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 /* Assumes caller has already pushed into the kernel filesystem context */
  31 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  32                         __u64 offset)
  33 {
  34         int rc = 0;
  35         struct mds_obd *mds = &req->rq_obd->u.mds;
  36         struct ptlrpc_bulk_desc *desc;
  37         struct ptlrpc_bulk_page *bulk;
  38         char *buf;
  39         ENTRY;
  40
  41         desc = ptlrpc_prep_bulk(req->rq_connection);
  42         if (desc == NULL)
  43                 GOTO(out, rc = -ENOMEM);
  44
  45         bulk = ptlrpc_prep_bulk_page(desc);
  46         if (bulk == NULL)
  47                 GOTO(cleanup_bulk, rc = -ENOMEM);
  48
  49         OBD_ALLOC(buf, PAGE_SIZE);
  50         if (buf == NULL)
  51                 GOTO(cleanup_bulk, rc = -ENOMEM);
  52
  53         rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
  54
  55         if (rc != PAGE_SIZE)
  56                 GOTO(cleanup_buf, rc = -EIO);
  57
  58         bulk->b_xid = req->rq_xid;
  59         bulk->b_buf = buf;
  60         bulk->b_buflen = PAGE_SIZE;
  61         desc->b_portal = MDS_BULK_PORTAL;
  62
  63         rc = ptlrpc_send_bulk(desc);
  64         if (rc)
  65                 GOTO(cleanup_buf, rc);
  66
  67         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
  68                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
  69                        OBD_FAIL_MDS_SENDPAGE, rc);
  70                 ptlrpc_abort_bulk(desc);
  71                 GOTO(cleanup_buf, rc);
  72         }
  73
  74         wait_event_interruptible(desc->b_waitq, ptlrpc_check_bulk_sent(desc));
  75         if (desc->b_flags & PTL_RPC_FL_INTR)
  76                 GOTO(cleanup_buf, rc = -EINTR);
  77
  78         EXIT;
  79  cleanup_buf:
  80         OBD_FREE(buf, PAGE_SIZE);
  81  cleanup_bulk:
  82         ptlrpc_free_bulk(desc);
  83  out:
  84         return rc;
  85 }
  86
  87 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
  88                               struct vfsmount **mnt)
  89 {
  90         /* stolen from NFS */
  91         struct super_block *sb = mds->mds_sb;
  92         unsigned long ino = fid->id;
  93         __u32 generation = fid->generation;
  94         struct inode *inode;
  95         struct list_head *lp;
  96         struct dentry *result;
  97
  98         if (ino == 0)
  99                 RETURN(ERR_PTR(-ESTALE));
 100
 101         inode = iget(sb, ino);
 102         if (inode == NULL)
 103                 RETURN(ERR_PTR(-ENOMEM));
 104
 105         CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
 106
 107         if (is_bad_inode(inode) ||
 108             (generation && inode->i_generation != generation)) {
 109                 /* we didn't find the right inode.. */
 110                 CERROR("bad inode %lu, link: %d ct: %d or version  %u/%u\n",
 111                        inode->i_ino, inode->i_nlink,
 112                        atomic_read(&inode->i_count), inode->i_generation,
 113                        generation);
 114                 LBUG();
 115                 iput(inode);
 116                 RETURN(ERR_PTR(-ESTALE));
 117         }
 118
 119         /* now to find a dentry.
 120          * If possible, get a well-connected one
 121          */
 122         if (mnt)
 123                 *mnt = mds->mds_vfsmnt;
 124         spin_lock(&dcache_lock);
 125         for (lp = inode->i_dentry.next; lp != &inode->i_dentry ; lp=lp->next) {
 126                 result = list_entry(lp,struct dentry, d_alias);
 127                 if (! (result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
 128                         dget_locked(result);
 129                         result->d_vfs_flags |= DCACHE_REFERENCED;
 130                         spin_unlock(&dcache_lock);
 131                         iput(inode);
 132                         if (mnt)
 133                                 mntget(*mnt);
 134                         return result;
 135                 }
 136         }
 137         spin_unlock(&dcache_lock);
 138         result = d_alloc_root(inode);
 139         if (result == NULL) {
 140                 iput(inode);
 141                 return ERR_PTR(-ENOMEM);
 142         }
 143         if (mnt)
 144                 mntget(*mnt);
 145         result->d_flags |= DCACHE_NFSD_DISCONNECTED;
 146         return result;
 147 }
 148
 149 static int mds_connect(struct ptlrpc_request *req, struct mds_obd **mdsp)
 150 {
 151         struct mds_obd *mds;
 152         char *uuid;
 153         int rc, i;
 154         ENTRY;
 155
 156         uuid = lustre_msg_buf(req->rq_reqmsg, 0);
 157         if (req->rq_reqmsg->buflens[0] > 37) {
 158                 /* Invalid UUID */
 159                 req->rq_status = -EINVAL;
 160                 RETURN(-EINVAL);
 161         }
 162
 163         i = obd_class_name2dev(uuid);
 164         if (i == -1) {
 165                 req->rq_status = -ENODEV;
 166                 RETURN(-ENODEV);
 167         }
 168
 169         *mdsp = mds = &(obd_dev[i].u.mds);
 170         if (mds != &(req->rq_obd->u.mds)) {
 171                 CERROR("device %d (%s) is not an mds\n", i, uuid);
 172                 req->rq_status = -ENODEV;
 173                 RETURN(-ENODEV);
 174         }
 175
 176         CDEBUG(D_INFO, "MDS connect from UUID '%s'\n",
 177                ptlrpc_req_to_uuid(req));
 178         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 179         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CONNECT_PACK)) {
 180                 req->rq_status = -ENOMEM;
 181                 RETURN(-ENOMEM);
 182         }
 183         req->rq_repmsg->target_id = i;
 184
 185         RETURN(0);
 186 }
 187
 188 /* FIXME: the error cases need fixing to avoid leaks */
 189 static int mds_getstatus(struct mds_obd *mds, struct ptlrpc_request *req)
 190 {
 191         struct mds_body *body;
 192         struct mds_client_info *mci;
 193         struct mds_client_data *mcd;
 194         int rc, size = sizeof(*body);
 195         ENTRY;
 196
 197         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 198         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 199                 CERROR("mds: out of memory for message: size=%d\n", size);
 200                 req->rq_status = -ENOMEM;
 201                 RETURN(0);
 202         }
 203
 204         body = lustre_msg_buf(req->rq_reqmsg, 0);
 205         mds_unpack_body(body);
 206         /* Anything we need to do here with the client's trans no or so? */
 207
 208         body = lustre_msg_buf(req->rq_repmsg, 0);
 209         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 210
 211         mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
 212         if (!mci) {
 213                 /* We don't have any old connection data for this client */
 214                 int rc;
 215
 216                 CDEBUG(D_INFO, "allocating new client data for UUID '%s'",
 217                        ptlrpc_req_to_uuid(req));
 218
 219                 OBD_ALLOC(mcd, sizeof(*mcd));
 220                 if (!mcd) {
 221                         CERROR("mds: out of memory for client data\n");
 222                         req->rq_status = -ENOMEM;
 223                         RETURN(0);
 224                 }
 225                 memcpy(mcd->mcd_uuid, ptlrpc_req_to_uuid(req),
 226                        sizeof(mcd->mcd_uuid));
 227                 rc = mds_client_add(mds, mcd, -1);
 228                 if (rc) {
 229                         req->rq_status = rc;
 230                         RETURN(0);
 231                 }
 232         } else {
 233                 /* We have old connection data for this client... */
 234                 mcd = mci->mci_mcd;
 235                 CDEBUG(D_INFO, "found existing data for UUID '%s' at #%d\n",
 236                        mcd->mcd_uuid, mci->mci_off);
 237         }
 238         /* mcd_last_xid is is stored in little endian on the disk and
 239            mds_pack_rep_body converts it to network order */
 240         body->last_xid = le32_to_cpu(mcd->mcd_last_xid);
 241         mds_pack_rep_body(req);
 242         RETURN(0);
 243 }
 244
 245 static int mds_disconnect(struct mds_obd *mds, struct ptlrpc_request *req)
 246 {
 247         struct mds_body *body;
 248         int rc;
 249         ENTRY;
 250
 251         body = lustre_msg_buf(req->rq_reqmsg, 0);
 252
 253         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 254         if (rc)
 255                 RETURN(rc);
 256
 257         RETURN(0);
 258 }
 259
 260 int mds_lock_callback(struct lustre_handle *lockh, struct ldlm_lock_desc *desc,
 261                       void *data, int data_len, struct ptlrpc_request **reqp)
 262 {
 263         ENTRY;
 264
 265         if (desc == NULL) {
 266                 /* Completion AST.  Do nothing */
 267                 RETURN(0);
 268         }
 269
 270         if (ldlm_cli_cancel(lockh) < 0)
 271                 LBUG();
 272         RETURN(0);
 273 }
 274
 275 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 276 {
 277         struct mds_obd *mds = &req->rq_obd->u.mds;
 278         struct obd_run_ctxt saved;
 279         struct mds_body *body;
 280         struct dentry *de = NULL, *dchild = NULL;
 281         struct inode *dir;
 282         struct lustre_handle lockh;
 283         char *name;
 284         int namelen, flags, lock_mode, rc = 0;
 285         __u64 res_id[3] = {0, 0, 0};
 286         ENTRY;
 287
 288         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 289                 LBUG();
 290
 291         if (req->rq_reqmsg->bufcount <= offset + 1) {
 292                 LBUG();
 293                 GOTO(out_pre_de, rc = -EINVAL);
 294         }
 295
 296         body = lustre_msg_buf(req->rq_reqmsg, offset);
 297         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 298         namelen = req->rq_reqmsg->buflens[offset + 1];
 299         /* requests were at offset 2, replies go back at 1 */
 300         if (offset)
 301                 offset = 1;
 302
 303         push_ctxt(&saved, &mds->mds_ctxt);
 304         de = mds_fid2dentry(mds, &body->fid1, NULL);
 305         if (IS_ERR(de)) {
 306                 LBUG();
 307                 GOTO(out_pre_de, rc = -ESTALE);
 308         }
 309
 310         dir = de->d_inode;
 311         CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
 312
 313         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 314         res_id[0] = dir->i_ino;
 315
 316         rc = ldlm_lock_match(mds->mds_local_namespace, res_id, LDLM_PLAIN,
 317                              NULL, 0, lock_mode, &lockh);
 318         if (rc == 0) {
 319                 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
 320                 rc = ldlm_cli_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
 321                                       NULL, mds->mds_local_namespace, NULL,
 322                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 323                                       &flags, (void *)mds_lock_callback,
 324                                       NULL, 0, &lockh);
 325                 if (rc != ELDLM_OK) {
 326                         CERROR("lock enqueue: err: %d\n", rc);
 327                         GOTO(out_create_de, rc = -EIO);
 328                 }
 329         }
 330         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 331
 332         down(&dir->i_sem);
 333         dchild = lookup_one_len(name, de, namelen - 1);
 334         if (IS_ERR(dchild)) {
 335                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 336                 up(&dir->i_sem);
 337                 LBUG();
 338                 GOTO(out_create_dchild, rc = -ESTALE);
 339         }
 340
 341         if (dchild->d_inode) {
 342                 struct mds_body *body;
 343                 struct obdo *obdo;
 344                 struct inode *inode = dchild->d_inode;
 345                 CERROR("child exists (dir %ld, name %s, ino %ld)\n",
 346                        dir->i_ino, name, dchild->d_inode->i_ino);
 347
 348                 body = lustre_msg_buf(req->rq_repmsg, offset);
 349                 mds_pack_inode2fid(&body->fid1, inode);
 350                 mds_pack_inode2body(body, inode);
 351                 if (S_ISREG(inode->i_mode)) {
 352                         obdo = lustre_msg_buf(req->rq_repmsg, offset + 1);
 353                         mds_fs_get_obdo(mds, inode, obdo);
 354                 }
 355                 /* now a normal case for intent locking */
 356                 rc = 0;
 357         } else {
 358                 rc = -ENOENT;
 359         }
 360
 361         EXIT;
 362 out_create_dchild:
 363         l_dput(dchild);
 364         up(&dir->i_sem);
 365         ldlm_lock_decref(&lockh, lock_mode);
 366 out_create_de:
 367         l_dput(de);
 368 out_pre_de:
 369         req->rq_status = rc;
 370         pop_ctxt(&saved);
 371         return 0;
 372 }
 373
 374
 375 static int mds_getattr(int offset, struct ptlrpc_request *req)
 376 {
 377         struct mds_obd *mds = &req->rq_obd->u.mds;
 378         struct obd_run_ctxt saved;
 379         struct dentry *de;
 380         struct inode *inode;
 381         struct mds_body *body;
 382         int rc, size[2] = {sizeof(*body)}, bufcount = 1;
 383         ENTRY;
 384
 385         body = lustre_msg_buf(req->rq_reqmsg, offset);
 386         push_ctxt(&saved, &mds->mds_ctxt);
 387         de = mds_fid2dentry(mds, &body->fid1, NULL);
 388         if (IS_ERR(de)) {
 389                 GOTO(out_pop, rc = -ENOENT);
 390         }
 391
 392         inode = de->d_inode;
 393         if (S_ISREG(body->fid1.f_type)) {
 394                 bufcount = 2;
 395                 size[1] = sizeof(struct obdo);
 396         } else if (body->valid & OBD_MD_LINKNAME) {
 397                 bufcount = 2;
 398                 size[1] = inode->i_size;
 399         }
 400
 401         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 402                              &req->rq_repmsg);
 403         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 404                 CERROR("mds: out of memory\n");
 405                 GOTO(out, rc);
 406         }
 407
 408         if (body->valid & OBD_MD_LINKNAME) {
 409                 char *tmp = lustre_msg_buf(req->rq_repmsg, 1);
 410
 411                 rc = inode->i_op->readlink(de, tmp, size[1]);
 412
 413                 if (rc < 0) {
 414                         CERROR("readlink failed: %d\n", rc);
 415                         GOTO(out, rc);
 416                 }
 417         }
 418
 419         body = lustre_msg_buf(req->rq_repmsg, 0);
 420         body->ino = inode->i_ino;
 421         body->generation = inode->i_generation;
 422         body->atime = inode->i_atime;
 423         body->ctime = inode->i_ctime;
 424         body->mtime = inode->i_mtime;
 425         body->uid = inode->i_uid;
 426         body->gid = inode->i_gid;
 427         body->size = inode->i_size;
 428         body->mode = inode->i_mode;
 429         body->nlink = inode->i_nlink;
 430         body->valid = ~0; /* FIXME: should be more selective */
 431
 432         if (S_ISREG(inode->i_mode)) {
 433                 rc = mds_fs_get_obdo(mds, inode,
 434                                      lustre_msg_buf(req->rq_repmsg, 1));
 435                 if (rc < 0) {
 436                         CERROR("mds_fs_get_obdo failed: %d\n", rc);
 437                         GOTO(out, rc);
 438                 }
 439         }
 440 out:
 441         l_dput(de);
 442 out_pop:
 443         pop_ctxt(&saved);
 444         req->rq_status = rc;
 445         RETURN(0);
 446 }
 447
 448 static int mds_statfs(struct ptlrpc_request *req)
 449 {
 450         struct mds_obd *mds = &req->rq_obd->u.mds;
 451         struct obd_statfs *osfs;
 452         struct statfs sfs;
 453         int rc, size = sizeof(*osfs);
 454         ENTRY;
 455
 456         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 457                              &req->rq_repmsg);
 458         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 459                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 460                 GOTO(out, rc);
 461         }
 462
 463         rc = vfs_statfs(mds->mds_sb, &sfs);
 464         if (rc) {
 465                 CERROR("mds: statfs failed: rc %d\n", rc);
 466                 GOTO(out, rc);
 467         }
 468         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 469         memset(osfs, 0, size);
 470         /* FIXME: hack until the MDS gets the OST data, next month */
 471         sfs.f_blocks = -1;
 472         sfs.f_bfree = -1;
 473         sfs.f_bavail = -1;
 474         obd_statfs_pack(osfs, &sfs);
 475
 476 out:
 477         req->rq_status = rc;
 478         RETURN(0);
 479 }
 480
 481 static int mds_open(struct ptlrpc_request *req)
 482 {
 483         struct dentry *de;
 484         struct mds_body *body;
 485         struct file *file;
 486         struct vfsmount *mnt;
 487         struct mds_obd *mds = &req->rq_obd->u.mds;
 488         struct mds_client_info *mci;
 489         __u32 flags;
 490         struct list_head *tmp;
 491         struct mds_file_data *mfd;
 492         int rc, size = sizeof(*body);
 493         ENTRY;
 494
 495         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 496         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 497                 CERROR("mds: out of memory\n");
 498                 req->rq_status = -ENOMEM;
 499                 RETURN(0);
 500         }
 501
 502         mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
 503         if (!mci) {
 504                 CERROR("mds: no mci!\n");
 505                 req->rq_status = -ENOTCONN;
 506                 RETURN(0);
 507         }
 508
 509         body = lustre_msg_buf(req->rq_reqmsg, 0);
 510
 511         /* was this animal open already? */
 512         /* XXX we should only check on re-open, or do a refcount... */
 513         list_for_each(tmp, &mci->mci_open_head) {
 514                 struct mds_file_data *fd;
 515                 fd = list_entry(tmp, struct mds_file_data, mfd_list);
 516                 if (body->extra == fd->mfd_clientfd &&
 517                     body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
 518                         CERROR("Re opening %Ld\n", body->fid1.id);
 519                         RETURN(0);
 520                 }
 521         }
 522
 523         OBD_ALLOC(mfd, sizeof(*mfd));
 524         if (!mfd) {
 525                 CERROR("mds: out of memory\n");
 526                 req->rq_status = -ENOMEM;
 527                 RETURN(0);
 528         }
 529
 530         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 531         if (IS_ERR(de)) {
 532                 req->rq_status = -ENOENT;
 533                 RETURN(0);
 534         }
 535
 536         /* check if this inode has seen a delayed object creation */
 537         if (req->rq_reqmsg->bufcount > 1) {
 538                 void *handle;
 539                 struct inode *inode = de->d_inode;
 540                 //struct iattr iattr;
 541                 struct obdo *obdo;
 542                 int rc;
 543
 544                 obdo = lustre_msg_buf(req->rq_reqmsg, 1);
 545                 //iattr.ia_valid = ATTR_MODE;
 546                 //iattr.ia_mode = inode->i_mode;
 547
 548                 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 549                 if (!handle) {
 550                         req->rq_status = -ENOMEM;
 551                         RETURN(0);
 552                 }
 553
 554                 /* XXX error handling */
 555                 rc = mds_fs_set_obdo(mds, inode, handle, obdo);
 556                 //                rc = mds_fs_setattr(mds, de, handle, &iattr);
 557                 if (!rc)
 558                         rc = mds_update_last_rcvd(mds, handle, req);
 559                 else {
 560                         req->rq_status = rc;
 561                         RETURN(0);
 562                 }
 563                 /* FIXME: need to return last_rcvd, last_committed */
 564
 565                 /* FIXME: keep rc intact */
 566                 rc = mds_fs_commit(mds, de->d_inode, handle);
 567                 if (rc) {
 568                         req->rq_status = rc;
 569                         RETURN(0);
 570                 }
 571         }
 572
 573         flags = body->flags;
 574         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 575         if (!file || IS_ERR(file)) {
 576                 req->rq_status = -EINVAL;
 577                 OBD_FREE(mfd, sizeof(*mfd));
 578                 RETURN(0);
 579         }
 580
 581         file->private_data = mfd;
 582         mfd->mfd_file = file;
 583         mfd->mfd_clientfd = body->extra;
 584         list_add(&mfd->mfd_list, &mci->mci_open_head);
 585
 586         body = lustre_msg_buf(req->rq_repmsg, 0);
 587         body->extra = (__u64) (unsigned long)file;
 588         RETURN(0);
 589 }
 590
 591 static int mds_close(struct ptlrpc_request *req)
 592 {
 593         struct dentry *de;
 594         struct mds_body *body;
 595         struct file *file;
 596         struct mds_obd *mds = &req->rq_obd->u.mds;
 597         struct vfsmount *mnt;
 598         struct mds_file_data *mfd;
 599         int rc;
 600         ENTRY;
 601
 602         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 603         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 604                 CERROR("mds: out of memory\n");
 605                 req->rq_status = -ENOMEM;
 606                 RETURN(0);
 607         }
 608
 609         body = lustre_msg_buf(req->rq_reqmsg, 0);
 610         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 611         if (IS_ERR(de)) {
 612                 req->rq_status = -ENOENT;
 613                 RETURN(0);
 614         }
 615
 616         file = (struct file *)(unsigned long)body->extra;
 617         if (!file->f_dentry)
 618                 LBUG();
 619         mfd = (struct mds_file_data *)file->private_data;
 620         list_del(&mfd->mfd_list);
 621         OBD_FREE(mfd, sizeof(*mfd));
 622
 623         req->rq_status = filp_close(file, 0);
 624         l_dput(de);
 625         mntput(mnt);
 626
 627         RETURN(0);
 628 }
 629
 630 static int mds_readpage(struct mds_obd *mds, struct ptlrpc_request *req)
 631 {
 632         struct vfsmount *mnt;
 633         struct dentry *de;
 634         struct file *file;
 635         struct mds_body *body;
 636         struct obd_run_ctxt saved;
 637         int rc, size = sizeof(*body);
 638         ENTRY;
 639
 640         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 641         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 642                 CERROR("mds: out of memory\n");
 643                 GOTO(out, rc = -ENOMEM);
 644         }
 645
 646         body = lustre_msg_buf(req->rq_reqmsg, 0);
 647         push_ctxt(&saved, &mds->mds_ctxt);
 648         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 649         if (IS_ERR(de))
 650                 GOTO(out_pop, rc = PTR_ERR(de));
 651
 652         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 653
 654         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 655         /* note: in case of an error, dentry_open puts dentry */
 656         if (IS_ERR(file))
 657                 GOTO(out_pop, rc = PTR_ERR(file));
 658
 659         /* to make this asynchronous make sure that the handling function
 660            doesn't send a reply when this function completes. Instead a
 661            callback function would send the reply */
 662         rc = mds_sendpage(req, file, body->size);
 663
 664         filp_close(file, 0);
 665 out_pop:
 666         pop_ctxt(&saved);
 667 out:
 668         req->rq_status = rc;
 669         RETURN(0);
 670 }
 671
 672 int mds_reint(int offset, struct ptlrpc_request *req)
 673 {
 674         int rc;
 675         struct mds_update_record rec;
 676
 677         rc = mds_update_unpack(req, offset, &rec);
 678         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 679                 CERROR("invalid record\n");
 680                 req->rq_status = -EINVAL;
 681                 RETURN(0);
 682         }
 683         /* rc will be used to interrupt a for loop over multiple records */
 684         rc = mds_reint_rec(&rec, offset, req);
 685         return rc;
 686 }
 687
 688 int mds_handle(struct obd_device *dev, struct ptlrpc_service *svc,
 689                struct ptlrpc_request *req)
 690 {
 691         struct mds_obd *mds = NULL;
 692         int rc;
 693         ENTRY;
 694
 695         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 696         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
 697                 CERROR("lustre_mds: Invalid request\n");
 698                 GOTO(out, rc);
 699         }
 700
 701         if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 702                 CERROR("lustre_mds: wrong packet type sent %d\n",
 703                        req->rq_reqmsg->type);
 704                 GOTO(out, rc = -EINVAL);
 705         }
 706
 707         if (req->rq_reqmsg->opc != MDS_CONNECT) {
 708                 int id = req->rq_reqmsg->target_id;
 709                 struct obd_device *obddev;
 710                 if (id < 0 || id > MAX_OBD_DEVICES)
 711                         GOTO(out, rc = -ENODEV);
 712                 obddev = &obd_dev[id];
 713                 if (strcmp(obddev->obd_type->typ_name, "mds") != 0)
 714                         GOTO(out, rc = -EINVAL);
 715                 mds = &obddev->u.mds;
 716                 req->rq_obd = obddev;
 717         }
 718
 719         switch (req->rq_reqmsg->opc) {
 720         case MDS_CONNECT:
 721                 CDEBUG(D_INODE, "connect\n");
 722                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
 723                 rc = mds_connect(req, &mds);
 724                 break;
 725
 726         case MDS_DISCONNECT:
 727                 CDEBUG(D_INODE, "disconnect\n");
 728                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
 729                 rc = mds_disconnect(mds, req);
 730                 break;
 731
 732         case MDS_GETSTATUS:
 733                 CDEBUG(D_INODE, "getstatus\n");
 734                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
 735                 rc = mds_getstatus(mds, req);
 736                 break;
 737
 738         case MDS_GETATTR:
 739                 CDEBUG(D_INODE, "getattr\n");
 740                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 741                 rc = mds_getattr(0, req);
 742                 break;
 743
 744         case MDS_STATFS:
 745                 CDEBUG(D_INODE, "statfs\n");
 746                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
 747                 rc = mds_statfs(req);
 748                 break;
 749
 750         case MDS_READPAGE:
 751                 CDEBUG(D_INODE, "readpage\n");
 752                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 753                 rc = mds_readpage(mds, req);
 754
 755                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 756                         return 0;
 757                 break;
 758
 759         case MDS_REINT: {
 760                 int size = sizeof(struct mds_body);
 761                 CDEBUG(D_INODE, "reint\n");
 762                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 763
 764                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 765                                      &req->rq_repmsg);
 766                 if (rc) {
 767                         rc = req->rq_status = -ENOMEM;
 768                         break;
 769                 }
 770                 rc = mds_reint(0, req);
 771                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
 772                 break;
 773         }
 774
 775         case MDS_OPEN:
 776                 CDEBUG(D_INODE, "open\n");
 777                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 778                 rc = mds_open(req);
 779                 break;
 780
 781         case MDS_CLOSE:
 782                 CDEBUG(D_INODE, "close\n");
 783                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 784                 rc = mds_close(req);
 785                 break;
 786
 787         default:
 788                 rc = ptlrpc_error(svc, req);
 789                 RETURN(rc);
 790         }
 791
 792         EXIT;
 793 out:
 794         /* Still not 100% sure whether we should reply with the server
 795          * last_rcvd or that of this client.  I'm not sure it even makes
 796          * a difference on a per-client basis, because last_rcvd is global
 797          * and we are not supposed to allow transactions while in recovery.
 798          */
 799         if (rc) {
 800                 ptlrpc_error(svc, req);
 801         } else {
 802                 req->rq_repmsg->last_rcvd = HTON__u64(mds->mds_last_rcvd);
 803                 req->rq_repmsg->last_committed =
 804                         HTON__u64(mds->mds_last_committed);
 805                 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
 806                        (unsigned long long)mds->mds_last_rcvd,
 807                        (unsigned long long)mds->mds_last_committed,
 808                        cpu_to_le32(req->rq_xid));
 809                 CDEBUG(D_NET, "sending reply\n");
 810                 ptlrpc_reply(svc, req);
 811         }
 812         return 0;
 813 }
 814
 815 /* Update the server data on disk.  This stores the new mount_count and
 816  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
 817  * then the server last_rcvd value may be less than that of the clients.
 818  * This will alert us that we may need to do client recovery.
 819  */
 820 static
 821 int mds_update_server_data(struct mds_obd *mds)
 822 {
 823         struct obd_run_ctxt saved;
 824         struct mds_server_data *msd = mds->mds_server_data;
 825         struct file *filp = mds->mds_rcvd_filp;
 826         loff_t off = 0;
 827         int rc;
 828
 829         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
 830         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
 831
 832         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
 833                (unsigned long long)mds->mds_mount_count,
 834                (unsigned long long)mds->mds_last_rcvd);
 835         push_ctxt(&saved, &mds->mds_ctxt);
 836         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
 837         if (rc != sizeof(*msd)) {
 838                 CERROR("error writing MDS server data: rc = %d\n", rc);
 839                 if (rc > 0)
 840                         RETURN(-EIO);
 841                 RETURN(rc);
 842         }
 843         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
 844         pop_ctxt(&saved);
 845         if (rc)
 846                 CERROR("error flushing MDS server data: rc = %d\n", rc);
 847
 848         return 0;
 849 }
 850
 851 /* Do recovery actions for the MDS */
 852 static int mds_recover(struct obd_device *obddev)
 853 {
 854         struct mds_obd *mds = &obddev->u.mds;
 855         int rc;
 856
 857         /* This happens at the end when recovery is complete */
 858         ++mds->mds_mount_count;
 859         rc = mds_update_server_data(mds);
 860
 861         return rc;
 862 }
 863
 864
 865 /* mount the file system (secretly) */
 866 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
 867 {
 868         struct obd_ioctl_data* data = buf;
 869         struct mds_obd *mds = &obddev->u.mds;
 870         struct vfsmount *mnt;
 871         int rc = 0;
 872         ENTRY;
 873
 874         MOD_INC_USE_COUNT;
 875 #ifdef CONFIG_DEV_RDONLY
 876         dev_clear_rdonly(2);
 877 #endif
 878         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 879                 GOTO(err_dec, rc = -EINVAL);
 880
 881         mds->mds_fstype = strdup(data->ioc_inlbuf2);
 882
 883         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
 884         if (IS_ERR(mnt)) {
 885                 rc = PTR_ERR(mnt);
 886                 CERROR("do_kern_mount failed: rc = %d\n", rc);
 887                 GOTO(err_kfree, rc);
 888         }
 889
 890         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 891         if (!mds->mds_sb)
 892                 GOTO(err_put, rc = -ENODEV);
 893
 894         rc = mds_fs_setup(mds, mnt);
 895         if (rc) {
 896                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
 897                 GOTO(err_put, rc);
 898         }
 899
 900         mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
 901                                            MDC_REPLY_PORTAL, "self",mds_handle);
 902         if (!mds->mds_service) {
 903                 CERROR("failed to start service\n");
 904                 GOTO(err_fs, rc = -EINVAL);
 905         }
 906
 907         rc = ptlrpc_start_thread(obddev, mds->mds_service, "lustre_mds");
 908         if (rc) {
 909                 CERROR("cannot start thread: rc = %d\n", rc);
 910                 GOTO(err_svc, rc);
 911         }
 912
 913         rc = -ENOENT;
 914         mds->mds_ldlm_conn = ptlrpc_uuid_to_connection("self");
 915         if (!mds->mds_ldlm_conn) {
 916                 mds_cleanup(obddev);
 917                 GOTO(err_thread, rc);
 918         }
 919
 920         obddev->obd_namespace =
 921                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
 922         if (obddev->obd_namespace == NULL) {
 923                 LBUG();
 924                 mds_cleanup(obddev);
 925                 GOTO(err_thread, rc);
 926         }
 927
 928         mds->mds_local_namespace =
 929                 ldlm_namespace_new("mds_client", LDLM_NAMESPACE_CLIENT);
 930         if (mds->mds_local_namespace == NULL) {
 931                 LBUG();
 932                 mds_cleanup(obddev);
 933                 GOTO(err_thread, rc);
 934         }
 935
 936         OBD_ALLOC(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
 937         if (mds->mds_ldlm_client == NULL) {
 938                 LBUG();
 939                 mds_cleanup(obddev);
 940                 GOTO(err_thread, rc);
 941         }
 942         ptlrpc_init_client(NULL, NULL, LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
 943                            mds->mds_ldlm_client);
 944         mds->mds_ldlm_client->cli_target_devno = obddev->obd_minor;
 945         mds->mds_ldlm_client->cli_name = "mds ldlm";
 946
 947         rc = mds_recover(obddev);
 948         if (rc)
 949                 GOTO(err_thread, rc);
 950
 951         RETURN(0);
 952
 953 err_thread:
 954         ptlrpc_stop_all_threads(mds->mds_service);
 955 err_svc:
 956         ptlrpc_unregister_service(mds->mds_service);
 957 err_fs:
 958         mds_fs_cleanup(mds);
 959 err_put:
 960         unlock_kernel();
 961         mntput(mds->mds_vfsmnt);
 962         mds->mds_sb = 0;
 963         lock_kernel();
 964 err_kfree:
 965         kfree(mds->mds_fstype);
 966 err_dec:
 967         MOD_DEC_USE_COUNT;
 968         return rc;
 969 }
 970
 971 static int mds_cleanup(struct obd_device * obddev)
 972 {
 973         struct super_block *sb;
 974         struct mds_obd *mds = &obddev->u.mds;
 975
 976         ENTRY;
 977
 978         if ( !list_empty(&obddev->obd_gen_clients) ) {
 979                 CERROR("still has clients!\n");
 980                 RETURN(-EBUSY);
 981         }
 982
 983         ptlrpc_stop_all_threads(mds->mds_service);
 984         ptlrpc_unregister_service(mds->mds_service);
 985
 986         sb = mds->mds_sb;
 987         if (!mds->mds_sb)
 988                 RETURN(0);
 989
 990         mds_update_server_data(mds);
 991
 992         if (mds->mds_rcvd_filp) {
 993                 int rc = filp_close(mds->mds_rcvd_filp, 0);
 994                 mds->mds_rcvd_filp = NULL;
 995
 996                 if (rc)
 997                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
 998         }
 999
1000         unlock_kernel();
1001         mntput(mds->mds_vfsmnt);
1002         mds->mds_sb = 0;
1003         kfree(mds->mds_fstype);
1004
1005         ldlm_namespace_free(mds->mds_local_namespace);
1006         ldlm_namespace_free(obddev->obd_namespace);
1007
1008         if (mds->mds_ldlm_conn != NULL)
1009                 ptlrpc_put_connection(mds->mds_ldlm_conn);
1010
1011         OBD_FREE(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
1012
1013         lock_kernel();
1014 #ifdef CONFIG_DEV_RDONLY
1015         dev_clear_rdonly(2);
1016 #endif
1017         mds_fs_cleanup(mds);
1018
1019         MOD_DEC_USE_COUNT;
1020         RETURN(0);
1021 }
1022
1023 /* use obd ops to offer management infrastructure */
1024 static struct obd_ops mds_obd_ops = {
1025         o_setup:       mds_setup,
1026         o_cleanup:     mds_cleanup,
1027 };
1028
1029 static int __init mds_init(void)
1030 {
1031         inter_module_register("mds_reint", THIS_MODULE, &mds_reint);
1032         inter_module_register("mds_getattr_name", THIS_MODULE,
1033                               &mds_getattr_name);
1034         obd_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1035         return 0;
1036 }
1037
1038 static void __exit mds_exit(void)
1039 {
1040         inter_module_unregister("mds_reint");
1041         inter_module_unregister("mds_getattr_name");
1042         obd_unregister_type(LUSTRE_MDS_NAME);
1043 }
1044
1045 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1046 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1047 MODULE_LICENSE("GPL");
1048
1049 module_init(mds_init);
1050 module_exit(mds_exit);