085 }
1086
1087 struct page *filter_get_page_write(struct inode *inode, unsigned long index,
1088                                    struct niobuf_local *lnb, int *pglocked)
1089 {
1090         struct address_space *mapping = inode->i_mapping;
1091         struct page *page;
1092         int rc;
1093
1094         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1095         if (*pglocked)
1096                 page = grab_cache_page_nowait(mapping, index); /* locked page */
1097         else
1098                 page = grab_cache_page(mapping, index); /* locked page */
1099
1100
1101         /* This page is currently locked, so get a temporary page instead. */
1102         /* XXX I believe this is a very dangerous thing to do - consider if
1103          *     we had multiple writers for the same file (definitely the case
1104          *     if we are using this codepath).  If writer A locks the page,
1105          *     writer B writes to a copy (as here), writer A drops the page
1106          *     lock, and writer C grabs the lock before B does, then B will
1107          *     later overwrite the data from C, even if C had LDLM locked
1108          *     and initiated the write after B did.
1109          */
1110         if (!page) {
1111                 unsigned long addr;
1112                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1113                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1114                 if (!addr) {
1115                         CERROR("no memory for a temp page\n");
1116                         LBUG();
1117                         GOTO(err, rc = -ENOMEM);
1118                 }
1119                 page = virt_to_page(addr);
1120                 kmap(page);
1121                 page->index = index;
1122                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1123         } else if (!IS_ERR(page)) {
1124                 (*pglocked)++;
1125                 kmap(page);
1126
1127                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1128                  * a no-op for most filesystems, because we write the whole
1129                  * page.  For partial-page I/O this will read in the page.
1130                  */
1131                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1132                 if (rc) {
1133                         CERROR("page index %lu, rc = %d\n", index, rc);
1134                         if (rc != -ENOSPC)
1135                                 LBUG();
1136                         GOTO(err_unlock, rc);
1137                 }
1138                 /* XXX not sure if we need this if we are overwriting page */
1139                 if (PageError(page)) {
1140                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1141                         LBUG();
1142                         GOTO(err_unlock, rc = -EIO);
1143                 }
1144         }
1145         return page;
1146
1147 err_unlock:
1148         unlock_page(page);
1149         lustre_put_page(page);
1150 err:
1151         return ERR_PTR(rc);
1152 }
1153
1154 /*
1155  * We need to balance prepare_write() calls with commit_write() calls.
1156  * If the page has been prepared, but we have no data for it, we don't
1157  * want to overwrite valid data on disk, but we still need to zero out
1158  * data for space which was newly allocated.  Like part of what happens
1159  * in __block_prepare_write() for newly allocated blocks.
1160  *
1161  * XXX currently __block_prepare_write() creates buffers for all the
1162  *     pages, and the filesystems mark these buffers as BH_New if they
1163  *     were newly allocated from disk. We use the BH_New flag similarly.
1164  */
1165 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1166                                int err)
1167 {
1168         if (err) {
1169                 unsigned block_start, block_end;
1170                 struct buffer_head *bh, *head = page->buffers;
1171                 unsigned blocksize = head->b_size;
1172                 void *addr = page_address(page);
1173
1174                 /* debugging: just seeing if this ever happens */
1175                 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1176                        page->index, page->mapping->host->i_ino, err);
1177
1178                 /* Currently one buffer per page, but in the future... */
1179                 for (bh = head, block_start = 0; bh != head || !block_start;
1180                      block_start = block_end, bh = bh->b_this_page) {
1181                         block_end = block_start + blocksize;
1182                         if (buffer_new(bh))
1183                                 memset(addr + block_start, 0, blocksize);
1184                 }
1185         }
1186
1187         return lustre_commit_write(page, from, to);
1188 }
1189
1190 static int filter_preprw(int cmd, struct lustre_handle *conn,
1191                          int objcount, struct obd_ioobj *obj,
1192                          int niocount, struct niobuf_remote *nb,
1193                          struct niobuf_local *res, void **desc_private)
1194 {
1195         struct obd_run_ctxt saved;
1196         struct obd_device *obd;
1197         struct obd_ioobj *o = obj;
1198         struct niobuf_remote *b = nb;
1199         struct niobuf_local *r = res;
1200         void *journal_save = NULL;
1201         int pglocked = 0;
1202         int rc = 0;
1203         int i;
1204         ENTRY;
1205
1206         obd = class_conn2obd(conn);
1207         if (!obd) {
1208                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1209                 RETURN(-EINVAL);
1210         }
1211         memset(res, 0, sizeof(*res) * niocount);
1212
1213         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1214
1215         if (cmd & OBD_BRW_WRITE) {
1216                 *desc_private = filter_journal_start(&journal_save,
1217                                                      &obd->u.filter,
1218                                                      objcount, obj, niocount,
1219                                                      nb);
1220                 if (IS_ERR(*desc_private))
1221                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1222         }
1223
1224         for (i = 0; i < objcount; i++, o++) {
1225                 struct dentry *dentry;
1226                 struct inode *inode;
1227                 int j;
1228
1229                 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1230                                            o->ioo_id, S_IFREG);
1231                 if (IS_ERR(dentry))
1232                         GOTO(out_clean, rc = PTR_ERR(dentry));
1233                 inode = dentry->d_inode;
1234                 if (!inode) {
1235                         CERROR("trying to BRW to non-existent file "LPU64"\n",
1236                                o->ioo_id);
1237                         f_dput(dentry);
1238                         GOTO(out_clean, rc = -ENOENT);
1239                 }
1240
1241                 for (j = 0; j < o->ioo_bufcnt; j++, b++, r++) {
1242                         unsigned long index = b->offset >> PAGE_SHIFT;
1243                         struct page *page;
1244
1245                         if (j == 0)
1246                                 r->dentry = dentry;
1247                         else
1248                                 r->dentry = dget(dentry);
1249
1250                         if (cmd & OBD_BRW_WRITE)
1251                                 page = filter_get_page_write(inode, index, r,
1252                                                              &pglocked);
1253                         else
1254                                 page = lustre_get_page_read(inode, index);
1255
1256                         if (IS_ERR(page)) {
1257                                 f_dput(dentry);
1258                                 GOTO(out_clean, rc = PTR_ERR(page));
1259                         }
1260
1261                         r->addr = page_address(page);
1262                         r->offset = b->offset;
1263                         r->page = page;
1264                         r->len = b->len;
1265                 }
1266         }
1267
1268 out_stop:
1269         if (cmd & OBD_BRW_WRITE) {
1270                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1271                                               *desc_private);
1272                 if (!rc)
1273                         rc = err;
1274         }
1275 out_ctxt:
1276         pop_ctxt(&saved);
1277         RETURN(rc);
1278 out_clean:
1279         while (r-- > res) {
1280                 CERROR("error cleanup on brw\n");
1281                 f_dput(r->dentry);
1282                 if (cmd & OBD_BRW_WRITE)
1283                         filter_commit_write(r->page, 0, PAGE_SIZE, rc);
1284                 else
1285                         lustre_put_page(r->page);
1286         }
1287         goto out_stop;
1288 }
1289
1290 static int filter_write_locked_page(struct niobuf_local *lnb)
1291 {
1292         struct page *lpage;
1293         int rc;
1294
1295         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1296         if (IS_ERR(lpage)) {
1297                 /* It is highly unlikely that we would ever get an error here.
1298                  * The page we want to get was previously locked, so it had to
1299                  * have already allocated the space, and we were just writing
1300                  * over the same data, so there would be no hole in the file.
1301                  *
1302                  * XXX: possibility of a race with truncate could exist, need
1303                  *      to check that.  There are no guarantees w.r.t.
1304                  *      write order even on a local filesystem, although the
1305                  *      normal response would be to return the number of bytes
1306                  *      successfully written and leave the rest to the app.
1307                  */
1308                 rc = PTR_ERR(lpage);
1309                 CERROR("error getting locked page index %ld: rc = %d\n",
1310                        lnb->page->index, rc);
1311                 GOTO(out, rc);
1312         }
1313
1314         /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1315          * lustre_commit_write() below, lnb->page was kmapped previously in
1316          * filter_get_page_write() and kunmapped in lustre_put_page() below.
1317          */
1318         memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1319         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1320         if (rc)
1321                 CERROR("error committing locked page %ld: rc = %d\n",
1322                        lnb->page->index, rc);
1323 out:
1324         lustre_put_page(lnb->page);
1325
1326         return rc;
1327 }
1328
1329 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1330                            int objcount, struct obd_ioobj *obj,
1331                            int niocount, struct niobuf_local *res,
1332                            void *private)
1333 {
1334         struct obd_run_ctxt saved;
1335         struct obd_ioobj *o;
1336         struct niobuf_local *r;
1337         struct obd_device *obd = class_conn2obd(conn);
1338         void *journal_save;
1339         int found_locked = 0;
1340         int rc = 0;
1341         int i;
1342         ENTRY;
1343
1344         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1345         journal_save = current->journal_info;
1346         LASSERT(!journal_save);
1347
1348         current->journal_info = private;
1349         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1350                 int j;
1351                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1352                         struct page *page = r->page;
1353
1354                         if (!page)
1355                                 LBUG();
1356
1357                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1358                                 found_locked++;
1359                                 continue;
1360                         }
1361
1362                         if (cmd & OBD_BRW_WRITE) {
1363                                 int err = filter_commit_write(page, 0,
1364                                                               r->len, 0);
1365
1366                                 if (!rc)
1367                                         rc = err;
1368                         } else
1369                                 lustre_put_page(page);
1370
1371                         f_dput(r->dentry);
1372                 }
1373         }
1374         current->journal_info = journal_save;
1375
1376         if (!found_locked)
1377                 goto out_ctxt;
1378
1379         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1380                 int j;
1381                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1382                         int err;
1383                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1384                                 continue;
1385
1386                         err = filter_write_locked_page(r);
1387                         if (!rc)
1388                                 rc = err;
1389                         f_dput(r->dentry);
1390                 }
1391         }
1392
1393 out_ctxt:
1394         pop_ctxt(&saved);
1395         RETURN(rc);
1396 }
1397
1398 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1399 {
1400         struct obd_device *obd = class_conn2obd(conn);
1401         struct statfs sfs;
1402         int rc;
1403
1404         ENTRY;
1405         rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1406         if (!rc)
1407                 statfs_pack(osfs, &sfs);
1408
1409         return rc;
1410 }
1411
1412 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1413                            void *key, obd_count *vallen, void **val)
1414 {
1415         struct obd_device *obd;
1416         ENTRY;
1417
1418         obd = class_conn2obd(conn);
1419         if (!obd) {
1420                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1421                 RETURN(-EINVAL);
1422         }
1423
1424         if ( keylen == strlen("blocksize") &&
1425              memcmp(key, "blocksize", keylen) == 0 ) {
1426                 *vallen = sizeof(long);
1427                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1428                 RETURN(0);
1429         }
1430
1431         if ( keylen == strlen("blocksize_bits") &&
1432              memcmp(key, "blocksize_bits", keylen) == 0 ){
1433                 *vallen = sizeof(long);
1434                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1435                 RETURN(0);
1436         }
1437
1438         if ( keylen == strlen("root_ino") &&
1439              memcmp(key, "root_ino", keylen) == 0 ){
1440                 *vallen = sizeof(obd_id);
1441                 *val = (void *)(obd_id)FILTER_ROOTINO;
1442                 RETURN(0);
1443         }
1444
1445         CDEBUG(D_IOCTL, "invalid key\n");
1446         RETURN(-EINVAL);
1447 }
1448
1449 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1450                   struct lustre_handle *src_conn, struct obdo *src,
1451                   obd_size count, obd_off offset)
1452 {
1453         struct page *page;
1454         struct lov_stripe_md srcmd, dstmd;
1455         unsigned long index = 0;
1456         int err = 0;
1457
1458         memset(&srcmd, 0, sizeof(srcmd));
1459         memset(&dstmd, 0, sizeof(dstmd));
1460         srcmd.lsm_object_id = src->o_id;
1461         dstmd.lsm_object_id = dst->o_id;
1462
1463         ENTRY;
1464         CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1465                ", dst: ino "LPU64"\n",
1466                src->o_id, src->o_blocks, src->o_size, dst->o_id);
1467         page = alloc_page(GFP_USER);
1468         if (page == NULL)
1469                 RETURN(-ENOMEM);
1470
1471         while (TryLockPage(page))
1472                 ___wait_on_page(page);
1473
1474         /* XXX with brw vector I/O, we could batch up reads and writes here,
1475          *     all we need to do is allocate multiple pages to handle the I/Os
1476          *     and arrays to handle the request parameters.
1477          */
1478         while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1479                 struct brw_page pg;
1480                 struct io_cb_data *cbd = ll_init_cb();
1481
1482                 if (!cbd) {
1483                         err = -ENOMEM;
1484                         EXIT;
1485                         break;
1486                 }
1487
1488                 pg.pg = page;
1489                 pg.count = PAGE_SIZE;
1490                 pg.off = (page->index) << PAGE_SHIFT;
1491                 pg.flag = 0;
1492
1493                 page->index = index;
1494                 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1495                               ll_sync_io_cb, cbd);
1496
1497                 if ( err ) {
1498                         EXIT;
1499                         break;
1500                 }
1501
1502                 cbd = ll_init_cb();
1503                 if (!cbd) {
1504                         err = -ENOMEM;
1505                         EXIT;
1506                         break;
1507                 }
1508                 pg.flag = OBD_BRW_CREATE;
1509                 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1510
1511                 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1512                               ll_sync_io_cb, cbd);
1513
1514                 /* XXX should handle dst->o_size, dst->o_blocks here */
1515                 if ( err ) {
1516                         EXIT;
1517                         break;
1518                 }
1519
1520                 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1521
1522                 index++;
1523         }
1524         dst->o_size = src->o_size;
1525         dst->o_blocks = src->o_blocks;
1526         dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1527         unlock_page(page);
1528         __free_page(page);
1529
1530         RETURN(err);
1531 }
1532
1533
1534 static struct obd_ops filter_obd_ops = {
1535         o_get_info:    filter_get_info,
1536         o_setup:       filter_setup,
1537         o_cleanup:     filter_cleanup,
1538         o_connect:     filter_connect,
1539         o_disconnect:  filter_disconnect,
1540         o_statfs:      filter_statfs,
1541         o_getattr:     filter_getattr,
1542         o_create:      filter_create,
1543         o_setattr:     filter_setattr,
1544         o_destroy:     filter_destroy,
1545         o_open:        filter_open,
1546         o_close:       filter_close,
1547         o_brw:         filter_pgcache_brw,
1548         o_punch:       filter_truncate,
1549         o_preprw:      filter_preprw,
1550         o_commitrw:    filter_commitrw
1551 #if 0
1552         o_preallocate: filter_preallocate_inodes,
1553         o_migrate:     filter_migrate,
1554         o_copy:        filter_copy_data,
1555         o_iterate:     filter_iterate
1556 #endif
1557 };
1558
1559
1560 static int __init obdfilter_init(void)
1561 {
1562         printk(KERN_INFO "Filtering OBD driver  v0.001, info@clusterfs.com\n");
1563         return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1564 }
1565
1566 static void __exit obdfilter_exit(void)
1567 {
1568         class_unregister_type(OBD_FILTER_DEVICENAME);
1569 }
1570
1571 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1572 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1573 MODULE_LICENSE("GPL");
1574
1575 module_init(obdfilter_init);
1576 module_exit(obdfilter_exit);