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 #include <linux/lprocfs_status.h>
  34
  35 extern struct lprocfs_vars status_class_var[];
  36 extern struct lprocfs_vars status_var_nm_1[];
  37
  38 static kmem_cache_t *filter_open_cache;
  39 static kmem_cache_t *filter_dentry_cache;
  40
  41 #define FILTER_ROOTINO 2
  42 #define FILTER_ROOTINO_STR __stringify(FILTER_ROOTINO)
  43
  44 #define S_SHIFT 12
  45 static char *obd_type_by_mode[S_IFMT >> S_SHIFT] = {
  46         [0]                     NULL,
  47         [S_IFREG >> S_SHIFT]    "R",
  48         [S_IFDIR >> S_SHIFT]    "D",
  49         [S_IFCHR >> S_SHIFT]    "C",
  50         [S_IFBLK >> S_SHIFT]    "B",
  51         [S_IFIFO >> S_SHIFT]    "F",
  52         [S_IFSOCK >> S_SHIFT]   "S",
  53         [S_IFLNK >> S_SHIFT]    "L"
  54 };
  55
  56 static inline const char *obd_mode_to_type(int mode)
  57 {
  58         return obd_type_by_mode[(mode & S_IFMT) >> S_SHIFT];
  59 }
  60
  61 /* write the pathname into the string */
  62 static int filter_id(char *buf, obd_id id, obd_mode mode)
  63 {
  64         return sprintf(buf, "O/%s/"LPU64, obd_mode_to_type(mode), id);
  65 }
  66
  67 static inline void f_dput(struct dentry *dentry)
  68 {
  69         /* Can't go inside filter_ddelete because it can block */
  70         CDEBUG(D_INODE, "putting %s: %p, count = %d\n",
  71                dentry->d_name.name, dentry, atomic_read(&dentry->d_count) - 1);
  72         LASSERT(atomic_read(&dentry->d_count) > 0);
  73
  74         dput(dentry);
  75 }
  76
  77 /* Not racy w.r.t. others, because we are the only user of this dentry */
  78 static void filter_drelease(struct dentry *dentry)
  79 {
  80         if (dentry->d_fsdata)
  81                 kmem_cache_free(filter_dentry_cache, dentry->d_fsdata);
  82 }
  83
  84 struct dentry_operations filter_dops = {
  85         .d_release = filter_drelease,
  86 };
  87
  88 /* setup the object store with correct subdirectories */
  89 static int filter_prep(struct obd_device *obd)
  90 {
  91         struct obd_run_ctxt saved;
  92         struct filter_obd *filter = &obd->u.filter;
  93         struct dentry *dentry;
  94         struct dentry *root;
  95         struct file *file;
  96         struct inode *inode;
  97         int rc = 0;
  98         __u64 lastobjid = 2;
  99         int mode = 0;
 100
 101         push_ctxt(&saved, &filter->fo_ctxt, NULL);
 102         dentry = simple_mkdir(current->fs->pwd, "O", 0700);
 103         CDEBUG(D_INODE, "got/created O: %p\n", dentry);
 104         if (IS_ERR(dentry)) {
 105                 rc = PTR_ERR(dentry);
 106                 CERROR("cannot open/create O: rc = %d\n", rc);
 107                 GOTO(out, rc);
 108         }
 109         filter->fo_dentry_O = dentry;
 110         dentry = simple_mkdir(current->fs->pwd, "P", 0700);
 111         CDEBUG(D_INODE, "got/created P: %p\n", dentry);
 112         if (IS_ERR(dentry)) {
 113                 rc = PTR_ERR(dentry);
 114                 CERROR("cannot open/create P: rc = %d\n", rc);
 115                 GOTO(out_O, rc);
 116         }
 117         f_dput(dentry);
 118         dentry = simple_mkdir(current->fs->pwd, "D", 0700);
 119         CDEBUG(D_INODE, "got/created D: %p\n", dentry);
 120         if (IS_ERR(dentry)) {
 121                 rc = PTR_ERR(dentry);
 122                 CERROR("cannot open/create D: rc = %d\n", rc);
 123                 GOTO(out_O, rc);
 124         }
 125
 126         root = simple_mknod(dentry, FILTER_ROOTINO_STR, S_IFREG | 0755);
 127         f_dput(dentry);
 128         if (IS_ERR(root)) {
 129                 rc = PTR_ERR(root);
 130                 CERROR("OBD filter: cannot open/create root %d: rc = %d\n",
 131                        FILTER_ROOTINO, rc);
 132                 GOTO(out_O, rc);
 133         }
 134         f_dput(root);
 135
 136         /*
 137          * Create directories and/or get dentries for each object type.
 138          * This saves us from having to do multiple lookups for each one.
 139          */
 140         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 141                 char *type = obd_type_by_mode[mode];
 142
 143                 if (!type) {
 144                         filter->fo_dentry_O_mode[mode] = NULL;
 145                         continue;
 146                 }
 147                 dentry = simple_mkdir(filter->fo_dentry_O, type, 0700);
 148                 CDEBUG(D_INODE, "got/created O/%s: %p\n", type, dentry);
 149                 if (IS_ERR(dentry)) {
 150                         rc = PTR_ERR(dentry);
 151                         CERROR("cannot create O/%s: rc = %d\n", type, rc);
 152                         GOTO(out_O_mode, rc);
 153                 }
 154                 filter->fo_dentry_O_mode[mode] = dentry;
 155         }
 156
 157         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 158         if ( !file || IS_ERR(file) ) {
 159                 rc = PTR_ERR(file);
 160                 CERROR("OBD filter: cannot open/create status %s: rc = %d\n",
 161                        "D/status", rc);
 162                 GOTO(out_O_mode, rc);
 163         }
 164
 165         /* steal operations */
 166         inode = file->f_dentry->d_inode;
 167         filter->fo_fop = file->f_op;
 168         filter->fo_iop = inode->i_op;
 169         filter->fo_aops = inode->i_mapping->a_ops;
 170
 171         if (inode->i_size == 0) {
 172                 __u64 disk_lastobjid = cpu_to_le64(lastobjid);
 173                 ssize_t retval = file->f_op->write(file,(char *)&disk_lastobjid,
 174                                                    sizeof(disk_lastobjid),
 175                                                    &file->f_pos);
 176                 if (retval != sizeof(disk_lastobjid)) {
 177                         CDEBUG(D_INODE,"OBD filter: error writing lastobjid\n");
 178                         filp_close(file, 0);
 179                         GOTO(out_O_mode, rc = -EIO);
 180                 }
 181         } else {
 182                 __u64 disk_lastobjid;
 183                 ssize_t retval = file->f_op->read(file, (char *)&disk_lastobjid,
 184                                                   sizeof(disk_lastobjid),
 185                                                   &file->f_pos);
 186                 if (retval != sizeof(disk_lastobjid)) {
 187                         CDEBUG(D_INODE,"OBD filter: error reading lastobjid\n");
 188                         filp_close(file, 0);
 189                         GOTO(out_O_mode, rc = -EIO);
 190                 }
 191                 lastobjid = le64_to_cpu(disk_lastobjid);
 192         }
 193         filter->fo_lastobjid = lastobjid;
 194         filp_close(file, 0);
 195
 196         rc = 0;
 197  out:
 198         pop_ctxt(&saved);
 199
 200         return(rc);
 201
 202 out_O_mode:
 203         while (mode-- > 0) {
 204                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 205                 if (dentry) {
 206                         f_dput(dentry);
 207                         filter->fo_dentry_O_mode[mode] = NULL;
 208                 }
 209         }
 210 out_O:
 211         f_dput(filter->fo_dentry_O);
 212         filter->fo_dentry_O = NULL;
 213         goto out;
 214 }
 215
 216 /* cleanup the filter: write last used object id to status file */
 217 static void filter_post(struct obd_device *obd)
 218 {
 219         struct obd_run_ctxt saved;
 220         struct filter_obd *filter = &obd->u.filter;
 221         __u64 disk_lastobjid;
 222         long rc;
 223         struct file *file;
 224         int mode;
 225
 226         push_ctxt(&saved, &filter->fo_ctxt, NULL);
 227         file = filp_open("D/status", O_RDWR | O_CREAT, 0700);
 228         if (IS_ERR(file)) {
 229                 CERROR("OBD filter: cannot create status file\n");
 230                 goto out;
 231         }
 232
 233         file->f_pos = 0;
 234         disk_lastobjid = cpu_to_le64(filter->fo_lastobjid);
 235         rc = file->f_op->write(file, (char *)&disk_lastobjid,
 236                        sizeof(disk_lastobjid), &file->f_pos);
 237         if (rc != sizeof(disk_lastobjid))
 238                 CERROR("OBD filter: error writing lastobjid: rc = %ld\n", rc);
 239
 240         rc = filp_close(file, NULL);
 241         if (rc)
 242                 CERROR("OBD filter: cannot close status file: rc = %ld\n", rc);
 243
 244         for (mode = 0; mode < (S_IFMT >> S_SHIFT); mode++) {
 245                 struct dentry *dentry = filter->fo_dentry_O_mode[mode];
 246                 if (dentry) {
 247                         f_dput(dentry);
 248                         filter->fo_dentry_O_mode[mode] = NULL;
 249                 }
 250         }
 251         f_dput(filter->fo_dentry_O);
 252 out:
 253         pop_ctxt(&saved);
 254 }
 255
 256
 257 static __u64 filter_next_id(struct obd_device *obd)
 258 {
 259         obd_id id;
 260
 261         spin_lock(&obd->u.filter.fo_objidlock);
 262         id = ++obd->u.filter.fo_lastobjid;
 263         spin_unlock(&obd->u.filter.fo_objidlock);
 264
 265         /* FIXME: write the lastobjid to disk here */
 266         return id;
 267 }
 268
 269 /* how to get files, dentries, inodes from object id's */
 270 /* parent i_sem is already held if needed for exclusivity */
 271 static struct dentry *filter_fid2dentry(struct obd_device *obd,
 272                                         struct dentry *dparent,
 273                                         __u64 id, __u32 type)
 274 {
 275         struct super_block *sb = obd->u.filter.fo_sb;
 276         struct dentry *dchild;
 277         char name[32];
 278         int len;
 279         ENTRY;
 280
 281         if (!sb || !sb->s_dev) {
 282                 CERROR("fatal: device not initialized.\n");
 283                 RETURN(ERR_PTR(-ENXIO));
 284         }
 285
 286         if (id == 0) {
 287                 CERROR("fatal: invalid object #0\n");
 288                 LBUG();
 289                 RETURN(ERR_PTR(-ESTALE));
 290         }
 291
 292         if (!(type & S_IFMT)) {
 293                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 294                        __FUNCTION__, id, type);
 295                 RETURN(ERR_PTR(-EINVAL));
 296         }
 297
 298         len = sprintf(name, LPU64, id);
 299         CDEBUG(D_INODE, "opening object O/%s/%s\n", obd_mode_to_type(type),
 300                name);
 301         dchild = lookup_one_len(name, dparent, len);
 302         if (IS_ERR(dchild)) {
 303                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 304                 RETURN(dchild);
 305         }
 306
 307         CDEBUG(D_INODE, "got child obj O/%s/%s: %p, count = %d\n",
 308                obd_mode_to_type(type), name, dchild,
 309                atomic_read(&dchild->d_count));
 310
 311         LASSERT(atomic_read(&dchild->d_count) > 0);
 312
 313         RETURN(dchild);
 314 }
 315
 316 static inline struct dentry *filter_parent(struct obd_device *obd,
 317                                            obd_mode mode)
 318 {
 319         struct filter_obd *filter = &obd->u.filter;
 320
 321         return filter->fo_dentry_O_mode[(mode & S_IFMT) >> S_SHIFT];
 322 }
 323
 324 static struct file *filter_obj_open(struct obd_export *export,
 325                                     __u64 id, __u32 type)
 326 {
 327         struct filter_obd *filter = &export->exp_obd->u.filter;
 328         struct super_block *sb = filter->fo_sb;
 329         struct dentry *dentry;
 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         dentry = file->f_dentry;
 376         spin_lock(&filter->fo_fddlock);
 377         if (dentry->d_fsdata) {
 378                 spin_unlock(&filter->fo_fddlock);
 379                 kmem_cache_free(filter_dentry_cache, fdd);
 380                 fdd = dentry->d_fsdata;
 381                 LASSERT(kmem_cache_validate(filter_dentry_cache, fdd));
 382                 /* should only happen during client recovery */
 383                 if (fdd->fdd_flags & FILTER_FLAG_DESTROY)
 384                         CDEBUG(D_INODE,"opening destroyed object "LPX64"\n",id);
 385                 atomic_inc(&fdd->fdd_open_count);
 386         } else {
 387                 atomic_set(&fdd->fdd_open_count, 1);
 388                 fdd->fdd_flags = 0;
 389                 /* If this is racy, then we can use {cmp}xchg and atomic_add */
 390                 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         if (!dentry->d_op)
 399                 dentry->d_op = &filter_dops;
 400         else
 401                 LASSERT(dentry->d_op == &filter_dops);
 402
 403         spin_lock(&fed->fed_lock);
 404         list_add(&ffd->ffd_export_list, &fed->fed_open_head);
 405         spin_unlock(&fed->fed_lock);
 406
 407         CDEBUG(D_INODE, "opening objid "LPX64": rc = %p\n", id, file);
 408
 409 out:
 410         RETURN(file);
 411
 412 out_fdd:
 413         kmem_cache_free(filter_dentry_cache, fdd);
 414 out_ffd:
 415         ffd->ffd_servercookie = DEAD_HANDLE_MAGIC;
 416         kmem_cache_free(filter_open_cache, ffd);
 417         goto out;
 418 }
 419
 420 /* Caller must hold i_sem on dir_dentry->d_inode */
 421 static int filter_destroy_internal(struct obd_device *obd,
 422                                    struct dentry *dir_dentry,
 423                                    struct dentry *object_dentry)
 424 {
 425         struct obd_run_ctxt saved;
 426         struct inode *inode = object_dentry->d_inode;
 427         int rc;
 428         ENTRY;
 429
 430         if (inode->i_nlink != 1 || atomic_read(&inode->i_count) != 1) {
 431                 CERROR("destroying objid %*s nlink = %d, count = %d\n",
 432                        object_dentry->d_name.len,
 433                        object_dentry->d_name.name,
 434                        inode->i_nlink, atomic_read(&inode->i_count));
 435         }
 436
 437         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 438         rc = vfs_unlink(dir_dentry->d_inode, object_dentry);
 439         /* XXX unlink from PENDING directory now too */
 440         pop_ctxt(&saved);
 441
 442         if (rc)
 443                 CERROR("error unlinking objid %*s: rc %d\n",
 444                        object_dentry->d_name.len,
 445                        object_dentry->d_name.name, rc);
 446
 447         RETURN(rc);
 448 }
 449
 450 static int filter_close_internal(struct obd_device *obd,
 451                                  struct filter_file_data *ffd)
 452 {
 453         struct file *filp = ffd->ffd_file;
 454         struct dentry *object_dentry = dget(filp->f_dentry);
 455         struct filter_dentry_data *fdd = object_dentry->d_fsdata;
 456         int rc, rc2 = 0;
 457         ENTRY;
 458
 459         LASSERT(filp->private_data == ffd);
 460         LASSERT(fdd);
 461
 462         rc = filp_close(filp, 0);
 463
 464         if (atomic_dec_and_test(&fdd->fdd_open_count) &&
 465             fdd->fdd_flags & FILTER_FLAG_DESTROY) {
 466                 struct dentry *dir_dentry = filter_parent(obd, S_IFREG);
 467
 468                 down(&dir_dentry->d_inode->i_sem);
 469                 rc2 = filter_destroy_internal(obd, dir_dentry, object_dentry);
 470                 if (rc2 && !rc)
 471                         rc = rc2;
 472                 up(&dir_dentry->d_inode->i_sem);
 473         }
 474
 475         f_dput(object_dentry);
 476         kmem_cache_free(filter_open_cache, ffd);
 477
 478         RETURN(rc);
 479 }
 480
 481 /* obd methods */
 482 static int filter_connect(struct lustre_handle *conn, struct obd_device *obd,
 483                           obd_uuid_t cluuid, struct recovd_obd *recovd,
 484                           ptlrpc_recovery_cb_t recover)
 485 {
 486         struct obd_export *exp;
 487         int rc;
 488
 489         ENTRY;
 490         MOD_INC_USE_COUNT;
 491         rc = class_connect(conn, obd, cluuid);
 492         if (rc)
 493                 GOTO(out_dec, rc);
 494         exp = class_conn2export(conn);
 495         LASSERT(exp);
 496
 497         INIT_LIST_HEAD(&exp->exp_filter_data.fed_open_head);
 498         spin_lock_init(&exp->exp_filter_data.fed_lock);
 499 out:
 500         RETURN(rc);
 501
 502 out_dec:
 503         MOD_DEC_USE_COUNT;
 504         goto out;
 505 }
 506
 507 static int filter_disconnect(struct lustre_handle *conn)
 508 {
 509         struct obd_export *exp = class_conn2export(conn);
 510         struct filter_export_data *fed;
 511         int rc;
 512         ENTRY;
 513
 514         LASSERT(exp);
 515         fed = &exp->exp_filter_data;
 516         spin_lock(&fed->fed_lock);
 517         while (!list_empty(&fed->fed_open_head)) {
 518                 struct filter_file_data *ffd;
 519
 520                 ffd = list_entry(fed->fed_open_head.next, typeof(*ffd),
 521                                  ffd_export_list);
 522                 list_del(&ffd->ffd_export_list);
 523                 spin_unlock(&fed->fed_lock);
 524
 525                 CERROR("force closing file %*s on disconnect\n",
 526                        ffd->ffd_file->f_dentry->d_name.len,
 527                        ffd->ffd_file->f_dentry->d_name.name);
 528
 529                 filter_close_internal(exp->exp_obd, ffd);
 530                 spin_lock(&fed->fed_lock);
 531         }
 532         spin_unlock(&fed->fed_lock);
 533
 534         ldlm_cancel_locks_for_export(exp);
 535         rc = class_disconnect(conn);
 536         if (!rc)
 537                 MOD_DEC_USE_COUNT;
 538
 539         /* XXX cleanup preallocated inodes */
 540         RETURN(rc);
 541 }
 542
 543 /* mount the file system (secretly) */
 544 static int filter_setup(struct obd_device *obd, obd_count len, void *buf)
 545 {
 546         struct obd_ioctl_data* data = buf;
 547         struct filter_obd *filter;
 548         struct vfsmount *mnt;
 549         int err = 0;
 550         ENTRY;
 551
 552         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 553                 RETURN(-EINVAL);
 554
 555         MOD_INC_USE_COUNT;
 556         mnt = do_kern_mount(data->ioc_inlbuf2, 0, data->ioc_inlbuf1, NULL);
 557         err = PTR_ERR(mnt);
 558         if (IS_ERR(mnt))
 559                 GOTO(err_dec, err);
 560
 561         filter = &obd->u.filter;;
 562         filter->fo_vfsmnt = mnt;
 563         filter->fo_fstype = strdup(data->ioc_inlbuf2);
 564         filter->fo_sb = mnt->mnt_root->d_inode->i_sb;
 565         CERROR("%s: mnt is %p\n", data->ioc_inlbuf1, filter->fo_vfsmnt);
 566         /* XXX is this even possible if do_kern_mount succeeded? */
 567         if (!filter->fo_sb)
 568                 GOTO(err_kfree, err = -ENODEV);
 569
 570         OBD_SET_CTXT_MAGIC(&filter->fo_ctxt);
 571         filter->fo_ctxt.pwdmnt = mnt;
 572         filter->fo_ctxt.pwd = mnt->mnt_root;
 573         filter->fo_ctxt.fs = get_ds();
 574
 575         err = filter_prep(obd);
 576         if (err)
 577                 GOTO(err_kfree, err);
 578         spin_lock_init(&filter->fo_fddlock);
 579         spin_lock_init(&filter->fo_objidlock);
 580         INIT_LIST_HEAD(&filter->fo_export_list);
 581
 582         obd->obd_namespace =
 583                 ldlm_namespace_new("filter-tgt", LDLM_NAMESPACE_SERVER);
 584         if (obd->obd_namespace == NULL)
 585                 LBUG();
 586
 587         ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
 588                            "filter_ldlm_cb_client", &obd->obd_ldlm_client);
 589
 590         RETURN(0);
 591
 592 err_kfree:
 593         kfree(filter->fo_fstype);
 594         unlock_kernel();
 595         mntput(filter->fo_vfsmnt);
 596         filter->fo_sb = 0;
 597         lock_kernel();
 598
 599 err_dec:
 600         MOD_DEC_USE_COUNT;
 601         return err;
 602 }
 603
 604
 605 static int filter_cleanup(struct obd_device *obd)
 606 {
 607         struct super_block *sb;
 608         ENTRY;
 609
 610         if (!list_empty(&obd->obd_exports)) {
 611                 CERROR("still has clients!\n");
 612                 class_disconnect_all(obd);
 613                 if (!list_empty(&obd->obd_exports)) {
 614                         CERROR("still has exports after forced cleanup?\n");
 615                         RETURN(-EBUSY);
 616                 }
 617         }
 618
 619         ldlm_namespace_free(obd->obd_namespace);
 620
 621         sb = obd->u.filter.fo_sb;
 622         if (!obd->u.filter.fo_sb)
 623                 RETURN(0);
 624
 625         filter_post(obd);
 626
 627         shrink_dcache_parent(sb->s_root);
 628         unlock_kernel();
 629         mntput(obd->u.filter.fo_vfsmnt);
 630         obd->u.filter.fo_sb = 0;
 631         kfree(obd->u.filter.fo_fstype);
 632
 633         lock_kernel();
 634
 635         MOD_DEC_USE_COUNT;
 636         RETURN(0);
 637 }
 638
 639
 640 static void filter_from_inode(struct obdo *oa, struct inode *inode, int valid)
 641 {
 642         int type = oa->o_mode & S_IFMT;
 643         ENTRY;
 644
 645         CDEBUG(D_INFO, "src inode %ld (%p), dst obdo %ld valid 0x%08x\n",
 646                inode->i_ino, inode, (long)oa->o_id, valid);
 647         /* Don't copy the inode number in place of the object ID */
 648         obdo_from_inode(oa, inode, valid);
 649         oa->o_mode &= ~S_IFMT;
 650         oa->o_mode |= type;
 651
 652         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
 653                 obd_rdev rdev = kdev_t_to_nr(inode->i_rdev);
 654                 oa->o_rdev = rdev;
 655                 oa->o_valid |= OBD_MD_FLRDEV;
 656         }
 657
 658         EXIT;
 659 }
 660
 661 static struct filter_file_data *filter_handle2ffd(struct lustre_handle *handle)
 662 {
 663         struct filter_file_data *ffd = NULL;
 664         ENTRY;
 665
 666         if (!handle || !handle->addr)
 667                 RETURN(NULL);
 668
 669         ffd = (struct filter_file_data *)(unsigned long)(handle->addr);
 670         if (!kmem_cache_validate(filter_open_cache, (void *)ffd))
 671                 RETURN(NULL);
 672
 673         if (ffd->ffd_servercookie != handle->cookie)
 674                 RETURN(NULL);
 675
 676         LASSERT(ffd->ffd_file->private_data == ffd);
 677         RETURN(ffd);
 678 }
 679
 680 static struct dentry *__filter_oa2dentry(struct lustre_handle *conn,
 681                                          struct obdo *oa, char *what)
 682 {
 683         struct dentry *dentry = NULL;
 684
 685         if (oa->o_valid & OBD_MD_FLHANDLE) {
 686                 struct lustre_handle *ost_handle = obdo_handle(oa);
 687                 struct filter_file_data *ffd = filter_handle2ffd(ost_handle);
 688
 689                 if (ffd)
 690                         dentry = dget(ffd->ffd_file->f_dentry);
 691         }
 692
 693         if (!dentry) {
 694                 struct obd_device *obd = class_conn2obd(conn);
 695                 if (!obd) {
 696                         CERROR("invalid client "LPX64"\n", conn->addr);
 697                         RETURN(ERR_PTR(-EINVAL));
 698                 }
 699                 dentry = filter_fid2dentry(obd, filter_parent(obd, oa->o_mode),
 700                                            oa->o_id, oa->o_mode);
 701         }
 702
 703         if (!dentry->d_inode) {
 704                 CERROR("%s on non-existent object: "LPU64"\n", what, oa->o_id);
 705                 f_dput(dentry);
 706                 dentry = ERR_PTR(-ENOENT);
 707         }
 708
 709         return dentry;
 710 }
 711
 712 #define filter_oa2dentry(conn, oa) __filter_oa2dentry(conn, oa, __FUNCTION__)
 713
 714 static int filter_getattr(struct lustre_handle *conn, struct obdo *oa,
 715                           struct lov_stripe_md *md)
 716 {
 717         struct dentry *dentry = NULL;
 718         int rc = 0;
 719         ENTRY;
 720
 721         dentry = filter_oa2dentry(conn, oa);
 722         if (IS_ERR(dentry))
 723                 RETURN(PTR_ERR(dentry));
 724
 725         filter_from_inode(oa, dentry->d_inode, oa->o_valid);
 726
 727         f_dput(dentry);
 728         RETURN(rc);
 729 }
 730
 731 static int filter_setattr(struct lustre_handle *conn, struct obdo *oa,
 732                           struct lov_stripe_md *md)
 733 {
 734         struct obd_run_ctxt saved;
 735         struct obd_device *obd = class_conn2obd(conn);
 736         struct dentry *dentry;
 737         struct iattr iattr;
 738         struct inode *inode;
 739         int rc;
 740         ENTRY;
 741
 742         dentry = filter_oa2dentry(conn, oa);
 743
 744         if (IS_ERR(dentry))
 745                 RETURN(PTR_ERR(dentry));
 746
 747         iattr_from_obdo(&iattr, oa, oa->o_valid);
 748         iattr.ia_mode = (iattr.ia_mode & ~S_IFMT) | S_IFREG;
 749         inode = dentry->d_inode;
 750
 751         lock_kernel();
 752         if (iattr.ia_valid & ATTR_SIZE)
 753                 down(&inode->i_sem);
 754         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 755         if (inode->i_op->setattr)
 756                 rc = inode->i_op->setattr(dentry, &iattr);
 757         else
 758                 rc = inode_setattr(inode, &iattr);
 759         pop_ctxt(&saved);
 760         if (iattr.ia_valid & ATTR_SIZE) {
 761                 up(&inode->i_sem);
 762                 oa->o_valid = OBD_MD_FLBLOCKS | OBD_MD_FLCTIME | OBD_MD_FLMTIME;
 763                 obdo_from_inode(oa, inode, oa->o_valid);
 764         }
 765         unlock_kernel();
 766
 767         f_dput(dentry);
 768         RETURN(rc);
 769 }
 770
 771 static int filter_open(struct lustre_handle *conn, struct obdo *oa,
 772                        struct lov_stripe_md *ea)
 773 {
 774         struct obd_export *export;
 775         struct lustre_handle *handle;
 776         struct filter_file_data *ffd;
 777         struct file *filp;
 778         int rc = 0;
 779         ENTRY;
 780
 781         export = class_conn2export(conn);
 782         if (!export) {
 783                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 784                 RETURN(-EINVAL);
 785         }
 786
 787         filp = filter_obj_open(export, oa->o_id, oa->o_mode);
 788         if (IS_ERR(filp))
 789                 GOTO(out, rc = PTR_ERR(filp));
 790
 791         filter_from_inode(oa, filp->f_dentry->d_inode, oa->o_valid);
 792
 793         ffd = filp->private_data;
 794         handle = obdo_handle(oa);
 795         handle->addr = (__u64)(unsigned long)ffd;
 796         handle->cookie = ffd->ffd_servercookie;
 797         oa->o_valid |= OBD_MD_FLHANDLE;
 798 out:
 799         RETURN(rc);
 800 } /* filter_open */
 801
 802 static int filter_close(struct lustre_handle *conn, struct obdo *oa,
 803                         struct lov_stripe_md *ea)
 804 {
 805         struct obd_export *exp;
 806         struct filter_file_data *ffd;
 807         struct filter_export_data *fed;
 808         int rc;
 809         ENTRY;
 810
 811         exp = class_conn2export(conn);
 812         if (!exp) {
 813                 CDEBUG(D_IOCTL, "fatal: invalid client "LPX64"\n", conn->addr);
 814                 RETURN(-EINVAL);
 815         }
 816
 817         if (!(oa->o_valid & OBD_MD_FLHANDLE)) {
 818                 CERROR("no handle for close of objid "LPX64"\n", oa->o_id);
 819                 RETURN(-EINVAL);
 820         }
 821
 822         ffd = filter_handle2ffd(obdo_handle(oa));
 823         if (!ffd) {
 824                 struct lustre_handle *handle = obdo_handle(oa);
 825                 CERROR("bad handle ("LPX64") or cookie ("LPX64") for close\n",
 826                        handle->addr, handle->cookie);
 827                 RETURN(-ESTALE);
 828         }
 829
 830         fed = &exp->exp_filter_data;
 831         spin_lock(&fed->fed_lock);
 832         list_del(&ffd->ffd_export_list);
 833         spin_unlock(&fed->fed_lock);
 834
 835         rc = filter_close_internal(exp->exp_obd, ffd);
 836
 837         RETURN(rc);
 838 } /* filter_close */
 839
 840 static int filter_create(struct lustre_handle *conn, struct obdo *oa,
 841                          struct lov_stripe_md **ea)
 842 {
 843         struct obd_device *obd = class_conn2obd(conn);
 844         char name[64];
 845         struct obd_run_ctxt saved;
 846         struct dentry *new;
 847         struct iattr;
 848         ENTRY;
 849
 850         if (!obd) {
 851                 CERROR("invalid client "LPX64"\n", conn->addr);
 852                 return -EINVAL;
 853         }
 854
 855         if (!(oa->o_mode & S_IFMT)) {
 856                 CERROR("OBD %s, object "LPU64" has bad type: %o\n",
 857                        __FUNCTION__, oa->o_id, oa->o_mode);
 858                 return -ENOENT;
 859         }
 860
 861         oa->o_id = filter_next_id(obd);
 862
 863         //filter_id(name, oa->o_id, oa->o_mode);
 864         sprintf(name, LPU64, oa->o_id);
 865         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
 866         new = simple_mknod(filter_parent(obd, oa->o_mode), name, oa->o_mode);
 867         pop_ctxt(&saved);
 868         if (IS_ERR(new)) {
 869                 CERROR("Error mknod obj %s, err %ld\n", name, PTR_ERR(new));
 870                 return -ENOENT;
 871         }
 872
 873         /* Set flags for fields we have set in the inode struct */
 874         oa->o_valid = OBD_MD_FLID | OBD_MD_FLBLKSZ | OBD_MD_FLBLOCKS |
 875                  OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME;
 876         filter_from_inode(oa, new->d_inode, oa->o_valid);
 877         f_dput(new);
 878
 879         return 0;
 880 }
 881
 882 static int filter_destroy(struct lustre_handle *conn, struct obdo *oa,
 883                           struct lov_stripe_md *ea)
 884 {
 885         struct obd_device *obd = class_conn2obd(conn);
 886         struct dentry *dir_dentry, *object_dentry;
 887         struct filter_dentry_data *fdd;
 888         int rc;
 889         ENTRY;
 890
 891         if (!obd) {
 892                 CERROR("invalid client "LPX64"\n", conn->addr);
 893                 RETURN(-EINVAL);
 894         }
 895
 896         CDEBUG(D_INODE, "destroying objid "LPX64"\n", oa->o_id);
 897
 898         dir_dentry = filter_parent(obd, oa->o_mode);
 899         down(&dir_dentry->d_inode->i_sem);
 900
 901         object_dentry = filter_oa2dentry(conn, oa);
 902         if (IS_ERR(object_dentry))
 903                 GOTO(out, rc = -ENOENT);
 904
 905         fdd = object_dentry->d_fsdata;
 906         if (fdd && atomic_read(&fdd->fdd_open_count)) {
 907                 if (!(fdd->fdd_flags & FILTER_FLAG_DESTROY)) {
 908                         fdd->fdd_flags |= FILTER_FLAG_DESTROY;
 909                         /* XXX put into PENDING directory in case of crash */
 910                         CDEBUG(D_INODE,
 911                                "defer destroy of %dx open objid "LPX64"\n",
 912                                atomic_read(&fdd->fdd_open_count), oa->o_id);
 913                 } else
 914                         CDEBUG(D_INODE,
 915                                "repeat destroy of %dx open objid "LPX64"\n",
 916                                atomic_read(&fdd->fdd_open_count), oa->o_id);
 917                 GOTO(out_dput, rc = 0);
 918         }
 919
 920         rc = filter_destroy_internal(obd, dir_dentry, object_dentry);
 921 out_dput:
 922         f_dput(object_dentry);
 923
 924         EXIT;
 925 out:
 926         up(&dir_dentry->d_inode->i_sem);
 927         return rc;
 928 }
 929
 930 /* NB count and offset are used for punch, but not truncate */
 931 static int filter_truncate(struct lustre_handle *conn, struct obdo *oa,
 932                            struct lov_stripe_md *lsm,
 933                            obd_off start, obd_off end)
 934 {
 935         int error;
 936         ENTRY;
 937
 938         if (end != OBD_OBJECT_EOF)
 939                 CERROR("PUNCH not supported, only truncate works\n");
 940
 941         CDEBUG(D_INODE, "calling truncate for object "LPX64", valid = %x, "
 942                "o_size = "LPD64"\n", oa->o_id, oa->o_valid, start);
 943         oa->o_size = start;
 944         error = filter_setattr(conn, oa, NULL);
 945         RETURN(error);
 946 }
 947
 948 static int filter_pgcache_brw(int cmd, struct lustre_handle *conn,
 949                               struct lov_stripe_md *lsm, obd_count oa_bufs,
 950                               struct brw_page *pga, brw_cb_t callback,
 951                               struct brw_cb_data *data)
 952 {
 953         struct obd_export       *export = class_conn2export(conn);
 954         struct obd_run_ctxt      saved;
 955         struct super_block      *sb;
 956         int                      pnum;          /* index to pages (bufs) */
 957         unsigned long            retval;
 958         int                      error;
 959         struct file             *file;
 960         int pg;
 961         ENTRY;
 962
 963         if (!export) {
 964                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
 965                 RETURN(-EINVAL);
 966         }
 967
 968         sb = export->exp_obd->u.filter.fo_sb;
 969         push_ctxt(&saved, &export->exp_obd->u.filter.fo_ctxt, NULL);
 970         pnum = 0; /* pnum indexes buf 0..num_pages */
 971
 972         file = filter_obj_open(export, lsm->lsm_object_id, S_IFREG);
 973         if (IS_ERR(file))
 974                 GOTO(out, retval = PTR_ERR(file));
 975
 976         /* count doubles as retval */
 977         for (pg = 0; pg < oa_bufs; pg++) {
 978                 CDEBUG(D_INODE, "OP %d obdo pgno: (%d) (%ld,"LPU64
 979                        ") off count ("LPU64",%d)\n",
 980                        cmd, pnum, file->f_dentry->d_inode->i_ino,
 981                        pga[pnum].off >> PAGE_CACHE_SHIFT, pga[pnum].off,
 982                        (int)pga[pnum].count);
 983                 if (cmd & OBD_BRW_WRITE) {
 984                         loff_t off;
 985                         char *buffer;
 986                         off = pga[pnum].off;
 987                         buffer = kmap(pga[pnum].pg);
 988                         retval = file->f_op->write(file, buffer,
 989                                                    pga[pnum].count,
 990                                                    &off);
 991                         kunmap(pga[pnum].pg);
 992                         CDEBUG(D_INODE, "retval %ld\n", retval);
 993                 } else {
 994                         loff_t off = pga[pnum].off;
 995                         char *buffer = kmap(pga[pnum].pg);
 996
 997                         if (off >= file->f_dentry->d_inode->i_size) {
 998                                 memset(buffer, 0, pga[pnum].count);
 999                                 retval = pga[pnum].count;
1000                         } else {
1001                                 retval = file->f_op->read(file, buffer,
1002                                                           pga[pnum].count, &off);
1003                         }
1004                         kunmap(pga[pnum].pg);
1005
1006                         if (retval != pga[pnum].count) {
1007                                 filp_close(file, 0);
1008                                 GOTO(out, retval = -EIO);
1009                         }
1010                         CDEBUG(D_INODE, "retval %ld\n", retval);
1011                 }
1012                 pnum++;
1013         }
1014         /* sizes and blocks are set by generic_file_write */
1015         /* ctimes/mtimes will follow with a setattr call */
1016         filp_close(file, 0);
1017
1018         /* XXX: do something with callback if it is set? */
1019
1020         EXIT;
1021 out:
1022         pop_ctxt(&saved);
1023         error = (retval >= 0) ? 0 : retval;
1024         return error;
1025 }
1026
1027 /*
1028  * Calculate the number of buffer credits needed to write multiple pages in
1029  * a single ext3/extN transaction.  No, this shouldn't be here, but as yet
1030  * ext3 doesn't have a nice API for calculating this sort of thing in advance.
1031  *
1032  * See comment above ext3_writepage_trans_blocks for details.  We assume
1033  * no data journaling is being done, but it does allow for all of the pages
1034  * being non-contiguous.  If we are guaranteed contiguous pages we could
1035  * reduce the number of (d)indirect blocks a lot.
1036  *
1037  * With N blocks per page and P pages, for each inode we have at most:
1038  * N*P indirect
1039  * min(N*P, blocksize/4 + 1) dindirect blocks
1040  * 1 tindirect
1041  *
1042  * For the entire filesystem, we have at most:
1043  * min(sum(nindir + P), ngroups) bitmap blocks (from the above)
1044  * min(sum(nindir + P), gdblocks) group descriptor blocks (from the above)
1045  * 1 inode block
1046  * 1 superblock
1047  * 2 * EXT3_SINGLEDATA_TRANS_BLOCKS for the quota files
1048  */
1049 static int ext3_credits_needed(struct super_block *sb, int objcount,
1050                                struct obd_ioobj *obj)
1051 {
1052         struct obd_ioobj *o = obj;
1053         int blockpp = 1 << (PAGE_CACHE_SHIFT - sb->s_blocksize_bits);
1054         int addrpp = EXT3_ADDR_PER_BLOCK(sb) * blockpp;
1055         int nbitmaps = 0;
1056         int ngdblocks = 0;
1057         int needed = objcount + 1;
1058         int i;
1059
1060         for (i = 0; i < objcount; i++, o++) {
1061                 int nblocks = o->ioo_bufcnt * blockpp;
1062                 int ndindirect = min(nblocks, addrpp + 1);
1063                 int nindir = nblocks + ndindirect + 1;
1064
1065                 nbitmaps += nindir + nblocks;
1066                 ngdblocks += nindir + nblocks;
1067
1068                 needed += nindir;
1069         }
1070
1071         /* Assumes ext3 and extN have same sb_info layout at the start. */
1072         if (nbitmaps > EXT3_SB(sb)->s_groups_count)
1073                 nbitmaps = EXT3_SB(sb)->s_groups_count;
1074         if (ngdblocks > EXT3_SB(sb)->s_gdb_count)
1075                 ngdblocks = EXT3_SB(sb)->s_gdb_count;
1076
1077         needed += nbitmaps + ngdblocks;
1078
1079 #ifdef CONFIG_QUOTA
1080         /* We assume that there will be 1 bit set in s_dquot.flags for each
1081          * quota file that is active.  This is at least true for now.
1082          */
1083         needed += hweight32(sb_any_quota_enabled(sb)) *
1084                 EXT3_SINGLEDATA_TRANS_BLOCKS;
1085 #endif
1086
1087         return needed;
1088 }
1089
1090 /* We have to start a huge journal transaction here to hold all of the
1091  * metadata for the pages being written here.  This is necessitated by
1092  * the fact that we do lots of prepare_write operations before we do
1093  * any of the matching commit_write operations, so even if we split
1094  * up to use "smaller" transactions none of them could complete until
1095  * all of them were opened.  By having a single journal transaction,
1096  * we eliminate duplicate reservations for common blocks like the
1097  * superblock and group descriptors or bitmaps.
1098  *
1099  * We will start the transaction here, but each prepare_write will
1100  * add a refcount to the transaction, and each commit_write will
1101  * remove a refcount.  The transaction will be closed when all of
1102  * the pages have been written.
1103  */
1104 static void *ext3_filter_journal_start(struct filter_obd *filter,
1105                                        int objcount, struct obd_ioobj *obj,
1106                                        int niocount, struct niobuf_remote *nb)
1107 {
1108         journal_t *journal = NULL;
1109         handle_t *handle = NULL;
1110         int needed;
1111
1112         /* It appears that some kernels have different values for
1113          * EXT*_MAX_GROUP_LOADED (either 8 or 32), so we cannot
1114          * assume anything after s_inode_bitmap_number is the same.
1115          */
1116         if (!strcmp(filter->fo_fstype, "ext3"))
1117                 journal = EXT3_SB(filter->fo_sb)->s_journal;
1118 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1119         else if (!strcmp(filter->fo_fstype, "extN"))
1120                 journal = EXTN_SB(filter->fo_sb)->s_journal;
1121 #endif
1122         needed = ext3_credits_needed(filter->fo_sb, objcount, obj);
1123
1124         /* The number of blocks we could _possibly_ dirty can very large.
1125          * We reduce our request if it is absurd (and we couldn't get that
1126          * many credits for a single handle anyways).
1127          *
1128          * At some point we have to limit the size of I/Os sent at one time,
1129          * increase the size of the journal, or we have to calculate the
1130          * actual journal requirements more carefully by checking all of
1131          * the blocks instead of being maximally pessimistic.  It remains to
1132          * be seen if this is a real problem or not.
1133          */
1134         if (needed > journal->j_max_transaction_buffers) {
1135                 CERROR("want too many journal credits (%d) using %d instead\n",
1136                        needed, journal->j_max_transaction_buffers);
1137                 needed = journal->j_max_transaction_buffers;
1138         }
1139
1140         lock_kernel();
1141         handle = journal_start(journal, needed);
1142         unlock_kernel();
1143         if (IS_ERR(handle))
1144                 CERROR("can't get handle for %d credits: rc = %ld\n", needed,
1145                        PTR_ERR(handle));
1146
1147         return(handle);
1148 }
1149
1150 static void *filter_journal_start(void **journal_save,
1151                                   struct filter_obd *filter,
1152                                   int objcount, struct obd_ioobj *obj,
1153                                   int niocount, struct niobuf_remote *nb)
1154 {
1155         void *handle = NULL;
1156
1157         /* This may not be necessary - we probably never have a
1158          * transaction started when we enter here, so we can
1159          * remove the saving of the journal state entirely.
1160          * For now leave it in just to see if it ever happens.
1161          */
1162         *journal_save = current->journal_info;
1163         if (*journal_save) {
1164                 CERROR("Already have handle %p???\n", *journal_save);
1165                 LBUG();
1166                 current->journal_info = NULL;
1167         }
1168
1169         if (!strcmp(filter->fo_fstype, "ext3") ||
1170             !strcmp(filter->fo_fstype, "extN"))
1171                 handle = ext3_filter_journal_start(filter, objcount, obj,
1172                                                    niocount, nb);
1173         return handle;
1174 }
1175
1176 static int ext3_filter_journal_stop(void *handle)
1177 {
1178         int rc;
1179
1180         /* We got a refcount on the handle for each call to prepare_write,
1181          * so we can drop the "parent" handle here to avoid the need for
1182          * osc to call back into filterobd to close the handle.  The
1183          * remaining references will be dropped in commit_write.
1184          */
1185         lock_kernel();
1186         rc = journal_stop((handle_t *)handle);
1187         unlock_kernel();
1188
1189         return rc;
1190 }
1191
1192 static int filter_journal_stop(void *journal_save, struct filter_obd *filter,
1193                                void *handle)
1194 {
1195         int rc = 0;
1196
1197         if (!strcmp(filter->fo_fstype, "ext3") ||
1198             !strcmp(filter->fo_fstype, "extN"))
1199                 rc = ext3_filter_journal_stop(handle);
1200
1201         if (rc)
1202                 CERROR("error on journal stop: rc = %d\n", rc);
1203
1204         current->journal_info = journal_save;
1205
1206         return rc;
1207 }
1208
1209 static inline void lustre_put_page(struct page *page)
1210 {
1211         kunmap(page);
1212         page_cache_release(page);
1213 }
1214
1215
1216 static struct page *
1217 lustre_get_page_read(struct inode *inode, struct niobuf_remote *rnb)
1218 {
1219         unsigned long index = rnb->offset >> PAGE_SHIFT;
1220         struct address_space *mapping = inode->i_mapping;
1221         struct page *page;
1222         int rc;
1223
1224         page = read_cache_page(mapping, index,
1225                                (filler_t*)mapping->a_ops->readpage, NULL);
1226         if (!IS_ERR(page)) {
1227                 wait_on_page(page);
1228                 kmap(page);
1229                 if (!PageUptodate(page)) {
1230                         CERROR("page index %lu not uptodate\n", index);
1231                         GOTO(err_page, rc = -EIO);
1232                 }
1233                 if (PageError(page)) {
1234                         CERROR("page index %lu has error\n", index);
1235                         GOTO(err_page, rc = -EIO);
1236                 }
1237         }
1238         return page;
1239
1240 err_page:
1241         lustre_put_page(page);
1242         return ERR_PTR(rc);
1243 }
1244
1245 static struct page *
1246 lustre_get_page_write(struct inode *inode, unsigned long index)
1247 {
1248         struct address_space *mapping = inode->i_mapping;
1249         struct page *page;
1250         int rc;
1251
1252         page = grab_cache_page(mapping, index); /* locked page */
1253
1254         if (!IS_ERR(page)) {
1255                 kmap(page);
1256                 /* Note: Called with "O" and "PAGE_SIZE" this is essentially
1257                  * a no-op for most filesystems, because we write the whole
1258                  * page.  For partial-page I/O this will read in the page.
1259                  */
1260                 rc = mapping->a_ops->prepare_write(NULL, page, 0, PAGE_SIZE);
1261                 if (rc) {
1262                         CERROR("page index %lu, rc = %d\n", index, rc);
1263                         if (rc != -ENOSPC)
1264                                 LBUG();
1265                         GOTO(err_unlock, rc);
1266                 }
1267                 /* XXX not sure if we need this if we are overwriting page */
1268                 if (PageError(page)) {
1269                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1270                         LBUG();
1271                         GOTO(err_unlock, rc = -EIO);
1272                 }
1273         }
1274         return page;
1275
1276 err_unlock:
1277         unlock_page(page);
1278         lustre_put_page(page);
1279         return ERR_PTR(rc);
1280 }
1281
1282 static int lustre_commit_write(struct page *page, unsigned from, unsigned to)
1283 {
1284         struct inode *inode = page->mapping->host;
1285         int err;
1286
1287         err = page->mapping->a_ops->commit_write(NULL, page, from, to);
1288 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1289         if (!err && IS_SYNC(inode))
1290                 err = waitfor_one_page(page);
1291 #else
1292 #warning ADD 2.5 waiting code here?
1293 #endif
1294         //SetPageUptodate(page); // the client commit_write will do this
1295
1296         SetPageReferenced(page);
1297         unlock_page(page);
1298         lustre_put_page(page);
1299         return err;
1300 }
1301
1302 struct page *filter_get_page_write(struct inode *inode,
1303                                    struct niobuf_remote *rnb,
1304                                    struct niobuf_local *lnb, int *pglocked)
1305 {
1306         unsigned long index = rnb->offset >> PAGE_SHIFT;
1307         struct address_space *mapping = inode->i_mapping;
1308
1309         struct page *page;
1310         int rc;
1311
1312         //ASSERT_PAGE_INDEX(index, GOTO(err, rc = -EINVAL));
1313         if (*pglocked)
1314                 page = grab_cache_page_nowait(mapping, index); /* locked page */
1315         else
1316                 page = grab_cache_page(mapping, index); /* locked page */
1317
1318
1319         /* This page is currently locked, so get a temporary page instead. */
1320         /* XXX I believe this is a very dangerous thing to do - consider if
1321          *     we had multiple writers for the same file (definitely the case
1322          *     if we are using this codepath).  If writer A locks the page,
1323          *     writer B writes to a copy (as here), writer A drops the page
1324          *     lock, and writer C grabs the lock before B does, then B will
1325          *     later overwrite the data from C, even if C had LDLM locked
1326          *     and initiated the write after B did.
1327          */
1328         if (!page) {
1329                 unsigned long addr;
1330                 CDEBUG(D_PAGE, "ino %ld page %ld locked\n", inode->i_ino,index);
1331                 addr = __get_free_pages(GFP_KERNEL, 0); /* locked page */
1332                 if (!addr) {
1333                         CERROR("no memory for a temp page\n");
1334                         LBUG();
1335                         GOTO(err, rc = -ENOMEM);
1336                 }
1337                 /* XXX debugging */
1338                 memset((void *)addr, 0xBA, PAGE_SIZE);
1339                 page = virt_to_page(addr);
1340                 kmap(page);
1341                 page->index = index;
1342                 lnb->flags |= N_LOCAL_TEMP_PAGE;
1343         } else if (!IS_ERR(page)) {
1344                 (*pglocked)++;
1345                 kmap(page);
1346
1347                 rc = mapping->a_ops->prepare_write(NULL, page,
1348                                                    rnb->offset % PAGE_SIZE,
1349                                                    rnb->len);
1350                 if (rc) {
1351                         CERROR("page index %lu, rc = %d\n", index, rc);
1352                         if (rc != -ENOSPC)
1353                                 LBUG();
1354                         GOTO(err_unlock, rc);
1355                 }
1356                 /* XXX not sure if we need this if we are overwriting page */
1357                 if (PageError(page)) {
1358                         CERROR("error on page index %lu, rc = %d\n", index, rc);
1359                         LBUG();
1360                         GOTO(err_unlock, rc = -EIO);
1361                 }
1362         }
1363         return page;
1364
1365 err_unlock:
1366         unlock_page(page);
1367         lustre_put_page(page);
1368 err:
1369         return ERR_PTR(rc);
1370 }
1371
1372 /*
1373  * We need to balance prepare_write() calls with commit_write() calls.
1374  * If the page has been prepared, but we have no data for it, we don't
1375  * want to overwrite valid data on disk, but we still need to zero out
1376  * data for space which was newly allocated.  Like part of what happens
1377  * in __block_prepare_write() for newly allocated blocks.
1378  *
1379  * XXX currently __block_prepare_write() creates buffers for all the
1380  *     pages, and the filesystems mark these buffers as BH_New if they
1381  *     were newly allocated from disk. We use the BH_New flag similarly.
1382  */
1383 static int filter_commit_write(struct page *page, unsigned from, unsigned to,
1384                                int err)
1385 {
1386 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1387         if (err) {
1388                 unsigned block_start, block_end;
1389                 struct buffer_head *bh, *head = page->buffers;
1390                 unsigned blocksize = head->b_size;
1391                 void *addr = page_address(page);
1392
1393                 /* debugging: just seeing if this ever happens */
1394                 CERROR("called filter_commit_write for obj %ld:%ld on err %d\n",
1395                        page->index, page->mapping->host->i_ino, err);
1396
1397                 /* Currently one buffer per page, but in the future... */
1398                 for (bh = head, block_start = 0; bh != head || !block_start;
1399                      block_start = block_end, bh = bh->b_this_page) {
1400                         block_end = block_start + blocksize;
1401                         if (buffer_new(bh))
1402                                 memset(addr + block_start, 0, blocksize);
1403                 }
1404         }
1405 #endif
1406         return lustre_commit_write(page, from, to);
1407 }
1408
1409 static int filter_preprw(int cmd, struct lustre_handle *conn,
1410                          int objcount, struct obd_ioobj *obj,
1411                          int niocount, struct niobuf_remote *nb,
1412                          struct niobuf_local *res, void **desc_private)
1413 {
1414         struct obd_run_ctxt saved;
1415         struct obd_device *obd;
1416         struct obd_ioobj *o = obj;
1417         struct niobuf_remote *rnb = nb;
1418         struct niobuf_local *lnb = res;
1419         void *journal_save = NULL;
1420         int pglocked = 0;
1421         int rc = 0;
1422         int i;
1423         ENTRY;
1424
1425         obd = class_conn2obd(conn);
1426         if (!obd) {
1427                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1428                 RETURN(-EINVAL);
1429         }
1430         memset(res, 0, sizeof(*res) * niocount);
1431
1432         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1433
1434         if (cmd & OBD_BRW_WRITE) {
1435                 *desc_private = filter_journal_start(&journal_save,
1436                                                      &obd->u.filter,
1437                                                      objcount, obj, niocount,
1438                                                      nb);
1439                 if (IS_ERR(*desc_private))
1440                         GOTO(out_ctxt, rc = PTR_ERR(*desc_private));
1441         }
1442
1443         for (i = 0; i < objcount; i++, o++) {
1444                 struct dentry *dentry;
1445                 struct inode *inode;
1446                 int j;
1447
1448                 dentry = filter_fid2dentry(obd, filter_parent(obd, S_IFREG),
1449                                            o->ioo_id, S_IFREG);
1450                 if (IS_ERR(dentry))
1451                         GOTO(out_clean, rc = PTR_ERR(dentry));
1452                 inode = dentry->d_inode;
1453                 if (!inode) {
1454                         CERROR("trying to BRW to non-existent file "LPU64"\n",
1455                                o->ioo_id);
1456                         f_dput(dentry);
1457                         GOTO(out_clean, rc = -ENOENT);
1458                 }
1459
1460                 for (j = 0; j < o->ioo_bufcnt; j++, rnb++, lnb++) {
1461                         struct page *page;
1462
1463                         if (j == 0)
1464                                 lnb->dentry = dentry;
1465                         else
1466                                 lnb->dentry = dget(dentry);
1467
1468                         if (cmd & OBD_BRW_WRITE)
1469                                 page = filter_get_page_write(inode, rnb, lnb,
1470                                                              &pglocked);
1471                         else
1472                                 page = lustre_get_page_read(inode, rnb);
1473
1474                         if (IS_ERR(page)) {
1475                                 f_dput(dentry);
1476                                 GOTO(out_clean, rc = PTR_ERR(page));
1477                         }
1478
1479                         lnb->addr = page_address(page);
1480                         lnb->offset = rnb->offset;
1481                         lnb->page = page;
1482                         lnb->len = rnb->len;
1483                 }
1484         }
1485
1486 out_stop:
1487         if (cmd & OBD_BRW_WRITE) {
1488                 int err = filter_journal_stop(journal_save, &obd->u.filter,
1489                                               *desc_private);
1490                 if (!rc)
1491                         rc = err;
1492         }
1493 out_ctxt:
1494         pop_ctxt(&saved);
1495         RETURN(rc);
1496 out_clean:
1497         while (lnb-- > res) {
1498                 CERROR("error cleanup on brw\n");
1499                 f_dput(lnb->dentry);
1500                 if (cmd & OBD_BRW_WRITE)
1501                         filter_commit_write(lnb->page, 0, PAGE_SIZE, rc);
1502                 else
1503                         lustre_put_page(lnb->page);
1504         }
1505         goto out_stop;
1506 }
1507
1508 static int filter_write_locked_page(struct niobuf_local *lnb)
1509 {
1510         struct page *lpage;
1511         int rc;
1512
1513         lpage = lustre_get_page_write(lnb->dentry->d_inode, lnb->page->index);
1514         if (IS_ERR(lpage)) {
1515                 /* It is highly unlikely that we would ever get an error here.
1516                  * The page we want to get was previously locked, so it had to
1517                  * have already allocated the space, and we were just writing
1518                  * over the same data, so there would be no hole in the file.
1519                  *
1520                  * XXX: possibility of a race with truncate could exist, need
1521                  *      to check that.  There are no guarantees w.r.t.
1522                  *      write order even on a local filesystem, although the
1523                  *      normal response would be to return the number of bytes
1524                  *      successfully written and leave the rest to the app.
1525                  */
1526                 rc = PTR_ERR(lpage);
1527                 CERROR("error getting locked page index %ld: rc = %d\n",
1528                        lnb->page->index, rc);
1529                 GOTO(out, rc);
1530         }
1531
1532         /* lpage is kmapped in lustre_get_page_write() above and kunmapped in
1533          * lustre_commit_write() below, lnb->page was kmapped previously in
1534          * filter_get_page_write() and kunmapped in lustre_put_page() below.
1535          */
1536         memcpy(page_address(lpage), page_address(lnb->page), PAGE_SIZE);
1537         rc = lustre_commit_write(lpage, 0, PAGE_SIZE);
1538         if (rc)
1539                 CERROR("error committing locked page %ld: rc = %d\n",
1540                        lnb->page->index, rc);
1541 out:
1542         lustre_put_page(lnb->page);
1543
1544         return rc;
1545 }
1546
1547 static int filter_commitrw(int cmd, struct lustre_handle *conn,
1548                            int objcount, struct obd_ioobj *obj,
1549                            int niocount, struct niobuf_local *res,
1550                            void *private)
1551 {
1552         struct obd_run_ctxt saved;
1553         struct obd_ioobj *o;
1554         struct niobuf_local *r;
1555         struct obd_device *obd = class_conn2obd(conn);
1556         void *journal_save;
1557         int found_locked = 0;
1558         int rc = 0;
1559         int i;
1560         ENTRY;
1561
1562         push_ctxt(&saved, &obd->u.filter.fo_ctxt, NULL);
1563         lock_kernel();
1564         journal_save = current->journal_info;
1565         LASSERT(!journal_save);
1566
1567         current->journal_info = private;
1568         unlock_kernel();
1569         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1570                 int j;
1571                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1572                         struct page *page = r->page;
1573
1574                         if (!page)
1575                                 LBUG();
1576
1577                         if (r->flags & N_LOCAL_TEMP_PAGE) {
1578                                 found_locked++;
1579                                 continue;
1580                         }
1581
1582                         if (cmd & OBD_BRW_WRITE) {
1583                                 int err = filter_commit_write(page, 0,
1584                                                               r->len, 0);
1585
1586                                 if (!rc)
1587                                         rc = err;
1588                         } else
1589                                 lustre_put_page(page);
1590
1591                         f_dput(r->dentry);
1592                 }
1593         }
1594         lock_kernel();
1595         current->journal_info = journal_save;
1596         unlock_kernel();
1597
1598         if (!found_locked)
1599                 goto out_ctxt;
1600
1601         for (i = 0, o = obj, r = res; i < objcount; i++, o++) {
1602                 int j;
1603                 for (j = 0 ; j < o->ioo_bufcnt ; j++, r++) {
1604                         int err;
1605                         if (!(r->flags & N_LOCAL_TEMP_PAGE))
1606                                 continue;
1607
1608                         err = filter_write_locked_page(r);
1609                         if (!rc)
1610                                 rc = err;
1611                         f_dput(r->dentry);
1612                 }
1613         }
1614
1615 out_ctxt:
1616         pop_ctxt(&saved);
1617         RETURN(rc);
1618 }
1619
1620 static int filter_statfs(struct lustre_handle *conn, struct obd_statfs *osfs)
1621 {
1622         struct obd_device *obd = class_conn2obd(conn);
1623         struct statfs sfs;
1624         int rc;
1625
1626         ENTRY;
1627         rc = vfs_statfs(obd->u.filter.fo_sb, &sfs);
1628         if (!rc)
1629                 statfs_pack(osfs, &sfs);
1630
1631         return rc;
1632 }
1633
1634 static int filter_get_info(struct lustre_handle *conn, obd_count keylen,
1635                            void *key, obd_count *vallen, void **val)
1636 {
1637         struct obd_device *obd;
1638         ENTRY;
1639
1640         obd = class_conn2obd(conn);
1641         if (!obd) {
1642                 CDEBUG(D_IOCTL, "invalid client "LPX64"\n", conn->addr);
1643                 RETURN(-EINVAL);
1644         }
1645
1646         if ( keylen == strlen("blocksize") &&
1647              memcmp(key, "blocksize", keylen) == 0 ) {
1648                 *vallen = sizeof(long);
1649                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize;
1650                 RETURN(0);
1651         }
1652
1653         if ( keylen == strlen("blocksize_bits") &&
1654              memcmp(key, "blocksize_bits", keylen) == 0 ){
1655                 *vallen = sizeof(long);
1656                 *val = (void *)(long)obd->u.filter.fo_sb->s_blocksize_bits;
1657                 RETURN(0);
1658         }
1659
1660         if ( keylen == strlen("root_ino") &&
1661              memcmp(key, "root_ino", keylen) == 0 ){
1662                 *vallen = sizeof(obd_id);
1663                 *val = (void *)(obd_id)FILTER_ROOTINO;
1664                 RETURN(0);
1665         }
1666
1667         CDEBUG(D_IOCTL, "invalid key\n");
1668         RETURN(-EINVAL);
1669 }
1670
1671 int filter_copy_data(struct lustre_handle *dst_conn, struct obdo *dst,
1672                   struct lustre_handle *src_conn, struct obdo *src,
1673                   obd_size count, obd_off offset)
1674 {
1675         struct page *page;
1676         struct lov_stripe_md srcmd, dstmd;
1677         unsigned long index = 0;
1678         int err = 0;
1679
1680         memset(&srcmd, 0, sizeof(srcmd));
1681         memset(&dstmd, 0, sizeof(dstmd));
1682         srcmd.lsm_object_id = src->o_id;
1683         dstmd.lsm_object_id = dst->o_id;
1684
1685         ENTRY;
1686         CDEBUG(D_INFO, "src: ino "LPU64" blocks "LPU64", size "LPU64
1687                ", dst: ino "LPU64"\n",
1688                src->o_id, src->o_blocks, src->o_size, dst->o_id);
1689         page = alloc_page(GFP_USER);
1690         if (page == NULL)
1691                 RETURN(-ENOMEM);
1692
1693 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1694         while (TryLockPage(page))
1695                 ___wait_on_page(page);
1696 #else
1697         wait_on_page_locked(page);
1698 #endif
1699
1700         /* XXX with brw vector I/O, we could batch up reads and writes here,
1701          *     all we need to do is allocate multiple pages to handle the I/Os
1702          *     and arrays to handle the request parameters.
1703          */
1704         while (index < ((src->o_size + PAGE_SIZE - 1) >> PAGE_SHIFT)) {
1705                 struct brw_page pg;
1706                 struct brw_cb_data *brw_cbd = ll_init_brw_cb_data();
1707
1708                 if (!brw_cbd) {
1709                         err = -ENOMEM;
1710                         EXIT;
1711                         break;
1712                 }
1713
1714                 pg.pg = page;
1715                 pg.count = PAGE_SIZE;
1716                 pg.off = (page->index) << PAGE_SHIFT;
1717                 pg.flag = 0;
1718
1719                 page->index = index;
1720                 err = obd_brw(OBD_BRW_READ, src_conn, &srcmd, 1, &pg,
1721                               ll_sync_brw_cb, brw_cbd);
1722
1723                 if ( err ) {
1724                         EXIT;
1725                         break;
1726                 }
1727
1728                 brw_cbd = ll_init_brw_cb_data();
1729                 if (!brw_cbd) {
1730                         err = -ENOMEM;
1731                         EXIT;
1732                         break;
1733                 }
1734                 pg.flag = OBD_BRW_CREATE;
1735                 CDEBUG(D_INFO, "Read page %ld ...\n", page->index);
1736
1737                 err = obd_brw(OBD_BRW_WRITE, dst_conn, &dstmd, 1, &pg,
1738                               ll_sync_brw_cb, brw_cbd);
1739
1740                 /* XXX should handle dst->o_size, dst->o_blocks here */
1741                 if ( err ) {
1742                         EXIT;
1743                         break;
1744                 }
1745
1746                 CDEBUG(D_INFO, "Wrote page %ld ...\n", page->index);
1747
1748                 index++;
1749         }
1750         dst->o_size = src->o_size;
1751         dst->o_blocks = src->o_blocks;
1752         dst->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1753         unlock_page(page);
1754         __free_page(page);
1755
1756         RETURN(err);
1757 }
1758 int filter_attach(struct obd_device *dev,
1759                    obd_count len, void *data)
1760 {
1761
1762         int rc;
1763         rc = lprocfs_reg_obd(dev, (struct lprocfs_vars*)status_var_nm_1,
1764                              (void*)dev);
1765         return rc;
1766 }
1767
1768 int filter_detach(struct obd_device *dev)
1769 {
1770         int rc;
1771         rc = lprocfs_dereg_obd(dev);
1772         return rc;
1773
1774 }
1775 static struct obd_ops filter_obd_ops = {
1776         o_attach:      filter_attach,
1777         o_detach:      filter_detach,
1778         o_get_info:    filter_get_info,
1779         o_setup:       filter_setup,
1780         o_cleanup:     filter_cleanup,
1781         o_connect:     filter_connect,
1782         o_disconnect:  filter_disconnect,
1783         o_statfs:      filter_statfs,
1784         o_getattr:     filter_getattr,
1785         o_create:      filter_create,
1786         o_setattr:     filter_setattr,
1787         o_destroy:     filter_destroy,
1788         o_open:        filter_open,
1789         o_close:       filter_close,
1790         o_brw:         filter_pgcache_brw,
1791         o_punch:       filter_truncate,
1792         o_preprw:      filter_preprw,
1793         o_commitrw:    filter_commitrw
1794 #if 0
1795         o_preallocate: filter_preallocate_inodes,
1796         o_migrate:     filter_migrate,
1797         o_copy:        filter_copy_data,
1798         o_iterate:     filter_iterate
1799 #endif
1800 };
1801
1802
1803 static int __init obdfilter_init(void)
1804 {
1805         printk(KERN_INFO "Filtering OBD driver  v0.001, info@clusterfs.com\n");
1806         filter_open_cache = kmem_cache_create("ll_filter_fdata",
1807                                               sizeof(struct filter_file_data),
1808                                               0, 0, NULL, NULL);
1809         if (!filter_open_cache)
1810                 RETURN(-ENOMEM);
1811
1812         filter_dentry_cache = kmem_cache_create("ll_filter_dentry",
1813                                         sizeof(struct filter_dentry_data),
1814                                         0, 0, NULL, NULL);
1815         if (!filter_dentry_cache) {
1816                 kmem_cache_destroy(filter_open_cache);
1817                 RETURN(-ENOMEM);
1818         }
1819
1820         return class_register_type(&filter_obd_ops,
1821                                    (struct lprocfs_vars*)status_class_var,
1822                                    OBD_FILTER_DEVICENAME);
1823 }
1824
1825 static void __exit obdfilter_exit(void)
1826 {
1827         class_unregister_type(OBD_FILTER_DEVICENAME);
1828         if (kmem_cache_destroy(filter_dentry_cache))
1829                 CERROR("couldn't free obdfilter dentry cache\n");
1830         if (kmem_cache_destroy(filter_open_cache))
1831                 CERROR("couldn't free obdfilter open cache\n");
1832 }
1833
1834 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
1835 MODULE_DESCRIPTION("Lustre Filtering OBD driver v1.0");
1836 MODULE_LICENSE("GPL");
1837
1838 module_init(obdfilter_init);
1839 module_exit(obdfilter_exit);