lustre/mds/handler.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/mds/handler.c
   5  *
   6  *  Lustre Metadata Server (mds) request handler
   7  *
   8  *  Copyright (C) 2001, 2002 Cluster File Systems, Inc.
   9  *
  10  *  This code is issued under the GNU General Public License.
  11  *  See the file COPYING in this distribution
  12  *
  13  *  by Peter Braam <braam@clusterfs.com> &
  14  *     Andreas Dilger <braam@clusterfs.com>
  15  *
  16  *  This server is single threaded at present (but can easily be multi threaded)
  17  *
  18  */
  19
  20 #define EXPORT_SYMTAB
  21 #define DEBUG_SUBSYSTEM S_MDS
  22
  23 #include <linux/module.h>
  24 #include <linux/lustre_mds.h>
  25 #include <linux/lustre_dlm.h>
  26 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
  27                                 struct ptlrpc_request *req);
  28 static int mds_cleanup(struct obd_device * obddev);
  29
  30 static struct mds_obd *mds_export2mds(struct obd_export *export)
  31 {
  32         return &export->export_obd->u.mds;
  33 }
  34
  35 /* Assumes caller has already pushed into the kernel filesystem context */
  36 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  37                         __u64 offset)
  38 {
  39         int rc = 0;
  40         struct mds_obd *mds = &req->rq_obd->u.mds;
  41         struct ptlrpc_bulk_desc *desc;
  42         struct ptlrpc_bulk_page *bulk;
  43         char *buf;
  44         ENTRY;
  45
  46         desc = ptlrpc_prep_bulk(req->rq_connection);
  47         if (desc == NULL)
  48                 GOTO(out, rc = -ENOMEM);
  49
  50         bulk = ptlrpc_prep_bulk_page(desc);
  51         if (bulk == NULL)
  52                 GOTO(cleanup_bulk, rc = -ENOMEM);
  53
  54         OBD_ALLOC(buf, PAGE_SIZE);
  55         if (buf == NULL)
  56                 GOTO(cleanup_bulk, rc = -ENOMEM);
  57
  58         rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
  59
  60         if (rc != PAGE_SIZE)
  61                 GOTO(cleanup_buf, rc = -EIO);
  62
  63         bulk->b_xid = req->rq_xid;
  64         bulk->b_buf = buf;
  65         bulk->b_buflen = PAGE_SIZE;
  66         desc->b_portal = MDS_BULK_PORTAL;
  67
  68         rc = ptlrpc_send_bulk(desc);
  69         if (rc)
  70                 GOTO(cleanup_buf, rc);
  71
  72         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
  73                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
  74                        OBD_FAIL_MDS_SENDPAGE, rc);
  75                 ptlrpc_abort_bulk(desc);
  76                 GOTO(cleanup_buf, rc);
  77         }
  78
  79         wait_event_interruptible(desc->b_waitq, ptlrpc_check_bulk_sent(desc));
  80         if (desc->b_flags & PTL_RPC_FL_INTR)
  81                 GOTO(cleanup_buf, rc = -EINTR);
  82
  83         EXIT;
  84  cleanup_buf:
  85         OBD_FREE(buf, PAGE_SIZE);
  86  cleanup_bulk:
  87         ptlrpc_free_bulk(desc);
  88  out:
  89         return rc;
  90 }
  91
  92 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
  93                               struct vfsmount **mnt)
  94 {
  95         /* stolen from NFS */
  96         struct super_block *sb = mds->mds_sb;
  97         unsigned long ino = fid->id;
  98         __u32 generation = fid->generation;
  99         struct inode *inode;
 100         struct list_head *lp;
 101         struct dentry *result;
 102
 103         if (ino == 0)
 104                 RETURN(ERR_PTR(-ESTALE));
 105
 106         inode = iget(sb, ino);
 107         if (inode == NULL)
 108                 RETURN(ERR_PTR(-ENOMEM));
 109
 110         CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
 111
 112         if (is_bad_inode(inode) ||
 113             (generation && inode->i_generation != generation)) {
 114                 /* we didn't find the right inode.. */
 115                 CERROR("bad inode %lu, link: %d ct: %d or version  %u/%u\n",
 116                        inode->i_ino, inode->i_nlink,
 117                        atomic_read(&inode->i_count), inode->i_generation,
 118                        generation);
 119                 LBUG();
 120                 iput(inode);
 121                 RETURN(ERR_PTR(-ESTALE));
 122         }
 123
 124         /* now to find a dentry.
 125          * If possible, get a well-connected one
 126          */
 127         if (mnt)
 128                 *mnt = mds->mds_vfsmnt;
 129         spin_lock(&dcache_lock);
 130         for (lp = inode->i_dentry.next; lp != &inode->i_dentry ; lp=lp->next) {
 131                 result = list_entry(lp,struct dentry, d_alias);
 132                 if (! (result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
 133                         dget_locked(result);
 134                         result->d_vfs_flags |= DCACHE_REFERENCED;
 135                         spin_unlock(&dcache_lock);
 136                         iput(inode);
 137                         if (mnt)
 138                                 mntget(*mnt);
 139                         return result;
 140                 }
 141         }
 142         spin_unlock(&dcache_lock);
 143         result = d_alloc_root(inode);
 144         if (result == NULL) {
 145                 iput(inode);
 146                 return ERR_PTR(-ENOMEM);
 147         }
 148         if (mnt)
 149                 mntget(*mnt);
 150         result->d_flags |= DCACHE_NFSD_DISCONNECTED;
 151         return result;
 152 }
 153
 154 static int mds_connect(struct obd_conn *conn, struct obd_device *obd)
 155 {
 156         int rc;
 157
 158         MOD_INC_USE_COUNT;
 159         rc = gen_connect(conn, obd);
 160
 161         if (rc)
 162                 MOD_DEC_USE_COUNT;
 163
 164         return rc;
 165 }
 166
 167 static int mds_disconnect(struct obd_conn *conn)
 168 {
 169         int rc;
 170
 171         rc = gen_disconnect(conn);
 172         if (!rc)
 173                 MOD_DEC_USE_COUNT;
 174
 175         return rc;
 176 }
 177
 178 /* FIXME: the error cases need fixing to avoid leaks */
 179 static int mds_getstatus(struct ptlrpc_request *req)
 180 {
 181         struct mds_obd *mds = mds_export2mds(req->rq_export);
 182         struct mds_body *body;
 183         struct mds_client_info *mci;
 184         struct mds_client_data *mcd;
 185         int rc, size = sizeof(*body);
 186         ENTRY;
 187
 188         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 189         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 190                 CERROR("mds: out of memory for message: size=%d\n", size);
 191                 req->rq_status = -ENOMEM;
 192                 RETURN(0);
 193         }
 194
 195         body = lustre_msg_buf(req->rq_reqmsg, 0);
 196         mds_unpack_body(body);
 197         /* Anything we need to do here with the client's trans no or so? */
 198
 199         body = lustre_msg_buf(req->rq_repmsg, 0);
 200         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 201
 202         mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
 203         if (!mci) {
 204                 /* We don't have any old connection data for this client */
 205                 int rc;
 206
 207                 CDEBUG(D_INFO, "allocating new client data for UUID '%s'",
 208                        ptlrpc_req_to_uuid(req));
 209
 210                 OBD_ALLOC(mcd, sizeof(*mcd));
 211                 if (!mcd) {
 212                         CERROR("mds: out of memory for client data\n");
 213                         req->rq_status = -ENOMEM;
 214                         RETURN(0);
 215                 }
 216                 memcpy(mcd->mcd_uuid, ptlrpc_req_to_uuid(req),
 217                        sizeof(mcd->mcd_uuid));
 218                 rc = mds_client_add(mds, mcd, -1);
 219                 if (rc) {
 220                         req->rq_status = rc;
 221                         RETURN(0);
 222                 }
 223         } else {
 224                 /* We have old connection data for this client... */
 225                 mcd = mci->mci_mcd;
 226                 CDEBUG(D_INFO, "found existing data for UUID '%s' at #%d\n",
 227                        mcd->mcd_uuid, mci->mci_off);
 228         }
 229         /* mcd_last_xid is is stored in little endian on the disk and
 230            mds_pack_rep_body converts it to network order */
 231         body->last_xid = le32_to_cpu(mcd->mcd_last_xid);
 232         mds_pack_rep_body(req);
 233         RETURN(0);
 234 }
 235
 236 int mds_lock_callback(struct lustre_handle *lockh, struct ldlm_lock_desc *desc,
 237                       void *data, int data_len, struct ptlrpc_request **reqp)
 238 {
 239         ENTRY;
 240
 241         if (desc == NULL) {
 242                 /* Completion AST.  Do nothing */
 243                 RETURN(0);
 244         }
 245
 246         if (ldlm_cli_cancel(lockh) < 0)
 247                 LBUG();
 248         RETURN(0);
 249 }
 250
 251 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 252 {
 253         struct mds_obd *mds = &req->rq_obd->u.mds;
 254         struct obd_run_ctxt saved;
 255         struct mds_body *body;
 256         struct dentry *de = NULL, *dchild = NULL;
 257         struct inode *dir;
 258         struct lustre_handle lockh;
 259         char *name;
 260         int namelen, flags, lock_mode, rc = 0;
 261         __u64 res_id[3] = {0, 0, 0};
 262         ENTRY;
 263
 264         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 265                 LBUG();
 266
 267         if (req->rq_reqmsg->bufcount <= offset + 1) {
 268                 LBUG();
 269                 GOTO(out_pre_de, rc = -EINVAL);
 270         }
 271
 272         body = lustre_msg_buf(req->rq_reqmsg, offset);
 273         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 274         namelen = req->rq_reqmsg->buflens[offset + 1];
 275         /* requests were at offset 2, replies go back at 1 */
 276         if (offset)
 277                 offset = 1;
 278
 279         push_ctxt(&saved, &mds->mds_ctxt);
 280         de = mds_fid2dentry(mds, &body->fid1, NULL);
 281         if (IS_ERR(de)) {
 282                 LBUG();
 283                 GOTO(out_pre_de, rc = -ESTALE);
 284         }
 285
 286         dir = de->d_inode;
 287         CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
 288
 289         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 290         res_id[0] = dir->i_ino;
 291
 292         rc = ldlm_lock_match(mds->mds_local_namespace, res_id, LDLM_PLAIN,
 293                              NULL, 0, lock_mode, &lockh);
 294         if (rc == 0) {
 295                 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
 296                 rc = ldlm_cli_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
 297                                       NULL, mds->mds_local_namespace, NULL,
 298                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 299                                       &flags, (void *)mds_lock_callback,
 300                                       NULL, 0, &lockh);
 301                 if (rc != ELDLM_OK) {
 302                         CERROR("lock enqueue: err: %d\n", rc);
 303                         GOTO(out_create_de, rc = -EIO);
 304                 }
 305         }
 306         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 307
 308         down(&dir->i_sem);
 309         dchild = lookup_one_len(name, de, namelen - 1);
 310         if (IS_ERR(dchild)) {
 311                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 312                 up(&dir->i_sem);
 313                 LBUG();
 314                 GOTO(out_create_dchild, rc = -ESTALE);
 315         }
 316
 317         if (dchild->d_inode) {
 318                 struct mds_body *body;
 319                 struct obdo *obdo;
 320                 struct inode *inode = dchild->d_inode;
 321                 CERROR("child exists (dir %ld, name %s, ino %ld)\n",
 322                        dir->i_ino, name, dchild->d_inode->i_ino);
 323
 324                 body = lustre_msg_buf(req->rq_repmsg, offset);
 325                 mds_pack_inode2fid(&body->fid1, inode);
 326                 mds_pack_inode2body(body, inode);
 327                 if (S_ISREG(inode->i_mode)) {
 328                         obdo = lustre_msg_buf(req->rq_repmsg, offset + 1);
 329                         mds_fs_get_obdo(mds, inode, obdo);
 330                 }
 331                 /* now a normal case for intent locking */
 332                 rc = 0;
 333         } else {
 334                 rc = -ENOENT;
 335         }
 336
 337         EXIT;
 338 out_create_dchild:
 339         l_dput(dchild);
 340         up(&dir->i_sem);
 341         ldlm_lock_decref(&lockh, lock_mode);
 342 out_create_de:
 343         l_dput(de);
 344 out_pre_de:
 345         req->rq_status = rc;
 346         pop_ctxt(&saved);
 347         return 0;
 348 }
 349
 350
 351 static int mds_getattr(int offset, struct ptlrpc_request *req)
 352 {
 353         struct mds_obd *mds = &req->rq_obd->u.mds;
 354         struct obd_run_ctxt saved;
 355         struct dentry *de;
 356         struct inode *inode;
 357         struct mds_body *body;
 358         int rc, size[2] = {sizeof(*body)}, bufcount = 1;
 359         ENTRY;
 360
 361         body = lustre_msg_buf(req->rq_reqmsg, offset);
 362         push_ctxt(&saved, &mds->mds_ctxt);
 363         de = mds_fid2dentry(mds, &body->fid1, NULL);
 364         if (IS_ERR(de)) {
 365                 GOTO(out_pop, rc = -ENOENT);
 366         }
 367
 368         inode = de->d_inode;
 369         if (S_ISREG(body->fid1.f_type)) {
 370                 bufcount = 2;
 371                 size[1] = sizeof(struct obdo);
 372         } else if (body->valid & OBD_MD_LINKNAME) {
 373                 bufcount = 2;
 374                 size[1] = inode->i_size;
 375         }
 376
 377         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 378                              &req->rq_repmsg);
 379         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 380                 CERROR("mds: out of memory\n");
 381                 GOTO(out, rc);
 382         }
 383
 384         if (body->valid & OBD_MD_LINKNAME) {
 385                 char *tmp = lustre_msg_buf(req->rq_repmsg, 1);
 386
 387                 rc = inode->i_op->readlink(de, tmp, size[1]);
 388
 389                 if (rc < 0) {
 390                         CERROR("readlink failed: %d\n", rc);
 391                         GOTO(out, rc);
 392                 }
 393         }
 394
 395         body = lustre_msg_buf(req->rq_repmsg, 0);
 396         body->ino = inode->i_ino;
 397         body->generation = inode->i_generation;
 398         body->atime = inode->i_atime;
 399         body->ctime = inode->i_ctime;
 400         body->mtime = inode->i_mtime;
 401         body->uid = inode->i_uid;
 402         body->gid = inode->i_gid;
 403         body->size = inode->i_size;
 404         body->mode = inode->i_mode;
 405         body->nlink = inode->i_nlink;
 406         body->valid = ~0; /* FIXME: should be more selective */
 407
 408         if (S_ISREG(inode->i_mode)) {
 409                 rc = mds_fs_get_obdo(mds, inode,
 410                                      lustre_msg_buf(req->rq_repmsg, 1));
 411                 if (rc < 0) {
 412                         CERROR("mds_fs_get_obdo failed: %d\n", rc);
 413                         GOTO(out, rc);
 414                 }
 415         }
 416 out:
 417         l_dput(de);
 418 out_pop:
 419         pop_ctxt(&saved);
 420         req->rq_status = rc;
 421         RETURN(0);
 422 }
 423
 424 static int mds_statfs(struct ptlrpc_request *req)
 425 {
 426         struct mds_obd *mds = &req->rq_obd->u.mds;
 427         struct obd_statfs *osfs;
 428         struct statfs sfs;
 429         int rc, size = sizeof(*osfs);
 430         ENTRY;
 431
 432         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 433                              &req->rq_repmsg);
 434         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 435                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 436                 GOTO(out, rc);
 437         }
 438
 439         rc = vfs_statfs(mds->mds_sb, &sfs);
 440         if (rc) {
 441                 CERROR("mds: statfs failed: rc %d\n", rc);
 442                 GOTO(out, rc);
 443         }
 444         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 445         memset(osfs, 0, size);
 446         /* FIXME: hack until the MDS gets the OST data, next month */
 447         sfs.f_blocks = -1;
 448         sfs.f_bfree = -1;
 449         sfs.f_bavail = -1;
 450         obd_statfs_pack(osfs, &sfs);
 451
 452 out:
 453         req->rq_status = rc;
 454         RETURN(0);
 455 }
 456
 457 static int mds_open(struct ptlrpc_request *req)
 458 {
 459         struct dentry *de;
 460         struct mds_body *body;
 461         struct file *file;
 462         struct vfsmount *mnt;
 463         struct mds_obd *mds = &req->rq_obd->u.mds;
 464         struct mds_client_info *mci;
 465         __u32 flags;
 466         struct list_head *tmp;
 467         struct mds_file_data *mfd;
 468         int rc, size = sizeof(*body);
 469         ENTRY;
 470
 471         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 472         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 473                 CERROR("mds: out of memory\n");
 474                 req->rq_status = -ENOMEM;
 475                 RETURN(0);
 476         }
 477
 478         mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
 479         if (!mci) {
 480                 CERROR("mds: no mci!\n");
 481                 req->rq_status = -ENOTCONN;
 482                 RETURN(0);
 483         }
 484
 485         body = lustre_msg_buf(req->rq_reqmsg, 0);
 486
 487         /* was this animal open already? */
 488         /* XXX we should only check on re-open, or do a refcount... */
 489         list_for_each(tmp, &mci->mci_open_head) {
 490                 struct mds_file_data *fd;
 491                 fd = list_entry(tmp, struct mds_file_data, mfd_list);
 492                 if (body->extra == fd->mfd_clientfd &&
 493                     body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
 494                         CERROR("Re opening %Ld\n", body->fid1.id);
 495                         RETURN(0);
 496                 }
 497         }
 498
 499         OBD_ALLOC(mfd, sizeof(*mfd));
 500         if (!mfd) {
 501                 CERROR("mds: out of memory\n");
 502                 req->rq_status = -ENOMEM;
 503                 RETURN(0);
 504         }
 505
 506         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 507         if (IS_ERR(de)) {
 508                 req->rq_status = -ENOENT;
 509                 RETURN(0);
 510         }
 511
 512         /* check if this inode has seen a delayed object creation */
 513         if (req->rq_reqmsg->bufcount > 1) {
 514                 void *handle;
 515                 struct inode *inode = de->d_inode;
 516                 //struct iattr iattr;
 517                 struct obdo *obdo;
 518                 int rc;
 519
 520                 obdo = lustre_msg_buf(req->rq_reqmsg, 1);
 521                 //iattr.ia_valid = ATTR_MODE;
 522                 //iattr.ia_mode = inode->i_mode;
 523
 524                 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 525                 if (!handle) {
 526                         req->rq_status = -ENOMEM;
 527                         RETURN(0);
 528                 }
 529
 530                 /* XXX error handling */
 531                 rc = mds_fs_set_obdo(mds, inode, handle, obdo);
 532                 //                rc = mds_fs_setattr(mds, de, handle, &iattr);
 533                 if (!rc) {
 534                         struct obd_run_ctxt saved;
 535                         push_ctxt(&saved, &mds->mds_ctxt);
 536                         rc = mds_update_last_rcvd(mds, handle, req);
 537                         pop_ctxt(&saved);
 538                 } else {
 539                         req->rq_status = rc;
 540                         RETURN(0);
 541                 }
 542                 /* FIXME: need to return last_rcvd, last_committed */
 543
 544                 /* FIXME: keep rc intact */
 545                 rc = mds_fs_commit(mds, de->d_inode, handle);
 546                 if (rc) {
 547                         req->rq_status = rc;
 548                         RETURN(0);
 549                 }
 550         }
 551
 552         flags = body->flags;
 553         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 554         if (!file || IS_ERR(file)) {
 555                 req->rq_status = -EINVAL;
 556                 OBD_FREE(mfd, sizeof(*mfd));
 557                 RETURN(0);
 558         }
 559
 560         file->private_data = mfd;
 561         mfd->mfd_file = file;
 562         mfd->mfd_clientfd = body->extra;
 563         list_add(&mfd->mfd_list, &mci->mci_open_head);
 564
 565         body = lustre_msg_buf(req->rq_repmsg, 0);
 566         body->extra = (__u64) (unsigned long)file;
 567         RETURN(0);
 568 }
 569
 570 static int mds_close(struct ptlrpc_request *req)
 571 {
 572         struct dentry *de;
 573         struct mds_body *body;
 574         struct file *file;
 575         struct mds_obd *mds = &req->rq_obd->u.mds;
 576         struct vfsmount *mnt;
 577         struct mds_file_data *mfd;
 578         int rc;
 579         ENTRY;
 580
 581         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 582         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 583                 CERROR("mds: out of memory\n");
 584                 req->rq_status = -ENOMEM;
 585                 RETURN(0);
 586         }
 587
 588         body = lustre_msg_buf(req->rq_reqmsg, 0);
 589         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 590         if (IS_ERR(de)) {
 591                 req->rq_status = -ENOENT;
 592                 RETURN(0);
 593         }
 594
 595         file = (struct file *)(unsigned long)body->extra;
 596         if (!file->f_dentry)
 597                 LBUG();
 598         mfd = (struct mds_file_data *)file->private_data;
 599         list_del(&mfd->mfd_list);
 600         OBD_FREE(mfd, sizeof(*mfd));
 601
 602         req->rq_status = filp_close(file, 0);
 603         l_dput(de);
 604         mntput(mnt);
 605
 606         RETURN(0);
 607 }
 608
 609 static int mds_readpage(struct ptlrpc_request *req)
 610 {
 611         struct mds_obd *mds = mds_export2mds(req->rq_export);
 612         struct vfsmount *mnt;
 613         struct dentry *de;
 614         struct file *file;
 615         struct mds_body *body;
 616         struct obd_run_ctxt saved;
 617         int rc, size = sizeof(*body);
 618         ENTRY;
 619
 620         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 621         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 622                 CERROR("mds: out of memory\n");
 623                 GOTO(out, rc = -ENOMEM);
 624         }
 625
 626         body = lustre_msg_buf(req->rq_reqmsg, 0);
 627         push_ctxt(&saved, &mds->mds_ctxt);
 628         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 629         if (IS_ERR(de))
 630                 GOTO(out_pop, rc = PTR_ERR(de));
 631
 632         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 633
 634         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 635         /* note: in case of an error, dentry_open puts dentry */
 636         if (IS_ERR(file))
 637                 GOTO(out_pop, rc = PTR_ERR(file));
 638
 639         /* to make this asynchronous make sure that the handling function
 640            doesn't send a reply when this function completes. Instead a
 641            callback function would send the reply */
 642         rc = mds_sendpage(req, file, body->size);
 643
 644         filp_close(file, 0);
 645 out_pop:
 646         pop_ctxt(&saved);
 647 out:
 648         req->rq_status = rc;
 649         RETURN(0);
 650 }
 651
 652 int mds_reint(int offset, struct ptlrpc_request *req)
 653 {
 654         int rc;
 655         struct mds_update_record rec;
 656
 657         rc = mds_update_unpack(req, offset, &rec);
 658         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 659                 CERROR("invalid record\n");
 660                 req->rq_status = -EINVAL;
 661                 RETURN(0);
 662         }
 663         /* rc will be used to interrupt a for loop over multiple records */
 664         rc = mds_reint_rec(&rec, offset, req);
 665         return rc;
 666 }
 667
 668 int mds_handle(struct ptlrpc_request *req)
 669 {
 670         int rc;
 671         ENTRY;
 672
 673         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 674         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
 675                 CERROR("lustre_mds: Invalid request\n");
 676                 GOTO(out, rc);
 677         }
 678
 679         if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 680                 CERROR("lustre_mds: wrong packet type sent %d\n",
 681                        req->rq_reqmsg->type);
 682                 GOTO(out, rc = -EINVAL);
 683         }
 684
 685         if (req->rq_reqmsg->opc != MDS_CONNECT &&
 686             req->rq_export == NULL)
 687                 GOTO(out, rc = -ENOTCONN);
 688
 689         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 690                 GOTO(out, rc = -EINVAL);
 691
 692         switch (req->rq_reqmsg->opc) {
 693         case MDS_CONNECT:
 694                 CDEBUG(D_INODE, "connect\n");
 695                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
 696                 rc = target_handle_connect(req);
 697                 break;
 698
 699         case MDS_DISCONNECT:
 700                 CDEBUG(D_INODE, "disconnect\n");
 701                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
 702                 rc = target_handle_disconnect(req);
 703                 break;
 704
 705         case MDS_GETSTATUS:
 706                 CDEBUG(D_INODE, "getstatus\n");
 707                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
 708                 rc = mds_getstatus(req);
 709                 break;
 710
 711         case MDS_GETATTR:
 712                 CDEBUG(D_INODE, "getattr\n");
 713                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 714                 rc = mds_getattr(0, req);
 715                 break;
 716
 717         case MDS_STATFS:
 718                 CDEBUG(D_INODE, "statfs\n");
 719                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
 720                 rc = mds_statfs(req);
 721                 break;
 722
 723         case MDS_READPAGE:
 724                 CDEBUG(D_INODE, "readpage\n");
 725                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 726                 rc = mds_readpage(req);
 727
 728                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 729                         return 0;
 730                 break;
 731
 732         case MDS_REINT: {
 733                 int size = sizeof(struct mds_body);
 734                 CDEBUG(D_INODE, "reint\n");
 735                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 736
 737                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 738                                      &req->rq_repmsg);
 739                 if (rc) {
 740                         rc = req->rq_status = -ENOMEM;
 741                         break;
 742                 }
 743                 rc = mds_reint(0, req);
 744                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
 745                 break;
 746         }
 747
 748         case MDS_OPEN:
 749                 CDEBUG(D_INODE, "open\n");
 750                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 751                 rc = mds_open(req);
 752                 break;
 753
 754         case MDS_CLOSE:
 755                 CDEBUG(D_INODE, "close\n");
 756                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 757                 rc = mds_close(req);
 758                 break;
 759
 760         default:
 761                 rc = ptlrpc_error(req->rq_svc, req);
 762                 RETURN(rc);
 763         }
 764
 765         EXIT;
 766 out:
 767         /* Still not 100% sure whether we should reply with the server
 768          * last_rcvd or that of this client.  I'm not sure it even makes
 769          * a difference on a per-client basis, because last_rcvd is global
 770          * and we are not supposed to allow transactions while in recovery.
 771          */
 772         if (rc) {
 773                 CERROR("mds: processing error %d\n", rc);
 774                 ptlrpc_error(req->rq_svc, req);
 775         } else {
 776                 struct mds_obd *mds = &req->rq_obd->u.mds;
 777                 req->rq_repmsg->last_rcvd = HTON__u64(mds->mds_last_rcvd);
 778                 req->rq_repmsg->last_committed =
 779                         HTON__u64(mds->mds_last_committed);
 780                 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
 781                        (unsigned long long)mds->mds_last_rcvd,
 782                        (unsigned long long)mds->mds_last_committed,
 783                        cpu_to_le32(req->rq_xid));
 784                 CDEBUG(D_NET, "sending reply\n");
 785                 ptlrpc_reply(req->rq_svc, req);
 786         }
 787         return 0;
 788 }
 789
 790 /* Update the server data on disk.  This stores the new mount_count and
 791  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
 792  * then the server last_rcvd value may be less than that of the clients.
 793  * This will alert us that we may need to do client recovery.
 794  */
 795 static
 796 int mds_update_server_data(struct mds_obd *mds)
 797 {
 798         struct obd_run_ctxt saved;
 799         struct mds_server_data *msd = mds->mds_server_data;
 800         struct file *filp = mds->mds_rcvd_filp;
 801         loff_t off = 0;
 802         int rc;
 803
 804         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
 805         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
 806
 807         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
 808                (unsigned long long)mds->mds_mount_count,
 809                (unsigned long long)mds->mds_last_rcvd);
 810         push_ctxt(&saved, &mds->mds_ctxt);
 811         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
 812         if (rc != sizeof(*msd)) {
 813                 CERROR("error writing MDS server data: rc = %d\n", rc);
 814                 if (rc > 0)
 815                         RETURN(-EIO);
 816                 RETURN(rc);
 817         }
 818         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
 819         pop_ctxt(&saved);
 820         if (rc)
 821                 CERROR("error flushing MDS server data: rc = %d\n", rc);
 822
 823         return 0;
 824 }
 825
 826 /* Do recovery actions for the MDS */
 827 static int mds_recover(struct obd_device *obddev)
 828 {
 829         struct mds_obd *mds = &obddev->u.mds;
 830         int rc;
 831
 832         /* This happens at the end when recovery is complete */
 833         ++mds->mds_mount_count;
 834         rc = mds_update_server_data(mds);
 835
 836         return rc;
 837 }
 838
 839
 840 /* mount the file system (secretly) */
 841 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
 842 {
 843         struct obd_ioctl_data* data = buf;
 844         struct mds_obd *mds = &obddev->u.mds;
 845         struct vfsmount *mnt;
 846         int rc = 0;
 847         ENTRY;
 848
 849         MOD_INC_USE_COUNT;
 850 #ifdef CONFIG_DEV_RDONLY
 851         dev_clear_rdonly(2);
 852 #endif
 853         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 854                 GOTO(err_dec, rc = -EINVAL);
 855
 856         mds->mds_fstype = strdup(data->ioc_inlbuf2);
 857
 858         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
 859         if (IS_ERR(mnt)) {
 860                 rc = PTR_ERR(mnt);
 861                 CERROR("do_kern_mount failed: rc = %d\n", rc);
 862                 GOTO(err_kfree, rc);
 863         }
 864
 865         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 866         if (!mds->mds_sb)
 867                 GOTO(err_put, rc = -ENODEV);
 868
 869         rc = mds_fs_setup(mds, mnt);
 870         if (rc) {
 871                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
 872                 GOTO(err_put, rc);
 873         }
 874
 875         mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
 876                                            MDC_REPLY_PORTAL, "self",mds_handle);
 877         if (!mds->mds_service) {
 878                 CERROR("failed to start service\n");
 879                 GOTO(err_fs, rc = -EINVAL);
 880         }
 881
 882         rc = ptlrpc_start_thread(obddev, mds->mds_service, "lustre_mds");
 883         if (rc) {
 884                 CERROR("cannot start thread: rc = %d\n", rc);
 885                 GOTO(err_svc, rc);
 886         }
 887
 888         rc = -ENOENT;
 889         mds->mds_ldlm_conn = ptlrpc_uuid_to_connection("self");
 890         if (!mds->mds_ldlm_conn) {
 891                 mds_cleanup(obddev);
 892                 GOTO(err_thread, rc);
 893         }
 894
 895         obddev->obd_namespace =
 896                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
 897         if (obddev->obd_namespace == NULL) {
 898                 LBUG();
 899                 mds_cleanup(obddev);
 900                 GOTO(err_thread, rc);
 901         }
 902
 903         mds->mds_local_namespace =
 904                 ldlm_namespace_new("mds_client", LDLM_NAMESPACE_CLIENT);
 905         if (mds->mds_local_namespace == NULL) {
 906                 LBUG();
 907                 mds_cleanup(obddev);
 908                 GOTO(err_thread, rc);
 909         }
 910
 911         OBD_ALLOC(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
 912         if (mds->mds_ldlm_client == NULL) {
 913                 LBUG();
 914                 mds_cleanup(obddev);
 915                 GOTO(err_thread, rc);
 916         }
 917         ptlrpc_init_client(NULL, NULL, LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
 918                            mds->mds_ldlm_client);
 919         mds->mds_ldlm_client->cli_target_devno = obddev->obd_minor;
 920         mds->mds_ldlm_client->cli_name = "mds ldlm";
 921
 922         rc = mds_recover(obddev);
 923         if (rc)
 924                 GOTO(err_thread, rc);
 925
 926         RETURN(0);
 927
 928 err_thread:
 929         ptlrpc_stop_all_threads(mds->mds_service);
 930 err_svc:
 931         ptlrpc_unregister_service(mds->mds_service);
 932 err_fs:
 933         mds_fs_cleanup(mds);
 934 err_put:
 935         unlock_kernel();
 936         mntput(mds->mds_vfsmnt);
 937         mds->mds_sb = 0;
 938         lock_kernel();
 939 err_kfree:
 940         kfree(mds->mds_fstype);
 941 err_dec:
 942         MOD_DEC_USE_COUNT;
 943         return rc;
 944 }
 945
 946 static int mds_cleanup(struct obd_device * obddev)
 947 {
 948         struct super_block *sb;
 949         struct mds_obd *mds = &obddev->u.mds;
 950
 951         ENTRY;
 952
 953         if ( !list_empty(&obddev->obd_exports) ) {
 954                 CERROR("still has exports!\n");
 955                 RETURN(-EBUSY);
 956         }
 957
 958         ptlrpc_stop_all_threads(mds->mds_service);
 959         ptlrpc_unregister_service(mds->mds_service);
 960
 961         sb = mds->mds_sb;
 962         if (!mds->mds_sb)
 963                 RETURN(0);
 964
 965         mds_update_server_data(mds);
 966
 967         if (mds->mds_rcvd_filp) {
 968                 int rc = filp_close(mds->mds_rcvd_filp, 0);
 969                 mds->mds_rcvd_filp = NULL;
 970
 971                 if (rc)
 972                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
 973         }
 974
 975         unlock_kernel();
 976         mntput(mds->mds_vfsmnt);
 977         mds->mds_sb = 0;
 978         kfree(mds->mds_fstype);
 979
 980         ldlm_namespace_free(mds->mds_local_namespace);
 981         ldlm_namespace_free(obddev->obd_namespace);
 982
 983         if (mds->mds_ldlm_conn != NULL)
 984                 ptlrpc_put_connection(mds->mds_ldlm_conn);
 985
 986         OBD_FREE(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
 987
 988         lock_kernel();
 989 #ifdef CONFIG_DEV_RDONLY
 990         dev_clear_rdonly(2);
 991 #endif
 992         mds_fs_cleanup(mds);
 993
 994         MOD_DEC_USE_COUNT;
 995         RETURN(0);
 996 }
 997
 998 /* use obd ops to offer management infrastructure */
 999 static struct obd_ops mds_obd_ops = {
1000         o_connect:     mds_connect,
1001         o_disconnect:  mds_disconnect,
1002         o_setup:       mds_setup,
1003         o_cleanup:     mds_cleanup,
1004 };
1005
1006 static int __init mds_init(void)
1007 {
1008         inter_module_register("mds_reint", THIS_MODULE, &mds_reint);
1009         inter_module_register("mds_getattr_name", THIS_MODULE,
1010                               &mds_getattr_name);
1011         obd_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1012         return 0;
1013 }
1014
1015 static void __exit mds_exit(void)
1016 {
1017         inter_module_unregister("mds_reint");
1018         inter_module_unregister("mds_getattr_name");
1019         obd_unregister_type(LUSTRE_MDS_NAME);
1020 }
1021
1022 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1023 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1024 MODULE_LICENSE("GPL");
1025
1026 module_init(mds_init);
1027 module_exit(mds_exit);