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->b_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->b_xid = req->rq_xid;
  88         bulk->b_buf = buf;
  89         bulk->b_buflen = PAGE_SIZE;
  90         desc->b_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->b_waitq, desc->b_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         list_for_each(p, &obd->obd_exports) {
 278                 exp = list_entry(p, struct obd_export, exp_chain);
 279                 mcd = exp->exp_mds_data.med_mcd;
 280                 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
 281                         CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
 282                                cluuid, exp);
 283                         LASSERT(exp->exp_obd == obd);
 284
 285                         exp->exp_rconnh.addr = conn->addr;
 286                         exp->exp_rconnh.cookie = conn->cookie;
 287                         conn->addr = (__u64) (unsigned long)exp;
 288                         conn->cookie = exp->exp_cookie;
 289                         CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
 290                                (long long)conn->addr, (long long)conn->cookie);
 291                         RETURN(0);
 292                 }
 293         }
 294         rc = class_connect(conn, obd, cluuid);
 295         if (rc)
 296                 GOTO(out_dec, rc);
 297         exp = class_conn2export(conn);
 298         LASSERT(exp);
 299
 300         OBD_ALLOC(mcd, sizeof(*mcd));
 301         if (!mcd) {
 302                 CERROR("mds: out of memory for client data\n");
 303                 GOTO(out_export, rc = -ENOMEM);
 304         }
 305         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
 306         exp->exp_mds_data.med_mcd = mcd;
 307         rc = mds_client_add(&exp->exp_mds_data, -1);
 308         if (rc)
 309                 GOTO(out_mdc, rc);
 310
 311         RETURN(0);
 312
 313 out_mdc:
 314         OBD_FREE(mcd, sizeof(*mcd));
 315 out_export:
 316         class_disconnect(conn);
 317 out_dec:
 318         MOD_DEC_USE_COUNT;
 319
 320         return rc;
 321 }
 322
 323 static int mds_disconnect(struct lustre_handle *conn)
 324 {
 325         struct obd_export *exp;
 326         int rc;
 327
 328         exp = class_conn2export(conn);
 329         if (!exp)
 330                 RETURN(-EINVAL);
 331
 332         rc = class_disconnect(conn);
 333         if (!rc)
 334                 MOD_DEC_USE_COUNT;
 335
 336         return rc;
 337 }
 338
 339 static int mds_getstatus(struct ptlrpc_request *req)
 340 {
 341         struct mds_obd *mds = mds_req2mds(req);
 342         struct mds_body *body;
 343         struct mds_export_data *med = &req->rq_export->exp_mds_data;
 344         int rc, size = sizeof(*body);
 345         ENTRY;
 346
 347         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 348         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
 349                 CERROR("mds: out of memory for message: size=%d\n", size);
 350                 req->rq_status = -ENOMEM;
 351                 RETURN(0);
 352         }
 353
 354         body = lustre_msg_buf(req->rq_reqmsg, 0);
 355         mds_unpack_body(body);
 356
 357         /* Anything we need to do here with the client's trans no or so? */
 358         body = lustre_msg_buf(req->rq_repmsg, 0);
 359         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 360
 361         LASSERT(med->med_mcd);
 362
 363         /* mcd_last_xid is is stored in little endian on the disk and
 364            mds_pack_rep_body converts it to network order */
 365         req->rq_repmsg->last_xid = le32_to_cpu(med->med_mcd->mcd_last_xid);
 366         mds_pack_rep_body(req);
 367         RETURN(0);
 368 }
 369
 370 static int mds_getlovinfo(struct ptlrpc_request *req)
 371 {
 372         struct mds_obd *mds = mds_req2mds(req);
 373         struct mds_status_req *streq;
 374         struct lov_desc *desc;
 375         int tgt_count;
 376         int rc, size[2] = {sizeof(*desc)};
 377         ENTRY;
 378
 379         streq = lustre_msg_buf(req->rq_reqmsg, 0);
 380         streq->flags = NTOH__u32(streq->flags);
 381         streq->repbuf = NTOH__u32(streq->repbuf);
 382         size[1] = streq->repbuf;
 383
 384         rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
 385         if (rc) {
 386                 CERROR("mds: out of memory for message: size=%d\n", size[1]);
 387                 req->rq_status = -ENOMEM;
 388                 RETURN(0);
 389         }
 390
 391         desc = lustre_msg_buf(req->rq_repmsg, 0);
 392         rc = mds_get_lovdesc(req->rq_obd, desc);
 393         if (rc != 0 ) {
 394                 CERROR("mds_get_lovdesc error %d", rc);
 395                 req->rq_status = rc;
 396                 RETURN(0);
 397         }
 398
 399         tgt_count = NTOH__u32(desc->ld_tgt_count);
 400         if (tgt_count * sizeof(uuid_t) > streq->repbuf) {
 401                 CERROR("too many targets, enlarge client buffers\n");
 402                 req->rq_status = -ENOSPC;
 403                 RETURN(0);
 404         }
 405
 406         mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
 407                 tgt_count * sizeof(struct lov_object_id);
 408         rc = mds_get_lovtgts(req->rq_obd, tgt_count,
 409                              lustre_msg_buf(req->rq_repmsg, 1));
 410         if (rc) {
 411                 CERROR("get_lovtgts error %d\n", rc);
 412                 req->rq_status = rc;
 413                 RETURN(0);
 414         }
 415         RETURN(0);
 416 }
 417
 418 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
 419                      void *data, __u32 data_len)
 420 {
 421         int do_ast;
 422         ENTRY;
 423
 424         l_lock(&lock->l_resource->lr_namespace->ns_lock);
 425         lock->l_flags |= LDLM_FL_CBPENDING;
 426         do_ast = (!lock->l_readers && !lock->l_writers);
 427         l_unlock(&lock->l_resource->lr_namespace->ns_lock);
 428
 429         if (do_ast) {
 430                 struct lustre_handle lockh;
 431                 int rc;
 432
 433                 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
 434                 ldlm_lock2handle(lock, &lockh);
 435                 rc = ldlm_cli_cancel(&lockh);
 436                 if (rc < 0)
 437                         CERROR("ldlm_cli_cancel: %d\n", rc);
 438         } else
 439                 LDLM_DEBUG(lock, "Lock still has references, will be"
 440                            "cancelled later");
 441         RETURN(0);
 442 }
 443
 444 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
 445                                 struct ptlrpc_request *req,
 446                                 int request_off, int reply_off)
 447 {
 448         struct mds_body *body;
 449         struct inode *inode = dentry->d_inode;
 450         int rc;
 451         ENTRY;
 452
 453         if (inode == NULL)
 454                 RETURN(-ENOENT);
 455
 456         body = lustre_msg_buf(req->rq_repmsg, reply_off);
 457
 458         mds_pack_inode2fid(&body->fid1, inode);
 459         mds_pack_inode2body(body, inode);
 460
 461         if (S_ISREG(inode->i_mode)) {
 462                 struct lov_mds_md *md;
 463
 464                 md = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 465                 md->lmd_easize = mds->mds_max_mdsize;
 466                 rc = mds_fs_get_md(mds, inode, md);
 467
 468                 if (rc < 0) {
 469                         if (rc == -ENODATA)
 470                                 RETURN(0);
 471                         CERROR("mds_fs_get_md failed: %d\n", rc);
 472                         RETURN(rc);
 473                 }
 474                 body->valid |= OBD_MD_FLEASIZE;
 475         } else if ((body->valid & OBD_MD_LINKNAME) && S_ISLNK(inode->i_mode)) {
 476                 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
 477                 int len = req->rq_repmsg->buflens[reply_off + 1];
 478
 479                 rc = inode->i_op->readlink(dentry, symname, len);
 480                 if (rc < 0) {
 481                         CERROR("readlink failed: %d\n", rc);
 482                         RETURN(rc);
 483                 } else
 484                         CDEBUG(D_INODE, "read symlink dest %s\n", symname);
 485
 486                 body->valid |= OBD_MD_LINKNAME;
 487         }
 488         RETURN(0);
 489 }
 490
 491 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 492 {
 493         struct mds_obd *mds = mds_req2mds(req);
 494         struct obd_device *obd = req->rq_export->exp_obd;
 495         struct obd_run_ctxt saved;
 496         struct mds_body *body;
 497         struct dentry *de = NULL, *dchild = NULL;
 498         struct inode *dir;
 499         struct lustre_handle lockh;
 500         char *name;
 501         int namelen, flags, lock_mode, rc = 0, old_offset = offset;
 502         __u64 res_id[3] = {0, 0, 0};
 503         ENTRY;
 504
 505         LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
 506
 507         if (req->rq_reqmsg->bufcount <= offset + 1) {
 508                 LBUG();
 509                 GOTO(out_pre_de, rc = -EINVAL);
 510         }
 511
 512         body = lustre_msg_buf(req->rq_reqmsg, offset);
 513         name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
 514         namelen = req->rq_reqmsg->buflens[offset + 1];
 515         /* requests were at offset 2, replies go back at 1 */
 516         if (offset)
 517                 offset = 1;
 518
 519         push_ctxt(&saved, &mds->mds_ctxt);
 520         de = mds_fid2dentry(mds, &body->fid1, NULL);
 521         if (IS_ERR(de)) {
 522                 LBUG();
 523                 GOTO(out_pre_de, rc = -ESTALE);
 524         }
 525
 526         dir = de->d_inode;
 527         CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
 528
 529         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
 530         res_id[0] = dir->i_ino;
 531
 532         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
 533                              NULL, 0, lock_mode, &lockh);
 534         if (rc == 0) {
 535                 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
 536                 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
 537                                       res_id, LDLM_PLAIN, NULL, 0, lock_mode,
 538                                       &flags, ldlm_completion_ast,
 539                                       mds_blocking_ast, NULL, 0, &lockh);
 540                 if (rc != ELDLM_OK) {
 541                         CERROR("lock enqueue: err: %d\n", rc);
 542                         GOTO(out_create_de, rc = -EIO);
 543                 }
 544         }
 545         ldlm_lock_dump((void *)(unsigned long)lockh.addr);
 546
 547         down(&dir->i_sem);
 548         dchild = lookup_one_len(name, de, namelen - 1);
 549         if (IS_ERR(dchild)) {
 550                 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
 551                 up(&dir->i_sem);
 552                 LBUG();
 553                 GOTO(out_create_dchild, rc = -ESTALE);
 554         }
 555
 556         rc = mds_getattr_internal(mds, dchild, req, old_offset, offset);
 557
 558         EXIT;
 559 out_create_dchild:
 560         l_dput(dchild);
 561         up(&dir->i_sem);
 562         ldlm_lock_decref(&lockh, lock_mode);
 563 out_create_de:
 564         l_dput(de);
 565 out_pre_de:
 566         req->rq_status = rc;
 567         pop_ctxt(&saved);
 568         return 0;
 569 }
 570
 571 static int mds_getattr(int offset, struct ptlrpc_request *req)
 572 {
 573         struct mds_obd *mds = mds_req2mds(req);
 574         struct obd_run_ctxt saved;
 575         struct dentry *de;
 576         struct inode *inode;
 577         struct mds_body *body;
 578         int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
 579         ENTRY;
 580
 581         body = lustre_msg_buf(req->rq_reqmsg, offset);
 582         push_ctxt(&saved, &mds->mds_ctxt);
 583         de = mds_fid2dentry(mds, &body->fid1, NULL);
 584         if (IS_ERR(de)) {
 585                 req->rq_status = -ENOENT;
 586                 GOTO(out_pop, rc = -ENOENT);
 587         }
 588
 589         inode = de->d_inode;
 590         if (S_ISREG(body->fid1.f_type)) {
 591                 bufcount = 2;
 592                 size[1] = mds->mds_max_mdsize;
 593         } else if (body->valid & OBD_MD_LINKNAME) {
 594                 bufcount = 2;
 595                 size[1] = inode->i_size;
 596         }
 597
 598         rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
 599                              &req->rq_repmsg);
 600         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
 601                 CERROR("out of memory or FAIL_MDS_GETATTR_PACK\n");
 602                 req->rq_status = rc;
 603                 GOTO(out, rc = 0);
 604         }
 605
 606         req->rq_status = mds_getattr_internal(mds, de, req, offset, 0);
 607
 608 out:
 609         l_dput(de);
 610 out_pop:
 611         pop_ctxt(&saved);
 612         RETURN(rc);
 613 }
 614
 615 static int mds_statfs(struct ptlrpc_request *req)
 616 {
 617         struct mds_obd *mds = mds_req2mds(req);
 618         struct obd_statfs *osfs;
 619         struct statfs sfs;
 620         int rc, size = sizeof(*osfs);
 621         ENTRY;
 622
 623         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 624                              &req->rq_repmsg);
 625         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
 626                 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
 627                 GOTO(out, rc);
 628         }
 629
 630         rc = vfs_statfs(mds->mds_sb, &sfs);
 631         if (rc) {
 632                 CERROR("mds: statfs failed: rc %d\n", rc);
 633                 GOTO(out, rc);
 634         }
 635         osfs = lustre_msg_buf(req->rq_repmsg, 0);
 636         memset(osfs, 0, size);
 637         obd_statfs_pack(osfs, &sfs);
 638
 639 out:
 640         req->rq_status = rc;
 641         RETURN(0);
 642 }
 643
 644 static int mds_open(struct ptlrpc_request *req)
 645 {
 646         struct dentry *de;
 647         struct mds_body *body;
 648         struct file *file;
 649         struct vfsmount *mnt;
 650         struct mds_obd *mds = mds_req2mds(req);
 651         struct mds_export_data *med;
 652         __u32 flags;
 653         struct list_head *tmp;
 654         struct mds_file_data *mfd;
 655         int rc, size = sizeof(*body);
 656         ENTRY;
 657
 658         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 659         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
 660                 CERROR("mds: out of memory\n");
 661                 req->rq_status = -ENOMEM;
 662                 RETURN(0);
 663         }
 664
 665         body = lustre_msg_buf(req->rq_reqmsg, 0);
 666
 667         /* was this animal open already? */
 668         /* XXX we should only check on re-open, or do a refcount... */
 669         med = &req->rq_export->exp_mds_data;
 670         list_for_each(tmp, &med->med_open_head) {
 671                 struct mds_file_data *fd;
 672                 fd = list_entry(tmp, struct mds_file_data, mfd_list);
 673                 if (body->extra == fd->mfd_clientfd &&
 674                     body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
 675                         CERROR("Re opening %Ld\n", body->fid1.id);
 676                         RETURN(0);
 677                 }
 678         }
 679
 680         OBD_ALLOC(mfd, sizeof(*mfd));
 681         if (!mfd) {
 682                 CERROR("mds: out of memory\n");
 683                 req->rq_status = -ENOMEM;
 684                 RETURN(0);
 685         }
 686
 687         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 688         if (IS_ERR(de)) {
 689                 req->rq_status = -ENOENT;
 690                 RETURN(0);
 691         }
 692
 693         /* check if this inode has seen a delayed object creation */
 694         if (req->rq_reqmsg->bufcount > 1) {
 695                 void *handle;
 696                 struct lov_mds_md *md;
 697                 struct inode *inode = de->d_inode;
 698                 int rc;
 699
 700                 md = lustre_msg_buf(req->rq_reqmsg, 1);
 701
 702                 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
 703                 if (!handle) {
 704                         req->rq_status = -ENOMEM;
 705                         RETURN(0);
 706                 }
 707
 708                 /* XXX error handling */
 709                 rc = mds_fs_set_md(mds, inode, handle, md);
 710                 if (!rc) {
 711                         struct obd_run_ctxt saved;
 712                         push_ctxt(&saved, &mds->mds_ctxt);
 713                         rc = mds_update_last_rcvd(mds, handle, req);
 714                         pop_ctxt(&saved);
 715                 } else {
 716                         req->rq_status = rc;
 717                         RETURN(0);
 718                 }
 719                 /* FIXME: need to return last_rcvd, last_committed */
 720
 721                 /* FIXME: keep rc intact */
 722                 rc = mds_fs_commit(mds, de->d_inode, handle);
 723                 if (rc) {
 724                         req->rq_status = rc;
 725                         RETURN(0);
 726                 }
 727         }
 728
 729         flags = body->flags;
 730         file = dentry_open(de, mnt, flags & ~O_DIRECT);
 731         if (!file || IS_ERR(file)) {
 732                 req->rq_status = -EINVAL;
 733                 OBD_FREE(mfd, sizeof(*mfd));
 734                 RETURN(0);
 735         }
 736
 737         file->private_data = mfd;
 738         mfd->mfd_file = file;
 739         mfd->mfd_clientfd = body->extra;
 740         list_add(&mfd->mfd_list, &med->med_open_head);
 741
 742         body = lustre_msg_buf(req->rq_repmsg, 0);
 743         /* FIXME: need to have cookies involved here */
 744         body->extra = (__u64) (unsigned long)file;
 745         RETURN(0);
 746 }
 747
 748 static int mds_close(struct ptlrpc_request *req)
 749 {
 750         struct dentry *de;
 751         struct mds_body *body;
 752         struct file *file;
 753         struct mds_obd *mds = mds_req2mds(req);
 754         struct vfsmount *mnt;
 755         struct mds_file_data *mfd;
 756         int rc;
 757         ENTRY;
 758
 759         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 760         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
 761                 CERROR("mds: out of memory\n");
 762                 req->rq_status = -ENOMEM;
 763                 RETURN(0);
 764         }
 765
 766         body = lustre_msg_buf(req->rq_reqmsg, 0);
 767         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 768         if (IS_ERR(de)) {
 769                 req->rq_status = -ENOENT;
 770                 RETURN(0);
 771         }
 772
 773         /* FIXME: need to have cookies involved here */
 774         file = (struct file *)(unsigned long)body->extra;
 775         if (!file->f_dentry)
 776                 LBUG();
 777         mfd = (struct mds_file_data *)file->private_data;
 778         list_del(&mfd->mfd_list);
 779         OBD_FREE(mfd, sizeof(*mfd));
 780
 781         req->rq_status = filp_close(file, 0);
 782         l_dput(de);
 783         mntput(mnt);
 784
 785         RETURN(0);
 786 }
 787
 788 static int mds_readpage(struct ptlrpc_request *req)
 789 {
 790         struct mds_obd *mds = mds_req2mds(req);
 791         struct vfsmount *mnt;
 792         struct dentry *de;
 793         struct file *file;
 794         struct mds_body *body;
 795         struct obd_run_ctxt saved;
 796         int rc, size = sizeof(*body);
 797         ENTRY;
 798
 799         rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
 800         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
 801                 CERROR("mds: out of memory\n");
 802                 GOTO(out, rc = -ENOMEM);
 803         }
 804
 805         body = lustre_msg_buf(req->rq_reqmsg, 0);
 806         push_ctxt(&saved, &mds->mds_ctxt);
 807         de = mds_fid2dentry(mds, &body->fid1, &mnt);
 808         if (IS_ERR(de))
 809                 GOTO(out_pop, rc = PTR_ERR(de));
 810
 811         CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
 812
 813         file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
 814         /* note: in case of an error, dentry_open puts dentry */
 815         if (IS_ERR(file))
 816                 GOTO(out_pop, rc = PTR_ERR(file));
 817
 818         /* to make this asynchronous make sure that the handling function
 819            doesn't send a reply when this function completes. Instead a
 820            callback function would send the reply */
 821         rc = mds_sendpage(req, file, body->size);
 822
 823         filp_close(file, 0);
 824 out_pop:
 825         pop_ctxt(&saved);
 826 out:
 827         req->rq_status = rc;
 828         RETURN(0);
 829 }
 830
 831 int mds_reint(int offset, struct ptlrpc_request *req)
 832 {
 833         int rc;
 834         struct mds_update_record rec;
 835
 836         rc = mds_update_unpack(req, offset, &rec);
 837         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
 838                 CERROR("invalid record\n");
 839                 req->rq_status = -EINVAL;
 840                 RETURN(0);
 841         }
 842         /* rc will be used to interrupt a for loop over multiple records */
 843         rc = mds_reint_rec(&rec, offset, req);
 844         return rc;
 845 }
 846
 847 int mds_handle(struct ptlrpc_request *req)
 848 {
 849         int rc;
 850         ENTRY;
 851
 852         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 853         if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
 854                 CERROR("lustre_mds: Invalid request\n");
 855                 GOTO(out, rc);
 856         }
 857
 858         if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
 859                 CERROR("lustre_mds: wrong packet type sent %d\n",
 860                        req->rq_reqmsg->type);
 861                 GOTO(out, rc = -EINVAL);
 862         }
 863
 864         if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
 865                 GOTO(out, rc = -ENOTCONN);
 866
 867         if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
 868                 GOTO(out, rc = -EINVAL);
 869
 870         switch (req->rq_reqmsg->opc) {
 871         case MDS_CONNECT:
 872                 CDEBUG(D_INODE, "connect\n");
 873                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
 874                 rc = target_handle_connect(req);
 875                 break;
 876
 877         case MDS_DISCONNECT:
 878                 CDEBUG(D_INODE, "disconnect\n");
 879                 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
 880                 rc = target_handle_disconnect(req);
 881                 goto out;
 882
 883         case MDS_GETSTATUS:
 884                 CDEBUG(D_INODE, "getstatus\n");
 885                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
 886                 rc = mds_getstatus(req);
 887                 break;
 888
 889         case MDS_GETLOVINFO:
 890                 CDEBUG(D_INODE, "getlovinfo\n");
 891                 rc = mds_getlovinfo(req);
 892                 break;
 893
 894         case MDS_GETATTR:
 895                 CDEBUG(D_INODE, "getattr\n");
 896                 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
 897                 rc = mds_getattr(0, req);
 898                 break;
 899
 900         case MDS_STATFS:
 901                 CDEBUG(D_INODE, "statfs\n");
 902                 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
 903                 rc = mds_statfs(req);
 904                 break;
 905
 906         case MDS_READPAGE:
 907                 CDEBUG(D_INODE, "readpage\n");
 908                 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
 909                 rc = mds_readpage(req);
 910
 911                 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
 912                         return 0;
 913                 break;
 914
 915         case MDS_REINT: {
 916                 int size = sizeof(struct mds_body);
 917                 CDEBUG(D_INODE, "reint\n");
 918                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
 919
 920                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
 921                                      &req->rq_repmsg);
 922                 if (rc) {
 923                         req->rq_status = rc;
 924                         break;
 925                 }
 926                 rc = mds_reint(0, req);
 927                 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
 928                 break;
 929                 }
 930
 931         case MDS_OPEN:
 932                 CDEBUG(D_INODE, "open\n");
 933                 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
 934                 rc = mds_open(req);
 935                 break;
 936
 937         case MDS_CLOSE:
 938                 CDEBUG(D_INODE, "close\n");
 939                 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
 940                 rc = mds_close(req);
 941                 break;
 942
 943         case LDLM_ENQUEUE:
 944                 CDEBUG(D_INODE, "enqueue\n");
 945                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
 946                 rc = ldlm_handle_enqueue(req);
 947                 if (rc)
 948                         break;
 949                 RETURN(0);
 950
 951         case LDLM_CONVERT:
 952                 CDEBUG(D_INODE, "convert\n");
 953                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
 954                 rc = ldlm_handle_convert(req);
 955                 if (rc)
 956                         break;
 957                 RETURN(0);
 958
 959         case LDLM_CANCEL:
 960                 CDEBUG(D_INODE, "cancel\n");
 961                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
 962                 rc = ldlm_handle_cancel(req);
 963                 if (rc)
 964                         break;
 965                 RETURN(0);
 966         case LDLM_BL_CALLBACK:
 967         case LDLM_CP_CALLBACK:
 968                 CDEBUG(D_INODE, "callback\n");
 969                 CERROR("callbacks should not happen on MDS\n");
 970                 LBUG();
 971                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
 972                 break;
 973         default:
 974                 rc = ptlrpc_error(req->rq_svc, req);
 975                 RETURN(rc);
 976         }
 977
 978         EXIT;
 979
 980         if (!rc) {
 981                 struct mds_obd *mds = mds_req2mds(req);
 982                 req->rq_repmsg->last_xid = HTON__u64(mds->mds_last_rcvd);
 983                 req->rq_repmsg->last_committed =
 984                         HTON__u64(mds->mds_last_committed);
 985                 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
 986                        (unsigned long long)mds->mds_last_rcvd,
 987                        (unsigned long long)mds->mds_last_committed,
 988                        cpu_to_le32(req->rq_xid));
 989         }
 990  out:
 991         /* Still not 100% sure whether we should reply with the server
 992          * last_rcvd or that of this client.  I'm not sure it even makes
 993          * a difference on a per-client basis, because last_rcvd is global
 994          * and we are not supposed to allow transactions while in recovery.
 995          */
 996         if (rc) {
 997                 CERROR("mds: processing error %d\n", rc);
 998                 ptlrpc_error(req->rq_svc, req);
 999         } else {
1000                 CDEBUG(D_NET, "sending reply\n");
1001                 ptlrpc_reply(req->rq_svc, req);
1002         }
1003         return 0;
1004 }
1005
1006 /* Update the server data on disk.  This stores the new mount_count and
1007  * also the last_rcvd value to disk.  If we don't have a clean shutdown,
1008  * then the server last_rcvd value may be less than that of the clients.
1009  * This will alert us that we may need to do client recovery.
1010  *
1011  * Assumes we are already in the server filesystem context.
1012  */
1013 static
1014 int mds_update_server_data(struct mds_obd *mds)
1015 {
1016         struct mds_server_data *msd = mds->mds_server_data;
1017         struct file *filp = mds->mds_rcvd_filp;
1018         loff_t off = 0;
1019         int rc;
1020
1021         msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1022         msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1023
1024         CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1025                (unsigned long long)mds->mds_mount_count,
1026                (unsigned long long)mds->mds_last_rcvd);
1027         rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1028         if (rc != sizeof(*msd)) {
1029                 CERROR("error writing MDS server data: rc = %d\n", rc);
1030                 if (rc > 0)
1031                         RETURN(-EIO);
1032                 RETURN(rc);
1033         }
1034         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1035         if (rc)
1036                 CERROR("error flushing MDS server data: rc = %d\n", rc);
1037
1038         return 0;
1039 }
1040
1041 /* Do recovery actions for the MDS */
1042 static int mds_recover(struct obd_device *obddev)
1043 {
1044         struct mds_obd *mds = &obddev->u.mds;
1045         struct obd_run_ctxt saved;
1046         int rc;
1047
1048         /* This happens at the end when recovery is complete */
1049         ++mds->mds_mount_count;
1050         push_ctxt(&saved, &mds->mds_ctxt);
1051         rc = mds_update_server_data(mds);
1052         pop_ctxt(&saved);
1053
1054         return rc;
1055 }
1056
1057 #define MDS_NUM_THREADS 8
1058 /* mount the file system (secretly) */
1059 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1060 {
1061         int i;
1062         struct obd_ioctl_data* data = buf;
1063         struct mds_obd *mds = &obddev->u.mds;
1064         struct vfsmount *mnt;
1065         int rc = 0;
1066         ENTRY;
1067
1068         MOD_INC_USE_COUNT;
1069 #ifdef CONFIG_DEV_RDONLY
1070         dev_clear_rdonly(2);
1071 #endif
1072         if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1073                 GOTO(err_dec, rc = -EINVAL);
1074
1075         mds->mds_fstype = strdup(data->ioc_inlbuf2);
1076
1077         mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1078         if (IS_ERR(mnt)) {
1079                 rc = PTR_ERR(mnt);
1080                 CERROR("do_kern_mount failed: rc = %d\n", rc);
1081                 GOTO(err_kfree, rc);
1082         }
1083
1084         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1085         if (!mds->mds_sb)
1086                 GOTO(err_put, rc = -ENODEV);
1087
1088         mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1089         rc = mds_fs_setup(obddev, mnt);
1090         if (rc) {
1091                 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1092                 GOTO(err_put, rc);
1093         }
1094
1095         mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
1096                                            MDC_REPLY_PORTAL, "self",mds_handle,
1097                                            "mds");
1098         if (!mds->mds_service) {
1099                 CERROR("failed to start service\n");
1100                 GOTO(err_fs, rc = -EINVAL);
1101         }
1102
1103         obddev->obd_namespace =
1104                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1105         if (obddev->obd_namespace == NULL) {
1106                 mds_cleanup(obddev);
1107                 GOTO(err_svc, rc = -ENOMEM);
1108         }
1109
1110         for (i = 0; i < MDS_NUM_THREADS; i++) {
1111                 char name[32];
1112                 sprintf(name, "lustre_MDS_%02d", i);
1113                 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1114                 if (rc) {
1115                         CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1116                         GOTO(err_thread, rc);
1117                 }
1118         }
1119
1120         rc = mds_recover(obddev);
1121         if (rc)
1122                 GOTO(err_thread, rc);
1123
1124         mds_destroy_export = mds_client_free;
1125
1126         RETURN(0);
1127
1128 err_thread:
1129         ptlrpc_stop_all_threads(mds->mds_service);
1130 err_svc:
1131         ptlrpc_unregister_service(mds->mds_service);
1132 err_fs:
1133         mds_fs_cleanup(obddev);
1134 err_put:
1135         unlock_kernel();
1136         mntput(mds->mds_vfsmnt);
1137         mds->mds_sb = 0;
1138         lock_kernel();
1139 err_kfree:
1140         kfree(mds->mds_fstype);
1141 err_dec:
1142         MOD_DEC_USE_COUNT;
1143         return rc;
1144 }
1145
1146 static int mds_cleanup(struct obd_device *obddev)
1147 {
1148         struct super_block *sb;
1149         struct mds_obd *mds = &obddev->u.mds;
1150         struct obd_run_ctxt saved;
1151         ENTRY;
1152
1153         ptlrpc_stop_all_threads(mds->mds_service);
1154         ptlrpc_unregister_service(mds->mds_service);
1155
1156         sb = mds->mds_sb;
1157         if (!mds->mds_sb)
1158                 RETURN(0);
1159
1160         push_ctxt(&saved, &mds->mds_ctxt);
1161         mds_update_server_data(mds);
1162
1163         if (mds->mds_rcvd_filp) {
1164                 int rc = filp_close(mds->mds_rcvd_filp, 0);
1165                 mds->mds_rcvd_filp = NULL;
1166
1167                 if (rc)
1168                         CERROR("last_rcvd file won't close, rc=%d\n", rc);
1169         }
1170         pop_ctxt(&saved);
1171
1172         unlock_kernel();
1173         mntput(mds->mds_vfsmnt);
1174         mds->mds_sb = 0;
1175         kfree(mds->mds_fstype);
1176
1177         ldlm_namespace_free(obddev->obd_namespace);
1178
1179         lock_kernel();
1180 #ifdef CONFIG_DEV_RDONLY
1181         dev_clear_rdonly(2);
1182 #endif
1183         mds_fs_cleanup(obddev);
1184
1185         MOD_DEC_USE_COUNT;
1186         RETURN(0);
1187 }
1188
1189 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1190                               ldlm_mode_t mode, void *data)
1191 {
1192         struct ptlrpc_request *req = req_cookie;
1193         int rc = 0;
1194         ENTRY;
1195
1196         if (!req_cookie)
1197                 RETURN(0);
1198
1199         if (req->rq_reqmsg->bufcount > 1) {
1200                 /* an intent needs to be considered */
1201                 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1202                 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1203                 struct mds_body *mds_rep;
1204                 struct ldlm_reply *rep;
1205                 __u64 new_resid[3] = {0, 0, 0}, old_res;
1206                 int rc, size[3] = {sizeof(struct ldlm_reply),
1207                                                   sizeof(struct mds_body),
1208                                                   mds->mds_max_mdsize};
1209
1210                 it->opc = NTOH__u64(it->opc);
1211
1212                 LDLM_DEBUG(lock, "intent policy, opc: %s",
1213                            ldlm_it2str(it->opc));
1214
1215                 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1216                                      &req->rq_repmsg);
1217                 if (rc) {
1218                         rc = req->rq_status = -ENOMEM;
1219                         RETURN(rc);
1220                 }
1221
1222                 rep = lustre_msg_buf(req->rq_repmsg, 0);
1223                 rep->lock_policy_res1 = 1;
1224
1225                 /* execute policy */
1226                 switch ((long)it->opc) {
1227                 case IT_CREAT|IT_OPEN:
1228                         rc = mds_reint(2, req);
1229                         if (rc || (req->rq_status != 0 &&
1230                                    req->rq_status != -EEXIST)) {
1231                                 rep->lock_policy_res2 = req->rq_status;
1232                                 RETURN(ELDLM_LOCK_ABORTED);
1233                         }
1234                         break;
1235                 case IT_CREAT:
1236                 case IT_LINK:
1237                 case IT_MKDIR:
1238                 case IT_MKNOD:
1239                 case IT_RENAME2:
1240                 case IT_RMDIR:
1241                 case IT_SYMLINK:
1242                 case IT_UNLINK:
1243                         rc = mds_reint(2, req);
1244                         if (rc || (req->rq_status != 0 &&
1245                                    req->rq_status != -EISDIR &&
1246                                    req->rq_status != -ENOTDIR)) {
1247                                 rep->lock_policy_res2 = req->rq_status;
1248                                 RETURN(ELDLM_LOCK_ABORTED);
1249                         }
1250                         break;
1251                 case IT_GETATTR:
1252                 case IT_LOOKUP:
1253                 case IT_OPEN:
1254                 case IT_READDIR:
1255                 case IT_READLINK:
1256                 case IT_RENAME:
1257                 case IT_SETATTR:
1258                         rc = mds_getattr_name(2, req);
1259                         /* FIXME: we need to sit down and decide on who should
1260                          * set req->rq_status, who should return negative and
1261                          * positive return values, and what they all mean. */
1262                         if (rc || req->rq_status != 0) {
1263                                 rep->lock_policy_res2 = req->rq_status;
1264                                 RETURN(ELDLM_LOCK_ABORTED);
1265                         }
1266                         break;
1267                 case IT_READDIR|IT_OPEN:
1268                         LBUG();
1269                         break;
1270                 default:
1271                         CERROR("Unhandled intent\n");
1272                         LBUG();
1273                 }
1274
1275                 if (it->opc == IT_UNLINK || it->opc == IT_RMDIR ||
1276                     it->opc == IT_RENAME || it->opc == IT_RENAME2)
1277                         RETURN(ELDLM_LOCK_ABORTED);
1278
1279                 rep->lock_policy_res2 = req->rq_status;
1280                 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1281
1282                 /* If the client is about to open a file that doesn't have an MD
1283                  * stripe record, it's going to need a write lock. */
1284                 if (it->opc & IT_OPEN) {
1285                         struct lov_mds_md *md =
1286                                 lustre_msg_buf(req->rq_repmsg, 2);
1287                         if (md->lmd_easize == 0) {
1288                                 LDLM_DEBUG(lock, "open with no EA; returning PW"
1289                                            " lock");
1290                                 lock->l_req_mode = LCK_PW;
1291                         }
1292                 }
1293
1294                 /* Give the client a lock on the child object, instead of the
1295                  * parent that it requested. */
1296                 new_resid[0] = NTOH__u32(mds_rep->ino);
1297                 if (new_resid[0] == 0)
1298                         LBUG();
1299                 old_res = lock->l_resource->lr_name[0];
1300
1301                 ldlm_lock_change_resource(lock, new_resid);
1302                 if (lock->l_resource == NULL) {
1303                         LBUG();
1304                         RETURN(-ENOMEM);
1305                 }
1306                 LDLM_DEBUG(lock, "intent policy, old res %ld",
1307                            (long)old_res);
1308                 RETURN(ELDLM_LOCK_CHANGED);
1309         } else {
1310                 int size = sizeof(struct ldlm_reply);
1311                 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1312                                      &req->rq_repmsg);
1313                 if (rc) {
1314                         LBUG();
1315                         RETURN(-ENOMEM);
1316                 }
1317         }
1318         RETURN(rc);
1319 }
1320
1321
1322 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1323                          int len, void *karg, void *uarg);
1324
1325 /* use obd ops to offer management infrastructure */
1326 static struct obd_ops mds_obd_ops = {
1327         o_connect:     mds_connect,
1328         o_disconnect:  mds_disconnect,
1329         o_setup:       mds_setup,
1330         o_cleanup:     mds_cleanup,
1331         o_iocontrol:   mds_iocontrol
1332 };
1333
1334 static int __init mds_init(void)
1335 {
1336         class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1337         ldlm_register_intent(ldlm_intent_policy);
1338         return 0;
1339 }
1340
1341 static void __exit mds_exit(void)
1342 {
1343         ldlm_unregister_intent();
1344         class_unregister_type(LUSTRE_MDS_NAME);
1345 }
1346
1347 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1348 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1349 MODULE_LICENSE("GPL");
1350
1351 module_init(mds_init);
1352 module_exit(mds_exit);