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 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
  27 #include <linux/extN_jbd.h>
  28 #endif
  29 #include <linux/quotaops.h>
  30 #include <linux/init.h>
  31 #include <linux/random.h>
  32 #include <linux/stringify.h>
  33
  34 static kmem_cache_t *filter_open_cache;
  35 static kmem_cache_t *filter_dentry_cache;
  36
  37 #define FILTER_ROOTINO 2
  38 #define FILTER_ROOTINO_STR __stringify(FILTER_ROOTINO)
  39
  40 #define S_SHIFT 12
  41 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
  42         [0]                     NULL,
  43         [S_IFREG >> S_SHIFT]    "R",
  44         [S_IFDIR >> S_SHIFT]    "D",
  45         [S_IFCHR >> S_SHIFT]    "C",
  46         [S_IFBLK >> S_SHIFT]    "B",
  47         [S_IFIFO >> S_SHIFT]    "F",
  48         [S_IFSOCK >> S_SHIFT]   "S",
  49         [S_IFLNK >> S_SHIFT]    "L"
  50 };
  51
  52 static inline const char *obd_mode_to_type(int mode)
  53 {
  54         return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
  55 }
  56
  57 /* write the pathname into the string */
  58 static int filter_id(char *buf, obd_id id, obd_mode mode)
  59 {
  60         return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
  61 }
  62
  63 static inline void f_dput(struct dentry *dentry)
  64 {
  65         /* Can't go inside filter_ddelete because it can block */
  66         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
  67                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
  68         LASSERT(atomic_read(&dentry->d_count) > 0);
  69
  70         dput(dentry);
  71 }
  72
  73 /* Not racy w.r.t. others, because we are the only user of this dentry */
  74 static void filter_drelease(struct dentry *dentry)
  75 {
  76         if (dentry->d_fsdata)
  77                 kmem_cache_free(filter_dentry_cache, dentry->d_fsdata);
  78 }
  79
  80 struct dentry_operations filter_dops = {
  81         .d_release = filter_drelease,
  82 };
  83
  84 /* setup the object store with correct subdirectories */
  85 static int filter_prep(struct obd_device *obd)
  86 {
  87         struct obd_run_ctxt saved;
  88         struct filter_obd *filter = &obd->u.filter;
  89         struct dentry *dentry;
  90         struct dentry *root;
  91         struct file *file;
  92         struct inode *inode;
  93         int rc = 0;
  94         __u64 lastobjid = 2;
  95         int mode = 0;
  96
  97         push_ctxt(&saved, &filter->fo_ctxt, NULL);
  98         dentry = simple_mkdir(current->fs->pwd, "O", 0700);
  99         CDEBUG(D_INODE, "got/created O: %p\n", dentry);
 100         if (IS_ERR(dentry)) {
 101                 rc = PTR_ERR(dentry);
 102                 CERROR("cannot open/create O: rc = %d\n", rc);
 103                 GOTO(out, rc);
 104         }
 105         filter->fo_dentry_O = dentry;
 106         dentry = simple_mkdir(current->fs->pwd, "P", 0700);
 107         CDEBUG(D_INODE, "got/created P: %p\n", dentry);
 108         if (IS_ERR(dentry)) {
 109                 rc = PTR_ERR(dentry);
 110                 CERROR("cannot open/create P: rc = %d\n", rc);
 111                 GOTO(out_O, rc);
 112         }
 113         f_dput(dentry);
 114         dentry = simple_mkdir(current->fs->pwd, "D", 0700);
 115         CDEBUG(D_INODE, "got/created D: %p\n", dentry);
 116         if (IS_ERR(dentry)) {
 117                 rc = PTR_ERR(dentry);
 118                 CERROR("cannot open/create D: rc = %d\n", rc);
 119                 GOTO(out_O, rc);
 120         }
 121
 122         root = simple_mknod(dentry, FILTER_ROOTINO_STR, S_IFREG | 0755);
 123         f_dput(dentry);
 124         if (IS_ERR(root)) {
 125                 rc = PTR_ERR(root);
 126                 CERROR("OBD filter: cannot open/create root %d: rc = %d\n",
 127                        FILTER_ROOTINO, rc);
 128                 GOTO(out_O, rc);
 129         }
 130         f_dput(root);
 131
 132         /*
 133          * Create directories and/or get dentries for each object type.
 134          * This saves us from having to do multiple lookups for each one.
 135          */
 136         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 137                 char *type = obd_type_by_mode[mode];
 138
 139                 if (!type) {
 140                         filter->fo_dentry_O_mode[mode] = NULL;
 141                         continue;
 142                 }
 143                 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
 144                 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
 145                 if (IS_ERR(dentry)) {
 146                         rc = PTR_ERR(dentry);
 147                         CERROR("cannot create O/%s: rc = %d\n", type, rc);
 148                         GOTO(out_O_mode, rc);
 149                 }
 150                 filter->fo_dentry_O_mode[mode] = dentry;
 151         }
 152
 153         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 154         if ( !file || IS_ERR(file) ) {
 155                 rc = PTR_ERR(file);
 156                 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
 157                        "D/status", rc);
 158                 GOTO(out_O_mode, rc);
 159         }
 160
 161         /* steal operations */
 162         inode = file->f_dentry->d_inode;
 163         filter->fo_fop = file->f_op;
 164         filter->fo_iop = inode->i_op;
 165         filter->fo_aops = inode->i_mapping->a_ops;
 166
 167         if (inode->i_size == 0) {
 168                 __u64 disk_lastobjid = cpu_to_le64(lastobjid);
 169                 ssize_t retval = file->f_op->write(file,(char *)&disk_lastobjid,
 170                                                    sizeof(disk_lastobjid),
 171                                                    &file->f_pos);
 172                 if (retval != sizeof(disk_lastobjid)) {
 173                         CDEBUG(D_INODE,"OBD filter: error writing lastobjid\n");
 174                         filp_close(file, 0);
 175                         GOTO(out_O_mode, rc = -EIO);
 176                 }
 177         } else {
 178                 __u64 disk_lastobjid;
 179                 ssize_t retval = file->f_op->read(file, (char *)&disk_lastobjid,
 180                                                   sizeof(disk_lastobjid),
 181                                                   &file->f_pos);
 182                 if (retval != sizeof(disk_lastobjid)) {
 183                         CDEBUG(D_INODE,"OBD filter: error reading lastobjid\n");
 184                         filp_close(file, 0);
 185                         GOTO(out_O_mode, rc = -EIO);
 186                 }
 187                 lastobjid = le64_to_cpu(disk_lastobjid);
 188         }
 189         filter->fo_lastobjid = lastobjid;
 190         filp_close(file, 0);
 191
 192         rc = 0;
 193  out:
 194         pop_ctxt(&saved);
 195
 196         return(rc);
 197
 198 out_O_mode:
 199         while (mode-- > 0) {
 200                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 201                 if (dentry) {
 202                         f_dput(dentry);
 203                         filter->fo_dentry_O_mode[mode] = NULL;
 204                 }
 205         }
 206 out_O:
 207         f_dput(filter->fo_dentry_O);
 208         filter->fo_dentry_O = NULL;
 209         goto out;
 210 }
 211
 212 /* cleanup the filter: write last used object id to status file */
 213 static void filter_post(struct obd_device *obd)
 214 {
 215         struct obd_run_ctxt saved;
 216         struct filter_obd *filter = &obd->u.filter;
 217         __u64 disk_lastobjid;
 218         long rc;
 219         struct file *file;
 220         int mode;
 221
 222         push_ctxt(&saved, &filter->fo_ctxt, NULL);
 223         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 224         if (IS_ERR(file)) {
 225                 CERROR("OBD filter: cannot create status file\n");
 226                 goto out;
 227         }
 228
 229         file->f_pos = 0;
 230         disk_lastobjid = cpu_to_le64(filter->fo_lastobjid);
 231         rc = file->f_op->write(file, (char *)&disk_lastobjid,
 232                        sizeof(disk_lastobjid), &file->f_pos);
 233         if (rc != sizeof(disk_lastobjid))
 234                 CERROR("OBD filter: error writing lastobjid: rc = %ld\n", rc);
 235
 236         rc = filp_close(file, NULL);
 237         if (rc)
 238                 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
 239
 240         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 241                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 242                 if (dentry) {
 243                         f_dput(dentry);
 244                         filter->fo_dentry_O_mode[mode] = NULL;
 245                 }
 246         }
 247         f_dput(filter->fo_dentry_O);
 248 out:
 249         pop_ctxt(&saved);
 250 }
 251
 252
 253 static __u64 filter_next_id(struct obd_device *obd)
 254 {
 255         obd_id id;
 256
 257         spin_lock(&obd->u.filter.fo_objidlock);
 258         id = ++obd->u.filter.fo_lastobjid;
 259         spin_unlock(&obd->u.filter.fo_objidlock);
 260
 261         /* FIXME: write the lastobjid to disk here */
 262         return id;
 263 }
 264
 265 /* how to get files, dentries, inodes from object id's */
 266 /* parent i_sem is already held if needed for exclusivity */
 267 static struct dentry *filter_fid2dentry(struct obd_device *obd,
 268                                         struct dentry *dparent,
 269                                         __u64 id, __u32 type)
 270 {
 271         struct super_block *sb = obd->u.filter.fo_sb;
 272         struct dentry *dchild;
 273         char name[32];
 274         int len;
 275         ENTRY;
 276
 277         if (!sb || !sb->s_dev) {
 278                 CERROR("fatal: device not initialized.\n");
 279                 RETURN(ERR_PTR(-ENXIO));
 280         }
 281
 282         if (id == 0) {
 283                 CERROR("fatal: invalid object #0\n");
 284                 LBUG();
 285                 RETURN(ERR_PTR(-ESTALE));
 286         }
 287
 288         if (!(type & S_IFMT)) {
 289                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 290                        __FUNCTION__, id, type);
 291                 RETURN(ERR_PTR(-EINVAL));
 292         }
 293
 294         len = sprintf(name, LPU64, id);
 295         CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
 296                name);
 297         dchild = lookup_one_len(name, dparent, len);
 298         if (IS_ERR(dchild)) {
 299                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 300                 RETURN(dchild);
 301         }
 302
 303         CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
 304                obd_mode_to_type(type), name, dchild,
 305                atomic_read(&dchild->d_count));
 306
 307         LASSERT(atomic_read(&dchild->d_count) > 0);
 308
 309         RETURN(dchild);
 310 }
 311
 312 static inline struct dentry *filter_parent(struct obd_device *obd,
 313                                            obd_mode mode)
 314 {
 315         struct filter_obd *filter = &obd->u.filter;
 316
 317         return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
 318 }
 319
 320 static struct file *filter_obj_open(struct obd_export *export,
 321                                     __u64 id, __u32 type)
 322 {
 323         struct filter_obd *filter = &export->exp_obd->u.filter;
 324         struct super_block *sb = filter->fo_sb;
 325         struct dentry *dentry;
 326         struct filter_export_data *fed = &export->exp_filter_data;
 327         struct filter_dentry_data *fdd;
 328         struct filter_file_data *ffd;
 329         struct obd_run_ctxt saved;
 330         char name[24];
 331         struct file *file;
 332         ENTRY;
 333
 334         if (!sb || !sb->s_dev) {
 335                 CERROR("fatal: device not initialized.\n");
 336                 RETURN(ERR_PTR(-ENXIO));
 337         }
 338
 339         if (!id) {
 340                 CERROR("fatal: invalid obdo "LPU64"\n", id);
 341                 RETURN(ERR_PTR(-ESTALE));
 342         }
 343
 344         if (!(type & S_IFMT)) {
 345                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 346                        __FUNCTION__, id, type);
 347                 RETURN(ERR_PTR(-EINVAL));
 348         }
 349
 350         ffd = kmem_cache_alloc(filter_open_cache, SLAB_KERNEL);
 351         if (!ffd) {
 352                 CERROR("obdfilter: out of memory\n");
 353                 RETURN(ERR_PTR(-ENOMEM));
 354         }
 355
 356         /* We preallocate this to avoid blocking while holding fo_fddlock */
 357         fdd = kmem_cache_alloc(filter_dentry_cache, SLAB_KERNEL);
 358         if (!fdd) {
 359                 CERROR("obdfilter: out of memory\n");
 360                 GOTO(out_ffd, file = ERR_PTR(-ENOMEM));
 361         }
 362
 363         filter_id(name, id, type);
 364         push_ctxt(&saved, &filter->fo_ctxt, NULL);
 365         file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
 366         pop_ctxt(&saved);
 367
 368         if (IS_ERR(file))
 369                 GOTO(out_fdd, file);
 370
 371         dentry = file->f_dentry;
 372         spin_lock(&filter->fo_fddlock);
 373         if (dentry->d_fsdata) {
 374                 spin_unlock(&filter->fo_fddlock);
 375                 kmem_cache_free(filter_dentry_cache, fdd);
 376                 fdd = dentry->d_fsdata;
 377                 LASSERT(kmem_cache_validate(filter_dentry_cache, fdd));
 378                 /* should only happen during client recovery */
 379                 if (fdd->fdd_flags & FILTER_FLAG_DESTROY)
 380                         CDEBUG(D_INODE,"opening destroyed object "LPX64"\n",id);
 381                 atomic_inc(&fdd->fdd_open_count);
 382         } else {
 383                 atomic_set(&fdd->fdd_open_count, 1);
 384                 fdd->fdd_flags = 0;
 385                 /* If this is racy, then we can use {cmp}xchg and atomic_add */
 386                 dentry->d_fsdata = fdd;
 387                 spin_unlock(&filter->fo_fddlock);
 388         }
 389
 390         get_random_bytes(&ffd->ffd_servercookie, sizeof(ffd->ffd_servercookie));
 391         ffd->ffd_file = file;
 392         file->private_data = ffd;
 393
 394         if (!dentry->d_op)
 395                 dentry->d_op = &filter_dops;
 396         else
 397                 LASSERT(dentry->d_op == &filter_dops);
 398
 399         spin_lock(&fed->fed_lock);
 400         list_add(&ffd->ffd_export_list, &fed->fed_open_head);
 401         spin_unlock(&fed->fed_lock);
 402
 403         CDEBUG(D_INODE, "opening objid "LPX64": rc = %p\n", id, file);
 404
 405 out:
 406         RETURN(file);
 407
 408 out_fdd:
 409         kmem_cache_free(filter_dentry_cache, fdd);
 410 out_ffd:
 411         ffd->ffd_servercookie = DEAD_HANDLE_MAGIC;
 412         kmem_cache_free(filter_open_cache, ffd);
 413         goto out;
 414 }
 415
 416 /* Caller must hold i_sem on dir_dentry->d_inode */
 417 static int filter_destroy_internal(struct obd_device *obd,
 418                                    struct dentry *dir_dentry,
 419                                    struct dentry *object_dentry)
 420 {
 421         struct obd_run_ctxt saved;
 422         struct inode *inode = object_dentry->d_inode;
 423         int rc;
 424         ENTRY;
 425
 426         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
 427                 CERROR("destroying objid %*s nlink = %d, count = %d\n",
 428                        object_dentry->d_name.len,
 429                        object_dentry->d_name.name,
 430                        inode->i_nlink, atomic_read(&inode->i_count));
 431         }
 432
 433         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 434         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 435         /* XXX unlink from PENDING directory now too */
 436         pop_ctxt(&saved);
 437
 438         if (rc)
 439                 CERROR("error unlinking objid %*s: rc %d\n",
 440                        object_dentry->d_name.len,
 441                        object_dentry->d_name.name, rc);
 442
 443         RETURN(rc);
 444 }
 445
 446 static int filter_close_internal(struct obd_device *obd,
 447                                  struct filter_file_data *ffd)
 448 {
 449         struct file *filp = ffd->ffd_file;
 450         struct dentry *object_dentry = dget(filp->f_dentry);
 451         struct filter_dentry_data *fdd = object_dentry->d_fsdata;
 452         int rc, rc2 = 0;
 453         ENTRY;
 454
 455         LASSERT(filp->private_data == ffd);
 456         LASSERT(fdd);
 457
 458         rc = filp_close(filp, 0);
 459
 460         if (atomic_dec_and_test(&fdd->fdd_open_count) &&
 461             fdd->fdd_flags & FILTER_FLAG_DESTROY) {
 462                 struct dentry *dir_dentry = filter_parent(obd, S_IFREG);
 463
 464                 down(&dir_dentry->d_inode->i_sem);
 465                 rc2 = filter_destroy_internal(obd, dir_dentry, object_dentry);
 466                 if (rc2 && !rc)
 467                         rc = rc2;
 468                 up(&dir_dentry->d_inode->i_sem);
 469         }
 470
 471         f_dput(object_dentry);
 472         kmem_cache_free(filter_open_cache, ffd);
 473
 474         RETURN(rc);
 475 }
 476
 477 /* obd methods */
 478 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
 479                           obd_uuid_t cluuid, struct recovd_obd *recovd,
 480                           ptlrpc_recovery_cb_t recover)
 481 {
 482         struct obd_export *exp;
 483         int rc;
 484
 485         ENTRY;
 486         MOD_INC_USE_COUNT;
 487         rc = class_connect(conn, obd, cluuid);
 488         if (rc)
 489                 GOTO(out_dec, rc);
 490         exp = class_conn2export(conn);
 491         LASSERT(exp);
 492
 493         INIT_LIST_HEAD(&exp->exp_filter_data.fed_open_head);
 494         spin_lock_init(&exp->exp_filter_data.fed_lock);
 495 out:
 496         RETURN(rc);
 497
 498 out_dec:
 499         MOD_DEC_USE_COUNT;
 500         goto out;
 501 }
 502
 503 static int filter_disconnect(struct lustre_handle *conn)
 504 {
 505         struct obd_export *exp = class_conn2export(conn);
 506         struct filter_export_data *fed;
 507         int rc;
 508         ENTRY;
 509
 510         LASSERT(exp);
 511         fed = &exp->exp_filter_data;
 512         spin_lock(&fed->fed_lock);
 513         while (!list_empty(&fed->fed_open_head)) {
 514                 struct filter_file_data *ffd;
 515
 516                 ffd = list_entry(fed->fed_open_head.next, typeof(*ffd),
 517                                  ffd_export_list);
 518                 list_del(&ffd->ffd_export_list);
 519                 spin_unlock(&fed->fed_lock);
 520
 521                 CERROR("force closing file %*s on disconnect\n",
 522                        ffd->ffd_file->f_dentry->d_name.len,
 523                        ffd->ffd_file->f_dentry->d_name.name);
 524
 525                 filter_close_internal(exp->exp_obd, ffd);
 526                 spin_lock(&fed->fed_lock);
 527         }
 528         spin_unlock(&fed->fed_lock);
 529
 530         ldlm_cancel_locks_for_export(exp);
 531         rc = class_disconnect(conn);
 532         if (!rc)
 533                 MOD_DEC_USE_COUNT;
 534
 535         /* XXX cleanup preallocated inodes */
 536         RETURN(rc);
 537 }
 538
 539 /* mount the file system (secretly) */
 540 static int filter_setup(struct obd_device *obd, obd_count len, void *buf)
 541 {
 542         struct obd_ioctl_data* data = buf;
 543         struct filter_obd *filter;
 544         struct vfsmount *mnt;
 545         int err = 0;
 546         ENTRY;
 547
 548         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 549                 RETURN(-EINVAL);
 550
 551         MOD_INC_USE_COUNT;
 552         mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
 553         err = PTR_ERR(mnt);
 554         if (IS_ERR(mnt))
 555                 GOTO(err_dec, err);
 556
 557         filter = &obd->u.filter;;
 558         filter->fo_vfsmnt = mnt;
 559         filter->fo_fstype = strdup(data->ioc_inlbuf2);
 560         filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
 561         CERROR("%s: mnt is %p\n", data->ioc_inlbuf1, filter->fo_vfsmnt);
 562         /* XXX is this even possible if do_kern_mount succeeded? */
 563         if (!filter->fo_sb)
 564                 GOTO(err_kfree, err = -ENODEV);
 565
 566         OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
 567         filter->fo_ctxt.pwdmnt = mnt;
 568         filter->fo_ctxt.pwd = mnt->mnt_root;
 569         filter->fo_ctxt.fs = get_ds();
 570
 571         err = filter_prep(obd);
 572         if (err)
 573                 GOTO(err_kfree, err);
 574         spin_lock_init(&filter->fo_fddlock);
 575         spin_lock_init(&filter->fo_objidlock);
 576         INIT_LIST_HEAD(&filter->fo_export_list);
 577
 578         obd->obd_namespace =
 579                 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
 580         if (obd->obd_namespace == NULL)
 581                 LBUG();
 582
 583         ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
 584                            "filter_ldlm_client", &obd->obd_ldlm_client);
 585
 586         RETURN(0);
 587
 588 err_kfree:
 589         kfree(filter->fo_fstype);
 590         unlock_kernel();
 591         mntput(filter->fo_vfsmnt);
 592         filter->fo_sb = 0;
 593         lock_kernel();
 594
 595 err_dec:
 596         MOD_DEC_USE_COUNT;
 597         return err;
 598 }
 599
 600
 601 static int filter_cleanup(struct obd_device *obd)
 602 {
 603         struct super_block *sb;
 604         ENTRY;
 605
 606         if (!list_empty(&obd->obd_exports)) {
 607                 CERROR("still has clients!\n");
 608                 class_disconnect_all(obd);
 609                 if (!list_empty(&obd->obd_exports)) {
 610                         CERROR("still has exports after forced cleanup?\n");
 611                         RETURN(-EBUSY);
 612                 }
 613         }
 614
 615         ldlm_namespace_free(obd->obd_namespace);
 616
 617         sb = obd->u.filter.fo_sb;
 618         if (!obd->u.filter.fo_sb)
 619                 RETURN(0);
 620
 621         filter_post(obd);
 622
 623         shrink_dcache_parent(sb->s_root);
 624         unlock_kernel();
 625         mntput(obd->u.filter.fo_vfsmnt);
 626         obd->u.filter.fo_sb = 0;
 627         kfree(obd->u.filter.fo_fstype);
 628
 629         lock_kernel();
 630
 631         MOD_DEC_USE_COUNT;
 632         RETURN(0);
 633 }
 634
 635
 636 static void filter_from_inode(struct obdo *oa, struct inode *inode, int valid)
 637 {
 638         int type = oa->o_mode & S_IFMT;
 639         ENTRY;
 640
 641         CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
 642                inode->i_ino, inode, (long)oa->o_id, valid);
 643         /* Don't copy the inode number in place of the object ID */
 644         obdo_from_inode(oa, inode, valid);
 645         oa->o_mode &= ~S_IFMT;
 646         oa->o_mode |= type;
 647
 648         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
 649                 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
 650                 oa->o_rdev = rdev;
 651                 oa->o_valid |= OBD_MD_FLRDEV;
 652         }
 653
 654         EXIT;
 655 }
 656
 657 static struct filter_file_data *filter_handle2ffd(struct lustre_handle *handle)
 658 {
 659         struct filter_file_data *ffd = NULL;
 660         ENTRY;
 661
 662         if (!handle || !handle->addr)
 663                 RETURN(NULL);
 664
 665         ffd = (struct filter_file_data *)(unsigned long)(handle->addr);
 666         if (!kmem_cache_validate(filter_open_cache, (void *)ffd))
 667                 RETURN(NULL);
 668
 669         if (ffd->ffd_servercookie != handle->cookie)
 670                 RETURN(NULL);
 671
 672         LASSERT(ffd->ffd_file->private_data == ffd);
 673         RETURN(ffd);
 674 }
 675
 676 static struct dentry *__filter_oa2dentry(struct lustre_handle *conn,
 677                                          struct obdo *oa, char *what)
 678 {
 679         struct dentry *dentry = NULL;
 680
 681         if (oa->o_valid & OBD_MD_FLHANDLE) {
 682                 struct lustre_handle *ost_handle = obdo_handle(oa);
 683                 struct filter_file_data *ffd = filter_handle2ffd(ost_handle);
 684
 685                 if (ffd)
 686                         dentry = dget(ffd->ffd_file->f_dentry);
 687         }
 688
 689         if (!dentry) {
 690                 struct obd_device *obd = class_conn2obd(conn);
 691                 if (!obd) {
 692                         CERROR("invalid client "LPX64"\n", conn->addr);
 693                         RETURN(ERR_PTR(-EINVAL));
 694                 }
 695                 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 696                                            oa->o_id, oa->o_mode);
 697         }
 698
 699         if (!dentry->d_inode) {
 700                 CERROR("%s on non-existent object: "LPU64"\n", what, oa->o_id);
 701                 f_dput(dentry);
 702                 dentry = ERR_PTR(-ENOENT);
 703         }
 704
 705         return dentry;
 706 }
 707
 708 #define filter_oa2dentry(conn, oa) __filter_oa2dentry(conn, oa, __FUNCTION__)
 709
 710 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
 711                           struct lov_stripe_md *md)
 712 {
 713         struct dentry *dentry = NULL;
 714         int rc = 0;
 715         ENTRY;
 716
 717         dentry = filter_oa2dentry(conn, oa);
 718         if (IS_ERR(dentry))
 719                 RETURN(PTR_ERR(dentry));
 720
 721         filter_from_inode(oa, dentry->d_inode, oa->o_valid);
 722
 723         f_dput(dentry);
 724         RETURN(rc);
 725 }
 726
 727 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
 728                           struct lov_stripe_md *md)
 729 {
 730         struct obd_run_ctxt saved;
 731         struct obd_device *obd = class_conn2obd(conn);
 732         struct dentry *dentry;
 733         struct iattr iattr;
 734         struct inode *inode;
 735         int rc;
 736         ENTRY;
 737
 738         dentry = filter_oa2dentry(conn, oa);
 739
 740         if (IS_ERR(dentry))
 741                 RETURN(PTR_ERR(dentry));
 742
 743         iattr_from_obdo(&iattr, oa, oa->o_valid);
 744         iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
 745         inode = dentry->d_inode;
 746
 747         lock_kernel();
 748         if (iattr.ia_valid & ATTR_SIZE)
 749                 down(&inode->i_sem);
 750         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 751         if (inode->i_op->setattr)
 752                 rc = inode->i_op->setattr(dentry, &iattr);
 753         else
 754                 rc = inode_setattr(inode, &iattr);
 755         pop_ctxt(&saved);
 756         if (iattr.ia_valid & ATTR_SIZE) {
 757                 up(&inode->i_sem);
 758                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 759                 obdo_from_inode(oa, inode, oa->o_valid);
 760         }
 761         unlock_kernel();
 762
 763         f_dput(dentry);
 764         RETURN(rc);
 765 }
 766
 767 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
 768                        struct lov_stripe_md *ea)
 769 {
 770         struct obd_export *export;
 771         struct lustre_handle *handle;
 772         struct filter_file_data *ffd;
 773         struct file *filp;
 774         int rc = 0;
 775         ENTRY;
 776
 777         export = class_conn2export(conn);
 778         if (!export) {
 779                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 780                 RETURN(-EINVAL);
 781         }
 782
 783         filp = filter_obj_open(export, oa->o_id, oa->o_mode);
 784         if (IS_ERR(filp))
 785                 GOTO(out, rc = PTR_ERR(filp));
 786
 787         filter_from_inode(oa, filp->f_dentry->d_inode, oa->o_valid);
 788
 789         ffd = filp->private_data;
 790         handle = obdo_handle(oa);
 791         handle->addr = (__u64)(unsigned long)ffd;
 792         handle->cookie = ffd->ffd_servercookie;
 793         oa->o_valid |= OBD_MD_FLHANDLE;
 794 out:
 795         RETURN(rc);
 796 } /* filter_open */
 797
 798 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
 799                         struct lov_stripe_md *ea)
 800 {
 801         struct obd_export *exp;
 802         struct filter_file_data *ffd;
 803         struct filter_export_data *fed;
 804         int rc;
 805         ENTRY;
 806
 807         exp = class_conn2export(conn);
 808         if (!exp) {
 809                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 810                 RETURN(-EINVAL);
 811         }
 812
 813         if (!(oa->o_valid & OBD_MD_FLHANDLE)) {
 814                 CERROR("no handle for close of objid "LPX64"\n", oa->o_id);
 815                 RETURN(-EINVAL);
 816         }
 817
 818         ffd = filter_handle2ffd(obdo_handle(oa));
 819         if (!ffd) {
 820                 struct lustre_handle *handle = obdo_handle(oa);
 821                 CERROR("bad handle ("LPX64") or cookie ("LPX64") for close\n",
 822                        handle->addr, handle->cookie);
 823                 RETURN(-ESTALE);
 824         }
 825
 826         fed = &exp->exp_filter_data;
 827         spin_lock(&fed->fed_lock);
 828         list_del(&ffd->ffd_export_list);
 829         spin_unlock(&fed->fed_lock);
 830
 831         rc = filter_close_internal(exp->exp_obd, ffd);
 832
 833         RETURN(rc);
 834 } /* filter_close */
 835
 836 static int filter_create(struct lustre_handle *conn, struct obdo *oa,
 837                          struct lov_stripe_md **ea)
 838 {
 839         struct obd_device *obd = class_conn2obd(conn);
 840         char name[64];
 841         struct obd_run_ctxt saved;
 842         struct dentry *new;
 843         struct iattr;
 844         ENTRY;
 845
 846         if (!obd) {
 847                 CERROR("invalid client "LPX64"\n", conn->addr);
 848                 return -EINVAL;
 849         }
 850
 851         if (!(oa->o_mode & S_IFMT)) {
 852                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 853                        __FUNCTION__, oa->o_id, oa->o_mode);
 854                 return -ENOENT;
 855         }
 856
 857         oa->o_id = filter_next_id(obd);
 858
 859         //filter_id(name, oa->o_id, oa->o_mode);
 860         sprintf(name, LPU64, oa->o_id);
 861         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 862         new = simple_mknod(filter_parent(obd, oa->o_mode), name, oa->o_mode);
 863         pop_ctxt(&saved);
 864         if (IS_ERR(new)) {
 865                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
 866                 return -ENOENT;
 867         }
 868
 869         /* Set flags for fields we have set in the inode struct */
 870         oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 871                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
 872         filter_from_inode(oa, new->d_inode, oa->o_valid);
 873         f_dput(new);
 874
 875         return 0;
 876 }
 877
 878 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
 879                           struct lov_stripe_md *ea)
 880 {
 881         struct obd_device *obd = class_conn2obd(conn);
 882         struct dentry *dir_dentry, *object_dentry;
 883         struct filter_dentry_data *fdd;
 884         int rc;
 885         ENTRY;
 886
 887         if (!obd) {
 888                 CERROR("invalid client "LPX64"\n", conn->addr);
 889                 RETURN(-EINVAL);
 890         }
 891
 892         CDEBUG(D_INODE, "destroying objid "LPX64"\n", oa->o_id);
 893
 894         dir_dentry = filter_parent(obd, oa->o_mode);
 895         down(&dir_dentry->d_inode->i_sem);
 896
 897         object_dentry = filter_oa2dentry(conn, oa);
 898         if (IS_ERR(object_dentry))
 899                 GOTO(out, rc = -ENOENT);
 900
 901         fdd = object_dentry->d_fsdata;
 902         if (fdd && atomic_read(&fdd->fdd_open_count)) {
 903                 if (!(fdd->fdd_flags & FILTER_FLAG_DESTROY)) {
 904                         fdd->fdd_flags |= FILTER_FLAG_DESTROY;
 905                         /* XXX put into PENDING directory in case of crash */
 906                         CDEBUG(D_INODE,
 907                                "defer destroy of %dx open objid "LPX64"\n",
 908                                atomic_read(&fdd->fdd_open_count), oa->o_id);
 909                 } else
 910                         CDEBUG(D_INODE,
 911                                "repeat destroy of %dx open objid "LPX64"\n",
 912                                atomic_read(&fdd->fdd_open_count), oa->o_id);
 913                 GOTO(out_dput, rc = 0);
 914         }
 915
 916         rc = filter_destroy_internal(obd, dir_dentry, object_dentry);
 917 out_dput:
 918         f_dput(object_dentry);
 919
 920         EXIT;
 921 out:
 922         up(&dir_dentry->d_inode->i_sem);
 923         return rc;
 924 }
 925
 926 /* NB count and offset are used for punch, but not truncate */
 927 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
 928                            struct lov_stripe_md *lsm,
 929                            obd_off start, obd_off end)
 930 {
 931         int error;
 932         ENTRY;
 933
 934         if (end != OBD_OBJECT_EOF)
 935                 CERROR("PUNCH not supported, only truncate works\n");
 936
 937         CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
 938                "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
 939         oa->o_size = start;
 940         error = filter_setattr(conn, oa, NULL);
 941         RETURN(error);
 942 }
 943
 944 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
 945                               struct lov_stripe_md *lsm, obd_count oa_bufs,
 946                               struct brw_page *pga, brw_callback_t callback,
 947                               struct io_cb_data *data)
 948 {
 949         struct obd_export       *export = class_conn2export(conn);
 950         struct obd_run_ctxt      saved;
 951         struct super_block      *sb;
 952         int                      pnum;          /* index to pages (bufs) */
 953         unsigned long            retval;
 954         int                      error;
 955         struct file             *file;
 956         int pg;
 957         ENTRY;
 958
 959         if (!export) {
 960                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
 961                 RETURN(-EINVAL);
 962         }
 963
 964         sb = export->exp_obd->u.filter.fo_sb;
 965         push_ctxt(&saved, &export->exp_obd->u.filter.fo_ctxt, NULL);
 966         pnum = 0; /* pnum indexes buf 0..num_pages */
 967
 968         file = filter_obj_open(export, lsm->lsm_object_id, S_IFREG);
 969         if (IS_ERR(file))
 970                 GOTO(out, retval = PTR_ERR(file));
 971
 972         /* count doubles as retval */
 973         for (pg = 0; pg < oa_bufs; pg++) {
 974                 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
 975                        ") off count ("LPU64",%d)\n",
 976                        cmd, pnum, file->f_dentry->d_inode->i_ino,
 977                        pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
 978                        (int)pga[pnum].count);
 979                 if (cmd & OBD_BRW_WRITE) {
 980                         loff_t off;
 981                         char *buffer;
 982                         off = pga[pnum].off;
 983                         buffer = kmap(pga[pnum].pg);
 984                         retval = file->f_op->write(file, buffer,
 985                                                    pga[pnum].count,
 986                                                    &off);
 987                         kunmap(pga[pnum].pg);
 988                         CDEBUG(D_INODE, "retval %ld\n", retval);
 989                 } else {
 990                         loff_t off = pga[pnum].off;
 991                         char *buffer = kmap(pga[pnum].pg);
 992
 993                         if (off >= file->f_dentry->d_inode->i_size) {
 994                                 memset(buffer, 0, pga[pnum].count);
 995                                 retval = pga[pnum].count;
 996                         } else {
 997                                 retval = file->f_op->read(file, buffer,
 998                                                           pga[pnum].count, &off);
 999                         }
1000                         kunmap(pga[pnum].pg);
1001
1002                         if (retval != pga[pnum].count) {
1003                                 filp_close(file, 0);
1004                                 GOTO(out, retval = -EIO);
1005                         }
1006                         CDEBUG(D_INODE, "retval %ld\n", retval);
1007                 }
1008                 pnum++;
1009         }
1010         /* sizes and blocks are set by generic_file_write */
1011         /* ctimes/mtimes will follow with a setattr call */
1012         filp_close(file, 0);
1013
1014         /* XXX: do something with callback if it is set? */
1015
1016         EXIT;
1017 out:
1018         pop_ctxt(&saved);
1019         error = (retval >= 0) ? 0 : retval;
1020         return error;
1021 }
1022
1023 /*
1024  * Calculate the number of buffer credits needed to write multiple pages in
1025  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
1026  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
1027  *
1028  * See comment above ext3_writepage_trans_blocks for details.  We assume
1029  * no data journaling is being done, but it does allow for all of the pages
1030  * being non-contiguous.  If we are guaranteed contiguous pages we could
1031  * reduce the number of (d)indirect blocks a lot.
1032  *
1033  * With N blocks per page and P pages, for each inode we have at most:
1034  * N*P indirect
1035  * min(N*P, blocksize/4 + 1) dindirect blocks
1036  * 1 tindirect
1037  *
1038  * For the entire filesystem, we have at most:
1039  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
1040  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
1041  * 1 inode block
1042  * 1 superblock
1043  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
1044  */
1045 static int ext3_credits_needed(struct super_block *sb, int objcount,
1046                                struct obd_ioobj *obj)
1047 {
1048         struct obd_ioobj *o = obj;
1049         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
1050         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
1051         int nbitmaps = 0;
1052         int ngdblocks = 0;
1053         int needed = objcount + 1;
1054         int i;
1055
1056         for (i = 0; i < objcount; i++, o++) {
1057                 int nblocks = o->ioo_bufcnt * blockpp;
1058                 int ndindirect = min(nblocks, addrpp + 1);
1059                 int nindir = nblocks + ndindirect + 1;
1060
1061                 nbitmaps += nindir + nblocks;
1062                 ngdblocks += nindir + nblocks;
1063
1064                 needed += nindir;
1065         }
1066
1067         /* Assumes ext3 and extN have same sb_info layout at the start. */
1068         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
1069                 nbitmaps = EXT3_SB(sb)->s_groups_count;
1070         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
1071                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
1072
1073         needed += nbitmaps + ngdblocks;
1074
1075 #ifdef CONFIG_QUOTA
1076         /* We assume that there will be 1 bit set in s_dquot.flags for each
1077          * quota file that is active.  This is at least true for now.
1078          */
1079         needed += hweight32(sb_any_quota_enabled(sb)) *
1080                 EXT3_SINGLEDATA_TRANS_BLOCKS;
1081 #endif
1082
1083         return needed;
1084 }
1085
1086 /* We have to start a huge journal transaction here to hold all of the
1087  * metadata for the pages being written here.  This is necessitated by
1088  * the fact that we do lots of prepare_write operations before we do
1089  * any of the matching commit_write operations, so even if we split
1090  * up to use "smaller" transactions none of them could complete until
1091  * all of them were opened.  By having a single journal transaction,
1092  * we eliminate duplicate reservations for common blocks like the
1093  * superblock and group descriptors or bitmaps.
1094  *
1095  * We will start the transaction here, but each prepare_write will
1096  * add a refcount to the transaction, and each commit_write will
1097  * remove a refcount.  The transaction will be closed when all of
1098  * the pages have been written.
1099  */
1100 static void *ext3_filter_journal_start(struct filter_obd *filter,
1101                                        int objcount, struct obd_ioobj *obj,
1102                                        int niocount, struct niobuf_remote *nb)
1103 {
1104         journal_t *journal = NULL;
1105         handle_t *handle = NULL;
1106         int needed;
1107
1108         /* It appears that some kernels have different values for
1109          * EXT*_MAX_GROUP_LOADED (either 8 or 32), so we cannot
1110          * assume anything after s_inode_bitmap_number is the same.
1111          */
1112         if (!strcmp(filter->fo_fstype, "ext3"))
1113                 journal = EXT3_SB(filter->fo_sb)->s_journal;
1114 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1115         else if (!strcmp(filter->fo_fstype, "extN"))
1116                 journal = EXTN_SB(filter->fo_sb)->s_journal;
1117 #endif
1118         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
1119
1120         /* The number of blocks we could _possibly_ dirty can very large.
1121          * We reduce our request if it is absurd (and we couldn't get that
1122          * many credits for a single handle anyways).
1123          *
1124          * At some point we have to limit the size of I/Os sent at one time,
1125          * increase the size of the journal, or we have to calculate the
1126          * actual journal requirements more carefully by checking all of
1127          * the blocks instead of being maximally pessimistic.  It remains to
1128          * be seen if this is a real problem or not.
1129          */
1130         if (needed > journal->j_max_transaction_buffers) {
1131                 CERROR("want too many journal credits (%d) using %d instead\n",
1132                        needed, journal->j_max_transaction_buffers);
1133                 needed = journal->j_max_transaction_buffers;
1134         }
1135
1136         lock_kernel();
1137         handle = journal_start(journal, needed);
1138         unlock_kernel();
1139         if (IS_ERR(handle))
1140                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
1141                        PTR_ERR(handle));
1142
1143         return(handle);
1144 }
1145
1146 static void *filter_journal_start(void **journal_save,
1147                                   struct filter_obd *filter,
1148                                   int objcount, struct obd_ioobj *obj,
1149                                   int niocount, struct niobuf_remote *nb)
1150 {
1151         void *handle = NULL;
1152
1153         /* This may not be necessary - we probably never have a
1154          * transaction started when we enter here, so we can
1155          * remove the saving of the journal state entirely.
1156          * For now leave it in just to see if it ever happens.
1157          */
1158         *journal_save = current->journal_info;
1159         if (*journal_save) {
1160                 CERROR("Already have handle %p???\n", *journal_save);
1161                 LBUG();
1162                 current->journal_info = NULL;
1163         }
1164
1165         if (!strcmp(filter->fo_fstype, "ext3") ||
1166             !strcmp(filter->fo_fstype, "extN"))
1167                 handle = ext3_filter_journal_start(filter, objcount, obj,
1168                                                    niocount, nb);
1169         return handle;
1170 }
1171
1172 static int ext3_filter_journal_stop(void *handle)
1173 {
1174         int rc;
1175
1176         /* We got a refcount on the handle for each call to prepare_write,
1177          * so we can drop the "parent" handle here to avoid the need for
1178          * osc to call back into filterobd to close the handle.  The
1179          * remaining references will be dropped in commit_write.
1180          */
1181         lock_kernel();
1182         rc = journal_stop((handle_t *)handle);
1183         unlock_kernel();
1184
1185         return rc;
1186 }
1187
1188 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
1189                                void *handle)
1190 {
1191         int rc = 0;
1192
1193         if (!strcmp(filter->fo_fstype, "ext3") ||
1194             !strcmp(filter->fo_fstype, "extN"))
1195                 rc = ext3_filter_journal_stop(handle);
1196
1197         if (rc)
1198                 CERROR("error on journal stop: rc = %d\n", rc);
1199
1200         current->journal_info = journal_save;
1201
1202         return rc;
1203 }
1204
1205 static inline void lustre_put_page(struct page *page)
1206 {
1207         kunmap(page);
1208         page_cache_release(page);
1209 }
1210
1211
1212 static struct page *
1213 lustre_get_page_read(struct inode *inode, struct niobuf_remote *rnb)
1214 {
1215         unsigned long index = rnb->offset >> PAGE_SHIFT;
1216         struct address_space *mapping = inode->i_mapping;
1217         struct page *page;
1218         int rc;
1219
1220         page = read_cache_page(mapping, index,
1221                                (filler_t*)mapping->a_ops->readpage, NULL);
1222         if (!IS_ERR(page)) {
1223                 wait_on_page(page);
1224                 kmap(page);
1225                 if (!PageUptodate(page)) {
1226                         CERROR("page index %lu not uptodate\n", index);
1227                         GOTO(err_page, rc = -EIO);
1228                 }
1229                 if (PageError(page)) {
1230                         CERROR("page index %lu has error\n", index);
1231                         GOTO(err_page, rc = -EIO);
1232                 }
1233         }
1234         return page;
1235
1236 err_page:
1237         lustre_put_page(page);
1238         return ERR_PTR(rc);
1239 }
1240
1241 static struct page *
1242 lustre_get_page_write(struct inode *inode, unsigned long index)
1243 {
1244         struct address_space *mapping = inode->i_mapping;
1245         struct page *page;
1246         int rc;
1247
1248         page = grab_cache_page(mapping, index); /* locked page */
1249
1250         if (!IS_ERR(page)) {
1251                 kmap(page);
1252                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1253                  * a no-op for most filesystems, because we write the whole
1254                  * page.  For partial-page I/O this will read in the page.
1255                  */
1256                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1257                 if (rc) {
1258                         CERROR("page index %lu, rc = %d\n", index, rc);
1259                         if (rc != -ENOSPC)
1260                                 LBUG();
1261                         GOTO(err_unlock, rc);
1262                 }
1263                 /* XXX not sure if we need this if we are overwriting page */
1264                 if (PageError(page)) {
1265                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1266                         LBUG();
1267                         GOTO(err_unlock, rc = -EIO);
1268                 }
1269         }
1270         return page;
1271
1272 err_unlock:
1273         unlock_page(page);
1274         lustre_put_page(page);
1275         return ERR_PTR(rc);
1276 }
1277
1278 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1279 {
1280         struct inode *inode = page->mapping->host;
1281         int err;
1282
1283         err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1284 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1285         if (!err && IS_SYNC(inode))
1286                 err = waitfor_one_page(page);
1287 #else
1288 #warning ADD 2.5 waiting code here?
1289 #endif
1290         //SetPageUptodate(page); // the client commit_write will do this
1291
1292         SetPageReferenced(page);
1293         unlock_page(page);
1294         lustre_put_page(page);
1295         return err;
1296 }
1297
1298 struct page *filter_get_page_write(struct inode *inode,
1299                                    struct niobuf_remote *rnb,
1300                                    struct niobuf_local *lnb, int *pglocked)
1301 {
1302         unsigned long index = rnb->offset >> PAGE_SHIFT;
1303         struct address_space *mapping = inode->i_mapping;
1304
1305         struct page *page;
1306         int rc;
1307
1308         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1309         if (*pglocked)
1310                 page = grab_cache_page_nowait(mapping, index); /* locked page */
1311         else
1312                 page = grab_cache_page(mapping, index); /* locked page */
1313
1314
1315         /* This page is currently locked, so get a temporary page instead. */
1316         /* XXX I believe this is a very dangerous thing to do - consider if
1317          *     we had multiple writers for the same file (definitely the case
1318          *     if we are using this codepath).  If writer A locks the page,
1319          *     writer B writes to a copy (as here), writer A drops the page
1320          *     lock, and writer C grabs the lock before B does, then B will
1321          *     later overwrite the data from C, even if C had LDLM locked
1322          *     and initiated the write after B did.
1323          */
1324         if (!page) {
1325                 unsigned long addr;
1326                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1327                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1328                 if (!addr) {
1329                         CERROR("no memory for a temp page\n");
1330                         LBUG();
1331                         GOTO(err, rc = -ENOMEM);
1332                 }
1333                 /* XXX debugging */
1334                 memset((void *)addr, 0xBA, PAGE_SIZE);
1335                 page = virt_to_page(addr);
1336                 kmap(page);
1337                 page->index = index;
1338                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1339         } else if (!IS_ERR(page)) {
1340                 (*pglocked)++;
1341                 kmap(page);
1342
1343                 rc = mapping->a_ops->prepare_write(NULL, page,
1344                                                    rnb->offset % PAGE_SIZE,
1345                                                    rnb->len);
1346                 if (rc) {
1347                         CERROR("page index %lu, rc = %d\n", index, rc);
1348                         if (rc != -ENOSPC)
1349                                 LBUG();
1350                         GOTO(err_unlock, rc);
1351                 }
1352                 /* XXX not sure if we need this if we are overwriting page */
1353                 if (PageError(page)) {
1354                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1355                         LBUG();
1356                         GOTO(err_unlock, rc = -EIO);
1357                 }
1358         }
1359         return page;
1360
1361 err_unlock:
1362         unlock_page(page);
1363         lustre_put_page(page);
1364 err:
1365         return ERR_PTR(rc);
1366 }
1367
1368 /*
1369  * We need to balance prepare_write() calls with commit_write() calls.
1370  * If the page has been prepared, but we have no data for it, we don't
1371  * want to overwrite valid data on disk, but we still need to zero out
1372  * data for space which was newly allocated.  Like part of what happens
1373  * in __block_prepare_write() for newly allocated blocks.
1374  *
1375  * XXX currently __block_prepare_write() creates buffers for all the
1376  *     pages, and the filesystems mark these buffers as BH_New if they
1377  *     were newly allocated from disk. We use the BH_New flag similarly.
1378  */
1379 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1380                                int err)
1381 {
1382 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1383         if (err) {
1384                 unsigned block_start, block_end;
1385                 struct buffer_head *bh, *head = page->buffers;
1386                 unsigned blocksize = head->b_size;
1387                 void *addr = page_address(page);
1388
1389                 /* debugging: just seeing if this ever happens */
1390                 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1391                        page->index, page->mapping->host->i_ino, err);
1392
1393                 /* Currently one buffer per page, but in the future... */
1394                 for (bh = head, block_start = 0; bh != head || !block_start;
1395                      block_start = block_end, bh = bh->b_this_page) {
1396                         block_end = block_start + blocksize;
1397                         if (buffer_new(bh))
1398                                 memset(addr + block_start, 0, blocksize);
1399                 }
1400         }
1401 #endif
1402         return lustre_commit_write(page, from, to);
1403 }
1404
1405 static int filter_preprw(int cmd, struct lustre_handle *conn,
1406                          int objcount, struct obd_ioobj *obj,
1407                          int niocount, struct niobuf_remote *nb,
1408                          struct niobuf_local *res, void **desc_private)
1409 {
1410         struct obd_run_ctxt saved;
1411         struct obd_device *obd;
1412         struct obd_ioobj *o = obj;
1413         struct niobuf_remote *rnb = nb;
1414         struct niobuf_local *lnb = res;
1415         void *journal_save = NULL;
1416         int pglocked = 0;
1417         int rc = 0;
1418         int i;
1419         ENTRY;
1420
1421         obd = class_conn2obd(conn);
1422         if (!obd) {
1423                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1424                 RETURN(-EINVAL);
1425         }
1426         memset(res, 0, sizeof(*res) * niocount);
1427
1428         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1429
1430         if (cmd & OBD_BRW_WRITE) {
1431                 *desc_private = filter_journal_start(&journal_save,
1432                                                      &obd->u.filter,
1433                                                      objcount, obj, niocount,
1434                                                      nb);
1435                 if (IS_ERR(*desc_private))
1436                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1437         }
1438
1439         for (i = 0; i < objcount; i++, o++) {
1440                 struct dentry *dentry;
1441                 struct inode *inode;
1442                 int j;
1443
1444                 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1445                                            o->ioo_id, S_IFREG);
1446                 if (IS_ERR(dentry))
1447                         GOTO(out_clean, rc = PTR_ERR(dentry));
1448                 inode = dentry->d_inode;
1449                 if (!inode) {
1450                         CERROR("trying to BRW to non-existent file "LPU64"\n",
1451                                o->ioo_id);
1452                         f_dput(dentry);
1453                         GOTO(out_clean, rc = -ENOENT);
1454                 }
1455
1456                 for (j = 0; j < o->ioo_bufcnt; j++, rnb++, lnb++) {
1457                         struct page *page;
1458
1459                         if (j == 0)
1460                                 lnb->dentry = dentry;
1461                         else
1462                                 lnb->dentry = dget(dentry);
1463
1464                         if (cmd & OBD_BRW_WRITE)
1465                                 page = filter_get_page_write(inode, rnb, lnb,
1466                                                              &pglocked);
1467                         else
1468                                 page = lustre_get_page_read(inode, rnb);
1469
1470                         if (IS_ERR(page)) {
1471                                 f_dput(dentry);
1472                                 GOTO(out_clean, rc = PTR_ERR(page));
1473                         }
1474
1475                         lnb->addr = page_address(page);
1476                         lnb->offset = rnb->offset;
1477                         lnb->page = page;
1478                         lnb->len = rnb->len;
1479                 }
1480         }
1481
1482 out_stop:
1483         if (cmd & OBD_BRW_WRITE) {
1484                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1485                                               *desc_private);
1486                 if (!rc)
1487                         rc = err;
1488         }
1489 out_ctxt:
1490         pop_ctxt(&saved);
1491         RETURN(rc);
1492 out_clean:
1493         while (lnb-- > res) {
1494                 CERROR("error cleanup on brw\n");
1495                 f_dput(lnb->dentry);
1496                 if (cmd & OBD_BRW_WRITE)
1497                         filter_commit_write(lnb->page, 0, PAGE_SIZE, rc);
1498                 else
1499                         lustre_put_page(lnb->page);
1500         }
1501         goto out_stop;
1502 }
1503
1504 static int filter_write_locked_page(struct niobuf_local *lnb)
1505 {
1506         struct page *lpage;
1507         int rc;
1508
1509         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1510         if (IS_ERR(lpage)) {
1511                 /* It is highly unlikely that we would ever get an error here.
1512                  * The page we want to get was previously locked, so it had to
1513                  * have already allocated the space, and we were just writing
1514                  * over the same data, so there would be no hole in the file.
1515                  *
1516                  * XXX: possibility of a race with truncate could exist, need
1517                  *      to check that.  There are no guarantees w.r.t.
1518                  *      write order even on a local filesystem, although the
1519                  *      normal response would be to return the number of bytes
1520                  *      successfully written and leave the rest to the app.
1521                  */
1522                 rc = PTR_ERR(lpage);
1523                 CERROR("error getting locked page index %ld: rc = %d\n",
1524                        lnb->page->index, rc);
1525                 GOTO(out, rc);
1526         }
1527
1528         /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1529          * lustre_commit_write() below, lnb->page was kmapped previously in
1530          * filter_get_page_write() and kunmapped in lustre_put_page() below.
1531          */
1532         memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1533         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1534         if (rc)
1535                 CERROR("error committing locked page %ld: rc = %d\n",
1536                        lnb->page->index, rc);
1537 out:
1538         lustre_put_page(lnb->page);
1539
1540         return rc;
1541 }
1542
1543 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1544                            int objcount, struct obd_ioobj *obj,
1545                            int niocount, struct niobuf_local *res,
1546                            void *private)
1547 {
1548         struct obd_run_ctxt saved;
1549         struct obd_ioobj *o;
1550         struct niobuf_local *r;
1551         struct obd_device *obd = class_conn2obd(conn);
1552         void *journal_save;
1553         int found_locked = 0;
1554         int rc = 0;
1555         int i;
1556         ENTRY;
1557
1558         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1559         lock_kernel();
1560         journal_save = current->journal_info;
1561         LASSERT(!journal_save);
1562
1563         current->journal_info = private;
1564         unlock_kernel();
1565         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1566                 int j;
1567                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1568                         struct page *page = r->page;
1569
1570                         if (!page)
1571                                 LBUG();
1572
1573                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1574                                 found_locked++;
1575                                 continue;
1576                         }
1577
1578                         if (cmd & OBD_BRW_WRITE) {
1579                                 int err = filter_commit_write(page, 0,
1580                                                               r->len, 0);
1581
1582                                 if (!rc)
1583                                         rc = err;
1584                         } else
1585                                 lustre_put_page(page);
1586
1587                         f_dput(r->dentry);
1588                 }
1589         }
1590         lock_kernel();
1591         current->journal_info = journal_save;
1592         unlock_kernel();
1593
1594         if (!found_locked)
1595                 goto out_ctxt;
1596
1597         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1598                 int j;
1599                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1600                         int err;
1601                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1602                                 continue;
1603
1604                         err = filter_write_locked_page(r);
1605                         if (!rc)
1606                                 rc = err;
1607                         f_dput(r->dentry);
1608                 }
1609         }
1610
1611 out_ctxt:
1612         pop_ctxt(&saved);
1613         RETURN(rc);
1614 }
1615
1616 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1617 {
1618         struct obd_device *obd = class_conn2obd(conn);
1619         struct statfs sfs;
1620         int rc;
1621
1622         ENTRY;
1623         rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1624         if (!rc)
1625                 statfs_pack(osfs, &sfs);
1626
1627         return rc;
1628 }
1629
1630 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1631                            void *key, obd_count *vallen, void **val)
1632 {
1633         struct obd_device *obd;
1634         ENTRY;
1635
1636         obd = class_conn2obd(conn);
1637         if (!obd) {
1638                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1639                 RETURN(-EINVAL);
1640         }
1641
1642         if ( keylen == strlen("blocksize") &&
1643              memcmp(key, "blocksize", keylen) == 0 ) {
1644                 *vallen = sizeof(long);
1645                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1646                 RETURN(0);
1647         }
1648
1649         if ( keylen == strlen("blocksize_bits") &&
1650              memcmp(key, "blocksize_bits", keylen) == 0 ){
1651                 *vallen = sizeof(long);
1652                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1653                 RETURN(0);
1654         }
1655
1656         if ( keylen == strlen("root_ino") &&
1657              memcmp(key, "root_ino", keylen) == 0 ){
1658                 *vallen = sizeof(obd_id);
1659                 *val = (void *)(obd_id)FILTER_ROOTINO;
1660                 RETURN(0);
1661         }
1662
1663         CDEBUG(D_IOCTL, "invalid key\n");
1664         RETURN(-EINVAL);
1665 }
1666
1667 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1668                   struct lustre_handle *src_conn, struct obdo *src,
1669                   obd_size count, obd_off offset)
1670 {
1671         struct page *page;
1672         struct lov_stripe_md srcmd, dstmd;
1673         unsigned long index = 0;
1674         int err = 0;
1675
1676         memset(&srcmd, 0, sizeof(srcmd));
1677         memset(&dstmd, 0, sizeof(dstmd));
1678         srcmd.lsm_object_id = src->o_id;
1679         dstmd.lsm_object_id = dst->o_id;
1680
1681         ENTRY;
1682         CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1683                ", dst: ino "LPU64"\n",
1684                src->o_id, src->o_blocks, src->o_size, dst->o_id);
1685         page = alloc_page(GFP_USER);
1686         if (page == NULL)
1687                 RETURN(-ENOMEM);
1688
1689 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1690         while (TryLockPage(page))
1691                 ___wait_on_page(page);
1692 #else
1693         wait_on_page_locked(page);
1694 #endif
1695
1696         /* XXX with brw vector I/O, we could batch up reads and writes here,
1697          *     all we need to do is allocate multiple pages to handle the I/Os
1698          *     and arrays to handle the request parameters.
1699          */
1700         while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1701                 struct brw_page pg;
1702                 struct io_cb_data *cbd = ll_init_cb();
1703
1704                 if (!cbd) {
1705                         err = -ENOMEM;
1706                         EXIT;
1707                         break;
1708                 }
1709
1710                 pg.pg = page;
1711                 pg.count = PAGE_SIZE;
1712                 pg.off = (page->index) << PAGE_SHIFT;
1713                 pg.flag = 0;
1714
1715                 page->index = index;
1716                 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1717                               ll_sync_io_cb, cbd);
1718
1719                 if ( err ) {
1720                         EXIT;
1721                         break;
1722                 }
1723
1724                 cbd = ll_init_cb();
1725                 if (!cbd) {
1726                         err = -ENOMEM;
1727                         EXIT;
1728                         break;
1729                 }
1730                 pg.flag = OBD_BRW_CREATE;
1731                 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1732
1733                 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1734                               ll_sync_io_cb, cbd);
1735
1736                 /* XXX should handle dst->o_size, dst->o_blocks here */
1737                 if ( err ) {
1738                         EXIT;
1739                         break;
1740                 }
1741
1742                 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1743
1744                 index++;
1745         }
1746         dst->o_size = src->o_size;
1747         dst->o_blocks = src->o_blocks;
1748         dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1749         unlock_page(page);
1750         __free_page(page);
1751
1752         RETURN(err);
1753 }
1754
1755
1756 static struct obd_ops filter_obd_ops = {
1757         o_get_info:    filter_get_info,
1758         o_setup:       filter_setup,
1759         o_cleanup:     filter_cleanup,
1760         o_connect:     filter_connect,
1761         o_disconnect:  filter_disconnect,
1762         o_statfs:      filter_statfs,
1763         o_getattr:     filter_getattr,
1764         o_create:      filter_create,
1765         o_setattr:     filter_setattr,
1766         o_destroy:     filter_destroy,
1767         o_open:        filter_open,
1768         o_close:       filter_close,
1769         o_brw:         filter_pgcache_brw,
1770         o_punch:       filter_truncate,
1771         o_preprw:      filter_preprw,
1772         o_commitrw:    filter_commitrw
1773 #if 0
1774         o_preallocate: filter_preallocate_inodes,
1775         o_migrate:     filter_migrate,
1776         o_copy:        filter_copy_data,
1777         o_iterate:     filter_iterate
1778 #endif
1779 };
1780
1781
1782 static int __init obdfilter_init(void)
1783 {
1784         printk(KERN_INFO "Filtering OBD driver  v0.001, info@clusterfs.com\n");
1785         filter_open_cache = kmem_cache_create("ll_filter_fdata",
1786                                               sizeof(struct filter_file_data),
1787                                               0, 0, NULL, NULL);
1788         if (!filter_open_cache)
1789                 RETURN(-ENOMEM);
1790
1791         filter_dentry_cache = kmem_cache_create("ll_filter_dentry",
1792                                         sizeof(struct filter_dentry_data),
1793                                         0, 0, NULL, NULL);
1794         if (!filter_dentry_cache) {
1795                 kmem_cache_destroy(filter_open_cache);
1796                 RETURN(-ENOMEM);
1797         }
1798
1799         return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1800 }
1801
1802 static void __exit obdfilter_exit(void)
1803 {
1804         class_unregister_type(OBD_FILTER_DEVICENAME);
1805         if (kmem_cache_destroy(filter_dentry_cache))
1806                 CERROR("couldn't free obdfilter dentry cache\n");
1807         if (kmem_cache_destroy(filter_open_cache))
1808                 CERROR("couldn't free obdfilter open cache\n");
1809 }
1810
1811 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1812 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1813 MODULE_LICENSE("GPL");
1814
1815 module_init(obdfilter_init);
1816 module_exit(obdfilter_exit);