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 #include <linux/init.h>
  27
  28 extern int mds_get_lovtgts(struct obd_device *obd, int tgt_count,
  29                            uuid_t *uuidarray);
  30 extern int mds_get_lovdesc(struct obd_device *obd, struct lov_desc *desc);
  31 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
  32                                 struct ptlrpc_request *req);
  33 static int mds_cleanup(struct obd_device * obddev);
  34
  35 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
  36 {
  37         return &req->rq_export->exp_obd->u.mds;
  38 }
  39
  40 /* Assumes caller has already pushed into the kernel filesystem context */
  41 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  42                         __u64 offset)
  43 {
  44         int rc = 0;
  45         struct mds_obd *mds = mds_req2mds(req);
  46         struct ptlrpc_bulk_desc *desc;
  47         struct ptlrpc_bulk_page *bulk;
  48         char *buf;
  49         ENTRY;
  50
  51         desc = ptlrpc_prep_bulk(req->rq_connection);
  52         if (desc == NULL)
  53                 GOTO(out, rc = -ENOMEM);
  54
  55         bulk = ptlrpc_prep_bulk_page(desc);
  56         if (bulk == NULL)
  57                 GOTO(cleanup_bulk, rc = -ENOMEM);
  58
  59         OBD_ALLOC(buf, PAGE_SIZE);
  60         if (buf == NULL)
  61                 GOTO(cleanup_bulk, rc = -ENOMEM);
  62
  63         rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
  64
  65         if (rc != PAGE_SIZE)
  66                 GOTO(cleanup_buf, rc = -EIO);
  67
  68         bulk->b_xid = req->rq_xid;
  69         bulk->b_buf = buf;
  70         bulk->b_buflen = PAGE_SIZE;
  71         desc->b_portal = MDS_BULK_PORTAL;
  72
  73         rc = ptlrpc_send_bulk(desc);
  74         if (rc)
  75                 GOTO(cleanup_buf, rc);
  76
  77         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
  78                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
  79                        OBD_FAIL_MDS_SENDPAGE, rc);
  80                 ptlrpc_abort_bulk(desc);
  81                 GOTO(cleanup_buf, rc);
  82         }
  83
  84         wait_event(desc->b_waitq, ptlrpc_check_bulk_sent(desc));
  85         if (desc->b_flags & PTL_RPC_FL_INTR)
  86                 GOTO(cleanup_buf, rc = -EINTR);
  87
  88         EXIT;
  89  cleanup_buf:
  90         OBD_FREE(buf, PAGE_SIZE);
  91  cleanup_bulk:
  92         ptlrpc_free_bulk(desc);
  93  out:
  94         return rc;
  95 }
  96
  97 /* 'dir' is a inode for which a lock has already been taken */
  98 struct dentry *mds_name2locked_dentry(struct mds_obd *mds, struct dentry *dir,
  99                                       struct vfsmount **mnt, char *name,
 100                                       int namelen, int lock_mode,
 101                                       struct lustre_handle *lockh,
 102                                       int dir_lock_mode)
 103 {
 104         struct dentry *dchild;
 105         int flags, rc;
 106         __u64 res_id[3] = {0};
 107         ENTRY;
 108
 109         down(&dir->d_inode->i_sem);
 110         dchild = lookup_one_len(name, dir, namelen);
 111         if (IS_ERR(dchild)) {
 112                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 113                 up(&dir->d_inode->i_sem);
 114                 LBUG();
 115                 RETURN(dchild);
 116         }
 117         if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
 118                 up(&dir->d_inode->i_sem);
 119                 ldlm_lock_decref(lockh, dir_lock_mode);
 120         }
 121
 122         if (lock_mode == 0 || !dchild->d_inode)
 123                 RETURN(dchild);
 124
 125         res_id[0] = dchild->d_inode->i_ino;
 126         rc = ldlm_match_or_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
 127                                    (struct lustre_handle *)&mds->mds_connh,
 128                                    NULL, mds->mds_local_namespace, NULL,
 129                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 130                                    &flags, (void *)mds_lock_callback, NULL,
 131                                    0, lockh);
 132         if (rc != ELDLM_OK) {
 133                 l_dput(dchild);
 134                 up(&dir->d_inode->i_sem);
 135                 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
 136         }
 137
 138         RETURN(dchild);
 139 }
 140
 141 struct dentry *mds_fid2locked_dentry(struct mds_obd *mds, struct ll_fid *fid,
 142                                      struct vfsmount **mnt, int lock_mode,
 143                                      struct lustre_handle *lockh)
 144 {
 145         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 146         int flags, rc;
 147         __u64 res_id[3] = {0};
 148         ENTRY;
 149
 150         if (IS_ERR(de))
 151                 RETURN(de);
 152
 153         res_id[0] = de->d_inode->i_ino;
 154         rc = ldlm_match_or_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
 155                                    (struct lustre_handle *)&mds->mds_connh,
 156                                    NULL, mds->mds_local_namespace, NULL,
 157                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 158                                    &flags, (void *)mds_lock_callback, NULL,
 159                                    0, lockh);
 160         if (rc != ELDLM_OK) {
 161                 l_dput(de);
 162                 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
 163         }
 164
 165         RETURN(retval);
 166 }
 167
 168 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 169                               struct vfsmount **mnt)
 170 {
 171         /* stolen from NFS */
 172         struct super_block *sb = mds->mds_sb;
 173         unsigned long ino = fid->id;
 174         __u32 generation = fid->generation;
 175         struct inode *inode;
 176         struct list_head *lp;
 177         struct dentry *result;
 178
 179         if (ino == 0)
 180                 RETURN(ERR_PTR(-ESTALE));
 181
 182         inode = iget(sb, ino);
 183         if (inode == NULL)
 184                 RETURN(ERR_PTR(-ENOMEM));
 185
 186         CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
 187
 188         if (is_bad_inode(inode) ||
 189             (generation && inode->i_generation != generation)) {
 190                 /* we didn't find the right inode.. */
 191                 CERROR("bad inode %lu, link: %d ct: %d or version  %u/%u\n",
 192                        inode->i_ino, inode->i_nlink,
 193                        atomic_read(&inode->i_count), inode->i_generation,
 194                        generation);
 195                 LBUG();
 196                 iput(inode);
 197                 RETURN(ERR_PTR(-ESTALE));
 198         }
 199
 200         /* now to find a dentry.
 201          * If possible, get a well-connected one
 202          */
 203         if (mnt)
 204                 *mnt = mds->mds_vfsmnt;
 205         spin_lock(&dcache_lock);
 206         for (lp = inode->i_dentry.next; lp != &inode->i_dentry ; lp=lp->next) {
 207                 result = list_entry(lp,struct dentry, d_alias);
 208                 if (! (result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
 209                         dget_locked(result);
 210                         result->d_vfs_flags |= DCACHE_REFERENCED;
 211                         spin_unlock(&dcache_lock);
 212                         iput(inode);
 213                         if (mnt)
 214                                 mntget(*mnt);
 215                         return result;
 216                 }
 217         }
 218         spin_unlock(&dcache_lock);
 219         result = d_alloc_root(inode);
 220         if (result == NULL) {
 221                 iput(inode);
 222                 return ERR_PTR(-ENOMEM);
 223         }
 224         if (mnt)
 225                 mntget(*mnt);
 226         result->d_flags |= DCACHE_NFSD_DISCONNECTED;
 227         return result;
 228 }
 229
 230 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd)
 231 {
 232         int rc;
 233
 234         MOD_INC_USE_COUNT;
 235         rc = class_connect(conn, obd);
 236
 237         if (rc)
 238                 MOD_DEC_USE_COUNT;
 239
 240         return rc;
 241 }
 242
 243 static int mds_disconnect(struct lustre_handle *conn)
 244 {
 245         int rc;
 246
 247         rc = class_disconnect(conn);
 248         if (!rc)
 249                 MOD_DEC_USE_COUNT;
 250
 251         return rc;
 252 }
 253
 254 /* FIXME: the error cases need fixing to avoid leaks */
 255 static int mds_getstatus(struct ptlrpc_request *req)
 256 {
 257         struct mds_obd *mds = mds_req2mds(req);
 258         struct mds_body *body;
 259         struct mds_client_info *mci;
 260         struct mds_client_data *mcd;
 261         int rc, size = sizeof(*body);
 262         ENTRY;
 263
 264         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 265         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 266                 CERROR("mds: out of memory for message: size=%d\n", size);
 267                 req->rq_status = -ENOMEM;
 268                 RETURN(0);
 269         }
 270
 271         body = lustre_msg_buf(req->rq_reqmsg, 0);
 272         mds_unpack_body(body);
 273
 274         /* Anything we need to do here with the client's trans no or so? */
 275         body = lustre_msg_buf(req->rq_repmsg, 0);
 276         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 277
 278         mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
 279         if (!mci) {
 280                 /* We don't have any old connection data for this client */
 281                 int rc;
 282
 283                 CDEBUG(D_INFO, "allocating new client data for UUID '%s'",
 284                        ptlrpc_req_to_uuid(req));
 285
 286                 OBD_ALLOC(mcd, sizeof(*mcd));
 287                 if (!mcd) {
 288                         CERROR("mds: out of memory for client data\n");
 289                         req->rq_status = -ENOMEM;
 290                         RETURN(0);
 291                 }
 292                 memcpy(mcd->mcd_uuid, ptlrpc_req_to_uuid(req),
 293                        sizeof(mcd->mcd_uuid));
 294                 rc = mds_client_add(mds, mcd, -1);
 295                 if (rc) {
 296                         req->rq_status = rc;
 297                         RETURN(0);
 298                 }
 299         } else {
 300                 /* We have old connection data for this client... */
 301                 mcd = mci->mci_mcd;
 302                 CDEBUG(D_INFO, "found existing data for UUID '%s' at #%d\n",
 303                        mcd->mcd_uuid, mci->mci_off);
 304         }
 305         /* mcd_last_xid is is stored in little endian on the disk and
 306            mds_pack_rep_body converts it to network order */
 307         body->last_xid = le32_to_cpu(mcd->mcd_last_xid);
 308         mds_pack_rep_body(req);
 309         RETURN(0);
 310 }
 311
 312 static int mds_getlovinfo(struct ptlrpc_request *req)
 313 {
 314         struct mds_obd *mds = mds_req2mds(req);
 315         struct mds_status_req *streq;
 316         struct lov_desc *desc;
 317         int tgt_count;
 318         int rc, size[2] = {sizeof(*desc)};
 319         ENTRY;
 320
 321         streq = lustre_msg_buf(req->rq_reqmsg, 0);
 322         streq->flags = NTOH__u32(streq->flags);
 323         streq->repbuf = NTOH__u32(streq->repbuf);
 324         size[1] = streq->repbuf;
 325
 326         rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
 327         if (rc) {
 328                 CERROR("mds: out of memory for message: size=%d\n", size[1]);
 329                 req->rq_status = -ENOMEM;
 330                 RETURN(0);
 331         }
 332
 333         desc = lustre_msg_buf(req->rq_repmsg, 0);
 334         rc = mds_get_lovdesc(req->rq_obd, desc);
 335         if (rc != 0 ) {
 336                 CERROR("mds_get_lovdesc error %d", rc);
 337                 req->rq_status = rc;
 338                 RETURN(0);
 339         }
 340
 341         tgt_count = NTOH__u32(desc->ld_tgt_count);
 342         if (tgt_count * sizeof(uuid_t) > streq->repbuf) {
 343                 CERROR("too many targets, enlarge client buffers\n");
 344                 req->rq_status = -ENOSPC;
 345                 RETURN(0);
 346         }
 347
 348         mds->mds_max_mdsize = sizeof(desc) + tgt_count * sizeof(uuid_t);
 349         rc = mds_get_lovtgts(req->rq_obd, tgt_count,
 350                              lustre_msg_buf(req->rq_repmsg, 1));
 351         if (rc) {
 352                 CERROR("get_lovtgts error %d\n", rc);
 353                 req->rq_status = rc;
 354                 RETURN(0);
 355         }
 356         RETURN(0);
 357 }
 358
 359 int mds_lock_callback(struct lustre_handle *lockh, struct ldlm_lock_desc *desc,
 360                       void *data, int data_len, struct ptlrpc_request **reqp)
 361 {
 362         ENTRY;
 363
 364         if (desc == NULL) {
 365                 /* Completion AST.  Do nothing */
 366                 RETURN(0);
 367         }
 368
 369         if (ldlm_cli_cancel(lockh) < 0)
 370                 LBUG();
 371         RETURN(0);
 372 }
 373
 374 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 375 {
 376         struct mds_obd *mds = mds_req2mds(req);
 377         struct obd_run_ctxt saved;
 378         struct mds_body *body;
 379         struct dentry *de = NULL, *dchild = NULL;
 380         struct inode *dir;
 381         struct lustre_handle lockh;
 382         char *name;
 383         int namelen, flags, lock_mode, rc = 0;
 384         __u64 res_id[3] = {0, 0, 0};
 385         ENTRY;
 386
 387         if (strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds") != 0)
 388                 LBUG();
 389
 390         if (req->rq_reqmsg->bufcount <= offset + 1) {
 391                 LBUG();
 392                 GOTO(out_pre_de, rc = -EINVAL);
 393         }
 394
 395         body = lustre_msg_buf(req->rq_reqmsg, offset);
 396         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 397         namelen = req->rq_reqmsg->buflens[offset + 1];
 398         /* requests were at offset 2, replies go back at 1 */
 399         if (offset)
 400                 offset = 1;
 401
 402         push_ctxt(&saved, &mds->mds_ctxt);
 403         de = mds_fid2dentry(mds, &body->fid1, NULL);
 404         if (IS_ERR(de)) {
 405                 LBUG();
 406                 GOTO(out_pre_de, rc = -ESTALE);
 407         }
 408
 409         dir = de->d_inode;
 410         CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
 411
 412         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 413         res_id[0] = dir->i_ino;
 414
 415         rc = ldlm_lock_match(mds->mds_local_namespace, res_id, LDLM_PLAIN,
 416                              NULL, 0, lock_mode, &lockh);
 417         if (rc == 0) {
 418                 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
 419                 rc = ldlm_cli_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
 420                                       (struct lustre_handle *)&mds->mds_connh,
 421                                       NULL, mds->mds_local_namespace, NULL,
 422                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 423                                       &flags, (void *)mds_lock_callback,
 424                                       NULL, 0, &lockh);
 425                 if (rc != ELDLM_OK) {
 426                         CERROR("lock enqueue: err: %d\n", rc);
 427                         GOTO(out_create_de, rc = -EIO);
 428                 }
 429         }
 430         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 431
 432         down(&dir->i_sem);
 433         dchild = lookup_one_len(name, de, namelen - 1);
 434         if (IS_ERR(dchild)) {
 435                 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
 436                 up(&dir->i_sem);
 437                 LBUG();
 438                 GOTO(out_create_dchild, rc = -ESTALE);
 439         }
 440
 441         if (dchild->d_inode) {
 442                 struct mds_body *body;
 443                 struct inode *inode = dchild->d_inode;
 444                 CDEBUG(D_INODE, "child exists (dir %ld, name %s, ino %ld)\n",
 445                        dir->i_ino, name, dchild->d_inode->i_ino);
 446
 447                 body = lustre_msg_buf(req->rq_repmsg, offset);
 448                 mds_pack_inode2fid(&body->fid1, inode);
 449                 mds_pack_inode2body(body, inode);
 450                 if (S_ISREG(inode->i_mode)) {
 451                         struct lov_stripe_md *md;
 452                         md = lustre_msg_buf(req->rq_repmsg, offset + 1);
 453                         md->lmd_size = mds->mds_max_mdsize;
 454                         mds_fs_get_md(mds, inode, md);
 455                 }
 456                 /* now a normal case for intent locking */
 457                 rc = 0;
 458         } else
 459                 rc = -ENOENT;
 460
 461         EXIT;
 462 out_create_dchild:
 463         l_dput(dchild);
 464         up(&dir->i_sem);
 465         ldlm_lock_decref(&lockh, lock_mode);
 466 out_create_de:
 467         l_dput(de);
 468 out_pre_de:
 469         req->rq_status = rc;
 470         pop_ctxt(&saved);
 471         return 0;
 472 }
 473
 474
 475 static int mds_getattr(int offset, struct ptlrpc_request *req)
 476 {
 477         struct mds_obd *mds = mds_req2mds(req);
 478         struct obd_run_ctxt saved;
 479         struct dentry *de;
 480         struct inode *inode;
 481         struct mds_body *body;
 482         int rc, size[2] = {sizeof(*body)}, bufcount = 1;
 483         ENTRY;
 484
 485         body = lustre_msg_buf(req->rq_reqmsg, offset);
 486         push_ctxt(&saved, &mds->mds_ctxt);
 487         de = mds_fid2dentry(mds, &body->fid1, NULL);
 488         if (IS_ERR(de)) {
 489                 GOTO(out_pop, rc = -ENOENT);
 490         }
 491
 492         inode = de->d_inode;
 493         if (S_ISREG(body->fid1.f_type)) {
 494                 bufcount = 2;
 495                 size[1] = mds->mds_max_mdsize;
 496         } else if (body->valid & OBD_MD_LINKNAME) {
 497                 bufcount = 2;
 498                 size[1] = inode->i_size;
 499         }
 500
 501         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 502                              &req->rq_repmsg);
 503         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 504                 CERROR("mds: out of memory\n");
 505                 GOTO(out, rc);
 506         }
 507
 508         if (body->valid & OBD_MD_LINKNAME) {
 509                 char *tmp = lustre_msg_buf(req->rq_repmsg, 1);
 510
 511                 rc = inode->i_op->readlink(de, tmp, size[1]);
 512
 513                 if (rc < 0) {
 514                         CERROR("readlink failed: %d\n", rc);
 515                         GOTO(out, rc);
 516                 }
 517         }
 518
 519         body = lustre_msg_buf(req->rq_repmsg, 0);
 520         body->ino = inode->i_ino;
 521         body->generation = inode->i_generation;
 522         body->atime = inode->i_atime;
 523         body->ctime = inode->i_ctime;
 524         body->mtime = inode->i_mtime;
 525         body->uid = inode->i_uid;
 526         body->gid = inode->i_gid;
 527         body->size = inode->i_size;
 528         body->mode = inode->i_mode;
 529         body->nlink = inode->i_nlink;
 530         body->valid = ~0; /* FIXME: should be more selective */
 531
 532         if (S_ISREG(inode->i_mode)) {
 533                 rc = mds_fs_get_md(mds, inode,
 534                                      lustre_msg_buf(req->rq_repmsg, 1));
 535                 if (rc < 0) {
 536                         CERROR("mds_fs_get_md failed: %d\n", rc);
 537                         GOTO(out, rc);
 538                 }
 539         }
 540 out:
 541         l_dput(de);
 542 out_pop:
 543         pop_ctxt(&saved);
 544         req->rq_status = rc;
 545         RETURN(0);
 546 }
 547
 548 static int mds_statfs(struct ptlrpc_request *req)
 549 {
 550         struct mds_obd *mds = mds_req2mds(req);
 551         struct obd_statfs *osfs;
 552         struct statfs sfs;
 553         int rc, size = sizeof(*osfs);
 554         ENTRY;
 555
 556         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 557                              &req->rq_repmsg);
 558         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 559                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 560                 GOTO(out, rc);
 561         }
 562
 563         rc = vfs_statfs(mds->mds_sb, &sfs);
 564         if (rc) {
 565                 CERROR("mds: statfs failed: rc %d\n", rc);
 566                 GOTO(out, rc);
 567         }
 568         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 569         memset(osfs, 0, size);
 570         obd_statfs_pack(osfs, &sfs);
 571
 572 out:
 573         req->rq_status = rc;
 574         RETURN(0);
 575 }
 576
 577 static int mds_open(struct ptlrpc_request *req)
 578 {
 579         struct dentry *de;
 580         struct mds_body *body;
 581         struct file *file;
 582         struct vfsmount *mnt;
 583         struct mds_obd *mds = mds_req2mds(req);
 584         struct mds_client_info *mci;
 585         __u32 flags;
 586         struct list_head *tmp;
 587         struct mds_file_data *mfd;
 588         int rc, size = sizeof(*body);
 589         ENTRY;
 590
 591         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 592         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 593                 CERROR("mds: out of memory\n");
 594                 req->rq_status = -ENOMEM;
 595                 RETURN(0);
 596         }
 597
 598         mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
 599         if (!mci) {
 600                 CERROR("mds: no mci!\n");
 601                 req->rq_status = -ENOTCONN;
 602                 RETURN(0);
 603         }
 604
 605         body = lustre_msg_buf(req->rq_reqmsg, 0);
 606
 607         /* was this animal open already? */
 608         /* XXX we should only check on re-open, or do a refcount... */
 609         list_for_each(tmp, &mci->mci_open_head) {
 610                 struct mds_file_data *fd;
 611                 fd = list_entry(tmp, struct mds_file_data, mfd_list);
 612                 if (body->extra == fd->mfd_clientfd &&
 613                     body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
 614                         CERROR("Re opening %Ld\n", body->fid1.id);
 615                         RETURN(0);
 616                 }
 617         }
 618
 619         OBD_ALLOC(mfd, sizeof(*mfd));
 620         if (!mfd) {
 621                 CERROR("mds: out of memory\n");
 622                 req->rq_status = -ENOMEM;
 623                 RETURN(0);
 624         }
 625
 626         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 627         if (IS_ERR(de)) {
 628                 req->rq_status = -ENOENT;
 629                 RETURN(0);
 630         }
 631
 632         /* check if this inode has seen a delayed object creation */
 633         if (req->rq_reqmsg->bufcount > 1) {
 634                 void *handle;
 635                 struct lov_stripe_md *md;
 636                 struct inode *inode = de->d_inode;
 637                 //struct iattr iattr;
 638                 int rc;
 639
 640                 md = lustre_msg_buf(req->rq_reqmsg, 1);
 641                 //iattr.ia_mode = inode->i_mode;
 642
 643                 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 644                 if (!handle) {
 645                         req->rq_status = -ENOMEM;
 646                         RETURN(0);
 647                 }
 648
 649                 /* XXX error handling */
 650                 rc = mds_fs_set_md(mds, inode, handle, md);
 651                 //                rc = mds_fs_setattr(mds, de, handle, &iattr);
 652                 if (!rc) {
 653                         struct obd_run_ctxt saved;
 654                         push_ctxt(&saved, &mds->mds_ctxt);
 655                         rc = mds_update_last_rcvd(mds, handle, req);
 656                         pop_ctxt(&saved);
 657                 } else {
 658                         req->rq_status = rc;
 659                         RETURN(0);
 660                 }
 661                 /* FIXME: need to return last_rcvd, last_committed */
 662
 663                 /* FIXME: keep rc intact */
 664                 rc = mds_fs_commit(mds, de->d_inode, handle);
 665                 if (rc) {
 666                         req->rq_status = rc;
 667                         RETURN(0);
 668                 }
 669         }
 670
 671         flags = body->flags;
 672         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 673         if (!file || IS_ERR(file)) {
 674                 req->rq_status = -EINVAL;
 675                 OBD_FREE(mfd, sizeof(*mfd));
 676                 RETURN(0);
 677         }
 678
 679         file->private_data = mfd;
 680         mfd->mfd_file = file;
 681         mfd->mfd_clientfd = body->extra;
 682         list_add(&mfd->mfd_list, &mci->mci_open_head);
 683
 684         body = lustre_msg_buf(req->rq_repmsg, 0);
 685         body->extra = (__u64) (unsigned long)file;
 686         RETURN(0);
 687 }
 688
 689 static int mds_close(struct ptlrpc_request *req)
 690 {
 691         struct dentry *de;
 692         struct mds_body *body;
 693         struct file *file;
 694         struct mds_obd *mds = mds_req2mds(req);
 695         struct vfsmount *mnt;
 696         struct mds_file_data *mfd;
 697         int rc;
 698         ENTRY;
 699
 700         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 701         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 702                 CERROR("mds: out of memory\n");
 703                 req->rq_status = -ENOMEM;
 704                 RETURN(0);
 705         }
 706
 707         body = lustre_msg_buf(req->rq_reqmsg, 0);
 708         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 709         if (IS_ERR(de)) {
 710                 req->rq_status = -ENOENT;
 711                 RETURN(0);
 712         }
 713
 714         file = (struct file *)(unsigned long)body->extra;
 715         if (!file->f_dentry)
 716                 LBUG();
 717         mfd = (struct mds_file_data *)file->private_data;
 718         list_del(&mfd->mfd_list);
 719         OBD_FREE(mfd, sizeof(*mfd));
 720
 721         req->rq_status = filp_close(file, 0);
 722         l_dput(de);
 723         mntput(mnt);
 724
 725         RETURN(0);
 726 }
 727
 728 static int mds_readpage(struct ptlrpc_request *req)
 729 {
 730         struct mds_obd *mds = mds_req2mds(req);
 731         struct vfsmount *mnt;
 732         struct dentry *de;
 733         struct file *file;
 734         struct mds_body *body;
 735         struct obd_run_ctxt saved;
 736         int rc, size = sizeof(*body);
 737         ENTRY;
 738
 739         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 740         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 741                 CERROR("mds: out of memory\n");
 742                 GOTO(out, rc = -ENOMEM);
 743         }
 744
 745         body = lustre_msg_buf(req->rq_reqmsg, 0);
 746         push_ctxt(&saved, &mds->mds_ctxt);
 747         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 748         if (IS_ERR(de))
 749                 GOTO(out_pop, rc = PTR_ERR(de));
 750
 751         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 752
 753         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 754         /* note: in case of an error, dentry_open puts dentry */
 755         if (IS_ERR(file))
 756                 GOTO(out_pop, rc = PTR_ERR(file));
 757
 758         /* to make this asynchronous make sure that the handling function
 759            doesn't send a reply when this function completes. Instead a
 760            callback function would send the reply */
 761         rc = mds_sendpage(req, file, body->size);
 762
 763         filp_close(file, 0);
 764 out_pop:
 765         pop_ctxt(&saved);
 766 out:
 767         req->rq_status = rc;
 768         RETURN(0);
 769 }
 770
 771 int mds_reint(int offset, struct ptlrpc_request *req)
 772 {
 773         int rc;
 774         struct mds_update_record rec;
 775
 776         rc = mds_update_unpack(req, offset, &rec);
 777         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 778                 CERROR("invalid record\n");
 779                 req->rq_status = -EINVAL;
 780                 RETURN(0);
 781         }
 782         /* rc will be used to interrupt a for loop over multiple records */
 783         rc = mds_reint_rec(&rec, offset, req);
 784         return rc;
 785 }
 786
 787 int mds_handle(struct ptlrpc_request *req)
 788 {
 789         int rc;
 790         ENTRY;
 791
 792         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 793         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
 794                 CERROR("lustre_mds: Invalid request\n");
 795                 GOTO(out, rc);
 796         }
 797
 798         if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 799                 CERROR("lustre_mds: wrong packet type sent %d\n",
 800                        req->rq_reqmsg->type);
 801                 GOTO(out, rc = -EINVAL);
 802         }
 803
 804         if (req->rq_reqmsg->opc != MDS_CONNECT &&
 805             req->rq_export == NULL)
 806                 GOTO(out, rc = -ENOTCONN);
 807
 808         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 809                 GOTO(out, rc = -EINVAL);
 810
 811         switch (req->rq_reqmsg->opc) {
 812         case MDS_CONNECT:
 813                 CDEBUG(D_INODE, "connect\n");
 814                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
 815                 rc = target_handle_connect(req);
 816                 break;
 817
 818         case MDS_DISCONNECT:
 819                 CDEBUG(D_INODE, "disconnect\n");
 820                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
 821                 rc = target_handle_disconnect(req);
 822                 goto out;
 823
 824         case MDS_GETSTATUS:
 825                 CDEBUG(D_INODE, "getstatus\n");
 826                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
 827                 rc = mds_getstatus(req);
 828                 break;
 829
 830         case MDS_GETLOVINFO:
 831                 CDEBUG(D_INODE, "getlovinfo\n");
 832                 rc = mds_getlovinfo(req);
 833                 break;
 834
 835         case MDS_GETATTR:
 836                 CDEBUG(D_INODE, "getattr\n");
 837                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 838                 rc = mds_getattr(0, req);
 839                 break;
 840
 841         case MDS_STATFS:
 842                 CDEBUG(D_INODE, "statfs\n");
 843                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
 844                 rc = mds_statfs(req);
 845                 break;
 846
 847         case MDS_READPAGE:
 848                 CDEBUG(D_INODE, "readpage\n");
 849                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 850                 rc = mds_readpage(req);
 851
 852                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 853                         return 0;
 854                 break;
 855
 856         case MDS_REINT: {
 857                 int size = sizeof(struct mds_body);
 858                 CDEBUG(D_INODE, "reint\n");
 859                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 860
 861                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 862                                      &req->rq_repmsg);
 863                 if (rc) {
 864                         rc = req->rq_status = -ENOMEM;
 865                         break;
 866                 }
 867                 rc = mds_reint(0, req);
 868                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
 869                 break;
 870         }
 871
 872         case MDS_OPEN:
 873                 CDEBUG(D_INODE, "open\n");
 874                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 875                 rc = mds_open(req);
 876                 break;
 877
 878         case MDS_CLOSE:
 879                 CDEBUG(D_INODE, "close\n");
 880                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 881                 rc = mds_close(req);
 882                 break;
 883
 884         default:
 885                 rc = ptlrpc_error(req->rq_svc, req);
 886                 RETURN(rc);
 887         }
 888
 889         EXIT;
 890
 891         if (!rc) {
 892                 struct mds_obd *mds = mds_req2mds(req);
 893                 req->rq_repmsg->last_rcvd = HTON__u64(mds->mds_last_rcvd);
 894                 req->rq_repmsg->last_committed =
 895                         HTON__u64(mds->mds_last_committed);
 896                 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
 897                        (unsigned long long)mds->mds_last_rcvd,
 898                        (unsigned long long)mds->mds_last_committed,
 899                        cpu_to_le32(req->rq_xid));
 900         }
 901  out:
 902         /* Still not 100% sure whether we should reply with the server
 903          * last_rcvd or that of this client.  I'm not sure it even makes
 904          * a difference on a per-client basis, because last_rcvd is global
 905          * and we are not supposed to allow transactions while in recovery.
 906          */
 907         if (rc) {
 908                 CERROR("mds: processing error %d\n", rc);
 909                 ptlrpc_error(req->rq_svc, req);
 910         } else {
 911                 CDEBUG(D_NET, "sending reply\n");
 912                 ptlrpc_reply(req->rq_svc, req);
 913         }
 914         return 0;
 915 }
 916
 917 /* Update the server data on disk.  This stores the new mount_count and
 918  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
 919  * then the server last_rcvd value may be less than that of the clients.
 920  * This will alert us that we may need to do client recovery.
 921  */
 922 static
 923 int mds_update_server_data(struct mds_obd *mds)
 924 {
 925         struct obd_run_ctxt saved;
 926         struct mds_server_data *msd = mds->mds_server_data;
 927         struct file *filp = mds->mds_rcvd_filp;
 928         loff_t off = 0;
 929         int rc;
 930
 931         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
 932         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
 933
 934         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
 935                (unsigned long long)mds->mds_mount_count,
 936                (unsigned long long)mds->mds_last_rcvd);
 937         push_ctxt(&saved, &mds->mds_ctxt);
 938         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
 939         if (rc != sizeof(*msd)) {
 940                 CERROR("error writing MDS server data: rc = %d\n", rc);
 941                 if (rc > 0)
 942                         RETURN(-EIO);
 943                 RETURN(rc);
 944         }
 945         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
 946         pop_ctxt(&saved);
 947         if (rc)
 948                 CERROR("error flushing MDS server data: rc = %d\n", rc);
 949
 950         return 0;
 951 }
 952
 953 /* Do recovery actions for the MDS */
 954 static int mds_recover(struct obd_device *obddev)
 955 {
 956         struct mds_obd *mds = &obddev->u.mds;
 957         int rc;
 958
 959         /* This happens at the end when recovery is complete */
 960         ++mds->mds_mount_count;
 961         rc = mds_update_server_data(mds);
 962
 963         return rc;
 964 }
 965
 966
 967 /* mount the file system (secretly) */
 968 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
 969 {
 970         struct obd_ioctl_data* data = buf;
 971         struct obd_export *export;
 972         struct mds_obd *mds = &obddev->u.mds;
 973         struct vfsmount *mnt;
 974         int rc = 0;
 975         ENTRY;
 976
 977         MOD_INC_USE_COUNT;
 978 #ifdef CONFIG_DEV_RDONLY
 979         dev_clear_rdonly(2);
 980 #endif
 981         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
 982                 GOTO(err_dec, rc = -EINVAL);
 983
 984         mds->mds_fstype = strdup(data->ioc_inlbuf2);
 985
 986         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
 987         if (IS_ERR(mnt)) {
 988                 rc = PTR_ERR(mnt);
 989                 CERROR("do_kern_mount failed: rc = %d\n", rc);
 990                 GOTO(err_kfree, rc);
 991         }
 992
 993         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
 994         if (!mds->mds_sb)
 995                 GOTO(err_put, rc = -ENODEV);
 996
 997         mds->mds_max_mdsize = sizeof(struct lov_stripe_md);
 998         rc = mds_fs_setup(mds, mnt);
 999         if (rc) {
1000                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1001                 GOTO(err_put, rc);
1002         }
1003
1004         mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
1005                                            MDC_REPLY_PORTAL, "self",mds_handle);
1006         if (!mds->mds_service) {
1007                 CERROR("failed to start service\n");
1008                 GOTO(err_fs, rc = -EINVAL);
1009         }
1010
1011         rc = ptlrpc_start_thread(obddev, mds->mds_service, "lustre_mds");
1012         if (rc) {
1013                 CERROR("cannot start thread: rc = %d\n", rc);
1014                 GOTO(err_svc, rc);
1015         }
1016
1017         rc = -ENOENT;
1018         mds->mds_ldlm_conn = ptlrpc_uuid_to_connection("self");
1019         if (!mds->mds_ldlm_conn) {
1020                 mds_cleanup(obddev);
1021                 GOTO(err_thread, rc);
1022         }
1023
1024         obddev->obd_namespace =
1025                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1026         if (obddev->obd_namespace == NULL) {
1027                 LBUG();
1028                 mds_cleanup(obddev);
1029                 GOTO(err_thread, rc);
1030         }
1031
1032         mds->mds_local_namespace =
1033                 ldlm_namespace_new("mds_client", LDLM_NAMESPACE_CLIENT);
1034         if (mds->mds_local_namespace == NULL) {
1035                 LBUG();
1036                 mds_cleanup(obddev);
1037                 GOTO(err_thread, rc);
1038         }
1039
1040         OBD_ALLOC(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
1041         if (mds->mds_ldlm_client == NULL) {
1042                 LBUG();
1043                 mds_cleanup(obddev);
1044                 GOTO(err_thread, rc);
1045         }
1046         ptlrpc_init_client(NULL, NULL, LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
1047                            mds->mds_ldlm_client);
1048         mds->mds_ldlm_client->cli_target_devno = obddev->obd_minor;
1049         mds->mds_ldlm_client->cli_name = "mds ldlm";
1050
1051         rc = mds_recover(obddev);
1052         if (rc)
1053                 GOTO(err_thread, rc);
1054
1055         rc = class_connect(&mds->mds_connh, obddev);
1056         if (rc)
1057                 GOTO(err_thread, rc);
1058         export = class_conn2export(&mds->mds_connh);
1059         if (!export)
1060                 LBUG();
1061         export->exp_connection = mds->mds_ldlm_conn;
1062
1063         RETURN(0);
1064
1065
1066
1067 err_thread:
1068         ptlrpc_stop_all_threads(mds->mds_service);
1069 err_svc:
1070         ptlrpc_unregister_service(mds->mds_service);
1071 err_fs:
1072         mds_fs_cleanup(mds);
1073 err_put:
1074         unlock_kernel();
1075         mntput(mds->mds_vfsmnt);
1076         mds->mds_sb = 0;
1077         lock_kernel();
1078 err_kfree:
1079         kfree(mds->mds_fstype);
1080 err_dec:
1081         MOD_DEC_USE_COUNT;
1082         return rc;
1083 }
1084
1085 static int mds_cleanup(struct obd_device * obddev)
1086 {
1087         struct super_block *sb;
1088         struct mds_obd *mds = &obddev->u.mds;
1089
1090         ENTRY;
1091         class_disconnect(&mds->mds_connh);
1092
1093
1094         if ( !list_empty(&obddev->obd_exports) ) {
1095                 CERROR("still has exports!\n");
1096                 RETURN(-EBUSY);
1097         }
1098
1099         ptlrpc_stop_all_threads(mds->mds_service);
1100         ptlrpc_unregister_service(mds->mds_service);
1101
1102         sb = mds->mds_sb;
1103         if (!mds->mds_sb)
1104                 RETURN(0);
1105
1106         mds_update_server_data(mds);
1107
1108         if (mds->mds_rcvd_filp) {
1109                 int rc = filp_close(mds->mds_rcvd_filp, 0);
1110                 mds->mds_rcvd_filp = NULL;
1111
1112                 if (rc)
1113                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
1114         }
1115
1116         unlock_kernel();
1117         mntput(mds->mds_vfsmnt);
1118         mds->mds_sb = 0;
1119         kfree(mds->mds_fstype);
1120
1121         ldlm_namespace_free(mds->mds_local_namespace);
1122         ldlm_namespace_free(obddev->obd_namespace);
1123
1124         if (mds->mds_ldlm_conn != NULL)
1125                 ptlrpc_put_connection(mds->mds_ldlm_conn);
1126
1127         OBD_FREE(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
1128
1129         lock_kernel();
1130 #ifdef CONFIG_DEV_RDONLY
1131         dev_clear_rdonly(2);
1132 #endif
1133         mds_fs_cleanup(mds);
1134
1135         MOD_DEC_USE_COUNT;
1136         RETURN(0);
1137 }
1138
1139 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1140                           int len, void *karg, void *uarg);
1141
1142 /* use obd ops to offer management infrastructure */
1143 static struct obd_ops mds_obd_ops = {
1144         o_connect:     mds_connect,
1145         o_disconnect:  mds_disconnect,
1146         o_setup:       mds_setup,
1147         o_cleanup:     mds_cleanup,
1148         o_iocontrol:   mds_iocontrol
1149 };
1150
1151 static int __init mds_init(void)
1152 {
1153         inter_module_register("mds_reint", THIS_MODULE, &mds_reint);
1154         inter_module_register("mds_getattr_name", THIS_MODULE,
1155                               &mds_getattr_name);
1156         class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1157         return 0;
1158 }
1159
1160 static void __exit mds_exit(void)
1161 {
1162         inter_module_unregister("mds_reint");
1163         inter_module_unregister("mds_getattr_name");
1164         class_unregister_type(LUSTRE_MDS_NAME);
1165 }
1166
1167 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1168 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1169 MODULE_LICENSE("GPL");
1170
1171 module_init(mds_init);
1172 module_exit(mds_exit);