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 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
  31 {
  32         return &req->rq_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 = mds_req2mds(req);
  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 lustre_handle *conn, struct obd_device *obd)
 155 {
 156         int rc;
 157
 158         MOD_INC_USE_COUNT;
 159         rc = class_connect(conn, obd);
 160
 161         if (rc)
 162                 MOD_DEC_USE_COUNT;
 163
 164         return rc;
 165 }
 166
 167 static int mds_disconnect(struct lustre_handle *conn)
 168 {
 169         int rc;
 170
 171         rc = class_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_req2mds(req);
 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, NULL) < 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 = mds_req2mds(req);
 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_export->export_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                                       (struct lustre_handle *)&mds->mds_connh,
 298                                       NULL, mds->mds_local_namespace, NULL,
 299                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 300                                       &flags, (void *)mds_lock_callback,
 301                                       NULL, 0, &lockh);
 302                 if (rc != ELDLM_OK) {
 303                         CERROR("lock enqueue: err: %d\n", rc);
 304                         GOTO(out_create_de, rc = -EIO);
 305                 }
 306         }
 307         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 308
 309         down(&dir->i_sem);
 310         dchild = lookup_one_len(name, de, namelen - 1);
 311         if (IS_ERR(dchild)) {
 312                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 313                 up(&dir->i_sem);
 314                 LBUG();
 315                 GOTO(out_create_dchild, rc = -ESTALE);
 316         }
 317
 318         if (dchild->d_inode) {
 319                 struct mds_body *body;
 320                 struct obdo *obdo;
 321                 struct inode *inode = dchild->d_inode;
 322                 CERROR("child exists (dir %ld, name %s, ino %ld)\n",
 323                        dir->i_ino, name, dchild->d_inode->i_ino);
 324
 325                 body = lustre_msg_buf(req->rq_repmsg, offset);
 326                 mds_pack_inode2fid(&body->fid1, inode);
 327                 mds_pack_inode2body(body, inode);
 328                 if (S_ISREG(inode->i_mode)) {
 329                         obdo = lustre_msg_buf(req->rq_repmsg, offset + 1);
 330                         mds_fs_get_obdo(mds, inode, obdo);
 331                 }
 332                 /* now a normal case for intent locking */
 333                 rc = 0;
 334         } else {
 335                 rc = -ENOENT;
 336         }
 337
 338         EXIT;
 339 out_create_dchild:
 340         l_dput(dchild);
 341         up(&dir->i_sem);
 342         ldlm_lock_decref(&lockh, lock_mode);
 343 out_create_de:
 344         l_dput(de);
 345 out_pre_de:
 346         req->rq_status = rc;
 347         pop_ctxt(&saved);
 348         return 0;
 349 }
 350
 351
 352 static int mds_getattr(int offset, struct ptlrpc_request *req)
 353 {
 354         struct mds_obd *mds = mds_req2mds(req);
 355         struct obd_run_ctxt saved;
 356         struct dentry *de;
 357         struct inode *inode;
 358         struct mds_body *body;
 359         int rc, size[2] = {sizeof(*body)}, bufcount = 1;
 360         ENTRY;
 361
 362         body = lustre_msg_buf(req->rq_reqmsg, offset);
 363         push_ctxt(&saved, &mds->mds_ctxt);
 364         de = mds_fid2dentry(mds, &body->fid1, NULL);
 365         if (IS_ERR(de)) {
 366                 GOTO(out_pop, rc = -ENOENT);
 367         }
 368
 369         inode = de->d_inode;
 370         if (S_ISREG(body->fid1.f_type)) {
 371                 bufcount = 2;
 372                 size[1] = sizeof(struct obdo);
 373         } else if (body->valid & OBD_MD_LINKNAME) {
 374                 bufcount = 2;
 375                 size[1] = inode->i_size;
 376         }
 377
 378         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 379                              &req->rq_repmsg);
 380         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 381                 CERROR("mds: out of memory\n");
 382                 GOTO(out, rc);
 383         }
 384
 385         if (body->valid & OBD_MD_LINKNAME) {
 386                 char *tmp = lustre_msg_buf(req->rq_repmsg, 1);
 387
 388                 rc = inode->i_op->readlink(de, tmp, size[1]);
 389
 390                 if (rc < 0) {
 391                         CERROR("readlink failed: %d\n", rc);
 392                         GOTO(out, rc);
 393                 }
 394         }
 395
 396         body = lustre_msg_buf(req->rq_repmsg, 0);
 397         body->ino = inode->i_ino;
 398         body->generation = inode->i_generation;
 399         body->atime = inode->i_atime;
 400         body->ctime = inode->i_ctime;
 401         body->mtime = inode->i_mtime;
 402         body->uid = inode->i_uid;
 403         body->gid = inode->i_gid;
 404         body->size = inode->i_size;
 405         body->mode = inode->i_mode;
 406         body->nlink = inode->i_nlink;
 407         body->valid = ~0; /* FIXME: should be more selective */
 408
 409         if (S_ISREG(inode->i_mode)) {
 410                 rc = mds_fs_get_obdo(mds, inode,
 411                                      lustre_msg_buf(req->rq_repmsg, 1));
 412                 if (rc < 0) {
 413                         CERROR("mds_fs_get_obdo failed: %d\n", rc);
 414                         GOTO(out, rc);
 415                 }
 416         }
 417 out:
 418         l_dput(de);
 419 out_pop:
 420         pop_ctxt(&saved);
 421         req->rq_status = rc;
 422         RETURN(0);
 423 }
 424
 425 static int mds_statfs(struct ptlrpc_request *req)
 426 {
 427         struct mds_obd *mds = mds_req2mds(req);
 428         struct obd_statfs *osfs;
 429         struct statfs sfs;
 430         int rc, size = sizeof(*osfs);
 431         ENTRY;
 432
 433         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 434                              &req->rq_repmsg);
 435         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 436                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 437                 GOTO(out, rc);
 438         }
 439
 440         rc = vfs_statfs(mds->mds_sb, &sfs);
 441         if (rc) {
 442                 CERROR("mds: statfs failed: rc %d\n", rc);
 443                 GOTO(out, rc);
 444         }
 445         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 446         memset(osfs, 0, size);
 447         /* FIXME: hack until the MDS gets the OST data, next month */
 448         sfs.f_blocks = -1;
 449         sfs.f_bfree = -1;
 450         sfs.f_bavail = -1;
 451         obd_statfs_pack(osfs, &sfs);
 452
 453 out:
 454         req->rq_status = rc;
 455         RETURN(0);
 456 }
 457
 458 static int mds_open(struct ptlrpc_request *req)
 459 {
 460         struct dentry *de;
 461         struct mds_body *body;
 462         struct file *file;
 463         struct vfsmount *mnt;
 464         struct mds_obd *mds = mds_req2mds(req);
 465         struct mds_client_info *mci;
 466         __u32 flags;
 467         struct list_head *tmp;
 468         struct mds_file_data *mfd;
 469         int rc, size = sizeof(*body);
 470         ENTRY;
 471
 472         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 473         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 474                 CERROR("mds: out of memory\n");
 475                 req->rq_status = -ENOMEM;
 476                 RETURN(0);
 477         }
 478
 479         mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
 480         if (!mci) {
 481                 CERROR("mds: no mci!\n");
 482                 req->rq_status = -ENOTCONN;
 483                 RETURN(0);
 484         }
 485
 486         body = lustre_msg_buf(req->rq_reqmsg, 0);
 487
 488         /* was this animal open already? */
 489         /* XXX we should only check on re-open, or do a refcount... */
 490         list_for_each(tmp, &mci->mci_open_head) {
 491                 struct mds_file_data *fd;
 492                 fd = list_entry(tmp, struct mds_file_data, mfd_list);
 493                 if (body->extra == fd->mfd_clientfd &&
 494                     body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
 495                         CERROR("Re opening %Ld\n", body->fid1.id);
 496                         RETURN(0);
 497                 }
 498         }
 499
 500         OBD_ALLOC(mfd, sizeof(*mfd));
 501         if (!mfd) {
 502                 CERROR("mds: out of memory\n");
 503                 req->rq_status = -ENOMEM;
 504                 RETURN(0);
 505         }
 506
 507         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 508         if (IS_ERR(de)) {
 509                 req->rq_status = -ENOENT;
 510                 RETURN(0);
 511         }
 512
 513         /* check if this inode has seen a delayed object creation */
 514         if (req->rq_reqmsg->bufcount > 1) {
 515                 void *handle;
 516                 struct inode *inode = de->d_inode;
 517                 //struct iattr iattr;
 518                 struct obdo *obdo;
 519                 int rc;
 520
 521                 obdo = lustre_msg_buf(req->rq_reqmsg, 1);
 522                 //iattr.ia_valid = ATTR_MODE;
 523                 //iattr.ia_mode = inode->i_mode;
 524
 525                 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 526                 if (!handle) {
 527                         req->rq_status = -ENOMEM;
 528                         RETURN(0);
 529                 }
 530
 531                 /* XXX error handling */
 532                 rc = mds_fs_set_obdo(mds, inode, handle, obdo);
 533                 //                rc = mds_fs_setattr(mds, de, handle, &iattr);
 534                 if (!rc) {
 535                         struct obd_run_ctxt saved;
 536                         push_ctxt(&saved, &mds->mds_ctxt);
 537                         rc = mds_update_last_rcvd(mds, handle, req);
 538                         pop_ctxt(&saved);
 539                 } else {
 540                         req->rq_status = rc;
 541                         RETURN(0);
 542                 }
 543                 /* FIXME: need to return last_rcvd, last_committed */
 544
 545                 /* FIXME: keep rc intact */
 546                 rc = mds_fs_commit(mds, de->d_inode, handle);
 547                 if (rc) {
 548                         req->rq_status = rc;
 549                         RETURN(0);
 550                 }
 551         }
 552
 553         flags = body->flags;
 554         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 555         if (!file || IS_ERR(file)) {
 556                 req->rq_status = -EINVAL;
 557                 OBD_FREE(mfd, sizeof(*mfd));
 558                 RETURN(0);
 559         }
 560
 561         file->private_data = mfd;
 562         mfd->mfd_file = file;
 563         mfd->mfd_clientfd = body->extra;
 564         list_add(&mfd->mfd_list, &mci->mci_open_head);
 565
 566         body = lustre_msg_buf(req->rq_repmsg, 0);
 567         body->extra = (__u64) (unsigned long)file;
 568         RETURN(0);
 569 }
 570
 571 static int mds_close(struct ptlrpc_request *req)
 572 {
 573         struct dentry *de;
 574         struct mds_body *body;
 575         struct file *file;
 576         struct mds_obd *mds = mds_req2mds(req);
 577         struct vfsmount *mnt;
 578         struct mds_file_data *mfd;
 579         int rc;
 580         ENTRY;
 581
 582         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 583         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 584                 CERROR("mds: out of memory\n");
 585                 req->rq_status = -ENOMEM;
 586                 RETURN(0);
 587         }
 588
 589         body = lustre_msg_buf(req->rq_reqmsg, 0);
 590         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 591         if (IS_ERR(de)) {
 592                 req->rq_status = -ENOENT;
 593                 RETURN(0);
 594         }
 595
 596         file = (struct file *)(unsigned long)body->extra;
 597         if (!file->f_dentry)
 598                 LBUG();
 599         mfd = (struct mds_file_data *)file->private_data;
 600         list_del(&mfd->mfd_list);
 601         OBD_FREE(mfd, sizeof(*mfd));
 602
 603         req->rq_status = filp_close(file, 0);
 604         l_dput(de);
 605         mntput(mnt);
 606
 607         RETURN(0);
 608 }
 609
 610 static int mds_readpage(struct ptlrpc_request *req)
 611 {
 612         struct mds_obd *mds = mds_req2mds(req);
 613         struct vfsmount *mnt;
 614         struct dentry *de;
 615         struct file *file;
 616         struct mds_body *body;
 617         struct obd_run_ctxt saved;
 618         int rc, size = sizeof(*body);
 619         ENTRY;
 620
 621         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 622         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 623                 CERROR("mds: out of memory\n");
 624                 GOTO(out, rc = -ENOMEM);
 625         }
 626
 627         body = lustre_msg_buf(req->rq_reqmsg, 0);
 628         push_ctxt(&saved, &mds->mds_ctxt);
 629         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 630         if (IS_ERR(de))
 631                 GOTO(out_pop, rc = PTR_ERR(de));
 632
 633         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 634
 635         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 636         /* note: in case of an error, dentry_open puts dentry */
 637         if (IS_ERR(file))
 638                 GOTO(out_pop, rc = PTR_ERR(file));
 639
 640         /* to make this asynchronous make sure that the handling function
 641            doesn't send a reply when this function completes. Instead a
 642            callback function would send the reply */
 643         rc = mds_sendpage(req, file, body->size);
 644
 645         filp_close(file, 0);
 646 out_pop:
 647         pop_ctxt(&saved);
 648 out:
 649         req->rq_status = rc;
 650         RETURN(0);
 651 }
 652
 653 int mds_reint(int offset, struct ptlrpc_request *req)
 654 {
 655         int rc;
 656         struct mds_update_record rec;
 657
 658         rc = mds_update_unpack(req, offset, &rec);
 659         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 660                 CERROR("invalid record\n");
 661                 req->rq_status = -EINVAL;
 662                 RETURN(0);
 663         }
 664         /* rc will be used to interrupt a for loop over multiple records */
 665         rc = mds_reint_rec(&rec, offset, req);
 666         return rc;
 667 }
 668
 669 int mds_handle(struct ptlrpc_request *req)
 670 {
 671         int rc;
 672         ENTRY;
 673
 674         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 675         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
 676                 CERROR("lustre_mds: Invalid request\n");
 677                 GOTO(out, rc);
 678         }
 679
 680         if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 681                 CERROR("lustre_mds: wrong packet type sent %d\n",
 682                        req->rq_reqmsg->type);
 683                 GOTO(out, rc = -EINVAL);
 684         }
 685
 686         if (req->rq_reqmsg->opc != MDS_CONNECT &&
 687             req->rq_export == NULL)
 688                 GOTO(out, rc = -ENOTCONN);
 689
 690         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 691                 GOTO(out, rc = -EINVAL);
 692
 693         switch (req->rq_reqmsg->opc) {
 694         case MDS_CONNECT:
 695                 CDEBUG(D_INODE, "connect\n");
 696                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
 697                 rc = target_handle_connect(req);
 698                 break;
 699
 700         case MDS_DISCONNECT:
 701                 CDEBUG(D_INODE, "disconnect\n");
 702                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
 703                 rc = target_handle_disconnect(req);
 704                 goto out;
 705
 706         case MDS_GETSTATUS:
 707                 CDEBUG(D_INODE, "getstatus\n");
 708                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
 709                 rc = mds_getstatus(req);
 710                 break;
 711
 712         case MDS_GETATTR:
 713                 CDEBUG(D_INODE, "getattr\n");
 714                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 715                 rc = mds_getattr(0, req);
 716                 break;
 717
 718         case MDS_STATFS:
 719                 CDEBUG(D_INODE, "statfs\n");
 720                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
 721                 rc = mds_statfs(req);
 722                 break;
 723
 724         case MDS_READPAGE:
 725                 CDEBUG(D_INODE, "readpage\n");
 726                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 727                 rc = mds_readpage(req);
 728
 729                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 730                         return 0;
 731                 break;
 732
 733         case MDS_REINT: {
 734                 int size = sizeof(struct mds_body);
 735                 CDEBUG(D_INODE, "reint\n");
 736                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 737
 738                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 739                                      &req->rq_repmsg);
 740                 if (rc) {
 741                         rc = req->rq_status = -ENOMEM;
 742                         break;
 743                 }
 744                 rc = mds_reint(0, req);
 745                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
 746                 break;
 747         }
 748
 749         case MDS_OPEN:
 750                 CDEBUG(D_INODE, "open\n");
 751                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 752                 rc = mds_open(req);
 753                 break;
 754
 755         case MDS_CLOSE:
 756                 CDEBUG(D_INODE, "close\n");
 757                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 758                 rc = mds_close(req);
 759                 break;
 760
 761         default:
 762                 rc = ptlrpc_error(req->rq_svc, req);
 763                 RETURN(rc);
 764         }
 765
 766         EXIT;
 767
 768         if (!rc) {
 769                 struct mds_obd *mds = mds_req2mds(req);
 770                 req->rq_repmsg->last_rcvd = HTON__u64(mds->mds_last_rcvd);
 771                 req->rq_repmsg->last_committed =
 772                         HTON__u64(mds->mds_last_committed);
 773                 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
 774                        (unsigned long long)mds->mds_last_rcvd,
 775                        (unsigned long long)mds->mds_last_committed,
 776                        cpu_to_le32(req->rq_xid));
 777         }
 778  out:
 779         /* Still not 100% sure whether we should reply with the server
 780          * last_rcvd or that of this client.  I'm not sure it even makes
 781          * a difference on a per-client basis, because last_rcvd is global
 782          * and we are not supposed to allow transactions while in recovery.
 783          */
 784         if (rc) {
 785                 CERROR("mds: processing error %d\n", rc);
 786                 ptlrpc_error(req->rq_svc, req);
 787         } else {
 788                 CDEBUG(D_NET, "sending reply\n");
 789                 ptlrpc_reply(req->rq_svc, req);
 790         }
 791         return 0;
 792 }
 793
 794 /* Update the server data on disk.  This stores the new mount_count and
 795  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
 796  * then the server last_rcvd value may be less than that of the clients.
 797  * This will alert us that we may need to do client recovery.
 798  */
 799 static
 800 int mds_update_server_data(struct mds_obd *mds)
 801 {
 802         struct obd_run_ctxt saved;
 803         struct mds_server_data *msd = mds->mds_server_data;
 804         struct file *filp = mds->mds_rcvd_filp;
 805         loff_t off = 0;
 806         int rc;
 807
 808         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
 809         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
 810
 811         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
 812                (unsigned long long)mds->mds_mount_count,
 813                (unsigned long long)mds->mds_last_rcvd);
 814         push_ctxt(&saved, &mds->mds_ctxt);
 815         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
 816         if (rc != sizeof(*msd)) {
 817                 CERROR("error writing MDS server data: rc = %d\n", rc);
 818                 if (rc > 0)
 819                         RETURN(-EIO);
 820                 RETURN(rc);
 821         }
 822         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
 823         pop_ctxt(&saved);
 824         if (rc)
 825                 CERROR("error flushing MDS server data: rc = %d\n", rc);
 826
 827         return 0;
 828 }
 829
 830 /* Do recovery actions for the MDS */
 831 static int mds_recover(struct obd_device *obddev)
 832 {
 833         struct mds_obd *mds = &obddev->u.mds;
 834         int rc;
 835
 836         /* This happens at the end when recovery is complete */
 837         ++mds->mds_mount_count;
 838         rc = mds_update_server_data(mds);
 839
 840         return rc;
 841 }
 842
 843
 844 /* mount the file system (secretly) */
 845 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
 846 {
 847         struct obd_ioctl_data* data = buf;
 848         struct obd_export *export;
 849         struct mds_obd *mds = &obddev->u.mds;
 850         struct vfsmount *mnt;
 851         int rc = 0;
 852         ENTRY;
 853
 854         MOD_INC_USE_COUNT;
 855 #ifdef CONFIG_DEV_RDONLY
 856         dev_clear_rdonly(2);
 857 #endif
 858         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 859                 GOTO(err_dec, rc = -EINVAL);
 860
 861         mds->mds_fstype = strdup(data->ioc_inlbuf2);
 862
 863         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
 864         if (IS_ERR(mnt)) {
 865                 rc = PTR_ERR(mnt);
 866                 CERROR("do_kern_mount failed: rc = %d\n", rc);
 867                 GOTO(err_kfree, rc);
 868         }
 869
 870         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 871         if (!mds->mds_sb)
 872                 GOTO(err_put, rc = -ENODEV);
 873
 874         rc = mds_fs_setup(mds, mnt);
 875         if (rc) {
 876                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
 877                 GOTO(err_put, rc);
 878         }
 879
 880         mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
 881                                            MDC_REPLY_PORTAL, "self",mds_handle);
 882         if (!mds->mds_service) {
 883                 CERROR("failed to start service\n");
 884                 GOTO(err_fs, rc = -EINVAL);
 885         }
 886
 887         rc = ptlrpc_start_thread(obddev, mds->mds_service, "lustre_mds");
 888         if (rc) {
 889                 CERROR("cannot start thread: rc = %d\n", rc);
 890                 GOTO(err_svc, rc);
 891         }
 892
 893         rc = -ENOENT;
 894         mds->mds_ldlm_conn = ptlrpc_uuid_to_connection("self");
 895         if (!mds->mds_ldlm_conn) {
 896                 mds_cleanup(obddev);
 897                 GOTO(err_thread, rc);
 898         }
 899
 900         obddev->obd_namespace =
 901                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
 902         if (obddev->obd_namespace == NULL) {
 903                 LBUG();
 904                 mds_cleanup(obddev);
 905                 GOTO(err_thread, rc);
 906         }
 907
 908         mds->mds_local_namespace =
 909                 ldlm_namespace_new("mds_client", LDLM_NAMESPACE_CLIENT);
 910         if (mds->mds_local_namespace == NULL) {
 911                 LBUG();
 912                 mds_cleanup(obddev);
 913                 GOTO(err_thread, rc);
 914         }
 915
 916         OBD_ALLOC(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
 917         if (mds->mds_ldlm_client == NULL) {
 918                 LBUG();
 919                 mds_cleanup(obddev);
 920                 GOTO(err_thread, rc);
 921         }
 922         ptlrpc_init_client(NULL, NULL, LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
 923                            mds->mds_ldlm_client);
 924         mds->mds_ldlm_client->cli_target_devno = obddev->obd_minor;
 925         mds->mds_ldlm_client->cli_name = "mds ldlm";
 926
 927         rc = mds_recover(obddev);
 928         if (rc)
 929                 GOTO(err_thread, rc);
 930
 931         rc = class_connect(&mds->mds_connh, obddev);
 932         if (rc)
 933                 GOTO(err_thread, rc);
 934         export = class_conn2export(&mds->mds_connh);
 935         if (!export)
 936                 LBUG();
 937         export->export_connection = mds->mds_ldlm_conn;
 938
 939         RETURN(0);
 940
 941
 942
 943 err_thread:
 944         ptlrpc_stop_all_threads(mds->mds_service);
 945 err_svc:
 946         ptlrpc_unregister_service(mds->mds_service);
 947 err_fs:
 948         mds_fs_cleanup(mds);
 949 err_put:
 950         unlock_kernel();
 951         mntput(mds->mds_vfsmnt);
 952         mds->mds_sb = 0;
 953         lock_kernel();
 954 err_kfree:
 955         kfree(mds->mds_fstype);
 956 err_dec:
 957         MOD_DEC_USE_COUNT;
 958         return rc;
 959 }
 960
 961 static int mds_cleanup(struct obd_device * obddev)
 962 {
 963         struct super_block *sb;
 964         struct mds_obd *mds = &obddev->u.mds;
 965
 966         ENTRY;
 967         class_disconnect(&mds->mds_connh);
 968
 969
 970         if ( !list_empty(&obddev->obd_exports) ) {
 971                 CERROR("still has exports!\n");
 972                 RETURN(-EBUSY);
 973         }
 974
 975         ptlrpc_stop_all_threads(mds->mds_service);
 976         ptlrpc_unregister_service(mds->mds_service);
 977
 978         sb = mds->mds_sb;
 979         if (!mds->mds_sb)
 980                 RETURN(0);
 981
 982         mds_update_server_data(mds);
 983
 984         if (mds->mds_rcvd_filp) {
 985                 int rc = filp_close(mds->mds_rcvd_filp, 0);
 986                 mds->mds_rcvd_filp = NULL;
 987
 988                 if (rc)
 989                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
 990         }
 991
 992         unlock_kernel();
 993         mntput(mds->mds_vfsmnt);
 994         mds->mds_sb = 0;
 995         kfree(mds->mds_fstype);
 996
 997         ldlm_namespace_free(mds->mds_local_namespace);
 998         ldlm_namespace_free(obddev->obd_namespace);
 999
1000         if (mds->mds_ldlm_conn != NULL)
1001                 ptlrpc_put_connection(mds->mds_ldlm_conn);
1002
1003         OBD_FREE(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
1004
1005         lock_kernel();
1006 #ifdef CONFIG_DEV_RDONLY
1007         dev_clear_rdonly(2);
1008 #endif
1009         mds_fs_cleanup(mds);
1010
1011         MOD_DEC_USE_COUNT;
1012         RETURN(0);
1013 }
1014
1015 /* use obd ops to offer management infrastructure */
1016 static struct obd_ops mds_obd_ops = {
1017         o_connect:     mds_connect,
1018         o_disconnect:  mds_disconnect,
1019         o_setup:       mds_setup,
1020         o_cleanup:     mds_cleanup,
1021 };
1022
1023 static int __init mds_init(void)
1024 {
1025         inter_module_register("mds_reint", THIS_MODULE, &mds_reint);
1026         inter_module_register("mds_getattr_name", THIS_MODULE,
1027                               &mds_getattr_name);
1028         class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1029         return 0;
1030 }
1031
1032 static void __exit mds_exit(void)
1033 {
1034         inter_module_unregister("mds_reint");
1035         inter_module_unregister("mds_getattr_name");
1036         class_unregister_type(LUSTRE_MDS_NAME);
1037 }
1038
1039 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1040 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1041 MODULE_LICENSE("GPL");
1042
1043 module_init(mds_init);
1044 module_exit(mds_exit);