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_get_lovtgts(struct obd_device *obd, uuid_t *uuidarray);
  27 extern int mds_get_lovdesc(struct obd_device *obd, struct lov_desc *desc);
  28 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
  29                                 struct ptlrpc_request *req);
  30 static int mds_cleanup(struct obd_device * obddev);
  31
  32 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
  33 {
  34         return &req->rq_export->exp_obd->u.mds;
  35 }
  36
  37 /* Assumes caller has already pushed into the kernel filesystem context */
  38 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  39                         __u64 offset)
  40 {
  41         int rc = 0;
  42         struct mds_obd *mds = mds_req2mds(req);
  43         struct ptlrpc_bulk_desc *desc;
  44         struct ptlrpc_bulk_page *bulk;
  45         char *buf;
  46         ENTRY;
  47
  48         desc = ptlrpc_prep_bulk(req->rq_connection);
  49         if (desc == NULL)
  50                 GOTO(out, rc = -ENOMEM);
  51
  52         bulk = ptlrpc_prep_bulk_page(desc);
  53         if (bulk == NULL)
  54                 GOTO(cleanup_bulk, rc = -ENOMEM);
  55
  56         OBD_ALLOC(buf, PAGE_SIZE);
  57         if (buf == NULL)
  58                 GOTO(cleanup_bulk, rc = -ENOMEM);
  59
  60         rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
  61
  62         if (rc != PAGE_SIZE)
  63                 GOTO(cleanup_buf, rc = -EIO);
  64
  65         bulk->b_xid = req->rq_xid;
  66         bulk->b_buf = buf;
  67         bulk->b_buflen = PAGE_SIZE;
  68         desc->b_portal = MDS_BULK_PORTAL;
  69
  70         rc = ptlrpc_send_bulk(desc);
  71         if (rc)
  72                 GOTO(cleanup_buf, rc);
  73
  74         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
  75                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
  76                        OBD_FAIL_MDS_SENDPAGE, rc);
  77                 ptlrpc_abort_bulk(desc);
  78                 GOTO(cleanup_buf, rc);
  79         }
  80
  81         wait_event(desc->b_waitq, ptlrpc_check_bulk_sent(desc));
  82         if (desc->b_flags & PTL_RPC_FL_INTR)
  83                 GOTO(cleanup_buf, rc = -EINTR);
  84
  85         EXIT;
  86  cleanup_buf:
  87         OBD_FREE(buf, PAGE_SIZE);
  88  cleanup_bulk:
  89         ptlrpc_free_bulk(desc);
  90  out:
  91         return rc;
  92 }
  93
  94 /* 'dir' is a inode for which a lock has already been taken */
  95 struct dentry *mds_name2locked_dentry(struct mds_obd *mds, struct dentry *dir,
  96                                       struct vfsmount **mnt, char *name,
  97                                       int namelen, int lock_mode,
  98                                       struct lustre_handle *lockh,
  99                                       int dir_lock_mode)
 100 {
 101         struct dentry *dchild;
 102         int flags, rc;
 103         __u64 res_id[3] = {0};
 104         ENTRY;
 105
 106         down(&dir->d_inode->i_sem);
 107         dchild = lookup_one_len(name, dir, namelen);
 108         if (IS_ERR(dchild)) {
 109                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 110                 up(&dir->d_inode->i_sem);
 111                 LBUG();
 112         }
 113         if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
 114                 up(&dir->d_inode->i_sem);
 115                 ldlm_lock_decref(lockh, dir_lock_mode);
 116         }
 117
 118         if (lock_mode == 0 || !dchild->d_inode)
 119                 RETURN(dchild);
 120
 121         res_id[0] = dchild->d_inode->i_ino;
 122         rc = ldlm_match_or_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
 123                                    (struct lustre_handle *)&mds->mds_connh,
 124                                    NULL, mds->mds_local_namespace, NULL,
 125                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 126                                    &flags, (void *)mds_lock_callback, NULL,
 127                                    0, lockh);
 128         if (rc != ELDLM_OK) {
 129                 l_dput(dchild);
 130                 RETURN(NULL);
 131         }
 132
 133         RETURN(dchild);
 134 }
 135
 136 struct dentry *mds_fid2locked_dentry(struct mds_obd *mds, struct ll_fid *fid,
 137                                      struct vfsmount **mnt, int lock_mode,
 138                                      struct lustre_handle *lockh)
 139 {
 140         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 141         int flags, rc;
 142         __u64 res_id[3] = {0};
 143         ENTRY;
 144
 145         if (IS_ERR(de))
 146                 RETURN(de);
 147
 148         res_id[0] = de->d_inode->i_ino;
 149         rc = ldlm_match_or_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
 150                                    (struct lustre_handle *)&mds->mds_connh,
 151                                    NULL, mds->mds_local_namespace, NULL,
 152                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 153                                    &flags, (void *)mds_lock_callback, NULL,
 154                                    0, lockh);
 155         if (rc != ELDLM_OK) {
 156                 l_dput(de);
 157                 retval = NULL;
 158         }
 159
 160         RETURN(retval);
 161 }
 162
 163 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 164                               struct vfsmount **mnt)
 165 {
 166         /* stolen from NFS */
 167         struct super_block *sb = mds->mds_sb;
 168         unsigned long ino = fid->id;
 169         __u32 generation = fid->generation;
 170         struct inode *inode;
 171         struct list_head *lp;
 172         struct dentry *result;
 173
 174         if (ino == 0)
 175                 RETURN(ERR_PTR(-ESTALE));
 176
 177         inode = iget(sb, ino);
 178         if (inode == NULL)
 179                 RETURN(ERR_PTR(-ENOMEM));
 180
 181         CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
 182
 183         if (is_bad_inode(inode) ||
 184             (generation && inode->i_generation != generation)) {
 185                 /* we didn't find the right inode.. */
 186                 CERROR("bad inode %lu, link: %d ct: %d or version  %u/%u\n",
 187                        inode->i_ino, inode->i_nlink,
 188                        atomic_read(&inode->i_count), inode->i_generation,
 189                        generation);
 190                 LBUG();
 191                 iput(inode);
 192                 RETURN(ERR_PTR(-ESTALE));
 193         }
 194
 195         /* now to find a dentry.
 196          * If possible, get a well-connected one
 197          */
 198         if (mnt)
 199                 *mnt = mds->mds_vfsmnt;
 200         spin_lock(&dcache_lock);
 201         for (lp = inode->i_dentry.next; lp != &inode->i_dentry ; lp=lp->next) {
 202                 result = list_entry(lp,struct dentry, d_alias);
 203                 if (! (result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
 204                         dget_locked(result);
 205                         result->d_vfs_flags |= DCACHE_REFERENCED;
 206                         spin_unlock(&dcache_lock);
 207                         iput(inode);
 208                         if (mnt)
 209                                 mntget(*mnt);
 210                         return result;
 211                 }
 212         }
 213         spin_unlock(&dcache_lock);
 214         result = d_alloc_root(inode);
 215         if (result == NULL) {
 216                 iput(inode);
 217                 return ERR_PTR(-ENOMEM);
 218         }
 219         if (mnt)
 220                 mntget(*mnt);
 221         result->d_flags |= DCACHE_NFSD_DISCONNECTED;
 222         return result;
 223 }
 224
 225 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd)
 226 {
 227         int rc;
 228
 229         MOD_INC_USE_COUNT;
 230         rc = class_connect(conn, obd);
 231
 232         if (rc)
 233                 MOD_DEC_USE_COUNT;
 234
 235         return rc;
 236 }
 237
 238 static int mds_disconnect(struct lustre_handle *conn)
 239 {
 240         int rc;
 241
 242         rc = class_disconnect(conn);
 243         if (!rc)
 244                 MOD_DEC_USE_COUNT;
 245
 246         return rc;
 247 }
 248
 249 /* FIXME: the error cases need fixing to avoid leaks */
 250 static int mds_getstatus(struct ptlrpc_request *req)
 251 {
 252         struct mds_obd *mds = mds_req2mds(req);
 253         struct mds_body *body;
 254         struct mds_client_info *mci;
 255         struct mds_client_data *mcd;
 256         int rc, size = sizeof(*body);
 257         ENTRY;
 258
 259         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 260         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 261                 CERROR("mds: out of memory for message: size=%d\n", size);
 262                 req->rq_status = -ENOMEM;
 263                 RETURN(0);
 264         }
 265
 266         body = lustre_msg_buf(req->rq_reqmsg, 0);
 267         mds_unpack_body(body);
 268
 269         /* Anything we need to do here with the client's trans no or so? */
 270         body = lustre_msg_buf(req->rq_repmsg, 0);
 271         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 272
 273         mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
 274         if (!mci) {
 275                 /* We don't have any old connection data for this client */
 276                 int rc;
 277
 278                 CDEBUG(D_INFO, "allocating new client data for UUID '%s'",
 279                        ptlrpc_req_to_uuid(req));
 280
 281                 OBD_ALLOC(mcd, sizeof(*mcd));
 282                 if (!mcd) {
 283                         CERROR("mds: out of memory for client data\n");
 284                         req->rq_status = -ENOMEM;
 285                         RETURN(0);
 286                 }
 287                 memcpy(mcd->mcd_uuid, ptlrpc_req_to_uuid(req),
 288                        sizeof(mcd->mcd_uuid));
 289                 rc = mds_client_add(mds, mcd, -1);
 290                 if (rc) {
 291                         req->rq_status = rc;
 292                         RETURN(0);
 293                 }
 294         } else {
 295                 /* We have old connection data for this client... */
 296                 mcd = mci->mci_mcd;
 297                 CDEBUG(D_INFO, "found existing data for UUID '%s' at #%d\n",
 298                        mcd->mcd_uuid, mci->mci_off);
 299         }
 300         /* mcd_last_xid is is stored in little endian on the disk and
 301            mds_pack_rep_body converts it to network order */
 302         body->last_xid = le32_to_cpu(mcd->mcd_last_xid);
 303         mds_pack_rep_body(req);
 304         RETURN(0);
 305 }
 306
 307 static int mds_lovinfo(struct ptlrpc_request *req)
 308 {
 309         struct mds_obd *mds = mds_req2mds(req);
 310         struct mds_status_req *streq;
 311         struct lov_desc *desc;
 312         int rc, size[2] = {sizeof(*desc)};
 313         ENTRY;
 314
 315         streq = lustre_msg_buf(req->rq_reqmsg, 0);
 316         streq->flags = NTOH__u32(streq->flags);
 317         streq->repbuf = NTOH__u32(streq->repbuf);
 318         size[1] = streq->repbuf;
 319
 320         rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
 321         if (rc) {
 322                 CERROR("mds: out of memory for message: size=%d\n", size[1]);
 323                 req->rq_status = -ENOMEM;
 324                 RETURN(0);
 325         }
 326
 327         desc = lustre_msg_buf(req->rq_repmsg, 0);
 328         rc = mds_get_lovdesc(req->rq_obd, desc);
 329         if (rc != 0 ) {
 330                 CERROR("get_lovdesc error %d", rc);
 331                 req->rq_status = rc;
 332                 RETURN(0);
 333         }
 334
 335         if (desc->ld_tgt_count * sizeof(uuid_t) > streq->repbuf) {
 336                 CERROR("too many targets, enlarge client buffers\n");
 337                 req->rq_status = -ENOSPC;
 338                 RETURN(0);
 339         }
 340
 341         mds->mds_max_mdsize = sizeof(desc) +
 342                 desc->ld_tgt_count * sizeof(uuid_t);
 343         rc = mds_get_lovtgts(req->rq_obd, lustre_msg_buf(req->rq_repmsg, 1));
 344         if (rc) {
 345                 CERROR("get_lovtgts error %d", rc);
 346                 req->rq_status = rc;
 347                 RETURN(0);
 348         }
 349         RETURN(0);
 350 }
 351
 352 int mds_lock_callback(struct lustre_handle *lockh, struct ldlm_lock_desc *desc,
 353                       void *data, int data_len, struct ptlrpc_request **reqp)
 354 {
 355         ENTRY;
 356
 357         if (desc == NULL) {
 358                 /* Completion AST.  Do nothing */
 359                 RETURN(0);
 360         }
 361
 362         if (ldlm_cli_cancel(lockh) < 0)
 363                 LBUG();
 364         RETURN(0);
 365 }
 366
 367 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 368 {
 369         struct mds_obd *mds = mds_req2mds(req);
 370         struct obd_run_ctxt saved;
 371         struct mds_body *body;
 372         struct dentry *de = NULL, *dchild = NULL;
 373         struct inode *dir;
 374         struct lustre_handle lockh;
 375         char *name;
 376         int namelen, flags, lock_mode, rc = 0;
 377         __u64 res_id[3] = {0, 0, 0};
 378         ENTRY;
 379
 380         if (strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds") != 0)
 381                 LBUG();
 382
 383         if (req->rq_reqmsg->bufcount <= offset + 1) {
 384                 LBUG();
 385                 GOTO(out_pre_de, rc = -EINVAL);
 386         }
 387
 388         body = lustre_msg_buf(req->rq_reqmsg, offset);
 389         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 390         namelen = req->rq_reqmsg->buflens[offset + 1];
 391         /* requests were at offset 2, replies go back at 1 */
 392         if (offset)
 393                 offset = 1;
 394
 395         push_ctxt(&saved, &mds->mds_ctxt);
 396         de = mds_fid2dentry(mds, &body->fid1, NULL);
 397         if (IS_ERR(de)) {
 398                 LBUG();
 399                 GOTO(out_pre_de, rc = -ESTALE);
 400         }
 401
 402         dir = de->d_inode;
 403         CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
 404
 405         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 406         res_id[0] = dir->i_ino;
 407
 408         rc = ldlm_lock_match(mds->mds_local_namespace, res_id, LDLM_PLAIN,
 409                              NULL, 0, lock_mode, &lockh);
 410         if (rc == 0) {
 411                 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
 412                 rc = ldlm_cli_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
 413                                       (struct lustre_handle *)&mds->mds_connh,
 414                                       NULL, mds->mds_local_namespace, NULL,
 415                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 416                                       &flags, (void *)mds_lock_callback,
 417                                       NULL, 0, &lockh);
 418                 if (rc != ELDLM_OK) {
 419                         CERROR("lock enqueue: err: %d\n", rc);
 420                         GOTO(out_create_de, rc = -EIO);
 421                 }
 422         }
 423         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 424
 425         down(&dir->i_sem);
 426         dchild = lookup_one_len(name, de, namelen - 1);
 427         if (IS_ERR(dchild)) {
 428                 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
 429                 up(&dir->i_sem);
 430                 LBUG();
 431                 GOTO(out_create_dchild, rc = -ESTALE);
 432         }
 433
 434         if (dchild->d_inode) {
 435                 struct mds_body *body;
 436                 struct inode *inode = dchild->d_inode;
 437                 CDEBUG(D_INODE, "child exists (dir %ld, name %s, ino %ld)\n",
 438                        dir->i_ino, name, dchild->d_inode->i_ino);
 439
 440                 body = lustre_msg_buf(req->rq_repmsg, offset);
 441                 mds_pack_inode2fid(&body->fid1, inode);
 442                 mds_pack_inode2body(body, inode);
 443                 if (S_ISREG(inode->i_mode)) {
 444                         struct lov_stripe_md *md;
 445                         md = lustre_msg_buf(req->rq_repmsg, offset + 1);
 446                         md->lmd_size = mds->mds_max_mdsize;
 447                         mds_fs_get_md(mds, inode, md);
 448                 }
 449                 /* now a normal case for intent locking */
 450                 rc = 0;
 451         } else
 452                 rc = -ENOENT;
 453
 454         EXIT;
 455 out_create_dchild:
 456         l_dput(dchild);
 457         up(&dir->i_sem);
 458         ldlm_lock_decref(&lockh, lock_mode);
 459 out_create_de:
 460         l_dput(de);
 461 out_pre_de:
 462         req->rq_status = rc;
 463         pop_ctxt(&saved);
 464         return 0;
 465 }
 466
 467
 468 static int mds_getattr(int offset, struct ptlrpc_request *req)
 469 {
 470         struct mds_obd *mds = mds_req2mds(req);
 471         struct obd_run_ctxt saved;
 472         struct dentry *de;
 473         struct inode *inode;
 474         struct mds_body *body;
 475         int rc, size[2] = {sizeof(*body)}, bufcount = 1;
 476         ENTRY;
 477
 478         body = lustre_msg_buf(req->rq_reqmsg, offset);
 479         push_ctxt(&saved, &mds->mds_ctxt);
 480         de = mds_fid2dentry(mds, &body->fid1, NULL);
 481         if (IS_ERR(de)) {
 482                 GOTO(out_pop, rc = -ENOENT);
 483         }
 484
 485         inode = de->d_inode;
 486         if (S_ISREG(body->fid1.f_type)) {
 487                 bufcount = 2;
 488                 size[1] = mds->mds_max_mdsize;
 489         } else if (body->valid & OBD_MD_LINKNAME) {
 490                 bufcount = 2;
 491                 size[1] = inode->i_size;
 492         }
 493
 494         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 495                              &req->rq_repmsg);
 496         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 497                 CERROR("mds: out of memory\n");
 498                 GOTO(out, rc);
 499         }
 500
 501         if (body->valid & OBD_MD_LINKNAME) {
 502                 char *tmp = lustre_msg_buf(req->rq_repmsg, 1);
 503
 504                 rc = inode->i_op->readlink(de, tmp, size[1]);
 505
 506                 if (rc < 0) {
 507                         CERROR("readlink failed: %d\n", rc);
 508                         GOTO(out, rc);
 509                 }
 510         }
 511
 512         body = lustre_msg_buf(req->rq_repmsg, 0);
 513         body->ino = inode->i_ino;
 514         body->generation = inode->i_generation;
 515         body->atime = inode->i_atime;
 516         body->ctime = inode->i_ctime;
 517         body->mtime = inode->i_mtime;
 518         body->uid = inode->i_uid;
 519         body->gid = inode->i_gid;
 520         body->size = inode->i_size;
 521         body->mode = inode->i_mode;
 522         body->nlink = inode->i_nlink;
 523         body->valid = ~0; /* FIXME: should be more selective */
 524
 525         if (S_ISREG(inode->i_mode)) {
 526                 rc = mds_fs_get_md(mds, inode,
 527                                      lustre_msg_buf(req->rq_repmsg, 1));
 528                 if (rc < 0) {
 529                         CERROR("mds_fs_get_md failed: %d\n", rc);
 530                         GOTO(out, rc);
 531                 }
 532         }
 533 out:
 534         l_dput(de);
 535 out_pop:
 536         pop_ctxt(&saved);
 537         req->rq_status = rc;
 538         RETURN(0);
 539 }
 540
 541 static int mds_statfs(struct ptlrpc_request *req)
 542 {
 543         struct mds_obd *mds = mds_req2mds(req);
 544         struct obd_statfs *osfs;
 545         struct statfs sfs;
 546         int rc, size = sizeof(*osfs);
 547         ENTRY;
 548
 549         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 550                              &req->rq_repmsg);
 551         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 552                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 553                 GOTO(out, rc);
 554         }
 555
 556         rc = vfs_statfs(mds->mds_sb, &sfs);
 557         if (rc) {
 558                 CERROR("mds: statfs failed: rc %d\n", rc);
 559                 GOTO(out, rc);
 560         }
 561         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 562         memset(osfs, 0, size);
 563         obd_statfs_pack(osfs, &sfs);
 564
 565 out:
 566         req->rq_status = rc;
 567         RETURN(0);
 568 }
 569
 570 static int mds_open(struct ptlrpc_request *req)
 571 {
 572         struct dentry *de;
 573         struct mds_body *body;
 574         struct file *file;
 575         struct vfsmount *mnt;
 576         struct mds_obd *mds = mds_req2mds(req);
 577         struct mds_client_info *mci;
 578         __u32 flags;
 579         struct list_head *tmp;
 580         struct mds_file_data *mfd;
 581         int rc, size = sizeof(*body);
 582         ENTRY;
 583
 584         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 585         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 586                 CERROR("mds: out of memory\n");
 587                 req->rq_status = -ENOMEM;
 588                 RETURN(0);
 589         }
 590
 591         mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
 592         if (!mci) {
 593                 CERROR("mds: no mci!\n");
 594                 req->rq_status = -ENOTCONN;
 595                 RETURN(0);
 596         }
 597
 598         body = lustre_msg_buf(req->rq_reqmsg, 0);
 599
 600         /* was this animal open already? */
 601         /* XXX we should only check on re-open, or do a refcount... */
 602         list_for_each(tmp, &mci->mci_open_head) {
 603                 struct mds_file_data *fd;
 604                 fd = list_entry(tmp, struct mds_file_data, mfd_list);
 605                 if (body->extra == fd->mfd_clientfd &&
 606                     body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
 607                         CERROR("Re opening %Ld\n", body->fid1.id);
 608                         RETURN(0);
 609                 }
 610         }
 611
 612         OBD_ALLOC(mfd, sizeof(*mfd));
 613         if (!mfd) {
 614                 CERROR("mds: out of memory\n");
 615                 req->rq_status = -ENOMEM;
 616                 RETURN(0);
 617         }
 618
 619         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 620         if (IS_ERR(de)) {
 621                 req->rq_status = -ENOENT;
 622                 RETURN(0);
 623         }
 624
 625         /* check if this inode has seen a delayed object creation */
 626         if (req->rq_reqmsg->bufcount > 1) {
 627                 void *handle;
 628                 struct lov_stripe_md *md;
 629                 struct inode *inode = de->d_inode;
 630                 //struct iattr iattr;
 631                 int rc;
 632
 633                 md = lustre_msg_buf(req->rq_reqmsg, 1);
 634                 //iattr.ia_mode = inode->i_mode;
 635
 636                 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 637                 if (!handle) {
 638                         req->rq_status = -ENOMEM;
 639                         RETURN(0);
 640                 }
 641
 642                 /* XXX error handling */
 643                 rc = mds_fs_set_md(mds, inode, handle, md);
 644                 //                rc = mds_fs_setattr(mds, de, handle, &iattr);
 645                 if (!rc) {
 646                         struct obd_run_ctxt saved;
 647                         push_ctxt(&saved, &mds->mds_ctxt);
 648                         rc = mds_update_last_rcvd(mds, handle, req);
 649                         pop_ctxt(&saved);
 650                 } else {
 651                         req->rq_status = rc;
 652                         RETURN(0);
 653                 }
 654                 /* FIXME: need to return last_rcvd, last_committed */
 655
 656                 /* FIXME: keep rc intact */
 657                 rc = mds_fs_commit(mds, de->d_inode, handle);
 658                 if (rc) {
 659                         req->rq_status = rc;
 660                         RETURN(0);
 661                 }
 662         }
 663
 664         flags = body->flags;
 665         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 666         if (!file || IS_ERR(file)) {
 667                 req->rq_status = -EINVAL;
 668                 OBD_FREE(mfd, sizeof(*mfd));
 669                 RETURN(0);
 670         }
 671
 672         file->private_data = mfd;
 673         mfd->mfd_file = file;
 674         mfd->mfd_clientfd = body->extra;
 675         list_add(&mfd->mfd_list, &mci->mci_open_head);
 676
 677         body = lustre_msg_buf(req->rq_repmsg, 0);
 678         body->extra = (__u64) (unsigned long)file;
 679         RETURN(0);
 680 }
 681
 682 static int mds_close(struct ptlrpc_request *req)
 683 {
 684         struct dentry *de;
 685         struct mds_body *body;
 686         struct file *file;
 687         struct mds_obd *mds = mds_req2mds(req);
 688         struct vfsmount *mnt;
 689         struct mds_file_data *mfd;
 690         int rc;
 691         ENTRY;
 692
 693         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 694         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 695                 CERROR("mds: out of memory\n");
 696                 req->rq_status = -ENOMEM;
 697                 RETURN(0);
 698         }
 699
 700         body = lustre_msg_buf(req->rq_reqmsg, 0);
 701         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 702         if (IS_ERR(de)) {
 703                 req->rq_status = -ENOENT;
 704                 RETURN(0);
 705         }
 706
 707         file = (struct file *)(unsigned long)body->extra;
 708         if (!file->f_dentry)
 709                 LBUG();
 710         mfd = (struct mds_file_data *)file->private_data;
 711         list_del(&mfd->mfd_list);
 712         OBD_FREE(mfd, sizeof(*mfd));
 713
 714         req->rq_status = filp_close(file, 0);
 715         l_dput(de);
 716         mntput(mnt);
 717
 718         RETURN(0);
 719 }
 720
 721 static int mds_readpage(struct ptlrpc_request *req)
 722 {
 723         struct mds_obd *mds = mds_req2mds(req);
 724         struct vfsmount *mnt;
 725         struct dentry *de;
 726         struct file *file;
 727         struct mds_body *body;
 728         struct obd_run_ctxt saved;
 729         int rc, size = sizeof(*body);
 730         ENTRY;
 731
 732         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 733         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 734                 CERROR("mds: out of memory\n");
 735                 GOTO(out, rc = -ENOMEM);
 736         }
 737
 738         body = lustre_msg_buf(req->rq_reqmsg, 0);
 739         push_ctxt(&saved, &mds->mds_ctxt);
 740         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 741         if (IS_ERR(de))
 742                 GOTO(out_pop, rc = PTR_ERR(de));
 743
 744         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 745
 746         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 747         /* note: in case of an error, dentry_open puts dentry */
 748         if (IS_ERR(file))
 749                 GOTO(out_pop, rc = PTR_ERR(file));
 750
 751         /* to make this asynchronous make sure that the handling function
 752            doesn't send a reply when this function completes. Instead a
 753            callback function would send the reply */
 754         rc = mds_sendpage(req, file, body->size);
 755
 756         filp_close(file, 0);
 757 out_pop:
 758         pop_ctxt(&saved);
 759 out:
 760         req->rq_status = rc;
 761         RETURN(0);
 762 }
 763
 764 int mds_reint(int offset, struct ptlrpc_request *req)
 765 {
 766         int rc;
 767         struct mds_update_record rec;
 768
 769         rc = mds_update_unpack(req, offset, &rec);
 770         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 771                 CERROR("invalid record\n");
 772                 req->rq_status = -EINVAL;
 773                 RETURN(0);
 774         }
 775         /* rc will be used to interrupt a for loop over multiple records */
 776         rc = mds_reint_rec(&rec, offset, req);
 777         return rc;
 778 }
 779
 780 int mds_handle(struct ptlrpc_request *req)
 781 {
 782         int rc;
 783         ENTRY;
 784
 785         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 786         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
 787                 CERROR("lustre_mds: Invalid request\n");
 788                 GOTO(out, rc);
 789         }
 790
 791         if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 792                 CERROR("lustre_mds: wrong packet type sent %d\n",
 793                        req->rq_reqmsg->type);
 794                 GOTO(out, rc = -EINVAL);
 795         }
 796
 797         if (req->rq_reqmsg->opc != MDS_CONNECT &&
 798             req->rq_export == NULL)
 799                 GOTO(out, rc = -ENOTCONN);
 800
 801         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 802                 GOTO(out, rc = -EINVAL);
 803
 804         switch (req->rq_reqmsg->opc) {
 805         case MDS_CONNECT:
 806                 CDEBUG(D_INODE, "connect\n");
 807                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
 808                 rc = target_handle_connect(req);
 809                 break;
 810
 811         case MDS_DISCONNECT:
 812                 CDEBUG(D_INODE, "disconnect\n");
 813                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
 814                 rc = target_handle_disconnect(req);
 815                 goto out;
 816
 817         case MDS_GETSTATUS:
 818                 CDEBUG(D_INODE, "getstatus\n");
 819                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
 820                 rc = mds_getstatus(req);
 821                 break;
 822
 823         case MDS_LOVINFO:
 824                 CDEBUG(D_INODE, "lovinfo\n");
 825                 rc = mds_lovinfo(req);
 826                 break;
 827
 828         case MDS_GETATTR:
 829                 CDEBUG(D_INODE, "getattr\n");
 830                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 831                 rc = mds_getattr(0, req);
 832                 break;
 833
 834         case MDS_STATFS:
 835                 CDEBUG(D_INODE, "statfs\n");
 836                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
 837                 rc = mds_statfs(req);
 838                 break;
 839
 840         case MDS_READPAGE:
 841                 CDEBUG(D_INODE, "readpage\n");
 842                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 843                 rc = mds_readpage(req);
 844
 845                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 846                         return 0;
 847                 break;
 848
 849         case MDS_REINT: {
 850                 int size = sizeof(struct mds_body);
 851                 CDEBUG(D_INODE, "reint\n");
 852                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 853
 854                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 855                                      &req->rq_repmsg);
 856                 if (rc) {
 857                         rc = req->rq_status = -ENOMEM;
 858                         break;
 859                 }
 860                 rc = mds_reint(0, req);
 861                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
 862                 break;
 863         }
 864
 865         case MDS_OPEN:
 866                 CDEBUG(D_INODE, "open\n");
 867                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 868                 rc = mds_open(req);
 869                 break;
 870
 871         case MDS_CLOSE:
 872                 CDEBUG(D_INODE, "close\n");
 873                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 874                 rc = mds_close(req);
 875                 break;
 876
 877         default:
 878                 rc = ptlrpc_error(req->rq_svc, req);
 879                 RETURN(rc);
 880         }
 881
 882         EXIT;
 883
 884         if (!rc) {
 885                 struct mds_obd *mds = mds_req2mds(req);
 886                 req->rq_repmsg->last_rcvd = HTON__u64(mds->mds_last_rcvd);
 887                 req->rq_repmsg->last_committed =
 888                         HTON__u64(mds->mds_last_committed);
 889                 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
 890                        (unsigned long long)mds->mds_last_rcvd,
 891                        (unsigned long long)mds->mds_last_committed,
 892                        cpu_to_le32(req->rq_xid));
 893         }
 894  out:
 895         /* Still not 100% sure whether we should reply with the server
 896          * last_rcvd or that of this client.  I'm not sure it even makes
 897          * a difference on a per-client basis, because last_rcvd is global
 898          * and we are not supposed to allow transactions while in recovery.
 899          */
 900         if (rc) {
 901                 CERROR("mds: processing error %d\n", rc);
 902                 ptlrpc_error(req->rq_svc, req);
 903         } else {
 904                 CDEBUG(D_NET, "sending reply\n");
 905                 ptlrpc_reply(req->rq_svc, req);
 906         }
 907         return 0;
 908 }
 909
 910 /* Update the server data on disk.  This stores the new mount_count and
 911  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
 912  * then the server last_rcvd value may be less than that of the clients.
 913  * This will alert us that we may need to do client recovery.
 914  */
 915 static
 916 int mds_update_server_data(struct mds_obd *mds)
 917 {
 918         struct obd_run_ctxt saved;
 919         struct mds_server_data *msd = mds->mds_server_data;
 920         struct file *filp = mds->mds_rcvd_filp;
 921         loff_t off = 0;
 922         int rc;
 923
 924         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
 925         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
 926
 927         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
 928                (unsigned long long)mds->mds_mount_count,
 929                (unsigned long long)mds->mds_last_rcvd);
 930         push_ctxt(&saved, &mds->mds_ctxt);
 931         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
 932         if (rc != sizeof(*msd)) {
 933                 CERROR("error writing MDS server data: rc = %d\n", rc);
 934                 if (rc > 0)
 935                         RETURN(-EIO);
 936                 RETURN(rc);
 937         }
 938         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
 939         pop_ctxt(&saved);
 940         if (rc)
 941                 CERROR("error flushing MDS server data: rc = %d\n", rc);
 942
 943         return 0;
 944 }
 945
 946 /* Do recovery actions for the MDS */
 947 static int mds_recover(struct obd_device *obddev)
 948 {
 949         struct mds_obd *mds = &obddev->u.mds;
 950         int rc;
 951
 952         /* This happens at the end when recovery is complete */
 953         ++mds->mds_mount_count;
 954         rc = mds_update_server_data(mds);
 955
 956         return rc;
 957 }
 958
 959
 960 /* mount the file system (secretly) */
 961 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
 962 {
 963         struct obd_ioctl_data* data = buf;
 964         struct obd_export *export;
 965         struct mds_obd *mds = &obddev->u.mds;
 966         struct vfsmount *mnt;
 967         int rc = 0;
 968         ENTRY;
 969
 970         MOD_INC_USE_COUNT;
 971 #ifdef CONFIG_DEV_RDONLY
 972         dev_clear_rdonly(2);
 973 #endif
 974         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 975                 GOTO(err_dec, rc = -EINVAL);
 976
 977         mds->mds_fstype = strdup(data->ioc_inlbuf2);
 978
 979         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
 980         if (IS_ERR(mnt)) {
 981                 rc = PTR_ERR(mnt);
 982                 CERROR("do_kern_mount failed: rc = %d\n", rc);
 983                 GOTO(err_kfree, rc);
 984         }
 985
 986         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 987         if (!mds->mds_sb)
 988                 GOTO(err_put, rc = -ENODEV);
 989
 990         mds->mds_max_mdsize = sizeof(struct lov_stripe_md);
 991         rc = mds_fs_setup(mds, mnt);
 992         if (rc) {
 993                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
 994                 GOTO(err_put, rc);
 995         }
 996
 997         mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
 998                                            MDC_REPLY_PORTAL, "self",mds_handle);
 999         if (!mds->mds_service) {
1000                 CERROR("failed to start service\n");
1001                 GOTO(err_fs, rc = -EINVAL);
1002         }
1003
1004         rc = ptlrpc_start_thread(obddev, mds->mds_service, "lustre_mds");
1005         if (rc) {
1006                 CERROR("cannot start thread: rc = %d\n", rc);
1007                 GOTO(err_svc, rc);
1008         }
1009
1010         rc = -ENOENT;
1011         mds->mds_ldlm_conn = ptlrpc_uuid_to_connection("self");
1012         if (!mds->mds_ldlm_conn) {
1013                 mds_cleanup(obddev);
1014                 GOTO(err_thread, rc);
1015         }
1016
1017         obddev->obd_namespace =
1018                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1019         if (obddev->obd_namespace == NULL) {
1020                 LBUG();
1021                 mds_cleanup(obddev);
1022                 GOTO(err_thread, rc);
1023         }
1024
1025         mds->mds_local_namespace =
1026                 ldlm_namespace_new("mds_client", LDLM_NAMESPACE_CLIENT);
1027         if (mds->mds_local_namespace == NULL) {
1028                 LBUG();
1029                 mds_cleanup(obddev);
1030                 GOTO(err_thread, rc);
1031         }
1032
1033         OBD_ALLOC(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
1034         if (mds->mds_ldlm_client == NULL) {
1035                 LBUG();
1036                 mds_cleanup(obddev);
1037                 GOTO(err_thread, rc);
1038         }
1039         ptlrpc_init_client(NULL, NULL, LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
1040                            mds->mds_ldlm_client);
1041         mds->mds_ldlm_client->cli_target_devno = obddev->obd_minor;
1042         mds->mds_ldlm_client->cli_name = "mds ldlm";
1043
1044         rc = mds_recover(obddev);
1045         if (rc)
1046                 GOTO(err_thread, rc);
1047
1048         rc = class_connect(&mds->mds_connh, obddev);
1049         if (rc)
1050                 GOTO(err_thread, rc);
1051         export = class_conn2export(&mds->mds_connh);
1052         if (!export)
1053                 LBUG();
1054         export->exp_connection = mds->mds_ldlm_conn;
1055
1056         RETURN(0);
1057
1058
1059
1060 err_thread:
1061         ptlrpc_stop_all_threads(mds->mds_service);
1062 err_svc:
1063         ptlrpc_unregister_service(mds->mds_service);
1064 err_fs:
1065         mds_fs_cleanup(mds);
1066 err_put:
1067         unlock_kernel();
1068         mntput(mds->mds_vfsmnt);
1069         mds->mds_sb = 0;
1070         lock_kernel();
1071 err_kfree:
1072         kfree(mds->mds_fstype);
1073 err_dec:
1074         MOD_DEC_USE_COUNT;
1075         return rc;
1076 }
1077
1078 static int mds_cleanup(struct obd_device * obddev)
1079 {
1080         struct super_block *sb;
1081         struct mds_obd *mds = &obddev->u.mds;
1082
1083         ENTRY;
1084         class_disconnect(&mds->mds_connh);
1085
1086
1087         if ( !list_empty(&obddev->obd_exports) ) {
1088                 CERROR("still has exports!\n");
1089                 RETURN(-EBUSY);
1090         }
1091
1092         ptlrpc_stop_all_threads(mds->mds_service);
1093         ptlrpc_unregister_service(mds->mds_service);
1094
1095         sb = mds->mds_sb;
1096         if (!mds->mds_sb)
1097                 RETURN(0);
1098
1099         mds_update_server_data(mds);
1100
1101         if (mds->mds_rcvd_filp) {
1102                 int rc = filp_close(mds->mds_rcvd_filp, 0);
1103                 mds->mds_rcvd_filp = NULL;
1104
1105                 if (rc)
1106                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
1107         }
1108
1109         unlock_kernel();
1110         mntput(mds->mds_vfsmnt);
1111         mds->mds_sb = 0;
1112         kfree(mds->mds_fstype);
1113
1114         ldlm_namespace_free(mds->mds_local_namespace);
1115         ldlm_namespace_free(obddev->obd_namespace);
1116
1117         if (mds->mds_ldlm_conn != NULL)
1118                 ptlrpc_put_connection(mds->mds_ldlm_conn);
1119
1120         OBD_FREE(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
1121
1122         lock_kernel();
1123 #ifdef CONFIG_DEV_RDONLY
1124         dev_clear_rdonly(2);
1125 #endif
1126         mds_fs_cleanup(mds);
1127
1128         MOD_DEC_USE_COUNT;
1129         RETURN(0);
1130 }
1131
1132 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1133                           int len, void *karg, void *uarg);
1134
1135 /* use obd ops to offer management infrastructure */
1136 static struct obd_ops mds_obd_ops = {
1137         o_connect:     mds_connect,
1138         o_disconnect:  mds_disconnect,
1139         o_setup:       mds_setup,
1140         o_cleanup:     mds_cleanup,
1141         o_iocontrol:   mds_iocontrol
1142 };
1143
1144 static int __init mds_init(void)
1145 {
1146         inter_module_register("mds_reint", THIS_MODULE, &mds_reint);
1147         inter_module_register("mds_getattr_name", THIS_MODULE,
1148                               &mds_getattr_name);
1149         class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1150         return 0;
1151 }
1152
1153 static void __exit mds_exit(void)
1154 {
1155         inter_module_unregister("mds_reint");
1156         inter_module_unregister("mds_getattr_name");
1157         class_unregister_type(LUSTRE_MDS_NAME);
1158 }
1159
1160 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1161 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1162 MODULE_LICENSE("GPL");
1163
1164 module_init(mds_init);
1165 module_exit(mds_exit);