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