lustre/mds/handler.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  lustre/mds/handler.c
   5  *  Lustre Metadata Server (mds) request handler
   6  *
   7  *  Copyright (c) 2001, 2002 Cluster File Systems, Inc.
   8  *   Author: Peter Braam <braam@clusterfs.com>
   9  *   Author: Andreas Dilger <adilger@clusterfs.com>
  10  *   Author: Phil Schwan <phil@clusterfs.com>
  11  *
  12  *   This file is part of Lustre, http://www.lustre.org.
  13  *
  14  *   Lustre is free software; you can redistribute it and/or
  15  *   modify it under the terms of version 2 of the GNU General Public
  16  *   License as published by the Free Software Foundation.
  17  *
  18  *   Lustre is distributed in the hope that it will be useful,
  19  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  20  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  21  *   GNU General Public License for more details.
  22  *
  23  *   You should have received a copy of the GNU General Public License
  24  *   along with Lustre; if not, write to the Free Software
  25  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  26  */
  27
  28 #define EXPORT_SYMTAB
  29 #define DEBUG_SUBSYSTEM S_MDS
  30
  31 #include <linux/module.h>
  32 #include <linux/lustre_mds.h>
  33 #include <linux/lustre_dlm.h>
  34 #include <linux/init.h>
  35 #include <linux/obd_class.h>
  36
  37 extern int mds_get_lovtgts(struct obd_device *obd, int tgt_count,
  38                            uuid_t *uuidarray);
  39 extern int mds_get_lovdesc(struct obd_device *obd, struct lov_desc *desc);
  40 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
  41                                 struct ptlrpc_request *req);
  42 static int mds_cleanup(struct obd_device * obddev);
  43
  44 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
  45 {
  46         return &req->rq_export->exp_obd->u.mds;
  47 }
  48
  49 static int mds_bulk_timeout(void *data)
  50 {
  51         struct ptlrpc_bulk_desc *desc = data;
  52
  53         ENTRY;
  54         CERROR("(not yet) starting recovery of client %p\n", desc->bd_client);
  55         RETURN(1);
  56 }
  57
  58 /* Assumes caller has already pushed into the kernel filesystem context */
  59 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
  60                         __u64 offset)
  61 {
  62         int rc = 0;
  63         struct mds_obd *mds = mds_req2mds(req);
  64         struct ptlrpc_bulk_desc *desc;
  65         struct ptlrpc_bulk_page *bulk;
  66         struct l_wait_info lwi;
  67         char *buf;
  68         ENTRY;
  69
  70         desc = ptlrpc_prep_bulk(req->rq_connection);
  71         if (desc == NULL)
  72                 GOTO(out, rc = -ENOMEM);
  73
  74         bulk = ptlrpc_prep_bulk_page(desc);
  75         if (bulk == NULL)
  76                 GOTO(cleanup_bulk, rc = -ENOMEM);
  77
  78         OBD_ALLOC(buf, PAGE_SIZE);
  79         if (buf == NULL)
  80                 GOTO(cleanup_bulk, rc = -ENOMEM);
  81
  82         rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
  83
  84         if (rc != PAGE_SIZE)
  85                 GOTO(cleanup_buf, rc = -EIO);
  86
  87         bulk->bp_xid = req->rq_xid;
  88         bulk->bp_buf = buf;
  89         bulk->bp_buflen = PAGE_SIZE;
  90         desc->bd_portal = MDS_BULK_PORTAL;
  91
  92         rc = ptlrpc_send_bulk(desc);
  93         if (rc)
  94                 GOTO(cleanup_buf, rc);
  95
  96         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
  97                 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
  98                        OBD_FAIL_MDS_SENDPAGE, rc);
  99                 ptlrpc_abort_bulk(desc);
 100                 GOTO(cleanup_buf, rc);
 101         }
 102
 103         lwi = LWI_TIMEOUT(obd_timeout * HZ, mds_bulk_timeout, desc);
 104         rc = l_wait_event(desc->bd_waitq, desc->bd_flags & PTL_BULK_FL_SENT, &lwi);
 105         if (rc) {
 106                 if (rc != -ETIMEDOUT)
 107                         LBUG();
 108                 GOTO(cleanup_buf, rc);
 109         }
 110
 111         EXIT;
 112  cleanup_buf:
 113         OBD_FREE(buf, PAGE_SIZE);
 114  cleanup_bulk:
 115         ptlrpc_free_bulk(desc);
 116  out:
 117         return rc;
 118 }
 119
 120 /*
 121  * Look up a named entry in a directory, and get an LDLM lock on it.
 122  * 'dir' is a inode for which an LDLM lock has already been taken.
 123  *
 124  * If we do not need an exclusive or write lock on this entry (e.g.
 125  * a read lock for attribute lookup only) then we do not hold the
 126  * directory on return.  It is up to the caller to know what type
 127  * of lock it is getting, and clean up appropriately.
 128  */
 129 struct dentry *mds_name2locked_dentry(struct obd_device *obd,
 130                                       struct dentry *dir, struct vfsmount **mnt,
 131                                       char *name, int namelen, int lock_mode,
 132                                       struct lustre_handle *lockh,
 133                                       int dir_lock_mode)
 134 {
 135         struct dentry *dchild;
 136         int flags, rc;
 137         __u64 res_id[3] = {0};
 138         ENTRY;
 139
 140         down(&dir->d_inode->i_sem);
 141         dchild = lookup_one_len(name, dir, namelen);
 142         if (IS_ERR(dchild)) {
 143                 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
 144                 up(&dir->d_inode->i_sem);
 145                 LBUG();
 146                 RETURN(dchild);
 147         }
 148         if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
 149                 up(&dir->d_inode->i_sem);
 150                 ldlm_lock_decref(lockh, dir_lock_mode);
 151         }
 152
 153         if (lock_mode == 0 || !dchild->d_inode)
 154                 RETURN(dchild);
 155
 156         res_id[0] = dchild->d_inode->i_ino;
 157         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 158                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 159                                    &flags, ldlm_completion_ast,
 160                                    mds_blocking_ast, NULL, 0, lockh);
 161         if (rc != ELDLM_OK) {
 162                 l_dput(dchild);
 163                 up(&dir->d_inode->i_sem);
 164                 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
 165         }
 166
 167         RETURN(dchild);
 168 }
 169
 170 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 171                                      struct vfsmount **mnt, int lock_mode,
 172                                      struct lustre_handle *lockh)
 173 {
 174         struct mds_obd *mds = &obd->u.mds;
 175         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
 176         int flags, rc;
 177         __u64 res_id[3] = {0};
 178         ENTRY;
 179
 180         if (IS_ERR(de))
 181                 RETURN(de);
 182
 183         res_id[0] = de->d_inode->i_ino;
 184         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 185                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 186                                    &flags, ldlm_completion_ast,
 187                                    mds_blocking_ast, NULL, 0, lockh);
 188         if (rc != ELDLM_OK) {
 189                 l_dput(de);
 190                 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
 191         }
 192
 193         RETURN(retval);
 194 }
 195
 196 /* Look up an entry by inode number. */
 197 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
 198                               struct vfsmount **mnt)
 199 {
 200         /* stolen from NFS */
 201         struct super_block *sb = mds->mds_sb;
 202         unsigned long ino = fid->id;
 203         __u32 generation = fid->generation;
 204         struct inode *inode;
 205         struct list_head *lp;
 206         struct dentry *result;
 207
 208         if (ino == 0)
 209                 RETURN(ERR_PTR(-ESTALE));
 210
 211         inode = iget(sb, ino);
 212         if (inode == NULL)
 213                 RETURN(ERR_PTR(-ENOMEM));
 214
 215         CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
 216
 217         if (is_bad_inode(inode) ||
 218             (generation && inode->i_generation != generation)) {
 219                 /* we didn't find the right inode.. */
 220                 CERROR("bad inode %lu, link: %d ct: %d or version  %u/%u\n",
 221                        inode->i_ino, inode->i_nlink,
 222                        atomic_read(&inode->i_count), inode->i_generation,
 223                        generation);
 224                 LBUG();
 225                 iput(inode);
 226                 RETURN(ERR_PTR(-ESTALE));
 227         }
 228
 229         /* now to find a dentry. If possible, get a well-connected one */
 230         if (mnt)
 231                 *mnt = mds->mds_vfsmnt;
 232         spin_lock(&dcache_lock);
 233         list_for_each(lp, &inode->i_dentry) {
 234                 result = list_entry(lp, struct dentry, d_alias);
 235                 if (!(result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
 236                         dget_locked(result);
 237                         result->d_vfs_flags |= DCACHE_REFERENCED;
 238                         spin_unlock(&dcache_lock);
 239                         iput(inode);
 240                         if (mnt)
 241                                 mntget(*mnt);
 242                         return result;
 243                 }
 244         }
 245         spin_unlock(&dcache_lock);
 246         result = d_alloc_root(inode);
 247         if (result == NULL) {
 248                 iput(inode);
 249                 return ERR_PTR(-ENOMEM);
 250         }
 251         if (mnt)
 252                 mntget(*mnt);
 253         result->d_flags |= DCACHE_NFSD_DISCONNECTED;
 254         return result;
 255 }
 256
 257 /* Establish a connection to the MDS.
 258  *
 259  * This will set up an export structure for the client to hold state data
 260  * about that client, like open files, the last operation number it did
 261  * on the server, etc.
 262  */
 263 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
 264                        char *cluuid)
 265 {
 266         struct obd_export *exp;
 267         struct mds_client_data *mcd;
 268         struct list_head *p;
 269         int rc;
 270         ENTRY;
 271
 272         if (!conn || !obd || !cluuid)
 273                 RETURN(-EINVAL);
 274
 275         MOD_INC_USE_COUNT;
 276
 277         spin_lock(&obd->obd_dev_lock);
 278         list_for_each(p, &obd->obd_exports) {
 279                 exp = list_entry(p, struct obd_export, exp_obd_chain);
 280                 mcd = exp->exp_mds_data.med_mcd;
 281                 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
 282                         LASSERT(exp->exp_obd == obd);
 283
 284                         if (!list_empty(&exp->exp_conn_chain)) {
 285                                 CERROR("existing uuid/export, list not empty!\n");
 286                                 spin_unlock(&obd->obd_dev_lock);
 287                                 RETURN(-EALREADY);
 288                         }
 289                         conn->addr = (__u64) (unsigned long)exp;
 290                         conn->cookie = exp->exp_cookie;
 291                         spin_unlock(&obd->obd_dev_lock);
 292                         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 293                                cluuid, exp);
 294                         CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
 295                                (long long)conn->addr, (long long)conn->cookie);
 296                         RETURN(0);
 297                 }
 298         }
 299         spin_unlock(&obd->obd_dev_lock);
 300         /* XXX There is a small race between checking the list and adding a
 301          * new connection for the same UUID, but the real threat (list
 302          * corruption when multiple different clients connect) is solved.
 303          */
 304         rc = class_connect(conn, obd, cluuid);
 305         if (rc)
 306                 GOTO(out_dec, rc);
 307         exp = class_conn2export(conn);
 308         LASSERT(exp);
 309
 310         OBD_ALLOC(mcd, sizeof(*mcd));
 311         if (!mcd) {
 312                 CERROR("mds: out of memory for client data\n");
 313                 GOTO(out_export, rc = -ENOMEM);
 314         }
 315         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 316         exp->exp_mds_data.med_mcd = mcd;
 317         rc = mds_client_add(&exp->exp_mds_data, -1);
 318         if (rc)
 319                 GOTO(out_mdc, rc);
 320
 321         RETURN(0);
 322
 323 out_mdc:
 324         OBD_FREE(mcd, sizeof(*mcd));
 325 out_export:
 326         class_disconnect(conn);
 327 out_dec:
 328         MOD_DEC_USE_COUNT;
 329
 330         return rc;
 331 }
 332
 333 static int mds_disconnect(struct lustre_handle *conn)
 334 {
 335         int rc;
 336
 337         rc = class_disconnect(conn);
 338         if (!rc)
 339                 MOD_DEC_USE_COUNT;
 340
 341         return rc;
 342 }
 343
 344 static int mds_getstatus(struct ptlrpc_request *req)
 345 {
 346         struct mds_obd *mds = mds_req2mds(req);
 347         struct mds_body *body;
 348         int rc, size = sizeof(*body);
 349         ENTRY;
 350
 351         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 352         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 353                 CERROR("mds: out of memory for message: size=%d\n", size);
 354                 req->rq_status = -ENOMEM;
 355                 RETURN(0);
 356         }
 357
 358         /* Anything we need to do here with the client's trans no or so? */
 359         body = lustre_msg_buf(req->rq_repmsg, 0);
 360         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 361
 362         /* the last_committed and last_xid fields are filled in for all
 363          * replies already - no need to do so here also.
 364          */
 365         RETURN(0);
 366 }
 367
 368 static int mds_getlovinfo(struct ptlrpc_request *req)
 369 {
 370         struct mds_obd *mds = mds_req2mds(req);
 371         struct mds_status_req *streq;
 372         struct lov_desc *desc;
 373         int tgt_count;
 374         int rc, size[2] = {sizeof(*desc)};
 375         ENTRY;
 376
 377         streq = lustre_msg_buf(req->rq_reqmsg, 0);
 378         streq->flags = NTOH__u32(streq->flags);
 379         streq->repbuf = NTOH__u32(streq->repbuf);
 380         size[1] = streq->repbuf;
 381
 382         rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
 383         if (rc) {
 384                 CERROR("mds: out of memory for message: size=%d\n", size[1]);
 385                 req->rq_status = -ENOMEM;
 386                 RETURN(0);
 387         }
 388
 389         desc = lustre_msg_buf(req->rq_repmsg, 0);
 390         rc = mds_get_lovdesc(req->rq_obd, desc);
 391         if (rc != 0 ) {
 392                 CERROR("mds_get_lovdesc error %d", rc);
 393                 req->rq_status = rc;
 394                 RETURN(0);
 395         }
 396
 397         tgt_count = NTOH__u32(desc->ld_tgt_count);
 398         if (tgt_count * sizeof(uuid_t) > streq->repbuf) {
 399                 CERROR("too many targets, enlarge client buffers\n");
 400                 req->rq_status = -ENOSPC;
 401                 RETURN(0);
 402         }
 403
 404         mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
 405                 tgt_count * sizeof(struct lov_object_id);
 406         rc = mds_get_lovtgts(req->rq_obd, tgt_count,
 407                              lustre_msg_buf(req->rq_repmsg, 1));
 408         if (rc) {
 409                 CERROR("get_lovtgts error %d\n", rc);
 410                 req->rq_status = rc;
 411                 RETURN(0);
 412         }
 413         RETURN(0);
 414 }
 415
 416 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 417                      void *data, __u32 data_len)
 418 {
 419         int do_ast;
 420         ENTRY;
 421
 422         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 423         lock->l_flags |= LDLM_FL_CBPENDING;
 424         do_ast = (!lock->l_readers && !lock->l_writers);
 425         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 426
 427         if (do_ast) {
 428                 struct lustre_handle lockh;
 429                 int rc;
 430
 431                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 432                 ldlm_lock2handle(lock, &lockh);
 433                 rc = ldlm_cli_cancel(&lockh);
 434                 if (rc < 0)
 435                         CERROR("ldlm_cli_cancel: %d\n", rc);
 436         } else
 437                 LDLM_DEBUG(lock, "Lock still has references, will be"
 438                            "cancelled later");
 439         RETURN(0);
 440 }
 441
 442 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
 443                                 struct ptlrpc_request *req,
 444                                 struct mds_body *reqbody, int reply_off)
 445 {
 446         struct mds_body *body;
 447         struct inode *inode = dentry->d_inode;
 448         int rc;
 449         ENTRY;
 450
 451         if (inode == NULL)
 452                 RETURN(-ENOENT);
 453
 454         body = lustre_msg_buf(req->rq_repmsg, reply_off);
 455
 456         mds_pack_inode2fid(&body->fid1, inode);
 457         mds_pack_inode2body(body, inode);
 458
 459         if (S_ISREG(inode->i_mode)) {
 460                 struct lov_mds_md *md;
 461
 462                 md = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 463                 md->lmd_easize = mds->mds_max_mdsize;
 464                 rc = mds_fs_get_md(mds, inode, md);
 465
 466                 if (rc < 0) {
 467                         if (rc == -ENODATA)
 468                                 RETURN(0);
 469                         CERROR("mds_fs_get_md failed: %d\n", rc);
 470                         RETURN(rc);
 471                 }
 472                 body->valid |= OBD_MD_FLEASIZE;
 473         } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
 474                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 475                 int len = req->rq_repmsg->buflens[reply_off + 1];
 476
 477                 rc = inode->i_op->readlink(dentry, symname, len);
 478                 if (rc < 0) {
 479                         CERROR("readlink failed: %d\n", rc);
 480                         RETURN(rc);
 481                 } else
 482                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 483
 484                 body->valid |= OBD_MD_LINKNAME;
 485         }
 486         RETURN(0);
 487 }
 488
 489 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 490 {
 491         struct mds_obd *mds = mds_req2mds(req);
 492         struct obd_device *obd = req->rq_export->exp_obd;
 493         struct obd_run_ctxt saved;
 494         struct mds_body *body;
 495         struct dentry *de = NULL, *dchild = NULL;
 496         struct inode *dir;
 497         struct lustre_handle lockh;
 498         char *name;
 499         int namelen, flags, lock_mode, rc = 0;
 500         __u64 res_id[3] = {0, 0, 0};
 501         ENTRY;
 502
 503         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 504
 505         if (req->rq_reqmsg->bufcount <= offset + 1) {
 506                 LBUG();
 507                 GOTO(out_pre_de, rc = -EINVAL);
 508         }
 509
 510         body = lustre_msg_buf(req->rq_reqmsg, offset);
 511         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 512         namelen = req->rq_reqmsg->buflens[offset + 1];
 513         /* requests were at offset 2, replies go back at 1 */
 514         if (offset)
 515                 offset = 1;
 516
 517         push_ctxt(&saved, &mds->mds_ctxt);
 518         de = mds_fid2dentry(mds, &body->fid1, NULL);
 519         if (IS_ERR(de)) {
 520                 LBUG();
 521                 GOTO(out_pre_de, rc = -ESTALE);
 522         }
 523
 524         dir = de->d_inode;
 525         CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
 526
 527         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 528         res_id[0] = dir->i_ino;
 529
 530         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 531                              NULL, 0, lock_mode, &lockh);
 532         if (rc == 0) {
 533                 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
 534                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 535                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 536                                       &flags, ldlm_completion_ast,
 537                                       mds_blocking_ast, NULL, 0, &lockh);
 538                 if (rc != ELDLM_OK) {
 539                         CERROR("lock enqueue: err: %d\n", rc);
 540                         GOTO(out_create_de, rc = -EIO);
 541                 }
 542         }
 543         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 544
 545         down(&dir->i_sem);
 546         dchild = lookup_one_len(name, de, namelen - 1);
 547         if (IS_ERR(dchild)) {
 548                 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
 549                 up(&dir->i_sem);
 550                 LBUG();
 551                 GOTO(out_create_dchild, rc = -ESTALE);
 552         }
 553
 554         rc = mds_getattr_internal(mds, dchild, req, body, offset);
 555
 556         EXIT;
 557 out_create_dchild:
 558         l_dput(dchild);
 559         up(&dir->i_sem);
 560         ldlm_lock_decref(&lockh, lock_mode);
 561 out_create_de:
 562         l_dput(de);
 563 out_pre_de:
 564         req->rq_status = rc;
 565         pop_ctxt(&saved);
 566         return 0;
 567 }
 568
 569 static int mds_getattr(int offset, struct ptlrpc_request *req)
 570 {
 571         struct mds_obd *mds = mds_req2mds(req);
 572         struct obd_run_ctxt saved;
 573         struct dentry *de;
 574         struct inode *inode;
 575         struct mds_body *body;
 576         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 577         ENTRY;
 578
 579         body = lustre_msg_buf(req->rq_reqmsg, offset);
 580         push_ctxt(&saved, &mds->mds_ctxt);
 581         de = mds_fid2dentry(mds, &body->fid1, NULL);
 582         if (IS_ERR(de)) {
 583                 req->rq_status = -ENOENT;
 584                 GOTO(out_pop, rc = -ENOENT);
 585         }
 586
 587         inode = de->d_inode;
 588         if (S_ISREG(body->fid1.f_type)) {
 589                 bufcount = 2;
 590                 size[1] = mds->mds_max_mdsize;
 591         } else if (body->valid & OBD_MD_LINKNAME) {
 592                 bufcount = 2;
 593                 size[1] = inode->i_size + 1;
 594         }
 595
 596         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 597                              &req->rq_repmsg);
 598         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 599                 CERROR("out of memory or FAIL_MDS_GETATTR_PACK\n");
 600                 req->rq_status = rc;
 601                 GOTO(out, rc = 0);
 602         }
 603
 604         req->rq_status = mds_getattr_internal(mds, de, req, body, 0);
 605
 606 out:
 607         l_dput(de);
 608 out_pop:
 609         pop_ctxt(&saved);
 610         RETURN(rc);
 611 }
 612
 613 static int mds_statfs(struct ptlrpc_request *req)
 614 {
 615         struct mds_obd *mds = mds_req2mds(req);
 616         struct obd_statfs *osfs;
 617         struct statfs sfs;
 618         int rc, size = sizeof(*osfs);
 619         ENTRY;
 620
 621         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 622                              &req->rq_repmsg);
 623         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 624                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 625                 GOTO(out, rc);
 626         }
 627
 628         rc = vfs_statfs(mds->mds_sb, &sfs);
 629         if (rc) {
 630                 CERROR("mds: statfs failed: rc %d\n", rc);
 631                 GOTO(out, rc);
 632         }
 633         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 634         memset(osfs, 0, size);
 635         statfs_pack(osfs, &sfs);
 636         obd_statfs_pack(osfs, osfs);
 637
 638 out:
 639         req->rq_status = rc;
 640         RETURN(0);
 641 }
 642
 643 static int mds_open(struct ptlrpc_request *req)
 644 {
 645         struct dentry *de;
 646         struct mds_body *body;
 647         struct file *file;
 648         struct vfsmount *mnt;
 649         struct mds_obd *mds = mds_req2mds(req);
 650         struct mds_export_data *med;
 651         __u32 flags;
 652         struct list_head *tmp;
 653         struct mds_file_data *mfd;
 654         int rc, size = sizeof(*body);
 655         ENTRY;
 656
 657         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 658         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 659                 CERROR("mds: out of memory\n");
 660                 req->rq_status = -ENOMEM;
 661                 RETURN(0);
 662         }
 663
 664         body = lustre_msg_buf(req->rq_reqmsg, 0);
 665
 666         /* was this animal open already? */
 667         /* XXX we should only check on re-open, or do a refcount... */
 668         med = &req->rq_export->exp_mds_data;
 669         list_for_each(tmp, &med->med_open_head) {
 670                 struct mds_file_data *fd;
 671                 fd = list_entry(tmp, struct mds_file_data, mfd_list);
 672                 if (body->extra == fd->mfd_clientfd &&
 673                     body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
 674                         CERROR("Re opening "LPD64"\n", body->fid1.id);
 675                         RETURN(0);
 676                 }
 677         }
 678
 679         OBD_ALLOC(mfd, sizeof(*mfd));
 680         if (!mfd) {
 681                 CERROR("mds: out of memory\n");
 682                 req->rq_status = -ENOMEM;
 683                 RETURN(0);
 684         }
 685
 686         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 687         if (IS_ERR(de)) {
 688                 req->rq_status = -ENOENT;
 689                 RETURN(0);
 690         }
 691
 692         /* check if this inode has seen a delayed object creation */
 693         if (req->rq_reqmsg->bufcount > 1) {
 694                 void *handle;
 695                 struct lov_mds_md *md;
 696                 struct inode *inode = de->d_inode;
 697                 int rc;
 698
 699                 md = lustre_msg_buf(req->rq_reqmsg, 1);
 700
 701                 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 702                 if (!handle) {
 703                         req->rq_status = -ENOMEM;
 704                         RETURN(0);
 705                 }
 706
 707                 /* XXX error handling */
 708                 rc = mds_fs_set_md(mds, inode, handle, md);
 709                 if (!rc) {
 710                         struct obd_run_ctxt saved;
 711                         push_ctxt(&saved, &mds->mds_ctxt);
 712                         rc = mds_update_last_rcvd(mds, handle, req);
 713                         pop_ctxt(&saved);
 714                 } else {
 715                         req->rq_status = rc;
 716                         RETURN(0);
 717                 }
 718                 /* FIXME: need to return last_rcvd, last_committed */
 719
 720                 /* FIXME: keep rc intact */
 721                 rc = mds_fs_commit(mds, de->d_inode, handle);
 722                 if (rc) {
 723                         req->rq_status = rc;
 724                         RETURN(0);
 725                 }
 726         }
 727
 728         flags = body->flags;
 729         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 730         if (!file || IS_ERR(file)) {
 731                 req->rq_status = -EINVAL;
 732                 OBD_FREE(mfd, sizeof(*mfd));
 733                 RETURN(0);
 734         }
 735
 736         file->private_data = mfd;
 737         mfd->mfd_file = file;
 738         mfd->mfd_clientfd = body->extra;
 739         list_add(&mfd->mfd_list, &med->med_open_head);
 740
 741         body = lustre_msg_buf(req->rq_repmsg, 0);
 742         /* FIXME: need to have cookies involved here */
 743         body->extra = (__u64) (unsigned long)file;
 744         RETURN(0);
 745 }
 746
 747 static int mds_close(struct ptlrpc_request *req)
 748 {
 749         struct dentry *de;
 750         struct mds_body *body;
 751         struct file *file;
 752         struct mds_obd *mds = mds_req2mds(req);
 753         struct vfsmount *mnt;
 754         struct mds_file_data *mfd;
 755         int rc;
 756         ENTRY;
 757
 758         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 759         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 760                 CERROR("mds: out of memory\n");
 761                 req->rq_status = -ENOMEM;
 762                 RETURN(0);
 763         }
 764
 765         body = lustre_msg_buf(req->rq_reqmsg, 0);
 766         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 767         if (IS_ERR(de)) {
 768                 req->rq_status = -ENOENT;
 769                 RETURN(0);
 770         }
 771
 772         /* FIXME: need to have cookies involved here */
 773         file = (struct file *)(unsigned long)body->extra;
 774         if (!file->f_dentry)
 775                 LBUG();
 776         mfd = (struct mds_file_data *)file->private_data;
 777         list_del(&mfd->mfd_list);
 778         OBD_FREE(mfd, sizeof(*mfd));
 779
 780         req->rq_status = filp_close(file, 0);
 781         l_dput(de);
 782         mntput(mnt);
 783
 784         RETURN(0);
 785 }
 786
 787 static int mds_readpage(struct ptlrpc_request *req)
 788 {
 789         struct mds_obd *mds = mds_req2mds(req);
 790         struct vfsmount *mnt;
 791         struct dentry *de;
 792         struct file *file;
 793         struct mds_body *body;
 794         struct obd_run_ctxt saved;
 795         int rc, size = sizeof(*body);
 796         ENTRY;
 797
 798         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 799         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 800                 CERROR("mds: out of memory\n");
 801                 GOTO(out, rc = -ENOMEM);
 802         }
 803
 804         body = lustre_msg_buf(req->rq_reqmsg, 0);
 805         push_ctxt(&saved, &mds->mds_ctxt);
 806         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 807         if (IS_ERR(de))
 808                 GOTO(out_pop, rc = PTR_ERR(de));
 809
 810         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 811
 812         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 813         /* note: in case of an error, dentry_open puts dentry */
 814         if (IS_ERR(file))
 815                 GOTO(out_pop, rc = PTR_ERR(file));
 816
 817         /* to make this asynchronous make sure that the handling function
 818            doesn't send a reply when this function completes. Instead a
 819            callback function would send the reply */
 820         rc = mds_sendpage(req, file, body->size);
 821
 822         filp_close(file, 0);
 823 out_pop:
 824         pop_ctxt(&saved);
 825 out:
 826         req->rq_status = rc;
 827         RETURN(0);
 828 }
 829
 830 int mds_reint(int offset, struct ptlrpc_request *req)
 831 {
 832         int rc;
 833         struct mds_update_record rec;
 834
 835         rc = mds_update_unpack(req, offset, &rec);
 836         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 837                 CERROR("invalid record\n");
 838                 req->rq_status = -EINVAL;
 839                 RETURN(0);
 840         }
 841         /* rc will be used to interrupt a for loop over multiple records */
 842         rc = mds_reint_rec(&rec, offset, req);
 843         return rc;
 844 }
 845
 846 int mds_handle(struct ptlrpc_request *req)
 847 {
 848         int rc;
 849         ENTRY;
 850
 851         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 852         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
 853                 CERROR("lustre_mds: Invalid request\n");
 854                 GOTO(out, rc);
 855         }
 856
 857         if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
 858                 GOTO(out, rc = -ENOTCONN);
 859
 860         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 861                 GOTO(out, rc = -EINVAL);
 862
 863         switch (req->rq_reqmsg->opc) {
 864         case MDS_CONNECT:
 865                 CDEBUG(D_INODE, "connect\n");
 866                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
 867                 rc = target_handle_connect(req);
 868                 break;
 869
 870         case MDS_DISCONNECT:
 871                 CDEBUG(D_INODE, "disconnect\n");
 872                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
 873                 rc = target_handle_disconnect(req);
 874                 goto out;
 875
 876         case MDS_GETSTATUS:
 877                 CDEBUG(D_INODE, "getstatus\n");
 878                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
 879                 rc = mds_getstatus(req);
 880                 break;
 881
 882         case MDS_GETLOVINFO:
 883                 CDEBUG(D_INODE, "getlovinfo\n");
 884                 rc = mds_getlovinfo(req);
 885                 break;
 886
 887         case MDS_GETATTR:
 888                 CDEBUG(D_INODE, "getattr\n");
 889                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 890                 rc = mds_getattr(0, req);
 891                 break;
 892
 893         case MDS_STATFS:
 894                 CDEBUG(D_INODE, "statfs\n");
 895                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
 896                 rc = mds_statfs(req);
 897                 break;
 898
 899         case MDS_READPAGE:
 900                 CDEBUG(D_INODE, "readpage\n");
 901                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 902                 rc = mds_readpage(req);
 903
 904                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 905                         return 0;
 906                 break;
 907
 908         case MDS_REINT: {
 909                 int size = sizeof(struct mds_body);
 910                 CDEBUG(D_INODE, "reint\n");
 911                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 912
 913                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 914                                      &req->rq_repmsg);
 915                 if (rc) {
 916                         req->rq_status = rc;
 917                         break;
 918                 }
 919                 rc = mds_reint(0, req);
 920                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
 921                 break;
 922                 }
 923
 924         case MDS_OPEN:
 925                 CDEBUG(D_INODE, "open\n");
 926                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 927                 rc = mds_open(req);
 928                 break;
 929
 930         case MDS_CLOSE:
 931                 CDEBUG(D_INODE, "close\n");
 932                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 933                 rc = mds_close(req);
 934                 break;
 935
 936         case LDLM_ENQUEUE:
 937                 CDEBUG(D_INODE, "enqueue\n");
 938                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
 939                 rc = ldlm_handle_enqueue(req);
 940                 if (rc)
 941                         break;
 942                 RETURN(0);
 943
 944         case LDLM_CONVERT:
 945                 CDEBUG(D_INODE, "convert\n");
 946                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
 947                 rc = ldlm_handle_convert(req);
 948                 if (rc)
 949                         break;
 950                 RETURN(0);
 951
 952         case LDLM_CANCEL:
 953                 CDEBUG(D_INODE, "cancel\n");
 954                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
 955                 rc = ldlm_handle_cancel(req);
 956                 if (rc)
 957                         break;
 958                 RETURN(0);
 959         case LDLM_BL_CALLBACK:
 960         case LDLM_CP_CALLBACK:
 961                 CDEBUG(D_INODE, "callback\n");
 962                 CERROR("callbacks should not happen on MDS\n");
 963                 LBUG();
 964                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 965                 break;
 966         default:
 967                 rc = ptlrpc_error(req->rq_svc, req);
 968                 RETURN(rc);
 969         }
 970
 971         EXIT;
 972
 973         if (!rc) {
 974                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
 975                 struct mds_obd *mds = mds_req2mds(req);
 976
 977                 req->rq_repmsg->last_xid =
 978                         HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
 979                 req->rq_repmsg->last_committed =
 980                         HTON__u64(mds->mds_last_committed);
 981                 CDEBUG(D_INFO, "last_rcvd ~%Lu, last_committed %Lu, xid %d\n",
 982                        (unsigned long long)mds->mds_last_rcvd,
 983                        (unsigned long long)mds->mds_last_committed,
 984                        cpu_to_le32(req->rq_xid));
 985         }
 986  out:
 987         if (rc) {
 988                 CERROR("mds: processing error %d\n", rc);
 989                 ptlrpc_error(req->rq_svc, req);
 990         } else {
 991                 CDEBUG(D_NET, "sending reply\n");
 992                 ptlrpc_reply(req->rq_svc, req);
 993         }
 994         return 0;
 995 }
 996
 997 /* Update the server data on disk.  This stores the new mount_count and
 998  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
 999  * then the server last_rcvd value may be less than that of the clients.
1000  * This will alert us that we may need to do client recovery.
1001  *
1002  * Assumes we are already in the server filesystem context.
1003  *
1004  * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
1005  */
1006 static
1007 int mds_update_server_data(struct mds_obd *mds)
1008 {
1009         struct mds_server_data *msd = mds->mds_server_data;
1010         struct file *filp = mds->mds_rcvd_filp;
1011         loff_t off = 0;
1012         int rc;
1013
1014         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1015         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1016
1017         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1018                (unsigned long long)mds->mds_mount_count,
1019                (unsigned long long)mds->mds_last_rcvd);
1020         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1021         if (rc != sizeof(*msd)) {
1022                 CERROR("error writing MDS server data: rc = %d\n", rc);
1023                 if (rc > 0)
1024                         RETURN(-EIO);
1025                 RETURN(rc);
1026         }
1027         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1028         if (rc)
1029                 CERROR("error flushing MDS server data: rc = %d\n", rc);
1030
1031         return 0;
1032 }
1033
1034 /* Do recovery actions for the MDS */
1035 static int mds_recover(struct obd_device *obddev)
1036 {
1037         struct mds_obd *mds = &obddev->u.mds;
1038         struct obd_run_ctxt saved;
1039         int rc;
1040
1041         /* This happens at the end when recovery is complete */
1042         ++mds->mds_mount_count;
1043         push_ctxt(&saved, &mds->mds_ctxt);
1044         rc = mds_update_server_data(mds);
1045         pop_ctxt(&saved);
1046
1047         return rc;
1048 }
1049
1050 #define MDS_NUM_THREADS 8
1051 /* mount the file system (secretly) */
1052 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1053 {
1054         int i;
1055         struct obd_ioctl_data* data = buf;
1056         struct mds_obd *mds = &obddev->u.mds;
1057         struct vfsmount *mnt;
1058         int rc = 0;
1059         ENTRY;
1060
1061         MOD_INC_USE_COUNT;
1062 #ifdef CONFIG_DEV_RDONLY
1063         dev_clear_rdonly(2);
1064 #endif
1065         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1066                 GOTO(err_dec, rc = -EINVAL);
1067
1068         mds->mds_fstype = strdup(data->ioc_inlbuf2);
1069
1070         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1071         if (IS_ERR(mnt)) {
1072                 rc = PTR_ERR(mnt);
1073                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1074                 GOTO(err_kfree, rc);
1075         }
1076
1077         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1078         if (!mds->mds_sb)
1079                 GOTO(err_put, rc = -ENODEV);
1080
1081         spin_lock_init(&mds->mds_last_lock);
1082         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1083         rc = mds_fs_setup(obddev, mnt);
1084         if (rc) {
1085                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1086                 GOTO(err_put, rc);
1087         }
1088
1089         mds->mds_service = ptlrpc_init_svc(1024, 640, MDS_REQUEST_PORTAL,
1090                                            MDC_REPLY_PORTAL, "self",mds_handle,
1091                                            "mds");
1092         if (!mds->mds_service) {
1093                 CERROR("failed to start service\n");
1094                 GOTO(err_fs, rc = -EINVAL);
1095         }
1096
1097         obddev->obd_namespace =
1098                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1099         if (obddev->obd_namespace == NULL) {
1100                 mds_cleanup(obddev);
1101                 GOTO(err_svc, rc = -ENOMEM);
1102         }
1103
1104         for (i = 0; i < MDS_NUM_THREADS; i++) {
1105                 char name[32];
1106                 sprintf(name, "lustre_MDS_%02d", i);
1107                 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1108                 if (rc) {
1109                         CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1110                         GOTO(err_thread, rc);
1111                 }
1112         }
1113
1114         rc = mds_recover(obddev);
1115         if (rc)
1116                 GOTO(err_thread, rc);
1117
1118         mds_destroy_export = mds_client_free;
1119
1120         ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
1121                            "mds_ldlm_client", &obddev->obd_ldlm_client);
1122
1123         RETURN(0);
1124
1125 err_thread:
1126         ptlrpc_stop_all_threads(mds->mds_service);
1127 err_svc:
1128         ptlrpc_unregister_service(mds->mds_service);
1129 err_fs:
1130         mds_fs_cleanup(obddev);
1131 err_put:
1132         unlock_kernel();
1133         mntput(mds->mds_vfsmnt);
1134         mds->mds_sb = 0;
1135         lock_kernel();
1136 err_kfree:
1137         kfree(mds->mds_fstype);
1138 err_dec:
1139         MOD_DEC_USE_COUNT;
1140         return rc;
1141 }
1142
1143 static int mds_cleanup(struct obd_device *obddev)
1144 {
1145         struct super_block *sb;
1146         struct mds_obd *mds = &obddev->u.mds;
1147         struct obd_run_ctxt saved;
1148         ENTRY;
1149
1150         ptlrpc_stop_all_threads(mds->mds_service);
1151         ptlrpc_unregister_service(mds->mds_service);
1152
1153         sb = mds->mds_sb;
1154         if (!mds->mds_sb)
1155                 RETURN(0);
1156
1157         push_ctxt(&saved, &mds->mds_ctxt);
1158         mds_update_server_data(mds);
1159
1160         if (mds->mds_rcvd_filp) {
1161                 int rc = filp_close(mds->mds_rcvd_filp, 0);
1162                 mds->mds_rcvd_filp = NULL;
1163
1164                 if (rc)
1165                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
1166         }
1167         pop_ctxt(&saved);
1168
1169         unlock_kernel();
1170         mntput(mds->mds_vfsmnt);
1171         mds->mds_sb = 0;
1172         kfree(mds->mds_fstype);
1173
1174         ldlm_namespace_free(obddev->obd_namespace);
1175
1176         lock_kernel();
1177 #ifdef CONFIG_DEV_RDONLY
1178         dev_clear_rdonly(2);
1179 #endif
1180         mds_fs_cleanup(obddev);
1181
1182         MOD_DEC_USE_COUNT;
1183         RETURN(0);
1184 }
1185
1186 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1187                               ldlm_mode_t mode, void *data)
1188 {
1189         struct ptlrpc_request *req = req_cookie;
1190         int rc = 0;
1191         ENTRY;
1192
1193         if (!req_cookie)
1194                 RETURN(0);
1195
1196         if (req->rq_reqmsg->bufcount > 1) {
1197                 /* an intent needs to be considered */
1198                 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1199                 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1200                 struct mds_body *mds_rep;
1201                 struct ldlm_reply *rep;
1202                 __u64 new_resid[3] = {0, 0, 0}, old_res;
1203                 int rc, size[3] = {sizeof(struct ldlm_reply),
1204                                                   sizeof(struct mds_body),
1205                                                   mds->mds_max_mdsize};
1206
1207                 it->opc = NTOH__u64(it->opc);
1208
1209                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1210                            ldlm_it2str(it->opc));
1211
1212                 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1213                                      &req->rq_repmsg);
1214                 if (rc) {
1215                         rc = req->rq_status = -ENOMEM;
1216                         RETURN(rc);
1217                 }
1218
1219                 rep = lustre_msg_buf(req->rq_repmsg, 0);
1220                 rep->lock_policy_res1 = 1;
1221
1222                 /* execute policy */
1223                 switch ((long)it->opc) {
1224                 case IT_CREAT|IT_OPEN:
1225                         rc = mds_reint(2, req);
1226                         if (rc || (req->rq_status != 0 &&
1227                                    req->rq_status != -EEXIST)) {
1228                                 rep->lock_policy_res2 = req->rq_status;
1229                                 RETURN(ELDLM_LOCK_ABORTED);
1230                         }
1231                         break;
1232                 case IT_CREAT:
1233                 case IT_LINK:
1234                 case IT_MKDIR:
1235                 case IT_MKNOD:
1236                 case IT_RENAME2:
1237                 case IT_RMDIR:
1238                 case IT_SYMLINK:
1239                 case IT_UNLINK:
1240                         rc = mds_reint(2, req);
1241                         if (rc || (req->rq_status != 0 &&
1242                                    req->rq_status != -EISDIR &&
1243                                    req->rq_status != -ENOTDIR)) {
1244                                 rep->lock_policy_res2 = req->rq_status;
1245                                 RETURN(ELDLM_LOCK_ABORTED);
1246                         }
1247                         break;
1248                 case IT_GETATTR:
1249                 case IT_LOOKUP:
1250                 case IT_OPEN:
1251                 case IT_READDIR:
1252                 case IT_READLINK:
1253                 case IT_RENAME:
1254                 case IT_SETATTR:
1255                         rc = mds_getattr_name(2, req);
1256                         /* FIXME: we need to sit down and decide on who should
1257                          * set req->rq_status, who should return negative and
1258                          * positive return values, and what they all mean. */
1259                         if (rc || req->rq_status != 0) {
1260                                 rep->lock_policy_res2 = req->rq_status;
1261                                 RETURN(ELDLM_LOCK_ABORTED);
1262                         }
1263                         break;
1264                 case IT_READDIR|IT_OPEN:
1265                         LBUG();
1266                         break;
1267                 default:
1268                         CERROR("Unhandled intent\n");
1269                         LBUG();
1270                 }
1271
1272                 if (it->opc == IT_UNLINK || it->opc == IT_RMDIR ||
1273                     it->opc == IT_RENAME || it->opc == IT_RENAME2)
1274                         RETURN(ELDLM_LOCK_ABORTED);
1275
1276                 rep->lock_policy_res2 = req->rq_status;
1277                 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1278
1279                 /* If the client is about to open a file that doesn't have an MD
1280                  * stripe record, it's going to need a write lock. */
1281                 if (it->opc & IT_OPEN) {
1282                         struct lov_mds_md *md =
1283                                 lustre_msg_buf(req->rq_repmsg, 2);
1284                         if (md->lmd_easize == 0) {
1285                                 LDLM_DEBUG(lock, "open with no EA; returning PW"
1286                                            " lock");
1287                                 lock->l_req_mode = LCK_PW;
1288                         }
1289                 }
1290
1291                 /* Give the client a lock on the child object, instead of the
1292                  * parent that it requested. */
1293                 new_resid[0] = NTOH__u32(mds_rep->ino);
1294                 if (new_resid[0] == 0)
1295                         LBUG();
1296                 old_res = lock->l_resource->lr_name[0];
1297
1298                 ldlm_lock_change_resource(lock, new_resid);
1299                 if (lock->l_resource == NULL) {
1300                         LBUG();
1301                         RETURN(-ENOMEM);
1302                 }
1303                 LDLM_DEBUG(lock, "intent policy, old res %ld",
1304                            (long)old_res);
1305                 RETURN(ELDLM_LOCK_CHANGED);
1306         } else {
1307                 int size = sizeof(struct ldlm_reply);
1308                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1309                                      &req->rq_repmsg);
1310                 if (rc) {
1311                         LBUG();
1312                         RETURN(-ENOMEM);
1313                 }
1314         }
1315         RETURN(rc);
1316 }
1317
1318
1319 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1320                          int len, void *karg, void *uarg);
1321
1322 /* use obd ops to offer management infrastructure */
1323 static struct obd_ops mds_obd_ops = {
1324         o_connect:     mds_connect,
1325         o_disconnect:  mds_disconnect,
1326         o_setup:       mds_setup,
1327         o_cleanup:     mds_cleanup,
1328         o_iocontrol:   mds_iocontrol
1329 };
1330
1331 static int __init mds_init(void)
1332 {
1333         class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1334         ldlm_register_intent(ldlm_intent_policy);
1335         return 0;
1336 }
1337
1338 static void __exit mds_exit(void)
1339 {
1340         ldlm_unregister_intent();
1341         class_unregister_type(LUSTRE_MDS_NAME);
1342 }
1343
1344 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1345 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1346 MODULE_LICENSE("GPL");
1347
1348 module_init(mds_init);
1349 module_exit(mds_exit);