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 ldlm_lock *lock, struct ldlm_lock *new,
 261                       void *data, int data_len, struct ptlrpc_request **reqp)
 262 {
 263         ENTRY;
 264
 265         if (new == NULL) {
 266                 /* Completion AST.  Do nothing */
 267                 RETURN(0);
 268         }
 269
 270         if (ldlm_cli_cancel(lock->l_client, lock) < 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 ldlm_lock *lock;
 283         struct lustre_handle lockh;
 284         char *name;
 285         int namelen, flags, lock_mode, rc = 0;
 286         __u64 res_id[3] = {0, 0, 0};
 287         ENTRY;
 288
 289         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 290                 LBUG();
 291
 292         if (req->rq_reqmsg->bufcount <= offset + 1) {
 293                 LBUG();
 294                 GOTO(out_pre_de, rc = -EINVAL);
 295         }
 296
 297         body = lustre_msg_buf(req->rq_reqmsg, offset);
 298         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 299         namelen = req->rq_reqmsg->buflens[offset + 1];
 300         /* requests were at offset 2, replies go back at 1 */
 301         if (offset)
 302                 offset = 1;
 303
 304         push_ctxt(&saved, &mds->mds_ctxt);
 305         de = mds_fid2dentry(mds, &body->fid1, NULL);
 306         if (IS_ERR(de)) {
 307                 LBUG();
 308                 GOTO(out_pre_de, rc = -ESTALE);
 309         }
 310
 311         dir = de->d_inode;
 312         CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
 313
 314         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 315         res_id[0] = dir->i_ino;
 316
 317         rc = ldlm_local_lock_match(mds->mds_local_namespace, res_id, LDLM_PLAIN,
 318                                    NULL, 0, lock_mode, &lockh);
 319         if (rc == 0) {
 320                 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
 321                 rc = ldlm_cli_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
 322                                       NULL, mds->mds_local_namespace, NULL,
 323                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 324                                       &flags, (void *)mds_lock_callback,
 325                                       NULL, 0, &lockh);
 326                 if (rc != ELDLM_OK) {
 327                         CERROR("lock enqueue: err: %d\n", rc);
 328                         GOTO(out_create_de, rc = -EIO);
 329                 }
 330         } else {
 331                 lock = lustre_handle2object(&lockh);
 332                 LDLM_DEBUG(lock, "matched");
 333         }
 334         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 335
 336         down(&dir->i_sem);
 337         dchild = lookup_one_len(name, de, namelen - 1);
 338         if (IS_ERR(dchild)) {
 339                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 340                 up(&dir->i_sem);
 341                 LBUG();
 342                 GOTO(out_create_dchild, rc = -ESTALE);
 343         }
 344
 345         if (dchild->d_inode) {
 346                 struct mds_body *body;
 347                 struct obdo *obdo;
 348                 struct inode *inode = dchild->d_inode;
 349                 CERROR("child exists (dir %ld, name %s, ino %ld)\n",
 350                        dir->i_ino, name, dchild->d_inode->i_ino);
 351
 352                 body = lustre_msg_buf(req->rq_repmsg, offset);
 353                 mds_pack_inode2fid(&body->fid1, inode);
 354                 mds_pack_inode2body(body, inode);
 355                 if (S_ISREG(inode->i_mode)) {
 356                         obdo = lustre_msg_buf(req->rq_repmsg, offset + 1);
 357                         mds_fs_get_obdo(mds, inode, obdo);
 358                 }
 359                 /* now a normal case for intent locking */
 360                 rc = 0;
 361         } else {
 362                 rc = -ENOENT;
 363         }
 364
 365         EXIT;
 366 out_create_dchild:
 367         l_dput(dchild);
 368         up(&dir->i_sem);
 369         lock = lustre_handle2object(&lockh);
 370         ldlm_lock_decref(lock, lock_mode);
 371 out_create_de:
 372         l_dput(de);
 373 out_pre_de:
 374         req->rq_status = rc;
 375         pop_ctxt(&saved);
 376         return 0;
 377 }
 378
 379
 380 static int mds_getattr(int offset, struct ptlrpc_request *req)
 381 {
 382         struct mds_obd *mds = &req->rq_obd->u.mds;
 383         struct obd_run_ctxt saved;
 384         struct dentry *de;
 385         struct inode *inode;
 386         struct mds_body *body;
 387         int rc, size[2] = {sizeof(*body)}, bufcount = 1;
 388         ENTRY;
 389
 390         body = lustre_msg_buf(req->rq_reqmsg, offset);
 391         push_ctxt(&saved, &mds->mds_ctxt);
 392         de = mds_fid2dentry(mds, &body->fid1, NULL);
 393         if (IS_ERR(de)) {
 394                 GOTO(out_pop, rc = -ENOENT);
 395         }
 396
 397         inode = de->d_inode;
 398         if (S_ISREG(body->fid1.f_type)) {
 399                 bufcount = 2;
 400                 size[1] = sizeof(struct obdo);
 401         } else if (body->valid & OBD_MD_LINKNAME) {
 402                 bufcount = 2;
 403                 size[1] = inode->i_size;
 404         }
 405
 406         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 407                              &req->rq_repmsg);
 408         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 409                 CERROR("mds: out of memory\n");
 410                 GOTO(out, rc);
 411         }
 412
 413         if (body->valid & OBD_MD_LINKNAME) {
 414                 char *tmp = lustre_msg_buf(req->rq_repmsg, 1);
 415
 416                 rc = inode->i_op->readlink(de, tmp, size[1]);
 417
 418                 if (rc < 0) {
 419                         CERROR("readlink failed: %d\n", rc);
 420                         GOTO(out, rc);
 421                 }
 422         }
 423
 424         body = lustre_msg_buf(req->rq_repmsg, 0);
 425         body->ino = inode->i_ino;
 426         body->generation = inode->i_generation;
 427         body->atime = inode->i_atime;
 428         body->ctime = inode->i_ctime;
 429         body->mtime = inode->i_mtime;
 430         body->uid = inode->i_uid;
 431         body->gid = inode->i_gid;
 432         body->size = inode->i_size;
 433         body->mode = inode->i_mode;
 434         body->nlink = inode->i_nlink;
 435         body->valid = ~0; /* FIXME: should be more selective */
 436
 437         if (S_ISREG(inode->i_mode)) {
 438                 rc = mds_fs_get_obdo(mds, inode,
 439                                      lustre_msg_buf(req->rq_repmsg, 1));
 440                 if (rc < 0) {
 441                         CERROR("mds_fs_get_obdo failed: %d\n", rc);
 442                         GOTO(out, rc);
 443                 }
 444         }
 445 out:
 446         l_dput(de);
 447 out_pop:
 448         pop_ctxt(&saved);
 449         req->rq_status = rc;
 450         RETURN(0);
 451 }
 452
 453 static int mds_statfs(struct ptlrpc_request *req)
 454 {
 455         struct mds_obd *mds = &req->rq_obd->u.mds;
 456         struct obd_statfs *osfs;
 457         struct statfs sfs;
 458         int rc, size = sizeof(*osfs);
 459         ENTRY;
 460
 461         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 462                              &req->rq_repmsg);
 463         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 464                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 465                 GOTO(out, rc);
 466         }
 467
 468         rc = vfs_statfs(mds->mds_sb, &sfs);
 469         if (rc) {
 470                 CERROR("mds: statfs failed: rc %d\n", rc);
 471                 GOTO(out, rc);
 472         }
 473         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 474         memset(osfs, 0, size);
 475         /* FIXME: hack until the MDS gets the OST data, next month */
 476         sfs.f_blocks = -1;
 477         sfs.f_bfree = -1;
 478         sfs.f_bavail = -1;
 479         obd_statfs_pack(osfs, &sfs);
 480
 481 out:
 482         req->rq_status = rc;
 483         RETURN(0);
 484 }
 485
 486 static int mds_open(struct ptlrpc_request *req)
 487 {
 488         struct dentry *de;
 489         struct mds_body *body;
 490         struct file *file;
 491         struct vfsmount *mnt;
 492         struct mds_obd *mds = &req->rq_obd->u.mds;
 493         struct mds_client_info *mci;
 494         __u32 flags;
 495         struct list_head *tmp;
 496         struct mds_file_data *mfd;
 497         int rc, size = sizeof(*body);
 498         ENTRY;
 499
 500         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 501         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 502                 CERROR("mds: out of memory\n");
 503                 req->rq_status = -ENOMEM;
 504                 RETURN(0);
 505         }
 506
 507         mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
 508         if (!mci) {
 509                 CERROR("mds: no mci!\n");
 510                 req->rq_status = -ENOTCONN;
 511                 RETURN(0);
 512         }
 513
 514         body = lustre_msg_buf(req->rq_reqmsg, 0);
 515
 516         /* was this animal open already? */
 517         /* XXX we chould only check on re-open, or do a refcount... */
 518         list_for_each(tmp, &mci->mci_open_head) {
 519                 struct mds_file_data *fd;
 520                 fd = list_entry(tmp, struct mds_file_data, mfd_list);
 521                 if (body->extra == fd->mfd_clientfd &&
 522                     body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
 523                         CERROR("Re opening %Ld\n", body->fid1.id);
 524                         RETURN(0);
 525                 }
 526         }
 527
 528         OBD_ALLOC(mfd, sizeof(*mfd));
 529         if (!mfd) {
 530                 CERROR("mds: out of memory\n");
 531                 req->rq_status = -ENOMEM;
 532                 RETURN(0);
 533         }
 534
 535         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 536         if (IS_ERR(de)) {
 537                 req->rq_status = -ENOENT;
 538                 RETURN(0);
 539         }
 540
 541         /* check if this inode has seen a delayed object creation */
 542         if (req->rq_reqmsg->bufcount > 1) {
 543                 void *handle;
 544                 struct inode *inode = de->d_inode;
 545                 //struct iattr iattr;
 546                 struct obdo *obdo;
 547                 int rc;
 548
 549                 obdo = lustre_msg_buf(req->rq_reqmsg, 1);
 550                 //iattr.ia_valid = ATTR_MODE;
 551                 //iattr.ia_mode = inode->i_mode;
 552
 553                 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 554                 if (!handle) {
 555                         req->rq_status = -ENOMEM;
 556                         RETURN(0);
 557                 }
 558
 559                 /* XXX error handling */
 560                 rc = mds_fs_set_obdo(mds, inode, handle, obdo);
 561                 //                rc = mds_fs_setattr(mds, de, handle, &iattr);
 562                 if (!rc)
 563                         rc = mds_update_last_rcvd(mds, handle, req);
 564                 else {
 565                         req->rq_status = rc;
 566                         RETURN(0);
 567                 }
 568                 /* FIXME: need to return last_rcvd, last_committed */
 569
 570                 /* FIXME: keep rc intact */
 571                 rc = mds_fs_commit(mds, de->d_inode, handle);
 572                 if (rc) {
 573                         req->rq_status = rc;
 574                         RETURN(0);
 575                 }
 576         }
 577
 578         flags = body->flags;
 579         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 580         if (!file || IS_ERR(file)) {
 581                 req->rq_status = -EINVAL;
 582                 OBD_FREE(mfd, sizeof(*mfd));
 583                 RETURN(0);
 584         }
 585
 586         file->private_data = mfd;
 587         mfd->mfd_file = file;
 588         mfd->mfd_clientfd = body->extra;
 589         list_add(&mfd->mfd_list, &mci->mci_open_head);
 590
 591         body = lustre_msg_buf(req->rq_repmsg, 0);
 592         body->extra = (__u64) (unsigned long)file;
 593         RETURN(0);
 594 }
 595
 596 static
 597 int mds_close(struct ptlrpc_request *req)
 598 {
 599         struct dentry *de;
 600         struct mds_body *body;
 601         struct file *file;
 602         struct mds_obd *mds = &req->rq_obd->u.mds;
 603         struct vfsmount *mnt;
 604         struct mds_file_data *mfd;
 605         int rc;
 606         ENTRY;
 607
 608         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 609         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 610                 CERROR("mds: out of memory\n");
 611                 req->rq_status = -ENOMEM;
 612                 RETURN(0);
 613         }
 614
 615         body = lustre_msg_buf(req->rq_reqmsg, 0);
 616         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 617         if (IS_ERR(de)) {
 618                 req->rq_status = -ENOENT;
 619                 RETURN(0);
 620         }
 621
 622         file = (struct file *)(unsigned long)body->extra;
 623         if (!file->f_dentry)
 624                 LBUG();
 625         mfd = (struct mds_file_data *)file->private_data;
 626         list_del(&mfd->mfd_list);
 627         OBD_FREE(mfd, sizeof(*mfd));
 628
 629         req->rq_status = filp_close(file, 0);
 630         l_dput(de);
 631         mntput(mnt);
 632
 633         RETURN(0);
 634 }
 635
 636 static int mds_readpage(struct mds_obd *mds, struct ptlrpc_request *req)
 637 {
 638         struct vfsmount *mnt;
 639         struct dentry *de;
 640         struct file *file;
 641         struct mds_body *body;
 642         struct obd_run_ctxt saved;
 643         int rc, size = sizeof(*body);
 644         ENTRY;
 645
 646         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 647         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 648                 CERROR("mds: out of memory\n");
 649                 GOTO(out, rc = -ENOMEM);
 650         }
 651
 652         body = lustre_msg_buf(req->rq_reqmsg, 0);
 653         push_ctxt(&saved, &mds->mds_ctxt);
 654         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 655         if (IS_ERR(de))
 656                 GOTO(out_pop, rc = PTR_ERR(de));
 657
 658         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 659
 660         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 661         /* note: in case of an error, dentry_open puts dentry */
 662         if (IS_ERR(file))
 663                 GOTO(out_pop, rc = PTR_ERR(file));
 664
 665         /* to make this asynchronous make sure that the handling function
 666            doesn't send a reply when this function completes. Instead a
 667            callback function would send the reply */
 668         rc = mds_sendpage(req, file, body->size);
 669
 670         filp_close(file, 0);
 671 out_pop:
 672         pop_ctxt(&saved);
 673 out:
 674         req->rq_status = rc;
 675         RETURN(0);
 676 }
 677
 678 int mds_reint(int offset, struct ptlrpc_request *req)
 679 {
 680         int rc;
 681         struct mds_update_record rec;
 682
 683         rc = mds_update_unpack(req, offset, &rec);
 684         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 685                 CERROR("invalid record\n");
 686                 req->rq_status = -EINVAL;
 687                 RETURN(0);
 688         }
 689         /* rc will be used to interrupt a for loop over multiple records */
 690         rc = mds_reint_rec(&rec, offset, req);
 691         return rc;
 692 }
 693
 694 int mds_handle(struct obd_device *dev, struct ptlrpc_service *svc,
 695                struct ptlrpc_request *req)
 696 {
 697         struct mds_obd *mds = NULL;
 698         int rc;
 699         ENTRY;
 700
 701         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 702         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
 703                 CERROR("lustre_mds: Invalid request\n");
 704                 GOTO(out, rc);
 705         }
 706
 707         if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 708                 CERROR("lustre_mds: wrong packet type sent %d\n",
 709                        req->rq_reqmsg->type);
 710                 GOTO(out, rc = -EINVAL);
 711         }
 712
 713         if (req->rq_reqmsg->opc != MDS_CONNECT) {
 714                 int id = req->rq_reqmsg->target_id;
 715                 struct obd_device *obddev;
 716                 if (id < 0 || id > MAX_OBD_DEVICES)
 717                         GOTO(out, rc = -ENODEV);
 718                 obddev = &obd_dev[id];
 719                 if (strcmp(obddev->obd_type->typ_name, "mds") != 0)
 720                         GOTO(out, rc = -EINVAL);
 721                 mds = &obddev->u.mds;
 722                 req->rq_obd = obddev;
 723         }
 724
 725         switch (req->rq_reqmsg->opc) {
 726         case MDS_CONNECT:
 727                 CDEBUG(D_INODE, "connect\n");
 728                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
 729                 rc = mds_connect(req, &mds);
 730                 break;
 731
 732         case MDS_DISCONNECT:
 733                 CDEBUG(D_INODE, "disconnect\n");
 734                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
 735                 rc = mds_disconnect(mds, req);
 736                 break;
 737
 738         case MDS_GETSTATUS:
 739                 CDEBUG(D_INODE, "getstatus\n");
 740                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
 741                 rc = mds_getstatus(mds, req);
 742                 break;
 743
 744         case MDS_GETATTR:
 745                 CDEBUG(D_INODE, "getattr\n");
 746                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 747                 rc = mds_getattr(0, req);
 748                 break;
 749
 750         case MDS_STATFS:
 751                 CDEBUG(D_INODE, "statfs\n");
 752                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
 753                 rc = mds_statfs(req);
 754                 break;
 755
 756         case MDS_READPAGE:
 757                 CDEBUG(D_INODE, "readpage\n");
 758                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 759                 rc = mds_readpage(mds, req);
 760
 761                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 762                         return 0;
 763                 break;
 764
 765         case MDS_REINT: {
 766                 int size = sizeof(struct mds_body);
 767                 CDEBUG(D_INODE, "reint\n");
 768                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 769
 770                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 771                                      &req->rq_repmsg);
 772                 if (rc) {
 773                         rc = req->rq_status = -ENOMEM;
 774                         break;
 775                 }
 776                 rc = mds_reint(0, req);
 777                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
 778                 break;
 779         }
 780
 781         case MDS_OPEN:
 782                 CDEBUG(D_INODE, "open\n");
 783                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 784                 rc = mds_open(req);
 785                 break;
 786
 787         case MDS_CLOSE:
 788                 CDEBUG(D_INODE, "close\n");
 789                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 790                 rc = mds_close(req);
 791                 break;
 792
 793         default:
 794                 rc = ptlrpc_error(svc, req);
 795                 RETURN(rc);
 796         }
 797
 798         EXIT;
 799 out:
 800         /* Still not 100% sure whether we should reply with the server
 801          * last_rcvd or that of this client.  I'm not sure it even makes
 802          * a difference on a per-client basis, because last_rcvd is global
 803          * and we are not supposed to allow transactions while in recovery.
 804          */
 805         if (rc) {
 806                 ptlrpc_error(svc, req);
 807         } else {
 808                 req->rq_repmsg->last_rcvd = HTON__u64(mds->mds_last_rcvd);
 809                 req->rq_repmsg->last_committed =
 810                         HTON__u64(mds->mds_last_committed);
 811                 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
 812                        (unsigned long long)mds->mds_last_rcvd,
 813                        (unsigned long long)mds->mds_last_committed,
 814                        cpu_to_le32(req->rq_xid));
 815                 CDEBUG(D_NET, "sending reply\n");
 816                 ptlrpc_reply(svc, req);
 817         }
 818         return 0;
 819 }
 820
 821 /* Update the server data on disk.  This stores the new mount_count and
 822  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
 823  * then the server last_rcvd value may be less than that of the clients.
 824  * This will alert us that we may need to do client recovery.
 825  */
 826 static
 827 int mds_update_server_data(struct mds_obd *mds)
 828 {
 829         struct obd_run_ctxt saved;
 830         struct mds_server_data *msd = mds->mds_server_data;
 831         struct file *filp = mds->mds_rcvd_filp;
 832         loff_t off = 0;
 833         int rc;
 834
 835         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
 836         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
 837
 838         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
 839                (unsigned long long)mds->mds_mount_count,
 840                (unsigned long long)mds->mds_last_rcvd);
 841         push_ctxt(&saved, &mds->mds_ctxt);
 842         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
 843         if (rc != sizeof(*msd)) {
 844                 CERROR("error writing MDS server data: rc = %d\n", rc);
 845                 if (rc > 0)
 846                         RETURN(-EIO);
 847                 RETURN(rc);
 848         }
 849         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
 850         pop_ctxt(&saved);
 851         if (rc)
 852                 CERROR("error flushing MDS server data: rc = %d\n", rc);
 853
 854         return 0;
 855 }
 856
 857 /* Do recovery actions for the MDS */
 858 static int mds_recover(struct obd_device *obddev)
 859 {
 860         struct mds_obd *mds = &obddev->u.mds;
 861         int rc;
 862
 863         /* This happens at the end when recovery is complete */
 864         ++mds->mds_mount_count;
 865         rc = mds_update_server_data(mds);
 866
 867         return rc;
 868 }
 869
 870
 871 /* mount the file system (secretly) */
 872 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
 873 {
 874         struct obd_ioctl_data* data = buf;
 875         struct mds_obd *mds = &obddev->u.mds;
 876         struct vfsmount *mnt;
 877         int rc = 0;
 878         ENTRY;
 879
 880         MOD_INC_USE_COUNT;
 881 #ifdef CONFIG_DEV_RDONLY
 882         dev_clear_rdonly(2);
 883 #endif
 884         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 885                 GOTO(err_dec, rc = -EINVAL);
 886
 887         mds->mds_fstype = strdup(data->ioc_inlbuf2);
 888
 889         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
 890         if (IS_ERR(mnt)) {
 891                 rc = PTR_ERR(mnt);
 892                 CERROR("do_kern_mount failed: rc = %d\n", rc);
 893                 GOTO(err_kfree, rc);
 894         }
 895
 896         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 897         if (!mds->mds_sb)
 898                 GOTO(err_put, rc = -ENODEV);
 899
 900         rc = mds_fs_setup(mds, mnt);
 901         if (rc) {
 902                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
 903                 GOTO(err_put, rc);
 904         }
 905
 906         mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
 907                                            MDC_REPLY_PORTAL, "self",mds_handle);
 908         if (!mds->mds_service) {
 909                 CERROR("failed to start service\n");
 910                 GOTO(err_fs, rc = -EINVAL);
 911         }
 912
 913         rc = ptlrpc_start_thread(obddev, mds->mds_service, "lustre_mds");
 914         if (rc) {
 915                 CERROR("cannot start thread: rc = %d\n", rc);
 916                 GOTO(err_svc, rc);
 917         }
 918
 919         rc = -ENOENT;
 920         mds->mds_ldlm_conn = ptlrpc_uuid_to_connection("self");
 921         if (!mds->mds_ldlm_conn) {
 922                 mds_cleanup(obddev);
 923                 GOTO(err_thread, rc);
 924         }
 925
 926         obddev->obd_namespace =
 927                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
 928         if (obddev->obd_namespace == NULL) {
 929                 LBUG();
 930                 mds_cleanup(obddev);
 931                 GOTO(err_thread, rc);
 932         }
 933
 934         mds->mds_local_namespace =
 935                 ldlm_namespace_new("mds_client", LDLM_NAMESPACE_CLIENT);
 936         if (mds->mds_local_namespace == NULL) {
 937                 LBUG();
 938                 mds_cleanup(obddev);
 939                 GOTO(err_thread, rc);
 940         }
 941
 942         OBD_ALLOC(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
 943         if (mds->mds_ldlm_client == NULL) {
 944                 LBUG();
 945                 mds_cleanup(obddev);
 946                 GOTO(err_thread, rc);
 947         }
 948         ptlrpc_init_client(NULL, NULL, LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
 949                            mds->mds_ldlm_client);
 950         mds->mds_ldlm_client->cli_target_devno = obddev->obd_minor;
 951         mds->mds_ldlm_client->cli_name = "mds ldlm";
 952
 953         rc = mds_recover(obddev);
 954         if (rc)
 955                 GOTO(err_thread, rc);
 956
 957         RETURN(0);
 958
 959 err_thread:
 960         ptlrpc_stop_all_threads(mds->mds_service);
 961 err_svc:
 962         ptlrpc_unregister_service(mds->mds_service);
 963 err_fs:
 964         mds_fs_cleanup(mds);
 965 err_put:
 966         unlock_kernel();
 967         mntput(mds->mds_vfsmnt);
 968         mds->mds_sb = 0;
 969         lock_kernel();
 970 err_kfree:
 971         kfree(mds->mds_fstype);
 972 err_dec:
 973         MOD_DEC_USE_COUNT;
 974         return rc;
 975 }
 976
 977 static int mds_cleanup(struct obd_device * obddev)
 978 {
 979         struct super_block *sb;
 980         struct mds_obd *mds = &obddev->u.mds;
 981
 982         ENTRY;
 983
 984         if ( !list_empty(&obddev->obd_gen_clients) ) {
 985                 CERROR("still has clients!\n");
 986                 RETURN(-EBUSY);
 987         }
 988
 989         ptlrpc_stop_all_threads(mds->mds_service);
 990         ptlrpc_unregister_service(mds->mds_service);
 991
 992         sb = mds->mds_sb;
 993         if (!mds->mds_sb)
 994                 RETURN(0);
 995
 996         mds_update_server_data(mds);
 997
 998         if (mds->mds_rcvd_filp) {
 999                 int rc = filp_close(mds->mds_rcvd_filp, 0);
1000                 mds->mds_rcvd_filp = NULL;
1001
1002                 if (rc)
1003                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
1004         }
1005
1006         unlock_kernel();
1007         mntput(mds->mds_vfsmnt);
1008         mds->mds_sb = 0;
1009         kfree(mds->mds_fstype);
1010
1011         ldlm_namespace_free(mds->mds_local_namespace);
1012         ldlm_namespace_free(obddev->obd_namespace);
1013
1014         if (mds->mds_ldlm_conn != NULL)
1015                 ptlrpc_put_connection(mds->mds_ldlm_conn);
1016
1017         OBD_FREE(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
1018
1019         lock_kernel();
1020 #ifdef CONFIG_DEV_RDONLY
1021         dev_clear_rdonly(2);
1022 #endif
1023         mds_fs_cleanup(mds);
1024
1025         MOD_DEC_USE_COUNT;
1026         RETURN(0);
1027 }
1028
1029 /* use obd ops to offer management infrastructure */
1030 static struct obd_ops mds_obd_ops = {
1031         o_setup:       mds_setup,
1032         o_cleanup:     mds_cleanup,
1033 };
1034
1035 static int __init mds_init(void)
1036 {
1037         inter_module_register("mds_reint", THIS_MODULE, &mds_reint);
1038         inter_module_register("mds_getattr_name", THIS_MODULE,
1039                               &mds_getattr_name);
1040         obd_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1041         return 0;
1042 }
1043
1044 static void __exit mds_exit(void)
1045 {
1046         inter_module_unregister("mds_reint");
1047         inter_module_unregister("mds_getattr_name");
1048         obd_unregister_type(LUSTRE_MDS_NAME);
1049 }
1050
1051 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1052 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1053 MODULE_LICENSE("GPL");
1054
1055 module_init(mds_init);
1056 module_exit(mds_exit);