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                 fdd->fdd_flags = 0;
 388                 /* If this is racy, then we can use {cmp}xchg and atomic_add */
 389                 file->f_dentry->d_fsdata = fdd;
 390                 spin_unlock(&filter->fo_fddlock);
 391         }
 392
 393         get_random_bytes(&ffd->ffd_servercookie, sizeof(ffd->ffd_servercookie));
 394         ffd->ffd_file = file;
 395         file->private_data = ffd;
 396
 397         spin_lock(&fed->fed_lock);
 398         list_add(&ffd->ffd_export_list, &fed->fed_open_head);
 399         spin_unlock(&fed->fed_lock);
 400
 401         CDEBUG(D_INODE, "opening objid "LPX64": rc = %p\n", id, file);
 402
 403 out:
 404         RETURN(file);
 405
 406 out_fdd:
 407         kmem_cache_free(filter_dentry_cache, fdd);
 408 out_ffd:
 409         ffd->ffd_servercookie = DEAD_HANDLE_MAGIC;
 410         kmem_cache_free(filter_open_cache, ffd);
 411         goto out;
 412 }
 413
 414 /* Caller must hold i_sem on dir_dentry->d_inode */
 415 static int filter_destroy_internal(struct obd_device *obd,
 416                                    struct dentry *dir_dentry,
 417                                    struct dentry *object_dentry)
 418 {
 419         struct obd_run_ctxt saved;
 420         struct inode *inode = object_dentry->d_inode;
 421         int rc;
 422         ENTRY;
 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         struct filter_dentry_data *fdd = object_dentry->d_fsdata;
 882         int rc;
 883         ENTRY;
 884
 885         if (!obd) {
 886                 CERROR("invalid client "LPX64"\n", conn->addr);
 887                 RETURN(-EINVAL);
 888         }
 889
 890         CDEBUG(D_INODE, "destroying objid "LPX64"\n", oa->o_id);
 891
 892         dir_dentry = filter_parent(obd, oa->o_mode);
 893         down(&dir_dentry->d_inode->i_sem);
 894
 895         object_dentry = filter_oa2dentry(conn, oa);
 896         if (IS_ERR(object_dentry))
 897                 GOTO(out, rc = -ENOENT);
 898
 899         if (fdd && atomic_read(&fdd->fdd_open_count)) {
 900                 struct filter_dentry_data *fdd = object_dentry->d_fsdata;
 901
 902                 if (!(fdd->fdd_flags & FILTER_FLAG_DESTROY)) {
 903                         fdd->fdd_flags |= FILTER_FLAG_DESTROY;
 904                         /* XXX put into PENDING directory in case of crash */
 905                         CDEBUG(D_INODE,
 906                                "defer destroy of %dx open objid "LPX64"\n",
 907                                atomic_read(&fdd->fdd_open_count), oa->o_id);
 908                 } else
 909                         CDEBUG(D_INODE,
 910                                "repeat destroy of %dx open objid "LPX64"\n",
 911                                atomic_read(&fdd->fdd_open_count), oa->o_id);
 912                 GOTO(out_dput, rc = 0);
 913         }
 914
 915         rc = filter_destroy_internal(obd, dir_dentry, object_dentry);
 916 out_dput:
 917         f_dput(object_dentry);
 918
 919         EXIT;
 920 out:
 921         up(&dir_dentry->d_inode->i_sem);
 922         return rc;
 923 }
 924
 925 /* NB count and offset are used for punch, but not truncate */
 926 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
 927                            struct lov_stripe_md *lsm,
 928                            obd_off start, obd_off end)
 929 {
 930         int error;
 931         ENTRY;
 932
 933         if (end != OBD_OBJECT_EOF)
 934                 CERROR("PUNCH not supported, only truncate works\n");
 935
 936         CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
 937                "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
 938         oa->o_size = start;
 939         error = filter_setattr(conn, oa, NULL);
 940         RETURN(error);
 941 }
 942
 943 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
 944                               struct lov_stripe_md *lsm, obd_count oa_bufs,
 945                               struct brw_page *pga, brw_callback_t callback,
 946                               struct io_cb_data *data)
 947 {
 948         struct obd_export       *export = class_conn2export(conn);
 949         struct obd_run_ctxt      saved;
 950         struct super_block      *sb;
 951         int                      pnum;          /* index to pages (bufs) */
 952         unsigned long            retval;
 953         int                      error;
 954         struct file             *file;
 955         int pg;
 956         ENTRY;
 957
 958         if (!export) {
 959                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
 960                 RETURN(-EINVAL);
 961         }
 962
 963         sb = export->exp_obd->u.filter.fo_sb;
 964         push_ctxt(&saved, &export->exp_obd->u.filter.fo_ctxt, NULL);
 965         pnum = 0; /* pnum indexes buf 0..num_pages */
 966
 967         file = filter_obj_open(export, lsm->lsm_object_id, S_IFREG);
 968         if (IS_ERR(file))
 969                 GOTO(out, retval = PTR_ERR(file));
 970
 971         /* count doubles as retval */
 972         for (pg = 0; pg < oa_bufs; pg++) {
 973                 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
 974                        ") off count ("LPU64",%d)\n",
 975                        cmd, pnum, file->f_dentry->d_inode->i_ino,
 976                        pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
 977                        (int)pga[pnum].count);
 978                 if (cmd & OBD_BRW_WRITE) {
 979                         loff_t off;
 980                         char *buffer;
 981                         off = pga[pnum].off;
 982                         buffer = kmap(pga[pnum].pg);
 983                         retval = file->f_op->write(file, buffer,
 984                                                    pga[pnum].count,
 985                                                    &off);
 986                         kunmap(pga[pnum].pg);
 987                         CDEBUG(D_INODE, "retval %ld\n", retval);
 988                 } else {
 989                         loff_t off = pga[pnum].off;
 990                         char *buffer = kmap(pga[pnum].pg);
 991
 992                         if (off >= file->f_dentry->d_inode->i_size) {
 993                                 memset(buffer, 0, pga[pnum].count);
 994                                 retval = pga[pnum].count;
 995                         } else {
 996                                 retval = file->f_op->read(file, buffer,
 997                                                           pga[pnum].count, &off);
 998                         }
 999                         kunmap(pga[pnum].pg);
1000
1001                         if (retval != pga[pnum].count) {
1002                                 filp_close(file, 0);
1003                                 GOTO(out, retval = -EIO);
1004                         }
1005                         CDEBUG(D_INODE, "retval %ld\n", retval);
1006                 }
1007                 pnum++;
1008         }
1009         /* sizes and blocks are set by generic_file_write */
1010         /* ctimes/mtimes will follow with a setattr call */
1011         filp_close(file, 0);
1012
1013         /* XXX: do something with callback if it is set? */
1014
1015         EXIT;
1016 out:
1017         pop_ctxt(&saved);
1018         error = (retval >= 0) ? 0 : retval;
1019         return error;
1020 }
1021
1022 /*
1023  * Calculate the number of buffer credits needed to write multiple pages in
1024  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
1025  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
1026  *
1027  * See comment above ext3_writepage_trans_blocks for details.  We assume
1028  * no data journaling is being done, but it does allow for all of the pages
1029  * being non-contiguous.  If we are guaranteed contiguous pages we could
1030  * reduce the number of (d)indirect blocks a lot.
1031  *
1032  * With N blocks per page and P pages, for each inode we have at most:
1033  * N*P indirect
1034  * min(N*P, blocksize/4 + 1) dindirect blocks
1035  * 1 tindirect
1036  *
1037  * For the entire filesystem, we have at most:
1038  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
1039  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
1040  * 1 inode block
1041  * 1 superblock
1042  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
1043  */
1044 static int ext3_credits_needed(struct super_block *sb, int objcount,
1045                                struct obd_ioobj *obj)
1046 {
1047         struct obd_ioobj *o = obj;
1048         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
1049         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
1050         int nbitmaps = 0;
1051         int ngdblocks = 0;
1052         int needed = objcount + 1;
1053         int i;
1054
1055         for (i = 0; i < objcount; i++, o++) {
1056                 int nblocks = o->ioo_bufcnt * blockpp;
1057                 int ndindirect = min(nblocks, addrpp + 1);
1058                 int nindir = nblocks + ndindirect + 1;
1059
1060                 nbitmaps += nindir + nblocks;
1061                 ngdblocks += nindir + nblocks;
1062
1063                 needed += nindir;
1064         }
1065
1066         /* Assumes ext3 and extN have same sb_info layout at the start. */
1067         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
1068                 nbitmaps = EXT3_SB(sb)->s_groups_count;
1069         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
1070                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
1071
1072         needed += nbitmaps + ngdblocks;
1073
1074 #ifdef CONFIG_QUOTA
1075         /* We assume that there will be 1 bit set in s_dquot.flags for each
1076          * quota file that is active.  This is at least true for now.
1077          */
1078         needed += hweight32(sb_any_quota_enabled(sb)) *
1079                 EXT3_SINGLEDATA_TRANS_BLOCKS;
1080 #endif
1081
1082         return needed;
1083 }
1084
1085 /* We have to start a huge journal transaction here to hold all of the
1086  * metadata for the pages being written here.  This is necessitated by
1087  * the fact that we do lots of prepare_write operations before we do
1088  * any of the matching commit_write operations, so even if we split
1089  * up to use "smaller" transactions none of them could complete until
1090  * all of them were opened.  By having a single journal transaction,
1091  * we eliminate duplicate reservations for common blocks like the
1092  * superblock and group descriptors or bitmaps.
1093  *
1094  * We will start the transaction here, but each prepare_write will
1095  * add a refcount to the transaction, and each commit_write will
1096  * remove a refcount.  The transaction will be closed when all of
1097  * the pages have been written.
1098  */
1099 static void *ext3_filter_journal_start(struct filter_obd *filter,
1100                                        int objcount, struct obd_ioobj *obj,
1101                                        int niocount, struct niobuf_remote *nb)
1102 {
1103         journal_t *journal = NULL;
1104         handle_t *handle = NULL;
1105         int needed;
1106
1107         /* It appears that some kernels have different values for
1108          * EXT*_MAX_GROUP_LOADED (either 8 or 32), so we cannot
1109          * assume anything after s_inode_bitmap_number is the same.
1110          */
1111         if (!strcmp(filter->fo_fstype, "ext3"))
1112                 journal = EXT3_SB(filter->fo_sb)->s_journal;
1113 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1114         else if (!strcmp(filter->fo_fstype, "extN"))
1115                 journal = EXTN_SB(filter->fo_sb)->s_journal;
1116 #endif
1117         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
1118
1119         /* The number of blocks we could _possibly_ dirty can very large.
1120          * We reduce our request if it is absurd (and we couldn't get that
1121          * many credits for a single handle anyways).
1122          *
1123          * At some point we have to limit the size of I/Os sent at one time,
1124          * increase the size of the journal, or we have to calculate the
1125          * actual journal requirements more carefully by checking all of
1126          * the blocks instead of being maximally pessimistic.  It remains to
1127          * be seen if this is a real problem or not.
1128          */
1129         if (needed > journal->j_max_transaction_buffers) {
1130                 CERROR("want too many journal credits (%d) using %d instead\n",
1131                        needed, journal->j_max_transaction_buffers);
1132                 needed = journal->j_max_transaction_buffers;
1133         }
1134
1135         lock_kernel();
1136         handle = journal_start(journal, needed);
1137         unlock_kernel();
1138         if (IS_ERR(handle))
1139                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
1140                        PTR_ERR(handle));
1141
1142         return(handle);
1143 }
1144
1145 static void *filter_journal_start(void **journal_save,
1146                                   struct filter_obd *filter,
1147                                   int objcount, struct obd_ioobj *obj,
1148                                   int niocount, struct niobuf_remote *nb)
1149 {
1150         void *handle = NULL;
1151
1152         /* This may not be necessary - we probably never have a
1153          * transaction started when we enter here, so we can
1154          * remove the saving of the journal state entirely.
1155          * For now leave it in just to see if it ever happens.
1156          */
1157         *journal_save = current->journal_info;
1158         if (*journal_save) {
1159                 CERROR("Already have handle %p???\n", *journal_save);
1160                 LBUG();
1161                 current->journal_info = NULL;
1162         }
1163
1164         if (!strcmp(filter->fo_fstype, "ext3") ||
1165             !strcmp(filter->fo_fstype, "extN"))
1166                 handle = ext3_filter_journal_start(filter, objcount, obj,
1167                                                    niocount, nb);
1168         return handle;
1169 }
1170
1171 static int ext3_filter_journal_stop(void *handle)
1172 {
1173         int rc;
1174
1175         /* We got a refcount on the handle for each call to prepare_write,
1176          * so we can drop the "parent" handle here to avoid the need for
1177          * osc to call back into filterobd to close the handle.  The
1178          * remaining references will be dropped in commit_write.
1179          */
1180         lock_kernel();
1181         rc = journal_stop((handle_t *)handle);
1182         unlock_kernel();
1183
1184         return rc;
1185 }
1186
1187 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
1188                                void *handle)
1189 {
1190         int rc = 0;
1191
1192         if (!strcmp(filter->fo_fstype, "ext3") ||
1193             !strcmp(filter->fo_fstype, "extN"))
1194                 rc = ext3_filter_journal_stop(handle);
1195
1196         if (rc)
1197                 CERROR("error on journal stop: rc = %d\n", rc);
1198
1199         current->journal_info = journal_save;
1200
1201         return rc;
1202 }
1203
1204 static inline void lustre_put_page(struct page *page)
1205 {
1206         kunmap(page);
1207         page_cache_release(page);
1208 }
1209
1210
1211 #ifndef PageUptodate
1212 #define PageUptodate(page) Page_Uptodate(page)
1213 #endif
1214 static struct page *
1215 lustre_get_page_read(struct inode *inode,
1216                      struct niobuf_remote *rnb)
1217 {
1218         unsigned long index = rnb->offset >> PAGE_SHIFT;
1219         struct address_space *mapping = inode->i_mapping;
1220         struct page *page;
1221         int rc;
1222
1223         page = read_cache_page(mapping, index,
1224                                (filler_t*)mapping->a_ops->readpage, NULL);
1225         if (!IS_ERR(page)) {
1226                 wait_on_page(page);
1227                 kmap(page);
1228                 if (!PageUptodate(page)) {
1229                         CERROR("page index %lu not uptodate\n", index);
1230                         GOTO(err_page, rc = -EIO);
1231                 }
1232                 if (PageError(page)) {
1233                         CERROR("page index %lu has error\n", index);
1234                         GOTO(err_page, rc = -EIO);
1235                 }
1236         }
1237         return page;
1238
1239 err_page:
1240         lustre_put_page(page);
1241         return ERR_PTR(rc);
1242 }
1243
1244 static struct page *
1245 lustre_get_page_write(struct inode *inode, unsigned long index)
1246 {
1247         struct address_space *mapping = inode->i_mapping;
1248         struct page *page;
1249         int rc;
1250
1251         page = grab_cache_page(mapping, index); /* locked page */
1252
1253         if (!IS_ERR(page)) {
1254                 kmap(page);
1255                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1256                  * a no-op for most filesystems, because we write the whole
1257                  * page.  For partial-page I/O this will read in the page.
1258                  */
1259                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1260                 if (rc) {
1261                         CERROR("page index %lu, rc = %d\n", index, rc);
1262                         if (rc != -ENOSPC)
1263                                 LBUG();
1264                         GOTO(err_unlock, rc);
1265                 }
1266                 /* XXX not sure if we need this if we are overwriting page */
1267                 if (PageError(page)) {
1268                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1269                         LBUG();
1270                         GOTO(err_unlock, rc = -EIO);
1271                 }
1272         }
1273         return page;
1274
1275 err_unlock:
1276         unlock_page(page);
1277         lustre_put_page(page);
1278         return ERR_PTR(rc);
1279 }
1280
1281 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1282 {
1283         struct inode *inode = page->mapping->host;
1284         int err;
1285
1286         err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1287 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1288         if (!err && IS_SYNC(inode))
1289                 err = waitfor_one_page(page);
1290 #else
1291 #warning ADD 2.5 waiting code here?
1292 #endif
1293         //SetPageUptodate(page); // the client commit_write will do this
1294
1295         SetPageReferenced(page);
1296         unlock_page(page);
1297         lustre_put_page(page);
1298         return err;
1299 }
1300
1301 struct page *filter_get_page_write(struct inode *inode,
1302                                    struct niobuf_remote *rnb,
1303                                    struct niobuf_local *lnb, int *pglocked)
1304 {
1305         unsigned long index = rnb->offset >> PAGE_SHIFT;
1306         struct address_space *mapping = inode->i_mapping;
1307
1308         struct page *page;
1309         int rc;
1310
1311         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1312         if (*pglocked)
1313                 page = grab_cache_page_nowait(mapping, index); /* locked page */
1314         else
1315                 page = grab_cache_page(mapping, index); /* locked page */
1316
1317
1318         /* This page is currently locked, so get a temporary page instead. */
1319         /* XXX I believe this is a very dangerous thing to do - consider if
1320          *     we had multiple writers for the same file (definitely the case
1321          *     if we are using this codepath).  If writer A locks the page,
1322          *     writer B writes to a copy (as here), writer A drops the page
1323          *     lock, and writer C grabs the lock before B does, then B will
1324          *     later overwrite the data from C, even if C had LDLM locked
1325          *     and initiated the write after B did.
1326          */
1327         if (!page) {
1328                 unsigned long addr;
1329                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1330                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1331                 if (!addr) {
1332                         CERROR("no memory for a temp page\n");
1333                         LBUG();
1334                         GOTO(err, rc = -ENOMEM);
1335                 }
1336                 /* XXX debugging */
1337                 memset((void *)addr, 0xBA, PAGE_SIZE);
1338                 page = virt_to_page(addr);
1339                 kmap(page);
1340                 page->index = index;
1341                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1342         } else if (!IS_ERR(page)) {
1343                 (*pglocked)++;
1344                 kmap(page);
1345
1346                 rc = mapping->a_ops->prepare_write(NULL, page,
1347                                                    rnb->offset % PAGE_SIZE,
1348                                                    rnb->len);
1349                 if (rc) {
1350                         CERROR("page index %lu, rc = %d\n", index, rc);
1351                         if (rc != -ENOSPC)
1352                                 LBUG();
1353                         GOTO(err_unlock, rc);
1354                 }
1355                 /* XXX not sure if we need this if we are overwriting page */
1356                 if (PageError(page)) {
1357                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1358                         LBUG();
1359                         GOTO(err_unlock, rc = -EIO);
1360                 }
1361         }
1362         return page;
1363
1364 err_unlock:
1365         unlock_page(page);
1366         lustre_put_page(page);
1367 err:
1368         return ERR_PTR(rc);
1369 }
1370
1371 /*
1372  * We need to balance prepare_write() calls with commit_write() calls.
1373  * If the page has been prepared, but we have no data for it, we don't
1374  * want to overwrite valid data on disk, but we still need to zero out
1375  * data for space which was newly allocated.  Like part of what happens
1376  * in __block_prepare_write() for newly allocated blocks.
1377  *
1378  * XXX currently __block_prepare_write() creates buffers for all the
1379  *     pages, and the filesystems mark these buffers as BH_New if they
1380  *     were newly allocated from disk. We use the BH_New flag similarly.
1381  */
1382 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1383                                int err)
1384 {
1385 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1386         if (err) {
1387                 unsigned block_start, block_end;
1388                 struct buffer_head *bh, *head = page->buffers;
1389                 unsigned blocksize = head->b_size;
1390                 void *addr = page_address(page);
1391
1392                 /* debugging: just seeing if this ever happens */
1393                 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1394                        page->index, page->mapping->host->i_ino, err);
1395
1396                 /* Currently one buffer per page, but in the future... */
1397                 for (bh = head, block_start = 0; bh != head || !block_start;
1398                      block_start = block_end, bh = bh->b_this_page) {
1399                         block_end = block_start + blocksize;
1400                         if (buffer_new(bh))
1401                                 memset(addr + block_start, 0, blocksize);
1402                 }
1403         }
1404 #endif
1405         return lustre_commit_write(page, from, to);
1406 }
1407
1408 static int filter_preprw(int cmd, struct lustre_handle *conn,
1409                          int objcount, struct obd_ioobj *obj,
1410                          int niocount, struct niobuf_remote *nb,
1411                          struct niobuf_local *res, void **desc_private)
1412 {
1413         struct obd_run_ctxt saved;
1414         struct obd_device *obd;
1415         struct obd_ioobj *o = obj;
1416         struct niobuf_remote *rnb = nb;
1417         struct niobuf_local *lnb = res;
1418         void *journal_save = NULL;
1419         int pglocked = 0;
1420         int rc = 0;
1421         int i;
1422         ENTRY;
1423
1424         obd = class_conn2obd(conn);
1425         if (!obd) {
1426                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1427                 RETURN(-EINVAL);
1428         }
1429         memset(res, 0, sizeof(*res) * niocount);
1430
1431         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1432
1433         if (cmd & OBD_BRW_WRITE) {
1434                 *desc_private = filter_journal_start(&journal_save,
1435                                                      &obd->u.filter,
1436                                                      objcount, obj, niocount,
1437                                                      nb);
1438                 if (IS_ERR(*desc_private))
1439                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1440         }
1441
1442         for (i = 0; i < objcount; i++, o++) {
1443                 struct dentry *dentry;
1444                 struct inode *inode;
1445                 int j;
1446
1447                 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1448                                            o->ioo_id, S_IFREG);
1449                 if (IS_ERR(dentry))
1450                         GOTO(out_clean, rc = PTR_ERR(dentry));
1451                 inode = dentry->d_inode;
1452                 if (!inode) {
1453                         CERROR("trying to BRW to non-existent file "LPU64"\n",
1454                                o->ioo_id);
1455                         f_dput(dentry);
1456                         GOTO(out_clean, rc = -ENOENT);
1457                 }
1458
1459                 for (j = 0; j < o->ioo_bufcnt; j++, rnb++, lnb++) {
1460                         struct page *page;
1461
1462                         if (j == 0)
1463                                 lnb->dentry = dentry;
1464                         else
1465                                 lnb->dentry = dget(dentry);
1466
1467                         if (cmd & OBD_BRW_WRITE)
1468                                 page = filter_get_page_write(inode, rnb, lnb,
1469                                                              &pglocked);
1470                         else
1471                                 page = lustre_get_page_read(inode, rnb);
1472
1473                         if (IS_ERR(page)) {
1474                                 f_dput(dentry);
1475                                 GOTO(out_clean, rc = PTR_ERR(page));
1476                         }
1477
1478                         lnb->addr = page_address(page);
1479                         lnb->offset = rnb->offset;
1480                         lnb->page = page;
1481                         lnb->len = rnb->len;
1482                 }
1483         }
1484
1485 out_stop:
1486         if (cmd & OBD_BRW_WRITE) {
1487                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1488                                               *desc_private);
1489                 if (!rc)
1490                         rc = err;
1491         }
1492 out_ctxt:
1493         pop_ctxt(&saved);
1494         RETURN(rc);
1495 out_clean:
1496         while (lnb-- > res) {
1497                 CERROR("error cleanup on brw\n");
1498                 f_dput(lnb->dentry);
1499                 if (cmd & OBD_BRW_WRITE)
1500                         filter_commit_write(lnb->page, 0, PAGE_SIZE, rc);
1501                 else
1502                         lustre_put_page(lnb->page);
1503         }
1504         goto out_stop;
1505 }
1506
1507 static int filter_write_locked_page(struct niobuf_local *lnb)
1508 {
1509         struct page *lpage;
1510         int rc;
1511
1512         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1513         if (IS_ERR(lpage)) {
1514                 /* It is highly unlikely that we would ever get an error here.
1515                  * The page we want to get was previously locked, so it had to
1516                  * have already allocated the space, and we were just writing
1517                  * over the same data, so there would be no hole in the file.
1518                  *
1519                  * XXX: possibility of a race with truncate could exist, need
1520                  *      to check that.  There are no guarantees w.r.t.
1521                  *      write order even on a local filesystem, although the
1522                  *      normal response would be to return the number of bytes
1523                  *      successfully written and leave the rest to the app.
1524                  */
1525                 rc = PTR_ERR(lpage);
1526                 CERROR("error getting locked page index %ld: rc = %d\n",
1527                        lnb->page->index, rc);
1528                 GOTO(out, rc);
1529         }
1530
1531         /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1532          * lustre_commit_write() below, lnb->page was kmapped previously in
1533          * filter_get_page_write() and kunmapped in lustre_put_page() below.
1534          */
1535         memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1536         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1537         if (rc)
1538                 CERROR("error committing locked page %ld: rc = %d\n",
1539                        lnb->page->index, rc);
1540 out:
1541         lustre_put_page(lnb->page);
1542
1543         return rc;
1544 }
1545
1546 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1547                            int objcount, struct obd_ioobj *obj,
1548                            int niocount, struct niobuf_local *res,
1549                            void *private)
1550 {
1551         struct obd_run_ctxt saved;
1552         struct obd_ioobj *o;
1553         struct niobuf_local *r;
1554         struct obd_device *obd = class_conn2obd(conn);
1555         void *journal_save;
1556         int found_locked = 0;
1557         int rc = 0;
1558         int i;
1559         ENTRY;
1560
1561         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1562         lock_kernel();
1563         journal_save = current->journal_info;
1564         LASSERT(!journal_save);
1565
1566         current->journal_info = private;
1567         unlock_kernel();
1568         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1569                 int j;
1570                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1571                         struct page *page = r->page;
1572
1573                         if (!page)
1574                                 LBUG();
1575
1576                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1577                                 found_locked++;
1578                                 continue;
1579                         }
1580
1581                         if (cmd & OBD_BRW_WRITE) {
1582                                 int err = filter_commit_write(page, 0,
1583                                                               r->len, 0);
1584
1585                                 if (!rc)
1586                                         rc = err;
1587                         } else
1588                                 lustre_put_page(page);
1589
1590                         f_dput(r->dentry);
1591                 }
1592         }
1593         lock_kernel();
1594         current->journal_info = journal_save;
1595         unlock_kernel();
1596
1597         if (!found_locked)
1598                 goto out_ctxt;
1599
1600         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1601                 int j;
1602                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1603                         int err;
1604                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1605                                 continue;
1606
1607                         err = filter_write_locked_page(r);
1608                         if (!rc)
1609                                 rc = err;
1610                         f_dput(r->dentry);
1611                 }
1612         }
1613
1614 out_ctxt:
1615         pop_ctxt(&saved);
1616         RETURN(rc);
1617 }
1618
1619 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1620 {
1621         struct obd_device *obd = class_conn2obd(conn);
1622         struct statfs sfs;
1623         int rc;
1624
1625         ENTRY;
1626         rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1627         if (!rc)
1628                 statfs_pack(osfs, &sfs);
1629
1630         return rc;
1631 }
1632
1633 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1634                            void *key, obd_count *vallen, void **val)
1635 {
1636         struct obd_device *obd;
1637         ENTRY;
1638
1639         obd = class_conn2obd(conn);
1640         if (!obd) {
1641                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1642                 RETURN(-EINVAL);
1643         }
1644
1645         if ( keylen == strlen("blocksize") &&
1646              memcmp(key, "blocksize", keylen) == 0 ) {
1647                 *vallen = sizeof(long);
1648                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1649                 RETURN(0);
1650         }
1651
1652         if ( keylen == strlen("blocksize_bits") &&
1653              memcmp(key, "blocksize_bits", keylen) == 0 ){
1654                 *vallen = sizeof(long);
1655                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1656                 RETURN(0);
1657         }
1658
1659         if ( keylen == strlen("root_ino") &&
1660              memcmp(key, "root_ino", keylen) == 0 ){
1661                 *vallen = sizeof(obd_id);
1662                 *val = (void *)(obd_id)FILTER_ROOTINO;
1663                 RETURN(0);
1664         }
1665
1666         CDEBUG(D_IOCTL, "invalid key\n");
1667         RETURN(-EINVAL);
1668 }
1669
1670 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1671                   struct lustre_handle *src_conn, struct obdo *src,
1672                   obd_size count, obd_off offset)
1673 {
1674         struct page *page;
1675         struct lov_stripe_md srcmd, dstmd;
1676         unsigned long index = 0;
1677         int err = 0;
1678
1679         memset(&srcmd, 0, sizeof(srcmd));
1680         memset(&dstmd, 0, sizeof(dstmd));
1681         srcmd.lsm_object_id = src->o_id;
1682         dstmd.lsm_object_id = dst->o_id;
1683
1684         ENTRY;
1685         CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1686                ", dst: ino "LPU64"\n",
1687                src->o_id, src->o_blocks, src->o_size, dst->o_id);
1688         page = alloc_page(GFP_USER);
1689         if (page == NULL)
1690                 RETURN(-ENOMEM);
1691
1692 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1693         while (TryLockPage(page))
1694                 ___wait_on_page(page);
1695 #else
1696         wait_on_page_locked(page);
1697 #endif
1698
1699         /* XXX with brw vector I/O, we could batch up reads and writes here,
1700          *     all we need to do is allocate multiple pages to handle the I/Os
1701          *     and arrays to handle the request parameters.
1702          */
1703         while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1704                 struct brw_page pg;
1705                 struct io_cb_data *cbd = ll_init_cb();
1706
1707                 if (!cbd) {
1708                         err = -ENOMEM;
1709                         EXIT;
1710                         break;
1711                 }
1712
1713                 pg.pg = page;
1714                 pg.count = PAGE_SIZE;
1715                 pg.off = (page->index) << PAGE_SHIFT;
1716                 pg.flag = 0;
1717
1718                 page->index = index;
1719                 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1720                               ll_sync_io_cb, cbd);
1721
1722                 if ( err ) {
1723                         EXIT;
1724                         break;
1725                 }
1726
1727                 cbd = ll_init_cb();
1728                 if (!cbd) {
1729                         err = -ENOMEM;
1730                         EXIT;
1731                         break;
1732                 }
1733                 pg.flag = OBD_BRW_CREATE;
1734                 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1735
1736                 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1737                               ll_sync_io_cb, cbd);
1738
1739                 /* XXX should handle dst->o_size, dst->o_blocks here */
1740                 if ( err ) {
1741                         EXIT;
1742                         break;
1743                 }
1744
1745                 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1746
1747                 index++;
1748         }
1749         dst->o_size = src->o_size;
1750         dst->o_blocks = src->o_blocks;
1751         dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1752         unlock_page(page);
1753         __free_page(page);
1754
1755         RETURN(err);
1756 }
1757
1758
1759 static struct obd_ops filter_obd_ops = {
1760         o_get_info:    filter_get_info,
1761         o_setup:       filter_setup,
1762         o_cleanup:     filter_cleanup,
1763         o_connect:     filter_connect,
1764         o_disconnect:  filter_disconnect,
1765         o_statfs:      filter_statfs,
1766         o_getattr:     filter_getattr,
1767         o_create:      filter_create,
1768         o_setattr:     filter_setattr,
1769         o_destroy:     filter_destroy,
1770         o_open:        filter_open,
1771         o_close:       filter_close,
1772         o_brw:         filter_pgcache_brw,
1773         o_punch:       filter_truncate,
1774         o_preprw:      filter_preprw,
1775         o_commitrw:    filter_commitrw
1776 #if 0
1777         o_preallocate: filter_preallocate_inodes,
1778         o_migrate:     filter_migrate,
1779         o_copy:        filter_copy_data,
1780         o_iterate:     filter_iterate
1781 #endif
1782 };
1783
1784
1785 static int __init obdfilter_init(void)
1786 {
1787         printk(KERN_INFO "Filtering OBD driver  v0.001, info@clusterfs.com\n");
1788         filter_open_cache = kmem_cache_create("ll_filter_fdata",
1789                                               sizeof(struct filter_file_data),
1790                                               0, 0, NULL, NULL);
1791         if (!filter_open_cache)
1792                 RETURN(-ENOMEM);
1793
1794         filter_dentry_cache = kmem_cache_create("ll_filter_dentry",
1795                                         sizeof(struct filter_dentry_data),
1796                                         0, 0, NULL, NULL);
1797         if (!filter_dentry_cache) {
1798                 kmem_cache_destroy(filter_open_cache);
1799                 RETURN(-ENOMEM);
1800         }
1801
1802         return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1803 }
1804
1805 static void __exit obdfilter_exit(void)
1806 {
1807         class_unregister_type(OBD_FILTER_DEVICENAME);
1808         if (kmem_cache_destroy(filter_dentry_cache))
1809                 CERROR("couldn't free obdfilter dentry cache\n");
1810         if (kmem_cache_destroy(filter_open_cache))
1811                 CERROR("couldn't free obdfilter open cache\n");
1812 }
1813
1814 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1815 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1816 MODULE_LICENSE("GPL");
1817
1818 module_init(obdfilter_init);
1819 module_exit(obdfilter_exit);