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_objidock);
 260
 261         /* FIXME: write the lastobjid to disk here */
 262         return id;
 263 }
 264
 265 /* how to get files, dentries, inodes from object id's */
 266 /* parent i_sem is already held if needed for exclusivity */
 267 static struct dentry *filter_fid2dentry(struct obd_device *obd,
 268                                         struct dentry *dparent,
 269                                         __u64 id, __u32 type)
 270 {
 271         struct super_block *sb = obd->u.filter.fo_sb;
 272         struct dentry *dchild;
 273         char name[32];
 274         int len;
 275         ENTRY;
 276
 277         if (!sb || !sb->s_dev) {
 278                 CERROR("fatal: device not initialized.\n");
 279                 RETURN(ERR_PTR(-ENXIO));
 280         }
 281
 282         if (id == 0) {
 283                 CERROR("fatal: invalid object #0\n");
 284                 LBUG();
 285                 RETURN(ERR_PTR(-ESTALE));
 286         }
 287
 288         if (!(type & S_IFMT)) {
 289                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 290                        __FUNCTION__, id, type);
 291                 RETURN(ERR_PTR(-EINVAL));
 292         }
 293
 294         len = sprintf(name, LPU64, id);
 295         CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
 296                name);
 297         dchild = lookup_one_len(name, dparent, len);
 298         if (IS_ERR(dchild)) {
 299                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 300                 RETURN(dchild);
 301         }
 302
 303         if (!dchild->d_op)
 304                 dchild->d_op = &filter_dops;
 305         else
 306                 LASSERT(dchild->d_op == &filter_dops);
 307
 308         CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
 309                obd_mode_to_type(type), name, dchild,
 310                atomic_read(&dchild->d_count));
 311
 312         LASSERT(atomic_read(&dchild->d_count) > 0);
 313
 314         RETURN(dchild);
 315 }
 316
 317 static inline struct dentry *filter_parent(struct obd_device *obd,
 318                                            obd_mode mode)
 319 {
 320         struct filter_obd *filter = &obd->u.filter;
 321
 322         return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
 323 }
 324
 325 static struct file *filter_obj_open(struct obd_export *export,
 326                                     __u64 id, __u32 type)
 327 {
 328         struct filter_obd *filter = &export->exp_obd->u.filter;
 329         struct super_block *sb = filter->fo_sb;
 330         struct filter_export_data *fed = &export->exp_filter_data;
 331         struct filter_dentry_data *fdd;
 332         struct filter_file_data *ffd;
 333         struct obd_run_ctxt saved;
 334         char name[24];
 335         struct file *file;
 336         ENTRY;
 337
 338         if (!sb || !sb->s_dev) {
 339                 CERROR("fatal: device not initialized.\n");
 340                 RETURN(ERR_PTR(-ENXIO));
 341         }
 342
 343         if (!id) {
 344                 CERROR("fatal: invalid obdo "LPU64"\n", id);
 345                 RETURN(ERR_PTR(-ESTALE));
 346         }
 347
 348         if (!(type & S_IFMT)) {
 349                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 350                        __FUNCTION__, id, type);
 351                 RETURN(ERR_PTR(-EINVAL));
 352         }
 353
 354         ffd = kmem_cache_alloc(filter_open_cache, SLAB_KERNEL);
 355         if (!ffd) {
 356                 CERROR("obdfilter: out of memory\n");
 357                 RETURN(ERR_PTR(-ENOMEM));
 358         }
 359
 360         /* We preallocate this to avoid blocking while holding fo_fddlock */
 361         fdd = kmem_cache_alloc(filter_dentry_cache, SLAB_KERNEL);
 362         if (!fdd) {
 363                 CERROR("obdfilter: out of memory\n");
 364                 GOTO(out_ffd, file = ERR_PTR(-ENOMEM));
 365         }
 366
 367         filter_id(name, id, type);
 368         push_ctxt(&saved, &filter->fo_ctxt, NULL);
 369         file = filp_open(name, O_RDONLY | O_LARGEFILE, 0 /* type? */);
 370         pop_ctxt(&saved);
 371
 372         if (IS_ERR(file))
 373                 GOTO(out_fdd, file);
 374
 375         spin_lock(&filter->fo_fddlock);
 376         if (file->f_dentry->d_fsdata) {
 377                 spin_unlock(&filter->fo_fddlock);
 378                 kmem_cache_free(filter_dentry_cache, fdd);
 379                 fdd = file->f_dentry->d_fsdata;
 380                 LASSERT(kmem_cache_validate(filter_dentry_cache, fdd));
 381                 /* should only happen during client recovery */
 382                 if (fdd->fdd_flags & FILTER_FLAG_DESTROY)
 383                         CDEBUG(D_INODE,"opening destroyed object "LPX64"\n",id);
 384                 atomic_inc(&fdd->fdd_open_count);
 385         } else {
 386                 atomic_set(&fdd->fdd_open_count, 1);
 387                 spin_lock_init(&fdd->fdd_lock);
 388                 fdd->fdd_flags = 0;
 389                 /* If this is racy, then we can use {cmp}xchg and atomic_add */
 390                 file->f_dentry->d_fsdata = fdd;
 391                 spin_unlock(&filter->fo_fddlock);
 392         }
 393
 394         get_random_bytes(&ffd->ffd_servercookie, sizeof(ffd->ffd_servercookie));
 395         ffd->ffd_file = file;
 396         file->private_data = ffd;
 397
 398         spin_lock(&fed->fed_lock);
 399         list_add(&ffd->ffd_export_list, &fed->fed_open_head);
 400         spin_unlock(&fed->fed_lock);
 401
 402         CDEBUG(D_INODE, "opening objid "LPX64": rc = %p\n", id, file);
 403
 404 out:
 405         RETURN(file);
 406
 407 out_fdd:
 408         kmem_cache_free(filter_dentry_cache, fdd);
 409 out_ffd:
 410         ffd->ffd_servercookie = DEAD_HANDLE_MAGIC;
 411         kmem_cache_free(filter_open_cache, ffd);
 412         goto out;
 413 }
 414
 415 /* Caller must hold i_sem on dir_dentry->d_inode */
 416 static int filter_destroy_internal(struct obd_device *obd,
 417                                    struct dentry *dir_dentry,
 418                                    struct dentry *object_dentry)
 419 {
 420         struct obd_run_ctxt saved;
 421         struct inode *inode = object_dentry->d_inode;
 422         int rc;
 423
 424         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
 425                 CERROR("destroying objid %*s nlink = %d, count = %d\n",
 426                        object_dentry->d_name.len,
 427                        object_dentry->d_name.name,
 428                        inode->i_nlink, atomic_read(&inode->i_count));
 429         }
 430
 431         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 432         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 433         /* XXX unlink from PENDING directory now too */
 434         pop_ctxt(&saved);
 435
 436         if (rc)
 437                 CERROR("error unlinking objid %*s: rc %d\n",
 438                        object_dentry->d_name.len,
 439                        object_dentry->d_name.name, rc);
 440
 441         return rc;
 442 }
 443
 444 static int filter_close_internal(struct obd_device *obd,
 445                                  struct filter_file_data *ffd)
 446 {
 447         struct file *filp = ffd->ffd_file;
 448         struct dentry *object_dentry = dget(filp->f_dentry);
 449         struct filter_dentry_data *fdd = object_dentry->d_fsdata;
 450         int rc, rc2 = 0;
 451         ENTRY;
 452
 453         LASSERT(filp->private_data == ffd);
 454         LASSERT(fdd);
 455
 456         rc = filp_close(filp, 0);
 457
 458         if (atomic_dec_and_test(&fdd->fdd_open_count) &&
 459             fdd->fdd_flags & FILTER_FLAG_DESTROY) {
 460                 struct dentry *dir_dentry = filter_parent(obd, S_IFREG);
 461
 462                 down(&dir_dentry->d_inode->i_sem);
 463                 rc2 = filter_destroy_internal(obd, dir_dentry, object_dentry);
 464                 if (rc2 && !rc)
 465                         rc = rc2;
 466                 up(&dir_dentry->d_inode->i_sem);
 467         }
 468
 469         f_dput(object_dentry);
 470         kmem_cache_free(filter_open_cache, ffd);
 471
 472         RETURN(rc);
 473 }
 474
 475 /* obd methods */
 476 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
 477                           obd_uuid_t cluuid, struct recovd_obd *recovd,
 478                           ptlrpc_recovery_cb_t recover)
 479 {
 480         struct obd_export *exp;
 481         int rc;
 482
 483         ENTRY;
 484         MOD_INC_USE_COUNT;
 485         rc = class_connect(conn, obd, cluuid);
 486         if (rc)
 487                 GOTO(out_dec, rc);
 488         exp = class_conn2export(conn);
 489         LASSERT(exp);
 490
 491         INIT_LIST_HEAD(&exp->exp_filter_data.fed_open_head);
 492         spin_lock_init(&exp->exp_filter_data.fed_lock);
 493 out:
 494         RETURN(rc);
 495
 496 out_dec:
 497         MOD_DEC_USE_COUNT;
 498         goto out;
 499 }
 500
 501 static int filter_disconnect(struct lustre_handle *conn)
 502 {
 503         struct obd_export *exp = class_conn2export(conn);
 504         struct filter_export_data *fed;
 505         int rc;
 506         ENTRY;
 507
 508         LASSERT(exp);
 509         fed = &exp->exp_filter_data;
 510         spin_lock(&exp->exp_filter_data);
 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(&exp->exp_filter_data);
 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(&exp->exp_filter_data);
 525         }
 526         spin_unlock(&exp->exp_filter_data);
 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_lock);
 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         int rc;
 801         ENTRY;
 802
 803         exp = class_conn2export(conn);
 804         if (!exp) {
 805                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 806                 RETURN(-EINVAL);
 807         }
 808
 809         if (!(oa->o_valid & OBD_MD_FLHANDLE)) {
 810                 CERROR("no handle for close of objid "LPX64"\n", oa->o_id);
 811                 RETURN(-EINVAL);
 812         }
 813
 814         ffd = filter_handle2ffd(obdo_handle(oa));
 815         if (!ffd) {
 816                 struct lustre_handle *handle = obdo_handle(oa);
 817                 CERROR("bad handle ("LPX64") or cookie ("LPX64") for close\n",
 818                        handle->addr, handle->cookie);
 819                 RETURN(-ESTALE);
 820         }
 821
 822         spin_lock(&exp->exp_filter_data);
 823         list_del(&ffd->ffd_export_list);
 824         spin_unlock(&exp->exp_filter_data);
 825
 826         rc = filter_close_internal(exp->exp_obd, ffd);
 827
 828         RETURN(rc);
 829 } /* filter_close */
 830
 831 static int filter_create(struct lustre_handle* conn, struct obdo *oa,
 832                          struct lov_stripe_md **ea)
 833 {
 834         struct obd_device *obd = class_conn2obd(conn);
 835         char name[64];
 836         struct obd_run_ctxt saved;
 837         struct dentry *new;
 838         int mode;
 839         struct iattr;
 840         ENTRY;
 841
 842         if (!obd) {
 843                 CERROR("invalid client "LPX64"\n", conn->addr);
 844                 return -EINVAL;
 845         }
 846
 847         if (!(oa->o_mode && S_IFMT)) {
 848                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 849                        __FUNCTION__, oa->o_id, oa->o_mode);
 850                 return -ENOENT;
 851         }
 852
 853         oa->o_id = filter_next_id(obd);
 854
 855         //filter_id(name, oa->o_id, oa->o_mode);
 856         sprintf(name, LPU64, oa->o_id);
 857         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 858         new = simple_mknod(filter_parent(obd, oa->o_mode), name, mode);
 859         pop_ctxt(&saved);
 860         if (IS_ERR(new)) {
 861                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
 862                 return -ENOENT;
 863         }
 864
 865         /* Set flags for fields we have set in the inode struct */
 866         oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 867                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
 868         filter_from_inode(oa, new->d_inode, oa->o_valid);
 869         f_dput(new);
 870
 871         return 0;
 872 }
 873
 874 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
 875                           struct lov_stripe_md *ea)
 876 {
 877         struct obd_device *obd = class_conn2obd(conn);
 878         struct dentry *dir_dentry, *object_dentry;
 879         int rc;
 880         ENTRY;
 881
 882         if (!obd) {
 883                 CERROR("invalid client "LPX64"\n", conn->addr);
 884                 RETURN(-EINVAL);
 885         }
 886
 887         CDEBUG(D_INODE, "destroying object "LPD64"\n", oa->o_id);
 888
 889         dir_dentry = filter_parent(obd, oa->o_mode);
 890         down(&dir_dentry->d_inode->i_sem);
 891
 892         object_dentry = filter_oa2dentry(conn, oa);
 893         if (IS_ERR(object_dentry))
 894                 GOTO(out, rc = -ENOENT);
 895
 896         if (object_dentry->d_fsdata) {
 897                 struct filter_dentry_data *fdd = object_dentry->d_fsdata;
 898
 899                 if (!(fdd->fdd_flags & FILTER_FLAG_DESTROY)) {
 900                         fdd->fdd_flags |= FILTER_FLAG_DESTROY;
 901                         /* XXX put into PENDING directory in case of crash */
 902                 }
 903                 GOTO(out_dput, rc = 0);
 904         }
 905
 906         rc = filter_destroy_internal(obd, dir_dentry, object_dentry);
 907 out_dput:
 908         f_dput(object_dentry);
 909
 910         EXIT;
 911 out:
 912         up(&dir_dentry->d_inode->i_sem);
 913         return rc;
 914 }
 915
 916 /* NB count and offset are used for punch, but not truncate */
 917 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
 918                            struct lov_stripe_md *lsm,
 919                            obd_off start, obd_off end)
 920 {
 921         int error;
 922         ENTRY;
 923
 924         if (end != OBD_OBJECT_EOF)
 925                 CERROR("PUNCH not supported, only truncate works\n");
 926
 927         CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
 928                "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
 929         oa->o_size = start;
 930         error = filter_setattr(conn, oa, NULL);
 931         RETURN(error);
 932 }
 933
 934 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
 935                               struct lov_stripe_md *lsm, obd_count oa_bufs,
 936                               struct brw_page *pga, brw_callback_t callback,
 937                               struct io_cb_data *data)
 938 {
 939         struct obd_export       *export = class_conn2export(conn);
 940         struct obd_run_ctxt      saved;
 941         struct super_block      *sb;
 942         int                      pnum;          /* index to pages (bufs) */
 943         unsigned long            retval;
 944         int                      error;
 945         struct file             *file;
 946         int pg;
 947         ENTRY;
 948
 949         if (!export) {
 950                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
 951                 RETURN(-EINVAL);
 952         }
 953
 954         sb = export->exp_obd->u.filter.fo_sb;
 955         push_ctxt(&saved, &export->exp_obd->u.filter.fo_ctxt, NULL);
 956         pnum = 0; /* pnum indexes buf 0..num_pages */
 957
 958         file = filter_obj_open(export, lsm->lsm_object_id, S_IFREG);
 959         if (IS_ERR(file))
 960                 GOTO(out, retval = PTR_ERR(file));
 961
 962         /* count doubles as retval */
 963         for (pg = 0; pg < oa_bufs; pg++) {
 964                 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
 965                        ") off count ("LPU64",%d)\n",
 966                        cmd, pnum, file->f_dentry->d_inode->i_ino,
 967                        pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
 968                        (int)pga[pnum].count);
 969                 if (cmd & OBD_BRW_WRITE) {
 970                         loff_t off;
 971                         char *buffer;
 972                         off = pga[pnum].off;
 973                         buffer = kmap(pga[pnum].pg);
 974                         retval = file->f_op->write(file, buffer,
 975                                                    pga[pnum].count,
 976                                                    &off);
 977                         kunmap(pga[pnum].pg);
 978                         CDEBUG(D_INODE, "retval %ld\n", retval);
 979                 } else {
 980                         loff_t off = pga[pnum].off;
 981                         char *buffer = kmap(pga[pnum].pg);
 982
 983                         if (off >= file->f_dentry->d_inode->i_size) {
 984                                 memset(buffer, 0, pga[pnum].count);
 985                                 retval = pga[pnum].count;
 986                         } else {
 987                                 retval = file->f_op->read(file, buffer,
 988                                                           pga[pnum].count, &off);
 989                         }
 990                         kunmap(pga[pnum].pg);
 991
 992                         if (retval != pga[pnum].count) {
 993                                 filp_close(file, 0);
 994                                 GOTO(out, retval = -EIO);
 995                         }
 996                         CDEBUG(D_INODE, "retval %ld\n", retval);
 997                 }
 998                 pnum++;
 999         }
1000         /* sizes and blocks are set by generic_file_write */
1001         /* ctimes/mtimes will follow with a setattr call */
1002         filp_close(file, 0);
1003
1004         /* XXX: do something with callback if it is set? */
1005
1006         EXIT;
1007 out:
1008         pop_ctxt(&saved);
1009         error = (retval >= 0) ? 0 : retval;
1010         return error;
1011 }
1012
1013 /*
1014  * Calculate the number of buffer credits needed to write multiple pages in
1015  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
1016  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
1017  *
1018  * See comment above ext3_writepage_trans_blocks for details.  We assume
1019  * no data journaling is being done, but it does allow for all of the pages
1020  * being non-contiguous.  If we are guaranteed contiguous pages we could
1021  * reduce the number of (d)indirect blocks a lot.
1022  *
1023  * With N blocks per page and P pages, for each inode we have at most:
1024  * N*P indirect
1025  * min(N*P, blocksize/4 + 1) dindirect blocks
1026  * 1 tindirect
1027  *
1028  * For the entire filesystem, we have at most:
1029  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
1030  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
1031  * 1 inode block
1032  * 1 superblock
1033  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
1034  */
1035 static int ext3_credits_needed(struct super_block *sb, int objcount,
1036                                struct obd_ioobj *obj)
1037 {
1038         struct obd_ioobj *o = obj;
1039         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
1040         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
1041         int nbitmaps = 0;
1042         int ngdblocks = 0;
1043         int needed = objcount + 1;
1044         int i;
1045
1046         for (i = 0; i < objcount; i++, o++) {
1047                 int nblocks = o->ioo_bufcnt * blockpp;
1048                 int ndindirect = min(nblocks, addrpp + 1);
1049                 int nindir = nblocks + ndindirect + 1;
1050
1051                 nbitmaps += nindir + nblocks;
1052                 ngdblocks += nindir + nblocks;
1053
1054                 needed += nindir;
1055         }
1056
1057         /* Assumes ext3 and extN have same sb_info layout at the start. */
1058         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
1059                 nbitmaps = EXT3_SB(sb)->s_groups_count;
1060         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
1061                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
1062
1063         needed += nbitmaps + ngdblocks;
1064
1065 #ifdef CONFIG_QUOTA
1066         /* We assume that there will be 1 bit set in s_dquot.flags for each
1067          * quota file that is active.  This is at least true for now.
1068          */
1069         needed += hweight32(sb_any_quota_enabled(sb)) *
1070                 EXT3_SINGLEDATA_TRANS_BLOCKS;
1071 #endif
1072
1073         return needed;
1074 }
1075
1076 /* We have to start a huge journal transaction here to hold all of the
1077  * metadata for the pages being written here.  This is necessitated by
1078  * the fact that we do lots of prepare_write operations before we do
1079  * any of the matching commit_write operations, so even if we split
1080  * up to use "smaller" transactions none of them could complete until
1081  * all of them were opened.  By having a single journal transaction,
1082  * we eliminate duplicate reservations for common blocks like the
1083  * superblock and group descriptors or bitmaps.
1084  *
1085  * We will start the transaction here, but each prepare_write will
1086  * add a refcount to the transaction, and each commit_write will
1087  * remove a refcount.  The transaction will be closed when all of
1088  * the pages have been written.
1089  */
1090 static void *ext3_filter_journal_start(struct filter_obd *filter,
1091                                        int objcount, struct obd_ioobj *obj,
1092                                        int niocount, struct niobuf_remote *nb)
1093 {
1094         journal_t *journal = NULL;
1095         handle_t *handle = NULL;
1096         int needed;
1097
1098         /* It appears that some kernels have different values for
1099          * EXT*_MAX_GROUP_LOADED (either 8 or 32), so we cannot
1100          * assume anything after s_inode_bitmap_number is the same.
1101          */
1102         if (!strcmp(filter->fo_fstype, "ext3"))
1103                 journal = EXT3_SB(filter->fo_sb)->s_journal;
1104 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1105         else if (!strcmp(filter->fo_fstype, "extN"))
1106                 journal = EXTN_SB(filter->fo_sb)->s_journal;
1107 #endif
1108         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
1109
1110         /* The number of blocks we could _possibly_ dirty can very large.
1111          * We reduce our request if it is absurd (and we couldn't get that
1112          * many credits for a single handle anyways).
1113          *
1114          * At some point we have to limit the size of I/Os sent at one time,
1115          * increase the size of the journal, or we have to calculate the
1116          * actual journal requirements more carefully by checking all of
1117          * the blocks instead of being maximally pessimistic.  It remains to
1118          * be seen if this is a real problem or not.
1119          */
1120         if (needed > journal->j_max_transaction_buffers) {
1121                 CERROR("want too many journal credits (%d) using %d instead\n",
1122                        needed, journal->j_max_transaction_buffers);
1123                 needed = journal->j_max_transaction_buffers;
1124         }
1125
1126         lock_kernel();
1127         handle = journal_start(journal, needed);
1128         unlock_kernel();
1129         if (IS_ERR(handle))
1130                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
1131                        PTR_ERR(handle));
1132
1133         return(handle);
1134 }
1135
1136 static void *filter_journal_start(void **journal_save,
1137                                   struct filter_obd *filter,
1138                                   int objcount, struct obd_ioobj *obj,
1139                                   int niocount, struct niobuf_remote *nb)
1140 {
1141         void *handle = NULL;
1142
1143         /* This may not be necessary - we probably never have a
1144          * transaction started when we enter here, so we can
1145          * remove the saving of the journal state entirely.
1146          * For now leave it in just to see if it ever happens.
1147          */
1148         *journal_save = current->journal_info;
1149         if (*journal_save) {
1150                 CERROR("Already have handle %p???\n", *journal_save);
1151                 LBUG();
1152                 current->journal_info = NULL;
1153         }
1154
1155         if (!strcmp(filter->fo_fstype, "ext3") ||
1156             !strcmp(filter->fo_fstype, "extN"))
1157                 handle = ext3_filter_journal_start(filter, objcount, obj,
1158                                                    niocount, nb);
1159         return handle;
1160 }
1161
1162 static int ext3_filter_journal_stop(void *handle)
1163 {
1164         int rc;
1165
1166         /* We got a refcount on the handle for each call to prepare_write,
1167          * so we can drop the "parent" handle here to avoid the need for
1168          * osc to call back into filterobd to close the handle.  The
1169          * remaining references will be dropped in commit_write.
1170          */
1171         lock_kernel();
1172         rc = journal_stop((handle_t *)handle);
1173         unlock_kernel();
1174
1175         return rc;
1176 }
1177
1178 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
1179                                void *handle)
1180 {
1181         int rc = 0;
1182
1183         if (!strcmp(filter->fo_fstype, "ext3") ||
1184             !strcmp(filter->fo_fstype, "extN"))
1185                 rc = ext3_filter_journal_stop(handle);
1186
1187         if (rc)
1188                 CERROR("error on journal stop: rc = %d\n", rc);
1189
1190         current->journal_info = journal_save;
1191
1192         return rc;
1193 }
1194
1195 static inline void lustre_put_page(struct page *page)
1196 {
1197         kunmap(page);
1198         page_cache_release(page);
1199 }
1200
1201
1202 #ifndef PageUptodate
1203 #define PageUptodate(page) Page_Uptodate(page)
1204 #endif
1205 static struct page *
1206 lustre_get_page_read(struct inode *inode,
1207                      struct niobuf_remote *rnb)
1208 {
1209         unsigned long index = rnb->offset >> PAGE_SHIFT;
1210         struct address_space *mapping = inode->i_mapping;
1211         struct page *page;
1212         int rc;
1213
1214         page = read_cache_page(mapping, index,
1215                                (filler_t*)mapping->a_ops->readpage, NULL);
1216         if (!IS_ERR(page)) {
1217                 wait_on_page(page);
1218                 kmap(page);
1219                 if (!PageUptodate(page)) {
1220                         CERROR("page index %lu not uptodate\n", index);
1221                         GOTO(err_page, rc = -EIO);
1222                 }
1223                 if (PageError(page)) {
1224                         CERROR("page index %lu has error\n", index);
1225                         GOTO(err_page, rc = -EIO);
1226                 }
1227         }
1228         return page;
1229
1230 err_page:
1231         lustre_put_page(page);
1232         return ERR_PTR(rc);
1233 }
1234
1235 static struct page *
1236 lustre_get_page_write(struct inode *inode, unsigned long index)
1237 {
1238         struct address_space *mapping = inode->i_mapping;
1239         struct page *page;
1240         int rc;
1241
1242         page = grab_cache_page(mapping, index); /* locked page */
1243
1244         if (!IS_ERR(page)) {
1245                 kmap(page);
1246                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1247                  * a no-op for most filesystems, because we write the whole
1248                  * page.  For partial-page I/O this will read in the page.
1249                  */
1250                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1251                 if (rc) {
1252                         CERROR("page index %lu, rc = %d\n", index, rc);
1253                         if (rc != -ENOSPC)
1254                                 LBUG();
1255                         GOTO(err_unlock, rc);
1256                 }
1257                 /* XXX not sure if we need this if we are overwriting page */
1258                 if (PageError(page)) {
1259                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1260                         LBUG();
1261                         GOTO(err_unlock, rc = -EIO);
1262                 }
1263         }
1264         return page;
1265
1266 err_unlock:
1267         unlock_page(page);
1268         lustre_put_page(page);
1269         return ERR_PTR(rc);
1270 }
1271
1272 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1273 {
1274         struct inode *inode = page->mapping->host;
1275         int err;
1276
1277         err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1278 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1279         if (!err && IS_SYNC(inode))
1280                 err = waitfor_one_page(page);
1281 #else
1282 #warning ADD 2.5 waiting code here?
1283 #endif
1284         //SetPageUptodate(page); // the client commit_write will do this
1285
1286         SetPageReferenced(page);
1287         unlock_page(page);
1288         lustre_put_page(page);
1289         return err;
1290 }
1291
1292 struct page *filter_get_page_write(struct inode *inode,
1293                                    struct niobuf_remote *rnb,
1294                                    struct niobuf_local *lnb, int *pglocked)
1295 {
1296         unsigned long index = rnb->offset >> PAGE_SHIFT;
1297         struct address_space *mapping = inode->i_mapping;
1298
1299         struct page *page;
1300         int rc;
1301
1302         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1303         if (*pglocked)
1304                 page = grab_cache_page_nowait(mapping, index); /* locked page */
1305         else
1306                 page = grab_cache_page(mapping, index); /* locked page */
1307
1308
1309         /* This page is currently locked, so get a temporary page instead. */
1310         /* XXX I believe this is a very dangerous thing to do - consider if
1311          *     we had multiple writers for the same file (definitely the case
1312          *     if we are using this codepath).  If writer A locks the page,
1313          *     writer B writes to a copy (as here), writer A drops the page
1314          *     lock, and writer C grabs the lock before B does, then B will
1315          *     later overwrite the data from C, even if C had LDLM locked
1316          *     and initiated the write after B did.
1317          */
1318         if (!page) {
1319                 unsigned long addr;
1320                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1321                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1322                 if (!addr) {
1323                         CERROR("no memory for a temp page\n");
1324                         LBUG();
1325                         GOTO(err, rc = -ENOMEM);
1326                 }
1327                 /* XXX debugging */
1328                 memset((void *)addr, 0xBA, PAGE_SIZE);
1329                 page = virt_to_page(addr);
1330                 kmap(page);
1331                 page->index = index;
1332                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1333         } else if (!IS_ERR(page)) {
1334                 (*pglocked)++;
1335                 kmap(page);
1336
1337                 rc = mapping->a_ops->prepare_write(NULL, page,
1338                                                    rnb->offset % PAGE_SIZE,
1339                                                    rnb->len);
1340                 if (rc) {
1341                         CERROR("page index %lu, rc = %d\n", index, rc);
1342                         if (rc != -ENOSPC)
1343                                 LBUG();
1344                         GOTO(err_unlock, rc);
1345                 }
1346                 /* XXX not sure if we need this if we are overwriting page */
1347                 if (PageError(page)) {
1348                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1349                         LBUG();
1350                         GOTO(err_unlock, rc = -EIO);
1351                 }
1352         }
1353         return page;
1354
1355 err_unlock:
1356         unlock_page(page);
1357         lustre_put_page(page);
1358 err:
1359         return ERR_PTR(rc);
1360 }
1361
1362 /*
1363  * We need to balance prepare_write() calls with commit_write() calls.
1364  * If the page has been prepared, but we have no data for it, we don't
1365  * want to overwrite valid data on disk, but we still need to zero out
1366  * data for space which was newly allocated.  Like part of what happens
1367  * in __block_prepare_write() for newly allocated blocks.
1368  *
1369  * XXX currently __block_prepare_write() creates buffers for all the
1370  *     pages, and the filesystems mark these buffers as BH_New if they
1371  *     were newly allocated from disk. We use the BH_New flag similarly.
1372  */
1373 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1374                                int err)
1375 {
1376 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1377         if (err) {
1378                 unsigned block_start, block_end;
1379                 struct buffer_head *bh, *head = page->buffers;
1380                 unsigned blocksize = head->b_size;
1381                 void *addr = page_address(page);
1382
1383                 /* debugging: just seeing if this ever happens */
1384                 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1385                        page->index, page->mapping->host->i_ino, err);
1386
1387                 /* Currently one buffer per page, but in the future... */
1388                 for (bh = head, block_start = 0; bh != head || !block_start;
1389                      block_start = block_end, bh = bh->b_this_page) {
1390                         block_end = block_start + blocksize;
1391                         if (buffer_new(bh))
1392                                 memset(addr + block_start, 0, blocksize);
1393                 }
1394         }
1395 #endif
1396         return lustre_commit_write(page, from, to);
1397 }
1398
1399 static int filter_preprw(int cmd, struct lustre_handle *conn,
1400                          int objcount, struct obd_ioobj *obj,
1401                          int niocount, struct niobuf_remote *nb,
1402                          struct niobuf_local *res, void **desc_private)
1403 {
1404         struct obd_run_ctxt saved;
1405         struct obd_device *obd;
1406         struct obd_ioobj *o = obj;
1407         struct niobuf_remote *rnb = nb;
1408         struct niobuf_local *lnb = res;
1409         void *journal_save = NULL;
1410         int pglocked = 0;
1411         int rc = 0;
1412         int i;
1413         ENTRY;
1414
1415         obd = class_conn2obd(conn);
1416         if (!obd) {
1417                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1418                 RETURN(-EINVAL);
1419         }
1420         memset(res, 0, sizeof(*res) * niocount);
1421
1422         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1423
1424         if (cmd & OBD_BRW_WRITE) {
1425                 *desc_private = filter_journal_start(&journal_save,
1426                                                      &obd->u.filter,
1427                                                      objcount, obj, niocount,
1428                                                      nb);
1429                 if (IS_ERR(*desc_private))
1430                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1431         }
1432
1433         for (i = 0; i < objcount; i++, o++) {
1434                 struct dentry *dentry;
1435                 struct inode *inode;
1436                 int j;
1437
1438                 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1439                                            o->ioo_id, S_IFREG);
1440                 if (IS_ERR(dentry))
1441                         GOTO(out_clean, rc = PTR_ERR(dentry));
1442                 inode = dentry->d_inode;
1443                 if (!inode) {
1444                         CERROR("trying to BRW to non-existent file "LPU64"\n",
1445                                o->ioo_id);
1446                         f_dput(dentry);
1447                         GOTO(out_clean, rc = -ENOENT);
1448                 }
1449
1450                 for (j = 0; j < o->ioo_bufcnt; j++, rnb++, lnb++) {
1451                         struct page *page;
1452
1453                         if (j == 0)
1454                                 lnb->dentry = dentry;
1455                         else
1456                                 lnb->dentry = dget(dentry);
1457
1458                         if (cmd & OBD_BRW_WRITE)
1459                                 page = filter_get_page_write(inode, rnb, lnb,
1460                                                              &pglocked);
1461                         else
1462                                 page = lustre_get_page_read(inode, rnb);
1463
1464                         if (IS_ERR(page)) {
1465                                 f_dput(dentry);
1466                                 GOTO(out_clean, rc = PTR_ERR(page));
1467                         }
1468
1469                         lnb->addr = page_address(page);
1470                         lnb->offset = rnb->offset;
1471                         lnb->page = page;
1472                         lnb->len = rnb->len;
1473                 }
1474         }
1475
1476 out_stop:
1477         if (cmd & OBD_BRW_WRITE) {
1478                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1479                                               *desc_private);
1480                 if (!rc)
1481                         rc = err;
1482         }
1483 out_ctxt:
1484         pop_ctxt(&saved);
1485         RETURN(rc);
1486 out_clean:
1487         while (lnb-- > res) {
1488                 CERROR("error cleanup on brw\n");
1489                 f_dput(lnb->dentry);
1490                 if (cmd & OBD_BRW_WRITE)
1491                         filter_commit_write(lnb->page, 0, PAGE_SIZE, rc);
1492                 else
1493                         lustre_put_page(lnb->page);
1494         }
1495         goto out_stop;
1496 }
1497
1498 static int filter_write_locked_page(struct niobuf_local *lnb)
1499 {
1500         struct page *lpage;
1501         int rc;
1502
1503         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1504         if (IS_ERR(lpage)) {
1505                 /* It is highly unlikely that we would ever get an error here.
1506                  * The page we want to get was previously locked, so it had to
1507                  * have already allocated the space, and we were just writing
1508                  * over the same data, so there would be no hole in the file.
1509                  *
1510                  * XXX: possibility of a race with truncate could exist, need
1511                  *      to check that.  There are no guarantees w.r.t.
1512                  *      write order even on a local filesystem, although the
1513                  *      normal response would be to return the number of bytes
1514                  *      successfully written and leave the rest to the app.
1515                  */
1516                 rc = PTR_ERR(lpage);
1517                 CERROR("error getting locked page index %ld: rc = %d\n",
1518                        lnb->page->index, rc);
1519                 GOTO(out, rc);
1520         }
1521
1522         /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1523          * lustre_commit_write() below, lnb->page was kmapped previously in
1524          * filter_get_page_write() and kunmapped in lustre_put_page() below.
1525          */
1526         memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1527         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1528         if (rc)
1529                 CERROR("error committing locked page %ld: rc = %d\n",
1530                        lnb->page->index, rc);
1531 out:
1532         lustre_put_page(lnb->page);
1533
1534         return rc;
1535 }
1536
1537 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1538                            int objcount, struct obd_ioobj *obj,
1539                            int niocount, struct niobuf_local *res,
1540                            void *private)
1541 {
1542         struct obd_run_ctxt saved;
1543         struct obd_ioobj *o;
1544         struct niobuf_local *r;
1545         struct obd_device *obd = class_conn2obd(conn);
1546         void *journal_save;
1547         int found_locked = 0;
1548         int rc = 0;
1549         int i;
1550         ENTRY;
1551
1552         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1553         lock_kernel();
1554         journal_save = current->journal_info;
1555         LASSERT(!journal_save);
1556
1557         current->journal_info = private;
1558         unlock_kernel();
1559         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1560                 int j;
1561                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1562                         struct page *page = r->page;
1563
1564                         if (!page)
1565                                 LBUG();
1566
1567                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1568                                 found_locked++;
1569                                 continue;
1570                         }
1571
1572                         if (cmd & OBD_BRW_WRITE) {
1573                                 int err = filter_commit_write(page, 0,
1574                                                               r->len, 0);
1575
1576                                 if (!rc)
1577                                         rc = err;
1578                         } else
1579                                 lustre_put_page(page);
1580
1581                         f_dput(r->dentry);
1582                 }
1583         }
1584         lock_kernel();
1585         current->journal_info = journal_save;
1586         unlock_kernel();
1587
1588         if (!found_locked)
1589                 goto out_ctxt;
1590
1591         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1592                 int j;
1593                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1594                         int err;
1595                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1596                                 continue;
1597
1598                         err = filter_write_locked_page(r);
1599                         if (!rc)
1600                                 rc = err;
1601                         f_dput(r->dentry);
1602                 }
1603         }
1604
1605 out_ctxt:
1606         pop_ctxt(&saved);
1607         RETURN(rc);
1608 }
1609
1610 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1611 {
1612         struct obd_device *obd = class_conn2obd(conn);
1613         struct statfs sfs;
1614         int rc;
1615
1616         ENTRY;
1617         rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1618         if (!rc)
1619                 statfs_pack(osfs, &sfs);
1620
1621         return rc;
1622 }
1623
1624 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1625                            void *key, obd_count *vallen, void **val)
1626 {
1627         struct obd_device *obd;
1628         ENTRY;
1629
1630         obd = class_conn2obd(conn);
1631         if (!obd) {
1632                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1633                 RETURN(-EINVAL);
1634         }
1635
1636         if ( keylen == strlen("blocksize") &&
1637              memcmp(key, "blocksize", keylen) == 0 ) {
1638                 *vallen = sizeof(long);
1639                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1640                 RETURN(0);
1641         }
1642
1643         if ( keylen == strlen("blocksize_bits") &&
1644              memcmp(key, "blocksize_bits", keylen) == 0 ){
1645                 *vallen = sizeof(long);
1646                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1647                 RETURN(0);
1648         }
1649
1650         if ( keylen == strlen("root_ino") &&
1651              memcmp(key, "root_ino", keylen) == 0 ){
1652                 *vallen = sizeof(obd_id);
1653                 *val = (void *)(obd_id)FILTER_ROOTINO;
1654                 RETURN(0);
1655         }
1656
1657         CDEBUG(D_IOCTL, "invalid key\n");
1658         RETURN(-EINVAL);
1659 }
1660
1661 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1662                   struct lustre_handle *src_conn, struct obdo *src,
1663                   obd_size count, obd_off offset)
1664 {
1665         struct page *page;
1666         struct lov_stripe_md srcmd, dstmd;
1667         unsigned long index = 0;
1668         int err = 0;
1669
1670         memset(&srcmd, 0, sizeof(srcmd));
1671         memset(&dstmd, 0, sizeof(dstmd));
1672         srcmd.lsm_object_id = src->o_id;
1673         dstmd.lsm_object_id = dst->o_id;
1674
1675         ENTRY;
1676         CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1677                ", dst: ino "LPU64"\n",
1678                src->o_id, src->o_blocks, src->o_size, dst->o_id);
1679         page = alloc_page(GFP_USER);
1680         if (page == NULL)
1681                 RETURN(-ENOMEM);
1682
1683 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1684         while (TryLockPage(page))
1685                 ___wait_on_page(page);
1686 #else
1687         wait_on_page_locked(page);
1688 #endif
1689
1690         /* XXX with brw vector I/O, we could batch up reads and writes here,
1691          *     all we need to do is allocate multiple pages to handle the I/Os
1692          *     and arrays to handle the request parameters.
1693          */
1694         while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1695                 struct brw_page pg;
1696                 struct io_cb_data *cbd = ll_init_cb();
1697
1698                 if (!cbd) {
1699                         err = -ENOMEM;
1700                         EXIT;
1701                         break;
1702                 }
1703
1704                 pg.pg = page;
1705                 pg.count = PAGE_SIZE;
1706                 pg.off = (page->index) << PAGE_SHIFT;
1707                 pg.flag = 0;
1708
1709                 page->index = index;
1710                 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1711                               ll_sync_io_cb, cbd);
1712
1713                 if ( err ) {
1714                         EXIT;
1715                         break;
1716                 }
1717
1718                 cbd = ll_init_cb();
1719                 if (!cbd) {
1720                         err = -ENOMEM;
1721                         EXIT;
1722                         break;
1723                 }
1724                 pg.flag = OBD_BRW_CREATE;
1725                 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1726
1727                 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1728                               ll_sync_io_cb, cbd);
1729
1730                 /* XXX should handle dst->o_size, dst->o_blocks here */
1731                 if ( err ) {
1732                         EXIT;
1733                         break;
1734                 }
1735
1736                 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1737
1738                 index++;
1739         }
1740         dst->o_size = src->o_size;
1741         dst->o_blocks = src->o_blocks;
1742         dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1743         unlock_page(page);
1744         __free_page(page);
1745
1746         RETURN(err);
1747 }
1748
1749
1750 static struct obd_ops filter_obd_ops = {
1751         o_get_info:    filter_get_info,
1752         o_setup:       filter_setup,
1753         o_cleanup:     filter_cleanup,
1754         o_connect:     filter_connect,
1755         o_disconnect:  filter_disconnect,
1756         o_statfs:      filter_statfs,
1757         o_getattr:     filter_getattr,
1758         o_create:      filter_create,
1759         o_setattr:     filter_setattr,
1760         o_destroy:     filter_destroy,
1761         o_open:        filter_open,
1762         o_close:       filter_close,
1763         o_brw:         filter_pgcache_brw,
1764         o_punch:       filter_truncate,
1765         o_preprw:      filter_preprw,
1766         o_commitrw:    filter_commitrw
1767 #if 0
1768         o_preallocate: filter_preallocate_inodes,
1769         o_migrate:     filter_migrate,
1770         o_copy:        filter_copy_data,
1771         o_iterate:     filter_iterate
1772 #endif
1773 };
1774
1775
1776 static int __init obdfilter_init(void)
1777 {
1778         printk(KERN_INFO "Filtering OBD driver  v0.001, info@clusterfs.com\n");
1779         filter_open_cache = kmem_cache_create("ll_filter_fdata",
1780                                               sizeof(struct filter_file_data),
1781                                               0, 0, NULL, NULL);
1782         if (!filter_open_cache)
1783                 RETURN(-ENOMEM);
1784
1785         filter_dentry_cache = kmem_cache_create("ll_filter_dentry",
1786                                         sizeof(struct filter_dentry_data),
1787                                         0, 0, NULL, NULL);
1788         if (!filter_dentry_cache) {
1789                 kmem_cache_destroy(filter_open_cache);
1790                 RETURN(-ENOMEM);
1791         }
1792
1793         return class_register_type(&filter_obd_ops, OBD_FILTER_DEVICENAME);
1794 }
1795
1796 static void __exit obdfilter_exit(void)
1797 {
1798         class_unregister_type(OBD_FILTER_DEVICENAME);
1799         if (kmem_cache_destroy(filter_dentry_cache))
1800                 CERROR("couldn't free obdfilter dentry cache\n");
1801         if (kmem_cache_destroy(filter_open_cache))
1802                 CERROR("couldn't free obdfilter open cache\n");
1803 }
1804
1805 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1806 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1807 MODULE_LICENSE("GPL");
1808
1809 module_init(obdfilter_init);
1810 module_exit(obdfilter_exit);