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