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                            obd_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                        obd_uuid_t 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         struct obd_export *export = class_conn2export(conn);
 337
 338         ldlm_cancel_locks_for_export(export);
 339         mds_client_free(export);
 340
 341         rc = class_disconnect(conn);
 342         if (!rc)
 343                 MOD_DEC_USE_COUNT;
 344
 345         return rc;
 346 }
 347
 348 static int mds_getstatus(struct ptlrpc_request *req)
 349 {
 350         struct mds_obd *mds = mds_req2mds(req);
 351         struct mds_body *body;
 352         int rc, size = sizeof(*body);
 353         ENTRY;
 354
 355         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 356         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 357                 CERROR("mds: out of memory for message: size=%d\n", size);
 358                 req->rq_status = -ENOMEM;
 359                 RETURN(0);
 360         }
 361
 362         /* Anything we need to do here with the client's trans no or so? */
 363         body = lustre_msg_buf(req->rq_repmsg, 0);
 364         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 365
 366         /* the last_committed and last_xid fields are filled in for all
 367          * replies already - no need to do so here also.
 368          */
 369         RETURN(0);
 370 }
 371
 372 static int mds_getlovinfo(struct ptlrpc_request *req)
 373 {
 374         struct mds_obd *mds = mds_req2mds(req);
 375         struct mds_status_req *streq;
 376         struct lov_desc *desc;
 377         int tgt_count;
 378         int rc, size[2] = {sizeof(*desc)};
 379         ENTRY;
 380
 381         streq = lustre_msg_buf(req->rq_reqmsg, 0);
 382         streq->flags = NTOH__u32(streq->flags);
 383         streq->repbuf = NTOH__u32(streq->repbuf);
 384         size[1] = streq->repbuf;
 385
 386         rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
 387         if (rc) {
 388                 CERROR("mds: out of memory for message: size=%d\n", size[1]);
 389                 req->rq_status = -ENOMEM;
 390                 RETURN(0);
 391         }
 392
 393         desc = lustre_msg_buf(req->rq_repmsg, 0);
 394         rc = mds_get_lovdesc(req->rq_obd, desc);
 395         if (rc != 0 ) {
 396                 CERROR("mds_get_lovdesc error %d", rc);
 397                 req->rq_status = rc;
 398                 RETURN(0);
 399         }
 400
 401         tgt_count = le32_to_cpu(desc->ld_tgt_count);
 402         if (tgt_count * sizeof(obd_uuid_t) > streq->repbuf) {
 403                 CERROR("too many targets, enlarge client buffers\n");
 404                 req->rq_status = -ENOSPC;
 405                 RETURN(0);
 406         }
 407
 408         mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
 409                 tgt_count * sizeof(struct lov_object_id);
 410         rc = mds_get_lovtgts(req->rq_obd, tgt_count,
 411                              lustre_msg_buf(req->rq_repmsg, 1));
 412         if (rc) {
 413                 CERROR("get_lovtgts error %d\n", rc);
 414                 req->rq_status = rc;
 415                 RETURN(0);
 416         }
 417         RETURN(0);
 418 }
 419
 420 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 421                      void *data, __u32 data_len, int flag)
 422 {
 423         int do_ast;
 424         ENTRY;
 425
 426         if (flag == LDLM_CB_CANCELING) {
 427                 /* Don't need to do anything here. */
 428                 RETURN(0);
 429         }
 430
 431         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 432         lock->l_flags |= LDLM_FL_CBPENDING;
 433         do_ast = (!lock->l_readers && !lock->l_writers);
 434         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 435
 436         if (do_ast) {
 437                 struct lustre_handle lockh;
 438                 int rc;
 439
 440                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 441                 ldlm_lock2handle(lock, &lockh);
 442                 rc = ldlm_cli_cancel(&lockh);
 443                 if (rc < 0)
 444                         CERROR("ldlm_cli_cancel: %d\n", rc);
 445         } else
 446                 LDLM_DEBUG(lock, "Lock still has references, will be"
 447                            "cancelled later");
 448         RETURN(0);
 449 }
 450
 451 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
 452                                 struct ptlrpc_request *req,
 453                                 struct mds_body *reqbody, int reply_off)
 454 {
 455         struct mds_body *body;
 456         struct inode *inode = dentry->d_inode;
 457         int rc;
 458         ENTRY;
 459
 460         if (inode == NULL)
 461                 RETURN(-ENOENT);
 462
 463         body = lustre_msg_buf(req->rq_repmsg, reply_off);
 464
 465         mds_pack_inode2fid(&body->fid1, inode);
 466         mds_pack_inode2body(body, inode);
 467
 468         if (S_ISREG(inode->i_mode)) {
 469                 struct lov_mds_md *lmm;
 470
 471                 lmm = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 472                 lmm->lmm_easize = mds->mds_max_mdsize;
 473                 rc = mds_fs_get_md(mds, inode, lmm);
 474
 475                 if (rc < 0) {
 476                         if (rc == -ENODATA)
 477                                 RETURN(0);
 478                         CERROR("mds_fs_get_md failed: %d\n", rc);
 479                         RETURN(rc);
 480                 }
 481                 body->valid |= OBD_MD_FLEASIZE;
 482         } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
 483                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 484                 int len = req->rq_repmsg->buflens[reply_off + 1];
 485
 486                 rc = inode->i_op->readlink(dentry, symname, len);
 487                 if (rc < 0) {
 488                         CERROR("readlink failed: %d\n", rc);
 489                         RETURN(rc);
 490                 } else
 491                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 492
 493                 body->valid |= OBD_MD_LINKNAME;
 494         }
 495         RETURN(0);
 496 }
 497
 498 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 499 {
 500         struct mds_obd *mds = mds_req2mds(req);
 501         struct obd_device *obd = req->rq_export->exp_obd;
 502         struct obd_run_ctxt saved;
 503         struct mds_body *body;
 504         struct dentry *de = NULL, *dchild = NULL;
 505         struct inode *dir;
 506         struct lustre_handle lockh;
 507         char *name;
 508         int namelen, flags, lock_mode, rc = 0;
 509         struct obd_ucred uc;
 510         __u64 res_id[3] = {0, 0, 0};
 511         ENTRY;
 512
 513         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 514
 515         if (req->rq_reqmsg->bufcount <= offset + 1) {
 516                 LBUG();
 517                 GOTO(out_pre_de, rc = -EINVAL);
 518         }
 519
 520         body = lustre_msg_buf(req->rq_reqmsg, offset);
 521         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 522         namelen = req->rq_reqmsg->buflens[offset + 1];
 523         /* requests were at offset 2, replies go back at 1 */
 524         if (offset)
 525                 offset = 1;
 526
 527         uc.ouc_fsuid = body->fsuid;
 528         uc.ouc_fsgid = body->fsgid;
 529         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 530         de = mds_fid2dentry(mds, &body->fid1, NULL);
 531         if (IS_ERR(de)) {
 532                 LBUG();
 533                 GOTO(out_pre_de, rc = -ESTALE);
 534         }
 535
 536         dir = de->d_inode;
 537         CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
 538
 539         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 540         res_id[0] = dir->i_ino;
 541
 542         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 543                              NULL, 0, lock_mode, &lockh);
 544         if (rc == 0) {
 545                 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
 546                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 547                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 548                                       &flags, ldlm_completion_ast,
 549                                       mds_blocking_ast, NULL, 0, &lockh);
 550                 if (rc != ELDLM_OK) {
 551                         CERROR("lock enqueue: err: %d\n", rc);
 552                         GOTO(out_create_de, rc = -EIO);
 553                 }
 554         }
 555         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 556
 557         down(&dir->i_sem);
 558         dchild = lookup_one_len(name, de, namelen - 1);
 559         if (IS_ERR(dchild)) {
 560                 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
 561                 up(&dir->i_sem);
 562                 LBUG();
 563                 GOTO(out_create_dchild, rc = -ESTALE);
 564         }
 565
 566         rc = mds_getattr_internal(mds, dchild, req, body, offset);
 567
 568         EXIT;
 569 out_create_dchild:
 570         l_dput(dchild);
 571         up(&dir->i_sem);
 572         ldlm_lock_decref(&lockh, lock_mode);
 573 out_create_de:
 574         l_dput(de);
 575 out_pre_de:
 576         req->rq_status = rc;
 577         pop_ctxt(&saved);
 578         return 0;
 579 }
 580
 581 static int mds_getattr(int offset, struct ptlrpc_request *req)
 582 {
 583         struct mds_obd *mds = mds_req2mds(req);
 584         struct obd_run_ctxt saved;
 585         struct dentry *de;
 586         struct inode *inode;
 587         struct mds_body *body;
 588         struct obd_ucred uc;
 589         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 590         ENTRY;
 591
 592         body = lustre_msg_buf(req->rq_reqmsg, offset);
 593         uc.ouc_fsuid = body->fsuid;
 594         uc.ouc_fsgid = body->fsgid;
 595         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 596         de = mds_fid2dentry(mds, &body->fid1, NULL);
 597         if (IS_ERR(de)) {
 598                 req->rq_status = -ENOENT;
 599                 rc = 0;
 600                 GOTO(out_pop, PTR_ERR(de));
 601         }
 602
 603         inode = de->d_inode;
 604         if (S_ISREG(body->fid1.f_type)) {
 605                 bufcount = 2;
 606                 size[1] = mds->mds_max_mdsize;
 607         } else if (body->valid & OBD_MD_LINKNAME) {
 608                 bufcount = 2;
 609                 size[1] = MIN(inode->i_size + 1, body->size);
 610                 CDEBUG(D_INODE, "symlink size: %d, reply space: %d\n",
 611                        inode->i_size + 1, body->size);
 612         }
 613
 614         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 615                 CERROR("failed GETATTR_PACK test\n");
 616                 req->rq_status = -ENOMEM;
 617                 GOTO(out, rc = -ENOMEM);
 618         }
 619
 620         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 621                              &req->rq_repmsg);
 622         if (rc) {
 623                 CERROR("out of memory or FAIL_MDS_GETATTR_PACK\n");
 624                 req->rq_status = rc;
 625                 GOTO(out, rc);
 626         }
 627
 628         req->rq_status = mds_getattr_internal(mds, de, req, body, 0);
 629
 630 out:
 631         l_dput(de);
 632 out_pop:
 633         pop_ctxt(&saved);
 634         RETURN(rc);
 635 }
 636
 637 static int mds_statfs(struct ptlrpc_request *req)
 638 {
 639         struct mds_obd *mds = mds_req2mds(req);
 640         struct obd_statfs *osfs;
 641         struct statfs sfs;
 642         int rc, size = sizeof(*osfs);
 643         ENTRY;
 644
 645         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 646                              &req->rq_repmsg);
 647         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 648                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 649                 GOTO(out, rc);
 650         }
 651
 652         rc = vfs_statfs(mds->mds_sb, &sfs);
 653         if (rc) {
 654                 CERROR("mds: statfs failed: rc %d\n", rc);
 655                 GOTO(out, rc);
 656         }
 657         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 658         memset(osfs, 0, size);
 659         statfs_pack(osfs, &sfs);
 660         obd_statfs_pack(osfs, osfs);
 661
 662 out:
 663         req->rq_status = rc;
 664         RETURN(0);
 665 }
 666
 667 static int mds_open(struct ptlrpc_request *req)
 668 {
 669         struct dentry *de;
 670         struct mds_body *body;
 671         struct file *file;
 672         struct vfsmount *mnt;
 673         struct mds_obd *mds = mds_req2mds(req);
 674         struct mds_export_data *med;
 675         __u32 flags;
 676         struct list_head *tmp;
 677         struct mds_file_data *mfd;
 678         int rc, size = sizeof(*body);
 679         ENTRY;
 680
 681         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 682         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 683                 CERROR("mds: out of memory\n");
 684                 req->rq_status = -ENOMEM;
 685                 RETURN(0);
 686         }
 687
 688         body = lustre_msg_buf(req->rq_reqmsg, 0);
 689
 690         /* was this animal open already? */
 691         /* XXX we should only check on re-open, or do a refcount... */
 692         med = &req->rq_export->exp_mds_data;
 693         list_for_each(tmp, &med->med_open_head) {
 694                 struct mds_file_data *fd;
 695                 fd = list_entry(tmp, struct mds_file_data, mfd_list);
 696                 if (body->extra == fd->mfd_clientfd &&
 697                     body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
 698                         CERROR("Re opening "LPD64"\n", body->fid1.id);
 699                         RETURN(0);
 700                 }
 701         }
 702
 703         OBD_ALLOC(mfd, sizeof(*mfd));
 704         if (!mfd) {
 705                 CERROR("mds: out of memory\n");
 706                 req->rq_status = -ENOMEM;
 707                 RETURN(0);
 708         }
 709
 710         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 711         if (IS_ERR(de)) {
 712                 req->rq_status = -ENOENT;
 713                 RETURN(0);
 714         }
 715
 716         /* check if this inode has seen a delayed object creation */
 717         if (req->rq_reqmsg->bufcount > 1) {
 718                 void *handle;
 719                 struct lov_mds_md *lmm;
 720                 struct inode *inode = de->d_inode;
 721                 struct obd_run_ctxt saved;
 722                 struct obd_ucred uc;
 723                 int rc, rc2;
 724
 725                 lmm = lustre_msg_buf(req->rq_reqmsg, 1);
 726
 727                 uc.ouc_fsuid = body->fsuid;
 728                 uc.ouc_fsgid = body->fsgid;
 729                 push_ctxt(&saved, &mds->mds_ctxt, &uc);
 730                 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 731                 if (!handle) {
 732                         pop_ctxt(&saved);
 733                         GOTO(out_md, rc = -ENOMEM);
 734                 }
 735
 736                 rc = mds_fs_set_md(mds, inode, handle, lmm);
 737                 if (!rc)
 738                         rc = mds_update_last_rcvd(mds, handle, req);
 739
 740                 rc2 = mds_fs_commit(mds, de->d_inode, handle);
 741                 if (rc2 && !rc)
 742                         rc = rc2;
 743                 pop_ctxt(&saved);
 744                 if (rc) {
 745 out_md:
 746                         req->rq_status = rc;
 747                         l_dput(de);
 748                         mntput(mnt);
 749                         RETURN(0);
 750                 }
 751         }
 752
 753         flags = body->flags;
 754         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 755         if (IS_ERR(file)) {
 756                 req->rq_status = PTR_ERR(file);
 757                 OBD_FREE(mfd, sizeof(*mfd));
 758                 RETURN(0);
 759         }
 760
 761         file->private_data = mfd;
 762         mfd->mfd_file = file;
 763         mfd->mfd_clientfd = body->extra;
 764         list_add(&mfd->mfd_list, &med->med_open_head);
 765
 766         body = lustre_msg_buf(req->rq_repmsg, 0);
 767         /* FIXME: need to have cookies involved here */
 768         body->extra = (__u64) (unsigned long)file;
 769         RETURN(0);
 770 }
 771
 772 static int mds_close(struct ptlrpc_request *req)
 773 {
 774         struct mds_body *body;
 775         struct file *file;
 776         struct mds_file_data *mfd;
 777         int rc;
 778         ENTRY;
 779
 780         body = lustre_msg_buf(req->rq_reqmsg, 0);
 781
 782         /* FIXME: need to have cookies involved here */
 783         file = (struct file *)(unsigned long)body->extra;
 784         if (!file->f_dentry)
 785                 LBUG();
 786         mfd = (struct mds_file_data *)file->private_data;
 787         list_del(&mfd->mfd_list);
 788         OBD_FREE(mfd, sizeof(*mfd));
 789
 790         req->rq_status = filp_close(file, 0);
 791
 792         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 793                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
 794                 req->rq_status = -ENOMEM;
 795                 RETURN(-ENOMEM);
 796         }
 797
 798         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 799         if (rc) {
 800                 CERROR("mds: lustre_pack_msg: rc = %d\n", rc);
 801                 req->rq_status = rc;
 802         }
 803
 804         RETURN(0);
 805 }
 806
 807 static int mds_readpage(struct ptlrpc_request *req)
 808 {
 809         struct mds_obd *mds = mds_req2mds(req);
 810         struct vfsmount *mnt;
 811         struct dentry *de;
 812         struct file *file;
 813         struct mds_body *body, *repbody;
 814         struct obd_run_ctxt saved;
 815         int rc, size = sizeof(*body);
 816         struct obd_ucred uc;
 817         ENTRY;
 818
 819         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 820         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 821                 CERROR("mds: out of memory\n");
 822                 GOTO(out, rc = -ENOMEM);
 823         }
 824
 825         body = lustre_msg_buf(req->rq_reqmsg, 0);
 826         uc.ouc_fsuid = body->fsuid;
 827         uc.ouc_fsgid = body->fsgid;
 828         push_ctxt(&saved, &mds->mds_ctxt, &uc);
 829         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 830         if (IS_ERR(de))
 831                 GOTO(out_pop, rc = PTR_ERR(de));
 832
 833         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 834
 835         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 836         /* note: in case of an error, dentry_open puts dentry */
 837         if (IS_ERR(file))
 838                 GOTO(out_pop, rc = PTR_ERR(file));
 839
 840         repbody = lustre_msg_buf(req->rq_repmsg, 0);
 841         repbody->size = file->f_dentry->d_inode->i_size;
 842         repbody->valid = OBD_MD_FLSIZE;
 843
 844         /* to make this asynchronous make sure that the handling function
 845            doesn't send a reply when this function completes. Instead a
 846            callback function would send the reply */
 847         /* note: in case of an error, dentry_open puts dentry */
 848         rc = mds_sendpage(req, file, body->size);
 849
 850         filp_close(file, 0);
 851 out_pop:
 852         pop_ctxt(&saved);
 853 out:
 854         req->rq_status = rc;
 855         RETURN(0);
 856 }
 857
 858 int mds_reint(int offset, struct ptlrpc_request *req)
 859 {
 860         int rc;
 861         struct mds_update_record rec;
 862
 863         rc = mds_update_unpack(req, offset, &rec);
 864         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 865                 CERROR("invalid record\n");
 866                 req->rq_status = -EINVAL;
 867                 RETURN(0);
 868         }
 869         /* rc will be used to interrupt a for loop over multiple records */
 870         rc = mds_reint_rec(&rec, offset, req);
 871         return rc;
 872 }
 873
 874 int mds_handle(struct ptlrpc_request *req)
 875 {
 876         int rc;
 877         ENTRY;
 878
 879         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 880         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
 881                 CERROR("lustre_mds: Invalid request\n");
 882                 GOTO(out, rc);
 883         }
 884
 885         if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
 886                 GOTO(out, rc = -ENOTCONN);
 887
 888         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 889                 GOTO(out, rc = -EINVAL);
 890
 891         switch (req->rq_reqmsg->opc) {
 892         case MDS_CONNECT:
 893                 CDEBUG(D_INODE, "connect\n");
 894                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
 895                 rc = target_handle_connect(req);
 896                 break;
 897
 898         case MDS_DISCONNECT:
 899                 CDEBUG(D_INODE, "disconnect\n");
 900                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
 901                 rc = target_handle_disconnect(req);
 902                 goto out;
 903
 904         case MDS_GETSTATUS:
 905                 CDEBUG(D_INODE, "getstatus\n");
 906                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
 907                 rc = mds_getstatus(req);
 908                 break;
 909
 910         case MDS_GETLOVINFO:
 911                 CDEBUG(D_INODE, "getlovinfo\n");
 912                 rc = mds_getlovinfo(req);
 913                 break;
 914
 915         case MDS_GETATTR:
 916                 CDEBUG(D_INODE, "getattr\n");
 917                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 918                 rc = mds_getattr(0, req);
 919                 break;
 920
 921         case MDS_STATFS:
 922                 CDEBUG(D_INODE, "statfs\n");
 923                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
 924                 rc = mds_statfs(req);
 925                 break;
 926
 927         case MDS_READPAGE:
 928                 CDEBUG(D_INODE, "readpage\n");
 929                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 930                 rc = mds_readpage(req);
 931
 932                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 933                         return 0;
 934                 break;
 935
 936         case MDS_REINT: {
 937                 int size = sizeof(struct mds_body);
 938                 CDEBUG(D_INODE, "reint\n");
 939                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 940
 941                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 942                                      &req->rq_repmsg);
 943                 if (rc) {
 944                         req->rq_status = rc;
 945                         break;
 946                 }
 947                 rc = mds_reint(0, req);
 948                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
 949                 break;
 950                 }
 951
 952         case MDS_OPEN:
 953                 CDEBUG(D_INODE, "open\n");
 954                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 955                 rc = mds_open(req);
 956                 break;
 957
 958         case MDS_CLOSE:
 959                 CDEBUG(D_INODE, "close\n");
 960                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 961                 rc = mds_close(req);
 962                 break;
 963
 964         case LDLM_ENQUEUE:
 965                 CDEBUG(D_INODE, "enqueue\n");
 966                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
 967                 rc = ldlm_handle_enqueue(req);
 968                 if (rc)
 969                         break;
 970                 RETURN(0);
 971
 972         case LDLM_CONVERT:
 973                 CDEBUG(D_INODE, "convert\n");
 974                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
 975                 rc = ldlm_handle_convert(req);
 976                 if (rc)
 977                         break;
 978                 RETURN(0);
 979
 980         case LDLM_CANCEL:
 981                 CDEBUG(D_INODE, "cancel\n");
 982                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
 983                 rc = ldlm_handle_cancel(req);
 984                 if (rc)
 985                         break;
 986                 RETURN(0);
 987         case LDLM_BL_CALLBACK:
 988         case LDLM_CP_CALLBACK:
 989                 CDEBUG(D_INODE, "callback\n");
 990                 CERROR("callbacks should not happen on MDS\n");
 991                 LBUG();
 992                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 993                 break;
 994         default:
 995                 rc = ptlrpc_error(req->rq_svc, req);
 996                 RETURN(rc);
 997         }
 998
 999         EXIT;
1000
1001         if (!rc) {
1002                 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1003                 struct mds_obd *mds = mds_req2mds(req);
1004
1005                 req->rq_repmsg->last_xid =
1006                         HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
1007                 req->rq_repmsg->last_committed =
1008                         HTON__u64(mds->mds_last_committed);
1009                 CDEBUG(D_INFO, "last_rcvd ~%Lu, last_committed %Lu, xid %d\n",
1010                        (unsigned long long)mds->mds_last_rcvd,
1011                        (unsigned long long)mds->mds_last_committed,
1012                        cpu_to_le32(req->rq_xid));
1013         }
1014  out:
1015         if (rc) {
1016                 CERROR("mds: processing error (opcode %d): %d\n",
1017                        req->rq_reqmsg->opc, rc);
1018                 ptlrpc_error(req->rq_svc, req);
1019         } else {
1020                 CDEBUG(D_NET, "sending reply\n");
1021                 ptlrpc_reply(req->rq_svc, req);
1022         }
1023         return 0;
1024 }
1025
1026 /* Update the server data on disk.  This stores the new mount_count and
1027  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1028  * then the server last_rcvd value may be less than that of the clients.
1029  * This will alert us that we may need to do client recovery.
1030  *
1031  * Assumes we are already in the server filesystem context.
1032  *
1033  * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
1034  */
1035 static
1036 int mds_update_server_data(struct mds_obd *mds)
1037 {
1038         struct mds_server_data *msd = mds->mds_server_data;
1039         struct file *filp = mds->mds_rcvd_filp;
1040         loff_t off = 0;
1041         int rc;
1042
1043         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1044         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1045
1046         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1047                (unsigned long long)mds->mds_mount_count,
1048                (unsigned long long)mds->mds_last_rcvd);
1049         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1050         if (rc != sizeof(*msd)) {
1051                 CERROR("error writing MDS server data: rc = %d\n", rc);
1052                 if (rc > 0)
1053                         RETURN(-EIO);
1054                 RETURN(rc);
1055         }
1056         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1057         if (rc)
1058                 CERROR("error flushing MDS server data: rc = %d\n", rc);
1059
1060         return 0;
1061 }
1062
1063 /* Do recovery actions for the MDS */
1064 static int mds_recover(struct obd_device *obddev)
1065 {
1066         struct mds_obd *mds = &obddev->u.mds;
1067         struct obd_run_ctxt saved;
1068         int rc;
1069
1070         /* This happens at the end when recovery is complete */
1071         ++mds->mds_mount_count;
1072         push_ctxt(&saved, &mds->mds_ctxt, NULL);
1073         rc = mds_update_server_data(mds);
1074         pop_ctxt(&saved);
1075
1076         return rc;
1077 }
1078
1079 #define MDS_NUM_THREADS 8
1080 /* mount the file system (secretly) */
1081 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1082 {
1083         int i;
1084         struct obd_ioctl_data* data = buf;
1085         struct mds_obd *mds = &obddev->u.mds;
1086         struct vfsmount *mnt;
1087         int rc = 0;
1088         ENTRY;
1089
1090         MOD_INC_USE_COUNT;
1091 #ifdef CONFIG_DEV_RDONLY
1092         dev_clear_rdonly(2);
1093 #endif
1094         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1095                 GOTO(err_dec, rc = -EINVAL);
1096
1097         mds->mds_fstype = strdup(data->ioc_inlbuf2);
1098
1099         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1100         if (IS_ERR(mnt)) {
1101                 rc = PTR_ERR(mnt);
1102                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1103                 GOTO(err_kfree, rc);
1104         }
1105
1106         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1107         if (!mds->mds_sb)
1108                 GOTO(err_put, rc = -ENODEV);
1109
1110         spin_lock_init(&mds->mds_last_lock);
1111         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1112         rc = mds_fs_setup(obddev, mnt);
1113         if (rc) {
1114                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1115                 GOTO(err_put, rc);
1116         }
1117
1118         mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1119                                            MDS_BUFSIZE, MDS_MAXREQSIZE,
1120                                            MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1121                                            "self", mds_handle, "mds");
1122         if (!mds->mds_service) {
1123                 CERROR("failed to start service\n");
1124                 GOTO(err_fs, rc = -EINVAL);
1125         }
1126
1127         obddev->obd_namespace =
1128                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1129         if (obddev->obd_namespace == NULL) {
1130                 mds_cleanup(obddev);
1131                 GOTO(err_svc, rc = -ENOMEM);
1132         }
1133
1134         for (i = 0; i < MDS_NUM_THREADS; i++) {
1135                 char name[32];
1136                 sprintf(name, "lustre_MDS_%02d", i);
1137                 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1138                 if (rc) {
1139                         CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1140                         GOTO(err_thread, rc);
1141                 }
1142         }
1143
1144         rc = mds_recover(obddev);
1145         if (rc)
1146                 GOTO(err_thread, rc);
1147
1148         ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
1149                            "mds_ldlm_client", &obddev->obd_ldlm_client);
1150
1151         RETURN(0);
1152
1153 err_thread:
1154         ptlrpc_stop_all_threads(mds->mds_service);
1155 err_svc:
1156         ptlrpc_unregister_service(mds->mds_service);
1157 err_fs:
1158         mds_fs_cleanup(obddev);
1159 err_put:
1160         unlock_kernel();
1161         mntput(mds->mds_vfsmnt);
1162         mds->mds_sb = 0;
1163         lock_kernel();
1164 err_kfree:
1165         kfree(mds->mds_fstype);
1166 err_dec:
1167         MOD_DEC_USE_COUNT;
1168         return rc;
1169 }
1170
1171 static int mds_cleanup(struct obd_device *obddev)
1172 {
1173         struct super_block *sb;
1174         struct mds_obd *mds = &obddev->u.mds;
1175         struct obd_run_ctxt saved;
1176         ENTRY;
1177
1178         ptlrpc_stop_all_threads(mds->mds_service);
1179         ptlrpc_unregister_service(mds->mds_service);
1180
1181         sb = mds->mds_sb;
1182         if (!mds->mds_sb)
1183                 RETURN(0);
1184
1185         push_ctxt(&saved, &mds->mds_ctxt, NULL);
1186         mds_update_server_data(mds);
1187
1188         if (mds->mds_rcvd_filp) {
1189                 int rc = filp_close(mds->mds_rcvd_filp, 0);
1190                 mds->mds_rcvd_filp = NULL;
1191
1192                 if (rc)
1193                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
1194         }
1195         pop_ctxt(&saved);
1196
1197         unlock_kernel();
1198         mntput(mds->mds_vfsmnt);
1199         mds->mds_sb = 0;
1200         kfree(mds->mds_fstype);
1201
1202         ldlm_namespace_free(obddev->obd_namespace);
1203
1204         lock_kernel();
1205 #ifdef CONFIG_DEV_RDONLY
1206         dev_clear_rdonly(2);
1207 #endif
1208         mds_fs_cleanup(obddev);
1209
1210         MOD_DEC_USE_COUNT;
1211         RETURN(0);
1212 }
1213
1214 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1215                               ldlm_mode_t mode, void *data)
1216 {
1217         struct ptlrpc_request *req = req_cookie;
1218         int rc = 0;
1219         ENTRY;
1220
1221         if (!req_cookie)
1222                 RETURN(0);
1223
1224         if (req->rq_reqmsg->bufcount > 1) {
1225                 /* an intent needs to be considered */
1226                 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1227                 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1228                 struct mds_body *mds_rep;
1229                 struct ldlm_reply *rep;
1230                 __u64 new_resid[3] = {0, 0, 0}, old_res;
1231                 int rc, size[3] = {sizeof(struct ldlm_reply),
1232                                                   sizeof(struct mds_body),
1233                                                   mds->mds_max_mdsize};
1234
1235                 it->opc = NTOH__u64(it->opc);
1236
1237                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1238                            ldlm_it2str(it->opc));
1239
1240                 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1241                                      &req->rq_repmsg);
1242                 if (rc) {
1243                         rc = req->rq_status = -ENOMEM;
1244                         RETURN(rc);
1245                 }
1246
1247                 rep = lustre_msg_buf(req->rq_repmsg, 0);
1248                 rep->lock_policy_res1 = 1;
1249
1250                 /* execute policy */
1251                 switch ((long)it->opc) {
1252                 case IT_CREAT|IT_OPEN:
1253                         rc = mds_reint(2, req);
1254                         if (rc || (req->rq_status != 0 &&
1255                                    req->rq_status != -EEXIST)) {
1256                                 rep->lock_policy_res2 = req->rq_status;
1257                                 RETURN(ELDLM_LOCK_ABORTED);
1258                         }
1259                         break;
1260                 case IT_CREAT:
1261                 case IT_MKDIR:
1262                 case IT_MKNOD:
1263                 case IT_RENAME2:
1264                 case IT_LINK2:
1265                 case IT_RMDIR:
1266                 case IT_SYMLINK:
1267                 case IT_UNLINK:
1268                         rc = mds_reint(2, req);
1269                         if (rc || (req->rq_status != 0 &&
1270                                    req->rq_status != -EISDIR &&
1271                                    req->rq_status != -ENOTDIR)) {
1272                                 rep->lock_policy_res2 = req->rq_status;
1273                                 RETURN(ELDLM_LOCK_ABORTED);
1274                         }
1275                         break;
1276                 case IT_GETATTR:
1277                 case IT_LOOKUP:
1278                 case IT_OPEN:
1279                 case IT_READDIR:
1280                 case IT_READLINK:
1281                 case IT_RENAME:
1282                 case IT_LINK:
1283                 case IT_SETATTR:
1284                         rc = mds_getattr_name(2, req);
1285                         /* FIXME: we need to sit down and decide on who should
1286                          * set req->rq_status, who should return negative and
1287                          * positive return values, and what they all mean. */
1288                         if (rc || req->rq_status != 0) {
1289                                 rep->lock_policy_res2 = req->rq_status;
1290                                 RETURN(ELDLM_LOCK_ABORTED);
1291                         }
1292                         break;
1293                 case IT_READDIR|IT_OPEN:
1294                         LBUG();
1295                         break;
1296                 default:
1297                         CERROR("Unhandled intent\n");
1298                         LBUG();
1299                 }
1300
1301                 /* We don't bother returning a lock to the client for a file
1302                  * or directory we are removing.
1303                  *
1304                  * As for link and rename, there is no reason for the client
1305                  * to get a lock on the target at this point.  If they are
1306                  * going to modify the file/directory later they will get a
1307                  * lock at that time.
1308                  */
1309                 if (it->opc & (IT_UNLINK | IT_RMDIR | IT_LINK | IT_LINK2 |
1310                                IT_RENAME | IT_RENAME2))
1311                         RETURN(ELDLM_LOCK_ABORTED);
1312
1313                 rep->lock_policy_res2 = req->rq_status;
1314                 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1315
1316                 /* If the client is about to open a file that doesn't have an MD
1317                  * stripe record, it's going to need a write lock. */
1318                 if (it->opc & IT_OPEN) {
1319                         struct lov_mds_md *lmm =
1320                                 lustre_msg_buf(req->rq_repmsg, 2);
1321                         if (lmm->lmm_easize == 0) {
1322                                 LDLM_DEBUG(lock, "open with no EA; returning PW"
1323                                            " lock");
1324                                 lock->l_req_mode = LCK_PW;
1325                         }
1326                 }
1327
1328                 /* Give the client a lock on the child object, instead of the
1329                  * parent that it requested. */
1330                 new_resid[0] = NTOH__u32(mds_rep->ino);
1331                 if (new_resid[0] == 0)
1332                         LBUG();
1333                 old_res = lock->l_resource->lr_name[0];
1334
1335                 ldlm_lock_change_resource(lock, new_resid);
1336                 if (lock->l_resource == NULL) {
1337                         LBUG();
1338                         RETURN(-ENOMEM);
1339                 }
1340                 LDLM_DEBUG(lock, "intent policy, old res %ld",
1341                            (long)old_res);
1342                 RETURN(ELDLM_LOCK_CHANGED);
1343         } else {
1344                 int size = sizeof(struct ldlm_reply);
1345                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1346                                      &req->rq_repmsg);
1347                 if (rc) {
1348                         LBUG();
1349                         RETURN(-ENOMEM);
1350                 }
1351         }
1352         RETURN(rc);
1353 }
1354
1355
1356 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1357                          int len, void *karg, void *uarg);
1358
1359 /* use obd ops to offer management infrastructure */
1360 static struct obd_ops mds_obd_ops = {
1361         o_connect:     mds_connect,
1362         o_disconnect:  mds_disconnect,
1363         o_setup:       mds_setup,
1364         o_cleanup:     mds_cleanup,
1365         o_iocontrol:   mds_iocontrol
1366 };
1367
1368 static int __init mds_init(void)
1369 {
1370         class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1371         ldlm_register_intent(ldlm_intent_policy);
1372         return 0;
1373 }
1374
1375 static void __exit mds_exit(void)
1376 {
1377         ldlm_unregister_intent();
1378         class_unregister_type(LUSTRE_MDS_NAME);
1379 }
1380
1381 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1382 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1383 MODULE_LICENSE("GPL");
1384
1385 module_init(mds_init);
1386 module_exit(mds_exit);